Commit 5ab551d662396f8437ec5aba12210b7a67eb492b
Exists in
ti-lsk-linux-4.1.y
and in
10 other branches
Merge branch 'sched-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip
Pull scheduler fixes from Ingo Molnar: "Misc fixes: group scheduling corner case fix, two deadline scheduler fixes, effective_load() overflow fix, nested sleep fix, 6144 CPUs system fix" * 'sched-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: sched/fair: Fix RCU stall upon -ENOMEM in sched_create_group() sched/deadline: Avoid double-accounting in case of missed deadlines sched/deadline: Fix migration of SCHED_DEADLINE tasks sched: Fix odd values in effective_load() calculations sched, fanotify: Deal with nested sleeps sched: Fix KMALLOC_MAX_SIZE overflow during cpumask allocation
Showing 4 changed files Inline Diff
fs/notify/fanotify/fanotify_user.c
1 | #include <linux/fanotify.h> | 1 | #include <linux/fanotify.h> |
2 | #include <linux/fcntl.h> | 2 | #include <linux/fcntl.h> |
3 | #include <linux/file.h> | 3 | #include <linux/file.h> |
4 | #include <linux/fs.h> | 4 | #include <linux/fs.h> |
5 | #include <linux/anon_inodes.h> | 5 | #include <linux/anon_inodes.h> |
6 | #include <linux/fsnotify_backend.h> | 6 | #include <linux/fsnotify_backend.h> |
7 | #include <linux/init.h> | 7 | #include <linux/init.h> |
8 | #include <linux/mount.h> | 8 | #include <linux/mount.h> |
9 | #include <linux/namei.h> | 9 | #include <linux/namei.h> |
10 | #include <linux/poll.h> | 10 | #include <linux/poll.h> |
11 | #include <linux/security.h> | 11 | #include <linux/security.h> |
12 | #include <linux/syscalls.h> | 12 | #include <linux/syscalls.h> |
13 | #include <linux/slab.h> | 13 | #include <linux/slab.h> |
14 | #include <linux/types.h> | 14 | #include <linux/types.h> |
15 | #include <linux/uaccess.h> | 15 | #include <linux/uaccess.h> |
16 | #include <linux/compat.h> | 16 | #include <linux/compat.h> |
17 | 17 | ||
18 | #include <asm/ioctls.h> | 18 | #include <asm/ioctls.h> |
19 | 19 | ||
20 | #include "../../mount.h" | 20 | #include "../../mount.h" |
21 | #include "../fdinfo.h" | 21 | #include "../fdinfo.h" |
22 | #include "fanotify.h" | 22 | #include "fanotify.h" |
23 | 23 | ||
24 | #define FANOTIFY_DEFAULT_MAX_EVENTS 16384 | 24 | #define FANOTIFY_DEFAULT_MAX_EVENTS 16384 |
25 | #define FANOTIFY_DEFAULT_MAX_MARKS 8192 | 25 | #define FANOTIFY_DEFAULT_MAX_MARKS 8192 |
26 | #define FANOTIFY_DEFAULT_MAX_LISTENERS 128 | 26 | #define FANOTIFY_DEFAULT_MAX_LISTENERS 128 |
27 | 27 | ||
28 | /* | 28 | /* |
29 | * All flags that may be specified in parameter event_f_flags of fanotify_init. | 29 | * All flags that may be specified in parameter event_f_flags of fanotify_init. |
30 | * | 30 | * |
31 | * Internal and external open flags are stored together in field f_flags of | 31 | * Internal and external open flags are stored together in field f_flags of |
32 | * struct file. Only external open flags shall be allowed in event_f_flags. | 32 | * struct file. Only external open flags shall be allowed in event_f_flags. |
33 | * Internal flags like FMODE_NONOTIFY, FMODE_EXEC, FMODE_NOCMTIME shall be | 33 | * Internal flags like FMODE_NONOTIFY, FMODE_EXEC, FMODE_NOCMTIME shall be |
34 | * excluded. | 34 | * excluded. |
35 | */ | 35 | */ |
36 | #define FANOTIFY_INIT_ALL_EVENT_F_BITS ( \ | 36 | #define FANOTIFY_INIT_ALL_EVENT_F_BITS ( \ |
37 | O_ACCMODE | O_APPEND | O_NONBLOCK | \ | 37 | O_ACCMODE | O_APPEND | O_NONBLOCK | \ |
38 | __O_SYNC | O_DSYNC | O_CLOEXEC | \ | 38 | __O_SYNC | O_DSYNC | O_CLOEXEC | \ |
39 | O_LARGEFILE | O_NOATIME ) | 39 | O_LARGEFILE | O_NOATIME ) |
40 | 40 | ||
41 | extern const struct fsnotify_ops fanotify_fsnotify_ops; | 41 | extern const struct fsnotify_ops fanotify_fsnotify_ops; |
42 | 42 | ||
43 | static struct kmem_cache *fanotify_mark_cache __read_mostly; | 43 | static struct kmem_cache *fanotify_mark_cache __read_mostly; |
44 | struct kmem_cache *fanotify_event_cachep __read_mostly; | 44 | struct kmem_cache *fanotify_event_cachep __read_mostly; |
45 | struct kmem_cache *fanotify_perm_event_cachep __read_mostly; | 45 | struct kmem_cache *fanotify_perm_event_cachep __read_mostly; |
46 | 46 | ||
47 | /* | 47 | /* |
48 | * Get an fsnotify notification event if one exists and is small | 48 | * Get an fsnotify notification event if one exists and is small |
49 | * enough to fit in "count". Return an error pointer if the count | 49 | * enough to fit in "count". Return an error pointer if the count |
50 | * is not large enough. | 50 | * is not large enough. |
51 | * | 51 | * |
52 | * Called with the group->notification_mutex held. | 52 | * Called with the group->notification_mutex held. |
53 | */ | 53 | */ |
54 | static struct fsnotify_event *get_one_event(struct fsnotify_group *group, | 54 | static struct fsnotify_event *get_one_event(struct fsnotify_group *group, |
55 | size_t count) | 55 | size_t count) |
56 | { | 56 | { |
57 | BUG_ON(!mutex_is_locked(&group->notification_mutex)); | 57 | BUG_ON(!mutex_is_locked(&group->notification_mutex)); |
58 | 58 | ||
59 | pr_debug("%s: group=%p count=%zd\n", __func__, group, count); | 59 | pr_debug("%s: group=%p count=%zd\n", __func__, group, count); |
60 | 60 | ||
61 | if (fsnotify_notify_queue_is_empty(group)) | 61 | if (fsnotify_notify_queue_is_empty(group)) |
62 | return NULL; | 62 | return NULL; |
63 | 63 | ||
64 | if (FAN_EVENT_METADATA_LEN > count) | 64 | if (FAN_EVENT_METADATA_LEN > count) |
65 | return ERR_PTR(-EINVAL); | 65 | return ERR_PTR(-EINVAL); |
66 | 66 | ||
67 | /* held the notification_mutex the whole time, so this is the | 67 | /* held the notification_mutex the whole time, so this is the |
68 | * same event we peeked above */ | 68 | * same event we peeked above */ |
69 | return fsnotify_remove_first_event(group); | 69 | return fsnotify_remove_first_event(group); |
70 | } | 70 | } |
71 | 71 | ||
72 | static int create_fd(struct fsnotify_group *group, | 72 | static int create_fd(struct fsnotify_group *group, |
73 | struct fanotify_event_info *event, | 73 | struct fanotify_event_info *event, |
74 | struct file **file) | 74 | struct file **file) |
75 | { | 75 | { |
76 | int client_fd; | 76 | int client_fd; |
77 | struct file *new_file; | 77 | struct file *new_file; |
78 | 78 | ||
79 | pr_debug("%s: group=%p event=%p\n", __func__, group, event); | 79 | pr_debug("%s: group=%p event=%p\n", __func__, group, event); |
80 | 80 | ||
81 | client_fd = get_unused_fd_flags(group->fanotify_data.f_flags); | 81 | client_fd = get_unused_fd_flags(group->fanotify_data.f_flags); |
82 | if (client_fd < 0) | 82 | if (client_fd < 0) |
83 | return client_fd; | 83 | return client_fd; |
84 | 84 | ||
85 | /* | 85 | /* |
86 | * we need a new file handle for the userspace program so it can read even if it was | 86 | * we need a new file handle for the userspace program so it can read even if it was |
87 | * originally opened O_WRONLY. | 87 | * originally opened O_WRONLY. |
88 | */ | 88 | */ |
89 | /* it's possible this event was an overflow event. in that case dentry and mnt | 89 | /* it's possible this event was an overflow event. in that case dentry and mnt |
90 | * are NULL; That's fine, just don't call dentry open */ | 90 | * are NULL; That's fine, just don't call dentry open */ |
91 | if (event->path.dentry && event->path.mnt) | 91 | if (event->path.dentry && event->path.mnt) |
92 | new_file = dentry_open(&event->path, | 92 | new_file = dentry_open(&event->path, |
93 | group->fanotify_data.f_flags | FMODE_NONOTIFY, | 93 | group->fanotify_data.f_flags | FMODE_NONOTIFY, |
94 | current_cred()); | 94 | current_cred()); |
95 | else | 95 | else |
96 | new_file = ERR_PTR(-EOVERFLOW); | 96 | new_file = ERR_PTR(-EOVERFLOW); |
97 | if (IS_ERR(new_file)) { | 97 | if (IS_ERR(new_file)) { |
98 | /* | 98 | /* |
99 | * we still send an event even if we can't open the file. this | 99 | * we still send an event even if we can't open the file. this |
100 | * can happen when say tasks are gone and we try to open their | 100 | * can happen when say tasks are gone and we try to open their |
101 | * /proc files or we try to open a WRONLY file like in sysfs | 101 | * /proc files or we try to open a WRONLY file like in sysfs |
102 | * we just send the errno to userspace since there isn't much | 102 | * we just send the errno to userspace since there isn't much |
103 | * else we can do. | 103 | * else we can do. |
104 | */ | 104 | */ |
105 | put_unused_fd(client_fd); | 105 | put_unused_fd(client_fd); |
106 | client_fd = PTR_ERR(new_file); | 106 | client_fd = PTR_ERR(new_file); |
107 | } else { | 107 | } else { |
108 | *file = new_file; | 108 | *file = new_file; |
109 | } | 109 | } |
110 | 110 | ||
111 | return client_fd; | 111 | return client_fd; |
112 | } | 112 | } |
113 | 113 | ||
114 | static int fill_event_metadata(struct fsnotify_group *group, | 114 | static int fill_event_metadata(struct fsnotify_group *group, |
115 | struct fanotify_event_metadata *metadata, | 115 | struct fanotify_event_metadata *metadata, |
116 | struct fsnotify_event *fsn_event, | 116 | struct fsnotify_event *fsn_event, |
117 | struct file **file) | 117 | struct file **file) |
118 | { | 118 | { |
119 | int ret = 0; | 119 | int ret = 0; |
120 | struct fanotify_event_info *event; | 120 | struct fanotify_event_info *event; |
121 | 121 | ||
122 | pr_debug("%s: group=%p metadata=%p event=%p\n", __func__, | 122 | pr_debug("%s: group=%p metadata=%p event=%p\n", __func__, |
123 | group, metadata, fsn_event); | 123 | group, metadata, fsn_event); |
124 | 124 | ||
125 | *file = NULL; | 125 | *file = NULL; |
126 | event = container_of(fsn_event, struct fanotify_event_info, fse); | 126 | event = container_of(fsn_event, struct fanotify_event_info, fse); |
127 | metadata->event_len = FAN_EVENT_METADATA_LEN; | 127 | metadata->event_len = FAN_EVENT_METADATA_LEN; |
128 | metadata->metadata_len = FAN_EVENT_METADATA_LEN; | 128 | metadata->metadata_len = FAN_EVENT_METADATA_LEN; |
129 | metadata->vers = FANOTIFY_METADATA_VERSION; | 129 | metadata->vers = FANOTIFY_METADATA_VERSION; |
130 | metadata->reserved = 0; | 130 | metadata->reserved = 0; |
131 | metadata->mask = fsn_event->mask & FAN_ALL_OUTGOING_EVENTS; | 131 | metadata->mask = fsn_event->mask & FAN_ALL_OUTGOING_EVENTS; |
132 | metadata->pid = pid_vnr(event->tgid); | 132 | metadata->pid = pid_vnr(event->tgid); |
133 | if (unlikely(fsn_event->mask & FAN_Q_OVERFLOW)) | 133 | if (unlikely(fsn_event->mask & FAN_Q_OVERFLOW)) |
134 | metadata->fd = FAN_NOFD; | 134 | metadata->fd = FAN_NOFD; |
135 | else { | 135 | else { |
136 | metadata->fd = create_fd(group, event, file); | 136 | metadata->fd = create_fd(group, event, file); |
137 | if (metadata->fd < 0) | 137 | if (metadata->fd < 0) |
138 | ret = metadata->fd; | 138 | ret = metadata->fd; |
139 | } | 139 | } |
140 | 140 | ||
141 | return ret; | 141 | return ret; |
142 | } | 142 | } |
143 | 143 | ||
144 | #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS | 144 | #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS |
145 | static struct fanotify_perm_event_info *dequeue_event( | 145 | static struct fanotify_perm_event_info *dequeue_event( |
146 | struct fsnotify_group *group, int fd) | 146 | struct fsnotify_group *group, int fd) |
147 | { | 147 | { |
148 | struct fanotify_perm_event_info *event, *return_e = NULL; | 148 | struct fanotify_perm_event_info *event, *return_e = NULL; |
149 | 149 | ||
150 | spin_lock(&group->fanotify_data.access_lock); | 150 | spin_lock(&group->fanotify_data.access_lock); |
151 | list_for_each_entry(event, &group->fanotify_data.access_list, | 151 | list_for_each_entry(event, &group->fanotify_data.access_list, |
152 | fae.fse.list) { | 152 | fae.fse.list) { |
153 | if (event->fd != fd) | 153 | if (event->fd != fd) |
154 | continue; | 154 | continue; |
155 | 155 | ||
156 | list_del_init(&event->fae.fse.list); | 156 | list_del_init(&event->fae.fse.list); |
157 | return_e = event; | 157 | return_e = event; |
158 | break; | 158 | break; |
159 | } | 159 | } |
160 | spin_unlock(&group->fanotify_data.access_lock); | 160 | spin_unlock(&group->fanotify_data.access_lock); |
161 | 161 | ||
162 | pr_debug("%s: found return_re=%p\n", __func__, return_e); | 162 | pr_debug("%s: found return_re=%p\n", __func__, return_e); |
163 | 163 | ||
164 | return return_e; | 164 | return return_e; |
165 | } | 165 | } |
166 | 166 | ||
167 | static int process_access_response(struct fsnotify_group *group, | 167 | static int process_access_response(struct fsnotify_group *group, |
168 | struct fanotify_response *response_struct) | 168 | struct fanotify_response *response_struct) |
169 | { | 169 | { |
170 | struct fanotify_perm_event_info *event; | 170 | struct fanotify_perm_event_info *event; |
171 | int fd = response_struct->fd; | 171 | int fd = response_struct->fd; |
172 | int response = response_struct->response; | 172 | int response = response_struct->response; |
173 | 173 | ||
174 | pr_debug("%s: group=%p fd=%d response=%d\n", __func__, group, | 174 | pr_debug("%s: group=%p fd=%d response=%d\n", __func__, group, |
175 | fd, response); | 175 | fd, response); |
176 | /* | 176 | /* |
177 | * make sure the response is valid, if invalid we do nothing and either | 177 | * make sure the response is valid, if invalid we do nothing and either |
178 | * userspace can send a valid response or we will clean it up after the | 178 | * userspace can send a valid response or we will clean it up after the |
179 | * timeout | 179 | * timeout |
180 | */ | 180 | */ |
181 | switch (response) { | 181 | switch (response) { |
182 | case FAN_ALLOW: | 182 | case FAN_ALLOW: |
183 | case FAN_DENY: | 183 | case FAN_DENY: |
184 | break; | 184 | break; |
185 | default: | 185 | default: |
186 | return -EINVAL; | 186 | return -EINVAL; |
187 | } | 187 | } |
188 | 188 | ||
189 | if (fd < 0) | 189 | if (fd < 0) |
190 | return -EINVAL; | 190 | return -EINVAL; |
191 | 191 | ||
192 | event = dequeue_event(group, fd); | 192 | event = dequeue_event(group, fd); |
193 | if (!event) | 193 | if (!event) |
194 | return -ENOENT; | 194 | return -ENOENT; |
195 | 195 | ||
196 | event->response = response; | 196 | event->response = response; |
197 | wake_up(&group->fanotify_data.access_waitq); | 197 | wake_up(&group->fanotify_data.access_waitq); |
198 | 198 | ||
199 | return 0; | 199 | return 0; |
200 | } | 200 | } |
201 | #endif | 201 | #endif |
202 | 202 | ||
203 | static ssize_t copy_event_to_user(struct fsnotify_group *group, | 203 | static ssize_t copy_event_to_user(struct fsnotify_group *group, |
204 | struct fsnotify_event *event, | 204 | struct fsnotify_event *event, |
205 | char __user *buf) | 205 | char __user *buf) |
206 | { | 206 | { |
207 | struct fanotify_event_metadata fanotify_event_metadata; | 207 | struct fanotify_event_metadata fanotify_event_metadata; |
208 | struct file *f; | 208 | struct file *f; |
209 | int fd, ret; | 209 | int fd, ret; |
210 | 210 | ||
211 | pr_debug("%s: group=%p event=%p\n", __func__, group, event); | 211 | pr_debug("%s: group=%p event=%p\n", __func__, group, event); |
212 | 212 | ||
213 | ret = fill_event_metadata(group, &fanotify_event_metadata, event, &f); | 213 | ret = fill_event_metadata(group, &fanotify_event_metadata, event, &f); |
214 | if (ret < 0) | 214 | if (ret < 0) |
215 | return ret; | 215 | return ret; |
216 | 216 | ||
217 | fd = fanotify_event_metadata.fd; | 217 | fd = fanotify_event_metadata.fd; |
218 | ret = -EFAULT; | 218 | ret = -EFAULT; |
219 | if (copy_to_user(buf, &fanotify_event_metadata, | 219 | if (copy_to_user(buf, &fanotify_event_metadata, |
220 | fanotify_event_metadata.event_len)) | 220 | fanotify_event_metadata.event_len)) |
221 | goto out_close_fd; | 221 | goto out_close_fd; |
222 | 222 | ||
223 | #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS | 223 | #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS |
224 | if (event->mask & FAN_ALL_PERM_EVENTS) | 224 | if (event->mask & FAN_ALL_PERM_EVENTS) |
225 | FANOTIFY_PE(event)->fd = fd; | 225 | FANOTIFY_PE(event)->fd = fd; |
226 | #endif | 226 | #endif |
227 | 227 | ||
228 | if (fd != FAN_NOFD) | 228 | if (fd != FAN_NOFD) |
229 | fd_install(fd, f); | 229 | fd_install(fd, f); |
230 | return fanotify_event_metadata.event_len; | 230 | return fanotify_event_metadata.event_len; |
231 | 231 | ||
232 | out_close_fd: | 232 | out_close_fd: |
233 | if (fd != FAN_NOFD) { | 233 | if (fd != FAN_NOFD) { |
234 | put_unused_fd(fd); | 234 | put_unused_fd(fd); |
235 | fput(f); | 235 | fput(f); |
236 | } | 236 | } |
237 | return ret; | 237 | return ret; |
238 | } | 238 | } |
239 | 239 | ||
240 | /* intofiy userspace file descriptor functions */ | 240 | /* intofiy userspace file descriptor functions */ |
241 | static unsigned int fanotify_poll(struct file *file, poll_table *wait) | 241 | static unsigned int fanotify_poll(struct file *file, poll_table *wait) |
242 | { | 242 | { |
243 | struct fsnotify_group *group = file->private_data; | 243 | struct fsnotify_group *group = file->private_data; |
244 | int ret = 0; | 244 | int ret = 0; |
245 | 245 | ||
246 | poll_wait(file, &group->notification_waitq, wait); | 246 | poll_wait(file, &group->notification_waitq, wait); |
247 | mutex_lock(&group->notification_mutex); | 247 | mutex_lock(&group->notification_mutex); |
248 | if (!fsnotify_notify_queue_is_empty(group)) | 248 | if (!fsnotify_notify_queue_is_empty(group)) |
249 | ret = POLLIN | POLLRDNORM; | 249 | ret = POLLIN | POLLRDNORM; |
250 | mutex_unlock(&group->notification_mutex); | 250 | mutex_unlock(&group->notification_mutex); |
251 | 251 | ||
252 | return ret; | 252 | return ret; |
253 | } | 253 | } |
254 | 254 | ||
255 | static ssize_t fanotify_read(struct file *file, char __user *buf, | 255 | static ssize_t fanotify_read(struct file *file, char __user *buf, |
256 | size_t count, loff_t *pos) | 256 | size_t count, loff_t *pos) |
257 | { | 257 | { |
258 | struct fsnotify_group *group; | 258 | struct fsnotify_group *group; |
259 | struct fsnotify_event *kevent; | 259 | struct fsnotify_event *kevent; |
260 | char __user *start; | 260 | char __user *start; |
261 | int ret; | 261 | int ret; |
262 | DEFINE_WAIT(wait); | 262 | DEFINE_WAIT_FUNC(wait, woken_wake_function); |
263 | 263 | ||
264 | start = buf; | 264 | start = buf; |
265 | group = file->private_data; | 265 | group = file->private_data; |
266 | 266 | ||
267 | pr_debug("%s: group=%p\n", __func__, group); | 267 | pr_debug("%s: group=%p\n", __func__, group); |
268 | 268 | ||
269 | add_wait_queue(&group->notification_waitq, &wait); | ||
269 | while (1) { | 270 | while (1) { |
270 | prepare_to_wait(&group->notification_waitq, &wait, TASK_INTERRUPTIBLE); | ||
271 | |||
272 | mutex_lock(&group->notification_mutex); | 271 | mutex_lock(&group->notification_mutex); |
273 | kevent = get_one_event(group, count); | 272 | kevent = get_one_event(group, count); |
274 | mutex_unlock(&group->notification_mutex); | 273 | mutex_unlock(&group->notification_mutex); |
275 | 274 | ||
276 | if (IS_ERR(kevent)) { | 275 | if (IS_ERR(kevent)) { |
277 | ret = PTR_ERR(kevent); | 276 | ret = PTR_ERR(kevent); |
278 | break; | 277 | break; |
279 | } | 278 | } |
280 | 279 | ||
281 | if (!kevent) { | 280 | if (!kevent) { |
282 | ret = -EAGAIN; | 281 | ret = -EAGAIN; |
283 | if (file->f_flags & O_NONBLOCK) | 282 | if (file->f_flags & O_NONBLOCK) |
284 | break; | 283 | break; |
285 | 284 | ||
286 | ret = -ERESTARTSYS; | 285 | ret = -ERESTARTSYS; |
287 | if (signal_pending(current)) | 286 | if (signal_pending(current)) |
288 | break; | 287 | break; |
289 | 288 | ||
290 | if (start != buf) | 289 | if (start != buf) |
291 | break; | 290 | break; |
292 | schedule(); | 291 | |
292 | wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); | ||
293 | continue; | 293 | continue; |
294 | } | 294 | } |
295 | 295 | ||
296 | ret = copy_event_to_user(group, kevent, buf); | 296 | ret = copy_event_to_user(group, kevent, buf); |
297 | /* | 297 | /* |
298 | * Permission events get queued to wait for response. Other | 298 | * Permission events get queued to wait for response. Other |
299 | * events can be destroyed now. | 299 | * events can be destroyed now. |
300 | */ | 300 | */ |
301 | if (!(kevent->mask & FAN_ALL_PERM_EVENTS)) { | 301 | if (!(kevent->mask & FAN_ALL_PERM_EVENTS)) { |
302 | fsnotify_destroy_event(group, kevent); | 302 | fsnotify_destroy_event(group, kevent); |
303 | if (ret < 0) | 303 | if (ret < 0) |
304 | break; | 304 | break; |
305 | } else { | 305 | } else { |
306 | #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS | 306 | #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS |
307 | if (ret < 0) { | 307 | if (ret < 0) { |
308 | FANOTIFY_PE(kevent)->response = FAN_DENY; | 308 | FANOTIFY_PE(kevent)->response = FAN_DENY; |
309 | wake_up(&group->fanotify_data.access_waitq); | 309 | wake_up(&group->fanotify_data.access_waitq); |
310 | break; | 310 | break; |
311 | } | 311 | } |
312 | spin_lock(&group->fanotify_data.access_lock); | 312 | spin_lock(&group->fanotify_data.access_lock); |
313 | list_add_tail(&kevent->list, | 313 | list_add_tail(&kevent->list, |
314 | &group->fanotify_data.access_list); | 314 | &group->fanotify_data.access_list); |
315 | spin_unlock(&group->fanotify_data.access_lock); | 315 | spin_unlock(&group->fanotify_data.access_lock); |
316 | #endif | 316 | #endif |
317 | } | 317 | } |
318 | buf += ret; | 318 | buf += ret; |
319 | count -= ret; | 319 | count -= ret; |
320 | } | 320 | } |
321 | remove_wait_queue(&group->notification_waitq, &wait); | ||
321 | 322 | ||
322 | finish_wait(&group->notification_waitq, &wait); | ||
323 | if (start != buf && ret != -EFAULT) | 323 | if (start != buf && ret != -EFAULT) |
324 | ret = buf - start; | 324 | ret = buf - start; |
325 | return ret; | 325 | return ret; |
326 | } | 326 | } |
327 | 327 | ||
328 | static ssize_t fanotify_write(struct file *file, const char __user *buf, size_t count, loff_t *pos) | 328 | static ssize_t fanotify_write(struct file *file, const char __user *buf, size_t count, loff_t *pos) |
329 | { | 329 | { |
330 | #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS | 330 | #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS |
331 | struct fanotify_response response = { .fd = -1, .response = -1 }; | 331 | struct fanotify_response response = { .fd = -1, .response = -1 }; |
332 | struct fsnotify_group *group; | 332 | struct fsnotify_group *group; |
333 | int ret; | 333 | int ret; |
334 | 334 | ||
335 | group = file->private_data; | 335 | group = file->private_data; |
336 | 336 | ||
337 | if (count > sizeof(response)) | 337 | if (count > sizeof(response)) |
338 | count = sizeof(response); | 338 | count = sizeof(response); |
339 | 339 | ||
340 | pr_debug("%s: group=%p count=%zu\n", __func__, group, count); | 340 | pr_debug("%s: group=%p count=%zu\n", __func__, group, count); |
341 | 341 | ||
342 | if (copy_from_user(&response, buf, count)) | 342 | if (copy_from_user(&response, buf, count)) |
343 | return -EFAULT; | 343 | return -EFAULT; |
344 | 344 | ||
345 | ret = process_access_response(group, &response); | 345 | ret = process_access_response(group, &response); |
346 | if (ret < 0) | 346 | if (ret < 0) |
347 | count = ret; | 347 | count = ret; |
348 | 348 | ||
349 | return count; | 349 | return count; |
350 | #else | 350 | #else |
351 | return -EINVAL; | 351 | return -EINVAL; |
352 | #endif | 352 | #endif |
353 | } | 353 | } |
354 | 354 | ||
355 | static int fanotify_release(struct inode *ignored, struct file *file) | 355 | static int fanotify_release(struct inode *ignored, struct file *file) |
356 | { | 356 | { |
357 | struct fsnotify_group *group = file->private_data; | 357 | struct fsnotify_group *group = file->private_data; |
358 | 358 | ||
359 | #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS | 359 | #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS |
360 | struct fanotify_perm_event_info *event, *next; | 360 | struct fanotify_perm_event_info *event, *next; |
361 | 361 | ||
362 | /* | 362 | /* |
363 | * There may be still new events arriving in the notification queue | 363 | * There may be still new events arriving in the notification queue |
364 | * but since userspace cannot use fanotify fd anymore, no event can | 364 | * but since userspace cannot use fanotify fd anymore, no event can |
365 | * enter or leave access_list by now. | 365 | * enter or leave access_list by now. |
366 | */ | 366 | */ |
367 | spin_lock(&group->fanotify_data.access_lock); | 367 | spin_lock(&group->fanotify_data.access_lock); |
368 | 368 | ||
369 | atomic_inc(&group->fanotify_data.bypass_perm); | 369 | atomic_inc(&group->fanotify_data.bypass_perm); |
370 | 370 | ||
371 | list_for_each_entry_safe(event, next, &group->fanotify_data.access_list, | 371 | list_for_each_entry_safe(event, next, &group->fanotify_data.access_list, |
372 | fae.fse.list) { | 372 | fae.fse.list) { |
373 | pr_debug("%s: found group=%p event=%p\n", __func__, group, | 373 | pr_debug("%s: found group=%p event=%p\n", __func__, group, |
374 | event); | 374 | event); |
375 | 375 | ||
376 | list_del_init(&event->fae.fse.list); | 376 | list_del_init(&event->fae.fse.list); |
377 | event->response = FAN_ALLOW; | 377 | event->response = FAN_ALLOW; |
378 | } | 378 | } |
379 | spin_unlock(&group->fanotify_data.access_lock); | 379 | spin_unlock(&group->fanotify_data.access_lock); |
380 | 380 | ||
381 | /* | 381 | /* |
382 | * Since bypass_perm is set, newly queued events will not wait for | 382 | * Since bypass_perm is set, newly queued events will not wait for |
383 | * access response. Wake up the already sleeping ones now. | 383 | * access response. Wake up the already sleeping ones now. |
384 | * synchronize_srcu() in fsnotify_destroy_group() will wait for all | 384 | * synchronize_srcu() in fsnotify_destroy_group() will wait for all |
385 | * processes sleeping in fanotify_handle_event() waiting for access | 385 | * processes sleeping in fanotify_handle_event() waiting for access |
386 | * response and thus also for all permission events to be freed. | 386 | * response and thus also for all permission events to be freed. |
387 | */ | 387 | */ |
388 | wake_up(&group->fanotify_data.access_waitq); | 388 | wake_up(&group->fanotify_data.access_waitq); |
389 | #endif | 389 | #endif |
390 | 390 | ||
391 | /* matches the fanotify_init->fsnotify_alloc_group */ | 391 | /* matches the fanotify_init->fsnotify_alloc_group */ |
392 | fsnotify_destroy_group(group); | 392 | fsnotify_destroy_group(group); |
393 | 393 | ||
394 | return 0; | 394 | return 0; |
395 | } | 395 | } |
396 | 396 | ||
397 | static long fanotify_ioctl(struct file *file, unsigned int cmd, unsigned long arg) | 397 | static long fanotify_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
398 | { | 398 | { |
399 | struct fsnotify_group *group; | 399 | struct fsnotify_group *group; |
400 | struct fsnotify_event *fsn_event; | 400 | struct fsnotify_event *fsn_event; |
401 | void __user *p; | 401 | void __user *p; |
402 | int ret = -ENOTTY; | 402 | int ret = -ENOTTY; |
403 | size_t send_len = 0; | 403 | size_t send_len = 0; |
404 | 404 | ||
405 | group = file->private_data; | 405 | group = file->private_data; |
406 | 406 | ||
407 | p = (void __user *) arg; | 407 | p = (void __user *) arg; |
408 | 408 | ||
409 | switch (cmd) { | 409 | switch (cmd) { |
410 | case FIONREAD: | 410 | case FIONREAD: |
411 | mutex_lock(&group->notification_mutex); | 411 | mutex_lock(&group->notification_mutex); |
412 | list_for_each_entry(fsn_event, &group->notification_list, list) | 412 | list_for_each_entry(fsn_event, &group->notification_list, list) |
413 | send_len += FAN_EVENT_METADATA_LEN; | 413 | send_len += FAN_EVENT_METADATA_LEN; |
414 | mutex_unlock(&group->notification_mutex); | 414 | mutex_unlock(&group->notification_mutex); |
415 | ret = put_user(send_len, (int __user *) p); | 415 | ret = put_user(send_len, (int __user *) p); |
416 | break; | 416 | break; |
417 | } | 417 | } |
418 | 418 | ||
419 | return ret; | 419 | return ret; |
420 | } | 420 | } |
421 | 421 | ||
422 | static const struct file_operations fanotify_fops = { | 422 | static const struct file_operations fanotify_fops = { |
423 | .show_fdinfo = fanotify_show_fdinfo, | 423 | .show_fdinfo = fanotify_show_fdinfo, |
424 | .poll = fanotify_poll, | 424 | .poll = fanotify_poll, |
425 | .read = fanotify_read, | 425 | .read = fanotify_read, |
426 | .write = fanotify_write, | 426 | .write = fanotify_write, |
427 | .fasync = NULL, | 427 | .fasync = NULL, |
428 | .release = fanotify_release, | 428 | .release = fanotify_release, |
429 | .unlocked_ioctl = fanotify_ioctl, | 429 | .unlocked_ioctl = fanotify_ioctl, |
430 | .compat_ioctl = fanotify_ioctl, | 430 | .compat_ioctl = fanotify_ioctl, |
431 | .llseek = noop_llseek, | 431 | .llseek = noop_llseek, |
432 | }; | 432 | }; |
433 | 433 | ||
434 | static void fanotify_free_mark(struct fsnotify_mark *fsn_mark) | 434 | static void fanotify_free_mark(struct fsnotify_mark *fsn_mark) |
435 | { | 435 | { |
436 | kmem_cache_free(fanotify_mark_cache, fsn_mark); | 436 | kmem_cache_free(fanotify_mark_cache, fsn_mark); |
437 | } | 437 | } |
438 | 438 | ||
439 | static int fanotify_find_path(int dfd, const char __user *filename, | 439 | static int fanotify_find_path(int dfd, const char __user *filename, |
440 | struct path *path, unsigned int flags) | 440 | struct path *path, unsigned int flags) |
441 | { | 441 | { |
442 | int ret; | 442 | int ret; |
443 | 443 | ||
444 | pr_debug("%s: dfd=%d filename=%p flags=%x\n", __func__, | 444 | pr_debug("%s: dfd=%d filename=%p flags=%x\n", __func__, |
445 | dfd, filename, flags); | 445 | dfd, filename, flags); |
446 | 446 | ||
447 | if (filename == NULL) { | 447 | if (filename == NULL) { |
448 | struct fd f = fdget(dfd); | 448 | struct fd f = fdget(dfd); |
449 | 449 | ||
450 | ret = -EBADF; | 450 | ret = -EBADF; |
451 | if (!f.file) | 451 | if (!f.file) |
452 | goto out; | 452 | goto out; |
453 | 453 | ||
454 | ret = -ENOTDIR; | 454 | ret = -ENOTDIR; |
455 | if ((flags & FAN_MARK_ONLYDIR) && | 455 | if ((flags & FAN_MARK_ONLYDIR) && |
456 | !(S_ISDIR(file_inode(f.file)->i_mode))) { | 456 | !(S_ISDIR(file_inode(f.file)->i_mode))) { |
457 | fdput(f); | 457 | fdput(f); |
458 | goto out; | 458 | goto out; |
459 | } | 459 | } |
460 | 460 | ||
461 | *path = f.file->f_path; | 461 | *path = f.file->f_path; |
462 | path_get(path); | 462 | path_get(path); |
463 | fdput(f); | 463 | fdput(f); |
464 | } else { | 464 | } else { |
465 | unsigned int lookup_flags = 0; | 465 | unsigned int lookup_flags = 0; |
466 | 466 | ||
467 | if (!(flags & FAN_MARK_DONT_FOLLOW)) | 467 | if (!(flags & FAN_MARK_DONT_FOLLOW)) |
468 | lookup_flags |= LOOKUP_FOLLOW; | 468 | lookup_flags |= LOOKUP_FOLLOW; |
469 | if (flags & FAN_MARK_ONLYDIR) | 469 | if (flags & FAN_MARK_ONLYDIR) |
470 | lookup_flags |= LOOKUP_DIRECTORY; | 470 | lookup_flags |= LOOKUP_DIRECTORY; |
471 | 471 | ||
472 | ret = user_path_at(dfd, filename, lookup_flags, path); | 472 | ret = user_path_at(dfd, filename, lookup_flags, path); |
473 | if (ret) | 473 | if (ret) |
474 | goto out; | 474 | goto out; |
475 | } | 475 | } |
476 | 476 | ||
477 | /* you can only watch an inode if you have read permissions on it */ | 477 | /* you can only watch an inode if you have read permissions on it */ |
478 | ret = inode_permission(path->dentry->d_inode, MAY_READ); | 478 | ret = inode_permission(path->dentry->d_inode, MAY_READ); |
479 | if (ret) | 479 | if (ret) |
480 | path_put(path); | 480 | path_put(path); |
481 | out: | 481 | out: |
482 | return ret; | 482 | return ret; |
483 | } | 483 | } |
484 | 484 | ||
485 | static __u32 fanotify_mark_remove_from_mask(struct fsnotify_mark *fsn_mark, | 485 | static __u32 fanotify_mark_remove_from_mask(struct fsnotify_mark *fsn_mark, |
486 | __u32 mask, | 486 | __u32 mask, |
487 | unsigned int flags, | 487 | unsigned int flags, |
488 | int *destroy) | 488 | int *destroy) |
489 | { | 489 | { |
490 | __u32 oldmask; | 490 | __u32 oldmask; |
491 | 491 | ||
492 | spin_lock(&fsn_mark->lock); | 492 | spin_lock(&fsn_mark->lock); |
493 | if (!(flags & FAN_MARK_IGNORED_MASK)) { | 493 | if (!(flags & FAN_MARK_IGNORED_MASK)) { |
494 | oldmask = fsn_mark->mask; | 494 | oldmask = fsn_mark->mask; |
495 | fsnotify_set_mark_mask_locked(fsn_mark, (oldmask & ~mask)); | 495 | fsnotify_set_mark_mask_locked(fsn_mark, (oldmask & ~mask)); |
496 | } else { | 496 | } else { |
497 | oldmask = fsn_mark->ignored_mask; | 497 | oldmask = fsn_mark->ignored_mask; |
498 | fsnotify_set_mark_ignored_mask_locked(fsn_mark, (oldmask & ~mask)); | 498 | fsnotify_set_mark_ignored_mask_locked(fsn_mark, (oldmask & ~mask)); |
499 | } | 499 | } |
500 | spin_unlock(&fsn_mark->lock); | 500 | spin_unlock(&fsn_mark->lock); |
501 | 501 | ||
502 | *destroy = !(oldmask & ~mask); | 502 | *destroy = !(oldmask & ~mask); |
503 | 503 | ||
504 | return mask & oldmask; | 504 | return mask & oldmask; |
505 | } | 505 | } |
506 | 506 | ||
507 | static int fanotify_remove_vfsmount_mark(struct fsnotify_group *group, | 507 | static int fanotify_remove_vfsmount_mark(struct fsnotify_group *group, |
508 | struct vfsmount *mnt, __u32 mask, | 508 | struct vfsmount *mnt, __u32 mask, |
509 | unsigned int flags) | 509 | unsigned int flags) |
510 | { | 510 | { |
511 | struct fsnotify_mark *fsn_mark = NULL; | 511 | struct fsnotify_mark *fsn_mark = NULL; |
512 | __u32 removed; | 512 | __u32 removed; |
513 | int destroy_mark; | 513 | int destroy_mark; |
514 | 514 | ||
515 | mutex_lock(&group->mark_mutex); | 515 | mutex_lock(&group->mark_mutex); |
516 | fsn_mark = fsnotify_find_vfsmount_mark(group, mnt); | 516 | fsn_mark = fsnotify_find_vfsmount_mark(group, mnt); |
517 | if (!fsn_mark) { | 517 | if (!fsn_mark) { |
518 | mutex_unlock(&group->mark_mutex); | 518 | mutex_unlock(&group->mark_mutex); |
519 | return -ENOENT; | 519 | return -ENOENT; |
520 | } | 520 | } |
521 | 521 | ||
522 | removed = fanotify_mark_remove_from_mask(fsn_mark, mask, flags, | 522 | removed = fanotify_mark_remove_from_mask(fsn_mark, mask, flags, |
523 | &destroy_mark); | 523 | &destroy_mark); |
524 | if (destroy_mark) | 524 | if (destroy_mark) |
525 | fsnotify_destroy_mark_locked(fsn_mark, group); | 525 | fsnotify_destroy_mark_locked(fsn_mark, group); |
526 | mutex_unlock(&group->mark_mutex); | 526 | mutex_unlock(&group->mark_mutex); |
527 | 527 | ||
528 | fsnotify_put_mark(fsn_mark); | 528 | fsnotify_put_mark(fsn_mark); |
529 | if (removed & real_mount(mnt)->mnt_fsnotify_mask) | 529 | if (removed & real_mount(mnt)->mnt_fsnotify_mask) |
530 | fsnotify_recalc_vfsmount_mask(mnt); | 530 | fsnotify_recalc_vfsmount_mask(mnt); |
531 | 531 | ||
532 | return 0; | 532 | return 0; |
533 | } | 533 | } |
534 | 534 | ||
535 | static int fanotify_remove_inode_mark(struct fsnotify_group *group, | 535 | static int fanotify_remove_inode_mark(struct fsnotify_group *group, |
536 | struct inode *inode, __u32 mask, | 536 | struct inode *inode, __u32 mask, |
537 | unsigned int flags) | 537 | unsigned int flags) |
538 | { | 538 | { |
539 | struct fsnotify_mark *fsn_mark = NULL; | 539 | struct fsnotify_mark *fsn_mark = NULL; |
540 | __u32 removed; | 540 | __u32 removed; |
541 | int destroy_mark; | 541 | int destroy_mark; |
542 | 542 | ||
543 | mutex_lock(&group->mark_mutex); | 543 | mutex_lock(&group->mark_mutex); |
544 | fsn_mark = fsnotify_find_inode_mark(group, inode); | 544 | fsn_mark = fsnotify_find_inode_mark(group, inode); |
545 | if (!fsn_mark) { | 545 | if (!fsn_mark) { |
546 | mutex_unlock(&group->mark_mutex); | 546 | mutex_unlock(&group->mark_mutex); |
547 | return -ENOENT; | 547 | return -ENOENT; |
548 | } | 548 | } |
549 | 549 | ||
550 | removed = fanotify_mark_remove_from_mask(fsn_mark, mask, flags, | 550 | removed = fanotify_mark_remove_from_mask(fsn_mark, mask, flags, |
551 | &destroy_mark); | 551 | &destroy_mark); |
552 | if (destroy_mark) | 552 | if (destroy_mark) |
553 | fsnotify_destroy_mark_locked(fsn_mark, group); | 553 | fsnotify_destroy_mark_locked(fsn_mark, group); |
554 | mutex_unlock(&group->mark_mutex); | 554 | mutex_unlock(&group->mark_mutex); |
555 | 555 | ||
556 | /* matches the fsnotify_find_inode_mark() */ | 556 | /* matches the fsnotify_find_inode_mark() */ |
557 | fsnotify_put_mark(fsn_mark); | 557 | fsnotify_put_mark(fsn_mark); |
558 | if (removed & inode->i_fsnotify_mask) | 558 | if (removed & inode->i_fsnotify_mask) |
559 | fsnotify_recalc_inode_mask(inode); | 559 | fsnotify_recalc_inode_mask(inode); |
560 | 560 | ||
561 | return 0; | 561 | return 0; |
562 | } | 562 | } |
563 | 563 | ||
564 | static __u32 fanotify_mark_add_to_mask(struct fsnotify_mark *fsn_mark, | 564 | static __u32 fanotify_mark_add_to_mask(struct fsnotify_mark *fsn_mark, |
565 | __u32 mask, | 565 | __u32 mask, |
566 | unsigned int flags) | 566 | unsigned int flags) |
567 | { | 567 | { |
568 | __u32 oldmask = -1; | 568 | __u32 oldmask = -1; |
569 | 569 | ||
570 | spin_lock(&fsn_mark->lock); | 570 | spin_lock(&fsn_mark->lock); |
571 | if (!(flags & FAN_MARK_IGNORED_MASK)) { | 571 | if (!(flags & FAN_MARK_IGNORED_MASK)) { |
572 | oldmask = fsn_mark->mask; | 572 | oldmask = fsn_mark->mask; |
573 | fsnotify_set_mark_mask_locked(fsn_mark, (oldmask | mask)); | 573 | fsnotify_set_mark_mask_locked(fsn_mark, (oldmask | mask)); |
574 | } else { | 574 | } else { |
575 | __u32 tmask = fsn_mark->ignored_mask | mask; | 575 | __u32 tmask = fsn_mark->ignored_mask | mask; |
576 | fsnotify_set_mark_ignored_mask_locked(fsn_mark, tmask); | 576 | fsnotify_set_mark_ignored_mask_locked(fsn_mark, tmask); |
577 | if (flags & FAN_MARK_IGNORED_SURV_MODIFY) | 577 | if (flags & FAN_MARK_IGNORED_SURV_MODIFY) |
578 | fsn_mark->flags |= FSNOTIFY_MARK_FLAG_IGNORED_SURV_MODIFY; | 578 | fsn_mark->flags |= FSNOTIFY_MARK_FLAG_IGNORED_SURV_MODIFY; |
579 | } | 579 | } |
580 | 580 | ||
581 | if (!(flags & FAN_MARK_ONDIR)) { | 581 | if (!(flags & FAN_MARK_ONDIR)) { |
582 | __u32 tmask = fsn_mark->ignored_mask | FAN_ONDIR; | 582 | __u32 tmask = fsn_mark->ignored_mask | FAN_ONDIR; |
583 | fsnotify_set_mark_ignored_mask_locked(fsn_mark, tmask); | 583 | fsnotify_set_mark_ignored_mask_locked(fsn_mark, tmask); |
584 | } | 584 | } |
585 | 585 | ||
586 | spin_unlock(&fsn_mark->lock); | 586 | spin_unlock(&fsn_mark->lock); |
587 | 587 | ||
588 | return mask & ~oldmask; | 588 | return mask & ~oldmask; |
589 | } | 589 | } |
590 | 590 | ||
591 | static struct fsnotify_mark *fanotify_add_new_mark(struct fsnotify_group *group, | 591 | static struct fsnotify_mark *fanotify_add_new_mark(struct fsnotify_group *group, |
592 | struct inode *inode, | 592 | struct inode *inode, |
593 | struct vfsmount *mnt) | 593 | struct vfsmount *mnt) |
594 | { | 594 | { |
595 | struct fsnotify_mark *mark; | 595 | struct fsnotify_mark *mark; |
596 | int ret; | 596 | int ret; |
597 | 597 | ||
598 | if (atomic_read(&group->num_marks) > group->fanotify_data.max_marks) | 598 | if (atomic_read(&group->num_marks) > group->fanotify_data.max_marks) |
599 | return ERR_PTR(-ENOSPC); | 599 | return ERR_PTR(-ENOSPC); |
600 | 600 | ||
601 | mark = kmem_cache_alloc(fanotify_mark_cache, GFP_KERNEL); | 601 | mark = kmem_cache_alloc(fanotify_mark_cache, GFP_KERNEL); |
602 | if (!mark) | 602 | if (!mark) |
603 | return ERR_PTR(-ENOMEM); | 603 | return ERR_PTR(-ENOMEM); |
604 | 604 | ||
605 | fsnotify_init_mark(mark, fanotify_free_mark); | 605 | fsnotify_init_mark(mark, fanotify_free_mark); |
606 | ret = fsnotify_add_mark_locked(mark, group, inode, mnt, 0); | 606 | ret = fsnotify_add_mark_locked(mark, group, inode, mnt, 0); |
607 | if (ret) { | 607 | if (ret) { |
608 | fsnotify_put_mark(mark); | 608 | fsnotify_put_mark(mark); |
609 | return ERR_PTR(ret); | 609 | return ERR_PTR(ret); |
610 | } | 610 | } |
611 | 611 | ||
612 | return mark; | 612 | return mark; |
613 | } | 613 | } |
614 | 614 | ||
615 | 615 | ||
616 | static int fanotify_add_vfsmount_mark(struct fsnotify_group *group, | 616 | static int fanotify_add_vfsmount_mark(struct fsnotify_group *group, |
617 | struct vfsmount *mnt, __u32 mask, | 617 | struct vfsmount *mnt, __u32 mask, |
618 | unsigned int flags) | 618 | unsigned int flags) |
619 | { | 619 | { |
620 | struct fsnotify_mark *fsn_mark; | 620 | struct fsnotify_mark *fsn_mark; |
621 | __u32 added; | 621 | __u32 added; |
622 | 622 | ||
623 | mutex_lock(&group->mark_mutex); | 623 | mutex_lock(&group->mark_mutex); |
624 | fsn_mark = fsnotify_find_vfsmount_mark(group, mnt); | 624 | fsn_mark = fsnotify_find_vfsmount_mark(group, mnt); |
625 | if (!fsn_mark) { | 625 | if (!fsn_mark) { |
626 | fsn_mark = fanotify_add_new_mark(group, NULL, mnt); | 626 | fsn_mark = fanotify_add_new_mark(group, NULL, mnt); |
627 | if (IS_ERR(fsn_mark)) { | 627 | if (IS_ERR(fsn_mark)) { |
628 | mutex_unlock(&group->mark_mutex); | 628 | mutex_unlock(&group->mark_mutex); |
629 | return PTR_ERR(fsn_mark); | 629 | return PTR_ERR(fsn_mark); |
630 | } | 630 | } |
631 | } | 631 | } |
632 | added = fanotify_mark_add_to_mask(fsn_mark, mask, flags); | 632 | added = fanotify_mark_add_to_mask(fsn_mark, mask, flags); |
633 | mutex_unlock(&group->mark_mutex); | 633 | mutex_unlock(&group->mark_mutex); |
634 | 634 | ||
635 | if (added & ~real_mount(mnt)->mnt_fsnotify_mask) | 635 | if (added & ~real_mount(mnt)->mnt_fsnotify_mask) |
636 | fsnotify_recalc_vfsmount_mask(mnt); | 636 | fsnotify_recalc_vfsmount_mask(mnt); |
637 | 637 | ||
638 | fsnotify_put_mark(fsn_mark); | 638 | fsnotify_put_mark(fsn_mark); |
639 | return 0; | 639 | return 0; |
640 | } | 640 | } |
641 | 641 | ||
642 | static int fanotify_add_inode_mark(struct fsnotify_group *group, | 642 | static int fanotify_add_inode_mark(struct fsnotify_group *group, |
643 | struct inode *inode, __u32 mask, | 643 | struct inode *inode, __u32 mask, |
644 | unsigned int flags) | 644 | unsigned int flags) |
645 | { | 645 | { |
646 | struct fsnotify_mark *fsn_mark; | 646 | struct fsnotify_mark *fsn_mark; |
647 | __u32 added; | 647 | __u32 added; |
648 | 648 | ||
649 | pr_debug("%s: group=%p inode=%p\n", __func__, group, inode); | 649 | pr_debug("%s: group=%p inode=%p\n", __func__, group, inode); |
650 | 650 | ||
651 | /* | 651 | /* |
652 | * If some other task has this inode open for write we should not add | 652 | * If some other task has this inode open for write we should not add |
653 | * an ignored mark, unless that ignored mark is supposed to survive | 653 | * an ignored mark, unless that ignored mark is supposed to survive |
654 | * modification changes anyway. | 654 | * modification changes anyway. |
655 | */ | 655 | */ |
656 | if ((flags & FAN_MARK_IGNORED_MASK) && | 656 | if ((flags & FAN_MARK_IGNORED_MASK) && |
657 | !(flags & FAN_MARK_IGNORED_SURV_MODIFY) && | 657 | !(flags & FAN_MARK_IGNORED_SURV_MODIFY) && |
658 | (atomic_read(&inode->i_writecount) > 0)) | 658 | (atomic_read(&inode->i_writecount) > 0)) |
659 | return 0; | 659 | return 0; |
660 | 660 | ||
661 | mutex_lock(&group->mark_mutex); | 661 | mutex_lock(&group->mark_mutex); |
662 | fsn_mark = fsnotify_find_inode_mark(group, inode); | 662 | fsn_mark = fsnotify_find_inode_mark(group, inode); |
663 | if (!fsn_mark) { | 663 | if (!fsn_mark) { |
664 | fsn_mark = fanotify_add_new_mark(group, inode, NULL); | 664 | fsn_mark = fanotify_add_new_mark(group, inode, NULL); |
665 | if (IS_ERR(fsn_mark)) { | 665 | if (IS_ERR(fsn_mark)) { |
666 | mutex_unlock(&group->mark_mutex); | 666 | mutex_unlock(&group->mark_mutex); |
667 | return PTR_ERR(fsn_mark); | 667 | return PTR_ERR(fsn_mark); |
668 | } | 668 | } |
669 | } | 669 | } |
670 | added = fanotify_mark_add_to_mask(fsn_mark, mask, flags); | 670 | added = fanotify_mark_add_to_mask(fsn_mark, mask, flags); |
671 | mutex_unlock(&group->mark_mutex); | 671 | mutex_unlock(&group->mark_mutex); |
672 | 672 | ||
673 | if (added & ~inode->i_fsnotify_mask) | 673 | if (added & ~inode->i_fsnotify_mask) |
674 | fsnotify_recalc_inode_mask(inode); | 674 | fsnotify_recalc_inode_mask(inode); |
675 | 675 | ||
676 | fsnotify_put_mark(fsn_mark); | 676 | fsnotify_put_mark(fsn_mark); |
677 | return 0; | 677 | return 0; |
678 | } | 678 | } |
679 | 679 | ||
680 | /* fanotify syscalls */ | 680 | /* fanotify syscalls */ |
681 | SYSCALL_DEFINE2(fanotify_init, unsigned int, flags, unsigned int, event_f_flags) | 681 | SYSCALL_DEFINE2(fanotify_init, unsigned int, flags, unsigned int, event_f_flags) |
682 | { | 682 | { |
683 | struct fsnotify_group *group; | 683 | struct fsnotify_group *group; |
684 | int f_flags, fd; | 684 | int f_flags, fd; |
685 | struct user_struct *user; | 685 | struct user_struct *user; |
686 | struct fanotify_event_info *oevent; | 686 | struct fanotify_event_info *oevent; |
687 | 687 | ||
688 | pr_debug("%s: flags=%d event_f_flags=%d\n", | 688 | pr_debug("%s: flags=%d event_f_flags=%d\n", |
689 | __func__, flags, event_f_flags); | 689 | __func__, flags, event_f_flags); |
690 | 690 | ||
691 | if (!capable(CAP_SYS_ADMIN)) | 691 | if (!capable(CAP_SYS_ADMIN)) |
692 | return -EPERM; | 692 | return -EPERM; |
693 | 693 | ||
694 | if (flags & ~FAN_ALL_INIT_FLAGS) | 694 | if (flags & ~FAN_ALL_INIT_FLAGS) |
695 | return -EINVAL; | 695 | return -EINVAL; |
696 | 696 | ||
697 | if (event_f_flags & ~FANOTIFY_INIT_ALL_EVENT_F_BITS) | 697 | if (event_f_flags & ~FANOTIFY_INIT_ALL_EVENT_F_BITS) |
698 | return -EINVAL; | 698 | return -EINVAL; |
699 | 699 | ||
700 | switch (event_f_flags & O_ACCMODE) { | 700 | switch (event_f_flags & O_ACCMODE) { |
701 | case O_RDONLY: | 701 | case O_RDONLY: |
702 | case O_RDWR: | 702 | case O_RDWR: |
703 | case O_WRONLY: | 703 | case O_WRONLY: |
704 | break; | 704 | break; |
705 | default: | 705 | default: |
706 | return -EINVAL; | 706 | return -EINVAL; |
707 | } | 707 | } |
708 | 708 | ||
709 | user = get_current_user(); | 709 | user = get_current_user(); |
710 | if (atomic_read(&user->fanotify_listeners) > FANOTIFY_DEFAULT_MAX_LISTENERS) { | 710 | if (atomic_read(&user->fanotify_listeners) > FANOTIFY_DEFAULT_MAX_LISTENERS) { |
711 | free_uid(user); | 711 | free_uid(user); |
712 | return -EMFILE; | 712 | return -EMFILE; |
713 | } | 713 | } |
714 | 714 | ||
715 | f_flags = O_RDWR | FMODE_NONOTIFY; | 715 | f_flags = O_RDWR | FMODE_NONOTIFY; |
716 | if (flags & FAN_CLOEXEC) | 716 | if (flags & FAN_CLOEXEC) |
717 | f_flags |= O_CLOEXEC; | 717 | f_flags |= O_CLOEXEC; |
718 | if (flags & FAN_NONBLOCK) | 718 | if (flags & FAN_NONBLOCK) |
719 | f_flags |= O_NONBLOCK; | 719 | f_flags |= O_NONBLOCK; |
720 | 720 | ||
721 | /* fsnotify_alloc_group takes a ref. Dropped in fanotify_release */ | 721 | /* fsnotify_alloc_group takes a ref. Dropped in fanotify_release */ |
722 | group = fsnotify_alloc_group(&fanotify_fsnotify_ops); | 722 | group = fsnotify_alloc_group(&fanotify_fsnotify_ops); |
723 | if (IS_ERR(group)) { | 723 | if (IS_ERR(group)) { |
724 | free_uid(user); | 724 | free_uid(user); |
725 | return PTR_ERR(group); | 725 | return PTR_ERR(group); |
726 | } | 726 | } |
727 | 727 | ||
728 | group->fanotify_data.user = user; | 728 | group->fanotify_data.user = user; |
729 | atomic_inc(&user->fanotify_listeners); | 729 | atomic_inc(&user->fanotify_listeners); |
730 | 730 | ||
731 | oevent = fanotify_alloc_event(NULL, FS_Q_OVERFLOW, NULL); | 731 | oevent = fanotify_alloc_event(NULL, FS_Q_OVERFLOW, NULL); |
732 | if (unlikely(!oevent)) { | 732 | if (unlikely(!oevent)) { |
733 | fd = -ENOMEM; | 733 | fd = -ENOMEM; |
734 | goto out_destroy_group; | 734 | goto out_destroy_group; |
735 | } | 735 | } |
736 | group->overflow_event = &oevent->fse; | 736 | group->overflow_event = &oevent->fse; |
737 | 737 | ||
738 | if (force_o_largefile()) | 738 | if (force_o_largefile()) |
739 | event_f_flags |= O_LARGEFILE; | 739 | event_f_flags |= O_LARGEFILE; |
740 | group->fanotify_data.f_flags = event_f_flags; | 740 | group->fanotify_data.f_flags = event_f_flags; |
741 | #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS | 741 | #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS |
742 | spin_lock_init(&group->fanotify_data.access_lock); | 742 | spin_lock_init(&group->fanotify_data.access_lock); |
743 | init_waitqueue_head(&group->fanotify_data.access_waitq); | 743 | init_waitqueue_head(&group->fanotify_data.access_waitq); |
744 | INIT_LIST_HEAD(&group->fanotify_data.access_list); | 744 | INIT_LIST_HEAD(&group->fanotify_data.access_list); |
745 | atomic_set(&group->fanotify_data.bypass_perm, 0); | 745 | atomic_set(&group->fanotify_data.bypass_perm, 0); |
746 | #endif | 746 | #endif |
747 | switch (flags & FAN_ALL_CLASS_BITS) { | 747 | switch (flags & FAN_ALL_CLASS_BITS) { |
748 | case FAN_CLASS_NOTIF: | 748 | case FAN_CLASS_NOTIF: |
749 | group->priority = FS_PRIO_0; | 749 | group->priority = FS_PRIO_0; |
750 | break; | 750 | break; |
751 | case FAN_CLASS_CONTENT: | 751 | case FAN_CLASS_CONTENT: |
752 | group->priority = FS_PRIO_1; | 752 | group->priority = FS_PRIO_1; |
753 | break; | 753 | break; |
754 | case FAN_CLASS_PRE_CONTENT: | 754 | case FAN_CLASS_PRE_CONTENT: |
755 | group->priority = FS_PRIO_2; | 755 | group->priority = FS_PRIO_2; |
756 | break; | 756 | break; |
757 | default: | 757 | default: |
758 | fd = -EINVAL; | 758 | fd = -EINVAL; |
759 | goto out_destroy_group; | 759 | goto out_destroy_group; |
760 | } | 760 | } |
761 | 761 | ||
762 | if (flags & FAN_UNLIMITED_QUEUE) { | 762 | if (flags & FAN_UNLIMITED_QUEUE) { |
763 | fd = -EPERM; | 763 | fd = -EPERM; |
764 | if (!capable(CAP_SYS_ADMIN)) | 764 | if (!capable(CAP_SYS_ADMIN)) |
765 | goto out_destroy_group; | 765 | goto out_destroy_group; |
766 | group->max_events = UINT_MAX; | 766 | group->max_events = UINT_MAX; |
767 | } else { | 767 | } else { |
768 | group->max_events = FANOTIFY_DEFAULT_MAX_EVENTS; | 768 | group->max_events = FANOTIFY_DEFAULT_MAX_EVENTS; |
769 | } | 769 | } |
770 | 770 | ||
771 | if (flags & FAN_UNLIMITED_MARKS) { | 771 | if (flags & FAN_UNLIMITED_MARKS) { |
772 | fd = -EPERM; | 772 | fd = -EPERM; |
773 | if (!capable(CAP_SYS_ADMIN)) | 773 | if (!capable(CAP_SYS_ADMIN)) |
774 | goto out_destroy_group; | 774 | goto out_destroy_group; |
775 | group->fanotify_data.max_marks = UINT_MAX; | 775 | group->fanotify_data.max_marks = UINT_MAX; |
776 | } else { | 776 | } else { |
777 | group->fanotify_data.max_marks = FANOTIFY_DEFAULT_MAX_MARKS; | 777 | group->fanotify_data.max_marks = FANOTIFY_DEFAULT_MAX_MARKS; |
778 | } | 778 | } |
779 | 779 | ||
780 | fd = anon_inode_getfd("[fanotify]", &fanotify_fops, group, f_flags); | 780 | fd = anon_inode_getfd("[fanotify]", &fanotify_fops, group, f_flags); |
781 | if (fd < 0) | 781 | if (fd < 0) |
782 | goto out_destroy_group; | 782 | goto out_destroy_group; |
783 | 783 | ||
784 | return fd; | 784 | return fd; |
785 | 785 | ||
786 | out_destroy_group: | 786 | out_destroy_group: |
787 | fsnotify_destroy_group(group); | 787 | fsnotify_destroy_group(group); |
788 | return fd; | 788 | return fd; |
789 | } | 789 | } |
790 | 790 | ||
791 | SYSCALL_DEFINE5(fanotify_mark, int, fanotify_fd, unsigned int, flags, | 791 | SYSCALL_DEFINE5(fanotify_mark, int, fanotify_fd, unsigned int, flags, |
792 | __u64, mask, int, dfd, | 792 | __u64, mask, int, dfd, |
793 | const char __user *, pathname) | 793 | const char __user *, pathname) |
794 | { | 794 | { |
795 | struct inode *inode = NULL; | 795 | struct inode *inode = NULL; |
796 | struct vfsmount *mnt = NULL; | 796 | struct vfsmount *mnt = NULL; |
797 | struct fsnotify_group *group; | 797 | struct fsnotify_group *group; |
798 | struct fd f; | 798 | struct fd f; |
799 | struct path path; | 799 | struct path path; |
800 | int ret; | 800 | int ret; |
801 | 801 | ||
802 | pr_debug("%s: fanotify_fd=%d flags=%x dfd=%d pathname=%p mask=%llx\n", | 802 | pr_debug("%s: fanotify_fd=%d flags=%x dfd=%d pathname=%p mask=%llx\n", |
803 | __func__, fanotify_fd, flags, dfd, pathname, mask); | 803 | __func__, fanotify_fd, flags, dfd, pathname, mask); |
804 | 804 | ||
805 | /* we only use the lower 32 bits as of right now. */ | 805 | /* we only use the lower 32 bits as of right now. */ |
806 | if (mask & ((__u64)0xffffffff << 32)) | 806 | if (mask & ((__u64)0xffffffff << 32)) |
807 | return -EINVAL; | 807 | return -EINVAL; |
808 | 808 | ||
809 | if (flags & ~FAN_ALL_MARK_FLAGS) | 809 | if (flags & ~FAN_ALL_MARK_FLAGS) |
810 | return -EINVAL; | 810 | return -EINVAL; |
811 | switch (flags & (FAN_MARK_ADD | FAN_MARK_REMOVE | FAN_MARK_FLUSH)) { | 811 | switch (flags & (FAN_MARK_ADD | FAN_MARK_REMOVE | FAN_MARK_FLUSH)) { |
812 | case FAN_MARK_ADD: /* fallthrough */ | 812 | case FAN_MARK_ADD: /* fallthrough */ |
813 | case FAN_MARK_REMOVE: | 813 | case FAN_MARK_REMOVE: |
814 | if (!mask) | 814 | if (!mask) |
815 | return -EINVAL; | 815 | return -EINVAL; |
816 | break; | 816 | break; |
817 | case FAN_MARK_FLUSH: | 817 | case FAN_MARK_FLUSH: |
818 | if (flags & ~(FAN_MARK_MOUNT | FAN_MARK_FLUSH)) | 818 | if (flags & ~(FAN_MARK_MOUNT | FAN_MARK_FLUSH)) |
819 | return -EINVAL; | 819 | return -EINVAL; |
820 | break; | 820 | break; |
821 | default: | 821 | default: |
822 | return -EINVAL; | 822 | return -EINVAL; |
823 | } | 823 | } |
824 | 824 | ||
825 | if (mask & FAN_ONDIR) { | 825 | if (mask & FAN_ONDIR) { |
826 | flags |= FAN_MARK_ONDIR; | 826 | flags |= FAN_MARK_ONDIR; |
827 | mask &= ~FAN_ONDIR; | 827 | mask &= ~FAN_ONDIR; |
828 | } | 828 | } |
829 | 829 | ||
830 | #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS | 830 | #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS |
831 | if (mask & ~(FAN_ALL_EVENTS | FAN_ALL_PERM_EVENTS | FAN_EVENT_ON_CHILD)) | 831 | if (mask & ~(FAN_ALL_EVENTS | FAN_ALL_PERM_EVENTS | FAN_EVENT_ON_CHILD)) |
832 | #else | 832 | #else |
833 | if (mask & ~(FAN_ALL_EVENTS | FAN_EVENT_ON_CHILD)) | 833 | if (mask & ~(FAN_ALL_EVENTS | FAN_EVENT_ON_CHILD)) |
834 | #endif | 834 | #endif |
835 | return -EINVAL; | 835 | return -EINVAL; |
836 | 836 | ||
837 | f = fdget(fanotify_fd); | 837 | f = fdget(fanotify_fd); |
838 | if (unlikely(!f.file)) | 838 | if (unlikely(!f.file)) |
839 | return -EBADF; | 839 | return -EBADF; |
840 | 840 | ||
841 | /* verify that this is indeed an fanotify instance */ | 841 | /* verify that this is indeed an fanotify instance */ |
842 | ret = -EINVAL; | 842 | ret = -EINVAL; |
843 | if (unlikely(f.file->f_op != &fanotify_fops)) | 843 | if (unlikely(f.file->f_op != &fanotify_fops)) |
844 | goto fput_and_out; | 844 | goto fput_and_out; |
845 | group = f.file->private_data; | 845 | group = f.file->private_data; |
846 | 846 | ||
847 | /* | 847 | /* |
848 | * group->priority == FS_PRIO_0 == FAN_CLASS_NOTIF. These are not | 848 | * group->priority == FS_PRIO_0 == FAN_CLASS_NOTIF. These are not |
849 | * allowed to set permissions events. | 849 | * allowed to set permissions events. |
850 | */ | 850 | */ |
851 | ret = -EINVAL; | 851 | ret = -EINVAL; |
852 | if (mask & FAN_ALL_PERM_EVENTS && | 852 | if (mask & FAN_ALL_PERM_EVENTS && |
853 | group->priority == FS_PRIO_0) | 853 | group->priority == FS_PRIO_0) |
854 | goto fput_and_out; | 854 | goto fput_and_out; |
855 | 855 | ||
856 | if (flags & FAN_MARK_FLUSH) { | 856 | if (flags & FAN_MARK_FLUSH) { |
857 | ret = 0; | 857 | ret = 0; |
858 | if (flags & FAN_MARK_MOUNT) | 858 | if (flags & FAN_MARK_MOUNT) |
859 | fsnotify_clear_vfsmount_marks_by_group(group); | 859 | fsnotify_clear_vfsmount_marks_by_group(group); |
860 | else | 860 | else |
861 | fsnotify_clear_inode_marks_by_group(group); | 861 | fsnotify_clear_inode_marks_by_group(group); |
862 | goto fput_and_out; | 862 | goto fput_and_out; |
863 | } | 863 | } |
864 | 864 | ||
865 | ret = fanotify_find_path(dfd, pathname, &path, flags); | 865 | ret = fanotify_find_path(dfd, pathname, &path, flags); |
866 | if (ret) | 866 | if (ret) |
867 | goto fput_and_out; | 867 | goto fput_and_out; |
868 | 868 | ||
869 | /* inode held in place by reference to path; group by fget on fd */ | 869 | /* inode held in place by reference to path; group by fget on fd */ |
870 | if (!(flags & FAN_MARK_MOUNT)) | 870 | if (!(flags & FAN_MARK_MOUNT)) |
871 | inode = path.dentry->d_inode; | 871 | inode = path.dentry->d_inode; |
872 | else | 872 | else |
873 | mnt = path.mnt; | 873 | mnt = path.mnt; |
874 | 874 | ||
875 | /* create/update an inode mark */ | 875 | /* create/update an inode mark */ |
876 | switch (flags & (FAN_MARK_ADD | FAN_MARK_REMOVE)) { | 876 | switch (flags & (FAN_MARK_ADD | FAN_MARK_REMOVE)) { |
877 | case FAN_MARK_ADD: | 877 | case FAN_MARK_ADD: |
878 | if (flags & FAN_MARK_MOUNT) | 878 | if (flags & FAN_MARK_MOUNT) |
879 | ret = fanotify_add_vfsmount_mark(group, mnt, mask, flags); | 879 | ret = fanotify_add_vfsmount_mark(group, mnt, mask, flags); |
880 | else | 880 | else |
881 | ret = fanotify_add_inode_mark(group, inode, mask, flags); | 881 | ret = fanotify_add_inode_mark(group, inode, mask, flags); |
882 | break; | 882 | break; |
883 | case FAN_MARK_REMOVE: | 883 | case FAN_MARK_REMOVE: |
884 | if (flags & FAN_MARK_MOUNT) | 884 | if (flags & FAN_MARK_MOUNT) |
885 | ret = fanotify_remove_vfsmount_mark(group, mnt, mask, flags); | 885 | ret = fanotify_remove_vfsmount_mark(group, mnt, mask, flags); |
886 | else | 886 | else |
887 | ret = fanotify_remove_inode_mark(group, inode, mask, flags); | 887 | ret = fanotify_remove_inode_mark(group, inode, mask, flags); |
888 | break; | 888 | break; |
889 | default: | 889 | default: |
890 | ret = -EINVAL; | 890 | ret = -EINVAL; |
891 | } | 891 | } |
892 | 892 | ||
893 | path_put(&path); | 893 | path_put(&path); |
894 | fput_and_out: | 894 | fput_and_out: |
895 | fdput(f); | 895 | fdput(f); |
896 | return ret; | 896 | return ret; |
897 | } | 897 | } |
898 | 898 | ||
899 | #ifdef CONFIG_COMPAT | 899 | #ifdef CONFIG_COMPAT |
900 | COMPAT_SYSCALL_DEFINE6(fanotify_mark, | 900 | COMPAT_SYSCALL_DEFINE6(fanotify_mark, |
901 | int, fanotify_fd, unsigned int, flags, | 901 | int, fanotify_fd, unsigned int, flags, |
902 | __u32, mask0, __u32, mask1, int, dfd, | 902 | __u32, mask0, __u32, mask1, int, dfd, |
903 | const char __user *, pathname) | 903 | const char __user *, pathname) |
904 | { | 904 | { |
905 | return sys_fanotify_mark(fanotify_fd, flags, | 905 | return sys_fanotify_mark(fanotify_fd, flags, |
906 | #ifdef __BIG_ENDIAN | 906 | #ifdef __BIG_ENDIAN |
907 | ((__u64)mask0 << 32) | mask1, | 907 | ((__u64)mask0 << 32) | mask1, |
908 | #else | 908 | #else |
909 | ((__u64)mask1 << 32) | mask0, | 909 | ((__u64)mask1 << 32) | mask0, |
910 | #endif | 910 | #endif |
911 | dfd, pathname); | 911 | dfd, pathname); |
912 | } | 912 | } |
913 | #endif | 913 | #endif |
914 | 914 | ||
915 | /* | 915 | /* |
916 | * fanotify_user_setup - Our initialization function. Note that we cannot return | 916 | * fanotify_user_setup - Our initialization function. Note that we cannot return |
917 | * error because we have compiled-in VFS hooks. So an (unlikely) failure here | 917 | * error because we have compiled-in VFS hooks. So an (unlikely) failure here |
918 | * must result in panic(). | 918 | * must result in panic(). |
919 | */ | 919 | */ |
920 | static int __init fanotify_user_setup(void) | 920 | static int __init fanotify_user_setup(void) |
921 | { | 921 | { |
922 | fanotify_mark_cache = KMEM_CACHE(fsnotify_mark, SLAB_PANIC); | 922 | fanotify_mark_cache = KMEM_CACHE(fsnotify_mark, SLAB_PANIC); |
923 | fanotify_event_cachep = KMEM_CACHE(fanotify_event_info, SLAB_PANIC); | 923 | fanotify_event_cachep = KMEM_CACHE(fanotify_event_info, SLAB_PANIC); |
924 | #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS | 924 | #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS |
925 | fanotify_perm_event_cachep = KMEM_CACHE(fanotify_perm_event_info, | 925 | fanotify_perm_event_cachep = KMEM_CACHE(fanotify_perm_event_info, |
926 | SLAB_PANIC); | 926 | SLAB_PANIC); |
927 | #endif | 927 | #endif |
928 | 928 | ||
929 | return 0; | 929 | return 0; |
kernel/sched/core.c
1 | /* | 1 | /* |
2 | * kernel/sched/core.c | 2 | * kernel/sched/core.c |
3 | * | 3 | * |
4 | * Kernel scheduler and related syscalls | 4 | * Kernel scheduler and related syscalls |
5 | * | 5 | * |
6 | * Copyright (C) 1991-2002 Linus Torvalds | 6 | * Copyright (C) 1991-2002 Linus Torvalds |
7 | * | 7 | * |
8 | * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and | 8 | * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and |
9 | * make semaphores SMP safe | 9 | * make semaphores SMP safe |
10 | * 1998-11-19 Implemented schedule_timeout() and related stuff | 10 | * 1998-11-19 Implemented schedule_timeout() and related stuff |
11 | * by Andrea Arcangeli | 11 | * by Andrea Arcangeli |
12 | * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar: | 12 | * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar: |
13 | * hybrid priority-list and round-robin design with | 13 | * hybrid priority-list and round-robin design with |
14 | * an array-switch method of distributing timeslices | 14 | * an array-switch method of distributing timeslices |
15 | * and per-CPU runqueues. Cleanups and useful suggestions | 15 | * and per-CPU runqueues. Cleanups and useful suggestions |
16 | * by Davide Libenzi, preemptible kernel bits by Robert Love. | 16 | * by Davide Libenzi, preemptible kernel bits by Robert Love. |
17 | * 2003-09-03 Interactivity tuning by Con Kolivas. | 17 | * 2003-09-03 Interactivity tuning by Con Kolivas. |
18 | * 2004-04-02 Scheduler domains code by Nick Piggin | 18 | * 2004-04-02 Scheduler domains code by Nick Piggin |
19 | * 2007-04-15 Work begun on replacing all interactivity tuning with a | 19 | * 2007-04-15 Work begun on replacing all interactivity tuning with a |
20 | * fair scheduling design by Con Kolivas. | 20 | * fair scheduling design by Con Kolivas. |
21 | * 2007-05-05 Load balancing (smp-nice) and other improvements | 21 | * 2007-05-05 Load balancing (smp-nice) and other improvements |
22 | * by Peter Williams | 22 | * by Peter Williams |
23 | * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith | 23 | * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith |
24 | * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri | 24 | * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri |
25 | * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, | 25 | * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, |
26 | * Thomas Gleixner, Mike Kravetz | 26 | * Thomas Gleixner, Mike Kravetz |
27 | */ | 27 | */ |
28 | 28 | ||
29 | #include <linux/mm.h> | 29 | #include <linux/mm.h> |
30 | #include <linux/module.h> | 30 | #include <linux/module.h> |
31 | #include <linux/nmi.h> | 31 | #include <linux/nmi.h> |
32 | #include <linux/init.h> | 32 | #include <linux/init.h> |
33 | #include <linux/uaccess.h> | 33 | #include <linux/uaccess.h> |
34 | #include <linux/highmem.h> | 34 | #include <linux/highmem.h> |
35 | #include <asm/mmu_context.h> | 35 | #include <asm/mmu_context.h> |
36 | #include <linux/interrupt.h> | 36 | #include <linux/interrupt.h> |
37 | #include <linux/capability.h> | 37 | #include <linux/capability.h> |
38 | #include <linux/completion.h> | 38 | #include <linux/completion.h> |
39 | #include <linux/kernel_stat.h> | 39 | #include <linux/kernel_stat.h> |
40 | #include <linux/debug_locks.h> | 40 | #include <linux/debug_locks.h> |
41 | #include <linux/perf_event.h> | 41 | #include <linux/perf_event.h> |
42 | #include <linux/security.h> | 42 | #include <linux/security.h> |
43 | #include <linux/notifier.h> | 43 | #include <linux/notifier.h> |
44 | #include <linux/profile.h> | 44 | #include <linux/profile.h> |
45 | #include <linux/freezer.h> | 45 | #include <linux/freezer.h> |
46 | #include <linux/vmalloc.h> | 46 | #include <linux/vmalloc.h> |
47 | #include <linux/blkdev.h> | 47 | #include <linux/blkdev.h> |
48 | #include <linux/delay.h> | 48 | #include <linux/delay.h> |
49 | #include <linux/pid_namespace.h> | 49 | #include <linux/pid_namespace.h> |
50 | #include <linux/smp.h> | 50 | #include <linux/smp.h> |
51 | #include <linux/threads.h> | 51 | #include <linux/threads.h> |
52 | #include <linux/timer.h> | 52 | #include <linux/timer.h> |
53 | #include <linux/rcupdate.h> | 53 | #include <linux/rcupdate.h> |
54 | #include <linux/cpu.h> | 54 | #include <linux/cpu.h> |
55 | #include <linux/cpuset.h> | 55 | #include <linux/cpuset.h> |
56 | #include <linux/percpu.h> | 56 | #include <linux/percpu.h> |
57 | #include <linux/proc_fs.h> | 57 | #include <linux/proc_fs.h> |
58 | #include <linux/seq_file.h> | 58 | #include <linux/seq_file.h> |
59 | #include <linux/sysctl.h> | 59 | #include <linux/sysctl.h> |
60 | #include <linux/syscalls.h> | 60 | #include <linux/syscalls.h> |
61 | #include <linux/times.h> | 61 | #include <linux/times.h> |
62 | #include <linux/tsacct_kern.h> | 62 | #include <linux/tsacct_kern.h> |
63 | #include <linux/kprobes.h> | 63 | #include <linux/kprobes.h> |
64 | #include <linux/delayacct.h> | 64 | #include <linux/delayacct.h> |
65 | #include <linux/unistd.h> | 65 | #include <linux/unistd.h> |
66 | #include <linux/pagemap.h> | 66 | #include <linux/pagemap.h> |
67 | #include <linux/hrtimer.h> | 67 | #include <linux/hrtimer.h> |
68 | #include <linux/tick.h> | 68 | #include <linux/tick.h> |
69 | #include <linux/debugfs.h> | 69 | #include <linux/debugfs.h> |
70 | #include <linux/ctype.h> | 70 | #include <linux/ctype.h> |
71 | #include <linux/ftrace.h> | 71 | #include <linux/ftrace.h> |
72 | #include <linux/slab.h> | 72 | #include <linux/slab.h> |
73 | #include <linux/init_task.h> | 73 | #include <linux/init_task.h> |
74 | #include <linux/binfmts.h> | 74 | #include <linux/binfmts.h> |
75 | #include <linux/context_tracking.h> | 75 | #include <linux/context_tracking.h> |
76 | #include <linux/compiler.h> | 76 | #include <linux/compiler.h> |
77 | 77 | ||
78 | #include <asm/switch_to.h> | 78 | #include <asm/switch_to.h> |
79 | #include <asm/tlb.h> | 79 | #include <asm/tlb.h> |
80 | #include <asm/irq_regs.h> | 80 | #include <asm/irq_regs.h> |
81 | #include <asm/mutex.h> | 81 | #include <asm/mutex.h> |
82 | #ifdef CONFIG_PARAVIRT | 82 | #ifdef CONFIG_PARAVIRT |
83 | #include <asm/paravirt.h> | 83 | #include <asm/paravirt.h> |
84 | #endif | 84 | #endif |
85 | 85 | ||
86 | #include "sched.h" | 86 | #include "sched.h" |
87 | #include "../workqueue_internal.h" | 87 | #include "../workqueue_internal.h" |
88 | #include "../smpboot.h" | 88 | #include "../smpboot.h" |
89 | 89 | ||
90 | #define CREATE_TRACE_POINTS | 90 | #define CREATE_TRACE_POINTS |
91 | #include <trace/events/sched.h> | 91 | #include <trace/events/sched.h> |
92 | 92 | ||
93 | void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) | 93 | void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) |
94 | { | 94 | { |
95 | unsigned long delta; | 95 | unsigned long delta; |
96 | ktime_t soft, hard, now; | 96 | ktime_t soft, hard, now; |
97 | 97 | ||
98 | for (;;) { | 98 | for (;;) { |
99 | if (hrtimer_active(period_timer)) | 99 | if (hrtimer_active(period_timer)) |
100 | break; | 100 | break; |
101 | 101 | ||
102 | now = hrtimer_cb_get_time(period_timer); | 102 | now = hrtimer_cb_get_time(period_timer); |
103 | hrtimer_forward(period_timer, now, period); | 103 | hrtimer_forward(period_timer, now, period); |
104 | 104 | ||
105 | soft = hrtimer_get_softexpires(period_timer); | 105 | soft = hrtimer_get_softexpires(period_timer); |
106 | hard = hrtimer_get_expires(period_timer); | 106 | hard = hrtimer_get_expires(period_timer); |
107 | delta = ktime_to_ns(ktime_sub(hard, soft)); | 107 | delta = ktime_to_ns(ktime_sub(hard, soft)); |
108 | __hrtimer_start_range_ns(period_timer, soft, delta, | 108 | __hrtimer_start_range_ns(period_timer, soft, delta, |
109 | HRTIMER_MODE_ABS_PINNED, 0); | 109 | HRTIMER_MODE_ABS_PINNED, 0); |
110 | } | 110 | } |
111 | } | 111 | } |
112 | 112 | ||
113 | DEFINE_MUTEX(sched_domains_mutex); | 113 | DEFINE_MUTEX(sched_domains_mutex); |
114 | DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); | 114 | DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); |
115 | 115 | ||
116 | static void update_rq_clock_task(struct rq *rq, s64 delta); | 116 | static void update_rq_clock_task(struct rq *rq, s64 delta); |
117 | 117 | ||
118 | void update_rq_clock(struct rq *rq) | 118 | void update_rq_clock(struct rq *rq) |
119 | { | 119 | { |
120 | s64 delta; | 120 | s64 delta; |
121 | 121 | ||
122 | if (rq->skip_clock_update > 0) | 122 | if (rq->skip_clock_update > 0) |
123 | return; | 123 | return; |
124 | 124 | ||
125 | delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; | 125 | delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; |
126 | if (delta < 0) | 126 | if (delta < 0) |
127 | return; | 127 | return; |
128 | rq->clock += delta; | 128 | rq->clock += delta; |
129 | update_rq_clock_task(rq, delta); | 129 | update_rq_clock_task(rq, delta); |
130 | } | 130 | } |
131 | 131 | ||
132 | /* | 132 | /* |
133 | * Debugging: various feature bits | 133 | * Debugging: various feature bits |
134 | */ | 134 | */ |
135 | 135 | ||
136 | #define SCHED_FEAT(name, enabled) \ | 136 | #define SCHED_FEAT(name, enabled) \ |
137 | (1UL << __SCHED_FEAT_##name) * enabled | | 137 | (1UL << __SCHED_FEAT_##name) * enabled | |
138 | 138 | ||
139 | const_debug unsigned int sysctl_sched_features = | 139 | const_debug unsigned int sysctl_sched_features = |
140 | #include "features.h" | 140 | #include "features.h" |
141 | 0; | 141 | 0; |
142 | 142 | ||
143 | #undef SCHED_FEAT | 143 | #undef SCHED_FEAT |
144 | 144 | ||
145 | #ifdef CONFIG_SCHED_DEBUG | 145 | #ifdef CONFIG_SCHED_DEBUG |
146 | #define SCHED_FEAT(name, enabled) \ | 146 | #define SCHED_FEAT(name, enabled) \ |
147 | #name , | 147 | #name , |
148 | 148 | ||
149 | static const char * const sched_feat_names[] = { | 149 | static const char * const sched_feat_names[] = { |
150 | #include "features.h" | 150 | #include "features.h" |
151 | }; | 151 | }; |
152 | 152 | ||
153 | #undef SCHED_FEAT | 153 | #undef SCHED_FEAT |
154 | 154 | ||
155 | static int sched_feat_show(struct seq_file *m, void *v) | 155 | static int sched_feat_show(struct seq_file *m, void *v) |
156 | { | 156 | { |
157 | int i; | 157 | int i; |
158 | 158 | ||
159 | for (i = 0; i < __SCHED_FEAT_NR; i++) { | 159 | for (i = 0; i < __SCHED_FEAT_NR; i++) { |
160 | if (!(sysctl_sched_features & (1UL << i))) | 160 | if (!(sysctl_sched_features & (1UL << i))) |
161 | seq_puts(m, "NO_"); | 161 | seq_puts(m, "NO_"); |
162 | seq_printf(m, "%s ", sched_feat_names[i]); | 162 | seq_printf(m, "%s ", sched_feat_names[i]); |
163 | } | 163 | } |
164 | seq_puts(m, "\n"); | 164 | seq_puts(m, "\n"); |
165 | 165 | ||
166 | return 0; | 166 | return 0; |
167 | } | 167 | } |
168 | 168 | ||
169 | #ifdef HAVE_JUMP_LABEL | 169 | #ifdef HAVE_JUMP_LABEL |
170 | 170 | ||
171 | #define jump_label_key__true STATIC_KEY_INIT_TRUE | 171 | #define jump_label_key__true STATIC_KEY_INIT_TRUE |
172 | #define jump_label_key__false STATIC_KEY_INIT_FALSE | 172 | #define jump_label_key__false STATIC_KEY_INIT_FALSE |
173 | 173 | ||
174 | #define SCHED_FEAT(name, enabled) \ | 174 | #define SCHED_FEAT(name, enabled) \ |
175 | jump_label_key__##enabled , | 175 | jump_label_key__##enabled , |
176 | 176 | ||
177 | struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { | 177 | struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { |
178 | #include "features.h" | 178 | #include "features.h" |
179 | }; | 179 | }; |
180 | 180 | ||
181 | #undef SCHED_FEAT | 181 | #undef SCHED_FEAT |
182 | 182 | ||
183 | static void sched_feat_disable(int i) | 183 | static void sched_feat_disable(int i) |
184 | { | 184 | { |
185 | if (static_key_enabled(&sched_feat_keys[i])) | 185 | if (static_key_enabled(&sched_feat_keys[i])) |
186 | static_key_slow_dec(&sched_feat_keys[i]); | 186 | static_key_slow_dec(&sched_feat_keys[i]); |
187 | } | 187 | } |
188 | 188 | ||
189 | static void sched_feat_enable(int i) | 189 | static void sched_feat_enable(int i) |
190 | { | 190 | { |
191 | if (!static_key_enabled(&sched_feat_keys[i])) | 191 | if (!static_key_enabled(&sched_feat_keys[i])) |
192 | static_key_slow_inc(&sched_feat_keys[i]); | 192 | static_key_slow_inc(&sched_feat_keys[i]); |
193 | } | 193 | } |
194 | #else | 194 | #else |
195 | static void sched_feat_disable(int i) { }; | 195 | static void sched_feat_disable(int i) { }; |
196 | static void sched_feat_enable(int i) { }; | 196 | static void sched_feat_enable(int i) { }; |
197 | #endif /* HAVE_JUMP_LABEL */ | 197 | #endif /* HAVE_JUMP_LABEL */ |
198 | 198 | ||
199 | static int sched_feat_set(char *cmp) | 199 | static int sched_feat_set(char *cmp) |
200 | { | 200 | { |
201 | int i; | 201 | int i; |
202 | int neg = 0; | 202 | int neg = 0; |
203 | 203 | ||
204 | if (strncmp(cmp, "NO_", 3) == 0) { | 204 | if (strncmp(cmp, "NO_", 3) == 0) { |
205 | neg = 1; | 205 | neg = 1; |
206 | cmp += 3; | 206 | cmp += 3; |
207 | } | 207 | } |
208 | 208 | ||
209 | for (i = 0; i < __SCHED_FEAT_NR; i++) { | 209 | for (i = 0; i < __SCHED_FEAT_NR; i++) { |
210 | if (strcmp(cmp, sched_feat_names[i]) == 0) { | 210 | if (strcmp(cmp, sched_feat_names[i]) == 0) { |
211 | if (neg) { | 211 | if (neg) { |
212 | sysctl_sched_features &= ~(1UL << i); | 212 | sysctl_sched_features &= ~(1UL << i); |
213 | sched_feat_disable(i); | 213 | sched_feat_disable(i); |
214 | } else { | 214 | } else { |
215 | sysctl_sched_features |= (1UL << i); | 215 | sysctl_sched_features |= (1UL << i); |
216 | sched_feat_enable(i); | 216 | sched_feat_enable(i); |
217 | } | 217 | } |
218 | break; | 218 | break; |
219 | } | 219 | } |
220 | } | 220 | } |
221 | 221 | ||
222 | return i; | 222 | return i; |
223 | } | 223 | } |
224 | 224 | ||
225 | static ssize_t | 225 | static ssize_t |
226 | sched_feat_write(struct file *filp, const char __user *ubuf, | 226 | sched_feat_write(struct file *filp, const char __user *ubuf, |
227 | size_t cnt, loff_t *ppos) | 227 | size_t cnt, loff_t *ppos) |
228 | { | 228 | { |
229 | char buf[64]; | 229 | char buf[64]; |
230 | char *cmp; | 230 | char *cmp; |
231 | int i; | 231 | int i; |
232 | struct inode *inode; | 232 | struct inode *inode; |
233 | 233 | ||
234 | if (cnt > 63) | 234 | if (cnt > 63) |
235 | cnt = 63; | 235 | cnt = 63; |
236 | 236 | ||
237 | if (copy_from_user(&buf, ubuf, cnt)) | 237 | if (copy_from_user(&buf, ubuf, cnt)) |
238 | return -EFAULT; | 238 | return -EFAULT; |
239 | 239 | ||
240 | buf[cnt] = 0; | 240 | buf[cnt] = 0; |
241 | cmp = strstrip(buf); | 241 | cmp = strstrip(buf); |
242 | 242 | ||
243 | /* Ensure the static_key remains in a consistent state */ | 243 | /* Ensure the static_key remains in a consistent state */ |
244 | inode = file_inode(filp); | 244 | inode = file_inode(filp); |
245 | mutex_lock(&inode->i_mutex); | 245 | mutex_lock(&inode->i_mutex); |
246 | i = sched_feat_set(cmp); | 246 | i = sched_feat_set(cmp); |
247 | mutex_unlock(&inode->i_mutex); | 247 | mutex_unlock(&inode->i_mutex); |
248 | if (i == __SCHED_FEAT_NR) | 248 | if (i == __SCHED_FEAT_NR) |
249 | return -EINVAL; | 249 | return -EINVAL; |
250 | 250 | ||
251 | *ppos += cnt; | 251 | *ppos += cnt; |
252 | 252 | ||
253 | return cnt; | 253 | return cnt; |
254 | } | 254 | } |
255 | 255 | ||
256 | static int sched_feat_open(struct inode *inode, struct file *filp) | 256 | static int sched_feat_open(struct inode *inode, struct file *filp) |
257 | { | 257 | { |
258 | return single_open(filp, sched_feat_show, NULL); | 258 | return single_open(filp, sched_feat_show, NULL); |
259 | } | 259 | } |
260 | 260 | ||
261 | static const struct file_operations sched_feat_fops = { | 261 | static const struct file_operations sched_feat_fops = { |
262 | .open = sched_feat_open, | 262 | .open = sched_feat_open, |
263 | .write = sched_feat_write, | 263 | .write = sched_feat_write, |
264 | .read = seq_read, | 264 | .read = seq_read, |
265 | .llseek = seq_lseek, | 265 | .llseek = seq_lseek, |
266 | .release = single_release, | 266 | .release = single_release, |
267 | }; | 267 | }; |
268 | 268 | ||
269 | static __init int sched_init_debug(void) | 269 | static __init int sched_init_debug(void) |
270 | { | 270 | { |
271 | debugfs_create_file("sched_features", 0644, NULL, NULL, | 271 | debugfs_create_file("sched_features", 0644, NULL, NULL, |
272 | &sched_feat_fops); | 272 | &sched_feat_fops); |
273 | 273 | ||
274 | return 0; | 274 | return 0; |
275 | } | 275 | } |
276 | late_initcall(sched_init_debug); | 276 | late_initcall(sched_init_debug); |
277 | #endif /* CONFIG_SCHED_DEBUG */ | 277 | #endif /* CONFIG_SCHED_DEBUG */ |
278 | 278 | ||
279 | /* | 279 | /* |
280 | * Number of tasks to iterate in a single balance run. | 280 | * Number of tasks to iterate in a single balance run. |
281 | * Limited because this is done with IRQs disabled. | 281 | * Limited because this is done with IRQs disabled. |
282 | */ | 282 | */ |
283 | const_debug unsigned int sysctl_sched_nr_migrate = 32; | 283 | const_debug unsigned int sysctl_sched_nr_migrate = 32; |
284 | 284 | ||
285 | /* | 285 | /* |
286 | * period over which we average the RT time consumption, measured | 286 | * period over which we average the RT time consumption, measured |
287 | * in ms. | 287 | * in ms. |
288 | * | 288 | * |
289 | * default: 1s | 289 | * default: 1s |
290 | */ | 290 | */ |
291 | const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; | 291 | const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; |
292 | 292 | ||
293 | /* | 293 | /* |
294 | * period over which we measure -rt task cpu usage in us. | 294 | * period over which we measure -rt task cpu usage in us. |
295 | * default: 1s | 295 | * default: 1s |
296 | */ | 296 | */ |
297 | unsigned int sysctl_sched_rt_period = 1000000; | 297 | unsigned int sysctl_sched_rt_period = 1000000; |
298 | 298 | ||
299 | __read_mostly int scheduler_running; | 299 | __read_mostly int scheduler_running; |
300 | 300 | ||
301 | /* | 301 | /* |
302 | * part of the period that we allow rt tasks to run in us. | 302 | * part of the period that we allow rt tasks to run in us. |
303 | * default: 0.95s | 303 | * default: 0.95s |
304 | */ | 304 | */ |
305 | int sysctl_sched_rt_runtime = 950000; | 305 | int sysctl_sched_rt_runtime = 950000; |
306 | 306 | ||
307 | /* | 307 | /* |
308 | * __task_rq_lock - lock the rq @p resides on. | 308 | * __task_rq_lock - lock the rq @p resides on. |
309 | */ | 309 | */ |
310 | static inline struct rq *__task_rq_lock(struct task_struct *p) | 310 | static inline struct rq *__task_rq_lock(struct task_struct *p) |
311 | __acquires(rq->lock) | 311 | __acquires(rq->lock) |
312 | { | 312 | { |
313 | struct rq *rq; | 313 | struct rq *rq; |
314 | 314 | ||
315 | lockdep_assert_held(&p->pi_lock); | 315 | lockdep_assert_held(&p->pi_lock); |
316 | 316 | ||
317 | for (;;) { | 317 | for (;;) { |
318 | rq = task_rq(p); | 318 | rq = task_rq(p); |
319 | raw_spin_lock(&rq->lock); | 319 | raw_spin_lock(&rq->lock); |
320 | if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) | 320 | if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) |
321 | return rq; | 321 | return rq; |
322 | raw_spin_unlock(&rq->lock); | 322 | raw_spin_unlock(&rq->lock); |
323 | 323 | ||
324 | while (unlikely(task_on_rq_migrating(p))) | 324 | while (unlikely(task_on_rq_migrating(p))) |
325 | cpu_relax(); | 325 | cpu_relax(); |
326 | } | 326 | } |
327 | } | 327 | } |
328 | 328 | ||
329 | /* | 329 | /* |
330 | * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. | 330 | * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. |
331 | */ | 331 | */ |
332 | static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) | 332 | static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) |
333 | __acquires(p->pi_lock) | 333 | __acquires(p->pi_lock) |
334 | __acquires(rq->lock) | 334 | __acquires(rq->lock) |
335 | { | 335 | { |
336 | struct rq *rq; | 336 | struct rq *rq; |
337 | 337 | ||
338 | for (;;) { | 338 | for (;;) { |
339 | raw_spin_lock_irqsave(&p->pi_lock, *flags); | 339 | raw_spin_lock_irqsave(&p->pi_lock, *flags); |
340 | rq = task_rq(p); | 340 | rq = task_rq(p); |
341 | raw_spin_lock(&rq->lock); | 341 | raw_spin_lock(&rq->lock); |
342 | if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) | 342 | if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) |
343 | return rq; | 343 | return rq; |
344 | raw_spin_unlock(&rq->lock); | 344 | raw_spin_unlock(&rq->lock); |
345 | raw_spin_unlock_irqrestore(&p->pi_lock, *flags); | 345 | raw_spin_unlock_irqrestore(&p->pi_lock, *flags); |
346 | 346 | ||
347 | while (unlikely(task_on_rq_migrating(p))) | 347 | while (unlikely(task_on_rq_migrating(p))) |
348 | cpu_relax(); | 348 | cpu_relax(); |
349 | } | 349 | } |
350 | } | 350 | } |
351 | 351 | ||
352 | static void __task_rq_unlock(struct rq *rq) | 352 | static void __task_rq_unlock(struct rq *rq) |
353 | __releases(rq->lock) | 353 | __releases(rq->lock) |
354 | { | 354 | { |
355 | raw_spin_unlock(&rq->lock); | 355 | raw_spin_unlock(&rq->lock); |
356 | } | 356 | } |
357 | 357 | ||
358 | static inline void | 358 | static inline void |
359 | task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) | 359 | task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) |
360 | __releases(rq->lock) | 360 | __releases(rq->lock) |
361 | __releases(p->pi_lock) | 361 | __releases(p->pi_lock) |
362 | { | 362 | { |
363 | raw_spin_unlock(&rq->lock); | 363 | raw_spin_unlock(&rq->lock); |
364 | raw_spin_unlock_irqrestore(&p->pi_lock, *flags); | 364 | raw_spin_unlock_irqrestore(&p->pi_lock, *flags); |
365 | } | 365 | } |
366 | 366 | ||
367 | /* | 367 | /* |
368 | * this_rq_lock - lock this runqueue and disable interrupts. | 368 | * this_rq_lock - lock this runqueue and disable interrupts. |
369 | */ | 369 | */ |
370 | static struct rq *this_rq_lock(void) | 370 | static struct rq *this_rq_lock(void) |
371 | __acquires(rq->lock) | 371 | __acquires(rq->lock) |
372 | { | 372 | { |
373 | struct rq *rq; | 373 | struct rq *rq; |
374 | 374 | ||
375 | local_irq_disable(); | 375 | local_irq_disable(); |
376 | rq = this_rq(); | 376 | rq = this_rq(); |
377 | raw_spin_lock(&rq->lock); | 377 | raw_spin_lock(&rq->lock); |
378 | 378 | ||
379 | return rq; | 379 | return rq; |
380 | } | 380 | } |
381 | 381 | ||
382 | #ifdef CONFIG_SCHED_HRTICK | 382 | #ifdef CONFIG_SCHED_HRTICK |
383 | /* | 383 | /* |
384 | * Use HR-timers to deliver accurate preemption points. | 384 | * Use HR-timers to deliver accurate preemption points. |
385 | */ | 385 | */ |
386 | 386 | ||
387 | static void hrtick_clear(struct rq *rq) | 387 | static void hrtick_clear(struct rq *rq) |
388 | { | 388 | { |
389 | if (hrtimer_active(&rq->hrtick_timer)) | 389 | if (hrtimer_active(&rq->hrtick_timer)) |
390 | hrtimer_cancel(&rq->hrtick_timer); | 390 | hrtimer_cancel(&rq->hrtick_timer); |
391 | } | 391 | } |
392 | 392 | ||
393 | /* | 393 | /* |
394 | * High-resolution timer tick. | 394 | * High-resolution timer tick. |
395 | * Runs from hardirq context with interrupts disabled. | 395 | * Runs from hardirq context with interrupts disabled. |
396 | */ | 396 | */ |
397 | static enum hrtimer_restart hrtick(struct hrtimer *timer) | 397 | static enum hrtimer_restart hrtick(struct hrtimer *timer) |
398 | { | 398 | { |
399 | struct rq *rq = container_of(timer, struct rq, hrtick_timer); | 399 | struct rq *rq = container_of(timer, struct rq, hrtick_timer); |
400 | 400 | ||
401 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); | 401 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); |
402 | 402 | ||
403 | raw_spin_lock(&rq->lock); | 403 | raw_spin_lock(&rq->lock); |
404 | update_rq_clock(rq); | 404 | update_rq_clock(rq); |
405 | rq->curr->sched_class->task_tick(rq, rq->curr, 1); | 405 | rq->curr->sched_class->task_tick(rq, rq->curr, 1); |
406 | raw_spin_unlock(&rq->lock); | 406 | raw_spin_unlock(&rq->lock); |
407 | 407 | ||
408 | return HRTIMER_NORESTART; | 408 | return HRTIMER_NORESTART; |
409 | } | 409 | } |
410 | 410 | ||
411 | #ifdef CONFIG_SMP | 411 | #ifdef CONFIG_SMP |
412 | 412 | ||
413 | static int __hrtick_restart(struct rq *rq) | 413 | static int __hrtick_restart(struct rq *rq) |
414 | { | 414 | { |
415 | struct hrtimer *timer = &rq->hrtick_timer; | 415 | struct hrtimer *timer = &rq->hrtick_timer; |
416 | ktime_t time = hrtimer_get_softexpires(timer); | 416 | ktime_t time = hrtimer_get_softexpires(timer); |
417 | 417 | ||
418 | return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); | 418 | return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); |
419 | } | 419 | } |
420 | 420 | ||
421 | /* | 421 | /* |
422 | * called from hardirq (IPI) context | 422 | * called from hardirq (IPI) context |
423 | */ | 423 | */ |
424 | static void __hrtick_start(void *arg) | 424 | static void __hrtick_start(void *arg) |
425 | { | 425 | { |
426 | struct rq *rq = arg; | 426 | struct rq *rq = arg; |
427 | 427 | ||
428 | raw_spin_lock(&rq->lock); | 428 | raw_spin_lock(&rq->lock); |
429 | __hrtick_restart(rq); | 429 | __hrtick_restart(rq); |
430 | rq->hrtick_csd_pending = 0; | 430 | rq->hrtick_csd_pending = 0; |
431 | raw_spin_unlock(&rq->lock); | 431 | raw_spin_unlock(&rq->lock); |
432 | } | 432 | } |
433 | 433 | ||
434 | /* | 434 | /* |
435 | * Called to set the hrtick timer state. | 435 | * Called to set the hrtick timer state. |
436 | * | 436 | * |
437 | * called with rq->lock held and irqs disabled | 437 | * called with rq->lock held and irqs disabled |
438 | */ | 438 | */ |
439 | void hrtick_start(struct rq *rq, u64 delay) | 439 | void hrtick_start(struct rq *rq, u64 delay) |
440 | { | 440 | { |
441 | struct hrtimer *timer = &rq->hrtick_timer; | 441 | struct hrtimer *timer = &rq->hrtick_timer; |
442 | ktime_t time; | 442 | ktime_t time; |
443 | s64 delta; | 443 | s64 delta; |
444 | 444 | ||
445 | /* | 445 | /* |
446 | * Don't schedule slices shorter than 10000ns, that just | 446 | * Don't schedule slices shorter than 10000ns, that just |
447 | * doesn't make sense and can cause timer DoS. | 447 | * doesn't make sense and can cause timer DoS. |
448 | */ | 448 | */ |
449 | delta = max_t(s64, delay, 10000LL); | 449 | delta = max_t(s64, delay, 10000LL); |
450 | time = ktime_add_ns(timer->base->get_time(), delta); | 450 | time = ktime_add_ns(timer->base->get_time(), delta); |
451 | 451 | ||
452 | hrtimer_set_expires(timer, time); | 452 | hrtimer_set_expires(timer, time); |
453 | 453 | ||
454 | if (rq == this_rq()) { | 454 | if (rq == this_rq()) { |
455 | __hrtick_restart(rq); | 455 | __hrtick_restart(rq); |
456 | } else if (!rq->hrtick_csd_pending) { | 456 | } else if (!rq->hrtick_csd_pending) { |
457 | smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd); | 457 | smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd); |
458 | rq->hrtick_csd_pending = 1; | 458 | rq->hrtick_csd_pending = 1; |
459 | } | 459 | } |
460 | } | 460 | } |
461 | 461 | ||
462 | static int | 462 | static int |
463 | hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) | 463 | hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) |
464 | { | 464 | { |
465 | int cpu = (int)(long)hcpu; | 465 | int cpu = (int)(long)hcpu; |
466 | 466 | ||
467 | switch (action) { | 467 | switch (action) { |
468 | case CPU_UP_CANCELED: | 468 | case CPU_UP_CANCELED: |
469 | case CPU_UP_CANCELED_FROZEN: | 469 | case CPU_UP_CANCELED_FROZEN: |
470 | case CPU_DOWN_PREPARE: | 470 | case CPU_DOWN_PREPARE: |
471 | case CPU_DOWN_PREPARE_FROZEN: | 471 | case CPU_DOWN_PREPARE_FROZEN: |
472 | case CPU_DEAD: | 472 | case CPU_DEAD: |
473 | case CPU_DEAD_FROZEN: | 473 | case CPU_DEAD_FROZEN: |
474 | hrtick_clear(cpu_rq(cpu)); | 474 | hrtick_clear(cpu_rq(cpu)); |
475 | return NOTIFY_OK; | 475 | return NOTIFY_OK; |
476 | } | 476 | } |
477 | 477 | ||
478 | return NOTIFY_DONE; | 478 | return NOTIFY_DONE; |
479 | } | 479 | } |
480 | 480 | ||
481 | static __init void init_hrtick(void) | 481 | static __init void init_hrtick(void) |
482 | { | 482 | { |
483 | hotcpu_notifier(hotplug_hrtick, 0); | 483 | hotcpu_notifier(hotplug_hrtick, 0); |
484 | } | 484 | } |
485 | #else | 485 | #else |
486 | /* | 486 | /* |
487 | * Called to set the hrtick timer state. | 487 | * Called to set the hrtick timer state. |
488 | * | 488 | * |
489 | * called with rq->lock held and irqs disabled | 489 | * called with rq->lock held and irqs disabled |
490 | */ | 490 | */ |
491 | void hrtick_start(struct rq *rq, u64 delay) | 491 | void hrtick_start(struct rq *rq, u64 delay) |
492 | { | 492 | { |
493 | __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, | 493 | __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, |
494 | HRTIMER_MODE_REL_PINNED, 0); | 494 | HRTIMER_MODE_REL_PINNED, 0); |
495 | } | 495 | } |
496 | 496 | ||
497 | static inline void init_hrtick(void) | 497 | static inline void init_hrtick(void) |
498 | { | 498 | { |
499 | } | 499 | } |
500 | #endif /* CONFIG_SMP */ | 500 | #endif /* CONFIG_SMP */ |
501 | 501 | ||
502 | static void init_rq_hrtick(struct rq *rq) | 502 | static void init_rq_hrtick(struct rq *rq) |
503 | { | 503 | { |
504 | #ifdef CONFIG_SMP | 504 | #ifdef CONFIG_SMP |
505 | rq->hrtick_csd_pending = 0; | 505 | rq->hrtick_csd_pending = 0; |
506 | 506 | ||
507 | rq->hrtick_csd.flags = 0; | 507 | rq->hrtick_csd.flags = 0; |
508 | rq->hrtick_csd.func = __hrtick_start; | 508 | rq->hrtick_csd.func = __hrtick_start; |
509 | rq->hrtick_csd.info = rq; | 509 | rq->hrtick_csd.info = rq; |
510 | #endif | 510 | #endif |
511 | 511 | ||
512 | hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | 512 | hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
513 | rq->hrtick_timer.function = hrtick; | 513 | rq->hrtick_timer.function = hrtick; |
514 | } | 514 | } |
515 | #else /* CONFIG_SCHED_HRTICK */ | 515 | #else /* CONFIG_SCHED_HRTICK */ |
516 | static inline void hrtick_clear(struct rq *rq) | 516 | static inline void hrtick_clear(struct rq *rq) |
517 | { | 517 | { |
518 | } | 518 | } |
519 | 519 | ||
520 | static inline void init_rq_hrtick(struct rq *rq) | 520 | static inline void init_rq_hrtick(struct rq *rq) |
521 | { | 521 | { |
522 | } | 522 | } |
523 | 523 | ||
524 | static inline void init_hrtick(void) | 524 | static inline void init_hrtick(void) |
525 | { | 525 | { |
526 | } | 526 | } |
527 | #endif /* CONFIG_SCHED_HRTICK */ | 527 | #endif /* CONFIG_SCHED_HRTICK */ |
528 | 528 | ||
529 | /* | 529 | /* |
530 | * cmpxchg based fetch_or, macro so it works for different integer types | 530 | * cmpxchg based fetch_or, macro so it works for different integer types |
531 | */ | 531 | */ |
532 | #define fetch_or(ptr, val) \ | 532 | #define fetch_or(ptr, val) \ |
533 | ({ typeof(*(ptr)) __old, __val = *(ptr); \ | 533 | ({ typeof(*(ptr)) __old, __val = *(ptr); \ |
534 | for (;;) { \ | 534 | for (;;) { \ |
535 | __old = cmpxchg((ptr), __val, __val | (val)); \ | 535 | __old = cmpxchg((ptr), __val, __val | (val)); \ |
536 | if (__old == __val) \ | 536 | if (__old == __val) \ |
537 | break; \ | 537 | break; \ |
538 | __val = __old; \ | 538 | __val = __old; \ |
539 | } \ | 539 | } \ |
540 | __old; \ | 540 | __old; \ |
541 | }) | 541 | }) |
542 | 542 | ||
543 | #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG) | 543 | #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG) |
544 | /* | 544 | /* |
545 | * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG, | 545 | * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG, |
546 | * this avoids any races wrt polling state changes and thereby avoids | 546 | * this avoids any races wrt polling state changes and thereby avoids |
547 | * spurious IPIs. | 547 | * spurious IPIs. |
548 | */ | 548 | */ |
549 | static bool set_nr_and_not_polling(struct task_struct *p) | 549 | static bool set_nr_and_not_polling(struct task_struct *p) |
550 | { | 550 | { |
551 | struct thread_info *ti = task_thread_info(p); | 551 | struct thread_info *ti = task_thread_info(p); |
552 | return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG); | 552 | return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG); |
553 | } | 553 | } |
554 | 554 | ||
555 | /* | 555 | /* |
556 | * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set. | 556 | * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set. |
557 | * | 557 | * |
558 | * If this returns true, then the idle task promises to call | 558 | * If this returns true, then the idle task promises to call |
559 | * sched_ttwu_pending() and reschedule soon. | 559 | * sched_ttwu_pending() and reschedule soon. |
560 | */ | 560 | */ |
561 | static bool set_nr_if_polling(struct task_struct *p) | 561 | static bool set_nr_if_polling(struct task_struct *p) |
562 | { | 562 | { |
563 | struct thread_info *ti = task_thread_info(p); | 563 | struct thread_info *ti = task_thread_info(p); |
564 | typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags); | 564 | typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags); |
565 | 565 | ||
566 | for (;;) { | 566 | for (;;) { |
567 | if (!(val & _TIF_POLLING_NRFLAG)) | 567 | if (!(val & _TIF_POLLING_NRFLAG)) |
568 | return false; | 568 | return false; |
569 | if (val & _TIF_NEED_RESCHED) | 569 | if (val & _TIF_NEED_RESCHED) |
570 | return true; | 570 | return true; |
571 | old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED); | 571 | old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED); |
572 | if (old == val) | 572 | if (old == val) |
573 | break; | 573 | break; |
574 | val = old; | 574 | val = old; |
575 | } | 575 | } |
576 | return true; | 576 | return true; |
577 | } | 577 | } |
578 | 578 | ||
579 | #else | 579 | #else |
580 | static bool set_nr_and_not_polling(struct task_struct *p) | 580 | static bool set_nr_and_not_polling(struct task_struct *p) |
581 | { | 581 | { |
582 | set_tsk_need_resched(p); | 582 | set_tsk_need_resched(p); |
583 | return true; | 583 | return true; |
584 | } | 584 | } |
585 | 585 | ||
586 | #ifdef CONFIG_SMP | 586 | #ifdef CONFIG_SMP |
587 | static bool set_nr_if_polling(struct task_struct *p) | 587 | static bool set_nr_if_polling(struct task_struct *p) |
588 | { | 588 | { |
589 | return false; | 589 | return false; |
590 | } | 590 | } |
591 | #endif | 591 | #endif |
592 | #endif | 592 | #endif |
593 | 593 | ||
594 | /* | 594 | /* |
595 | * resched_curr - mark rq's current task 'to be rescheduled now'. | 595 | * resched_curr - mark rq's current task 'to be rescheduled now'. |
596 | * | 596 | * |
597 | * On UP this means the setting of the need_resched flag, on SMP it | 597 | * On UP this means the setting of the need_resched flag, on SMP it |
598 | * might also involve a cross-CPU call to trigger the scheduler on | 598 | * might also involve a cross-CPU call to trigger the scheduler on |
599 | * the target CPU. | 599 | * the target CPU. |
600 | */ | 600 | */ |
601 | void resched_curr(struct rq *rq) | 601 | void resched_curr(struct rq *rq) |
602 | { | 602 | { |
603 | struct task_struct *curr = rq->curr; | 603 | struct task_struct *curr = rq->curr; |
604 | int cpu; | 604 | int cpu; |
605 | 605 | ||
606 | lockdep_assert_held(&rq->lock); | 606 | lockdep_assert_held(&rq->lock); |
607 | 607 | ||
608 | if (test_tsk_need_resched(curr)) | 608 | if (test_tsk_need_resched(curr)) |
609 | return; | 609 | return; |
610 | 610 | ||
611 | cpu = cpu_of(rq); | 611 | cpu = cpu_of(rq); |
612 | 612 | ||
613 | if (cpu == smp_processor_id()) { | 613 | if (cpu == smp_processor_id()) { |
614 | set_tsk_need_resched(curr); | 614 | set_tsk_need_resched(curr); |
615 | set_preempt_need_resched(); | 615 | set_preempt_need_resched(); |
616 | return; | 616 | return; |
617 | } | 617 | } |
618 | 618 | ||
619 | if (set_nr_and_not_polling(curr)) | 619 | if (set_nr_and_not_polling(curr)) |
620 | smp_send_reschedule(cpu); | 620 | smp_send_reschedule(cpu); |
621 | else | 621 | else |
622 | trace_sched_wake_idle_without_ipi(cpu); | 622 | trace_sched_wake_idle_without_ipi(cpu); |
623 | } | 623 | } |
624 | 624 | ||
625 | void resched_cpu(int cpu) | 625 | void resched_cpu(int cpu) |
626 | { | 626 | { |
627 | struct rq *rq = cpu_rq(cpu); | 627 | struct rq *rq = cpu_rq(cpu); |
628 | unsigned long flags; | 628 | unsigned long flags; |
629 | 629 | ||
630 | if (!raw_spin_trylock_irqsave(&rq->lock, flags)) | 630 | if (!raw_spin_trylock_irqsave(&rq->lock, flags)) |
631 | return; | 631 | return; |
632 | resched_curr(rq); | 632 | resched_curr(rq); |
633 | raw_spin_unlock_irqrestore(&rq->lock, flags); | 633 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
634 | } | 634 | } |
635 | 635 | ||
636 | #ifdef CONFIG_SMP | 636 | #ifdef CONFIG_SMP |
637 | #ifdef CONFIG_NO_HZ_COMMON | 637 | #ifdef CONFIG_NO_HZ_COMMON |
638 | /* | 638 | /* |
639 | * In the semi idle case, use the nearest busy cpu for migrating timers | 639 | * In the semi idle case, use the nearest busy cpu for migrating timers |
640 | * from an idle cpu. This is good for power-savings. | 640 | * from an idle cpu. This is good for power-savings. |
641 | * | 641 | * |
642 | * We don't do similar optimization for completely idle system, as | 642 | * We don't do similar optimization for completely idle system, as |
643 | * selecting an idle cpu will add more delays to the timers than intended | 643 | * selecting an idle cpu will add more delays to the timers than intended |
644 | * (as that cpu's timer base may not be uptodate wrt jiffies etc). | 644 | * (as that cpu's timer base may not be uptodate wrt jiffies etc). |
645 | */ | 645 | */ |
646 | int get_nohz_timer_target(int pinned) | 646 | int get_nohz_timer_target(int pinned) |
647 | { | 647 | { |
648 | int cpu = smp_processor_id(); | 648 | int cpu = smp_processor_id(); |
649 | int i; | 649 | int i; |
650 | struct sched_domain *sd; | 650 | struct sched_domain *sd; |
651 | 651 | ||
652 | if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu)) | 652 | if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu)) |
653 | return cpu; | 653 | return cpu; |
654 | 654 | ||
655 | rcu_read_lock(); | 655 | rcu_read_lock(); |
656 | for_each_domain(cpu, sd) { | 656 | for_each_domain(cpu, sd) { |
657 | for_each_cpu(i, sched_domain_span(sd)) { | 657 | for_each_cpu(i, sched_domain_span(sd)) { |
658 | if (!idle_cpu(i)) { | 658 | if (!idle_cpu(i)) { |
659 | cpu = i; | 659 | cpu = i; |
660 | goto unlock; | 660 | goto unlock; |
661 | } | 661 | } |
662 | } | 662 | } |
663 | } | 663 | } |
664 | unlock: | 664 | unlock: |
665 | rcu_read_unlock(); | 665 | rcu_read_unlock(); |
666 | return cpu; | 666 | return cpu; |
667 | } | 667 | } |
668 | /* | 668 | /* |
669 | * When add_timer_on() enqueues a timer into the timer wheel of an | 669 | * When add_timer_on() enqueues a timer into the timer wheel of an |
670 | * idle CPU then this timer might expire before the next timer event | 670 | * idle CPU then this timer might expire before the next timer event |
671 | * which is scheduled to wake up that CPU. In case of a completely | 671 | * which is scheduled to wake up that CPU. In case of a completely |
672 | * idle system the next event might even be infinite time into the | 672 | * idle system the next event might even be infinite time into the |
673 | * future. wake_up_idle_cpu() ensures that the CPU is woken up and | 673 | * future. wake_up_idle_cpu() ensures that the CPU is woken up and |
674 | * leaves the inner idle loop so the newly added timer is taken into | 674 | * leaves the inner idle loop so the newly added timer is taken into |
675 | * account when the CPU goes back to idle and evaluates the timer | 675 | * account when the CPU goes back to idle and evaluates the timer |
676 | * wheel for the next timer event. | 676 | * wheel for the next timer event. |
677 | */ | 677 | */ |
678 | static void wake_up_idle_cpu(int cpu) | 678 | static void wake_up_idle_cpu(int cpu) |
679 | { | 679 | { |
680 | struct rq *rq = cpu_rq(cpu); | 680 | struct rq *rq = cpu_rq(cpu); |
681 | 681 | ||
682 | if (cpu == smp_processor_id()) | 682 | if (cpu == smp_processor_id()) |
683 | return; | 683 | return; |
684 | 684 | ||
685 | if (set_nr_and_not_polling(rq->idle)) | 685 | if (set_nr_and_not_polling(rq->idle)) |
686 | smp_send_reschedule(cpu); | 686 | smp_send_reschedule(cpu); |
687 | else | 687 | else |
688 | trace_sched_wake_idle_without_ipi(cpu); | 688 | trace_sched_wake_idle_without_ipi(cpu); |
689 | } | 689 | } |
690 | 690 | ||
691 | static bool wake_up_full_nohz_cpu(int cpu) | 691 | static bool wake_up_full_nohz_cpu(int cpu) |
692 | { | 692 | { |
693 | /* | 693 | /* |
694 | * We just need the target to call irq_exit() and re-evaluate | 694 | * We just need the target to call irq_exit() and re-evaluate |
695 | * the next tick. The nohz full kick at least implies that. | 695 | * the next tick. The nohz full kick at least implies that. |
696 | * If needed we can still optimize that later with an | 696 | * If needed we can still optimize that later with an |
697 | * empty IRQ. | 697 | * empty IRQ. |
698 | */ | 698 | */ |
699 | if (tick_nohz_full_cpu(cpu)) { | 699 | if (tick_nohz_full_cpu(cpu)) { |
700 | if (cpu != smp_processor_id() || | 700 | if (cpu != smp_processor_id() || |
701 | tick_nohz_tick_stopped()) | 701 | tick_nohz_tick_stopped()) |
702 | tick_nohz_full_kick_cpu(cpu); | 702 | tick_nohz_full_kick_cpu(cpu); |
703 | return true; | 703 | return true; |
704 | } | 704 | } |
705 | 705 | ||
706 | return false; | 706 | return false; |
707 | } | 707 | } |
708 | 708 | ||
709 | void wake_up_nohz_cpu(int cpu) | 709 | void wake_up_nohz_cpu(int cpu) |
710 | { | 710 | { |
711 | if (!wake_up_full_nohz_cpu(cpu)) | 711 | if (!wake_up_full_nohz_cpu(cpu)) |
712 | wake_up_idle_cpu(cpu); | 712 | wake_up_idle_cpu(cpu); |
713 | } | 713 | } |
714 | 714 | ||
715 | static inline bool got_nohz_idle_kick(void) | 715 | static inline bool got_nohz_idle_kick(void) |
716 | { | 716 | { |
717 | int cpu = smp_processor_id(); | 717 | int cpu = smp_processor_id(); |
718 | 718 | ||
719 | if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) | 719 | if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) |
720 | return false; | 720 | return false; |
721 | 721 | ||
722 | if (idle_cpu(cpu) && !need_resched()) | 722 | if (idle_cpu(cpu) && !need_resched()) |
723 | return true; | 723 | return true; |
724 | 724 | ||
725 | /* | 725 | /* |
726 | * We can't run Idle Load Balance on this CPU for this time so we | 726 | * We can't run Idle Load Balance on this CPU for this time so we |
727 | * cancel it and clear NOHZ_BALANCE_KICK | 727 | * cancel it and clear NOHZ_BALANCE_KICK |
728 | */ | 728 | */ |
729 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); | 729 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); |
730 | return false; | 730 | return false; |
731 | } | 731 | } |
732 | 732 | ||
733 | #else /* CONFIG_NO_HZ_COMMON */ | 733 | #else /* CONFIG_NO_HZ_COMMON */ |
734 | 734 | ||
735 | static inline bool got_nohz_idle_kick(void) | 735 | static inline bool got_nohz_idle_kick(void) |
736 | { | 736 | { |
737 | return false; | 737 | return false; |
738 | } | 738 | } |
739 | 739 | ||
740 | #endif /* CONFIG_NO_HZ_COMMON */ | 740 | #endif /* CONFIG_NO_HZ_COMMON */ |
741 | 741 | ||
742 | #ifdef CONFIG_NO_HZ_FULL | 742 | #ifdef CONFIG_NO_HZ_FULL |
743 | bool sched_can_stop_tick(void) | 743 | bool sched_can_stop_tick(void) |
744 | { | 744 | { |
745 | /* | 745 | /* |
746 | * More than one running task need preemption. | 746 | * More than one running task need preemption. |
747 | * nr_running update is assumed to be visible | 747 | * nr_running update is assumed to be visible |
748 | * after IPI is sent from wakers. | 748 | * after IPI is sent from wakers. |
749 | */ | 749 | */ |
750 | if (this_rq()->nr_running > 1) | 750 | if (this_rq()->nr_running > 1) |
751 | return false; | 751 | return false; |
752 | 752 | ||
753 | return true; | 753 | return true; |
754 | } | 754 | } |
755 | #endif /* CONFIG_NO_HZ_FULL */ | 755 | #endif /* CONFIG_NO_HZ_FULL */ |
756 | 756 | ||
757 | void sched_avg_update(struct rq *rq) | 757 | void sched_avg_update(struct rq *rq) |
758 | { | 758 | { |
759 | s64 period = sched_avg_period(); | 759 | s64 period = sched_avg_period(); |
760 | 760 | ||
761 | while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { | 761 | while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { |
762 | /* | 762 | /* |
763 | * Inline assembly required to prevent the compiler | 763 | * Inline assembly required to prevent the compiler |
764 | * optimising this loop into a divmod call. | 764 | * optimising this loop into a divmod call. |
765 | * See __iter_div_u64_rem() for another example of this. | 765 | * See __iter_div_u64_rem() for another example of this. |
766 | */ | 766 | */ |
767 | asm("" : "+rm" (rq->age_stamp)); | 767 | asm("" : "+rm" (rq->age_stamp)); |
768 | rq->age_stamp += period; | 768 | rq->age_stamp += period; |
769 | rq->rt_avg /= 2; | 769 | rq->rt_avg /= 2; |
770 | } | 770 | } |
771 | } | 771 | } |
772 | 772 | ||
773 | #endif /* CONFIG_SMP */ | 773 | #endif /* CONFIG_SMP */ |
774 | 774 | ||
775 | #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ | 775 | #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ |
776 | (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) | 776 | (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) |
777 | /* | 777 | /* |
778 | * Iterate task_group tree rooted at *from, calling @down when first entering a | 778 | * Iterate task_group tree rooted at *from, calling @down when first entering a |
779 | * node and @up when leaving it for the final time. | 779 | * node and @up when leaving it for the final time. |
780 | * | 780 | * |
781 | * Caller must hold rcu_lock or sufficient equivalent. | 781 | * Caller must hold rcu_lock or sufficient equivalent. |
782 | */ | 782 | */ |
783 | int walk_tg_tree_from(struct task_group *from, | 783 | int walk_tg_tree_from(struct task_group *from, |
784 | tg_visitor down, tg_visitor up, void *data) | 784 | tg_visitor down, tg_visitor up, void *data) |
785 | { | 785 | { |
786 | struct task_group *parent, *child; | 786 | struct task_group *parent, *child; |
787 | int ret; | 787 | int ret; |
788 | 788 | ||
789 | parent = from; | 789 | parent = from; |
790 | 790 | ||
791 | down: | 791 | down: |
792 | ret = (*down)(parent, data); | 792 | ret = (*down)(parent, data); |
793 | if (ret) | 793 | if (ret) |
794 | goto out; | 794 | goto out; |
795 | list_for_each_entry_rcu(child, &parent->children, siblings) { | 795 | list_for_each_entry_rcu(child, &parent->children, siblings) { |
796 | parent = child; | 796 | parent = child; |
797 | goto down; | 797 | goto down; |
798 | 798 | ||
799 | up: | 799 | up: |
800 | continue; | 800 | continue; |
801 | } | 801 | } |
802 | ret = (*up)(parent, data); | 802 | ret = (*up)(parent, data); |
803 | if (ret || parent == from) | 803 | if (ret || parent == from) |
804 | goto out; | 804 | goto out; |
805 | 805 | ||
806 | child = parent; | 806 | child = parent; |
807 | parent = parent->parent; | 807 | parent = parent->parent; |
808 | if (parent) | 808 | if (parent) |
809 | goto up; | 809 | goto up; |
810 | out: | 810 | out: |
811 | return ret; | 811 | return ret; |
812 | } | 812 | } |
813 | 813 | ||
814 | int tg_nop(struct task_group *tg, void *data) | 814 | int tg_nop(struct task_group *tg, void *data) |
815 | { | 815 | { |
816 | return 0; | 816 | return 0; |
817 | } | 817 | } |
818 | #endif | 818 | #endif |
819 | 819 | ||
820 | static void set_load_weight(struct task_struct *p) | 820 | static void set_load_weight(struct task_struct *p) |
821 | { | 821 | { |
822 | int prio = p->static_prio - MAX_RT_PRIO; | 822 | int prio = p->static_prio - MAX_RT_PRIO; |
823 | struct load_weight *load = &p->se.load; | 823 | struct load_weight *load = &p->se.load; |
824 | 824 | ||
825 | /* | 825 | /* |
826 | * SCHED_IDLE tasks get minimal weight: | 826 | * SCHED_IDLE tasks get minimal weight: |
827 | */ | 827 | */ |
828 | if (p->policy == SCHED_IDLE) { | 828 | if (p->policy == SCHED_IDLE) { |
829 | load->weight = scale_load(WEIGHT_IDLEPRIO); | 829 | load->weight = scale_load(WEIGHT_IDLEPRIO); |
830 | load->inv_weight = WMULT_IDLEPRIO; | 830 | load->inv_weight = WMULT_IDLEPRIO; |
831 | return; | 831 | return; |
832 | } | 832 | } |
833 | 833 | ||
834 | load->weight = scale_load(prio_to_weight[prio]); | 834 | load->weight = scale_load(prio_to_weight[prio]); |
835 | load->inv_weight = prio_to_wmult[prio]; | 835 | load->inv_weight = prio_to_wmult[prio]; |
836 | } | 836 | } |
837 | 837 | ||
838 | static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) | 838 | static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) |
839 | { | 839 | { |
840 | update_rq_clock(rq); | 840 | update_rq_clock(rq); |
841 | sched_info_queued(rq, p); | 841 | sched_info_queued(rq, p); |
842 | p->sched_class->enqueue_task(rq, p, flags); | 842 | p->sched_class->enqueue_task(rq, p, flags); |
843 | } | 843 | } |
844 | 844 | ||
845 | static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) | 845 | static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) |
846 | { | 846 | { |
847 | update_rq_clock(rq); | 847 | update_rq_clock(rq); |
848 | sched_info_dequeued(rq, p); | 848 | sched_info_dequeued(rq, p); |
849 | p->sched_class->dequeue_task(rq, p, flags); | 849 | p->sched_class->dequeue_task(rq, p, flags); |
850 | } | 850 | } |
851 | 851 | ||
852 | void activate_task(struct rq *rq, struct task_struct *p, int flags) | 852 | void activate_task(struct rq *rq, struct task_struct *p, int flags) |
853 | { | 853 | { |
854 | if (task_contributes_to_load(p)) | 854 | if (task_contributes_to_load(p)) |
855 | rq->nr_uninterruptible--; | 855 | rq->nr_uninterruptible--; |
856 | 856 | ||
857 | enqueue_task(rq, p, flags); | 857 | enqueue_task(rq, p, flags); |
858 | } | 858 | } |
859 | 859 | ||
860 | void deactivate_task(struct rq *rq, struct task_struct *p, int flags) | 860 | void deactivate_task(struct rq *rq, struct task_struct *p, int flags) |
861 | { | 861 | { |
862 | if (task_contributes_to_load(p)) | 862 | if (task_contributes_to_load(p)) |
863 | rq->nr_uninterruptible++; | 863 | rq->nr_uninterruptible++; |
864 | 864 | ||
865 | dequeue_task(rq, p, flags); | 865 | dequeue_task(rq, p, flags); |
866 | } | 866 | } |
867 | 867 | ||
868 | static void update_rq_clock_task(struct rq *rq, s64 delta) | 868 | static void update_rq_clock_task(struct rq *rq, s64 delta) |
869 | { | 869 | { |
870 | /* | 870 | /* |
871 | * In theory, the compile should just see 0 here, and optimize out the call | 871 | * In theory, the compile should just see 0 here, and optimize out the call |
872 | * to sched_rt_avg_update. But I don't trust it... | 872 | * to sched_rt_avg_update. But I don't trust it... |
873 | */ | 873 | */ |
874 | #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) | 874 | #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) |
875 | s64 steal = 0, irq_delta = 0; | 875 | s64 steal = 0, irq_delta = 0; |
876 | #endif | 876 | #endif |
877 | #ifdef CONFIG_IRQ_TIME_ACCOUNTING | 877 | #ifdef CONFIG_IRQ_TIME_ACCOUNTING |
878 | irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; | 878 | irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; |
879 | 879 | ||
880 | /* | 880 | /* |
881 | * Since irq_time is only updated on {soft,}irq_exit, we might run into | 881 | * Since irq_time is only updated on {soft,}irq_exit, we might run into |
882 | * this case when a previous update_rq_clock() happened inside a | 882 | * this case when a previous update_rq_clock() happened inside a |
883 | * {soft,}irq region. | 883 | * {soft,}irq region. |
884 | * | 884 | * |
885 | * When this happens, we stop ->clock_task and only update the | 885 | * When this happens, we stop ->clock_task and only update the |
886 | * prev_irq_time stamp to account for the part that fit, so that a next | 886 | * prev_irq_time stamp to account for the part that fit, so that a next |
887 | * update will consume the rest. This ensures ->clock_task is | 887 | * update will consume the rest. This ensures ->clock_task is |
888 | * monotonic. | 888 | * monotonic. |
889 | * | 889 | * |
890 | * It does however cause some slight miss-attribution of {soft,}irq | 890 | * It does however cause some slight miss-attribution of {soft,}irq |
891 | * time, a more accurate solution would be to update the irq_time using | 891 | * time, a more accurate solution would be to update the irq_time using |
892 | * the current rq->clock timestamp, except that would require using | 892 | * the current rq->clock timestamp, except that would require using |
893 | * atomic ops. | 893 | * atomic ops. |
894 | */ | 894 | */ |
895 | if (irq_delta > delta) | 895 | if (irq_delta > delta) |
896 | irq_delta = delta; | 896 | irq_delta = delta; |
897 | 897 | ||
898 | rq->prev_irq_time += irq_delta; | 898 | rq->prev_irq_time += irq_delta; |
899 | delta -= irq_delta; | 899 | delta -= irq_delta; |
900 | #endif | 900 | #endif |
901 | #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING | 901 | #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING |
902 | if (static_key_false((¶virt_steal_rq_enabled))) { | 902 | if (static_key_false((¶virt_steal_rq_enabled))) { |
903 | steal = paravirt_steal_clock(cpu_of(rq)); | 903 | steal = paravirt_steal_clock(cpu_of(rq)); |
904 | steal -= rq->prev_steal_time_rq; | 904 | steal -= rq->prev_steal_time_rq; |
905 | 905 | ||
906 | if (unlikely(steal > delta)) | 906 | if (unlikely(steal > delta)) |
907 | steal = delta; | 907 | steal = delta; |
908 | 908 | ||
909 | rq->prev_steal_time_rq += steal; | 909 | rq->prev_steal_time_rq += steal; |
910 | delta -= steal; | 910 | delta -= steal; |
911 | } | 911 | } |
912 | #endif | 912 | #endif |
913 | 913 | ||
914 | rq->clock_task += delta; | 914 | rq->clock_task += delta; |
915 | 915 | ||
916 | #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) | 916 | #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) |
917 | if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY)) | 917 | if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY)) |
918 | sched_rt_avg_update(rq, irq_delta + steal); | 918 | sched_rt_avg_update(rq, irq_delta + steal); |
919 | #endif | 919 | #endif |
920 | } | 920 | } |
921 | 921 | ||
922 | void sched_set_stop_task(int cpu, struct task_struct *stop) | 922 | void sched_set_stop_task(int cpu, struct task_struct *stop) |
923 | { | 923 | { |
924 | struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; | 924 | struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; |
925 | struct task_struct *old_stop = cpu_rq(cpu)->stop; | 925 | struct task_struct *old_stop = cpu_rq(cpu)->stop; |
926 | 926 | ||
927 | if (stop) { | 927 | if (stop) { |
928 | /* | 928 | /* |
929 | * Make it appear like a SCHED_FIFO task, its something | 929 | * Make it appear like a SCHED_FIFO task, its something |
930 | * userspace knows about and won't get confused about. | 930 | * userspace knows about and won't get confused about. |
931 | * | 931 | * |
932 | * Also, it will make PI more or less work without too | 932 | * Also, it will make PI more or less work without too |
933 | * much confusion -- but then, stop work should not | 933 | * much confusion -- but then, stop work should not |
934 | * rely on PI working anyway. | 934 | * rely on PI working anyway. |
935 | */ | 935 | */ |
936 | sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); | 936 | sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); |
937 | 937 | ||
938 | stop->sched_class = &stop_sched_class; | 938 | stop->sched_class = &stop_sched_class; |
939 | } | 939 | } |
940 | 940 | ||
941 | cpu_rq(cpu)->stop = stop; | 941 | cpu_rq(cpu)->stop = stop; |
942 | 942 | ||
943 | if (old_stop) { | 943 | if (old_stop) { |
944 | /* | 944 | /* |
945 | * Reset it back to a normal scheduling class so that | 945 | * Reset it back to a normal scheduling class so that |
946 | * it can die in pieces. | 946 | * it can die in pieces. |
947 | */ | 947 | */ |
948 | old_stop->sched_class = &rt_sched_class; | 948 | old_stop->sched_class = &rt_sched_class; |
949 | } | 949 | } |
950 | } | 950 | } |
951 | 951 | ||
952 | /* | 952 | /* |
953 | * __normal_prio - return the priority that is based on the static prio | 953 | * __normal_prio - return the priority that is based on the static prio |
954 | */ | 954 | */ |
955 | static inline int __normal_prio(struct task_struct *p) | 955 | static inline int __normal_prio(struct task_struct *p) |
956 | { | 956 | { |
957 | return p->static_prio; | 957 | return p->static_prio; |
958 | } | 958 | } |
959 | 959 | ||
960 | /* | 960 | /* |
961 | * Calculate the expected normal priority: i.e. priority | 961 | * Calculate the expected normal priority: i.e. priority |
962 | * without taking RT-inheritance into account. Might be | 962 | * without taking RT-inheritance into account. Might be |
963 | * boosted by interactivity modifiers. Changes upon fork, | 963 | * boosted by interactivity modifiers. Changes upon fork, |
964 | * setprio syscalls, and whenever the interactivity | 964 | * setprio syscalls, and whenever the interactivity |
965 | * estimator recalculates. | 965 | * estimator recalculates. |
966 | */ | 966 | */ |
967 | static inline int normal_prio(struct task_struct *p) | 967 | static inline int normal_prio(struct task_struct *p) |
968 | { | 968 | { |
969 | int prio; | 969 | int prio; |
970 | 970 | ||
971 | if (task_has_dl_policy(p)) | 971 | if (task_has_dl_policy(p)) |
972 | prio = MAX_DL_PRIO-1; | 972 | prio = MAX_DL_PRIO-1; |
973 | else if (task_has_rt_policy(p)) | 973 | else if (task_has_rt_policy(p)) |
974 | prio = MAX_RT_PRIO-1 - p->rt_priority; | 974 | prio = MAX_RT_PRIO-1 - p->rt_priority; |
975 | else | 975 | else |
976 | prio = __normal_prio(p); | 976 | prio = __normal_prio(p); |
977 | return prio; | 977 | return prio; |
978 | } | 978 | } |
979 | 979 | ||
980 | /* | 980 | /* |
981 | * Calculate the current priority, i.e. the priority | 981 | * Calculate the current priority, i.e. the priority |
982 | * taken into account by the scheduler. This value might | 982 | * taken into account by the scheduler. This value might |
983 | * be boosted by RT tasks, or might be boosted by | 983 | * be boosted by RT tasks, or might be boosted by |
984 | * interactivity modifiers. Will be RT if the task got | 984 | * interactivity modifiers. Will be RT if the task got |
985 | * RT-boosted. If not then it returns p->normal_prio. | 985 | * RT-boosted. If not then it returns p->normal_prio. |
986 | */ | 986 | */ |
987 | static int effective_prio(struct task_struct *p) | 987 | static int effective_prio(struct task_struct *p) |
988 | { | 988 | { |
989 | p->normal_prio = normal_prio(p); | 989 | p->normal_prio = normal_prio(p); |
990 | /* | 990 | /* |
991 | * If we are RT tasks or we were boosted to RT priority, | 991 | * If we are RT tasks or we were boosted to RT priority, |
992 | * keep the priority unchanged. Otherwise, update priority | 992 | * keep the priority unchanged. Otherwise, update priority |
993 | * to the normal priority: | 993 | * to the normal priority: |
994 | */ | 994 | */ |
995 | if (!rt_prio(p->prio)) | 995 | if (!rt_prio(p->prio)) |
996 | return p->normal_prio; | 996 | return p->normal_prio; |
997 | return p->prio; | 997 | return p->prio; |
998 | } | 998 | } |
999 | 999 | ||
1000 | /** | 1000 | /** |
1001 | * task_curr - is this task currently executing on a CPU? | 1001 | * task_curr - is this task currently executing on a CPU? |
1002 | * @p: the task in question. | 1002 | * @p: the task in question. |
1003 | * | 1003 | * |
1004 | * Return: 1 if the task is currently executing. 0 otherwise. | 1004 | * Return: 1 if the task is currently executing. 0 otherwise. |
1005 | */ | 1005 | */ |
1006 | inline int task_curr(const struct task_struct *p) | 1006 | inline int task_curr(const struct task_struct *p) |
1007 | { | 1007 | { |
1008 | return cpu_curr(task_cpu(p)) == p; | 1008 | return cpu_curr(task_cpu(p)) == p; |
1009 | } | 1009 | } |
1010 | 1010 | ||
1011 | /* | 1011 | /* |
1012 | * Can drop rq->lock because from sched_class::switched_from() methods drop it. | 1012 | * Can drop rq->lock because from sched_class::switched_from() methods drop it. |
1013 | */ | 1013 | */ |
1014 | static inline void check_class_changed(struct rq *rq, struct task_struct *p, | 1014 | static inline void check_class_changed(struct rq *rq, struct task_struct *p, |
1015 | const struct sched_class *prev_class, | 1015 | const struct sched_class *prev_class, |
1016 | int oldprio) | 1016 | int oldprio) |
1017 | { | 1017 | { |
1018 | if (prev_class != p->sched_class) { | 1018 | if (prev_class != p->sched_class) { |
1019 | if (prev_class->switched_from) | 1019 | if (prev_class->switched_from) |
1020 | prev_class->switched_from(rq, p); | 1020 | prev_class->switched_from(rq, p); |
1021 | /* Possble rq->lock 'hole'. */ | 1021 | /* Possble rq->lock 'hole'. */ |
1022 | p->sched_class->switched_to(rq, p); | 1022 | p->sched_class->switched_to(rq, p); |
1023 | } else if (oldprio != p->prio || dl_task(p)) | 1023 | } else if (oldprio != p->prio || dl_task(p)) |
1024 | p->sched_class->prio_changed(rq, p, oldprio); | 1024 | p->sched_class->prio_changed(rq, p, oldprio); |
1025 | } | 1025 | } |
1026 | 1026 | ||
1027 | void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) | 1027 | void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) |
1028 | { | 1028 | { |
1029 | const struct sched_class *class; | 1029 | const struct sched_class *class; |
1030 | 1030 | ||
1031 | if (p->sched_class == rq->curr->sched_class) { | 1031 | if (p->sched_class == rq->curr->sched_class) { |
1032 | rq->curr->sched_class->check_preempt_curr(rq, p, flags); | 1032 | rq->curr->sched_class->check_preempt_curr(rq, p, flags); |
1033 | } else { | 1033 | } else { |
1034 | for_each_class(class) { | 1034 | for_each_class(class) { |
1035 | if (class == rq->curr->sched_class) | 1035 | if (class == rq->curr->sched_class) |
1036 | break; | 1036 | break; |
1037 | if (class == p->sched_class) { | 1037 | if (class == p->sched_class) { |
1038 | resched_curr(rq); | 1038 | resched_curr(rq); |
1039 | break; | 1039 | break; |
1040 | } | 1040 | } |
1041 | } | 1041 | } |
1042 | } | 1042 | } |
1043 | 1043 | ||
1044 | /* | 1044 | /* |
1045 | * A queue event has occurred, and we're going to schedule. In | 1045 | * A queue event has occurred, and we're going to schedule. In |
1046 | * this case, we can save a useless back to back clock update. | 1046 | * this case, we can save a useless back to back clock update. |
1047 | */ | 1047 | */ |
1048 | if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr)) | 1048 | if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr)) |
1049 | rq->skip_clock_update = 1; | 1049 | rq->skip_clock_update = 1; |
1050 | } | 1050 | } |
1051 | 1051 | ||
1052 | #ifdef CONFIG_SMP | 1052 | #ifdef CONFIG_SMP |
1053 | void set_task_cpu(struct task_struct *p, unsigned int new_cpu) | 1053 | void set_task_cpu(struct task_struct *p, unsigned int new_cpu) |
1054 | { | 1054 | { |
1055 | #ifdef CONFIG_SCHED_DEBUG | 1055 | #ifdef CONFIG_SCHED_DEBUG |
1056 | /* | 1056 | /* |
1057 | * We should never call set_task_cpu() on a blocked task, | 1057 | * We should never call set_task_cpu() on a blocked task, |
1058 | * ttwu() will sort out the placement. | 1058 | * ttwu() will sort out the placement. |
1059 | */ | 1059 | */ |
1060 | WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && | 1060 | WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && |
1061 | !p->on_rq); | 1061 | !p->on_rq); |
1062 | 1062 | ||
1063 | #ifdef CONFIG_LOCKDEP | 1063 | #ifdef CONFIG_LOCKDEP |
1064 | /* | 1064 | /* |
1065 | * The caller should hold either p->pi_lock or rq->lock, when changing | 1065 | * The caller should hold either p->pi_lock or rq->lock, when changing |
1066 | * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. | 1066 | * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. |
1067 | * | 1067 | * |
1068 | * sched_move_task() holds both and thus holding either pins the cgroup, | 1068 | * sched_move_task() holds both and thus holding either pins the cgroup, |
1069 | * see task_group(). | 1069 | * see task_group(). |
1070 | * | 1070 | * |
1071 | * Furthermore, all task_rq users should acquire both locks, see | 1071 | * Furthermore, all task_rq users should acquire both locks, see |
1072 | * task_rq_lock(). | 1072 | * task_rq_lock(). |
1073 | */ | 1073 | */ |
1074 | WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || | 1074 | WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || |
1075 | lockdep_is_held(&task_rq(p)->lock))); | 1075 | lockdep_is_held(&task_rq(p)->lock))); |
1076 | #endif | 1076 | #endif |
1077 | #endif | 1077 | #endif |
1078 | 1078 | ||
1079 | trace_sched_migrate_task(p, new_cpu); | 1079 | trace_sched_migrate_task(p, new_cpu); |
1080 | 1080 | ||
1081 | if (task_cpu(p) != new_cpu) { | 1081 | if (task_cpu(p) != new_cpu) { |
1082 | if (p->sched_class->migrate_task_rq) | 1082 | if (p->sched_class->migrate_task_rq) |
1083 | p->sched_class->migrate_task_rq(p, new_cpu); | 1083 | p->sched_class->migrate_task_rq(p, new_cpu); |
1084 | p->se.nr_migrations++; | 1084 | p->se.nr_migrations++; |
1085 | perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); | 1085 | perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); |
1086 | } | 1086 | } |
1087 | 1087 | ||
1088 | __set_task_cpu(p, new_cpu); | 1088 | __set_task_cpu(p, new_cpu); |
1089 | } | 1089 | } |
1090 | 1090 | ||
1091 | static void __migrate_swap_task(struct task_struct *p, int cpu) | 1091 | static void __migrate_swap_task(struct task_struct *p, int cpu) |
1092 | { | 1092 | { |
1093 | if (task_on_rq_queued(p)) { | 1093 | if (task_on_rq_queued(p)) { |
1094 | struct rq *src_rq, *dst_rq; | 1094 | struct rq *src_rq, *dst_rq; |
1095 | 1095 | ||
1096 | src_rq = task_rq(p); | 1096 | src_rq = task_rq(p); |
1097 | dst_rq = cpu_rq(cpu); | 1097 | dst_rq = cpu_rq(cpu); |
1098 | 1098 | ||
1099 | deactivate_task(src_rq, p, 0); | 1099 | deactivate_task(src_rq, p, 0); |
1100 | set_task_cpu(p, cpu); | 1100 | set_task_cpu(p, cpu); |
1101 | activate_task(dst_rq, p, 0); | 1101 | activate_task(dst_rq, p, 0); |
1102 | check_preempt_curr(dst_rq, p, 0); | 1102 | check_preempt_curr(dst_rq, p, 0); |
1103 | } else { | 1103 | } else { |
1104 | /* | 1104 | /* |
1105 | * Task isn't running anymore; make it appear like we migrated | 1105 | * Task isn't running anymore; make it appear like we migrated |
1106 | * it before it went to sleep. This means on wakeup we make the | 1106 | * it before it went to sleep. This means on wakeup we make the |
1107 | * previous cpu our targer instead of where it really is. | 1107 | * previous cpu our targer instead of where it really is. |
1108 | */ | 1108 | */ |
1109 | p->wake_cpu = cpu; | 1109 | p->wake_cpu = cpu; |
1110 | } | 1110 | } |
1111 | } | 1111 | } |
1112 | 1112 | ||
1113 | struct migration_swap_arg { | 1113 | struct migration_swap_arg { |
1114 | struct task_struct *src_task, *dst_task; | 1114 | struct task_struct *src_task, *dst_task; |
1115 | int src_cpu, dst_cpu; | 1115 | int src_cpu, dst_cpu; |
1116 | }; | 1116 | }; |
1117 | 1117 | ||
1118 | static int migrate_swap_stop(void *data) | 1118 | static int migrate_swap_stop(void *data) |
1119 | { | 1119 | { |
1120 | struct migration_swap_arg *arg = data; | 1120 | struct migration_swap_arg *arg = data; |
1121 | struct rq *src_rq, *dst_rq; | 1121 | struct rq *src_rq, *dst_rq; |
1122 | int ret = -EAGAIN; | 1122 | int ret = -EAGAIN; |
1123 | 1123 | ||
1124 | src_rq = cpu_rq(arg->src_cpu); | 1124 | src_rq = cpu_rq(arg->src_cpu); |
1125 | dst_rq = cpu_rq(arg->dst_cpu); | 1125 | dst_rq = cpu_rq(arg->dst_cpu); |
1126 | 1126 | ||
1127 | double_raw_lock(&arg->src_task->pi_lock, | 1127 | double_raw_lock(&arg->src_task->pi_lock, |
1128 | &arg->dst_task->pi_lock); | 1128 | &arg->dst_task->pi_lock); |
1129 | double_rq_lock(src_rq, dst_rq); | 1129 | double_rq_lock(src_rq, dst_rq); |
1130 | if (task_cpu(arg->dst_task) != arg->dst_cpu) | 1130 | if (task_cpu(arg->dst_task) != arg->dst_cpu) |
1131 | goto unlock; | 1131 | goto unlock; |
1132 | 1132 | ||
1133 | if (task_cpu(arg->src_task) != arg->src_cpu) | 1133 | if (task_cpu(arg->src_task) != arg->src_cpu) |
1134 | goto unlock; | 1134 | goto unlock; |
1135 | 1135 | ||
1136 | if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) | 1136 | if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) |
1137 | goto unlock; | 1137 | goto unlock; |
1138 | 1138 | ||
1139 | if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) | 1139 | if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) |
1140 | goto unlock; | 1140 | goto unlock; |
1141 | 1141 | ||
1142 | __migrate_swap_task(arg->src_task, arg->dst_cpu); | 1142 | __migrate_swap_task(arg->src_task, arg->dst_cpu); |
1143 | __migrate_swap_task(arg->dst_task, arg->src_cpu); | 1143 | __migrate_swap_task(arg->dst_task, arg->src_cpu); |
1144 | 1144 | ||
1145 | ret = 0; | 1145 | ret = 0; |
1146 | 1146 | ||
1147 | unlock: | 1147 | unlock: |
1148 | double_rq_unlock(src_rq, dst_rq); | 1148 | double_rq_unlock(src_rq, dst_rq); |
1149 | raw_spin_unlock(&arg->dst_task->pi_lock); | 1149 | raw_spin_unlock(&arg->dst_task->pi_lock); |
1150 | raw_spin_unlock(&arg->src_task->pi_lock); | 1150 | raw_spin_unlock(&arg->src_task->pi_lock); |
1151 | 1151 | ||
1152 | return ret; | 1152 | return ret; |
1153 | } | 1153 | } |
1154 | 1154 | ||
1155 | /* | 1155 | /* |
1156 | * Cross migrate two tasks | 1156 | * Cross migrate two tasks |
1157 | */ | 1157 | */ |
1158 | int migrate_swap(struct task_struct *cur, struct task_struct *p) | 1158 | int migrate_swap(struct task_struct *cur, struct task_struct *p) |
1159 | { | 1159 | { |
1160 | struct migration_swap_arg arg; | 1160 | struct migration_swap_arg arg; |
1161 | int ret = -EINVAL; | 1161 | int ret = -EINVAL; |
1162 | 1162 | ||
1163 | arg = (struct migration_swap_arg){ | 1163 | arg = (struct migration_swap_arg){ |
1164 | .src_task = cur, | 1164 | .src_task = cur, |
1165 | .src_cpu = task_cpu(cur), | 1165 | .src_cpu = task_cpu(cur), |
1166 | .dst_task = p, | 1166 | .dst_task = p, |
1167 | .dst_cpu = task_cpu(p), | 1167 | .dst_cpu = task_cpu(p), |
1168 | }; | 1168 | }; |
1169 | 1169 | ||
1170 | if (arg.src_cpu == arg.dst_cpu) | 1170 | if (arg.src_cpu == arg.dst_cpu) |
1171 | goto out; | 1171 | goto out; |
1172 | 1172 | ||
1173 | /* | 1173 | /* |
1174 | * These three tests are all lockless; this is OK since all of them | 1174 | * These three tests are all lockless; this is OK since all of them |
1175 | * will be re-checked with proper locks held further down the line. | 1175 | * will be re-checked with proper locks held further down the line. |
1176 | */ | 1176 | */ |
1177 | if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) | 1177 | if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) |
1178 | goto out; | 1178 | goto out; |
1179 | 1179 | ||
1180 | if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) | 1180 | if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) |
1181 | goto out; | 1181 | goto out; |
1182 | 1182 | ||
1183 | if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) | 1183 | if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) |
1184 | goto out; | 1184 | goto out; |
1185 | 1185 | ||
1186 | trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); | 1186 | trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); |
1187 | ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); | 1187 | ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); |
1188 | 1188 | ||
1189 | out: | 1189 | out: |
1190 | return ret; | 1190 | return ret; |
1191 | } | 1191 | } |
1192 | 1192 | ||
1193 | struct migration_arg { | 1193 | struct migration_arg { |
1194 | struct task_struct *task; | 1194 | struct task_struct *task; |
1195 | int dest_cpu; | 1195 | int dest_cpu; |
1196 | }; | 1196 | }; |
1197 | 1197 | ||
1198 | static int migration_cpu_stop(void *data); | 1198 | static int migration_cpu_stop(void *data); |
1199 | 1199 | ||
1200 | /* | 1200 | /* |
1201 | * wait_task_inactive - wait for a thread to unschedule. | 1201 | * wait_task_inactive - wait for a thread to unschedule. |
1202 | * | 1202 | * |
1203 | * If @match_state is nonzero, it's the @p->state value just checked and | 1203 | * If @match_state is nonzero, it's the @p->state value just checked and |
1204 | * not expected to change. If it changes, i.e. @p might have woken up, | 1204 | * not expected to change. If it changes, i.e. @p might have woken up, |
1205 | * then return zero. When we succeed in waiting for @p to be off its CPU, | 1205 | * then return zero. When we succeed in waiting for @p to be off its CPU, |
1206 | * we return a positive number (its total switch count). If a second call | 1206 | * we return a positive number (its total switch count). If a second call |
1207 | * a short while later returns the same number, the caller can be sure that | 1207 | * a short while later returns the same number, the caller can be sure that |
1208 | * @p has remained unscheduled the whole time. | 1208 | * @p has remained unscheduled the whole time. |
1209 | * | 1209 | * |
1210 | * The caller must ensure that the task *will* unschedule sometime soon, | 1210 | * The caller must ensure that the task *will* unschedule sometime soon, |
1211 | * else this function might spin for a *long* time. This function can't | 1211 | * else this function might spin for a *long* time. This function can't |
1212 | * be called with interrupts off, or it may introduce deadlock with | 1212 | * be called with interrupts off, or it may introduce deadlock with |
1213 | * smp_call_function() if an IPI is sent by the same process we are | 1213 | * smp_call_function() if an IPI is sent by the same process we are |
1214 | * waiting to become inactive. | 1214 | * waiting to become inactive. |
1215 | */ | 1215 | */ |
1216 | unsigned long wait_task_inactive(struct task_struct *p, long match_state) | 1216 | unsigned long wait_task_inactive(struct task_struct *p, long match_state) |
1217 | { | 1217 | { |
1218 | unsigned long flags; | 1218 | unsigned long flags; |
1219 | int running, queued; | 1219 | int running, queued; |
1220 | unsigned long ncsw; | 1220 | unsigned long ncsw; |
1221 | struct rq *rq; | 1221 | struct rq *rq; |
1222 | 1222 | ||
1223 | for (;;) { | 1223 | for (;;) { |
1224 | /* | 1224 | /* |
1225 | * We do the initial early heuristics without holding | 1225 | * We do the initial early heuristics without holding |
1226 | * any task-queue locks at all. We'll only try to get | 1226 | * any task-queue locks at all. We'll only try to get |
1227 | * the runqueue lock when things look like they will | 1227 | * the runqueue lock when things look like they will |
1228 | * work out! | 1228 | * work out! |
1229 | */ | 1229 | */ |
1230 | rq = task_rq(p); | 1230 | rq = task_rq(p); |
1231 | 1231 | ||
1232 | /* | 1232 | /* |
1233 | * If the task is actively running on another CPU | 1233 | * If the task is actively running on another CPU |
1234 | * still, just relax and busy-wait without holding | 1234 | * still, just relax and busy-wait without holding |
1235 | * any locks. | 1235 | * any locks. |
1236 | * | 1236 | * |
1237 | * NOTE! Since we don't hold any locks, it's not | 1237 | * NOTE! Since we don't hold any locks, it's not |
1238 | * even sure that "rq" stays as the right runqueue! | 1238 | * even sure that "rq" stays as the right runqueue! |
1239 | * But we don't care, since "task_running()" will | 1239 | * But we don't care, since "task_running()" will |
1240 | * return false if the runqueue has changed and p | 1240 | * return false if the runqueue has changed and p |
1241 | * is actually now running somewhere else! | 1241 | * is actually now running somewhere else! |
1242 | */ | 1242 | */ |
1243 | while (task_running(rq, p)) { | 1243 | while (task_running(rq, p)) { |
1244 | if (match_state && unlikely(p->state != match_state)) | 1244 | if (match_state && unlikely(p->state != match_state)) |
1245 | return 0; | 1245 | return 0; |
1246 | cpu_relax(); | 1246 | cpu_relax(); |
1247 | } | 1247 | } |
1248 | 1248 | ||
1249 | /* | 1249 | /* |
1250 | * Ok, time to look more closely! We need the rq | 1250 | * Ok, time to look more closely! We need the rq |
1251 | * lock now, to be *sure*. If we're wrong, we'll | 1251 | * lock now, to be *sure*. If we're wrong, we'll |
1252 | * just go back and repeat. | 1252 | * just go back and repeat. |
1253 | */ | 1253 | */ |
1254 | rq = task_rq_lock(p, &flags); | 1254 | rq = task_rq_lock(p, &flags); |
1255 | trace_sched_wait_task(p); | 1255 | trace_sched_wait_task(p); |
1256 | running = task_running(rq, p); | 1256 | running = task_running(rq, p); |
1257 | queued = task_on_rq_queued(p); | 1257 | queued = task_on_rq_queued(p); |
1258 | ncsw = 0; | 1258 | ncsw = 0; |
1259 | if (!match_state || p->state == match_state) | 1259 | if (!match_state || p->state == match_state) |
1260 | ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ | 1260 | ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ |
1261 | task_rq_unlock(rq, p, &flags); | 1261 | task_rq_unlock(rq, p, &flags); |
1262 | 1262 | ||
1263 | /* | 1263 | /* |
1264 | * If it changed from the expected state, bail out now. | 1264 | * If it changed from the expected state, bail out now. |
1265 | */ | 1265 | */ |
1266 | if (unlikely(!ncsw)) | 1266 | if (unlikely(!ncsw)) |
1267 | break; | 1267 | break; |
1268 | 1268 | ||
1269 | /* | 1269 | /* |
1270 | * Was it really running after all now that we | 1270 | * Was it really running after all now that we |
1271 | * checked with the proper locks actually held? | 1271 | * checked with the proper locks actually held? |
1272 | * | 1272 | * |
1273 | * Oops. Go back and try again.. | 1273 | * Oops. Go back and try again.. |
1274 | */ | 1274 | */ |
1275 | if (unlikely(running)) { | 1275 | if (unlikely(running)) { |
1276 | cpu_relax(); | 1276 | cpu_relax(); |
1277 | continue; | 1277 | continue; |
1278 | } | 1278 | } |
1279 | 1279 | ||
1280 | /* | 1280 | /* |
1281 | * It's not enough that it's not actively running, | 1281 | * It's not enough that it's not actively running, |
1282 | * it must be off the runqueue _entirely_, and not | 1282 | * it must be off the runqueue _entirely_, and not |
1283 | * preempted! | 1283 | * preempted! |
1284 | * | 1284 | * |
1285 | * So if it was still runnable (but just not actively | 1285 | * So if it was still runnable (but just not actively |
1286 | * running right now), it's preempted, and we should | 1286 | * running right now), it's preempted, and we should |
1287 | * yield - it could be a while. | 1287 | * yield - it could be a while. |
1288 | */ | 1288 | */ |
1289 | if (unlikely(queued)) { | 1289 | if (unlikely(queued)) { |
1290 | ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); | 1290 | ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); |
1291 | 1291 | ||
1292 | set_current_state(TASK_UNINTERRUPTIBLE); | 1292 | set_current_state(TASK_UNINTERRUPTIBLE); |
1293 | schedule_hrtimeout(&to, HRTIMER_MODE_REL); | 1293 | schedule_hrtimeout(&to, HRTIMER_MODE_REL); |
1294 | continue; | 1294 | continue; |
1295 | } | 1295 | } |
1296 | 1296 | ||
1297 | /* | 1297 | /* |
1298 | * Ahh, all good. It wasn't running, and it wasn't | 1298 | * Ahh, all good. It wasn't running, and it wasn't |
1299 | * runnable, which means that it will never become | 1299 | * runnable, which means that it will never become |
1300 | * running in the future either. We're all done! | 1300 | * running in the future either. We're all done! |
1301 | */ | 1301 | */ |
1302 | break; | 1302 | break; |
1303 | } | 1303 | } |
1304 | 1304 | ||
1305 | return ncsw; | 1305 | return ncsw; |
1306 | } | 1306 | } |
1307 | 1307 | ||
1308 | /*** | 1308 | /*** |
1309 | * kick_process - kick a running thread to enter/exit the kernel | 1309 | * kick_process - kick a running thread to enter/exit the kernel |
1310 | * @p: the to-be-kicked thread | 1310 | * @p: the to-be-kicked thread |
1311 | * | 1311 | * |
1312 | * Cause a process which is running on another CPU to enter | 1312 | * Cause a process which is running on another CPU to enter |
1313 | * kernel-mode, without any delay. (to get signals handled.) | 1313 | * kernel-mode, without any delay. (to get signals handled.) |
1314 | * | 1314 | * |
1315 | * NOTE: this function doesn't have to take the runqueue lock, | 1315 | * NOTE: this function doesn't have to take the runqueue lock, |
1316 | * because all it wants to ensure is that the remote task enters | 1316 | * because all it wants to ensure is that the remote task enters |
1317 | * the kernel. If the IPI races and the task has been migrated | 1317 | * the kernel. If the IPI races and the task has been migrated |
1318 | * to another CPU then no harm is done and the purpose has been | 1318 | * to another CPU then no harm is done and the purpose has been |
1319 | * achieved as well. | 1319 | * achieved as well. |
1320 | */ | 1320 | */ |
1321 | void kick_process(struct task_struct *p) | 1321 | void kick_process(struct task_struct *p) |
1322 | { | 1322 | { |
1323 | int cpu; | 1323 | int cpu; |
1324 | 1324 | ||
1325 | preempt_disable(); | 1325 | preempt_disable(); |
1326 | cpu = task_cpu(p); | 1326 | cpu = task_cpu(p); |
1327 | if ((cpu != smp_processor_id()) && task_curr(p)) | 1327 | if ((cpu != smp_processor_id()) && task_curr(p)) |
1328 | smp_send_reschedule(cpu); | 1328 | smp_send_reschedule(cpu); |
1329 | preempt_enable(); | 1329 | preempt_enable(); |
1330 | } | 1330 | } |
1331 | EXPORT_SYMBOL_GPL(kick_process); | 1331 | EXPORT_SYMBOL_GPL(kick_process); |
1332 | #endif /* CONFIG_SMP */ | 1332 | #endif /* CONFIG_SMP */ |
1333 | 1333 | ||
1334 | #ifdef CONFIG_SMP | 1334 | #ifdef CONFIG_SMP |
1335 | /* | 1335 | /* |
1336 | * ->cpus_allowed is protected by both rq->lock and p->pi_lock | 1336 | * ->cpus_allowed is protected by both rq->lock and p->pi_lock |
1337 | */ | 1337 | */ |
1338 | static int select_fallback_rq(int cpu, struct task_struct *p) | 1338 | static int select_fallback_rq(int cpu, struct task_struct *p) |
1339 | { | 1339 | { |
1340 | int nid = cpu_to_node(cpu); | 1340 | int nid = cpu_to_node(cpu); |
1341 | const struct cpumask *nodemask = NULL; | 1341 | const struct cpumask *nodemask = NULL; |
1342 | enum { cpuset, possible, fail } state = cpuset; | 1342 | enum { cpuset, possible, fail } state = cpuset; |
1343 | int dest_cpu; | 1343 | int dest_cpu; |
1344 | 1344 | ||
1345 | /* | 1345 | /* |
1346 | * If the node that the cpu is on has been offlined, cpu_to_node() | 1346 | * If the node that the cpu is on has been offlined, cpu_to_node() |
1347 | * will return -1. There is no cpu on the node, and we should | 1347 | * will return -1. There is no cpu on the node, and we should |
1348 | * select the cpu on the other node. | 1348 | * select the cpu on the other node. |
1349 | */ | 1349 | */ |
1350 | if (nid != -1) { | 1350 | if (nid != -1) { |
1351 | nodemask = cpumask_of_node(nid); | 1351 | nodemask = cpumask_of_node(nid); |
1352 | 1352 | ||
1353 | /* Look for allowed, online CPU in same node. */ | 1353 | /* Look for allowed, online CPU in same node. */ |
1354 | for_each_cpu(dest_cpu, nodemask) { | 1354 | for_each_cpu(dest_cpu, nodemask) { |
1355 | if (!cpu_online(dest_cpu)) | 1355 | if (!cpu_online(dest_cpu)) |
1356 | continue; | 1356 | continue; |
1357 | if (!cpu_active(dest_cpu)) | 1357 | if (!cpu_active(dest_cpu)) |
1358 | continue; | 1358 | continue; |
1359 | if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) | 1359 | if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) |
1360 | return dest_cpu; | 1360 | return dest_cpu; |
1361 | } | 1361 | } |
1362 | } | 1362 | } |
1363 | 1363 | ||
1364 | for (;;) { | 1364 | for (;;) { |
1365 | /* Any allowed, online CPU? */ | 1365 | /* Any allowed, online CPU? */ |
1366 | for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { | 1366 | for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { |
1367 | if (!cpu_online(dest_cpu)) | 1367 | if (!cpu_online(dest_cpu)) |
1368 | continue; | 1368 | continue; |
1369 | if (!cpu_active(dest_cpu)) | 1369 | if (!cpu_active(dest_cpu)) |
1370 | continue; | 1370 | continue; |
1371 | goto out; | 1371 | goto out; |
1372 | } | 1372 | } |
1373 | 1373 | ||
1374 | switch (state) { | 1374 | switch (state) { |
1375 | case cpuset: | 1375 | case cpuset: |
1376 | /* No more Mr. Nice Guy. */ | 1376 | /* No more Mr. Nice Guy. */ |
1377 | cpuset_cpus_allowed_fallback(p); | 1377 | cpuset_cpus_allowed_fallback(p); |
1378 | state = possible; | 1378 | state = possible; |
1379 | break; | 1379 | break; |
1380 | 1380 | ||
1381 | case possible: | 1381 | case possible: |
1382 | do_set_cpus_allowed(p, cpu_possible_mask); | 1382 | do_set_cpus_allowed(p, cpu_possible_mask); |
1383 | state = fail; | 1383 | state = fail; |
1384 | break; | 1384 | break; |
1385 | 1385 | ||
1386 | case fail: | 1386 | case fail: |
1387 | BUG(); | 1387 | BUG(); |
1388 | break; | 1388 | break; |
1389 | } | 1389 | } |
1390 | } | 1390 | } |
1391 | 1391 | ||
1392 | out: | 1392 | out: |
1393 | if (state != cpuset) { | 1393 | if (state != cpuset) { |
1394 | /* | 1394 | /* |
1395 | * Don't tell them about moving exiting tasks or | 1395 | * Don't tell them about moving exiting tasks or |
1396 | * kernel threads (both mm NULL), since they never | 1396 | * kernel threads (both mm NULL), since they never |
1397 | * leave kernel. | 1397 | * leave kernel. |
1398 | */ | 1398 | */ |
1399 | if (p->mm && printk_ratelimit()) { | 1399 | if (p->mm && printk_ratelimit()) { |
1400 | printk_deferred("process %d (%s) no longer affine to cpu%d\n", | 1400 | printk_deferred("process %d (%s) no longer affine to cpu%d\n", |
1401 | task_pid_nr(p), p->comm, cpu); | 1401 | task_pid_nr(p), p->comm, cpu); |
1402 | } | 1402 | } |
1403 | } | 1403 | } |
1404 | 1404 | ||
1405 | return dest_cpu; | 1405 | return dest_cpu; |
1406 | } | 1406 | } |
1407 | 1407 | ||
1408 | /* | 1408 | /* |
1409 | * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. | 1409 | * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. |
1410 | */ | 1410 | */ |
1411 | static inline | 1411 | static inline |
1412 | int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) | 1412 | int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) |
1413 | { | 1413 | { |
1414 | if (p->nr_cpus_allowed > 1) | 1414 | if (p->nr_cpus_allowed > 1) |
1415 | cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); | 1415 | cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); |
1416 | 1416 | ||
1417 | /* | 1417 | /* |
1418 | * In order not to call set_task_cpu() on a blocking task we need | 1418 | * In order not to call set_task_cpu() on a blocking task we need |
1419 | * to rely on ttwu() to place the task on a valid ->cpus_allowed | 1419 | * to rely on ttwu() to place the task on a valid ->cpus_allowed |
1420 | * cpu. | 1420 | * cpu. |
1421 | * | 1421 | * |
1422 | * Since this is common to all placement strategies, this lives here. | 1422 | * Since this is common to all placement strategies, this lives here. |
1423 | * | 1423 | * |
1424 | * [ this allows ->select_task() to simply return task_cpu(p) and | 1424 | * [ this allows ->select_task() to simply return task_cpu(p) and |
1425 | * not worry about this generic constraint ] | 1425 | * not worry about this generic constraint ] |
1426 | */ | 1426 | */ |
1427 | if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || | 1427 | if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || |
1428 | !cpu_online(cpu))) | 1428 | !cpu_online(cpu))) |
1429 | cpu = select_fallback_rq(task_cpu(p), p); | 1429 | cpu = select_fallback_rq(task_cpu(p), p); |
1430 | 1430 | ||
1431 | return cpu; | 1431 | return cpu; |
1432 | } | 1432 | } |
1433 | 1433 | ||
1434 | static void update_avg(u64 *avg, u64 sample) | 1434 | static void update_avg(u64 *avg, u64 sample) |
1435 | { | 1435 | { |
1436 | s64 diff = sample - *avg; | 1436 | s64 diff = sample - *avg; |
1437 | *avg += diff >> 3; | 1437 | *avg += diff >> 3; |
1438 | } | 1438 | } |
1439 | #endif | 1439 | #endif |
1440 | 1440 | ||
1441 | static void | 1441 | static void |
1442 | ttwu_stat(struct task_struct *p, int cpu, int wake_flags) | 1442 | ttwu_stat(struct task_struct *p, int cpu, int wake_flags) |
1443 | { | 1443 | { |
1444 | #ifdef CONFIG_SCHEDSTATS | 1444 | #ifdef CONFIG_SCHEDSTATS |
1445 | struct rq *rq = this_rq(); | 1445 | struct rq *rq = this_rq(); |
1446 | 1446 | ||
1447 | #ifdef CONFIG_SMP | 1447 | #ifdef CONFIG_SMP |
1448 | int this_cpu = smp_processor_id(); | 1448 | int this_cpu = smp_processor_id(); |
1449 | 1449 | ||
1450 | if (cpu == this_cpu) { | 1450 | if (cpu == this_cpu) { |
1451 | schedstat_inc(rq, ttwu_local); | 1451 | schedstat_inc(rq, ttwu_local); |
1452 | schedstat_inc(p, se.statistics.nr_wakeups_local); | 1452 | schedstat_inc(p, se.statistics.nr_wakeups_local); |
1453 | } else { | 1453 | } else { |
1454 | struct sched_domain *sd; | 1454 | struct sched_domain *sd; |
1455 | 1455 | ||
1456 | schedstat_inc(p, se.statistics.nr_wakeups_remote); | 1456 | schedstat_inc(p, se.statistics.nr_wakeups_remote); |
1457 | rcu_read_lock(); | 1457 | rcu_read_lock(); |
1458 | for_each_domain(this_cpu, sd) { | 1458 | for_each_domain(this_cpu, sd) { |
1459 | if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { | 1459 | if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { |
1460 | schedstat_inc(sd, ttwu_wake_remote); | 1460 | schedstat_inc(sd, ttwu_wake_remote); |
1461 | break; | 1461 | break; |
1462 | } | 1462 | } |
1463 | } | 1463 | } |
1464 | rcu_read_unlock(); | 1464 | rcu_read_unlock(); |
1465 | } | 1465 | } |
1466 | 1466 | ||
1467 | if (wake_flags & WF_MIGRATED) | 1467 | if (wake_flags & WF_MIGRATED) |
1468 | schedstat_inc(p, se.statistics.nr_wakeups_migrate); | 1468 | schedstat_inc(p, se.statistics.nr_wakeups_migrate); |
1469 | 1469 | ||
1470 | #endif /* CONFIG_SMP */ | 1470 | #endif /* CONFIG_SMP */ |
1471 | 1471 | ||
1472 | schedstat_inc(rq, ttwu_count); | 1472 | schedstat_inc(rq, ttwu_count); |
1473 | schedstat_inc(p, se.statistics.nr_wakeups); | 1473 | schedstat_inc(p, se.statistics.nr_wakeups); |
1474 | 1474 | ||
1475 | if (wake_flags & WF_SYNC) | 1475 | if (wake_flags & WF_SYNC) |
1476 | schedstat_inc(p, se.statistics.nr_wakeups_sync); | 1476 | schedstat_inc(p, se.statistics.nr_wakeups_sync); |
1477 | 1477 | ||
1478 | #endif /* CONFIG_SCHEDSTATS */ | 1478 | #endif /* CONFIG_SCHEDSTATS */ |
1479 | } | 1479 | } |
1480 | 1480 | ||
1481 | static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) | 1481 | static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) |
1482 | { | 1482 | { |
1483 | activate_task(rq, p, en_flags); | 1483 | activate_task(rq, p, en_flags); |
1484 | p->on_rq = TASK_ON_RQ_QUEUED; | 1484 | p->on_rq = TASK_ON_RQ_QUEUED; |
1485 | 1485 | ||
1486 | /* if a worker is waking up, notify workqueue */ | 1486 | /* if a worker is waking up, notify workqueue */ |
1487 | if (p->flags & PF_WQ_WORKER) | 1487 | if (p->flags & PF_WQ_WORKER) |
1488 | wq_worker_waking_up(p, cpu_of(rq)); | 1488 | wq_worker_waking_up(p, cpu_of(rq)); |
1489 | } | 1489 | } |
1490 | 1490 | ||
1491 | /* | 1491 | /* |
1492 | * Mark the task runnable and perform wakeup-preemption. | 1492 | * Mark the task runnable and perform wakeup-preemption. |
1493 | */ | 1493 | */ |
1494 | static void | 1494 | static void |
1495 | ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) | 1495 | ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
1496 | { | 1496 | { |
1497 | check_preempt_curr(rq, p, wake_flags); | 1497 | check_preempt_curr(rq, p, wake_flags); |
1498 | trace_sched_wakeup(p, true); | 1498 | trace_sched_wakeup(p, true); |
1499 | 1499 | ||
1500 | p->state = TASK_RUNNING; | 1500 | p->state = TASK_RUNNING; |
1501 | #ifdef CONFIG_SMP | 1501 | #ifdef CONFIG_SMP |
1502 | if (p->sched_class->task_woken) | 1502 | if (p->sched_class->task_woken) |
1503 | p->sched_class->task_woken(rq, p); | 1503 | p->sched_class->task_woken(rq, p); |
1504 | 1504 | ||
1505 | if (rq->idle_stamp) { | 1505 | if (rq->idle_stamp) { |
1506 | u64 delta = rq_clock(rq) - rq->idle_stamp; | 1506 | u64 delta = rq_clock(rq) - rq->idle_stamp; |
1507 | u64 max = 2*rq->max_idle_balance_cost; | 1507 | u64 max = 2*rq->max_idle_balance_cost; |
1508 | 1508 | ||
1509 | update_avg(&rq->avg_idle, delta); | 1509 | update_avg(&rq->avg_idle, delta); |
1510 | 1510 | ||
1511 | if (rq->avg_idle > max) | 1511 | if (rq->avg_idle > max) |
1512 | rq->avg_idle = max; | 1512 | rq->avg_idle = max; |
1513 | 1513 | ||
1514 | rq->idle_stamp = 0; | 1514 | rq->idle_stamp = 0; |
1515 | } | 1515 | } |
1516 | #endif | 1516 | #endif |
1517 | } | 1517 | } |
1518 | 1518 | ||
1519 | static void | 1519 | static void |
1520 | ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) | 1520 | ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) |
1521 | { | 1521 | { |
1522 | #ifdef CONFIG_SMP | 1522 | #ifdef CONFIG_SMP |
1523 | if (p->sched_contributes_to_load) | 1523 | if (p->sched_contributes_to_load) |
1524 | rq->nr_uninterruptible--; | 1524 | rq->nr_uninterruptible--; |
1525 | #endif | 1525 | #endif |
1526 | 1526 | ||
1527 | ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); | 1527 | ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); |
1528 | ttwu_do_wakeup(rq, p, wake_flags); | 1528 | ttwu_do_wakeup(rq, p, wake_flags); |
1529 | } | 1529 | } |
1530 | 1530 | ||
1531 | /* | 1531 | /* |
1532 | * Called in case the task @p isn't fully descheduled from its runqueue, | 1532 | * Called in case the task @p isn't fully descheduled from its runqueue, |
1533 | * in this case we must do a remote wakeup. Its a 'light' wakeup though, | 1533 | * in this case we must do a remote wakeup. Its a 'light' wakeup though, |
1534 | * since all we need to do is flip p->state to TASK_RUNNING, since | 1534 | * since all we need to do is flip p->state to TASK_RUNNING, since |
1535 | * the task is still ->on_rq. | 1535 | * the task is still ->on_rq. |
1536 | */ | 1536 | */ |
1537 | static int ttwu_remote(struct task_struct *p, int wake_flags) | 1537 | static int ttwu_remote(struct task_struct *p, int wake_flags) |
1538 | { | 1538 | { |
1539 | struct rq *rq; | 1539 | struct rq *rq; |
1540 | int ret = 0; | 1540 | int ret = 0; |
1541 | 1541 | ||
1542 | rq = __task_rq_lock(p); | 1542 | rq = __task_rq_lock(p); |
1543 | if (task_on_rq_queued(p)) { | 1543 | if (task_on_rq_queued(p)) { |
1544 | /* check_preempt_curr() may use rq clock */ | 1544 | /* check_preempt_curr() may use rq clock */ |
1545 | update_rq_clock(rq); | 1545 | update_rq_clock(rq); |
1546 | ttwu_do_wakeup(rq, p, wake_flags); | 1546 | ttwu_do_wakeup(rq, p, wake_flags); |
1547 | ret = 1; | 1547 | ret = 1; |
1548 | } | 1548 | } |
1549 | __task_rq_unlock(rq); | 1549 | __task_rq_unlock(rq); |
1550 | 1550 | ||
1551 | return ret; | 1551 | return ret; |
1552 | } | 1552 | } |
1553 | 1553 | ||
1554 | #ifdef CONFIG_SMP | 1554 | #ifdef CONFIG_SMP |
1555 | void sched_ttwu_pending(void) | 1555 | void sched_ttwu_pending(void) |
1556 | { | 1556 | { |
1557 | struct rq *rq = this_rq(); | 1557 | struct rq *rq = this_rq(); |
1558 | struct llist_node *llist = llist_del_all(&rq->wake_list); | 1558 | struct llist_node *llist = llist_del_all(&rq->wake_list); |
1559 | struct task_struct *p; | 1559 | struct task_struct *p; |
1560 | unsigned long flags; | 1560 | unsigned long flags; |
1561 | 1561 | ||
1562 | if (!llist) | 1562 | if (!llist) |
1563 | return; | 1563 | return; |
1564 | 1564 | ||
1565 | raw_spin_lock_irqsave(&rq->lock, flags); | 1565 | raw_spin_lock_irqsave(&rq->lock, flags); |
1566 | 1566 | ||
1567 | while (llist) { | 1567 | while (llist) { |
1568 | p = llist_entry(llist, struct task_struct, wake_entry); | 1568 | p = llist_entry(llist, struct task_struct, wake_entry); |
1569 | llist = llist_next(llist); | 1569 | llist = llist_next(llist); |
1570 | ttwu_do_activate(rq, p, 0); | 1570 | ttwu_do_activate(rq, p, 0); |
1571 | } | 1571 | } |
1572 | 1572 | ||
1573 | raw_spin_unlock_irqrestore(&rq->lock, flags); | 1573 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
1574 | } | 1574 | } |
1575 | 1575 | ||
1576 | void scheduler_ipi(void) | 1576 | void scheduler_ipi(void) |
1577 | { | 1577 | { |
1578 | /* | 1578 | /* |
1579 | * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting | 1579 | * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting |
1580 | * TIF_NEED_RESCHED remotely (for the first time) will also send | 1580 | * TIF_NEED_RESCHED remotely (for the first time) will also send |
1581 | * this IPI. | 1581 | * this IPI. |
1582 | */ | 1582 | */ |
1583 | preempt_fold_need_resched(); | 1583 | preempt_fold_need_resched(); |
1584 | 1584 | ||
1585 | if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) | 1585 | if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) |
1586 | return; | 1586 | return; |
1587 | 1587 | ||
1588 | /* | 1588 | /* |
1589 | * Not all reschedule IPI handlers call irq_enter/irq_exit, since | 1589 | * Not all reschedule IPI handlers call irq_enter/irq_exit, since |
1590 | * traditionally all their work was done from the interrupt return | 1590 | * traditionally all their work was done from the interrupt return |
1591 | * path. Now that we actually do some work, we need to make sure | 1591 | * path. Now that we actually do some work, we need to make sure |
1592 | * we do call them. | 1592 | * we do call them. |
1593 | * | 1593 | * |
1594 | * Some archs already do call them, luckily irq_enter/exit nest | 1594 | * Some archs already do call them, luckily irq_enter/exit nest |
1595 | * properly. | 1595 | * properly. |
1596 | * | 1596 | * |
1597 | * Arguably we should visit all archs and update all handlers, | 1597 | * Arguably we should visit all archs and update all handlers, |
1598 | * however a fair share of IPIs are still resched only so this would | 1598 | * however a fair share of IPIs are still resched only so this would |
1599 | * somewhat pessimize the simple resched case. | 1599 | * somewhat pessimize the simple resched case. |
1600 | */ | 1600 | */ |
1601 | irq_enter(); | 1601 | irq_enter(); |
1602 | sched_ttwu_pending(); | 1602 | sched_ttwu_pending(); |
1603 | 1603 | ||
1604 | /* | 1604 | /* |
1605 | * Check if someone kicked us for doing the nohz idle load balance. | 1605 | * Check if someone kicked us for doing the nohz idle load balance. |
1606 | */ | 1606 | */ |
1607 | if (unlikely(got_nohz_idle_kick())) { | 1607 | if (unlikely(got_nohz_idle_kick())) { |
1608 | this_rq()->idle_balance = 1; | 1608 | this_rq()->idle_balance = 1; |
1609 | raise_softirq_irqoff(SCHED_SOFTIRQ); | 1609 | raise_softirq_irqoff(SCHED_SOFTIRQ); |
1610 | } | 1610 | } |
1611 | irq_exit(); | 1611 | irq_exit(); |
1612 | } | 1612 | } |
1613 | 1613 | ||
1614 | static void ttwu_queue_remote(struct task_struct *p, int cpu) | 1614 | static void ttwu_queue_remote(struct task_struct *p, int cpu) |
1615 | { | 1615 | { |
1616 | struct rq *rq = cpu_rq(cpu); | 1616 | struct rq *rq = cpu_rq(cpu); |
1617 | 1617 | ||
1618 | if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) { | 1618 | if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) { |
1619 | if (!set_nr_if_polling(rq->idle)) | 1619 | if (!set_nr_if_polling(rq->idle)) |
1620 | smp_send_reschedule(cpu); | 1620 | smp_send_reschedule(cpu); |
1621 | else | 1621 | else |
1622 | trace_sched_wake_idle_without_ipi(cpu); | 1622 | trace_sched_wake_idle_without_ipi(cpu); |
1623 | } | 1623 | } |
1624 | } | 1624 | } |
1625 | 1625 | ||
1626 | void wake_up_if_idle(int cpu) | 1626 | void wake_up_if_idle(int cpu) |
1627 | { | 1627 | { |
1628 | struct rq *rq = cpu_rq(cpu); | 1628 | struct rq *rq = cpu_rq(cpu); |
1629 | unsigned long flags; | 1629 | unsigned long flags; |
1630 | 1630 | ||
1631 | rcu_read_lock(); | 1631 | rcu_read_lock(); |
1632 | 1632 | ||
1633 | if (!is_idle_task(rcu_dereference(rq->curr))) | 1633 | if (!is_idle_task(rcu_dereference(rq->curr))) |
1634 | goto out; | 1634 | goto out; |
1635 | 1635 | ||
1636 | if (set_nr_if_polling(rq->idle)) { | 1636 | if (set_nr_if_polling(rq->idle)) { |
1637 | trace_sched_wake_idle_without_ipi(cpu); | 1637 | trace_sched_wake_idle_without_ipi(cpu); |
1638 | } else { | 1638 | } else { |
1639 | raw_spin_lock_irqsave(&rq->lock, flags); | 1639 | raw_spin_lock_irqsave(&rq->lock, flags); |
1640 | if (is_idle_task(rq->curr)) | 1640 | if (is_idle_task(rq->curr)) |
1641 | smp_send_reschedule(cpu); | 1641 | smp_send_reschedule(cpu); |
1642 | /* Else cpu is not in idle, do nothing here */ | 1642 | /* Else cpu is not in idle, do nothing here */ |
1643 | raw_spin_unlock_irqrestore(&rq->lock, flags); | 1643 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
1644 | } | 1644 | } |
1645 | 1645 | ||
1646 | out: | 1646 | out: |
1647 | rcu_read_unlock(); | 1647 | rcu_read_unlock(); |
1648 | } | 1648 | } |
1649 | 1649 | ||
1650 | bool cpus_share_cache(int this_cpu, int that_cpu) | 1650 | bool cpus_share_cache(int this_cpu, int that_cpu) |
1651 | { | 1651 | { |
1652 | return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); | 1652 | return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); |
1653 | } | 1653 | } |
1654 | #endif /* CONFIG_SMP */ | 1654 | #endif /* CONFIG_SMP */ |
1655 | 1655 | ||
1656 | static void ttwu_queue(struct task_struct *p, int cpu) | 1656 | static void ttwu_queue(struct task_struct *p, int cpu) |
1657 | { | 1657 | { |
1658 | struct rq *rq = cpu_rq(cpu); | 1658 | struct rq *rq = cpu_rq(cpu); |
1659 | 1659 | ||
1660 | #if defined(CONFIG_SMP) | 1660 | #if defined(CONFIG_SMP) |
1661 | if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { | 1661 | if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { |
1662 | sched_clock_cpu(cpu); /* sync clocks x-cpu */ | 1662 | sched_clock_cpu(cpu); /* sync clocks x-cpu */ |
1663 | ttwu_queue_remote(p, cpu); | 1663 | ttwu_queue_remote(p, cpu); |
1664 | return; | 1664 | return; |
1665 | } | 1665 | } |
1666 | #endif | 1666 | #endif |
1667 | 1667 | ||
1668 | raw_spin_lock(&rq->lock); | 1668 | raw_spin_lock(&rq->lock); |
1669 | ttwu_do_activate(rq, p, 0); | 1669 | ttwu_do_activate(rq, p, 0); |
1670 | raw_spin_unlock(&rq->lock); | 1670 | raw_spin_unlock(&rq->lock); |
1671 | } | 1671 | } |
1672 | 1672 | ||
1673 | /** | 1673 | /** |
1674 | * try_to_wake_up - wake up a thread | 1674 | * try_to_wake_up - wake up a thread |
1675 | * @p: the thread to be awakened | 1675 | * @p: the thread to be awakened |
1676 | * @state: the mask of task states that can be woken | 1676 | * @state: the mask of task states that can be woken |
1677 | * @wake_flags: wake modifier flags (WF_*) | 1677 | * @wake_flags: wake modifier flags (WF_*) |
1678 | * | 1678 | * |
1679 | * Put it on the run-queue if it's not already there. The "current" | 1679 | * Put it on the run-queue if it's not already there. The "current" |
1680 | * thread is always on the run-queue (except when the actual | 1680 | * thread is always on the run-queue (except when the actual |
1681 | * re-schedule is in progress), and as such you're allowed to do | 1681 | * re-schedule is in progress), and as such you're allowed to do |
1682 | * the simpler "current->state = TASK_RUNNING" to mark yourself | 1682 | * the simpler "current->state = TASK_RUNNING" to mark yourself |
1683 | * runnable without the overhead of this. | 1683 | * runnable without the overhead of this. |
1684 | * | 1684 | * |
1685 | * Return: %true if @p was woken up, %false if it was already running. | 1685 | * Return: %true if @p was woken up, %false if it was already running. |
1686 | * or @state didn't match @p's state. | 1686 | * or @state didn't match @p's state. |
1687 | */ | 1687 | */ |
1688 | static int | 1688 | static int |
1689 | try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) | 1689 | try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) |
1690 | { | 1690 | { |
1691 | unsigned long flags; | 1691 | unsigned long flags; |
1692 | int cpu, success = 0; | 1692 | int cpu, success = 0; |
1693 | 1693 | ||
1694 | /* | 1694 | /* |
1695 | * If we are going to wake up a thread waiting for CONDITION we | 1695 | * If we are going to wake up a thread waiting for CONDITION we |
1696 | * need to ensure that CONDITION=1 done by the caller can not be | 1696 | * need to ensure that CONDITION=1 done by the caller can not be |
1697 | * reordered with p->state check below. This pairs with mb() in | 1697 | * reordered with p->state check below. This pairs with mb() in |
1698 | * set_current_state() the waiting thread does. | 1698 | * set_current_state() the waiting thread does. |
1699 | */ | 1699 | */ |
1700 | smp_mb__before_spinlock(); | 1700 | smp_mb__before_spinlock(); |
1701 | raw_spin_lock_irqsave(&p->pi_lock, flags); | 1701 | raw_spin_lock_irqsave(&p->pi_lock, flags); |
1702 | if (!(p->state & state)) | 1702 | if (!(p->state & state)) |
1703 | goto out; | 1703 | goto out; |
1704 | 1704 | ||
1705 | success = 1; /* we're going to change ->state */ | 1705 | success = 1; /* we're going to change ->state */ |
1706 | cpu = task_cpu(p); | 1706 | cpu = task_cpu(p); |
1707 | 1707 | ||
1708 | if (p->on_rq && ttwu_remote(p, wake_flags)) | 1708 | if (p->on_rq && ttwu_remote(p, wake_flags)) |
1709 | goto stat; | 1709 | goto stat; |
1710 | 1710 | ||
1711 | #ifdef CONFIG_SMP | 1711 | #ifdef CONFIG_SMP |
1712 | /* | 1712 | /* |
1713 | * If the owning (remote) cpu is still in the middle of schedule() with | 1713 | * If the owning (remote) cpu is still in the middle of schedule() with |
1714 | * this task as prev, wait until its done referencing the task. | 1714 | * this task as prev, wait until its done referencing the task. |
1715 | */ | 1715 | */ |
1716 | while (p->on_cpu) | 1716 | while (p->on_cpu) |
1717 | cpu_relax(); | 1717 | cpu_relax(); |
1718 | /* | 1718 | /* |
1719 | * Pairs with the smp_wmb() in finish_lock_switch(). | 1719 | * Pairs with the smp_wmb() in finish_lock_switch(). |
1720 | */ | 1720 | */ |
1721 | smp_rmb(); | 1721 | smp_rmb(); |
1722 | 1722 | ||
1723 | p->sched_contributes_to_load = !!task_contributes_to_load(p); | 1723 | p->sched_contributes_to_load = !!task_contributes_to_load(p); |
1724 | p->state = TASK_WAKING; | 1724 | p->state = TASK_WAKING; |
1725 | 1725 | ||
1726 | if (p->sched_class->task_waking) | 1726 | if (p->sched_class->task_waking) |
1727 | p->sched_class->task_waking(p); | 1727 | p->sched_class->task_waking(p); |
1728 | 1728 | ||
1729 | cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); | 1729 | cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); |
1730 | if (task_cpu(p) != cpu) { | 1730 | if (task_cpu(p) != cpu) { |
1731 | wake_flags |= WF_MIGRATED; | 1731 | wake_flags |= WF_MIGRATED; |
1732 | set_task_cpu(p, cpu); | 1732 | set_task_cpu(p, cpu); |
1733 | } | 1733 | } |
1734 | #endif /* CONFIG_SMP */ | 1734 | #endif /* CONFIG_SMP */ |
1735 | 1735 | ||
1736 | ttwu_queue(p, cpu); | 1736 | ttwu_queue(p, cpu); |
1737 | stat: | 1737 | stat: |
1738 | ttwu_stat(p, cpu, wake_flags); | 1738 | ttwu_stat(p, cpu, wake_flags); |
1739 | out: | 1739 | out: |
1740 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); | 1740 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
1741 | 1741 | ||
1742 | return success; | 1742 | return success; |
1743 | } | 1743 | } |
1744 | 1744 | ||
1745 | /** | 1745 | /** |
1746 | * try_to_wake_up_local - try to wake up a local task with rq lock held | 1746 | * try_to_wake_up_local - try to wake up a local task with rq lock held |
1747 | * @p: the thread to be awakened | 1747 | * @p: the thread to be awakened |
1748 | * | 1748 | * |
1749 | * Put @p on the run-queue if it's not already there. The caller must | 1749 | * Put @p on the run-queue if it's not already there. The caller must |
1750 | * ensure that this_rq() is locked, @p is bound to this_rq() and not | 1750 | * ensure that this_rq() is locked, @p is bound to this_rq() and not |
1751 | * the current task. | 1751 | * the current task. |
1752 | */ | 1752 | */ |
1753 | static void try_to_wake_up_local(struct task_struct *p) | 1753 | static void try_to_wake_up_local(struct task_struct *p) |
1754 | { | 1754 | { |
1755 | struct rq *rq = task_rq(p); | 1755 | struct rq *rq = task_rq(p); |
1756 | 1756 | ||
1757 | if (WARN_ON_ONCE(rq != this_rq()) || | 1757 | if (WARN_ON_ONCE(rq != this_rq()) || |
1758 | WARN_ON_ONCE(p == current)) | 1758 | WARN_ON_ONCE(p == current)) |
1759 | return; | 1759 | return; |
1760 | 1760 | ||
1761 | lockdep_assert_held(&rq->lock); | 1761 | lockdep_assert_held(&rq->lock); |
1762 | 1762 | ||
1763 | if (!raw_spin_trylock(&p->pi_lock)) { | 1763 | if (!raw_spin_trylock(&p->pi_lock)) { |
1764 | raw_spin_unlock(&rq->lock); | 1764 | raw_spin_unlock(&rq->lock); |
1765 | raw_spin_lock(&p->pi_lock); | 1765 | raw_spin_lock(&p->pi_lock); |
1766 | raw_spin_lock(&rq->lock); | 1766 | raw_spin_lock(&rq->lock); |
1767 | } | 1767 | } |
1768 | 1768 | ||
1769 | if (!(p->state & TASK_NORMAL)) | 1769 | if (!(p->state & TASK_NORMAL)) |
1770 | goto out; | 1770 | goto out; |
1771 | 1771 | ||
1772 | if (!task_on_rq_queued(p)) | 1772 | if (!task_on_rq_queued(p)) |
1773 | ttwu_activate(rq, p, ENQUEUE_WAKEUP); | 1773 | ttwu_activate(rq, p, ENQUEUE_WAKEUP); |
1774 | 1774 | ||
1775 | ttwu_do_wakeup(rq, p, 0); | 1775 | ttwu_do_wakeup(rq, p, 0); |
1776 | ttwu_stat(p, smp_processor_id(), 0); | 1776 | ttwu_stat(p, smp_processor_id(), 0); |
1777 | out: | 1777 | out: |
1778 | raw_spin_unlock(&p->pi_lock); | 1778 | raw_spin_unlock(&p->pi_lock); |
1779 | } | 1779 | } |
1780 | 1780 | ||
1781 | /** | 1781 | /** |
1782 | * wake_up_process - Wake up a specific process | 1782 | * wake_up_process - Wake up a specific process |
1783 | * @p: The process to be woken up. | 1783 | * @p: The process to be woken up. |
1784 | * | 1784 | * |
1785 | * Attempt to wake up the nominated process and move it to the set of runnable | 1785 | * Attempt to wake up the nominated process and move it to the set of runnable |
1786 | * processes. | 1786 | * processes. |
1787 | * | 1787 | * |
1788 | * Return: 1 if the process was woken up, 0 if it was already running. | 1788 | * Return: 1 if the process was woken up, 0 if it was already running. |
1789 | * | 1789 | * |
1790 | * It may be assumed that this function implies a write memory barrier before | 1790 | * It may be assumed that this function implies a write memory barrier before |
1791 | * changing the task state if and only if any tasks are woken up. | 1791 | * changing the task state if and only if any tasks are woken up. |
1792 | */ | 1792 | */ |
1793 | int wake_up_process(struct task_struct *p) | 1793 | int wake_up_process(struct task_struct *p) |
1794 | { | 1794 | { |
1795 | WARN_ON(task_is_stopped_or_traced(p)); | 1795 | WARN_ON(task_is_stopped_or_traced(p)); |
1796 | return try_to_wake_up(p, TASK_NORMAL, 0); | 1796 | return try_to_wake_up(p, TASK_NORMAL, 0); |
1797 | } | 1797 | } |
1798 | EXPORT_SYMBOL(wake_up_process); | 1798 | EXPORT_SYMBOL(wake_up_process); |
1799 | 1799 | ||
1800 | int wake_up_state(struct task_struct *p, unsigned int state) | 1800 | int wake_up_state(struct task_struct *p, unsigned int state) |
1801 | { | 1801 | { |
1802 | return try_to_wake_up(p, state, 0); | 1802 | return try_to_wake_up(p, state, 0); |
1803 | } | 1803 | } |
1804 | 1804 | ||
1805 | /* | 1805 | /* |
1806 | * This function clears the sched_dl_entity static params. | 1806 | * This function clears the sched_dl_entity static params. |
1807 | */ | 1807 | */ |
1808 | void __dl_clear_params(struct task_struct *p) | 1808 | void __dl_clear_params(struct task_struct *p) |
1809 | { | 1809 | { |
1810 | struct sched_dl_entity *dl_se = &p->dl; | 1810 | struct sched_dl_entity *dl_se = &p->dl; |
1811 | 1811 | ||
1812 | dl_se->dl_runtime = 0; | 1812 | dl_se->dl_runtime = 0; |
1813 | dl_se->dl_deadline = 0; | 1813 | dl_se->dl_deadline = 0; |
1814 | dl_se->dl_period = 0; | 1814 | dl_se->dl_period = 0; |
1815 | dl_se->flags = 0; | 1815 | dl_se->flags = 0; |
1816 | dl_se->dl_bw = 0; | 1816 | dl_se->dl_bw = 0; |
1817 | } | 1817 | } |
1818 | 1818 | ||
1819 | /* | 1819 | /* |
1820 | * Perform scheduler related setup for a newly forked process p. | 1820 | * Perform scheduler related setup for a newly forked process p. |
1821 | * p is forked by current. | 1821 | * p is forked by current. |
1822 | * | 1822 | * |
1823 | * __sched_fork() is basic setup used by init_idle() too: | 1823 | * __sched_fork() is basic setup used by init_idle() too: |
1824 | */ | 1824 | */ |
1825 | static void __sched_fork(unsigned long clone_flags, struct task_struct *p) | 1825 | static void __sched_fork(unsigned long clone_flags, struct task_struct *p) |
1826 | { | 1826 | { |
1827 | p->on_rq = 0; | 1827 | p->on_rq = 0; |
1828 | 1828 | ||
1829 | p->se.on_rq = 0; | 1829 | p->se.on_rq = 0; |
1830 | p->se.exec_start = 0; | 1830 | p->se.exec_start = 0; |
1831 | p->se.sum_exec_runtime = 0; | 1831 | p->se.sum_exec_runtime = 0; |
1832 | p->se.prev_sum_exec_runtime = 0; | 1832 | p->se.prev_sum_exec_runtime = 0; |
1833 | p->se.nr_migrations = 0; | 1833 | p->se.nr_migrations = 0; |
1834 | p->se.vruntime = 0; | 1834 | p->se.vruntime = 0; |
1835 | INIT_LIST_HEAD(&p->se.group_node); | 1835 | INIT_LIST_HEAD(&p->se.group_node); |
1836 | 1836 | ||
1837 | #ifdef CONFIG_SCHEDSTATS | 1837 | #ifdef CONFIG_SCHEDSTATS |
1838 | memset(&p->se.statistics, 0, sizeof(p->se.statistics)); | 1838 | memset(&p->se.statistics, 0, sizeof(p->se.statistics)); |
1839 | #endif | 1839 | #endif |
1840 | 1840 | ||
1841 | RB_CLEAR_NODE(&p->dl.rb_node); | 1841 | RB_CLEAR_NODE(&p->dl.rb_node); |
1842 | hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | 1842 | hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
1843 | __dl_clear_params(p); | 1843 | __dl_clear_params(p); |
1844 | 1844 | ||
1845 | INIT_LIST_HEAD(&p->rt.run_list); | 1845 | INIT_LIST_HEAD(&p->rt.run_list); |
1846 | 1846 | ||
1847 | #ifdef CONFIG_PREEMPT_NOTIFIERS | 1847 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
1848 | INIT_HLIST_HEAD(&p->preempt_notifiers); | 1848 | INIT_HLIST_HEAD(&p->preempt_notifiers); |
1849 | #endif | 1849 | #endif |
1850 | 1850 | ||
1851 | #ifdef CONFIG_NUMA_BALANCING | 1851 | #ifdef CONFIG_NUMA_BALANCING |
1852 | if (p->mm && atomic_read(&p->mm->mm_users) == 1) { | 1852 | if (p->mm && atomic_read(&p->mm->mm_users) == 1) { |
1853 | p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | 1853 | p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); |
1854 | p->mm->numa_scan_seq = 0; | 1854 | p->mm->numa_scan_seq = 0; |
1855 | } | 1855 | } |
1856 | 1856 | ||
1857 | if (clone_flags & CLONE_VM) | 1857 | if (clone_flags & CLONE_VM) |
1858 | p->numa_preferred_nid = current->numa_preferred_nid; | 1858 | p->numa_preferred_nid = current->numa_preferred_nid; |
1859 | else | 1859 | else |
1860 | p->numa_preferred_nid = -1; | 1860 | p->numa_preferred_nid = -1; |
1861 | 1861 | ||
1862 | p->node_stamp = 0ULL; | 1862 | p->node_stamp = 0ULL; |
1863 | p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; | 1863 | p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; |
1864 | p->numa_scan_period = sysctl_numa_balancing_scan_delay; | 1864 | p->numa_scan_period = sysctl_numa_balancing_scan_delay; |
1865 | p->numa_work.next = &p->numa_work; | 1865 | p->numa_work.next = &p->numa_work; |
1866 | p->numa_faults = NULL; | 1866 | p->numa_faults = NULL; |
1867 | p->last_task_numa_placement = 0; | 1867 | p->last_task_numa_placement = 0; |
1868 | p->last_sum_exec_runtime = 0; | 1868 | p->last_sum_exec_runtime = 0; |
1869 | 1869 | ||
1870 | p->numa_group = NULL; | 1870 | p->numa_group = NULL; |
1871 | #endif /* CONFIG_NUMA_BALANCING */ | 1871 | #endif /* CONFIG_NUMA_BALANCING */ |
1872 | } | 1872 | } |
1873 | 1873 | ||
1874 | #ifdef CONFIG_NUMA_BALANCING | 1874 | #ifdef CONFIG_NUMA_BALANCING |
1875 | #ifdef CONFIG_SCHED_DEBUG | 1875 | #ifdef CONFIG_SCHED_DEBUG |
1876 | void set_numabalancing_state(bool enabled) | 1876 | void set_numabalancing_state(bool enabled) |
1877 | { | 1877 | { |
1878 | if (enabled) | 1878 | if (enabled) |
1879 | sched_feat_set("NUMA"); | 1879 | sched_feat_set("NUMA"); |
1880 | else | 1880 | else |
1881 | sched_feat_set("NO_NUMA"); | 1881 | sched_feat_set("NO_NUMA"); |
1882 | } | 1882 | } |
1883 | #else | 1883 | #else |
1884 | __read_mostly bool numabalancing_enabled; | 1884 | __read_mostly bool numabalancing_enabled; |
1885 | 1885 | ||
1886 | void set_numabalancing_state(bool enabled) | 1886 | void set_numabalancing_state(bool enabled) |
1887 | { | 1887 | { |
1888 | numabalancing_enabled = enabled; | 1888 | numabalancing_enabled = enabled; |
1889 | } | 1889 | } |
1890 | #endif /* CONFIG_SCHED_DEBUG */ | 1890 | #endif /* CONFIG_SCHED_DEBUG */ |
1891 | 1891 | ||
1892 | #ifdef CONFIG_PROC_SYSCTL | 1892 | #ifdef CONFIG_PROC_SYSCTL |
1893 | int sysctl_numa_balancing(struct ctl_table *table, int write, | 1893 | int sysctl_numa_balancing(struct ctl_table *table, int write, |
1894 | void __user *buffer, size_t *lenp, loff_t *ppos) | 1894 | void __user *buffer, size_t *lenp, loff_t *ppos) |
1895 | { | 1895 | { |
1896 | struct ctl_table t; | 1896 | struct ctl_table t; |
1897 | int err; | 1897 | int err; |
1898 | int state = numabalancing_enabled; | 1898 | int state = numabalancing_enabled; |
1899 | 1899 | ||
1900 | if (write && !capable(CAP_SYS_ADMIN)) | 1900 | if (write && !capable(CAP_SYS_ADMIN)) |
1901 | return -EPERM; | 1901 | return -EPERM; |
1902 | 1902 | ||
1903 | t = *table; | 1903 | t = *table; |
1904 | t.data = &state; | 1904 | t.data = &state; |
1905 | err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); | 1905 | err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); |
1906 | if (err < 0) | 1906 | if (err < 0) |
1907 | return err; | 1907 | return err; |
1908 | if (write) | 1908 | if (write) |
1909 | set_numabalancing_state(state); | 1909 | set_numabalancing_state(state); |
1910 | return err; | 1910 | return err; |
1911 | } | 1911 | } |
1912 | #endif | 1912 | #endif |
1913 | #endif | 1913 | #endif |
1914 | 1914 | ||
1915 | /* | 1915 | /* |
1916 | * fork()/clone()-time setup: | 1916 | * fork()/clone()-time setup: |
1917 | */ | 1917 | */ |
1918 | int sched_fork(unsigned long clone_flags, struct task_struct *p) | 1918 | int sched_fork(unsigned long clone_flags, struct task_struct *p) |
1919 | { | 1919 | { |
1920 | unsigned long flags; | 1920 | unsigned long flags; |
1921 | int cpu = get_cpu(); | 1921 | int cpu = get_cpu(); |
1922 | 1922 | ||
1923 | __sched_fork(clone_flags, p); | 1923 | __sched_fork(clone_flags, p); |
1924 | /* | 1924 | /* |
1925 | * We mark the process as running here. This guarantees that | 1925 | * We mark the process as running here. This guarantees that |
1926 | * nobody will actually run it, and a signal or other external | 1926 | * nobody will actually run it, and a signal or other external |
1927 | * event cannot wake it up and insert it on the runqueue either. | 1927 | * event cannot wake it up and insert it on the runqueue either. |
1928 | */ | 1928 | */ |
1929 | p->state = TASK_RUNNING; | 1929 | p->state = TASK_RUNNING; |
1930 | 1930 | ||
1931 | /* | 1931 | /* |
1932 | * Make sure we do not leak PI boosting priority to the child. | 1932 | * Make sure we do not leak PI boosting priority to the child. |
1933 | */ | 1933 | */ |
1934 | p->prio = current->normal_prio; | 1934 | p->prio = current->normal_prio; |
1935 | 1935 | ||
1936 | /* | 1936 | /* |
1937 | * Revert to default priority/policy on fork if requested. | 1937 | * Revert to default priority/policy on fork if requested. |
1938 | */ | 1938 | */ |
1939 | if (unlikely(p->sched_reset_on_fork)) { | 1939 | if (unlikely(p->sched_reset_on_fork)) { |
1940 | if (task_has_dl_policy(p) || task_has_rt_policy(p)) { | 1940 | if (task_has_dl_policy(p) || task_has_rt_policy(p)) { |
1941 | p->policy = SCHED_NORMAL; | 1941 | p->policy = SCHED_NORMAL; |
1942 | p->static_prio = NICE_TO_PRIO(0); | 1942 | p->static_prio = NICE_TO_PRIO(0); |
1943 | p->rt_priority = 0; | 1943 | p->rt_priority = 0; |
1944 | } else if (PRIO_TO_NICE(p->static_prio) < 0) | 1944 | } else if (PRIO_TO_NICE(p->static_prio) < 0) |
1945 | p->static_prio = NICE_TO_PRIO(0); | 1945 | p->static_prio = NICE_TO_PRIO(0); |
1946 | 1946 | ||
1947 | p->prio = p->normal_prio = __normal_prio(p); | 1947 | p->prio = p->normal_prio = __normal_prio(p); |
1948 | set_load_weight(p); | 1948 | set_load_weight(p); |
1949 | 1949 | ||
1950 | /* | 1950 | /* |
1951 | * We don't need the reset flag anymore after the fork. It has | 1951 | * We don't need the reset flag anymore after the fork. It has |
1952 | * fulfilled its duty: | 1952 | * fulfilled its duty: |
1953 | */ | 1953 | */ |
1954 | p->sched_reset_on_fork = 0; | 1954 | p->sched_reset_on_fork = 0; |
1955 | } | 1955 | } |
1956 | 1956 | ||
1957 | if (dl_prio(p->prio)) { | 1957 | if (dl_prio(p->prio)) { |
1958 | put_cpu(); | 1958 | put_cpu(); |
1959 | return -EAGAIN; | 1959 | return -EAGAIN; |
1960 | } else if (rt_prio(p->prio)) { | 1960 | } else if (rt_prio(p->prio)) { |
1961 | p->sched_class = &rt_sched_class; | 1961 | p->sched_class = &rt_sched_class; |
1962 | } else { | 1962 | } else { |
1963 | p->sched_class = &fair_sched_class; | 1963 | p->sched_class = &fair_sched_class; |
1964 | } | 1964 | } |
1965 | 1965 | ||
1966 | if (p->sched_class->task_fork) | 1966 | if (p->sched_class->task_fork) |
1967 | p->sched_class->task_fork(p); | 1967 | p->sched_class->task_fork(p); |
1968 | 1968 | ||
1969 | /* | 1969 | /* |
1970 | * The child is not yet in the pid-hash so no cgroup attach races, | 1970 | * The child is not yet in the pid-hash so no cgroup attach races, |
1971 | * and the cgroup is pinned to this child due to cgroup_fork() | 1971 | * and the cgroup is pinned to this child due to cgroup_fork() |
1972 | * is ran before sched_fork(). | 1972 | * is ran before sched_fork(). |
1973 | * | 1973 | * |
1974 | * Silence PROVE_RCU. | 1974 | * Silence PROVE_RCU. |
1975 | */ | 1975 | */ |
1976 | raw_spin_lock_irqsave(&p->pi_lock, flags); | 1976 | raw_spin_lock_irqsave(&p->pi_lock, flags); |
1977 | set_task_cpu(p, cpu); | 1977 | set_task_cpu(p, cpu); |
1978 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); | 1978 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
1979 | 1979 | ||
1980 | #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) | 1980 | #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) |
1981 | if (likely(sched_info_on())) | 1981 | if (likely(sched_info_on())) |
1982 | memset(&p->sched_info, 0, sizeof(p->sched_info)); | 1982 | memset(&p->sched_info, 0, sizeof(p->sched_info)); |
1983 | #endif | 1983 | #endif |
1984 | #if defined(CONFIG_SMP) | 1984 | #if defined(CONFIG_SMP) |
1985 | p->on_cpu = 0; | 1985 | p->on_cpu = 0; |
1986 | #endif | 1986 | #endif |
1987 | init_task_preempt_count(p); | 1987 | init_task_preempt_count(p); |
1988 | #ifdef CONFIG_SMP | 1988 | #ifdef CONFIG_SMP |
1989 | plist_node_init(&p->pushable_tasks, MAX_PRIO); | 1989 | plist_node_init(&p->pushable_tasks, MAX_PRIO); |
1990 | RB_CLEAR_NODE(&p->pushable_dl_tasks); | 1990 | RB_CLEAR_NODE(&p->pushable_dl_tasks); |
1991 | #endif | 1991 | #endif |
1992 | 1992 | ||
1993 | put_cpu(); | 1993 | put_cpu(); |
1994 | return 0; | 1994 | return 0; |
1995 | } | 1995 | } |
1996 | 1996 | ||
1997 | unsigned long to_ratio(u64 period, u64 runtime) | 1997 | unsigned long to_ratio(u64 period, u64 runtime) |
1998 | { | 1998 | { |
1999 | if (runtime == RUNTIME_INF) | 1999 | if (runtime == RUNTIME_INF) |
2000 | return 1ULL << 20; | 2000 | return 1ULL << 20; |
2001 | 2001 | ||
2002 | /* | 2002 | /* |
2003 | * Doing this here saves a lot of checks in all | 2003 | * Doing this here saves a lot of checks in all |
2004 | * the calling paths, and returning zero seems | 2004 | * the calling paths, and returning zero seems |
2005 | * safe for them anyway. | 2005 | * safe for them anyway. |
2006 | */ | 2006 | */ |
2007 | if (period == 0) | 2007 | if (period == 0) |
2008 | return 0; | 2008 | return 0; |
2009 | 2009 | ||
2010 | return div64_u64(runtime << 20, period); | 2010 | return div64_u64(runtime << 20, period); |
2011 | } | 2011 | } |
2012 | 2012 | ||
2013 | #ifdef CONFIG_SMP | 2013 | #ifdef CONFIG_SMP |
2014 | inline struct dl_bw *dl_bw_of(int i) | 2014 | inline struct dl_bw *dl_bw_of(int i) |
2015 | { | 2015 | { |
2016 | rcu_lockdep_assert(rcu_read_lock_sched_held(), | 2016 | rcu_lockdep_assert(rcu_read_lock_sched_held(), |
2017 | "sched RCU must be held"); | 2017 | "sched RCU must be held"); |
2018 | return &cpu_rq(i)->rd->dl_bw; | 2018 | return &cpu_rq(i)->rd->dl_bw; |
2019 | } | 2019 | } |
2020 | 2020 | ||
2021 | static inline int dl_bw_cpus(int i) | 2021 | static inline int dl_bw_cpus(int i) |
2022 | { | 2022 | { |
2023 | struct root_domain *rd = cpu_rq(i)->rd; | 2023 | struct root_domain *rd = cpu_rq(i)->rd; |
2024 | int cpus = 0; | 2024 | int cpus = 0; |
2025 | 2025 | ||
2026 | rcu_lockdep_assert(rcu_read_lock_sched_held(), | 2026 | rcu_lockdep_assert(rcu_read_lock_sched_held(), |
2027 | "sched RCU must be held"); | 2027 | "sched RCU must be held"); |
2028 | for_each_cpu_and(i, rd->span, cpu_active_mask) | 2028 | for_each_cpu_and(i, rd->span, cpu_active_mask) |
2029 | cpus++; | 2029 | cpus++; |
2030 | 2030 | ||
2031 | return cpus; | 2031 | return cpus; |
2032 | } | 2032 | } |
2033 | #else | 2033 | #else |
2034 | inline struct dl_bw *dl_bw_of(int i) | 2034 | inline struct dl_bw *dl_bw_of(int i) |
2035 | { | 2035 | { |
2036 | return &cpu_rq(i)->dl.dl_bw; | 2036 | return &cpu_rq(i)->dl.dl_bw; |
2037 | } | 2037 | } |
2038 | 2038 | ||
2039 | static inline int dl_bw_cpus(int i) | 2039 | static inline int dl_bw_cpus(int i) |
2040 | { | 2040 | { |
2041 | return 1; | 2041 | return 1; |
2042 | } | 2042 | } |
2043 | #endif | 2043 | #endif |
2044 | 2044 | ||
2045 | /* | 2045 | /* |
2046 | * We must be sure that accepting a new task (or allowing changing the | 2046 | * We must be sure that accepting a new task (or allowing changing the |
2047 | * parameters of an existing one) is consistent with the bandwidth | 2047 | * parameters of an existing one) is consistent with the bandwidth |
2048 | * constraints. If yes, this function also accordingly updates the currently | 2048 | * constraints. If yes, this function also accordingly updates the currently |
2049 | * allocated bandwidth to reflect the new situation. | 2049 | * allocated bandwidth to reflect the new situation. |
2050 | * | 2050 | * |
2051 | * This function is called while holding p's rq->lock. | 2051 | * This function is called while holding p's rq->lock. |
2052 | */ | 2052 | */ |
2053 | static int dl_overflow(struct task_struct *p, int policy, | 2053 | static int dl_overflow(struct task_struct *p, int policy, |
2054 | const struct sched_attr *attr) | 2054 | const struct sched_attr *attr) |
2055 | { | 2055 | { |
2056 | 2056 | ||
2057 | struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); | 2057 | struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); |
2058 | u64 period = attr->sched_period ?: attr->sched_deadline; | 2058 | u64 period = attr->sched_period ?: attr->sched_deadline; |
2059 | u64 runtime = attr->sched_runtime; | 2059 | u64 runtime = attr->sched_runtime; |
2060 | u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; | 2060 | u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; |
2061 | int cpus, err = -1; | 2061 | int cpus, err = -1; |
2062 | 2062 | ||
2063 | if (new_bw == p->dl.dl_bw) | 2063 | if (new_bw == p->dl.dl_bw) |
2064 | return 0; | 2064 | return 0; |
2065 | 2065 | ||
2066 | /* | 2066 | /* |
2067 | * Either if a task, enters, leave, or stays -deadline but changes | 2067 | * Either if a task, enters, leave, or stays -deadline but changes |
2068 | * its parameters, we may need to update accordingly the total | 2068 | * its parameters, we may need to update accordingly the total |
2069 | * allocated bandwidth of the container. | 2069 | * allocated bandwidth of the container. |
2070 | */ | 2070 | */ |
2071 | raw_spin_lock(&dl_b->lock); | 2071 | raw_spin_lock(&dl_b->lock); |
2072 | cpus = dl_bw_cpus(task_cpu(p)); | 2072 | cpus = dl_bw_cpus(task_cpu(p)); |
2073 | if (dl_policy(policy) && !task_has_dl_policy(p) && | 2073 | if (dl_policy(policy) && !task_has_dl_policy(p) && |
2074 | !__dl_overflow(dl_b, cpus, 0, new_bw)) { | 2074 | !__dl_overflow(dl_b, cpus, 0, new_bw)) { |
2075 | __dl_add(dl_b, new_bw); | 2075 | __dl_add(dl_b, new_bw); |
2076 | err = 0; | 2076 | err = 0; |
2077 | } else if (dl_policy(policy) && task_has_dl_policy(p) && | 2077 | } else if (dl_policy(policy) && task_has_dl_policy(p) && |
2078 | !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { | 2078 | !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { |
2079 | __dl_clear(dl_b, p->dl.dl_bw); | 2079 | __dl_clear(dl_b, p->dl.dl_bw); |
2080 | __dl_add(dl_b, new_bw); | 2080 | __dl_add(dl_b, new_bw); |
2081 | err = 0; | 2081 | err = 0; |
2082 | } else if (!dl_policy(policy) && task_has_dl_policy(p)) { | 2082 | } else if (!dl_policy(policy) && task_has_dl_policy(p)) { |
2083 | __dl_clear(dl_b, p->dl.dl_bw); | 2083 | __dl_clear(dl_b, p->dl.dl_bw); |
2084 | err = 0; | 2084 | err = 0; |
2085 | } | 2085 | } |
2086 | raw_spin_unlock(&dl_b->lock); | 2086 | raw_spin_unlock(&dl_b->lock); |
2087 | 2087 | ||
2088 | return err; | 2088 | return err; |
2089 | } | 2089 | } |
2090 | 2090 | ||
2091 | extern void init_dl_bw(struct dl_bw *dl_b); | 2091 | extern void init_dl_bw(struct dl_bw *dl_b); |
2092 | 2092 | ||
2093 | /* | 2093 | /* |
2094 | * wake_up_new_task - wake up a newly created task for the first time. | 2094 | * wake_up_new_task - wake up a newly created task for the first time. |
2095 | * | 2095 | * |
2096 | * This function will do some initial scheduler statistics housekeeping | 2096 | * This function will do some initial scheduler statistics housekeeping |
2097 | * that must be done for every newly created context, then puts the task | 2097 | * that must be done for every newly created context, then puts the task |
2098 | * on the runqueue and wakes it. | 2098 | * on the runqueue and wakes it. |
2099 | */ | 2099 | */ |
2100 | void wake_up_new_task(struct task_struct *p) | 2100 | void wake_up_new_task(struct task_struct *p) |
2101 | { | 2101 | { |
2102 | unsigned long flags; | 2102 | unsigned long flags; |
2103 | struct rq *rq; | 2103 | struct rq *rq; |
2104 | 2104 | ||
2105 | raw_spin_lock_irqsave(&p->pi_lock, flags); | 2105 | raw_spin_lock_irqsave(&p->pi_lock, flags); |
2106 | #ifdef CONFIG_SMP | 2106 | #ifdef CONFIG_SMP |
2107 | /* | 2107 | /* |
2108 | * Fork balancing, do it here and not earlier because: | 2108 | * Fork balancing, do it here and not earlier because: |
2109 | * - cpus_allowed can change in the fork path | 2109 | * - cpus_allowed can change in the fork path |
2110 | * - any previously selected cpu might disappear through hotplug | 2110 | * - any previously selected cpu might disappear through hotplug |
2111 | */ | 2111 | */ |
2112 | set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); | 2112 | set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); |
2113 | #endif | 2113 | #endif |
2114 | 2114 | ||
2115 | /* Initialize new task's runnable average */ | 2115 | /* Initialize new task's runnable average */ |
2116 | init_task_runnable_average(p); | 2116 | init_task_runnable_average(p); |
2117 | rq = __task_rq_lock(p); | 2117 | rq = __task_rq_lock(p); |
2118 | activate_task(rq, p, 0); | 2118 | activate_task(rq, p, 0); |
2119 | p->on_rq = TASK_ON_RQ_QUEUED; | 2119 | p->on_rq = TASK_ON_RQ_QUEUED; |
2120 | trace_sched_wakeup_new(p, true); | 2120 | trace_sched_wakeup_new(p, true); |
2121 | check_preempt_curr(rq, p, WF_FORK); | 2121 | check_preempt_curr(rq, p, WF_FORK); |
2122 | #ifdef CONFIG_SMP | 2122 | #ifdef CONFIG_SMP |
2123 | if (p->sched_class->task_woken) | 2123 | if (p->sched_class->task_woken) |
2124 | p->sched_class->task_woken(rq, p); | 2124 | p->sched_class->task_woken(rq, p); |
2125 | #endif | 2125 | #endif |
2126 | task_rq_unlock(rq, p, &flags); | 2126 | task_rq_unlock(rq, p, &flags); |
2127 | } | 2127 | } |
2128 | 2128 | ||
2129 | #ifdef CONFIG_PREEMPT_NOTIFIERS | 2129 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
2130 | 2130 | ||
2131 | /** | 2131 | /** |
2132 | * preempt_notifier_register - tell me when current is being preempted & rescheduled | 2132 | * preempt_notifier_register - tell me when current is being preempted & rescheduled |
2133 | * @notifier: notifier struct to register | 2133 | * @notifier: notifier struct to register |
2134 | */ | 2134 | */ |
2135 | void preempt_notifier_register(struct preempt_notifier *notifier) | 2135 | void preempt_notifier_register(struct preempt_notifier *notifier) |
2136 | { | 2136 | { |
2137 | hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); | 2137 | hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); |
2138 | } | 2138 | } |
2139 | EXPORT_SYMBOL_GPL(preempt_notifier_register); | 2139 | EXPORT_SYMBOL_GPL(preempt_notifier_register); |
2140 | 2140 | ||
2141 | /** | 2141 | /** |
2142 | * preempt_notifier_unregister - no longer interested in preemption notifications | 2142 | * preempt_notifier_unregister - no longer interested in preemption notifications |
2143 | * @notifier: notifier struct to unregister | 2143 | * @notifier: notifier struct to unregister |
2144 | * | 2144 | * |
2145 | * This is safe to call from within a preemption notifier. | 2145 | * This is safe to call from within a preemption notifier. |
2146 | */ | 2146 | */ |
2147 | void preempt_notifier_unregister(struct preempt_notifier *notifier) | 2147 | void preempt_notifier_unregister(struct preempt_notifier *notifier) |
2148 | { | 2148 | { |
2149 | hlist_del(¬ifier->link); | 2149 | hlist_del(¬ifier->link); |
2150 | } | 2150 | } |
2151 | EXPORT_SYMBOL_GPL(preempt_notifier_unregister); | 2151 | EXPORT_SYMBOL_GPL(preempt_notifier_unregister); |
2152 | 2152 | ||
2153 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) | 2153 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) |
2154 | { | 2154 | { |
2155 | struct preempt_notifier *notifier; | 2155 | struct preempt_notifier *notifier; |
2156 | 2156 | ||
2157 | hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) | 2157 | hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) |
2158 | notifier->ops->sched_in(notifier, raw_smp_processor_id()); | 2158 | notifier->ops->sched_in(notifier, raw_smp_processor_id()); |
2159 | } | 2159 | } |
2160 | 2160 | ||
2161 | static void | 2161 | static void |
2162 | fire_sched_out_preempt_notifiers(struct task_struct *curr, | 2162 | fire_sched_out_preempt_notifiers(struct task_struct *curr, |
2163 | struct task_struct *next) | 2163 | struct task_struct *next) |
2164 | { | 2164 | { |
2165 | struct preempt_notifier *notifier; | 2165 | struct preempt_notifier *notifier; |
2166 | 2166 | ||
2167 | hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) | 2167 | hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) |
2168 | notifier->ops->sched_out(notifier, next); | 2168 | notifier->ops->sched_out(notifier, next); |
2169 | } | 2169 | } |
2170 | 2170 | ||
2171 | #else /* !CONFIG_PREEMPT_NOTIFIERS */ | 2171 | #else /* !CONFIG_PREEMPT_NOTIFIERS */ |
2172 | 2172 | ||
2173 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) | 2173 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) |
2174 | { | 2174 | { |
2175 | } | 2175 | } |
2176 | 2176 | ||
2177 | static void | 2177 | static void |
2178 | fire_sched_out_preempt_notifiers(struct task_struct *curr, | 2178 | fire_sched_out_preempt_notifiers(struct task_struct *curr, |
2179 | struct task_struct *next) | 2179 | struct task_struct *next) |
2180 | { | 2180 | { |
2181 | } | 2181 | } |
2182 | 2182 | ||
2183 | #endif /* CONFIG_PREEMPT_NOTIFIERS */ | 2183 | #endif /* CONFIG_PREEMPT_NOTIFIERS */ |
2184 | 2184 | ||
2185 | /** | 2185 | /** |
2186 | * prepare_task_switch - prepare to switch tasks | 2186 | * prepare_task_switch - prepare to switch tasks |
2187 | * @rq: the runqueue preparing to switch | 2187 | * @rq: the runqueue preparing to switch |
2188 | * @prev: the current task that is being switched out | 2188 | * @prev: the current task that is being switched out |
2189 | * @next: the task we are going to switch to. | 2189 | * @next: the task we are going to switch to. |
2190 | * | 2190 | * |
2191 | * This is called with the rq lock held and interrupts off. It must | 2191 | * This is called with the rq lock held and interrupts off. It must |
2192 | * be paired with a subsequent finish_task_switch after the context | 2192 | * be paired with a subsequent finish_task_switch after the context |
2193 | * switch. | 2193 | * switch. |
2194 | * | 2194 | * |
2195 | * prepare_task_switch sets up locking and calls architecture specific | 2195 | * prepare_task_switch sets up locking and calls architecture specific |
2196 | * hooks. | 2196 | * hooks. |
2197 | */ | 2197 | */ |
2198 | static inline void | 2198 | static inline void |
2199 | prepare_task_switch(struct rq *rq, struct task_struct *prev, | 2199 | prepare_task_switch(struct rq *rq, struct task_struct *prev, |
2200 | struct task_struct *next) | 2200 | struct task_struct *next) |
2201 | { | 2201 | { |
2202 | trace_sched_switch(prev, next); | 2202 | trace_sched_switch(prev, next); |
2203 | sched_info_switch(rq, prev, next); | 2203 | sched_info_switch(rq, prev, next); |
2204 | perf_event_task_sched_out(prev, next); | 2204 | perf_event_task_sched_out(prev, next); |
2205 | fire_sched_out_preempt_notifiers(prev, next); | 2205 | fire_sched_out_preempt_notifiers(prev, next); |
2206 | prepare_lock_switch(rq, next); | 2206 | prepare_lock_switch(rq, next); |
2207 | prepare_arch_switch(next); | 2207 | prepare_arch_switch(next); |
2208 | } | 2208 | } |
2209 | 2209 | ||
2210 | /** | 2210 | /** |
2211 | * finish_task_switch - clean up after a task-switch | 2211 | * finish_task_switch - clean up after a task-switch |
2212 | * @prev: the thread we just switched away from. | 2212 | * @prev: the thread we just switched away from. |
2213 | * | 2213 | * |
2214 | * finish_task_switch must be called after the context switch, paired | 2214 | * finish_task_switch must be called after the context switch, paired |
2215 | * with a prepare_task_switch call before the context switch. | 2215 | * with a prepare_task_switch call before the context switch. |
2216 | * finish_task_switch will reconcile locking set up by prepare_task_switch, | 2216 | * finish_task_switch will reconcile locking set up by prepare_task_switch, |
2217 | * and do any other architecture-specific cleanup actions. | 2217 | * and do any other architecture-specific cleanup actions. |
2218 | * | 2218 | * |
2219 | * Note that we may have delayed dropping an mm in context_switch(). If | 2219 | * Note that we may have delayed dropping an mm in context_switch(). If |
2220 | * so, we finish that here outside of the runqueue lock. (Doing it | 2220 | * so, we finish that here outside of the runqueue lock. (Doing it |
2221 | * with the lock held can cause deadlocks; see schedule() for | 2221 | * with the lock held can cause deadlocks; see schedule() for |
2222 | * details.) | 2222 | * details.) |
2223 | * | 2223 | * |
2224 | * The context switch have flipped the stack from under us and restored the | 2224 | * The context switch have flipped the stack from under us and restored the |
2225 | * local variables which were saved when this task called schedule() in the | 2225 | * local variables which were saved when this task called schedule() in the |
2226 | * past. prev == current is still correct but we need to recalculate this_rq | 2226 | * past. prev == current is still correct but we need to recalculate this_rq |
2227 | * because prev may have moved to another CPU. | 2227 | * because prev may have moved to another CPU. |
2228 | */ | 2228 | */ |
2229 | static struct rq *finish_task_switch(struct task_struct *prev) | 2229 | static struct rq *finish_task_switch(struct task_struct *prev) |
2230 | __releases(rq->lock) | 2230 | __releases(rq->lock) |
2231 | { | 2231 | { |
2232 | struct rq *rq = this_rq(); | 2232 | struct rq *rq = this_rq(); |
2233 | struct mm_struct *mm = rq->prev_mm; | 2233 | struct mm_struct *mm = rq->prev_mm; |
2234 | long prev_state; | 2234 | long prev_state; |
2235 | 2235 | ||
2236 | rq->prev_mm = NULL; | 2236 | rq->prev_mm = NULL; |
2237 | 2237 | ||
2238 | /* | 2238 | /* |
2239 | * A task struct has one reference for the use as "current". | 2239 | * A task struct has one reference for the use as "current". |
2240 | * If a task dies, then it sets TASK_DEAD in tsk->state and calls | 2240 | * If a task dies, then it sets TASK_DEAD in tsk->state and calls |
2241 | * schedule one last time. The schedule call will never return, and | 2241 | * schedule one last time. The schedule call will never return, and |
2242 | * the scheduled task must drop that reference. | 2242 | * the scheduled task must drop that reference. |
2243 | * The test for TASK_DEAD must occur while the runqueue locks are | 2243 | * The test for TASK_DEAD must occur while the runqueue locks are |
2244 | * still held, otherwise prev could be scheduled on another cpu, die | 2244 | * still held, otherwise prev could be scheduled on another cpu, die |
2245 | * there before we look at prev->state, and then the reference would | 2245 | * there before we look at prev->state, and then the reference would |
2246 | * be dropped twice. | 2246 | * be dropped twice. |
2247 | * Manfred Spraul <manfred@colorfullife.com> | 2247 | * Manfred Spraul <manfred@colorfullife.com> |
2248 | */ | 2248 | */ |
2249 | prev_state = prev->state; | 2249 | prev_state = prev->state; |
2250 | vtime_task_switch(prev); | 2250 | vtime_task_switch(prev); |
2251 | finish_arch_switch(prev); | 2251 | finish_arch_switch(prev); |
2252 | perf_event_task_sched_in(prev, current); | 2252 | perf_event_task_sched_in(prev, current); |
2253 | finish_lock_switch(rq, prev); | 2253 | finish_lock_switch(rq, prev); |
2254 | finish_arch_post_lock_switch(); | 2254 | finish_arch_post_lock_switch(); |
2255 | 2255 | ||
2256 | fire_sched_in_preempt_notifiers(current); | 2256 | fire_sched_in_preempt_notifiers(current); |
2257 | if (mm) | 2257 | if (mm) |
2258 | mmdrop(mm); | 2258 | mmdrop(mm); |
2259 | if (unlikely(prev_state == TASK_DEAD)) { | 2259 | if (unlikely(prev_state == TASK_DEAD)) { |
2260 | if (prev->sched_class->task_dead) | 2260 | if (prev->sched_class->task_dead) |
2261 | prev->sched_class->task_dead(prev); | 2261 | prev->sched_class->task_dead(prev); |
2262 | 2262 | ||
2263 | /* | 2263 | /* |
2264 | * Remove function-return probe instances associated with this | 2264 | * Remove function-return probe instances associated with this |
2265 | * task and put them back on the free list. | 2265 | * task and put them back on the free list. |
2266 | */ | 2266 | */ |
2267 | kprobe_flush_task(prev); | 2267 | kprobe_flush_task(prev); |
2268 | put_task_struct(prev); | 2268 | put_task_struct(prev); |
2269 | } | 2269 | } |
2270 | 2270 | ||
2271 | tick_nohz_task_switch(current); | 2271 | tick_nohz_task_switch(current); |
2272 | return rq; | 2272 | return rq; |
2273 | } | 2273 | } |
2274 | 2274 | ||
2275 | #ifdef CONFIG_SMP | 2275 | #ifdef CONFIG_SMP |
2276 | 2276 | ||
2277 | /* rq->lock is NOT held, but preemption is disabled */ | 2277 | /* rq->lock is NOT held, but preemption is disabled */ |
2278 | static inline void post_schedule(struct rq *rq) | 2278 | static inline void post_schedule(struct rq *rq) |
2279 | { | 2279 | { |
2280 | if (rq->post_schedule) { | 2280 | if (rq->post_schedule) { |
2281 | unsigned long flags; | 2281 | unsigned long flags; |
2282 | 2282 | ||
2283 | raw_spin_lock_irqsave(&rq->lock, flags); | 2283 | raw_spin_lock_irqsave(&rq->lock, flags); |
2284 | if (rq->curr->sched_class->post_schedule) | 2284 | if (rq->curr->sched_class->post_schedule) |
2285 | rq->curr->sched_class->post_schedule(rq); | 2285 | rq->curr->sched_class->post_schedule(rq); |
2286 | raw_spin_unlock_irqrestore(&rq->lock, flags); | 2286 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
2287 | 2287 | ||
2288 | rq->post_schedule = 0; | 2288 | rq->post_schedule = 0; |
2289 | } | 2289 | } |
2290 | } | 2290 | } |
2291 | 2291 | ||
2292 | #else | 2292 | #else |
2293 | 2293 | ||
2294 | static inline void post_schedule(struct rq *rq) | 2294 | static inline void post_schedule(struct rq *rq) |
2295 | { | 2295 | { |
2296 | } | 2296 | } |
2297 | 2297 | ||
2298 | #endif | 2298 | #endif |
2299 | 2299 | ||
2300 | /** | 2300 | /** |
2301 | * schedule_tail - first thing a freshly forked thread must call. | 2301 | * schedule_tail - first thing a freshly forked thread must call. |
2302 | * @prev: the thread we just switched away from. | 2302 | * @prev: the thread we just switched away from. |
2303 | */ | 2303 | */ |
2304 | asmlinkage __visible void schedule_tail(struct task_struct *prev) | 2304 | asmlinkage __visible void schedule_tail(struct task_struct *prev) |
2305 | __releases(rq->lock) | 2305 | __releases(rq->lock) |
2306 | { | 2306 | { |
2307 | struct rq *rq; | 2307 | struct rq *rq; |
2308 | 2308 | ||
2309 | /* finish_task_switch() drops rq->lock and enables preemtion */ | 2309 | /* finish_task_switch() drops rq->lock and enables preemtion */ |
2310 | preempt_disable(); | 2310 | preempt_disable(); |
2311 | rq = finish_task_switch(prev); | 2311 | rq = finish_task_switch(prev); |
2312 | post_schedule(rq); | 2312 | post_schedule(rq); |
2313 | preempt_enable(); | 2313 | preempt_enable(); |
2314 | 2314 | ||
2315 | if (current->set_child_tid) | 2315 | if (current->set_child_tid) |
2316 | put_user(task_pid_vnr(current), current->set_child_tid); | 2316 | put_user(task_pid_vnr(current), current->set_child_tid); |
2317 | } | 2317 | } |
2318 | 2318 | ||
2319 | /* | 2319 | /* |
2320 | * context_switch - switch to the new MM and the new thread's register state. | 2320 | * context_switch - switch to the new MM and the new thread's register state. |
2321 | */ | 2321 | */ |
2322 | static inline struct rq * | 2322 | static inline struct rq * |
2323 | context_switch(struct rq *rq, struct task_struct *prev, | 2323 | context_switch(struct rq *rq, struct task_struct *prev, |
2324 | struct task_struct *next) | 2324 | struct task_struct *next) |
2325 | { | 2325 | { |
2326 | struct mm_struct *mm, *oldmm; | 2326 | struct mm_struct *mm, *oldmm; |
2327 | 2327 | ||
2328 | prepare_task_switch(rq, prev, next); | 2328 | prepare_task_switch(rq, prev, next); |
2329 | 2329 | ||
2330 | mm = next->mm; | 2330 | mm = next->mm; |
2331 | oldmm = prev->active_mm; | 2331 | oldmm = prev->active_mm; |
2332 | /* | 2332 | /* |
2333 | * For paravirt, this is coupled with an exit in switch_to to | 2333 | * For paravirt, this is coupled with an exit in switch_to to |
2334 | * combine the page table reload and the switch backend into | 2334 | * combine the page table reload and the switch backend into |
2335 | * one hypercall. | 2335 | * one hypercall. |
2336 | */ | 2336 | */ |
2337 | arch_start_context_switch(prev); | 2337 | arch_start_context_switch(prev); |
2338 | 2338 | ||
2339 | if (!mm) { | 2339 | if (!mm) { |
2340 | next->active_mm = oldmm; | 2340 | next->active_mm = oldmm; |
2341 | atomic_inc(&oldmm->mm_count); | 2341 | atomic_inc(&oldmm->mm_count); |
2342 | enter_lazy_tlb(oldmm, next); | 2342 | enter_lazy_tlb(oldmm, next); |
2343 | } else | 2343 | } else |
2344 | switch_mm(oldmm, mm, next); | 2344 | switch_mm(oldmm, mm, next); |
2345 | 2345 | ||
2346 | if (!prev->mm) { | 2346 | if (!prev->mm) { |
2347 | prev->active_mm = NULL; | 2347 | prev->active_mm = NULL; |
2348 | rq->prev_mm = oldmm; | 2348 | rq->prev_mm = oldmm; |
2349 | } | 2349 | } |
2350 | /* | 2350 | /* |
2351 | * Since the runqueue lock will be released by the next | 2351 | * Since the runqueue lock will be released by the next |
2352 | * task (which is an invalid locking op but in the case | 2352 | * task (which is an invalid locking op but in the case |
2353 | * of the scheduler it's an obvious special-case), so we | 2353 | * of the scheduler it's an obvious special-case), so we |
2354 | * do an early lockdep release here: | 2354 | * do an early lockdep release here: |
2355 | */ | 2355 | */ |
2356 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); | 2356 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
2357 | 2357 | ||
2358 | context_tracking_task_switch(prev, next); | 2358 | context_tracking_task_switch(prev, next); |
2359 | /* Here we just switch the register state and the stack. */ | 2359 | /* Here we just switch the register state and the stack. */ |
2360 | switch_to(prev, next, prev); | 2360 | switch_to(prev, next, prev); |
2361 | barrier(); | 2361 | barrier(); |
2362 | 2362 | ||
2363 | return finish_task_switch(prev); | 2363 | return finish_task_switch(prev); |
2364 | } | 2364 | } |
2365 | 2365 | ||
2366 | /* | 2366 | /* |
2367 | * nr_running and nr_context_switches: | 2367 | * nr_running and nr_context_switches: |
2368 | * | 2368 | * |
2369 | * externally visible scheduler statistics: current number of runnable | 2369 | * externally visible scheduler statistics: current number of runnable |
2370 | * threads, total number of context switches performed since bootup. | 2370 | * threads, total number of context switches performed since bootup. |
2371 | */ | 2371 | */ |
2372 | unsigned long nr_running(void) | 2372 | unsigned long nr_running(void) |
2373 | { | 2373 | { |
2374 | unsigned long i, sum = 0; | 2374 | unsigned long i, sum = 0; |
2375 | 2375 | ||
2376 | for_each_online_cpu(i) | 2376 | for_each_online_cpu(i) |
2377 | sum += cpu_rq(i)->nr_running; | 2377 | sum += cpu_rq(i)->nr_running; |
2378 | 2378 | ||
2379 | return sum; | 2379 | return sum; |
2380 | } | 2380 | } |
2381 | 2381 | ||
2382 | /* | 2382 | /* |
2383 | * Check if only the current task is running on the cpu. | 2383 | * Check if only the current task is running on the cpu. |
2384 | */ | 2384 | */ |
2385 | bool single_task_running(void) | 2385 | bool single_task_running(void) |
2386 | { | 2386 | { |
2387 | if (cpu_rq(smp_processor_id())->nr_running == 1) | 2387 | if (cpu_rq(smp_processor_id())->nr_running == 1) |
2388 | return true; | 2388 | return true; |
2389 | else | 2389 | else |
2390 | return false; | 2390 | return false; |
2391 | } | 2391 | } |
2392 | EXPORT_SYMBOL(single_task_running); | 2392 | EXPORT_SYMBOL(single_task_running); |
2393 | 2393 | ||
2394 | unsigned long long nr_context_switches(void) | 2394 | unsigned long long nr_context_switches(void) |
2395 | { | 2395 | { |
2396 | int i; | 2396 | int i; |
2397 | unsigned long long sum = 0; | 2397 | unsigned long long sum = 0; |
2398 | 2398 | ||
2399 | for_each_possible_cpu(i) | 2399 | for_each_possible_cpu(i) |
2400 | sum += cpu_rq(i)->nr_switches; | 2400 | sum += cpu_rq(i)->nr_switches; |
2401 | 2401 | ||
2402 | return sum; | 2402 | return sum; |
2403 | } | 2403 | } |
2404 | 2404 | ||
2405 | unsigned long nr_iowait(void) | 2405 | unsigned long nr_iowait(void) |
2406 | { | 2406 | { |
2407 | unsigned long i, sum = 0; | 2407 | unsigned long i, sum = 0; |
2408 | 2408 | ||
2409 | for_each_possible_cpu(i) | 2409 | for_each_possible_cpu(i) |
2410 | sum += atomic_read(&cpu_rq(i)->nr_iowait); | 2410 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
2411 | 2411 | ||
2412 | return sum; | 2412 | return sum; |
2413 | } | 2413 | } |
2414 | 2414 | ||
2415 | unsigned long nr_iowait_cpu(int cpu) | 2415 | unsigned long nr_iowait_cpu(int cpu) |
2416 | { | 2416 | { |
2417 | struct rq *this = cpu_rq(cpu); | 2417 | struct rq *this = cpu_rq(cpu); |
2418 | return atomic_read(&this->nr_iowait); | 2418 | return atomic_read(&this->nr_iowait); |
2419 | } | 2419 | } |
2420 | 2420 | ||
2421 | void get_iowait_load(unsigned long *nr_waiters, unsigned long *load) | 2421 | void get_iowait_load(unsigned long *nr_waiters, unsigned long *load) |
2422 | { | 2422 | { |
2423 | struct rq *this = this_rq(); | 2423 | struct rq *this = this_rq(); |
2424 | *nr_waiters = atomic_read(&this->nr_iowait); | 2424 | *nr_waiters = atomic_read(&this->nr_iowait); |
2425 | *load = this->cpu_load[0]; | 2425 | *load = this->cpu_load[0]; |
2426 | } | 2426 | } |
2427 | 2427 | ||
2428 | #ifdef CONFIG_SMP | 2428 | #ifdef CONFIG_SMP |
2429 | 2429 | ||
2430 | /* | 2430 | /* |
2431 | * sched_exec - execve() is a valuable balancing opportunity, because at | 2431 | * sched_exec - execve() is a valuable balancing opportunity, because at |
2432 | * this point the task has the smallest effective memory and cache footprint. | 2432 | * this point the task has the smallest effective memory and cache footprint. |
2433 | */ | 2433 | */ |
2434 | void sched_exec(void) | 2434 | void sched_exec(void) |
2435 | { | 2435 | { |
2436 | struct task_struct *p = current; | 2436 | struct task_struct *p = current; |
2437 | unsigned long flags; | 2437 | unsigned long flags; |
2438 | int dest_cpu; | 2438 | int dest_cpu; |
2439 | 2439 | ||
2440 | raw_spin_lock_irqsave(&p->pi_lock, flags); | 2440 | raw_spin_lock_irqsave(&p->pi_lock, flags); |
2441 | dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); | 2441 | dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); |
2442 | if (dest_cpu == smp_processor_id()) | 2442 | if (dest_cpu == smp_processor_id()) |
2443 | goto unlock; | 2443 | goto unlock; |
2444 | 2444 | ||
2445 | if (likely(cpu_active(dest_cpu))) { | 2445 | if (likely(cpu_active(dest_cpu))) { |
2446 | struct migration_arg arg = { p, dest_cpu }; | 2446 | struct migration_arg arg = { p, dest_cpu }; |
2447 | 2447 | ||
2448 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); | 2448 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
2449 | stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); | 2449 | stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); |
2450 | return; | 2450 | return; |
2451 | } | 2451 | } |
2452 | unlock: | 2452 | unlock: |
2453 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); | 2453 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
2454 | } | 2454 | } |
2455 | 2455 | ||
2456 | #endif | 2456 | #endif |
2457 | 2457 | ||
2458 | DEFINE_PER_CPU(struct kernel_stat, kstat); | 2458 | DEFINE_PER_CPU(struct kernel_stat, kstat); |
2459 | DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); | 2459 | DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); |
2460 | 2460 | ||
2461 | EXPORT_PER_CPU_SYMBOL(kstat); | 2461 | EXPORT_PER_CPU_SYMBOL(kstat); |
2462 | EXPORT_PER_CPU_SYMBOL(kernel_cpustat); | 2462 | EXPORT_PER_CPU_SYMBOL(kernel_cpustat); |
2463 | 2463 | ||
2464 | /* | 2464 | /* |
2465 | * Return accounted runtime for the task. | 2465 | * Return accounted runtime for the task. |
2466 | * In case the task is currently running, return the runtime plus current's | 2466 | * In case the task is currently running, return the runtime plus current's |
2467 | * pending runtime that have not been accounted yet. | 2467 | * pending runtime that have not been accounted yet. |
2468 | */ | 2468 | */ |
2469 | unsigned long long task_sched_runtime(struct task_struct *p) | 2469 | unsigned long long task_sched_runtime(struct task_struct *p) |
2470 | { | 2470 | { |
2471 | unsigned long flags; | 2471 | unsigned long flags; |
2472 | struct rq *rq; | 2472 | struct rq *rq; |
2473 | u64 ns; | 2473 | u64 ns; |
2474 | 2474 | ||
2475 | #if defined(CONFIG_64BIT) && defined(CONFIG_SMP) | 2475 | #if defined(CONFIG_64BIT) && defined(CONFIG_SMP) |
2476 | /* | 2476 | /* |
2477 | * 64-bit doesn't need locks to atomically read a 64bit value. | 2477 | * 64-bit doesn't need locks to atomically read a 64bit value. |
2478 | * So we have a optimization chance when the task's delta_exec is 0. | 2478 | * So we have a optimization chance when the task's delta_exec is 0. |
2479 | * Reading ->on_cpu is racy, but this is ok. | 2479 | * Reading ->on_cpu is racy, but this is ok. |
2480 | * | 2480 | * |
2481 | * If we race with it leaving cpu, we'll take a lock. So we're correct. | 2481 | * If we race with it leaving cpu, we'll take a lock. So we're correct. |
2482 | * If we race with it entering cpu, unaccounted time is 0. This is | 2482 | * If we race with it entering cpu, unaccounted time is 0. This is |
2483 | * indistinguishable from the read occurring a few cycles earlier. | 2483 | * indistinguishable from the read occurring a few cycles earlier. |
2484 | * If we see ->on_cpu without ->on_rq, the task is leaving, and has | 2484 | * If we see ->on_cpu without ->on_rq, the task is leaving, and has |
2485 | * been accounted, so we're correct here as well. | 2485 | * been accounted, so we're correct here as well. |
2486 | */ | 2486 | */ |
2487 | if (!p->on_cpu || !task_on_rq_queued(p)) | 2487 | if (!p->on_cpu || !task_on_rq_queued(p)) |
2488 | return p->se.sum_exec_runtime; | 2488 | return p->se.sum_exec_runtime; |
2489 | #endif | 2489 | #endif |
2490 | 2490 | ||
2491 | rq = task_rq_lock(p, &flags); | 2491 | rq = task_rq_lock(p, &flags); |
2492 | /* | 2492 | /* |
2493 | * Must be ->curr _and_ ->on_rq. If dequeued, we would | 2493 | * Must be ->curr _and_ ->on_rq. If dequeued, we would |
2494 | * project cycles that may never be accounted to this | 2494 | * project cycles that may never be accounted to this |
2495 | * thread, breaking clock_gettime(). | 2495 | * thread, breaking clock_gettime(). |
2496 | */ | 2496 | */ |
2497 | if (task_current(rq, p) && task_on_rq_queued(p)) { | 2497 | if (task_current(rq, p) && task_on_rq_queued(p)) { |
2498 | update_rq_clock(rq); | 2498 | update_rq_clock(rq); |
2499 | p->sched_class->update_curr(rq); | 2499 | p->sched_class->update_curr(rq); |
2500 | } | 2500 | } |
2501 | ns = p->se.sum_exec_runtime; | 2501 | ns = p->se.sum_exec_runtime; |
2502 | task_rq_unlock(rq, p, &flags); | 2502 | task_rq_unlock(rq, p, &flags); |
2503 | 2503 | ||
2504 | return ns; | 2504 | return ns; |
2505 | } | 2505 | } |
2506 | 2506 | ||
2507 | /* | 2507 | /* |
2508 | * This function gets called by the timer code, with HZ frequency. | 2508 | * This function gets called by the timer code, with HZ frequency. |
2509 | * We call it with interrupts disabled. | 2509 | * We call it with interrupts disabled. |
2510 | */ | 2510 | */ |
2511 | void scheduler_tick(void) | 2511 | void scheduler_tick(void) |
2512 | { | 2512 | { |
2513 | int cpu = smp_processor_id(); | 2513 | int cpu = smp_processor_id(); |
2514 | struct rq *rq = cpu_rq(cpu); | 2514 | struct rq *rq = cpu_rq(cpu); |
2515 | struct task_struct *curr = rq->curr; | 2515 | struct task_struct *curr = rq->curr; |
2516 | 2516 | ||
2517 | sched_clock_tick(); | 2517 | sched_clock_tick(); |
2518 | 2518 | ||
2519 | raw_spin_lock(&rq->lock); | 2519 | raw_spin_lock(&rq->lock); |
2520 | update_rq_clock(rq); | 2520 | update_rq_clock(rq); |
2521 | curr->sched_class->task_tick(rq, curr, 0); | 2521 | curr->sched_class->task_tick(rq, curr, 0); |
2522 | update_cpu_load_active(rq); | 2522 | update_cpu_load_active(rq); |
2523 | raw_spin_unlock(&rq->lock); | 2523 | raw_spin_unlock(&rq->lock); |
2524 | 2524 | ||
2525 | perf_event_task_tick(); | 2525 | perf_event_task_tick(); |
2526 | 2526 | ||
2527 | #ifdef CONFIG_SMP | 2527 | #ifdef CONFIG_SMP |
2528 | rq->idle_balance = idle_cpu(cpu); | 2528 | rq->idle_balance = idle_cpu(cpu); |
2529 | trigger_load_balance(rq); | 2529 | trigger_load_balance(rq); |
2530 | #endif | 2530 | #endif |
2531 | rq_last_tick_reset(rq); | 2531 | rq_last_tick_reset(rq); |
2532 | } | 2532 | } |
2533 | 2533 | ||
2534 | #ifdef CONFIG_NO_HZ_FULL | 2534 | #ifdef CONFIG_NO_HZ_FULL |
2535 | /** | 2535 | /** |
2536 | * scheduler_tick_max_deferment | 2536 | * scheduler_tick_max_deferment |
2537 | * | 2537 | * |
2538 | * Keep at least one tick per second when a single | 2538 | * Keep at least one tick per second when a single |
2539 | * active task is running because the scheduler doesn't | 2539 | * active task is running because the scheduler doesn't |
2540 | * yet completely support full dynticks environment. | 2540 | * yet completely support full dynticks environment. |
2541 | * | 2541 | * |
2542 | * This makes sure that uptime, CFS vruntime, load | 2542 | * This makes sure that uptime, CFS vruntime, load |
2543 | * balancing, etc... continue to move forward, even | 2543 | * balancing, etc... continue to move forward, even |
2544 | * with a very low granularity. | 2544 | * with a very low granularity. |
2545 | * | 2545 | * |
2546 | * Return: Maximum deferment in nanoseconds. | 2546 | * Return: Maximum deferment in nanoseconds. |
2547 | */ | 2547 | */ |
2548 | u64 scheduler_tick_max_deferment(void) | 2548 | u64 scheduler_tick_max_deferment(void) |
2549 | { | 2549 | { |
2550 | struct rq *rq = this_rq(); | 2550 | struct rq *rq = this_rq(); |
2551 | unsigned long next, now = ACCESS_ONCE(jiffies); | 2551 | unsigned long next, now = ACCESS_ONCE(jiffies); |
2552 | 2552 | ||
2553 | next = rq->last_sched_tick + HZ; | 2553 | next = rq->last_sched_tick + HZ; |
2554 | 2554 | ||
2555 | if (time_before_eq(next, now)) | 2555 | if (time_before_eq(next, now)) |
2556 | return 0; | 2556 | return 0; |
2557 | 2557 | ||
2558 | return jiffies_to_nsecs(next - now); | 2558 | return jiffies_to_nsecs(next - now); |
2559 | } | 2559 | } |
2560 | #endif | 2560 | #endif |
2561 | 2561 | ||
2562 | notrace unsigned long get_parent_ip(unsigned long addr) | 2562 | notrace unsigned long get_parent_ip(unsigned long addr) |
2563 | { | 2563 | { |
2564 | if (in_lock_functions(addr)) { | 2564 | if (in_lock_functions(addr)) { |
2565 | addr = CALLER_ADDR2; | 2565 | addr = CALLER_ADDR2; |
2566 | if (in_lock_functions(addr)) | 2566 | if (in_lock_functions(addr)) |
2567 | addr = CALLER_ADDR3; | 2567 | addr = CALLER_ADDR3; |
2568 | } | 2568 | } |
2569 | return addr; | 2569 | return addr; |
2570 | } | 2570 | } |
2571 | 2571 | ||
2572 | #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ | 2572 | #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ |
2573 | defined(CONFIG_PREEMPT_TRACER)) | 2573 | defined(CONFIG_PREEMPT_TRACER)) |
2574 | 2574 | ||
2575 | void preempt_count_add(int val) | 2575 | void preempt_count_add(int val) |
2576 | { | 2576 | { |
2577 | #ifdef CONFIG_DEBUG_PREEMPT | 2577 | #ifdef CONFIG_DEBUG_PREEMPT |
2578 | /* | 2578 | /* |
2579 | * Underflow? | 2579 | * Underflow? |
2580 | */ | 2580 | */ |
2581 | if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) | 2581 | if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) |
2582 | return; | 2582 | return; |
2583 | #endif | 2583 | #endif |
2584 | __preempt_count_add(val); | 2584 | __preempt_count_add(val); |
2585 | #ifdef CONFIG_DEBUG_PREEMPT | 2585 | #ifdef CONFIG_DEBUG_PREEMPT |
2586 | /* | 2586 | /* |
2587 | * Spinlock count overflowing soon? | 2587 | * Spinlock count overflowing soon? |
2588 | */ | 2588 | */ |
2589 | DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= | 2589 | DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= |
2590 | PREEMPT_MASK - 10); | 2590 | PREEMPT_MASK - 10); |
2591 | #endif | 2591 | #endif |
2592 | if (preempt_count() == val) { | 2592 | if (preempt_count() == val) { |
2593 | unsigned long ip = get_parent_ip(CALLER_ADDR1); | 2593 | unsigned long ip = get_parent_ip(CALLER_ADDR1); |
2594 | #ifdef CONFIG_DEBUG_PREEMPT | 2594 | #ifdef CONFIG_DEBUG_PREEMPT |
2595 | current->preempt_disable_ip = ip; | 2595 | current->preempt_disable_ip = ip; |
2596 | #endif | 2596 | #endif |
2597 | trace_preempt_off(CALLER_ADDR0, ip); | 2597 | trace_preempt_off(CALLER_ADDR0, ip); |
2598 | } | 2598 | } |
2599 | } | 2599 | } |
2600 | EXPORT_SYMBOL(preempt_count_add); | 2600 | EXPORT_SYMBOL(preempt_count_add); |
2601 | NOKPROBE_SYMBOL(preempt_count_add); | 2601 | NOKPROBE_SYMBOL(preempt_count_add); |
2602 | 2602 | ||
2603 | void preempt_count_sub(int val) | 2603 | void preempt_count_sub(int val) |
2604 | { | 2604 | { |
2605 | #ifdef CONFIG_DEBUG_PREEMPT | 2605 | #ifdef CONFIG_DEBUG_PREEMPT |
2606 | /* | 2606 | /* |
2607 | * Underflow? | 2607 | * Underflow? |
2608 | */ | 2608 | */ |
2609 | if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) | 2609 | if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) |
2610 | return; | 2610 | return; |
2611 | /* | 2611 | /* |
2612 | * Is the spinlock portion underflowing? | 2612 | * Is the spinlock portion underflowing? |
2613 | */ | 2613 | */ |
2614 | if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && | 2614 | if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && |
2615 | !(preempt_count() & PREEMPT_MASK))) | 2615 | !(preempt_count() & PREEMPT_MASK))) |
2616 | return; | 2616 | return; |
2617 | #endif | 2617 | #endif |
2618 | 2618 | ||
2619 | if (preempt_count() == val) | 2619 | if (preempt_count() == val) |
2620 | trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); | 2620 | trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); |
2621 | __preempt_count_sub(val); | 2621 | __preempt_count_sub(val); |
2622 | } | 2622 | } |
2623 | EXPORT_SYMBOL(preempt_count_sub); | 2623 | EXPORT_SYMBOL(preempt_count_sub); |
2624 | NOKPROBE_SYMBOL(preempt_count_sub); | 2624 | NOKPROBE_SYMBOL(preempt_count_sub); |
2625 | 2625 | ||
2626 | #endif | 2626 | #endif |
2627 | 2627 | ||
2628 | /* | 2628 | /* |
2629 | * Print scheduling while atomic bug: | 2629 | * Print scheduling while atomic bug: |
2630 | */ | 2630 | */ |
2631 | static noinline void __schedule_bug(struct task_struct *prev) | 2631 | static noinline void __schedule_bug(struct task_struct *prev) |
2632 | { | 2632 | { |
2633 | if (oops_in_progress) | 2633 | if (oops_in_progress) |
2634 | return; | 2634 | return; |
2635 | 2635 | ||
2636 | printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", | 2636 | printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", |
2637 | prev->comm, prev->pid, preempt_count()); | 2637 | prev->comm, prev->pid, preempt_count()); |
2638 | 2638 | ||
2639 | debug_show_held_locks(prev); | 2639 | debug_show_held_locks(prev); |
2640 | print_modules(); | 2640 | print_modules(); |
2641 | if (irqs_disabled()) | 2641 | if (irqs_disabled()) |
2642 | print_irqtrace_events(prev); | 2642 | print_irqtrace_events(prev); |
2643 | #ifdef CONFIG_DEBUG_PREEMPT | 2643 | #ifdef CONFIG_DEBUG_PREEMPT |
2644 | if (in_atomic_preempt_off()) { | 2644 | if (in_atomic_preempt_off()) { |
2645 | pr_err("Preemption disabled at:"); | 2645 | pr_err("Preemption disabled at:"); |
2646 | print_ip_sym(current->preempt_disable_ip); | 2646 | print_ip_sym(current->preempt_disable_ip); |
2647 | pr_cont("\n"); | 2647 | pr_cont("\n"); |
2648 | } | 2648 | } |
2649 | #endif | 2649 | #endif |
2650 | dump_stack(); | 2650 | dump_stack(); |
2651 | add_taint(TAINT_WARN, LOCKDEP_STILL_OK); | 2651 | add_taint(TAINT_WARN, LOCKDEP_STILL_OK); |
2652 | } | 2652 | } |
2653 | 2653 | ||
2654 | /* | 2654 | /* |
2655 | * Various schedule()-time debugging checks and statistics: | 2655 | * Various schedule()-time debugging checks and statistics: |
2656 | */ | 2656 | */ |
2657 | static inline void schedule_debug(struct task_struct *prev) | 2657 | static inline void schedule_debug(struct task_struct *prev) |
2658 | { | 2658 | { |
2659 | #ifdef CONFIG_SCHED_STACK_END_CHECK | 2659 | #ifdef CONFIG_SCHED_STACK_END_CHECK |
2660 | BUG_ON(unlikely(task_stack_end_corrupted(prev))); | 2660 | BUG_ON(unlikely(task_stack_end_corrupted(prev))); |
2661 | #endif | 2661 | #endif |
2662 | /* | 2662 | /* |
2663 | * Test if we are atomic. Since do_exit() needs to call into | 2663 | * Test if we are atomic. Since do_exit() needs to call into |
2664 | * schedule() atomically, we ignore that path. Otherwise whine | 2664 | * schedule() atomically, we ignore that path. Otherwise whine |
2665 | * if we are scheduling when we should not. | 2665 | * if we are scheduling when we should not. |
2666 | */ | 2666 | */ |
2667 | if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) | 2667 | if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) |
2668 | __schedule_bug(prev); | 2668 | __schedule_bug(prev); |
2669 | rcu_sleep_check(); | 2669 | rcu_sleep_check(); |
2670 | 2670 | ||
2671 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); | 2671 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); |
2672 | 2672 | ||
2673 | schedstat_inc(this_rq(), sched_count); | 2673 | schedstat_inc(this_rq(), sched_count); |
2674 | } | 2674 | } |
2675 | 2675 | ||
2676 | /* | 2676 | /* |
2677 | * Pick up the highest-prio task: | 2677 | * Pick up the highest-prio task: |
2678 | */ | 2678 | */ |
2679 | static inline struct task_struct * | 2679 | static inline struct task_struct * |
2680 | pick_next_task(struct rq *rq, struct task_struct *prev) | 2680 | pick_next_task(struct rq *rq, struct task_struct *prev) |
2681 | { | 2681 | { |
2682 | const struct sched_class *class = &fair_sched_class; | 2682 | const struct sched_class *class = &fair_sched_class; |
2683 | struct task_struct *p; | 2683 | struct task_struct *p; |
2684 | 2684 | ||
2685 | /* | 2685 | /* |
2686 | * Optimization: we know that if all tasks are in | 2686 | * Optimization: we know that if all tasks are in |
2687 | * the fair class we can call that function directly: | 2687 | * the fair class we can call that function directly: |
2688 | */ | 2688 | */ |
2689 | if (likely(prev->sched_class == class && | 2689 | if (likely(prev->sched_class == class && |
2690 | rq->nr_running == rq->cfs.h_nr_running)) { | 2690 | rq->nr_running == rq->cfs.h_nr_running)) { |
2691 | p = fair_sched_class.pick_next_task(rq, prev); | 2691 | p = fair_sched_class.pick_next_task(rq, prev); |
2692 | if (unlikely(p == RETRY_TASK)) | 2692 | if (unlikely(p == RETRY_TASK)) |
2693 | goto again; | 2693 | goto again; |
2694 | 2694 | ||
2695 | /* assumes fair_sched_class->next == idle_sched_class */ | 2695 | /* assumes fair_sched_class->next == idle_sched_class */ |
2696 | if (unlikely(!p)) | 2696 | if (unlikely(!p)) |
2697 | p = idle_sched_class.pick_next_task(rq, prev); | 2697 | p = idle_sched_class.pick_next_task(rq, prev); |
2698 | 2698 | ||
2699 | return p; | 2699 | return p; |
2700 | } | 2700 | } |
2701 | 2701 | ||
2702 | again: | 2702 | again: |
2703 | for_each_class(class) { | 2703 | for_each_class(class) { |
2704 | p = class->pick_next_task(rq, prev); | 2704 | p = class->pick_next_task(rq, prev); |
2705 | if (p) { | 2705 | if (p) { |
2706 | if (unlikely(p == RETRY_TASK)) | 2706 | if (unlikely(p == RETRY_TASK)) |
2707 | goto again; | 2707 | goto again; |
2708 | return p; | 2708 | return p; |
2709 | } | 2709 | } |
2710 | } | 2710 | } |
2711 | 2711 | ||
2712 | BUG(); /* the idle class will always have a runnable task */ | 2712 | BUG(); /* the idle class will always have a runnable task */ |
2713 | } | 2713 | } |
2714 | 2714 | ||
2715 | /* | 2715 | /* |
2716 | * __schedule() is the main scheduler function. | 2716 | * __schedule() is the main scheduler function. |
2717 | * | 2717 | * |
2718 | * The main means of driving the scheduler and thus entering this function are: | 2718 | * The main means of driving the scheduler and thus entering this function are: |
2719 | * | 2719 | * |
2720 | * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. | 2720 | * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. |
2721 | * | 2721 | * |
2722 | * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return | 2722 | * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return |
2723 | * paths. For example, see arch/x86/entry_64.S. | 2723 | * paths. For example, see arch/x86/entry_64.S. |
2724 | * | 2724 | * |
2725 | * To drive preemption between tasks, the scheduler sets the flag in timer | 2725 | * To drive preemption between tasks, the scheduler sets the flag in timer |
2726 | * interrupt handler scheduler_tick(). | 2726 | * interrupt handler scheduler_tick(). |
2727 | * | 2727 | * |
2728 | * 3. Wakeups don't really cause entry into schedule(). They add a | 2728 | * 3. Wakeups don't really cause entry into schedule(). They add a |
2729 | * task to the run-queue and that's it. | 2729 | * task to the run-queue and that's it. |
2730 | * | 2730 | * |
2731 | * Now, if the new task added to the run-queue preempts the current | 2731 | * Now, if the new task added to the run-queue preempts the current |
2732 | * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets | 2732 | * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets |
2733 | * called on the nearest possible occasion: | 2733 | * called on the nearest possible occasion: |
2734 | * | 2734 | * |
2735 | * - If the kernel is preemptible (CONFIG_PREEMPT=y): | 2735 | * - If the kernel is preemptible (CONFIG_PREEMPT=y): |
2736 | * | 2736 | * |
2737 | * - in syscall or exception context, at the next outmost | 2737 | * - in syscall or exception context, at the next outmost |
2738 | * preempt_enable(). (this might be as soon as the wake_up()'s | 2738 | * preempt_enable(). (this might be as soon as the wake_up()'s |
2739 | * spin_unlock()!) | 2739 | * spin_unlock()!) |
2740 | * | 2740 | * |
2741 | * - in IRQ context, return from interrupt-handler to | 2741 | * - in IRQ context, return from interrupt-handler to |
2742 | * preemptible context | 2742 | * preemptible context |
2743 | * | 2743 | * |
2744 | * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) | 2744 | * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) |
2745 | * then at the next: | 2745 | * then at the next: |
2746 | * | 2746 | * |
2747 | * - cond_resched() call | 2747 | * - cond_resched() call |
2748 | * - explicit schedule() call | 2748 | * - explicit schedule() call |
2749 | * - return from syscall or exception to user-space | 2749 | * - return from syscall or exception to user-space |
2750 | * - return from interrupt-handler to user-space | 2750 | * - return from interrupt-handler to user-space |
2751 | */ | 2751 | */ |
2752 | static void __sched __schedule(void) | 2752 | static void __sched __schedule(void) |
2753 | { | 2753 | { |
2754 | struct task_struct *prev, *next; | 2754 | struct task_struct *prev, *next; |
2755 | unsigned long *switch_count; | 2755 | unsigned long *switch_count; |
2756 | struct rq *rq; | 2756 | struct rq *rq; |
2757 | int cpu; | 2757 | int cpu; |
2758 | 2758 | ||
2759 | need_resched: | 2759 | need_resched: |
2760 | preempt_disable(); | 2760 | preempt_disable(); |
2761 | cpu = smp_processor_id(); | 2761 | cpu = smp_processor_id(); |
2762 | rq = cpu_rq(cpu); | 2762 | rq = cpu_rq(cpu); |
2763 | rcu_note_context_switch(); | 2763 | rcu_note_context_switch(); |
2764 | prev = rq->curr; | 2764 | prev = rq->curr; |
2765 | 2765 | ||
2766 | schedule_debug(prev); | 2766 | schedule_debug(prev); |
2767 | 2767 | ||
2768 | if (sched_feat(HRTICK)) | 2768 | if (sched_feat(HRTICK)) |
2769 | hrtick_clear(rq); | 2769 | hrtick_clear(rq); |
2770 | 2770 | ||
2771 | /* | 2771 | /* |
2772 | * Make sure that signal_pending_state()->signal_pending() below | 2772 | * Make sure that signal_pending_state()->signal_pending() below |
2773 | * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) | 2773 | * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) |
2774 | * done by the caller to avoid the race with signal_wake_up(). | 2774 | * done by the caller to avoid the race with signal_wake_up(). |
2775 | */ | 2775 | */ |
2776 | smp_mb__before_spinlock(); | 2776 | smp_mb__before_spinlock(); |
2777 | raw_spin_lock_irq(&rq->lock); | 2777 | raw_spin_lock_irq(&rq->lock); |
2778 | 2778 | ||
2779 | switch_count = &prev->nivcsw; | 2779 | switch_count = &prev->nivcsw; |
2780 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { | 2780 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { |
2781 | if (unlikely(signal_pending_state(prev->state, prev))) { | 2781 | if (unlikely(signal_pending_state(prev->state, prev))) { |
2782 | prev->state = TASK_RUNNING; | 2782 | prev->state = TASK_RUNNING; |
2783 | } else { | 2783 | } else { |
2784 | deactivate_task(rq, prev, DEQUEUE_SLEEP); | 2784 | deactivate_task(rq, prev, DEQUEUE_SLEEP); |
2785 | prev->on_rq = 0; | 2785 | prev->on_rq = 0; |
2786 | 2786 | ||
2787 | /* | 2787 | /* |
2788 | * If a worker went to sleep, notify and ask workqueue | 2788 | * If a worker went to sleep, notify and ask workqueue |
2789 | * whether it wants to wake up a task to maintain | 2789 | * whether it wants to wake up a task to maintain |
2790 | * concurrency. | 2790 | * concurrency. |
2791 | */ | 2791 | */ |
2792 | if (prev->flags & PF_WQ_WORKER) { | 2792 | if (prev->flags & PF_WQ_WORKER) { |
2793 | struct task_struct *to_wakeup; | 2793 | struct task_struct *to_wakeup; |
2794 | 2794 | ||
2795 | to_wakeup = wq_worker_sleeping(prev, cpu); | 2795 | to_wakeup = wq_worker_sleeping(prev, cpu); |
2796 | if (to_wakeup) | 2796 | if (to_wakeup) |
2797 | try_to_wake_up_local(to_wakeup); | 2797 | try_to_wake_up_local(to_wakeup); |
2798 | } | 2798 | } |
2799 | } | 2799 | } |
2800 | switch_count = &prev->nvcsw; | 2800 | switch_count = &prev->nvcsw; |
2801 | } | 2801 | } |
2802 | 2802 | ||
2803 | if (task_on_rq_queued(prev) || rq->skip_clock_update < 0) | 2803 | if (task_on_rq_queued(prev) || rq->skip_clock_update < 0) |
2804 | update_rq_clock(rq); | 2804 | update_rq_clock(rq); |
2805 | 2805 | ||
2806 | next = pick_next_task(rq, prev); | 2806 | next = pick_next_task(rq, prev); |
2807 | clear_tsk_need_resched(prev); | 2807 | clear_tsk_need_resched(prev); |
2808 | clear_preempt_need_resched(); | 2808 | clear_preempt_need_resched(); |
2809 | rq->skip_clock_update = 0; | 2809 | rq->skip_clock_update = 0; |
2810 | 2810 | ||
2811 | if (likely(prev != next)) { | 2811 | if (likely(prev != next)) { |
2812 | rq->nr_switches++; | 2812 | rq->nr_switches++; |
2813 | rq->curr = next; | 2813 | rq->curr = next; |
2814 | ++*switch_count; | 2814 | ++*switch_count; |
2815 | 2815 | ||
2816 | rq = context_switch(rq, prev, next); /* unlocks the rq */ | 2816 | rq = context_switch(rq, prev, next); /* unlocks the rq */ |
2817 | cpu = cpu_of(rq); | 2817 | cpu = cpu_of(rq); |
2818 | } else | 2818 | } else |
2819 | raw_spin_unlock_irq(&rq->lock); | 2819 | raw_spin_unlock_irq(&rq->lock); |
2820 | 2820 | ||
2821 | post_schedule(rq); | 2821 | post_schedule(rq); |
2822 | 2822 | ||
2823 | sched_preempt_enable_no_resched(); | 2823 | sched_preempt_enable_no_resched(); |
2824 | if (need_resched()) | 2824 | if (need_resched()) |
2825 | goto need_resched; | 2825 | goto need_resched; |
2826 | } | 2826 | } |
2827 | 2827 | ||
2828 | static inline void sched_submit_work(struct task_struct *tsk) | 2828 | static inline void sched_submit_work(struct task_struct *tsk) |
2829 | { | 2829 | { |
2830 | if (!tsk->state || tsk_is_pi_blocked(tsk)) | 2830 | if (!tsk->state || tsk_is_pi_blocked(tsk)) |
2831 | return; | 2831 | return; |
2832 | /* | 2832 | /* |
2833 | * If we are going to sleep and we have plugged IO queued, | 2833 | * If we are going to sleep and we have plugged IO queued, |
2834 | * make sure to submit it to avoid deadlocks. | 2834 | * make sure to submit it to avoid deadlocks. |
2835 | */ | 2835 | */ |
2836 | if (blk_needs_flush_plug(tsk)) | 2836 | if (blk_needs_flush_plug(tsk)) |
2837 | blk_schedule_flush_plug(tsk); | 2837 | blk_schedule_flush_plug(tsk); |
2838 | } | 2838 | } |
2839 | 2839 | ||
2840 | asmlinkage __visible void __sched schedule(void) | 2840 | asmlinkage __visible void __sched schedule(void) |
2841 | { | 2841 | { |
2842 | struct task_struct *tsk = current; | 2842 | struct task_struct *tsk = current; |
2843 | 2843 | ||
2844 | sched_submit_work(tsk); | 2844 | sched_submit_work(tsk); |
2845 | __schedule(); | 2845 | __schedule(); |
2846 | } | 2846 | } |
2847 | EXPORT_SYMBOL(schedule); | 2847 | EXPORT_SYMBOL(schedule); |
2848 | 2848 | ||
2849 | #ifdef CONFIG_CONTEXT_TRACKING | 2849 | #ifdef CONFIG_CONTEXT_TRACKING |
2850 | asmlinkage __visible void __sched schedule_user(void) | 2850 | asmlinkage __visible void __sched schedule_user(void) |
2851 | { | 2851 | { |
2852 | /* | 2852 | /* |
2853 | * If we come here after a random call to set_need_resched(), | 2853 | * If we come here after a random call to set_need_resched(), |
2854 | * or we have been woken up remotely but the IPI has not yet arrived, | 2854 | * or we have been woken up remotely but the IPI has not yet arrived, |
2855 | * we haven't yet exited the RCU idle mode. Do it here manually until | 2855 | * we haven't yet exited the RCU idle mode. Do it here manually until |
2856 | * we find a better solution. | 2856 | * we find a better solution. |
2857 | * | 2857 | * |
2858 | * NB: There are buggy callers of this function. Ideally we | 2858 | * NB: There are buggy callers of this function. Ideally we |
2859 | * should warn if prev_state != IN_USER, but that will trigger | 2859 | * should warn if prev_state != IN_USER, but that will trigger |
2860 | * too frequently to make sense yet. | 2860 | * too frequently to make sense yet. |
2861 | */ | 2861 | */ |
2862 | enum ctx_state prev_state = exception_enter(); | 2862 | enum ctx_state prev_state = exception_enter(); |
2863 | schedule(); | 2863 | schedule(); |
2864 | exception_exit(prev_state); | 2864 | exception_exit(prev_state); |
2865 | } | 2865 | } |
2866 | #endif | 2866 | #endif |
2867 | 2867 | ||
2868 | /** | 2868 | /** |
2869 | * schedule_preempt_disabled - called with preemption disabled | 2869 | * schedule_preempt_disabled - called with preemption disabled |
2870 | * | 2870 | * |
2871 | * Returns with preemption disabled. Note: preempt_count must be 1 | 2871 | * Returns with preemption disabled. Note: preempt_count must be 1 |
2872 | */ | 2872 | */ |
2873 | void __sched schedule_preempt_disabled(void) | 2873 | void __sched schedule_preempt_disabled(void) |
2874 | { | 2874 | { |
2875 | sched_preempt_enable_no_resched(); | 2875 | sched_preempt_enable_no_resched(); |
2876 | schedule(); | 2876 | schedule(); |
2877 | preempt_disable(); | 2877 | preempt_disable(); |
2878 | } | 2878 | } |
2879 | 2879 | ||
2880 | #ifdef CONFIG_PREEMPT | 2880 | #ifdef CONFIG_PREEMPT |
2881 | /* | 2881 | /* |
2882 | * this is the entry point to schedule() from in-kernel preemption | 2882 | * this is the entry point to schedule() from in-kernel preemption |
2883 | * off of preempt_enable. Kernel preemptions off return from interrupt | 2883 | * off of preempt_enable. Kernel preemptions off return from interrupt |
2884 | * occur there and call schedule directly. | 2884 | * occur there and call schedule directly. |
2885 | */ | 2885 | */ |
2886 | asmlinkage __visible void __sched notrace preempt_schedule(void) | 2886 | asmlinkage __visible void __sched notrace preempt_schedule(void) |
2887 | { | 2887 | { |
2888 | /* | 2888 | /* |
2889 | * If there is a non-zero preempt_count or interrupts are disabled, | 2889 | * If there is a non-zero preempt_count or interrupts are disabled, |
2890 | * we do not want to preempt the current task. Just return.. | 2890 | * we do not want to preempt the current task. Just return.. |
2891 | */ | 2891 | */ |
2892 | if (likely(!preemptible())) | 2892 | if (likely(!preemptible())) |
2893 | return; | 2893 | return; |
2894 | 2894 | ||
2895 | do { | 2895 | do { |
2896 | __preempt_count_add(PREEMPT_ACTIVE); | 2896 | __preempt_count_add(PREEMPT_ACTIVE); |
2897 | __schedule(); | 2897 | __schedule(); |
2898 | __preempt_count_sub(PREEMPT_ACTIVE); | 2898 | __preempt_count_sub(PREEMPT_ACTIVE); |
2899 | 2899 | ||
2900 | /* | 2900 | /* |
2901 | * Check again in case we missed a preemption opportunity | 2901 | * Check again in case we missed a preemption opportunity |
2902 | * between schedule and now. | 2902 | * between schedule and now. |
2903 | */ | 2903 | */ |
2904 | barrier(); | 2904 | barrier(); |
2905 | } while (need_resched()); | 2905 | } while (need_resched()); |
2906 | } | 2906 | } |
2907 | NOKPROBE_SYMBOL(preempt_schedule); | 2907 | NOKPROBE_SYMBOL(preempt_schedule); |
2908 | EXPORT_SYMBOL(preempt_schedule); | 2908 | EXPORT_SYMBOL(preempt_schedule); |
2909 | 2909 | ||
2910 | #ifdef CONFIG_CONTEXT_TRACKING | 2910 | #ifdef CONFIG_CONTEXT_TRACKING |
2911 | /** | 2911 | /** |
2912 | * preempt_schedule_context - preempt_schedule called by tracing | 2912 | * preempt_schedule_context - preempt_schedule called by tracing |
2913 | * | 2913 | * |
2914 | * The tracing infrastructure uses preempt_enable_notrace to prevent | 2914 | * The tracing infrastructure uses preempt_enable_notrace to prevent |
2915 | * recursion and tracing preempt enabling caused by the tracing | 2915 | * recursion and tracing preempt enabling caused by the tracing |
2916 | * infrastructure itself. But as tracing can happen in areas coming | 2916 | * infrastructure itself. But as tracing can happen in areas coming |
2917 | * from userspace or just about to enter userspace, a preempt enable | 2917 | * from userspace or just about to enter userspace, a preempt enable |
2918 | * can occur before user_exit() is called. This will cause the scheduler | 2918 | * can occur before user_exit() is called. This will cause the scheduler |
2919 | * to be called when the system is still in usermode. | 2919 | * to be called when the system is still in usermode. |
2920 | * | 2920 | * |
2921 | * To prevent this, the preempt_enable_notrace will use this function | 2921 | * To prevent this, the preempt_enable_notrace will use this function |
2922 | * instead of preempt_schedule() to exit user context if needed before | 2922 | * instead of preempt_schedule() to exit user context if needed before |
2923 | * calling the scheduler. | 2923 | * calling the scheduler. |
2924 | */ | 2924 | */ |
2925 | asmlinkage __visible void __sched notrace preempt_schedule_context(void) | 2925 | asmlinkage __visible void __sched notrace preempt_schedule_context(void) |
2926 | { | 2926 | { |
2927 | enum ctx_state prev_ctx; | 2927 | enum ctx_state prev_ctx; |
2928 | 2928 | ||
2929 | if (likely(!preemptible())) | 2929 | if (likely(!preemptible())) |
2930 | return; | 2930 | return; |
2931 | 2931 | ||
2932 | do { | 2932 | do { |
2933 | __preempt_count_add(PREEMPT_ACTIVE); | 2933 | __preempt_count_add(PREEMPT_ACTIVE); |
2934 | /* | 2934 | /* |
2935 | * Needs preempt disabled in case user_exit() is traced | 2935 | * Needs preempt disabled in case user_exit() is traced |
2936 | * and the tracer calls preempt_enable_notrace() causing | 2936 | * and the tracer calls preempt_enable_notrace() causing |
2937 | * an infinite recursion. | 2937 | * an infinite recursion. |
2938 | */ | 2938 | */ |
2939 | prev_ctx = exception_enter(); | 2939 | prev_ctx = exception_enter(); |
2940 | __schedule(); | 2940 | __schedule(); |
2941 | exception_exit(prev_ctx); | 2941 | exception_exit(prev_ctx); |
2942 | 2942 | ||
2943 | __preempt_count_sub(PREEMPT_ACTIVE); | 2943 | __preempt_count_sub(PREEMPT_ACTIVE); |
2944 | barrier(); | 2944 | barrier(); |
2945 | } while (need_resched()); | 2945 | } while (need_resched()); |
2946 | } | 2946 | } |
2947 | EXPORT_SYMBOL_GPL(preempt_schedule_context); | 2947 | EXPORT_SYMBOL_GPL(preempt_schedule_context); |
2948 | #endif /* CONFIG_CONTEXT_TRACKING */ | 2948 | #endif /* CONFIG_CONTEXT_TRACKING */ |
2949 | 2949 | ||
2950 | #endif /* CONFIG_PREEMPT */ | 2950 | #endif /* CONFIG_PREEMPT */ |
2951 | 2951 | ||
2952 | /* | 2952 | /* |
2953 | * this is the entry point to schedule() from kernel preemption | 2953 | * this is the entry point to schedule() from kernel preemption |
2954 | * off of irq context. | 2954 | * off of irq context. |
2955 | * Note, that this is called and return with irqs disabled. This will | 2955 | * Note, that this is called and return with irqs disabled. This will |
2956 | * protect us against recursive calling from irq. | 2956 | * protect us against recursive calling from irq. |
2957 | */ | 2957 | */ |
2958 | asmlinkage __visible void __sched preempt_schedule_irq(void) | 2958 | asmlinkage __visible void __sched preempt_schedule_irq(void) |
2959 | { | 2959 | { |
2960 | enum ctx_state prev_state; | 2960 | enum ctx_state prev_state; |
2961 | 2961 | ||
2962 | /* Catch callers which need to be fixed */ | 2962 | /* Catch callers which need to be fixed */ |
2963 | BUG_ON(preempt_count() || !irqs_disabled()); | 2963 | BUG_ON(preempt_count() || !irqs_disabled()); |
2964 | 2964 | ||
2965 | prev_state = exception_enter(); | 2965 | prev_state = exception_enter(); |
2966 | 2966 | ||
2967 | do { | 2967 | do { |
2968 | __preempt_count_add(PREEMPT_ACTIVE); | 2968 | __preempt_count_add(PREEMPT_ACTIVE); |
2969 | local_irq_enable(); | 2969 | local_irq_enable(); |
2970 | __schedule(); | 2970 | __schedule(); |
2971 | local_irq_disable(); | 2971 | local_irq_disable(); |
2972 | __preempt_count_sub(PREEMPT_ACTIVE); | 2972 | __preempt_count_sub(PREEMPT_ACTIVE); |
2973 | 2973 | ||
2974 | /* | 2974 | /* |
2975 | * Check again in case we missed a preemption opportunity | 2975 | * Check again in case we missed a preemption opportunity |
2976 | * between schedule and now. | 2976 | * between schedule and now. |
2977 | */ | 2977 | */ |
2978 | barrier(); | 2978 | barrier(); |
2979 | } while (need_resched()); | 2979 | } while (need_resched()); |
2980 | 2980 | ||
2981 | exception_exit(prev_state); | 2981 | exception_exit(prev_state); |
2982 | } | 2982 | } |
2983 | 2983 | ||
2984 | int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, | 2984 | int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, |
2985 | void *key) | 2985 | void *key) |
2986 | { | 2986 | { |
2987 | return try_to_wake_up(curr->private, mode, wake_flags); | 2987 | return try_to_wake_up(curr->private, mode, wake_flags); |
2988 | } | 2988 | } |
2989 | EXPORT_SYMBOL(default_wake_function); | 2989 | EXPORT_SYMBOL(default_wake_function); |
2990 | 2990 | ||
2991 | #ifdef CONFIG_RT_MUTEXES | 2991 | #ifdef CONFIG_RT_MUTEXES |
2992 | 2992 | ||
2993 | /* | 2993 | /* |
2994 | * rt_mutex_setprio - set the current priority of a task | 2994 | * rt_mutex_setprio - set the current priority of a task |
2995 | * @p: task | 2995 | * @p: task |
2996 | * @prio: prio value (kernel-internal form) | 2996 | * @prio: prio value (kernel-internal form) |
2997 | * | 2997 | * |
2998 | * This function changes the 'effective' priority of a task. It does | 2998 | * This function changes the 'effective' priority of a task. It does |
2999 | * not touch ->normal_prio like __setscheduler(). | 2999 | * not touch ->normal_prio like __setscheduler(). |
3000 | * | 3000 | * |
3001 | * Used by the rt_mutex code to implement priority inheritance | 3001 | * Used by the rt_mutex code to implement priority inheritance |
3002 | * logic. Call site only calls if the priority of the task changed. | 3002 | * logic. Call site only calls if the priority of the task changed. |
3003 | */ | 3003 | */ |
3004 | void rt_mutex_setprio(struct task_struct *p, int prio) | 3004 | void rt_mutex_setprio(struct task_struct *p, int prio) |
3005 | { | 3005 | { |
3006 | int oldprio, queued, running, enqueue_flag = 0; | 3006 | int oldprio, queued, running, enqueue_flag = 0; |
3007 | struct rq *rq; | 3007 | struct rq *rq; |
3008 | const struct sched_class *prev_class; | 3008 | const struct sched_class *prev_class; |
3009 | 3009 | ||
3010 | BUG_ON(prio > MAX_PRIO); | 3010 | BUG_ON(prio > MAX_PRIO); |
3011 | 3011 | ||
3012 | rq = __task_rq_lock(p); | 3012 | rq = __task_rq_lock(p); |
3013 | 3013 | ||
3014 | /* | 3014 | /* |
3015 | * Idle task boosting is a nono in general. There is one | 3015 | * Idle task boosting is a nono in general. There is one |
3016 | * exception, when PREEMPT_RT and NOHZ is active: | 3016 | * exception, when PREEMPT_RT and NOHZ is active: |
3017 | * | 3017 | * |
3018 | * The idle task calls get_next_timer_interrupt() and holds | 3018 | * The idle task calls get_next_timer_interrupt() and holds |
3019 | * the timer wheel base->lock on the CPU and another CPU wants | 3019 | * the timer wheel base->lock on the CPU and another CPU wants |
3020 | * to access the timer (probably to cancel it). We can safely | 3020 | * to access the timer (probably to cancel it). We can safely |
3021 | * ignore the boosting request, as the idle CPU runs this code | 3021 | * ignore the boosting request, as the idle CPU runs this code |
3022 | * with interrupts disabled and will complete the lock | 3022 | * with interrupts disabled and will complete the lock |
3023 | * protected section without being interrupted. So there is no | 3023 | * protected section without being interrupted. So there is no |
3024 | * real need to boost. | 3024 | * real need to boost. |
3025 | */ | 3025 | */ |
3026 | if (unlikely(p == rq->idle)) { | 3026 | if (unlikely(p == rq->idle)) { |
3027 | WARN_ON(p != rq->curr); | 3027 | WARN_ON(p != rq->curr); |
3028 | WARN_ON(p->pi_blocked_on); | 3028 | WARN_ON(p->pi_blocked_on); |
3029 | goto out_unlock; | 3029 | goto out_unlock; |
3030 | } | 3030 | } |
3031 | 3031 | ||
3032 | trace_sched_pi_setprio(p, prio); | 3032 | trace_sched_pi_setprio(p, prio); |
3033 | oldprio = p->prio; | 3033 | oldprio = p->prio; |
3034 | prev_class = p->sched_class; | 3034 | prev_class = p->sched_class; |
3035 | queued = task_on_rq_queued(p); | 3035 | queued = task_on_rq_queued(p); |
3036 | running = task_current(rq, p); | 3036 | running = task_current(rq, p); |
3037 | if (queued) | 3037 | if (queued) |
3038 | dequeue_task(rq, p, 0); | 3038 | dequeue_task(rq, p, 0); |
3039 | if (running) | 3039 | if (running) |
3040 | put_prev_task(rq, p); | 3040 | put_prev_task(rq, p); |
3041 | 3041 | ||
3042 | /* | 3042 | /* |
3043 | * Boosting condition are: | 3043 | * Boosting condition are: |
3044 | * 1. -rt task is running and holds mutex A | 3044 | * 1. -rt task is running and holds mutex A |
3045 | * --> -dl task blocks on mutex A | 3045 | * --> -dl task blocks on mutex A |
3046 | * | 3046 | * |
3047 | * 2. -dl task is running and holds mutex A | 3047 | * 2. -dl task is running and holds mutex A |
3048 | * --> -dl task blocks on mutex A and could preempt the | 3048 | * --> -dl task blocks on mutex A and could preempt the |
3049 | * running task | 3049 | * running task |
3050 | */ | 3050 | */ |
3051 | if (dl_prio(prio)) { | 3051 | if (dl_prio(prio)) { |
3052 | struct task_struct *pi_task = rt_mutex_get_top_task(p); | 3052 | struct task_struct *pi_task = rt_mutex_get_top_task(p); |
3053 | if (!dl_prio(p->normal_prio) || | 3053 | if (!dl_prio(p->normal_prio) || |
3054 | (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) { | 3054 | (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) { |
3055 | p->dl.dl_boosted = 1; | 3055 | p->dl.dl_boosted = 1; |
3056 | p->dl.dl_throttled = 0; | 3056 | p->dl.dl_throttled = 0; |
3057 | enqueue_flag = ENQUEUE_REPLENISH; | 3057 | enqueue_flag = ENQUEUE_REPLENISH; |
3058 | } else | 3058 | } else |
3059 | p->dl.dl_boosted = 0; | 3059 | p->dl.dl_boosted = 0; |
3060 | p->sched_class = &dl_sched_class; | 3060 | p->sched_class = &dl_sched_class; |
3061 | } else if (rt_prio(prio)) { | 3061 | } else if (rt_prio(prio)) { |
3062 | if (dl_prio(oldprio)) | 3062 | if (dl_prio(oldprio)) |
3063 | p->dl.dl_boosted = 0; | 3063 | p->dl.dl_boosted = 0; |
3064 | if (oldprio < prio) | 3064 | if (oldprio < prio) |
3065 | enqueue_flag = ENQUEUE_HEAD; | 3065 | enqueue_flag = ENQUEUE_HEAD; |
3066 | p->sched_class = &rt_sched_class; | 3066 | p->sched_class = &rt_sched_class; |
3067 | } else { | 3067 | } else { |
3068 | if (dl_prio(oldprio)) | 3068 | if (dl_prio(oldprio)) |
3069 | p->dl.dl_boosted = 0; | 3069 | p->dl.dl_boosted = 0; |
3070 | p->sched_class = &fair_sched_class; | 3070 | p->sched_class = &fair_sched_class; |
3071 | } | 3071 | } |
3072 | 3072 | ||
3073 | p->prio = prio; | 3073 | p->prio = prio; |
3074 | 3074 | ||
3075 | if (running) | 3075 | if (running) |
3076 | p->sched_class->set_curr_task(rq); | 3076 | p->sched_class->set_curr_task(rq); |
3077 | if (queued) | 3077 | if (queued) |
3078 | enqueue_task(rq, p, enqueue_flag); | 3078 | enqueue_task(rq, p, enqueue_flag); |
3079 | 3079 | ||
3080 | check_class_changed(rq, p, prev_class, oldprio); | 3080 | check_class_changed(rq, p, prev_class, oldprio); |
3081 | out_unlock: | 3081 | out_unlock: |
3082 | __task_rq_unlock(rq); | 3082 | __task_rq_unlock(rq); |
3083 | } | 3083 | } |
3084 | #endif | 3084 | #endif |
3085 | 3085 | ||
3086 | void set_user_nice(struct task_struct *p, long nice) | 3086 | void set_user_nice(struct task_struct *p, long nice) |
3087 | { | 3087 | { |
3088 | int old_prio, delta, queued; | 3088 | int old_prio, delta, queued; |
3089 | unsigned long flags; | 3089 | unsigned long flags; |
3090 | struct rq *rq; | 3090 | struct rq *rq; |
3091 | 3091 | ||
3092 | if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) | 3092 | if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) |
3093 | return; | 3093 | return; |
3094 | /* | 3094 | /* |
3095 | * We have to be careful, if called from sys_setpriority(), | 3095 | * We have to be careful, if called from sys_setpriority(), |
3096 | * the task might be in the middle of scheduling on another CPU. | 3096 | * the task might be in the middle of scheduling on another CPU. |
3097 | */ | 3097 | */ |
3098 | rq = task_rq_lock(p, &flags); | 3098 | rq = task_rq_lock(p, &flags); |
3099 | /* | 3099 | /* |
3100 | * The RT priorities are set via sched_setscheduler(), but we still | 3100 | * The RT priorities are set via sched_setscheduler(), but we still |
3101 | * allow the 'normal' nice value to be set - but as expected | 3101 | * allow the 'normal' nice value to be set - but as expected |
3102 | * it wont have any effect on scheduling until the task is | 3102 | * it wont have any effect on scheduling until the task is |
3103 | * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: | 3103 | * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: |
3104 | */ | 3104 | */ |
3105 | if (task_has_dl_policy(p) || task_has_rt_policy(p)) { | 3105 | if (task_has_dl_policy(p) || task_has_rt_policy(p)) { |
3106 | p->static_prio = NICE_TO_PRIO(nice); | 3106 | p->static_prio = NICE_TO_PRIO(nice); |
3107 | goto out_unlock; | 3107 | goto out_unlock; |
3108 | } | 3108 | } |
3109 | queued = task_on_rq_queued(p); | 3109 | queued = task_on_rq_queued(p); |
3110 | if (queued) | 3110 | if (queued) |
3111 | dequeue_task(rq, p, 0); | 3111 | dequeue_task(rq, p, 0); |
3112 | 3112 | ||
3113 | p->static_prio = NICE_TO_PRIO(nice); | 3113 | p->static_prio = NICE_TO_PRIO(nice); |
3114 | set_load_weight(p); | 3114 | set_load_weight(p); |
3115 | old_prio = p->prio; | 3115 | old_prio = p->prio; |
3116 | p->prio = effective_prio(p); | 3116 | p->prio = effective_prio(p); |
3117 | delta = p->prio - old_prio; | 3117 | delta = p->prio - old_prio; |
3118 | 3118 | ||
3119 | if (queued) { | 3119 | if (queued) { |
3120 | enqueue_task(rq, p, 0); | 3120 | enqueue_task(rq, p, 0); |
3121 | /* | 3121 | /* |
3122 | * If the task increased its priority or is running and | 3122 | * If the task increased its priority or is running and |
3123 | * lowered its priority, then reschedule its CPU: | 3123 | * lowered its priority, then reschedule its CPU: |
3124 | */ | 3124 | */ |
3125 | if (delta < 0 || (delta > 0 && task_running(rq, p))) | 3125 | if (delta < 0 || (delta > 0 && task_running(rq, p))) |
3126 | resched_curr(rq); | 3126 | resched_curr(rq); |
3127 | } | 3127 | } |
3128 | out_unlock: | 3128 | out_unlock: |
3129 | task_rq_unlock(rq, p, &flags); | 3129 | task_rq_unlock(rq, p, &flags); |
3130 | } | 3130 | } |
3131 | EXPORT_SYMBOL(set_user_nice); | 3131 | EXPORT_SYMBOL(set_user_nice); |
3132 | 3132 | ||
3133 | /* | 3133 | /* |
3134 | * can_nice - check if a task can reduce its nice value | 3134 | * can_nice - check if a task can reduce its nice value |
3135 | * @p: task | 3135 | * @p: task |
3136 | * @nice: nice value | 3136 | * @nice: nice value |
3137 | */ | 3137 | */ |
3138 | int can_nice(const struct task_struct *p, const int nice) | 3138 | int can_nice(const struct task_struct *p, const int nice) |
3139 | { | 3139 | { |
3140 | /* convert nice value [19,-20] to rlimit style value [1,40] */ | 3140 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
3141 | int nice_rlim = nice_to_rlimit(nice); | 3141 | int nice_rlim = nice_to_rlimit(nice); |
3142 | 3142 | ||
3143 | return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || | 3143 | return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || |
3144 | capable(CAP_SYS_NICE)); | 3144 | capable(CAP_SYS_NICE)); |
3145 | } | 3145 | } |
3146 | 3146 | ||
3147 | #ifdef __ARCH_WANT_SYS_NICE | 3147 | #ifdef __ARCH_WANT_SYS_NICE |
3148 | 3148 | ||
3149 | /* | 3149 | /* |
3150 | * sys_nice - change the priority of the current process. | 3150 | * sys_nice - change the priority of the current process. |
3151 | * @increment: priority increment | 3151 | * @increment: priority increment |
3152 | * | 3152 | * |
3153 | * sys_setpriority is a more generic, but much slower function that | 3153 | * sys_setpriority is a more generic, but much slower function that |
3154 | * does similar things. | 3154 | * does similar things. |
3155 | */ | 3155 | */ |
3156 | SYSCALL_DEFINE1(nice, int, increment) | 3156 | SYSCALL_DEFINE1(nice, int, increment) |
3157 | { | 3157 | { |
3158 | long nice, retval; | 3158 | long nice, retval; |
3159 | 3159 | ||
3160 | /* | 3160 | /* |
3161 | * Setpriority might change our priority at the same moment. | 3161 | * Setpriority might change our priority at the same moment. |
3162 | * We don't have to worry. Conceptually one call occurs first | 3162 | * We don't have to worry. Conceptually one call occurs first |
3163 | * and we have a single winner. | 3163 | * and we have a single winner. |
3164 | */ | 3164 | */ |
3165 | increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH); | 3165 | increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH); |
3166 | nice = task_nice(current) + increment; | 3166 | nice = task_nice(current) + increment; |
3167 | 3167 | ||
3168 | nice = clamp_val(nice, MIN_NICE, MAX_NICE); | 3168 | nice = clamp_val(nice, MIN_NICE, MAX_NICE); |
3169 | if (increment < 0 && !can_nice(current, nice)) | 3169 | if (increment < 0 && !can_nice(current, nice)) |
3170 | return -EPERM; | 3170 | return -EPERM; |
3171 | 3171 | ||
3172 | retval = security_task_setnice(current, nice); | 3172 | retval = security_task_setnice(current, nice); |
3173 | if (retval) | 3173 | if (retval) |
3174 | return retval; | 3174 | return retval; |
3175 | 3175 | ||
3176 | set_user_nice(current, nice); | 3176 | set_user_nice(current, nice); |
3177 | return 0; | 3177 | return 0; |
3178 | } | 3178 | } |
3179 | 3179 | ||
3180 | #endif | 3180 | #endif |
3181 | 3181 | ||
3182 | /** | 3182 | /** |
3183 | * task_prio - return the priority value of a given task. | 3183 | * task_prio - return the priority value of a given task. |
3184 | * @p: the task in question. | 3184 | * @p: the task in question. |
3185 | * | 3185 | * |
3186 | * Return: The priority value as seen by users in /proc. | 3186 | * Return: The priority value as seen by users in /proc. |
3187 | * RT tasks are offset by -200. Normal tasks are centered | 3187 | * RT tasks are offset by -200. Normal tasks are centered |
3188 | * around 0, value goes from -16 to +15. | 3188 | * around 0, value goes from -16 to +15. |
3189 | */ | 3189 | */ |
3190 | int task_prio(const struct task_struct *p) | 3190 | int task_prio(const struct task_struct *p) |
3191 | { | 3191 | { |
3192 | return p->prio - MAX_RT_PRIO; | 3192 | return p->prio - MAX_RT_PRIO; |
3193 | } | 3193 | } |
3194 | 3194 | ||
3195 | /** | 3195 | /** |
3196 | * idle_cpu - is a given cpu idle currently? | 3196 | * idle_cpu - is a given cpu idle currently? |
3197 | * @cpu: the processor in question. | 3197 | * @cpu: the processor in question. |
3198 | * | 3198 | * |
3199 | * Return: 1 if the CPU is currently idle. 0 otherwise. | 3199 | * Return: 1 if the CPU is currently idle. 0 otherwise. |
3200 | */ | 3200 | */ |
3201 | int idle_cpu(int cpu) | 3201 | int idle_cpu(int cpu) |
3202 | { | 3202 | { |
3203 | struct rq *rq = cpu_rq(cpu); | 3203 | struct rq *rq = cpu_rq(cpu); |
3204 | 3204 | ||
3205 | if (rq->curr != rq->idle) | 3205 | if (rq->curr != rq->idle) |
3206 | return 0; | 3206 | return 0; |
3207 | 3207 | ||
3208 | if (rq->nr_running) | 3208 | if (rq->nr_running) |
3209 | return 0; | 3209 | return 0; |
3210 | 3210 | ||
3211 | #ifdef CONFIG_SMP | 3211 | #ifdef CONFIG_SMP |
3212 | if (!llist_empty(&rq->wake_list)) | 3212 | if (!llist_empty(&rq->wake_list)) |
3213 | return 0; | 3213 | return 0; |
3214 | #endif | 3214 | #endif |
3215 | 3215 | ||
3216 | return 1; | 3216 | return 1; |
3217 | } | 3217 | } |
3218 | 3218 | ||
3219 | /** | 3219 | /** |
3220 | * idle_task - return the idle task for a given cpu. | 3220 | * idle_task - return the idle task for a given cpu. |
3221 | * @cpu: the processor in question. | 3221 | * @cpu: the processor in question. |
3222 | * | 3222 | * |
3223 | * Return: The idle task for the cpu @cpu. | 3223 | * Return: The idle task for the cpu @cpu. |
3224 | */ | 3224 | */ |
3225 | struct task_struct *idle_task(int cpu) | 3225 | struct task_struct *idle_task(int cpu) |
3226 | { | 3226 | { |
3227 | return cpu_rq(cpu)->idle; | 3227 | return cpu_rq(cpu)->idle; |
3228 | } | 3228 | } |
3229 | 3229 | ||
3230 | /** | 3230 | /** |
3231 | * find_process_by_pid - find a process with a matching PID value. | 3231 | * find_process_by_pid - find a process with a matching PID value. |
3232 | * @pid: the pid in question. | 3232 | * @pid: the pid in question. |
3233 | * | 3233 | * |
3234 | * The task of @pid, if found. %NULL otherwise. | 3234 | * The task of @pid, if found. %NULL otherwise. |
3235 | */ | 3235 | */ |
3236 | static struct task_struct *find_process_by_pid(pid_t pid) | 3236 | static struct task_struct *find_process_by_pid(pid_t pid) |
3237 | { | 3237 | { |
3238 | return pid ? find_task_by_vpid(pid) : current; | 3238 | return pid ? find_task_by_vpid(pid) : current; |
3239 | } | 3239 | } |
3240 | 3240 | ||
3241 | /* | 3241 | /* |
3242 | * This function initializes the sched_dl_entity of a newly becoming | 3242 | * This function initializes the sched_dl_entity of a newly becoming |
3243 | * SCHED_DEADLINE task. | 3243 | * SCHED_DEADLINE task. |
3244 | * | 3244 | * |
3245 | * Only the static values are considered here, the actual runtime and the | 3245 | * Only the static values are considered here, the actual runtime and the |
3246 | * absolute deadline will be properly calculated when the task is enqueued | 3246 | * absolute deadline will be properly calculated when the task is enqueued |
3247 | * for the first time with its new policy. | 3247 | * for the first time with its new policy. |
3248 | */ | 3248 | */ |
3249 | static void | 3249 | static void |
3250 | __setparam_dl(struct task_struct *p, const struct sched_attr *attr) | 3250 | __setparam_dl(struct task_struct *p, const struct sched_attr *attr) |
3251 | { | 3251 | { |
3252 | struct sched_dl_entity *dl_se = &p->dl; | 3252 | struct sched_dl_entity *dl_se = &p->dl; |
3253 | 3253 | ||
3254 | init_dl_task_timer(dl_se); | 3254 | init_dl_task_timer(dl_se); |
3255 | dl_se->dl_runtime = attr->sched_runtime; | 3255 | dl_se->dl_runtime = attr->sched_runtime; |
3256 | dl_se->dl_deadline = attr->sched_deadline; | 3256 | dl_se->dl_deadline = attr->sched_deadline; |
3257 | dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; | 3257 | dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; |
3258 | dl_se->flags = attr->sched_flags; | 3258 | dl_se->flags = attr->sched_flags; |
3259 | dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); | 3259 | dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); |
3260 | dl_se->dl_throttled = 0; | 3260 | dl_se->dl_throttled = 0; |
3261 | dl_se->dl_new = 1; | 3261 | dl_se->dl_new = 1; |
3262 | dl_se->dl_yielded = 0; | 3262 | dl_se->dl_yielded = 0; |
3263 | } | 3263 | } |
3264 | 3264 | ||
3265 | /* | 3265 | /* |
3266 | * sched_setparam() passes in -1 for its policy, to let the functions | 3266 | * sched_setparam() passes in -1 for its policy, to let the functions |
3267 | * it calls know not to change it. | 3267 | * it calls know not to change it. |
3268 | */ | 3268 | */ |
3269 | #define SETPARAM_POLICY -1 | 3269 | #define SETPARAM_POLICY -1 |
3270 | 3270 | ||
3271 | static void __setscheduler_params(struct task_struct *p, | 3271 | static void __setscheduler_params(struct task_struct *p, |
3272 | const struct sched_attr *attr) | 3272 | const struct sched_attr *attr) |
3273 | { | 3273 | { |
3274 | int policy = attr->sched_policy; | 3274 | int policy = attr->sched_policy; |
3275 | 3275 | ||
3276 | if (policy == SETPARAM_POLICY) | 3276 | if (policy == SETPARAM_POLICY) |
3277 | policy = p->policy; | 3277 | policy = p->policy; |
3278 | 3278 | ||
3279 | p->policy = policy; | 3279 | p->policy = policy; |
3280 | 3280 | ||
3281 | if (dl_policy(policy)) | 3281 | if (dl_policy(policy)) |
3282 | __setparam_dl(p, attr); | 3282 | __setparam_dl(p, attr); |
3283 | else if (fair_policy(policy)) | 3283 | else if (fair_policy(policy)) |
3284 | p->static_prio = NICE_TO_PRIO(attr->sched_nice); | 3284 | p->static_prio = NICE_TO_PRIO(attr->sched_nice); |
3285 | 3285 | ||
3286 | /* | 3286 | /* |
3287 | * __sched_setscheduler() ensures attr->sched_priority == 0 when | 3287 | * __sched_setscheduler() ensures attr->sched_priority == 0 when |
3288 | * !rt_policy. Always setting this ensures that things like | 3288 | * !rt_policy. Always setting this ensures that things like |
3289 | * getparam()/getattr() don't report silly values for !rt tasks. | 3289 | * getparam()/getattr() don't report silly values for !rt tasks. |
3290 | */ | 3290 | */ |
3291 | p->rt_priority = attr->sched_priority; | 3291 | p->rt_priority = attr->sched_priority; |
3292 | p->normal_prio = normal_prio(p); | 3292 | p->normal_prio = normal_prio(p); |
3293 | set_load_weight(p); | 3293 | set_load_weight(p); |
3294 | } | 3294 | } |
3295 | 3295 | ||
3296 | /* Actually do priority change: must hold pi & rq lock. */ | 3296 | /* Actually do priority change: must hold pi & rq lock. */ |
3297 | static void __setscheduler(struct rq *rq, struct task_struct *p, | 3297 | static void __setscheduler(struct rq *rq, struct task_struct *p, |
3298 | const struct sched_attr *attr) | 3298 | const struct sched_attr *attr) |
3299 | { | 3299 | { |
3300 | __setscheduler_params(p, attr); | 3300 | __setscheduler_params(p, attr); |
3301 | 3301 | ||
3302 | /* | 3302 | /* |
3303 | * If we get here, there was no pi waiters boosting the | 3303 | * If we get here, there was no pi waiters boosting the |
3304 | * task. It is safe to use the normal prio. | 3304 | * task. It is safe to use the normal prio. |
3305 | */ | 3305 | */ |
3306 | p->prio = normal_prio(p); | 3306 | p->prio = normal_prio(p); |
3307 | 3307 | ||
3308 | if (dl_prio(p->prio)) | 3308 | if (dl_prio(p->prio)) |
3309 | p->sched_class = &dl_sched_class; | 3309 | p->sched_class = &dl_sched_class; |
3310 | else if (rt_prio(p->prio)) | 3310 | else if (rt_prio(p->prio)) |
3311 | p->sched_class = &rt_sched_class; | 3311 | p->sched_class = &rt_sched_class; |
3312 | else | 3312 | else |
3313 | p->sched_class = &fair_sched_class; | 3313 | p->sched_class = &fair_sched_class; |
3314 | } | 3314 | } |
3315 | 3315 | ||
3316 | static void | 3316 | static void |
3317 | __getparam_dl(struct task_struct *p, struct sched_attr *attr) | 3317 | __getparam_dl(struct task_struct *p, struct sched_attr *attr) |
3318 | { | 3318 | { |
3319 | struct sched_dl_entity *dl_se = &p->dl; | 3319 | struct sched_dl_entity *dl_se = &p->dl; |
3320 | 3320 | ||
3321 | attr->sched_priority = p->rt_priority; | 3321 | attr->sched_priority = p->rt_priority; |
3322 | attr->sched_runtime = dl_se->dl_runtime; | 3322 | attr->sched_runtime = dl_se->dl_runtime; |
3323 | attr->sched_deadline = dl_se->dl_deadline; | 3323 | attr->sched_deadline = dl_se->dl_deadline; |
3324 | attr->sched_period = dl_se->dl_period; | 3324 | attr->sched_period = dl_se->dl_period; |
3325 | attr->sched_flags = dl_se->flags; | 3325 | attr->sched_flags = dl_se->flags; |
3326 | } | 3326 | } |
3327 | 3327 | ||
3328 | /* | 3328 | /* |
3329 | * This function validates the new parameters of a -deadline task. | 3329 | * This function validates the new parameters of a -deadline task. |
3330 | * We ask for the deadline not being zero, and greater or equal | 3330 | * We ask for the deadline not being zero, and greater or equal |
3331 | * than the runtime, as well as the period of being zero or | 3331 | * than the runtime, as well as the period of being zero or |
3332 | * greater than deadline. Furthermore, we have to be sure that | 3332 | * greater than deadline. Furthermore, we have to be sure that |
3333 | * user parameters are above the internal resolution of 1us (we | 3333 | * user parameters are above the internal resolution of 1us (we |
3334 | * check sched_runtime only since it is always the smaller one) and | 3334 | * check sched_runtime only since it is always the smaller one) and |
3335 | * below 2^63 ns (we have to check both sched_deadline and | 3335 | * below 2^63 ns (we have to check both sched_deadline and |
3336 | * sched_period, as the latter can be zero). | 3336 | * sched_period, as the latter can be zero). |
3337 | */ | 3337 | */ |
3338 | static bool | 3338 | static bool |
3339 | __checkparam_dl(const struct sched_attr *attr) | 3339 | __checkparam_dl(const struct sched_attr *attr) |
3340 | { | 3340 | { |
3341 | /* deadline != 0 */ | 3341 | /* deadline != 0 */ |
3342 | if (attr->sched_deadline == 0) | 3342 | if (attr->sched_deadline == 0) |
3343 | return false; | 3343 | return false; |
3344 | 3344 | ||
3345 | /* | 3345 | /* |
3346 | * Since we truncate DL_SCALE bits, make sure we're at least | 3346 | * Since we truncate DL_SCALE bits, make sure we're at least |
3347 | * that big. | 3347 | * that big. |
3348 | */ | 3348 | */ |
3349 | if (attr->sched_runtime < (1ULL << DL_SCALE)) | 3349 | if (attr->sched_runtime < (1ULL << DL_SCALE)) |
3350 | return false; | 3350 | return false; |
3351 | 3351 | ||
3352 | /* | 3352 | /* |
3353 | * Since we use the MSB for wrap-around and sign issues, make | 3353 | * Since we use the MSB for wrap-around and sign issues, make |
3354 | * sure it's not set (mind that period can be equal to zero). | 3354 | * sure it's not set (mind that period can be equal to zero). |
3355 | */ | 3355 | */ |
3356 | if (attr->sched_deadline & (1ULL << 63) || | 3356 | if (attr->sched_deadline & (1ULL << 63) || |
3357 | attr->sched_period & (1ULL << 63)) | 3357 | attr->sched_period & (1ULL << 63)) |
3358 | return false; | 3358 | return false; |
3359 | 3359 | ||
3360 | /* runtime <= deadline <= period (if period != 0) */ | 3360 | /* runtime <= deadline <= period (if period != 0) */ |
3361 | if ((attr->sched_period != 0 && | 3361 | if ((attr->sched_period != 0 && |
3362 | attr->sched_period < attr->sched_deadline) || | 3362 | attr->sched_period < attr->sched_deadline) || |
3363 | attr->sched_deadline < attr->sched_runtime) | 3363 | attr->sched_deadline < attr->sched_runtime) |
3364 | return false; | 3364 | return false; |
3365 | 3365 | ||
3366 | return true; | 3366 | return true; |
3367 | } | 3367 | } |
3368 | 3368 | ||
3369 | /* | 3369 | /* |
3370 | * check the target process has a UID that matches the current process's | 3370 | * check the target process has a UID that matches the current process's |
3371 | */ | 3371 | */ |
3372 | static bool check_same_owner(struct task_struct *p) | 3372 | static bool check_same_owner(struct task_struct *p) |
3373 | { | 3373 | { |
3374 | const struct cred *cred = current_cred(), *pcred; | 3374 | const struct cred *cred = current_cred(), *pcred; |
3375 | bool match; | 3375 | bool match; |
3376 | 3376 | ||
3377 | rcu_read_lock(); | 3377 | rcu_read_lock(); |
3378 | pcred = __task_cred(p); | 3378 | pcred = __task_cred(p); |
3379 | match = (uid_eq(cred->euid, pcred->euid) || | 3379 | match = (uid_eq(cred->euid, pcred->euid) || |
3380 | uid_eq(cred->euid, pcred->uid)); | 3380 | uid_eq(cred->euid, pcred->uid)); |
3381 | rcu_read_unlock(); | 3381 | rcu_read_unlock(); |
3382 | return match; | 3382 | return match; |
3383 | } | 3383 | } |
3384 | 3384 | ||
3385 | static int __sched_setscheduler(struct task_struct *p, | 3385 | static int __sched_setscheduler(struct task_struct *p, |
3386 | const struct sched_attr *attr, | 3386 | const struct sched_attr *attr, |
3387 | bool user) | 3387 | bool user) |
3388 | { | 3388 | { |
3389 | int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 : | 3389 | int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 : |
3390 | MAX_RT_PRIO - 1 - attr->sched_priority; | 3390 | MAX_RT_PRIO - 1 - attr->sched_priority; |
3391 | int retval, oldprio, oldpolicy = -1, queued, running; | 3391 | int retval, oldprio, oldpolicy = -1, queued, running; |
3392 | int policy = attr->sched_policy; | 3392 | int policy = attr->sched_policy; |
3393 | unsigned long flags; | 3393 | unsigned long flags; |
3394 | const struct sched_class *prev_class; | 3394 | const struct sched_class *prev_class; |
3395 | struct rq *rq; | 3395 | struct rq *rq; |
3396 | int reset_on_fork; | 3396 | int reset_on_fork; |
3397 | 3397 | ||
3398 | /* may grab non-irq protected spin_locks */ | 3398 | /* may grab non-irq protected spin_locks */ |
3399 | BUG_ON(in_interrupt()); | 3399 | BUG_ON(in_interrupt()); |
3400 | recheck: | 3400 | recheck: |
3401 | /* double check policy once rq lock held */ | 3401 | /* double check policy once rq lock held */ |
3402 | if (policy < 0) { | 3402 | if (policy < 0) { |
3403 | reset_on_fork = p->sched_reset_on_fork; | 3403 | reset_on_fork = p->sched_reset_on_fork; |
3404 | policy = oldpolicy = p->policy; | 3404 | policy = oldpolicy = p->policy; |
3405 | } else { | 3405 | } else { |
3406 | reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); | 3406 | reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); |
3407 | 3407 | ||
3408 | if (policy != SCHED_DEADLINE && | 3408 | if (policy != SCHED_DEADLINE && |
3409 | policy != SCHED_FIFO && policy != SCHED_RR && | 3409 | policy != SCHED_FIFO && policy != SCHED_RR && |
3410 | policy != SCHED_NORMAL && policy != SCHED_BATCH && | 3410 | policy != SCHED_NORMAL && policy != SCHED_BATCH && |
3411 | policy != SCHED_IDLE) | 3411 | policy != SCHED_IDLE) |
3412 | return -EINVAL; | 3412 | return -EINVAL; |
3413 | } | 3413 | } |
3414 | 3414 | ||
3415 | if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) | 3415 | if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) |
3416 | return -EINVAL; | 3416 | return -EINVAL; |
3417 | 3417 | ||
3418 | /* | 3418 | /* |
3419 | * Valid priorities for SCHED_FIFO and SCHED_RR are | 3419 | * Valid priorities for SCHED_FIFO and SCHED_RR are |
3420 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, | 3420 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, |
3421 | * SCHED_BATCH and SCHED_IDLE is 0. | 3421 | * SCHED_BATCH and SCHED_IDLE is 0. |
3422 | */ | 3422 | */ |
3423 | if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || | 3423 | if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || |
3424 | (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) | 3424 | (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) |
3425 | return -EINVAL; | 3425 | return -EINVAL; |
3426 | if ((dl_policy(policy) && !__checkparam_dl(attr)) || | 3426 | if ((dl_policy(policy) && !__checkparam_dl(attr)) || |
3427 | (rt_policy(policy) != (attr->sched_priority != 0))) | 3427 | (rt_policy(policy) != (attr->sched_priority != 0))) |
3428 | return -EINVAL; | 3428 | return -EINVAL; |
3429 | 3429 | ||
3430 | /* | 3430 | /* |
3431 | * Allow unprivileged RT tasks to decrease priority: | 3431 | * Allow unprivileged RT tasks to decrease priority: |
3432 | */ | 3432 | */ |
3433 | if (user && !capable(CAP_SYS_NICE)) { | 3433 | if (user && !capable(CAP_SYS_NICE)) { |
3434 | if (fair_policy(policy)) { | 3434 | if (fair_policy(policy)) { |
3435 | if (attr->sched_nice < task_nice(p) && | 3435 | if (attr->sched_nice < task_nice(p) && |
3436 | !can_nice(p, attr->sched_nice)) | 3436 | !can_nice(p, attr->sched_nice)) |
3437 | return -EPERM; | 3437 | return -EPERM; |
3438 | } | 3438 | } |
3439 | 3439 | ||
3440 | if (rt_policy(policy)) { | 3440 | if (rt_policy(policy)) { |
3441 | unsigned long rlim_rtprio = | 3441 | unsigned long rlim_rtprio = |
3442 | task_rlimit(p, RLIMIT_RTPRIO); | 3442 | task_rlimit(p, RLIMIT_RTPRIO); |
3443 | 3443 | ||
3444 | /* can't set/change the rt policy */ | 3444 | /* can't set/change the rt policy */ |
3445 | if (policy != p->policy && !rlim_rtprio) | 3445 | if (policy != p->policy && !rlim_rtprio) |
3446 | return -EPERM; | 3446 | return -EPERM; |
3447 | 3447 | ||
3448 | /* can't increase priority */ | 3448 | /* can't increase priority */ |
3449 | if (attr->sched_priority > p->rt_priority && | 3449 | if (attr->sched_priority > p->rt_priority && |
3450 | attr->sched_priority > rlim_rtprio) | 3450 | attr->sched_priority > rlim_rtprio) |
3451 | return -EPERM; | 3451 | return -EPERM; |
3452 | } | 3452 | } |
3453 | 3453 | ||
3454 | /* | 3454 | /* |
3455 | * Can't set/change SCHED_DEADLINE policy at all for now | 3455 | * Can't set/change SCHED_DEADLINE policy at all for now |
3456 | * (safest behavior); in the future we would like to allow | 3456 | * (safest behavior); in the future we would like to allow |
3457 | * unprivileged DL tasks to increase their relative deadline | 3457 | * unprivileged DL tasks to increase their relative deadline |
3458 | * or reduce their runtime (both ways reducing utilization) | 3458 | * or reduce their runtime (both ways reducing utilization) |
3459 | */ | 3459 | */ |
3460 | if (dl_policy(policy)) | 3460 | if (dl_policy(policy)) |
3461 | return -EPERM; | 3461 | return -EPERM; |
3462 | 3462 | ||
3463 | /* | 3463 | /* |
3464 | * Treat SCHED_IDLE as nice 20. Only allow a switch to | 3464 | * Treat SCHED_IDLE as nice 20. Only allow a switch to |
3465 | * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. | 3465 | * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. |
3466 | */ | 3466 | */ |
3467 | if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { | 3467 | if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { |
3468 | if (!can_nice(p, task_nice(p))) | 3468 | if (!can_nice(p, task_nice(p))) |
3469 | return -EPERM; | 3469 | return -EPERM; |
3470 | } | 3470 | } |
3471 | 3471 | ||
3472 | /* can't change other user's priorities */ | 3472 | /* can't change other user's priorities */ |
3473 | if (!check_same_owner(p)) | 3473 | if (!check_same_owner(p)) |
3474 | return -EPERM; | 3474 | return -EPERM; |
3475 | 3475 | ||
3476 | /* Normal users shall not reset the sched_reset_on_fork flag */ | 3476 | /* Normal users shall not reset the sched_reset_on_fork flag */ |
3477 | if (p->sched_reset_on_fork && !reset_on_fork) | 3477 | if (p->sched_reset_on_fork && !reset_on_fork) |
3478 | return -EPERM; | 3478 | return -EPERM; |
3479 | } | 3479 | } |
3480 | 3480 | ||
3481 | if (user) { | 3481 | if (user) { |
3482 | retval = security_task_setscheduler(p); | 3482 | retval = security_task_setscheduler(p); |
3483 | if (retval) | 3483 | if (retval) |
3484 | return retval; | 3484 | return retval; |
3485 | } | 3485 | } |
3486 | 3486 | ||
3487 | /* | 3487 | /* |
3488 | * make sure no PI-waiters arrive (or leave) while we are | 3488 | * make sure no PI-waiters arrive (or leave) while we are |
3489 | * changing the priority of the task: | 3489 | * changing the priority of the task: |
3490 | * | 3490 | * |
3491 | * To be able to change p->policy safely, the appropriate | 3491 | * To be able to change p->policy safely, the appropriate |
3492 | * runqueue lock must be held. | 3492 | * runqueue lock must be held. |
3493 | */ | 3493 | */ |
3494 | rq = task_rq_lock(p, &flags); | 3494 | rq = task_rq_lock(p, &flags); |
3495 | 3495 | ||
3496 | /* | 3496 | /* |
3497 | * Changing the policy of the stop threads its a very bad idea | 3497 | * Changing the policy of the stop threads its a very bad idea |
3498 | */ | 3498 | */ |
3499 | if (p == rq->stop) { | 3499 | if (p == rq->stop) { |
3500 | task_rq_unlock(rq, p, &flags); | 3500 | task_rq_unlock(rq, p, &flags); |
3501 | return -EINVAL; | 3501 | return -EINVAL; |
3502 | } | 3502 | } |
3503 | 3503 | ||
3504 | /* | 3504 | /* |
3505 | * If not changing anything there's no need to proceed further, | 3505 | * If not changing anything there's no need to proceed further, |
3506 | * but store a possible modification of reset_on_fork. | 3506 | * but store a possible modification of reset_on_fork. |
3507 | */ | 3507 | */ |
3508 | if (unlikely(policy == p->policy)) { | 3508 | if (unlikely(policy == p->policy)) { |
3509 | if (fair_policy(policy) && attr->sched_nice != task_nice(p)) | 3509 | if (fair_policy(policy) && attr->sched_nice != task_nice(p)) |
3510 | goto change; | 3510 | goto change; |
3511 | if (rt_policy(policy) && attr->sched_priority != p->rt_priority) | 3511 | if (rt_policy(policy) && attr->sched_priority != p->rt_priority) |
3512 | goto change; | 3512 | goto change; |
3513 | if (dl_policy(policy)) | 3513 | if (dl_policy(policy)) |
3514 | goto change; | 3514 | goto change; |
3515 | 3515 | ||
3516 | p->sched_reset_on_fork = reset_on_fork; | 3516 | p->sched_reset_on_fork = reset_on_fork; |
3517 | task_rq_unlock(rq, p, &flags); | 3517 | task_rq_unlock(rq, p, &flags); |
3518 | return 0; | 3518 | return 0; |
3519 | } | 3519 | } |
3520 | change: | 3520 | change: |
3521 | 3521 | ||
3522 | if (user) { | 3522 | if (user) { |
3523 | #ifdef CONFIG_RT_GROUP_SCHED | 3523 | #ifdef CONFIG_RT_GROUP_SCHED |
3524 | /* | 3524 | /* |
3525 | * Do not allow realtime tasks into groups that have no runtime | 3525 | * Do not allow realtime tasks into groups that have no runtime |
3526 | * assigned. | 3526 | * assigned. |
3527 | */ | 3527 | */ |
3528 | if (rt_bandwidth_enabled() && rt_policy(policy) && | 3528 | if (rt_bandwidth_enabled() && rt_policy(policy) && |
3529 | task_group(p)->rt_bandwidth.rt_runtime == 0 && | 3529 | task_group(p)->rt_bandwidth.rt_runtime == 0 && |
3530 | !task_group_is_autogroup(task_group(p))) { | 3530 | !task_group_is_autogroup(task_group(p))) { |
3531 | task_rq_unlock(rq, p, &flags); | 3531 | task_rq_unlock(rq, p, &flags); |
3532 | return -EPERM; | 3532 | return -EPERM; |
3533 | } | 3533 | } |
3534 | #endif | 3534 | #endif |
3535 | #ifdef CONFIG_SMP | 3535 | #ifdef CONFIG_SMP |
3536 | if (dl_bandwidth_enabled() && dl_policy(policy)) { | 3536 | if (dl_bandwidth_enabled() && dl_policy(policy)) { |
3537 | cpumask_t *span = rq->rd->span; | 3537 | cpumask_t *span = rq->rd->span; |
3538 | 3538 | ||
3539 | /* | 3539 | /* |
3540 | * Don't allow tasks with an affinity mask smaller than | 3540 | * Don't allow tasks with an affinity mask smaller than |
3541 | * the entire root_domain to become SCHED_DEADLINE. We | 3541 | * the entire root_domain to become SCHED_DEADLINE. We |
3542 | * will also fail if there's no bandwidth available. | 3542 | * will also fail if there's no bandwidth available. |
3543 | */ | 3543 | */ |
3544 | if (!cpumask_subset(span, &p->cpus_allowed) || | 3544 | if (!cpumask_subset(span, &p->cpus_allowed) || |
3545 | rq->rd->dl_bw.bw == 0) { | 3545 | rq->rd->dl_bw.bw == 0) { |
3546 | task_rq_unlock(rq, p, &flags); | 3546 | task_rq_unlock(rq, p, &flags); |
3547 | return -EPERM; | 3547 | return -EPERM; |
3548 | } | 3548 | } |
3549 | } | 3549 | } |
3550 | #endif | 3550 | #endif |
3551 | } | 3551 | } |
3552 | 3552 | ||
3553 | /* recheck policy now with rq lock held */ | 3553 | /* recheck policy now with rq lock held */ |
3554 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { | 3554 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { |
3555 | policy = oldpolicy = -1; | 3555 | policy = oldpolicy = -1; |
3556 | task_rq_unlock(rq, p, &flags); | 3556 | task_rq_unlock(rq, p, &flags); |
3557 | goto recheck; | 3557 | goto recheck; |
3558 | } | 3558 | } |
3559 | 3559 | ||
3560 | /* | 3560 | /* |
3561 | * If setscheduling to SCHED_DEADLINE (or changing the parameters | 3561 | * If setscheduling to SCHED_DEADLINE (or changing the parameters |
3562 | * of a SCHED_DEADLINE task) we need to check if enough bandwidth | 3562 | * of a SCHED_DEADLINE task) we need to check if enough bandwidth |
3563 | * is available. | 3563 | * is available. |
3564 | */ | 3564 | */ |
3565 | if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { | 3565 | if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { |
3566 | task_rq_unlock(rq, p, &flags); | 3566 | task_rq_unlock(rq, p, &flags); |
3567 | return -EBUSY; | 3567 | return -EBUSY; |
3568 | } | 3568 | } |
3569 | 3569 | ||
3570 | p->sched_reset_on_fork = reset_on_fork; | 3570 | p->sched_reset_on_fork = reset_on_fork; |
3571 | oldprio = p->prio; | 3571 | oldprio = p->prio; |
3572 | 3572 | ||
3573 | /* | 3573 | /* |
3574 | * Special case for priority boosted tasks. | 3574 | * Special case for priority boosted tasks. |
3575 | * | 3575 | * |
3576 | * If the new priority is lower or equal (user space view) | 3576 | * If the new priority is lower or equal (user space view) |
3577 | * than the current (boosted) priority, we just store the new | 3577 | * than the current (boosted) priority, we just store the new |
3578 | * normal parameters and do not touch the scheduler class and | 3578 | * normal parameters and do not touch the scheduler class and |
3579 | * the runqueue. This will be done when the task deboost | 3579 | * the runqueue. This will be done when the task deboost |
3580 | * itself. | 3580 | * itself. |
3581 | */ | 3581 | */ |
3582 | if (rt_mutex_check_prio(p, newprio)) { | 3582 | if (rt_mutex_check_prio(p, newprio)) { |
3583 | __setscheduler_params(p, attr); | 3583 | __setscheduler_params(p, attr); |
3584 | task_rq_unlock(rq, p, &flags); | 3584 | task_rq_unlock(rq, p, &flags); |
3585 | return 0; | 3585 | return 0; |
3586 | } | 3586 | } |
3587 | 3587 | ||
3588 | queued = task_on_rq_queued(p); | 3588 | queued = task_on_rq_queued(p); |
3589 | running = task_current(rq, p); | 3589 | running = task_current(rq, p); |
3590 | if (queued) | 3590 | if (queued) |
3591 | dequeue_task(rq, p, 0); | 3591 | dequeue_task(rq, p, 0); |
3592 | if (running) | 3592 | if (running) |
3593 | put_prev_task(rq, p); | 3593 | put_prev_task(rq, p); |
3594 | 3594 | ||
3595 | prev_class = p->sched_class; | 3595 | prev_class = p->sched_class; |
3596 | __setscheduler(rq, p, attr); | 3596 | __setscheduler(rq, p, attr); |
3597 | 3597 | ||
3598 | if (running) | 3598 | if (running) |
3599 | p->sched_class->set_curr_task(rq); | 3599 | p->sched_class->set_curr_task(rq); |
3600 | if (queued) { | 3600 | if (queued) { |
3601 | /* | 3601 | /* |
3602 | * We enqueue to tail when the priority of a task is | 3602 | * We enqueue to tail when the priority of a task is |
3603 | * increased (user space view). | 3603 | * increased (user space view). |
3604 | */ | 3604 | */ |
3605 | enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0); | 3605 | enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0); |
3606 | } | 3606 | } |
3607 | 3607 | ||
3608 | check_class_changed(rq, p, prev_class, oldprio); | 3608 | check_class_changed(rq, p, prev_class, oldprio); |
3609 | task_rq_unlock(rq, p, &flags); | 3609 | task_rq_unlock(rq, p, &flags); |
3610 | 3610 | ||
3611 | rt_mutex_adjust_pi(p); | 3611 | rt_mutex_adjust_pi(p); |
3612 | 3612 | ||
3613 | return 0; | 3613 | return 0; |
3614 | } | 3614 | } |
3615 | 3615 | ||
3616 | static int _sched_setscheduler(struct task_struct *p, int policy, | 3616 | static int _sched_setscheduler(struct task_struct *p, int policy, |
3617 | const struct sched_param *param, bool check) | 3617 | const struct sched_param *param, bool check) |
3618 | { | 3618 | { |
3619 | struct sched_attr attr = { | 3619 | struct sched_attr attr = { |
3620 | .sched_policy = policy, | 3620 | .sched_policy = policy, |
3621 | .sched_priority = param->sched_priority, | 3621 | .sched_priority = param->sched_priority, |
3622 | .sched_nice = PRIO_TO_NICE(p->static_prio), | 3622 | .sched_nice = PRIO_TO_NICE(p->static_prio), |
3623 | }; | 3623 | }; |
3624 | 3624 | ||
3625 | /* Fixup the legacy SCHED_RESET_ON_FORK hack. */ | 3625 | /* Fixup the legacy SCHED_RESET_ON_FORK hack. */ |
3626 | if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) { | 3626 | if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) { |
3627 | attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; | 3627 | attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; |
3628 | policy &= ~SCHED_RESET_ON_FORK; | 3628 | policy &= ~SCHED_RESET_ON_FORK; |
3629 | attr.sched_policy = policy; | 3629 | attr.sched_policy = policy; |
3630 | } | 3630 | } |
3631 | 3631 | ||
3632 | return __sched_setscheduler(p, &attr, check); | 3632 | return __sched_setscheduler(p, &attr, check); |
3633 | } | 3633 | } |
3634 | /** | 3634 | /** |
3635 | * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. | 3635 | * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. |
3636 | * @p: the task in question. | 3636 | * @p: the task in question. |
3637 | * @policy: new policy. | 3637 | * @policy: new policy. |
3638 | * @param: structure containing the new RT priority. | 3638 | * @param: structure containing the new RT priority. |
3639 | * | 3639 | * |
3640 | * Return: 0 on success. An error code otherwise. | 3640 | * Return: 0 on success. An error code otherwise. |
3641 | * | 3641 | * |
3642 | * NOTE that the task may be already dead. | 3642 | * NOTE that the task may be already dead. |
3643 | */ | 3643 | */ |
3644 | int sched_setscheduler(struct task_struct *p, int policy, | 3644 | int sched_setscheduler(struct task_struct *p, int policy, |
3645 | const struct sched_param *param) | 3645 | const struct sched_param *param) |
3646 | { | 3646 | { |
3647 | return _sched_setscheduler(p, policy, param, true); | 3647 | return _sched_setscheduler(p, policy, param, true); |
3648 | } | 3648 | } |
3649 | EXPORT_SYMBOL_GPL(sched_setscheduler); | 3649 | EXPORT_SYMBOL_GPL(sched_setscheduler); |
3650 | 3650 | ||
3651 | int sched_setattr(struct task_struct *p, const struct sched_attr *attr) | 3651 | int sched_setattr(struct task_struct *p, const struct sched_attr *attr) |
3652 | { | 3652 | { |
3653 | return __sched_setscheduler(p, attr, true); | 3653 | return __sched_setscheduler(p, attr, true); |
3654 | } | 3654 | } |
3655 | EXPORT_SYMBOL_GPL(sched_setattr); | 3655 | EXPORT_SYMBOL_GPL(sched_setattr); |
3656 | 3656 | ||
3657 | /** | 3657 | /** |
3658 | * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. | 3658 | * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. |
3659 | * @p: the task in question. | 3659 | * @p: the task in question. |
3660 | * @policy: new policy. | 3660 | * @policy: new policy. |
3661 | * @param: structure containing the new RT priority. | 3661 | * @param: structure containing the new RT priority. |
3662 | * | 3662 | * |
3663 | * Just like sched_setscheduler, only don't bother checking if the | 3663 | * Just like sched_setscheduler, only don't bother checking if the |
3664 | * current context has permission. For example, this is needed in | 3664 | * current context has permission. For example, this is needed in |
3665 | * stop_machine(): we create temporary high priority worker threads, | 3665 | * stop_machine(): we create temporary high priority worker threads, |
3666 | * but our caller might not have that capability. | 3666 | * but our caller might not have that capability. |
3667 | * | 3667 | * |
3668 | * Return: 0 on success. An error code otherwise. | 3668 | * Return: 0 on success. An error code otherwise. |
3669 | */ | 3669 | */ |
3670 | int sched_setscheduler_nocheck(struct task_struct *p, int policy, | 3670 | int sched_setscheduler_nocheck(struct task_struct *p, int policy, |
3671 | const struct sched_param *param) | 3671 | const struct sched_param *param) |
3672 | { | 3672 | { |
3673 | return _sched_setscheduler(p, policy, param, false); | 3673 | return _sched_setscheduler(p, policy, param, false); |
3674 | } | 3674 | } |
3675 | 3675 | ||
3676 | static int | 3676 | static int |
3677 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | 3677 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) |
3678 | { | 3678 | { |
3679 | struct sched_param lparam; | 3679 | struct sched_param lparam; |
3680 | struct task_struct *p; | 3680 | struct task_struct *p; |
3681 | int retval; | 3681 | int retval; |
3682 | 3682 | ||
3683 | if (!param || pid < 0) | 3683 | if (!param || pid < 0) |
3684 | return -EINVAL; | 3684 | return -EINVAL; |
3685 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) | 3685 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) |
3686 | return -EFAULT; | 3686 | return -EFAULT; |
3687 | 3687 | ||
3688 | rcu_read_lock(); | 3688 | rcu_read_lock(); |
3689 | retval = -ESRCH; | 3689 | retval = -ESRCH; |
3690 | p = find_process_by_pid(pid); | 3690 | p = find_process_by_pid(pid); |
3691 | if (p != NULL) | 3691 | if (p != NULL) |
3692 | retval = sched_setscheduler(p, policy, &lparam); | 3692 | retval = sched_setscheduler(p, policy, &lparam); |
3693 | rcu_read_unlock(); | 3693 | rcu_read_unlock(); |
3694 | 3694 | ||
3695 | return retval; | 3695 | return retval; |
3696 | } | 3696 | } |
3697 | 3697 | ||
3698 | /* | 3698 | /* |
3699 | * Mimics kernel/events/core.c perf_copy_attr(). | 3699 | * Mimics kernel/events/core.c perf_copy_attr(). |
3700 | */ | 3700 | */ |
3701 | static int sched_copy_attr(struct sched_attr __user *uattr, | 3701 | static int sched_copy_attr(struct sched_attr __user *uattr, |
3702 | struct sched_attr *attr) | 3702 | struct sched_attr *attr) |
3703 | { | 3703 | { |
3704 | u32 size; | 3704 | u32 size; |
3705 | int ret; | 3705 | int ret; |
3706 | 3706 | ||
3707 | if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) | 3707 | if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) |
3708 | return -EFAULT; | 3708 | return -EFAULT; |
3709 | 3709 | ||
3710 | /* | 3710 | /* |
3711 | * zero the full structure, so that a short copy will be nice. | 3711 | * zero the full structure, so that a short copy will be nice. |
3712 | */ | 3712 | */ |
3713 | memset(attr, 0, sizeof(*attr)); | 3713 | memset(attr, 0, sizeof(*attr)); |
3714 | 3714 | ||
3715 | ret = get_user(size, &uattr->size); | 3715 | ret = get_user(size, &uattr->size); |
3716 | if (ret) | 3716 | if (ret) |
3717 | return ret; | 3717 | return ret; |
3718 | 3718 | ||
3719 | if (size > PAGE_SIZE) /* silly large */ | 3719 | if (size > PAGE_SIZE) /* silly large */ |
3720 | goto err_size; | 3720 | goto err_size; |
3721 | 3721 | ||
3722 | if (!size) /* abi compat */ | 3722 | if (!size) /* abi compat */ |
3723 | size = SCHED_ATTR_SIZE_VER0; | 3723 | size = SCHED_ATTR_SIZE_VER0; |
3724 | 3724 | ||
3725 | if (size < SCHED_ATTR_SIZE_VER0) | 3725 | if (size < SCHED_ATTR_SIZE_VER0) |
3726 | goto err_size; | 3726 | goto err_size; |
3727 | 3727 | ||
3728 | /* | 3728 | /* |
3729 | * If we're handed a bigger struct than we know of, | 3729 | * If we're handed a bigger struct than we know of, |
3730 | * ensure all the unknown bits are 0 - i.e. new | 3730 | * ensure all the unknown bits are 0 - i.e. new |
3731 | * user-space does not rely on any kernel feature | 3731 | * user-space does not rely on any kernel feature |
3732 | * extensions we dont know about yet. | 3732 | * extensions we dont know about yet. |
3733 | */ | 3733 | */ |
3734 | if (size > sizeof(*attr)) { | 3734 | if (size > sizeof(*attr)) { |
3735 | unsigned char __user *addr; | 3735 | unsigned char __user *addr; |
3736 | unsigned char __user *end; | 3736 | unsigned char __user *end; |
3737 | unsigned char val; | 3737 | unsigned char val; |
3738 | 3738 | ||
3739 | addr = (void __user *)uattr + sizeof(*attr); | 3739 | addr = (void __user *)uattr + sizeof(*attr); |
3740 | end = (void __user *)uattr + size; | 3740 | end = (void __user *)uattr + size; |
3741 | 3741 | ||
3742 | for (; addr < end; addr++) { | 3742 | for (; addr < end; addr++) { |
3743 | ret = get_user(val, addr); | 3743 | ret = get_user(val, addr); |
3744 | if (ret) | 3744 | if (ret) |
3745 | return ret; | 3745 | return ret; |
3746 | if (val) | 3746 | if (val) |
3747 | goto err_size; | 3747 | goto err_size; |
3748 | } | 3748 | } |
3749 | size = sizeof(*attr); | 3749 | size = sizeof(*attr); |
3750 | } | 3750 | } |
3751 | 3751 | ||
3752 | ret = copy_from_user(attr, uattr, size); | 3752 | ret = copy_from_user(attr, uattr, size); |
3753 | if (ret) | 3753 | if (ret) |
3754 | return -EFAULT; | 3754 | return -EFAULT; |
3755 | 3755 | ||
3756 | /* | 3756 | /* |
3757 | * XXX: do we want to be lenient like existing syscalls; or do we want | 3757 | * XXX: do we want to be lenient like existing syscalls; or do we want |
3758 | * to be strict and return an error on out-of-bounds values? | 3758 | * to be strict and return an error on out-of-bounds values? |
3759 | */ | 3759 | */ |
3760 | attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); | 3760 | attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); |
3761 | 3761 | ||
3762 | return 0; | 3762 | return 0; |
3763 | 3763 | ||
3764 | err_size: | 3764 | err_size: |
3765 | put_user(sizeof(*attr), &uattr->size); | 3765 | put_user(sizeof(*attr), &uattr->size); |
3766 | return -E2BIG; | 3766 | return -E2BIG; |
3767 | } | 3767 | } |
3768 | 3768 | ||
3769 | /** | 3769 | /** |
3770 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority | 3770 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority |
3771 | * @pid: the pid in question. | 3771 | * @pid: the pid in question. |
3772 | * @policy: new policy. | 3772 | * @policy: new policy. |
3773 | * @param: structure containing the new RT priority. | 3773 | * @param: structure containing the new RT priority. |
3774 | * | 3774 | * |
3775 | * Return: 0 on success. An error code otherwise. | 3775 | * Return: 0 on success. An error code otherwise. |
3776 | */ | 3776 | */ |
3777 | SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, | 3777 | SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, |
3778 | struct sched_param __user *, param) | 3778 | struct sched_param __user *, param) |
3779 | { | 3779 | { |
3780 | /* negative values for policy are not valid */ | 3780 | /* negative values for policy are not valid */ |
3781 | if (policy < 0) | 3781 | if (policy < 0) |
3782 | return -EINVAL; | 3782 | return -EINVAL; |
3783 | 3783 | ||
3784 | return do_sched_setscheduler(pid, policy, param); | 3784 | return do_sched_setscheduler(pid, policy, param); |
3785 | } | 3785 | } |
3786 | 3786 | ||
3787 | /** | 3787 | /** |
3788 | * sys_sched_setparam - set/change the RT priority of a thread | 3788 | * sys_sched_setparam - set/change the RT priority of a thread |
3789 | * @pid: the pid in question. | 3789 | * @pid: the pid in question. |
3790 | * @param: structure containing the new RT priority. | 3790 | * @param: structure containing the new RT priority. |
3791 | * | 3791 | * |
3792 | * Return: 0 on success. An error code otherwise. | 3792 | * Return: 0 on success. An error code otherwise. |
3793 | */ | 3793 | */ |
3794 | SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) | 3794 | SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) |
3795 | { | 3795 | { |
3796 | return do_sched_setscheduler(pid, SETPARAM_POLICY, param); | 3796 | return do_sched_setscheduler(pid, SETPARAM_POLICY, param); |
3797 | } | 3797 | } |
3798 | 3798 | ||
3799 | /** | 3799 | /** |
3800 | * sys_sched_setattr - same as above, but with extended sched_attr | 3800 | * sys_sched_setattr - same as above, but with extended sched_attr |
3801 | * @pid: the pid in question. | 3801 | * @pid: the pid in question. |
3802 | * @uattr: structure containing the extended parameters. | 3802 | * @uattr: structure containing the extended parameters. |
3803 | * @flags: for future extension. | 3803 | * @flags: for future extension. |
3804 | */ | 3804 | */ |
3805 | SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, | 3805 | SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, |
3806 | unsigned int, flags) | 3806 | unsigned int, flags) |
3807 | { | 3807 | { |
3808 | struct sched_attr attr; | 3808 | struct sched_attr attr; |
3809 | struct task_struct *p; | 3809 | struct task_struct *p; |
3810 | int retval; | 3810 | int retval; |
3811 | 3811 | ||
3812 | if (!uattr || pid < 0 || flags) | 3812 | if (!uattr || pid < 0 || flags) |
3813 | return -EINVAL; | 3813 | return -EINVAL; |
3814 | 3814 | ||
3815 | retval = sched_copy_attr(uattr, &attr); | 3815 | retval = sched_copy_attr(uattr, &attr); |
3816 | if (retval) | 3816 | if (retval) |
3817 | return retval; | 3817 | return retval; |
3818 | 3818 | ||
3819 | if ((int)attr.sched_policy < 0) | 3819 | if ((int)attr.sched_policy < 0) |
3820 | return -EINVAL; | 3820 | return -EINVAL; |
3821 | 3821 | ||
3822 | rcu_read_lock(); | 3822 | rcu_read_lock(); |
3823 | retval = -ESRCH; | 3823 | retval = -ESRCH; |
3824 | p = find_process_by_pid(pid); | 3824 | p = find_process_by_pid(pid); |
3825 | if (p != NULL) | 3825 | if (p != NULL) |
3826 | retval = sched_setattr(p, &attr); | 3826 | retval = sched_setattr(p, &attr); |
3827 | rcu_read_unlock(); | 3827 | rcu_read_unlock(); |
3828 | 3828 | ||
3829 | return retval; | 3829 | return retval; |
3830 | } | 3830 | } |
3831 | 3831 | ||
3832 | /** | 3832 | /** |
3833 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread | 3833 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread |
3834 | * @pid: the pid in question. | 3834 | * @pid: the pid in question. |
3835 | * | 3835 | * |
3836 | * Return: On success, the policy of the thread. Otherwise, a negative error | 3836 | * Return: On success, the policy of the thread. Otherwise, a negative error |
3837 | * code. | 3837 | * code. |
3838 | */ | 3838 | */ |
3839 | SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) | 3839 | SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) |
3840 | { | 3840 | { |
3841 | struct task_struct *p; | 3841 | struct task_struct *p; |
3842 | int retval; | 3842 | int retval; |
3843 | 3843 | ||
3844 | if (pid < 0) | 3844 | if (pid < 0) |
3845 | return -EINVAL; | 3845 | return -EINVAL; |
3846 | 3846 | ||
3847 | retval = -ESRCH; | 3847 | retval = -ESRCH; |
3848 | rcu_read_lock(); | 3848 | rcu_read_lock(); |
3849 | p = find_process_by_pid(pid); | 3849 | p = find_process_by_pid(pid); |
3850 | if (p) { | 3850 | if (p) { |
3851 | retval = security_task_getscheduler(p); | 3851 | retval = security_task_getscheduler(p); |
3852 | if (!retval) | 3852 | if (!retval) |
3853 | retval = p->policy | 3853 | retval = p->policy |
3854 | | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); | 3854 | | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); |
3855 | } | 3855 | } |
3856 | rcu_read_unlock(); | 3856 | rcu_read_unlock(); |
3857 | return retval; | 3857 | return retval; |
3858 | } | 3858 | } |
3859 | 3859 | ||
3860 | /** | 3860 | /** |
3861 | * sys_sched_getparam - get the RT priority of a thread | 3861 | * sys_sched_getparam - get the RT priority of a thread |
3862 | * @pid: the pid in question. | 3862 | * @pid: the pid in question. |
3863 | * @param: structure containing the RT priority. | 3863 | * @param: structure containing the RT priority. |
3864 | * | 3864 | * |
3865 | * Return: On success, 0 and the RT priority is in @param. Otherwise, an error | 3865 | * Return: On success, 0 and the RT priority is in @param. Otherwise, an error |
3866 | * code. | 3866 | * code. |
3867 | */ | 3867 | */ |
3868 | SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) | 3868 | SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) |
3869 | { | 3869 | { |
3870 | struct sched_param lp = { .sched_priority = 0 }; | 3870 | struct sched_param lp = { .sched_priority = 0 }; |
3871 | struct task_struct *p; | 3871 | struct task_struct *p; |
3872 | int retval; | 3872 | int retval; |
3873 | 3873 | ||
3874 | if (!param || pid < 0) | 3874 | if (!param || pid < 0) |
3875 | return -EINVAL; | 3875 | return -EINVAL; |
3876 | 3876 | ||
3877 | rcu_read_lock(); | 3877 | rcu_read_lock(); |
3878 | p = find_process_by_pid(pid); | 3878 | p = find_process_by_pid(pid); |
3879 | retval = -ESRCH; | 3879 | retval = -ESRCH; |
3880 | if (!p) | 3880 | if (!p) |
3881 | goto out_unlock; | 3881 | goto out_unlock; |
3882 | 3882 | ||
3883 | retval = security_task_getscheduler(p); | 3883 | retval = security_task_getscheduler(p); |
3884 | if (retval) | 3884 | if (retval) |
3885 | goto out_unlock; | 3885 | goto out_unlock; |
3886 | 3886 | ||
3887 | if (task_has_rt_policy(p)) | 3887 | if (task_has_rt_policy(p)) |
3888 | lp.sched_priority = p->rt_priority; | 3888 | lp.sched_priority = p->rt_priority; |
3889 | rcu_read_unlock(); | 3889 | rcu_read_unlock(); |
3890 | 3890 | ||
3891 | /* | 3891 | /* |
3892 | * This one might sleep, we cannot do it with a spinlock held ... | 3892 | * This one might sleep, we cannot do it with a spinlock held ... |
3893 | */ | 3893 | */ |
3894 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; | 3894 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; |
3895 | 3895 | ||
3896 | return retval; | 3896 | return retval; |
3897 | 3897 | ||
3898 | out_unlock: | 3898 | out_unlock: |
3899 | rcu_read_unlock(); | 3899 | rcu_read_unlock(); |
3900 | return retval; | 3900 | return retval; |
3901 | } | 3901 | } |
3902 | 3902 | ||
3903 | static int sched_read_attr(struct sched_attr __user *uattr, | 3903 | static int sched_read_attr(struct sched_attr __user *uattr, |
3904 | struct sched_attr *attr, | 3904 | struct sched_attr *attr, |
3905 | unsigned int usize) | 3905 | unsigned int usize) |
3906 | { | 3906 | { |
3907 | int ret; | 3907 | int ret; |
3908 | 3908 | ||
3909 | if (!access_ok(VERIFY_WRITE, uattr, usize)) | 3909 | if (!access_ok(VERIFY_WRITE, uattr, usize)) |
3910 | return -EFAULT; | 3910 | return -EFAULT; |
3911 | 3911 | ||
3912 | /* | 3912 | /* |
3913 | * If we're handed a smaller struct than we know of, | 3913 | * If we're handed a smaller struct than we know of, |
3914 | * ensure all the unknown bits are 0 - i.e. old | 3914 | * ensure all the unknown bits are 0 - i.e. old |
3915 | * user-space does not get uncomplete information. | 3915 | * user-space does not get uncomplete information. |
3916 | */ | 3916 | */ |
3917 | if (usize < sizeof(*attr)) { | 3917 | if (usize < sizeof(*attr)) { |
3918 | unsigned char *addr; | 3918 | unsigned char *addr; |
3919 | unsigned char *end; | 3919 | unsigned char *end; |
3920 | 3920 | ||
3921 | addr = (void *)attr + usize; | 3921 | addr = (void *)attr + usize; |
3922 | end = (void *)attr + sizeof(*attr); | 3922 | end = (void *)attr + sizeof(*attr); |
3923 | 3923 | ||
3924 | for (; addr < end; addr++) { | 3924 | for (; addr < end; addr++) { |
3925 | if (*addr) | 3925 | if (*addr) |
3926 | return -EFBIG; | 3926 | return -EFBIG; |
3927 | } | 3927 | } |
3928 | 3928 | ||
3929 | attr->size = usize; | 3929 | attr->size = usize; |
3930 | } | 3930 | } |
3931 | 3931 | ||
3932 | ret = copy_to_user(uattr, attr, attr->size); | 3932 | ret = copy_to_user(uattr, attr, attr->size); |
3933 | if (ret) | 3933 | if (ret) |
3934 | return -EFAULT; | 3934 | return -EFAULT; |
3935 | 3935 | ||
3936 | return 0; | 3936 | return 0; |
3937 | } | 3937 | } |
3938 | 3938 | ||
3939 | /** | 3939 | /** |
3940 | * sys_sched_getattr - similar to sched_getparam, but with sched_attr | 3940 | * sys_sched_getattr - similar to sched_getparam, but with sched_attr |
3941 | * @pid: the pid in question. | 3941 | * @pid: the pid in question. |
3942 | * @uattr: structure containing the extended parameters. | 3942 | * @uattr: structure containing the extended parameters. |
3943 | * @size: sizeof(attr) for fwd/bwd comp. | 3943 | * @size: sizeof(attr) for fwd/bwd comp. |
3944 | * @flags: for future extension. | 3944 | * @flags: for future extension. |
3945 | */ | 3945 | */ |
3946 | SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, | 3946 | SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, |
3947 | unsigned int, size, unsigned int, flags) | 3947 | unsigned int, size, unsigned int, flags) |
3948 | { | 3948 | { |
3949 | struct sched_attr attr = { | 3949 | struct sched_attr attr = { |
3950 | .size = sizeof(struct sched_attr), | 3950 | .size = sizeof(struct sched_attr), |
3951 | }; | 3951 | }; |
3952 | struct task_struct *p; | 3952 | struct task_struct *p; |
3953 | int retval; | 3953 | int retval; |
3954 | 3954 | ||
3955 | if (!uattr || pid < 0 || size > PAGE_SIZE || | 3955 | if (!uattr || pid < 0 || size > PAGE_SIZE || |
3956 | size < SCHED_ATTR_SIZE_VER0 || flags) | 3956 | size < SCHED_ATTR_SIZE_VER0 || flags) |
3957 | return -EINVAL; | 3957 | return -EINVAL; |
3958 | 3958 | ||
3959 | rcu_read_lock(); | 3959 | rcu_read_lock(); |
3960 | p = find_process_by_pid(pid); | 3960 | p = find_process_by_pid(pid); |
3961 | retval = -ESRCH; | 3961 | retval = -ESRCH; |
3962 | if (!p) | 3962 | if (!p) |
3963 | goto out_unlock; | 3963 | goto out_unlock; |
3964 | 3964 | ||
3965 | retval = security_task_getscheduler(p); | 3965 | retval = security_task_getscheduler(p); |
3966 | if (retval) | 3966 | if (retval) |
3967 | goto out_unlock; | 3967 | goto out_unlock; |
3968 | 3968 | ||
3969 | attr.sched_policy = p->policy; | 3969 | attr.sched_policy = p->policy; |
3970 | if (p->sched_reset_on_fork) | 3970 | if (p->sched_reset_on_fork) |
3971 | attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; | 3971 | attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; |
3972 | if (task_has_dl_policy(p)) | 3972 | if (task_has_dl_policy(p)) |
3973 | __getparam_dl(p, &attr); | 3973 | __getparam_dl(p, &attr); |
3974 | else if (task_has_rt_policy(p)) | 3974 | else if (task_has_rt_policy(p)) |
3975 | attr.sched_priority = p->rt_priority; | 3975 | attr.sched_priority = p->rt_priority; |
3976 | else | 3976 | else |
3977 | attr.sched_nice = task_nice(p); | 3977 | attr.sched_nice = task_nice(p); |
3978 | 3978 | ||
3979 | rcu_read_unlock(); | 3979 | rcu_read_unlock(); |
3980 | 3980 | ||
3981 | retval = sched_read_attr(uattr, &attr, size); | 3981 | retval = sched_read_attr(uattr, &attr, size); |
3982 | return retval; | 3982 | return retval; |
3983 | 3983 | ||
3984 | out_unlock: | 3984 | out_unlock: |
3985 | rcu_read_unlock(); | 3985 | rcu_read_unlock(); |
3986 | return retval; | 3986 | return retval; |
3987 | } | 3987 | } |
3988 | 3988 | ||
3989 | long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) | 3989 | long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) |
3990 | { | 3990 | { |
3991 | cpumask_var_t cpus_allowed, new_mask; | 3991 | cpumask_var_t cpus_allowed, new_mask; |
3992 | struct task_struct *p; | 3992 | struct task_struct *p; |
3993 | int retval; | 3993 | int retval; |
3994 | 3994 | ||
3995 | rcu_read_lock(); | 3995 | rcu_read_lock(); |
3996 | 3996 | ||
3997 | p = find_process_by_pid(pid); | 3997 | p = find_process_by_pid(pid); |
3998 | if (!p) { | 3998 | if (!p) { |
3999 | rcu_read_unlock(); | 3999 | rcu_read_unlock(); |
4000 | return -ESRCH; | 4000 | return -ESRCH; |
4001 | } | 4001 | } |
4002 | 4002 | ||
4003 | /* Prevent p going away */ | 4003 | /* Prevent p going away */ |
4004 | get_task_struct(p); | 4004 | get_task_struct(p); |
4005 | rcu_read_unlock(); | 4005 | rcu_read_unlock(); |
4006 | 4006 | ||
4007 | if (p->flags & PF_NO_SETAFFINITY) { | 4007 | if (p->flags & PF_NO_SETAFFINITY) { |
4008 | retval = -EINVAL; | 4008 | retval = -EINVAL; |
4009 | goto out_put_task; | 4009 | goto out_put_task; |
4010 | } | 4010 | } |
4011 | if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { | 4011 | if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { |
4012 | retval = -ENOMEM; | 4012 | retval = -ENOMEM; |
4013 | goto out_put_task; | 4013 | goto out_put_task; |
4014 | } | 4014 | } |
4015 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { | 4015 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { |
4016 | retval = -ENOMEM; | 4016 | retval = -ENOMEM; |
4017 | goto out_free_cpus_allowed; | 4017 | goto out_free_cpus_allowed; |
4018 | } | 4018 | } |
4019 | retval = -EPERM; | 4019 | retval = -EPERM; |
4020 | if (!check_same_owner(p)) { | 4020 | if (!check_same_owner(p)) { |
4021 | rcu_read_lock(); | 4021 | rcu_read_lock(); |
4022 | if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { | 4022 | if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { |
4023 | rcu_read_unlock(); | 4023 | rcu_read_unlock(); |
4024 | goto out_free_new_mask; | 4024 | goto out_free_new_mask; |
4025 | } | 4025 | } |
4026 | rcu_read_unlock(); | 4026 | rcu_read_unlock(); |
4027 | } | 4027 | } |
4028 | 4028 | ||
4029 | retval = security_task_setscheduler(p); | 4029 | retval = security_task_setscheduler(p); |
4030 | if (retval) | 4030 | if (retval) |
4031 | goto out_free_new_mask; | 4031 | goto out_free_new_mask; |
4032 | 4032 | ||
4033 | 4033 | ||
4034 | cpuset_cpus_allowed(p, cpus_allowed); | 4034 | cpuset_cpus_allowed(p, cpus_allowed); |
4035 | cpumask_and(new_mask, in_mask, cpus_allowed); | 4035 | cpumask_and(new_mask, in_mask, cpus_allowed); |
4036 | 4036 | ||
4037 | /* | 4037 | /* |
4038 | * Since bandwidth control happens on root_domain basis, | 4038 | * Since bandwidth control happens on root_domain basis, |
4039 | * if admission test is enabled, we only admit -deadline | 4039 | * if admission test is enabled, we only admit -deadline |
4040 | * tasks allowed to run on all the CPUs in the task's | 4040 | * tasks allowed to run on all the CPUs in the task's |
4041 | * root_domain. | 4041 | * root_domain. |
4042 | */ | 4042 | */ |
4043 | #ifdef CONFIG_SMP | 4043 | #ifdef CONFIG_SMP |
4044 | if (task_has_dl_policy(p) && dl_bandwidth_enabled()) { | 4044 | if (task_has_dl_policy(p) && dl_bandwidth_enabled()) { |
4045 | rcu_read_lock(); | 4045 | rcu_read_lock(); |
4046 | if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) { | 4046 | if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) { |
4047 | retval = -EBUSY; | 4047 | retval = -EBUSY; |
4048 | rcu_read_unlock(); | 4048 | rcu_read_unlock(); |
4049 | goto out_free_new_mask; | 4049 | goto out_free_new_mask; |
4050 | } | 4050 | } |
4051 | rcu_read_unlock(); | 4051 | rcu_read_unlock(); |
4052 | } | 4052 | } |
4053 | #endif | 4053 | #endif |
4054 | again: | 4054 | again: |
4055 | retval = set_cpus_allowed_ptr(p, new_mask); | 4055 | retval = set_cpus_allowed_ptr(p, new_mask); |
4056 | 4056 | ||
4057 | if (!retval) { | 4057 | if (!retval) { |
4058 | cpuset_cpus_allowed(p, cpus_allowed); | 4058 | cpuset_cpus_allowed(p, cpus_allowed); |
4059 | if (!cpumask_subset(new_mask, cpus_allowed)) { | 4059 | if (!cpumask_subset(new_mask, cpus_allowed)) { |
4060 | /* | 4060 | /* |
4061 | * We must have raced with a concurrent cpuset | 4061 | * We must have raced with a concurrent cpuset |
4062 | * update. Just reset the cpus_allowed to the | 4062 | * update. Just reset the cpus_allowed to the |
4063 | * cpuset's cpus_allowed | 4063 | * cpuset's cpus_allowed |
4064 | */ | 4064 | */ |
4065 | cpumask_copy(new_mask, cpus_allowed); | 4065 | cpumask_copy(new_mask, cpus_allowed); |
4066 | goto again; | 4066 | goto again; |
4067 | } | 4067 | } |
4068 | } | 4068 | } |
4069 | out_free_new_mask: | 4069 | out_free_new_mask: |
4070 | free_cpumask_var(new_mask); | 4070 | free_cpumask_var(new_mask); |
4071 | out_free_cpus_allowed: | 4071 | out_free_cpus_allowed: |
4072 | free_cpumask_var(cpus_allowed); | 4072 | free_cpumask_var(cpus_allowed); |
4073 | out_put_task: | 4073 | out_put_task: |
4074 | put_task_struct(p); | 4074 | put_task_struct(p); |
4075 | return retval; | 4075 | return retval; |
4076 | } | 4076 | } |
4077 | 4077 | ||
4078 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, | 4078 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, |
4079 | struct cpumask *new_mask) | 4079 | struct cpumask *new_mask) |
4080 | { | 4080 | { |
4081 | if (len < cpumask_size()) | 4081 | if (len < cpumask_size()) |
4082 | cpumask_clear(new_mask); | 4082 | cpumask_clear(new_mask); |
4083 | else if (len > cpumask_size()) | 4083 | else if (len > cpumask_size()) |
4084 | len = cpumask_size(); | 4084 | len = cpumask_size(); |
4085 | 4085 | ||
4086 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; | 4086 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; |
4087 | } | 4087 | } |
4088 | 4088 | ||
4089 | /** | 4089 | /** |
4090 | * sys_sched_setaffinity - set the cpu affinity of a process | 4090 | * sys_sched_setaffinity - set the cpu affinity of a process |
4091 | * @pid: pid of the process | 4091 | * @pid: pid of the process |
4092 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | 4092 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr |
4093 | * @user_mask_ptr: user-space pointer to the new cpu mask | 4093 | * @user_mask_ptr: user-space pointer to the new cpu mask |
4094 | * | 4094 | * |
4095 | * Return: 0 on success. An error code otherwise. | 4095 | * Return: 0 on success. An error code otherwise. |
4096 | */ | 4096 | */ |
4097 | SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, | 4097 | SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, |
4098 | unsigned long __user *, user_mask_ptr) | 4098 | unsigned long __user *, user_mask_ptr) |
4099 | { | 4099 | { |
4100 | cpumask_var_t new_mask; | 4100 | cpumask_var_t new_mask; |
4101 | int retval; | 4101 | int retval; |
4102 | 4102 | ||
4103 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) | 4103 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) |
4104 | return -ENOMEM; | 4104 | return -ENOMEM; |
4105 | 4105 | ||
4106 | retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); | 4106 | retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); |
4107 | if (retval == 0) | 4107 | if (retval == 0) |
4108 | retval = sched_setaffinity(pid, new_mask); | 4108 | retval = sched_setaffinity(pid, new_mask); |
4109 | free_cpumask_var(new_mask); | 4109 | free_cpumask_var(new_mask); |
4110 | return retval; | 4110 | return retval; |
4111 | } | 4111 | } |
4112 | 4112 | ||
4113 | long sched_getaffinity(pid_t pid, struct cpumask *mask) | 4113 | long sched_getaffinity(pid_t pid, struct cpumask *mask) |
4114 | { | 4114 | { |
4115 | struct task_struct *p; | 4115 | struct task_struct *p; |
4116 | unsigned long flags; | 4116 | unsigned long flags; |
4117 | int retval; | 4117 | int retval; |
4118 | 4118 | ||
4119 | rcu_read_lock(); | 4119 | rcu_read_lock(); |
4120 | 4120 | ||
4121 | retval = -ESRCH; | 4121 | retval = -ESRCH; |
4122 | p = find_process_by_pid(pid); | 4122 | p = find_process_by_pid(pid); |
4123 | if (!p) | 4123 | if (!p) |
4124 | goto out_unlock; | 4124 | goto out_unlock; |
4125 | 4125 | ||
4126 | retval = security_task_getscheduler(p); | 4126 | retval = security_task_getscheduler(p); |
4127 | if (retval) | 4127 | if (retval) |
4128 | goto out_unlock; | 4128 | goto out_unlock; |
4129 | 4129 | ||
4130 | raw_spin_lock_irqsave(&p->pi_lock, flags); | 4130 | raw_spin_lock_irqsave(&p->pi_lock, flags); |
4131 | cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); | 4131 | cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); |
4132 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); | 4132 | raw_spin_unlock_irqrestore(&p->pi_lock, flags); |
4133 | 4133 | ||
4134 | out_unlock: | 4134 | out_unlock: |
4135 | rcu_read_unlock(); | 4135 | rcu_read_unlock(); |
4136 | 4136 | ||
4137 | return retval; | 4137 | return retval; |
4138 | } | 4138 | } |
4139 | 4139 | ||
4140 | /** | 4140 | /** |
4141 | * sys_sched_getaffinity - get the cpu affinity of a process | 4141 | * sys_sched_getaffinity - get the cpu affinity of a process |
4142 | * @pid: pid of the process | 4142 | * @pid: pid of the process |
4143 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | 4143 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr |
4144 | * @user_mask_ptr: user-space pointer to hold the current cpu mask | 4144 | * @user_mask_ptr: user-space pointer to hold the current cpu mask |
4145 | * | 4145 | * |
4146 | * Return: 0 on success. An error code otherwise. | 4146 | * Return: 0 on success. An error code otherwise. |
4147 | */ | 4147 | */ |
4148 | SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, | 4148 | SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, |
4149 | unsigned long __user *, user_mask_ptr) | 4149 | unsigned long __user *, user_mask_ptr) |
4150 | { | 4150 | { |
4151 | int ret; | 4151 | int ret; |
4152 | cpumask_var_t mask; | 4152 | cpumask_var_t mask; |
4153 | 4153 | ||
4154 | if ((len * BITS_PER_BYTE) < nr_cpu_ids) | 4154 | if ((len * BITS_PER_BYTE) < nr_cpu_ids) |
4155 | return -EINVAL; | 4155 | return -EINVAL; |
4156 | if (len & (sizeof(unsigned long)-1)) | 4156 | if (len & (sizeof(unsigned long)-1)) |
4157 | return -EINVAL; | 4157 | return -EINVAL; |
4158 | 4158 | ||
4159 | if (!alloc_cpumask_var(&mask, GFP_KERNEL)) | 4159 | if (!alloc_cpumask_var(&mask, GFP_KERNEL)) |
4160 | return -ENOMEM; | 4160 | return -ENOMEM; |
4161 | 4161 | ||
4162 | ret = sched_getaffinity(pid, mask); | 4162 | ret = sched_getaffinity(pid, mask); |
4163 | if (ret == 0) { | 4163 | if (ret == 0) { |
4164 | size_t retlen = min_t(size_t, len, cpumask_size()); | 4164 | size_t retlen = min_t(size_t, len, cpumask_size()); |
4165 | 4165 | ||
4166 | if (copy_to_user(user_mask_ptr, mask, retlen)) | 4166 | if (copy_to_user(user_mask_ptr, mask, retlen)) |
4167 | ret = -EFAULT; | 4167 | ret = -EFAULT; |
4168 | else | 4168 | else |
4169 | ret = retlen; | 4169 | ret = retlen; |
4170 | } | 4170 | } |
4171 | free_cpumask_var(mask); | 4171 | free_cpumask_var(mask); |
4172 | 4172 | ||
4173 | return ret; | 4173 | return ret; |
4174 | } | 4174 | } |
4175 | 4175 | ||
4176 | /** | 4176 | /** |
4177 | * sys_sched_yield - yield the current processor to other threads. | 4177 | * sys_sched_yield - yield the current processor to other threads. |
4178 | * | 4178 | * |
4179 | * This function yields the current CPU to other tasks. If there are no | 4179 | * This function yields the current CPU to other tasks. If there are no |
4180 | * other threads running on this CPU then this function will return. | 4180 | * other threads running on this CPU then this function will return. |
4181 | * | 4181 | * |
4182 | * Return: 0. | 4182 | * Return: 0. |
4183 | */ | 4183 | */ |
4184 | SYSCALL_DEFINE0(sched_yield) | 4184 | SYSCALL_DEFINE0(sched_yield) |
4185 | { | 4185 | { |
4186 | struct rq *rq = this_rq_lock(); | 4186 | struct rq *rq = this_rq_lock(); |
4187 | 4187 | ||
4188 | schedstat_inc(rq, yld_count); | 4188 | schedstat_inc(rq, yld_count); |
4189 | current->sched_class->yield_task(rq); | 4189 | current->sched_class->yield_task(rq); |
4190 | 4190 | ||
4191 | /* | 4191 | /* |
4192 | * Since we are going to call schedule() anyway, there's | 4192 | * Since we are going to call schedule() anyway, there's |
4193 | * no need to preempt or enable interrupts: | 4193 | * no need to preempt or enable interrupts: |
4194 | */ | 4194 | */ |
4195 | __release(rq->lock); | 4195 | __release(rq->lock); |
4196 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); | 4196 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
4197 | do_raw_spin_unlock(&rq->lock); | 4197 | do_raw_spin_unlock(&rq->lock); |
4198 | sched_preempt_enable_no_resched(); | 4198 | sched_preempt_enable_no_resched(); |
4199 | 4199 | ||
4200 | schedule(); | 4200 | schedule(); |
4201 | 4201 | ||
4202 | return 0; | 4202 | return 0; |
4203 | } | 4203 | } |
4204 | 4204 | ||
4205 | static void __cond_resched(void) | 4205 | static void __cond_resched(void) |
4206 | { | 4206 | { |
4207 | __preempt_count_add(PREEMPT_ACTIVE); | 4207 | __preempt_count_add(PREEMPT_ACTIVE); |
4208 | __schedule(); | 4208 | __schedule(); |
4209 | __preempt_count_sub(PREEMPT_ACTIVE); | 4209 | __preempt_count_sub(PREEMPT_ACTIVE); |
4210 | } | 4210 | } |
4211 | 4211 | ||
4212 | int __sched _cond_resched(void) | 4212 | int __sched _cond_resched(void) |
4213 | { | 4213 | { |
4214 | if (should_resched()) { | 4214 | if (should_resched()) { |
4215 | __cond_resched(); | 4215 | __cond_resched(); |
4216 | return 1; | 4216 | return 1; |
4217 | } | 4217 | } |
4218 | return 0; | 4218 | return 0; |
4219 | } | 4219 | } |
4220 | EXPORT_SYMBOL(_cond_resched); | 4220 | EXPORT_SYMBOL(_cond_resched); |
4221 | 4221 | ||
4222 | /* | 4222 | /* |
4223 | * __cond_resched_lock() - if a reschedule is pending, drop the given lock, | 4223 | * __cond_resched_lock() - if a reschedule is pending, drop the given lock, |
4224 | * call schedule, and on return reacquire the lock. | 4224 | * call schedule, and on return reacquire the lock. |
4225 | * | 4225 | * |
4226 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level | 4226 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level |
4227 | * operations here to prevent schedule() from being called twice (once via | 4227 | * operations here to prevent schedule() from being called twice (once via |
4228 | * spin_unlock(), once by hand). | 4228 | * spin_unlock(), once by hand). |
4229 | */ | 4229 | */ |
4230 | int __cond_resched_lock(spinlock_t *lock) | 4230 | int __cond_resched_lock(spinlock_t *lock) |
4231 | { | 4231 | { |
4232 | int resched = should_resched(); | 4232 | int resched = should_resched(); |
4233 | int ret = 0; | 4233 | int ret = 0; |
4234 | 4234 | ||
4235 | lockdep_assert_held(lock); | 4235 | lockdep_assert_held(lock); |
4236 | 4236 | ||
4237 | if (spin_needbreak(lock) || resched) { | 4237 | if (spin_needbreak(lock) || resched) { |
4238 | spin_unlock(lock); | 4238 | spin_unlock(lock); |
4239 | if (resched) | 4239 | if (resched) |
4240 | __cond_resched(); | 4240 | __cond_resched(); |
4241 | else | 4241 | else |
4242 | cpu_relax(); | 4242 | cpu_relax(); |
4243 | ret = 1; | 4243 | ret = 1; |
4244 | spin_lock(lock); | 4244 | spin_lock(lock); |
4245 | } | 4245 | } |
4246 | return ret; | 4246 | return ret; |
4247 | } | 4247 | } |
4248 | EXPORT_SYMBOL(__cond_resched_lock); | 4248 | EXPORT_SYMBOL(__cond_resched_lock); |
4249 | 4249 | ||
4250 | int __sched __cond_resched_softirq(void) | 4250 | int __sched __cond_resched_softirq(void) |
4251 | { | 4251 | { |
4252 | BUG_ON(!in_softirq()); | 4252 | BUG_ON(!in_softirq()); |
4253 | 4253 | ||
4254 | if (should_resched()) { | 4254 | if (should_resched()) { |
4255 | local_bh_enable(); | 4255 | local_bh_enable(); |
4256 | __cond_resched(); | 4256 | __cond_resched(); |
4257 | local_bh_disable(); | 4257 | local_bh_disable(); |
4258 | return 1; | 4258 | return 1; |
4259 | } | 4259 | } |
4260 | return 0; | 4260 | return 0; |
4261 | } | 4261 | } |
4262 | EXPORT_SYMBOL(__cond_resched_softirq); | 4262 | EXPORT_SYMBOL(__cond_resched_softirq); |
4263 | 4263 | ||
4264 | /** | 4264 | /** |
4265 | * yield - yield the current processor to other threads. | 4265 | * yield - yield the current processor to other threads. |
4266 | * | 4266 | * |
4267 | * Do not ever use this function, there's a 99% chance you're doing it wrong. | 4267 | * Do not ever use this function, there's a 99% chance you're doing it wrong. |
4268 | * | 4268 | * |
4269 | * The scheduler is at all times free to pick the calling task as the most | 4269 | * The scheduler is at all times free to pick the calling task as the most |
4270 | * eligible task to run, if removing the yield() call from your code breaks | 4270 | * eligible task to run, if removing the yield() call from your code breaks |
4271 | * it, its already broken. | 4271 | * it, its already broken. |
4272 | * | 4272 | * |
4273 | * Typical broken usage is: | 4273 | * Typical broken usage is: |
4274 | * | 4274 | * |
4275 | * while (!event) | 4275 | * while (!event) |
4276 | * yield(); | 4276 | * yield(); |
4277 | * | 4277 | * |
4278 | * where one assumes that yield() will let 'the other' process run that will | 4278 | * where one assumes that yield() will let 'the other' process run that will |
4279 | * make event true. If the current task is a SCHED_FIFO task that will never | 4279 | * make event true. If the current task is a SCHED_FIFO task that will never |
4280 | * happen. Never use yield() as a progress guarantee!! | 4280 | * happen. Never use yield() as a progress guarantee!! |
4281 | * | 4281 | * |
4282 | * If you want to use yield() to wait for something, use wait_event(). | 4282 | * If you want to use yield() to wait for something, use wait_event(). |
4283 | * If you want to use yield() to be 'nice' for others, use cond_resched(). | 4283 | * If you want to use yield() to be 'nice' for others, use cond_resched(). |
4284 | * If you still want to use yield(), do not! | 4284 | * If you still want to use yield(), do not! |
4285 | */ | 4285 | */ |
4286 | void __sched yield(void) | 4286 | void __sched yield(void) |
4287 | { | 4287 | { |
4288 | set_current_state(TASK_RUNNING); | 4288 | set_current_state(TASK_RUNNING); |
4289 | sys_sched_yield(); | 4289 | sys_sched_yield(); |
4290 | } | 4290 | } |
4291 | EXPORT_SYMBOL(yield); | 4291 | EXPORT_SYMBOL(yield); |
4292 | 4292 | ||
4293 | /** | 4293 | /** |
4294 | * yield_to - yield the current processor to another thread in | 4294 | * yield_to - yield the current processor to another thread in |
4295 | * your thread group, or accelerate that thread toward the | 4295 | * your thread group, or accelerate that thread toward the |
4296 | * processor it's on. | 4296 | * processor it's on. |
4297 | * @p: target task | 4297 | * @p: target task |
4298 | * @preempt: whether task preemption is allowed or not | 4298 | * @preempt: whether task preemption is allowed or not |
4299 | * | 4299 | * |
4300 | * It's the caller's job to ensure that the target task struct | 4300 | * It's the caller's job to ensure that the target task struct |
4301 | * can't go away on us before we can do any checks. | 4301 | * can't go away on us before we can do any checks. |
4302 | * | 4302 | * |
4303 | * Return: | 4303 | * Return: |
4304 | * true (>0) if we indeed boosted the target task. | 4304 | * true (>0) if we indeed boosted the target task. |
4305 | * false (0) if we failed to boost the target. | 4305 | * false (0) if we failed to boost the target. |
4306 | * -ESRCH if there's no task to yield to. | 4306 | * -ESRCH if there's no task to yield to. |
4307 | */ | 4307 | */ |
4308 | int __sched yield_to(struct task_struct *p, bool preempt) | 4308 | int __sched yield_to(struct task_struct *p, bool preempt) |
4309 | { | 4309 | { |
4310 | struct task_struct *curr = current; | 4310 | struct task_struct *curr = current; |
4311 | struct rq *rq, *p_rq; | 4311 | struct rq *rq, *p_rq; |
4312 | unsigned long flags; | 4312 | unsigned long flags; |
4313 | int yielded = 0; | 4313 | int yielded = 0; |
4314 | 4314 | ||
4315 | local_irq_save(flags); | 4315 | local_irq_save(flags); |
4316 | rq = this_rq(); | 4316 | rq = this_rq(); |
4317 | 4317 | ||
4318 | again: | 4318 | again: |
4319 | p_rq = task_rq(p); | 4319 | p_rq = task_rq(p); |
4320 | /* | 4320 | /* |
4321 | * If we're the only runnable task on the rq and target rq also | 4321 | * If we're the only runnable task on the rq and target rq also |
4322 | * has only one task, there's absolutely no point in yielding. | 4322 | * has only one task, there's absolutely no point in yielding. |
4323 | */ | 4323 | */ |
4324 | if (rq->nr_running == 1 && p_rq->nr_running == 1) { | 4324 | if (rq->nr_running == 1 && p_rq->nr_running == 1) { |
4325 | yielded = -ESRCH; | 4325 | yielded = -ESRCH; |
4326 | goto out_irq; | 4326 | goto out_irq; |
4327 | } | 4327 | } |
4328 | 4328 | ||
4329 | double_rq_lock(rq, p_rq); | 4329 | double_rq_lock(rq, p_rq); |
4330 | if (task_rq(p) != p_rq) { | 4330 | if (task_rq(p) != p_rq) { |
4331 | double_rq_unlock(rq, p_rq); | 4331 | double_rq_unlock(rq, p_rq); |
4332 | goto again; | 4332 | goto again; |
4333 | } | 4333 | } |
4334 | 4334 | ||
4335 | if (!curr->sched_class->yield_to_task) | 4335 | if (!curr->sched_class->yield_to_task) |
4336 | goto out_unlock; | 4336 | goto out_unlock; |
4337 | 4337 | ||
4338 | if (curr->sched_class != p->sched_class) | 4338 | if (curr->sched_class != p->sched_class) |
4339 | goto out_unlock; | 4339 | goto out_unlock; |
4340 | 4340 | ||
4341 | if (task_running(p_rq, p) || p->state) | 4341 | if (task_running(p_rq, p) || p->state) |
4342 | goto out_unlock; | 4342 | goto out_unlock; |
4343 | 4343 | ||
4344 | yielded = curr->sched_class->yield_to_task(rq, p, preempt); | 4344 | yielded = curr->sched_class->yield_to_task(rq, p, preempt); |
4345 | if (yielded) { | 4345 | if (yielded) { |
4346 | schedstat_inc(rq, yld_count); | 4346 | schedstat_inc(rq, yld_count); |
4347 | /* | 4347 | /* |
4348 | * Make p's CPU reschedule; pick_next_entity takes care of | 4348 | * Make p's CPU reschedule; pick_next_entity takes care of |
4349 | * fairness. | 4349 | * fairness. |
4350 | */ | 4350 | */ |
4351 | if (preempt && rq != p_rq) | 4351 | if (preempt && rq != p_rq) |
4352 | resched_curr(p_rq); | 4352 | resched_curr(p_rq); |
4353 | } | 4353 | } |
4354 | 4354 | ||
4355 | out_unlock: | 4355 | out_unlock: |
4356 | double_rq_unlock(rq, p_rq); | 4356 | double_rq_unlock(rq, p_rq); |
4357 | out_irq: | 4357 | out_irq: |
4358 | local_irq_restore(flags); | 4358 | local_irq_restore(flags); |
4359 | 4359 | ||
4360 | if (yielded > 0) | 4360 | if (yielded > 0) |
4361 | schedule(); | 4361 | schedule(); |
4362 | 4362 | ||
4363 | return yielded; | 4363 | return yielded; |
4364 | } | 4364 | } |
4365 | EXPORT_SYMBOL_GPL(yield_to); | 4365 | EXPORT_SYMBOL_GPL(yield_to); |
4366 | 4366 | ||
4367 | /* | 4367 | /* |
4368 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so | 4368 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so |
4369 | * that process accounting knows that this is a task in IO wait state. | 4369 | * that process accounting knows that this is a task in IO wait state. |
4370 | */ | 4370 | */ |
4371 | void __sched io_schedule(void) | 4371 | void __sched io_schedule(void) |
4372 | { | 4372 | { |
4373 | struct rq *rq = raw_rq(); | 4373 | struct rq *rq = raw_rq(); |
4374 | 4374 | ||
4375 | delayacct_blkio_start(); | 4375 | delayacct_blkio_start(); |
4376 | atomic_inc(&rq->nr_iowait); | 4376 | atomic_inc(&rq->nr_iowait); |
4377 | blk_flush_plug(current); | 4377 | blk_flush_plug(current); |
4378 | current->in_iowait = 1; | 4378 | current->in_iowait = 1; |
4379 | schedule(); | 4379 | schedule(); |
4380 | current->in_iowait = 0; | 4380 | current->in_iowait = 0; |
4381 | atomic_dec(&rq->nr_iowait); | 4381 | atomic_dec(&rq->nr_iowait); |
4382 | delayacct_blkio_end(); | 4382 | delayacct_blkio_end(); |
4383 | } | 4383 | } |
4384 | EXPORT_SYMBOL(io_schedule); | 4384 | EXPORT_SYMBOL(io_schedule); |
4385 | 4385 | ||
4386 | long __sched io_schedule_timeout(long timeout) | 4386 | long __sched io_schedule_timeout(long timeout) |
4387 | { | 4387 | { |
4388 | struct rq *rq = raw_rq(); | 4388 | struct rq *rq = raw_rq(); |
4389 | long ret; | 4389 | long ret; |
4390 | 4390 | ||
4391 | delayacct_blkio_start(); | 4391 | delayacct_blkio_start(); |
4392 | atomic_inc(&rq->nr_iowait); | 4392 | atomic_inc(&rq->nr_iowait); |
4393 | blk_flush_plug(current); | 4393 | blk_flush_plug(current); |
4394 | current->in_iowait = 1; | 4394 | current->in_iowait = 1; |
4395 | ret = schedule_timeout(timeout); | 4395 | ret = schedule_timeout(timeout); |
4396 | current->in_iowait = 0; | 4396 | current->in_iowait = 0; |
4397 | atomic_dec(&rq->nr_iowait); | 4397 | atomic_dec(&rq->nr_iowait); |
4398 | delayacct_blkio_end(); | 4398 | delayacct_blkio_end(); |
4399 | return ret; | 4399 | return ret; |
4400 | } | 4400 | } |
4401 | 4401 | ||
4402 | /** | 4402 | /** |
4403 | * sys_sched_get_priority_max - return maximum RT priority. | 4403 | * sys_sched_get_priority_max - return maximum RT priority. |
4404 | * @policy: scheduling class. | 4404 | * @policy: scheduling class. |
4405 | * | 4405 | * |
4406 | * Return: On success, this syscall returns the maximum | 4406 | * Return: On success, this syscall returns the maximum |
4407 | * rt_priority that can be used by a given scheduling class. | 4407 | * rt_priority that can be used by a given scheduling class. |
4408 | * On failure, a negative error code is returned. | 4408 | * On failure, a negative error code is returned. |
4409 | */ | 4409 | */ |
4410 | SYSCALL_DEFINE1(sched_get_priority_max, int, policy) | 4410 | SYSCALL_DEFINE1(sched_get_priority_max, int, policy) |
4411 | { | 4411 | { |
4412 | int ret = -EINVAL; | 4412 | int ret = -EINVAL; |
4413 | 4413 | ||
4414 | switch (policy) { | 4414 | switch (policy) { |
4415 | case SCHED_FIFO: | 4415 | case SCHED_FIFO: |
4416 | case SCHED_RR: | 4416 | case SCHED_RR: |
4417 | ret = MAX_USER_RT_PRIO-1; | 4417 | ret = MAX_USER_RT_PRIO-1; |
4418 | break; | 4418 | break; |
4419 | case SCHED_DEADLINE: | 4419 | case SCHED_DEADLINE: |
4420 | case SCHED_NORMAL: | 4420 | case SCHED_NORMAL: |
4421 | case SCHED_BATCH: | 4421 | case SCHED_BATCH: |
4422 | case SCHED_IDLE: | 4422 | case SCHED_IDLE: |
4423 | ret = 0; | 4423 | ret = 0; |
4424 | break; | 4424 | break; |
4425 | } | 4425 | } |
4426 | return ret; | 4426 | return ret; |
4427 | } | 4427 | } |
4428 | 4428 | ||
4429 | /** | 4429 | /** |
4430 | * sys_sched_get_priority_min - return minimum RT priority. | 4430 | * sys_sched_get_priority_min - return minimum RT priority. |
4431 | * @policy: scheduling class. | 4431 | * @policy: scheduling class. |
4432 | * | 4432 | * |
4433 | * Return: On success, this syscall returns the minimum | 4433 | * Return: On success, this syscall returns the minimum |
4434 | * rt_priority that can be used by a given scheduling class. | 4434 | * rt_priority that can be used by a given scheduling class. |
4435 | * On failure, a negative error code is returned. | 4435 | * On failure, a negative error code is returned. |
4436 | */ | 4436 | */ |
4437 | SYSCALL_DEFINE1(sched_get_priority_min, int, policy) | 4437 | SYSCALL_DEFINE1(sched_get_priority_min, int, policy) |
4438 | { | 4438 | { |
4439 | int ret = -EINVAL; | 4439 | int ret = -EINVAL; |
4440 | 4440 | ||
4441 | switch (policy) { | 4441 | switch (policy) { |
4442 | case SCHED_FIFO: | 4442 | case SCHED_FIFO: |
4443 | case SCHED_RR: | 4443 | case SCHED_RR: |
4444 | ret = 1; | 4444 | ret = 1; |
4445 | break; | 4445 | break; |
4446 | case SCHED_DEADLINE: | 4446 | case SCHED_DEADLINE: |
4447 | case SCHED_NORMAL: | 4447 | case SCHED_NORMAL: |
4448 | case SCHED_BATCH: | 4448 | case SCHED_BATCH: |
4449 | case SCHED_IDLE: | 4449 | case SCHED_IDLE: |
4450 | ret = 0; | 4450 | ret = 0; |
4451 | } | 4451 | } |
4452 | return ret; | 4452 | return ret; |
4453 | } | 4453 | } |
4454 | 4454 | ||
4455 | /** | 4455 | /** |
4456 | * sys_sched_rr_get_interval - return the default timeslice of a process. | 4456 | * sys_sched_rr_get_interval - return the default timeslice of a process. |
4457 | * @pid: pid of the process. | 4457 | * @pid: pid of the process. |
4458 | * @interval: userspace pointer to the timeslice value. | 4458 | * @interval: userspace pointer to the timeslice value. |
4459 | * | 4459 | * |
4460 | * this syscall writes the default timeslice value of a given process | 4460 | * this syscall writes the default timeslice value of a given process |
4461 | * into the user-space timespec buffer. A value of '0' means infinity. | 4461 | * into the user-space timespec buffer. A value of '0' means infinity. |
4462 | * | 4462 | * |
4463 | * Return: On success, 0 and the timeslice is in @interval. Otherwise, | 4463 | * Return: On success, 0 and the timeslice is in @interval. Otherwise, |
4464 | * an error code. | 4464 | * an error code. |
4465 | */ | 4465 | */ |
4466 | SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, | 4466 | SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, |
4467 | struct timespec __user *, interval) | 4467 | struct timespec __user *, interval) |
4468 | { | 4468 | { |
4469 | struct task_struct *p; | 4469 | struct task_struct *p; |
4470 | unsigned int time_slice; | 4470 | unsigned int time_slice; |
4471 | unsigned long flags; | 4471 | unsigned long flags; |
4472 | struct rq *rq; | 4472 | struct rq *rq; |
4473 | int retval; | 4473 | int retval; |
4474 | struct timespec t; | 4474 | struct timespec t; |
4475 | 4475 | ||
4476 | if (pid < 0) | 4476 | if (pid < 0) |
4477 | return -EINVAL; | 4477 | return -EINVAL; |
4478 | 4478 | ||
4479 | retval = -ESRCH; | 4479 | retval = -ESRCH; |
4480 | rcu_read_lock(); | 4480 | rcu_read_lock(); |
4481 | p = find_process_by_pid(pid); | 4481 | p = find_process_by_pid(pid); |
4482 | if (!p) | 4482 | if (!p) |
4483 | goto out_unlock; | 4483 | goto out_unlock; |
4484 | 4484 | ||
4485 | retval = security_task_getscheduler(p); | 4485 | retval = security_task_getscheduler(p); |
4486 | if (retval) | 4486 | if (retval) |
4487 | goto out_unlock; | 4487 | goto out_unlock; |
4488 | 4488 | ||
4489 | rq = task_rq_lock(p, &flags); | 4489 | rq = task_rq_lock(p, &flags); |
4490 | time_slice = 0; | 4490 | time_slice = 0; |
4491 | if (p->sched_class->get_rr_interval) | 4491 | if (p->sched_class->get_rr_interval) |
4492 | time_slice = p->sched_class->get_rr_interval(rq, p); | 4492 | time_slice = p->sched_class->get_rr_interval(rq, p); |
4493 | task_rq_unlock(rq, p, &flags); | 4493 | task_rq_unlock(rq, p, &flags); |
4494 | 4494 | ||
4495 | rcu_read_unlock(); | 4495 | rcu_read_unlock(); |
4496 | jiffies_to_timespec(time_slice, &t); | 4496 | jiffies_to_timespec(time_slice, &t); |
4497 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; | 4497 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; |
4498 | return retval; | 4498 | return retval; |
4499 | 4499 | ||
4500 | out_unlock: | 4500 | out_unlock: |
4501 | rcu_read_unlock(); | 4501 | rcu_read_unlock(); |
4502 | return retval; | 4502 | return retval; |
4503 | } | 4503 | } |
4504 | 4504 | ||
4505 | static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; | 4505 | static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; |
4506 | 4506 | ||
4507 | void sched_show_task(struct task_struct *p) | 4507 | void sched_show_task(struct task_struct *p) |
4508 | { | 4508 | { |
4509 | unsigned long free = 0; | 4509 | unsigned long free = 0; |
4510 | int ppid; | 4510 | int ppid; |
4511 | unsigned state; | 4511 | unsigned state; |
4512 | 4512 | ||
4513 | state = p->state ? __ffs(p->state) + 1 : 0; | 4513 | state = p->state ? __ffs(p->state) + 1 : 0; |
4514 | printk(KERN_INFO "%-15.15s %c", p->comm, | 4514 | printk(KERN_INFO "%-15.15s %c", p->comm, |
4515 | state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); | 4515 | state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); |
4516 | #if BITS_PER_LONG == 32 | 4516 | #if BITS_PER_LONG == 32 |
4517 | if (state == TASK_RUNNING) | 4517 | if (state == TASK_RUNNING) |
4518 | printk(KERN_CONT " running "); | 4518 | printk(KERN_CONT " running "); |
4519 | else | 4519 | else |
4520 | printk(KERN_CONT " %08lx ", thread_saved_pc(p)); | 4520 | printk(KERN_CONT " %08lx ", thread_saved_pc(p)); |
4521 | #else | 4521 | #else |
4522 | if (state == TASK_RUNNING) | 4522 | if (state == TASK_RUNNING) |
4523 | printk(KERN_CONT " running task "); | 4523 | printk(KERN_CONT " running task "); |
4524 | else | 4524 | else |
4525 | printk(KERN_CONT " %016lx ", thread_saved_pc(p)); | 4525 | printk(KERN_CONT " %016lx ", thread_saved_pc(p)); |
4526 | #endif | 4526 | #endif |
4527 | #ifdef CONFIG_DEBUG_STACK_USAGE | 4527 | #ifdef CONFIG_DEBUG_STACK_USAGE |
4528 | free = stack_not_used(p); | 4528 | free = stack_not_used(p); |
4529 | #endif | 4529 | #endif |
4530 | ppid = 0; | 4530 | ppid = 0; |
4531 | rcu_read_lock(); | 4531 | rcu_read_lock(); |
4532 | if (pid_alive(p)) | 4532 | if (pid_alive(p)) |
4533 | ppid = task_pid_nr(rcu_dereference(p->real_parent)); | 4533 | ppid = task_pid_nr(rcu_dereference(p->real_parent)); |
4534 | rcu_read_unlock(); | 4534 | rcu_read_unlock(); |
4535 | printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, | 4535 | printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, |
4536 | task_pid_nr(p), ppid, | 4536 | task_pid_nr(p), ppid, |
4537 | (unsigned long)task_thread_info(p)->flags); | 4537 | (unsigned long)task_thread_info(p)->flags); |
4538 | 4538 | ||
4539 | print_worker_info(KERN_INFO, p); | 4539 | print_worker_info(KERN_INFO, p); |
4540 | show_stack(p, NULL); | 4540 | show_stack(p, NULL); |
4541 | } | 4541 | } |
4542 | 4542 | ||
4543 | void show_state_filter(unsigned long state_filter) | 4543 | void show_state_filter(unsigned long state_filter) |
4544 | { | 4544 | { |
4545 | struct task_struct *g, *p; | 4545 | struct task_struct *g, *p; |
4546 | 4546 | ||
4547 | #if BITS_PER_LONG == 32 | 4547 | #if BITS_PER_LONG == 32 |
4548 | printk(KERN_INFO | 4548 | printk(KERN_INFO |
4549 | " task PC stack pid father\n"); | 4549 | " task PC stack pid father\n"); |
4550 | #else | 4550 | #else |
4551 | printk(KERN_INFO | 4551 | printk(KERN_INFO |
4552 | " task PC stack pid father\n"); | 4552 | " task PC stack pid father\n"); |
4553 | #endif | 4553 | #endif |
4554 | rcu_read_lock(); | 4554 | rcu_read_lock(); |
4555 | for_each_process_thread(g, p) { | 4555 | for_each_process_thread(g, p) { |
4556 | /* | 4556 | /* |
4557 | * reset the NMI-timeout, listing all files on a slow | 4557 | * reset the NMI-timeout, listing all files on a slow |
4558 | * console might take a lot of time: | 4558 | * console might take a lot of time: |
4559 | */ | 4559 | */ |
4560 | touch_nmi_watchdog(); | 4560 | touch_nmi_watchdog(); |
4561 | if (!state_filter || (p->state & state_filter)) | 4561 | if (!state_filter || (p->state & state_filter)) |
4562 | sched_show_task(p); | 4562 | sched_show_task(p); |
4563 | } | 4563 | } |
4564 | 4564 | ||
4565 | touch_all_softlockup_watchdogs(); | 4565 | touch_all_softlockup_watchdogs(); |
4566 | 4566 | ||
4567 | #ifdef CONFIG_SCHED_DEBUG | 4567 | #ifdef CONFIG_SCHED_DEBUG |
4568 | sysrq_sched_debug_show(); | 4568 | sysrq_sched_debug_show(); |
4569 | #endif | 4569 | #endif |
4570 | rcu_read_unlock(); | 4570 | rcu_read_unlock(); |
4571 | /* | 4571 | /* |
4572 | * Only show locks if all tasks are dumped: | 4572 | * Only show locks if all tasks are dumped: |
4573 | */ | 4573 | */ |
4574 | if (!state_filter) | 4574 | if (!state_filter) |
4575 | debug_show_all_locks(); | 4575 | debug_show_all_locks(); |
4576 | } | 4576 | } |
4577 | 4577 | ||
4578 | void init_idle_bootup_task(struct task_struct *idle) | 4578 | void init_idle_bootup_task(struct task_struct *idle) |
4579 | { | 4579 | { |
4580 | idle->sched_class = &idle_sched_class; | 4580 | idle->sched_class = &idle_sched_class; |
4581 | } | 4581 | } |
4582 | 4582 | ||
4583 | /** | 4583 | /** |
4584 | * init_idle - set up an idle thread for a given CPU | 4584 | * init_idle - set up an idle thread for a given CPU |
4585 | * @idle: task in question | 4585 | * @idle: task in question |
4586 | * @cpu: cpu the idle task belongs to | 4586 | * @cpu: cpu the idle task belongs to |
4587 | * | 4587 | * |
4588 | * NOTE: this function does not set the idle thread's NEED_RESCHED | 4588 | * NOTE: this function does not set the idle thread's NEED_RESCHED |
4589 | * flag, to make booting more robust. | 4589 | * flag, to make booting more robust. |
4590 | */ | 4590 | */ |
4591 | void init_idle(struct task_struct *idle, int cpu) | 4591 | void init_idle(struct task_struct *idle, int cpu) |
4592 | { | 4592 | { |
4593 | struct rq *rq = cpu_rq(cpu); | 4593 | struct rq *rq = cpu_rq(cpu); |
4594 | unsigned long flags; | 4594 | unsigned long flags; |
4595 | 4595 | ||
4596 | raw_spin_lock_irqsave(&rq->lock, flags); | 4596 | raw_spin_lock_irqsave(&rq->lock, flags); |
4597 | 4597 | ||
4598 | __sched_fork(0, idle); | 4598 | __sched_fork(0, idle); |
4599 | idle->state = TASK_RUNNING; | 4599 | idle->state = TASK_RUNNING; |
4600 | idle->se.exec_start = sched_clock(); | 4600 | idle->se.exec_start = sched_clock(); |
4601 | 4601 | ||
4602 | do_set_cpus_allowed(idle, cpumask_of(cpu)); | 4602 | do_set_cpus_allowed(idle, cpumask_of(cpu)); |
4603 | /* | 4603 | /* |
4604 | * We're having a chicken and egg problem, even though we are | 4604 | * We're having a chicken and egg problem, even though we are |
4605 | * holding rq->lock, the cpu isn't yet set to this cpu so the | 4605 | * holding rq->lock, the cpu isn't yet set to this cpu so the |
4606 | * lockdep check in task_group() will fail. | 4606 | * lockdep check in task_group() will fail. |
4607 | * | 4607 | * |
4608 | * Similar case to sched_fork(). / Alternatively we could | 4608 | * Similar case to sched_fork(). / Alternatively we could |
4609 | * use task_rq_lock() here and obtain the other rq->lock. | 4609 | * use task_rq_lock() here and obtain the other rq->lock. |
4610 | * | 4610 | * |
4611 | * Silence PROVE_RCU | 4611 | * Silence PROVE_RCU |
4612 | */ | 4612 | */ |
4613 | rcu_read_lock(); | 4613 | rcu_read_lock(); |
4614 | __set_task_cpu(idle, cpu); | 4614 | __set_task_cpu(idle, cpu); |
4615 | rcu_read_unlock(); | 4615 | rcu_read_unlock(); |
4616 | 4616 | ||
4617 | rq->curr = rq->idle = idle; | 4617 | rq->curr = rq->idle = idle; |
4618 | idle->on_rq = TASK_ON_RQ_QUEUED; | 4618 | idle->on_rq = TASK_ON_RQ_QUEUED; |
4619 | #if defined(CONFIG_SMP) | 4619 | #if defined(CONFIG_SMP) |
4620 | idle->on_cpu = 1; | 4620 | idle->on_cpu = 1; |
4621 | #endif | 4621 | #endif |
4622 | raw_spin_unlock_irqrestore(&rq->lock, flags); | 4622 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
4623 | 4623 | ||
4624 | /* Set the preempt count _outside_ the spinlocks! */ | 4624 | /* Set the preempt count _outside_ the spinlocks! */ |
4625 | init_idle_preempt_count(idle, cpu); | 4625 | init_idle_preempt_count(idle, cpu); |
4626 | 4626 | ||
4627 | /* | 4627 | /* |
4628 | * The idle tasks have their own, simple scheduling class: | 4628 | * The idle tasks have their own, simple scheduling class: |
4629 | */ | 4629 | */ |
4630 | idle->sched_class = &idle_sched_class; | 4630 | idle->sched_class = &idle_sched_class; |
4631 | ftrace_graph_init_idle_task(idle, cpu); | 4631 | ftrace_graph_init_idle_task(idle, cpu); |
4632 | vtime_init_idle(idle, cpu); | 4632 | vtime_init_idle(idle, cpu); |
4633 | #if defined(CONFIG_SMP) | 4633 | #if defined(CONFIG_SMP) |
4634 | sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); | 4634 | sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); |
4635 | #endif | 4635 | #endif |
4636 | } | 4636 | } |
4637 | 4637 | ||
4638 | int cpuset_cpumask_can_shrink(const struct cpumask *cur, | 4638 | int cpuset_cpumask_can_shrink(const struct cpumask *cur, |
4639 | const struct cpumask *trial) | 4639 | const struct cpumask *trial) |
4640 | { | 4640 | { |
4641 | int ret = 1, trial_cpus; | 4641 | int ret = 1, trial_cpus; |
4642 | struct dl_bw *cur_dl_b; | 4642 | struct dl_bw *cur_dl_b; |
4643 | unsigned long flags; | 4643 | unsigned long flags; |
4644 | 4644 | ||
4645 | rcu_read_lock_sched(); | 4645 | rcu_read_lock_sched(); |
4646 | cur_dl_b = dl_bw_of(cpumask_any(cur)); | 4646 | cur_dl_b = dl_bw_of(cpumask_any(cur)); |
4647 | trial_cpus = cpumask_weight(trial); | 4647 | trial_cpus = cpumask_weight(trial); |
4648 | 4648 | ||
4649 | raw_spin_lock_irqsave(&cur_dl_b->lock, flags); | 4649 | raw_spin_lock_irqsave(&cur_dl_b->lock, flags); |
4650 | if (cur_dl_b->bw != -1 && | 4650 | if (cur_dl_b->bw != -1 && |
4651 | cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw) | 4651 | cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw) |
4652 | ret = 0; | 4652 | ret = 0; |
4653 | raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags); | 4653 | raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags); |
4654 | rcu_read_unlock_sched(); | 4654 | rcu_read_unlock_sched(); |
4655 | 4655 | ||
4656 | return ret; | 4656 | return ret; |
4657 | } | 4657 | } |
4658 | 4658 | ||
4659 | int task_can_attach(struct task_struct *p, | 4659 | int task_can_attach(struct task_struct *p, |
4660 | const struct cpumask *cs_cpus_allowed) | 4660 | const struct cpumask *cs_cpus_allowed) |
4661 | { | 4661 | { |
4662 | int ret = 0; | 4662 | int ret = 0; |
4663 | 4663 | ||
4664 | /* | 4664 | /* |
4665 | * Kthreads which disallow setaffinity shouldn't be moved | 4665 | * Kthreads which disallow setaffinity shouldn't be moved |
4666 | * to a new cpuset; we don't want to change their cpu | 4666 | * to a new cpuset; we don't want to change their cpu |
4667 | * affinity and isolating such threads by their set of | 4667 | * affinity and isolating such threads by their set of |
4668 | * allowed nodes is unnecessary. Thus, cpusets are not | 4668 | * allowed nodes is unnecessary. Thus, cpusets are not |
4669 | * applicable for such threads. This prevents checking for | 4669 | * applicable for such threads. This prevents checking for |
4670 | * success of set_cpus_allowed_ptr() on all attached tasks | 4670 | * success of set_cpus_allowed_ptr() on all attached tasks |
4671 | * before cpus_allowed may be changed. | 4671 | * before cpus_allowed may be changed. |
4672 | */ | 4672 | */ |
4673 | if (p->flags & PF_NO_SETAFFINITY) { | 4673 | if (p->flags & PF_NO_SETAFFINITY) { |
4674 | ret = -EINVAL; | 4674 | ret = -EINVAL; |
4675 | goto out; | 4675 | goto out; |
4676 | } | 4676 | } |
4677 | 4677 | ||
4678 | #ifdef CONFIG_SMP | 4678 | #ifdef CONFIG_SMP |
4679 | if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span, | 4679 | if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span, |
4680 | cs_cpus_allowed)) { | 4680 | cs_cpus_allowed)) { |
4681 | unsigned int dest_cpu = cpumask_any_and(cpu_active_mask, | 4681 | unsigned int dest_cpu = cpumask_any_and(cpu_active_mask, |
4682 | cs_cpus_allowed); | 4682 | cs_cpus_allowed); |
4683 | struct dl_bw *dl_b; | 4683 | struct dl_bw *dl_b; |
4684 | bool overflow; | 4684 | bool overflow; |
4685 | int cpus; | 4685 | int cpus; |
4686 | unsigned long flags; | 4686 | unsigned long flags; |
4687 | 4687 | ||
4688 | rcu_read_lock_sched(); | 4688 | rcu_read_lock_sched(); |
4689 | dl_b = dl_bw_of(dest_cpu); | 4689 | dl_b = dl_bw_of(dest_cpu); |
4690 | raw_spin_lock_irqsave(&dl_b->lock, flags); | 4690 | raw_spin_lock_irqsave(&dl_b->lock, flags); |
4691 | cpus = dl_bw_cpus(dest_cpu); | 4691 | cpus = dl_bw_cpus(dest_cpu); |
4692 | overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw); | 4692 | overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw); |
4693 | if (overflow) | 4693 | if (overflow) |
4694 | ret = -EBUSY; | 4694 | ret = -EBUSY; |
4695 | else { | 4695 | else { |
4696 | /* | 4696 | /* |
4697 | * We reserve space for this task in the destination | 4697 | * We reserve space for this task in the destination |
4698 | * root_domain, as we can't fail after this point. | 4698 | * root_domain, as we can't fail after this point. |
4699 | * We will free resources in the source root_domain | 4699 | * We will free resources in the source root_domain |
4700 | * later on (see set_cpus_allowed_dl()). | 4700 | * later on (see set_cpus_allowed_dl()). |
4701 | */ | 4701 | */ |
4702 | __dl_add(dl_b, p->dl.dl_bw); | 4702 | __dl_add(dl_b, p->dl.dl_bw); |
4703 | } | 4703 | } |
4704 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); | 4704 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); |
4705 | rcu_read_unlock_sched(); | 4705 | rcu_read_unlock_sched(); |
4706 | 4706 | ||
4707 | } | 4707 | } |
4708 | #endif | 4708 | #endif |
4709 | out: | 4709 | out: |
4710 | return ret; | 4710 | return ret; |
4711 | } | 4711 | } |
4712 | 4712 | ||
4713 | #ifdef CONFIG_SMP | 4713 | #ifdef CONFIG_SMP |
4714 | /* | 4714 | /* |
4715 | * move_queued_task - move a queued task to new rq. | 4715 | * move_queued_task - move a queued task to new rq. |
4716 | * | 4716 | * |
4717 | * Returns (locked) new rq. Old rq's lock is released. | 4717 | * Returns (locked) new rq. Old rq's lock is released. |
4718 | */ | 4718 | */ |
4719 | static struct rq *move_queued_task(struct task_struct *p, int new_cpu) | 4719 | static struct rq *move_queued_task(struct task_struct *p, int new_cpu) |
4720 | { | 4720 | { |
4721 | struct rq *rq = task_rq(p); | 4721 | struct rq *rq = task_rq(p); |
4722 | 4722 | ||
4723 | lockdep_assert_held(&rq->lock); | 4723 | lockdep_assert_held(&rq->lock); |
4724 | 4724 | ||
4725 | dequeue_task(rq, p, 0); | 4725 | dequeue_task(rq, p, 0); |
4726 | p->on_rq = TASK_ON_RQ_MIGRATING; | 4726 | p->on_rq = TASK_ON_RQ_MIGRATING; |
4727 | set_task_cpu(p, new_cpu); | 4727 | set_task_cpu(p, new_cpu); |
4728 | raw_spin_unlock(&rq->lock); | 4728 | raw_spin_unlock(&rq->lock); |
4729 | 4729 | ||
4730 | rq = cpu_rq(new_cpu); | 4730 | rq = cpu_rq(new_cpu); |
4731 | 4731 | ||
4732 | raw_spin_lock(&rq->lock); | 4732 | raw_spin_lock(&rq->lock); |
4733 | BUG_ON(task_cpu(p) != new_cpu); | 4733 | BUG_ON(task_cpu(p) != new_cpu); |
4734 | p->on_rq = TASK_ON_RQ_QUEUED; | 4734 | p->on_rq = TASK_ON_RQ_QUEUED; |
4735 | enqueue_task(rq, p, 0); | 4735 | enqueue_task(rq, p, 0); |
4736 | check_preempt_curr(rq, p, 0); | 4736 | check_preempt_curr(rq, p, 0); |
4737 | 4737 | ||
4738 | return rq; | 4738 | return rq; |
4739 | } | 4739 | } |
4740 | 4740 | ||
4741 | void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) | 4741 | void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) |
4742 | { | 4742 | { |
4743 | if (p->sched_class && p->sched_class->set_cpus_allowed) | 4743 | if (p->sched_class && p->sched_class->set_cpus_allowed) |
4744 | p->sched_class->set_cpus_allowed(p, new_mask); | 4744 | p->sched_class->set_cpus_allowed(p, new_mask); |
4745 | 4745 | ||
4746 | cpumask_copy(&p->cpus_allowed, new_mask); | 4746 | cpumask_copy(&p->cpus_allowed, new_mask); |
4747 | p->nr_cpus_allowed = cpumask_weight(new_mask); | 4747 | p->nr_cpus_allowed = cpumask_weight(new_mask); |
4748 | } | 4748 | } |
4749 | 4749 | ||
4750 | /* | 4750 | /* |
4751 | * This is how migration works: | 4751 | * This is how migration works: |
4752 | * | 4752 | * |
4753 | * 1) we invoke migration_cpu_stop() on the target CPU using | 4753 | * 1) we invoke migration_cpu_stop() on the target CPU using |
4754 | * stop_one_cpu(). | 4754 | * stop_one_cpu(). |
4755 | * 2) stopper starts to run (implicitly forcing the migrated thread | 4755 | * 2) stopper starts to run (implicitly forcing the migrated thread |
4756 | * off the CPU) | 4756 | * off the CPU) |
4757 | * 3) it checks whether the migrated task is still in the wrong runqueue. | 4757 | * 3) it checks whether the migrated task is still in the wrong runqueue. |
4758 | * 4) if it's in the wrong runqueue then the migration thread removes | 4758 | * 4) if it's in the wrong runqueue then the migration thread removes |
4759 | * it and puts it into the right queue. | 4759 | * it and puts it into the right queue. |
4760 | * 5) stopper completes and stop_one_cpu() returns and the migration | 4760 | * 5) stopper completes and stop_one_cpu() returns and the migration |
4761 | * is done. | 4761 | * is done. |
4762 | */ | 4762 | */ |
4763 | 4763 | ||
4764 | /* | 4764 | /* |
4765 | * Change a given task's CPU affinity. Migrate the thread to a | 4765 | * Change a given task's CPU affinity. Migrate the thread to a |
4766 | * proper CPU and schedule it away if the CPU it's executing on | 4766 | * proper CPU and schedule it away if the CPU it's executing on |
4767 | * is removed from the allowed bitmask. | 4767 | * is removed from the allowed bitmask. |
4768 | * | 4768 | * |
4769 | * NOTE: the caller must have a valid reference to the task, the | 4769 | * NOTE: the caller must have a valid reference to the task, the |
4770 | * task must not exit() & deallocate itself prematurely. The | 4770 | * task must not exit() & deallocate itself prematurely. The |
4771 | * call is not atomic; no spinlocks may be held. | 4771 | * call is not atomic; no spinlocks may be held. |
4772 | */ | 4772 | */ |
4773 | int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) | 4773 | int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) |
4774 | { | 4774 | { |
4775 | unsigned long flags; | 4775 | unsigned long flags; |
4776 | struct rq *rq; | 4776 | struct rq *rq; |
4777 | unsigned int dest_cpu; | 4777 | unsigned int dest_cpu; |
4778 | int ret = 0; | 4778 | int ret = 0; |
4779 | 4779 | ||
4780 | rq = task_rq_lock(p, &flags); | 4780 | rq = task_rq_lock(p, &flags); |
4781 | 4781 | ||
4782 | if (cpumask_equal(&p->cpus_allowed, new_mask)) | 4782 | if (cpumask_equal(&p->cpus_allowed, new_mask)) |
4783 | goto out; | 4783 | goto out; |
4784 | 4784 | ||
4785 | if (!cpumask_intersects(new_mask, cpu_active_mask)) { | 4785 | if (!cpumask_intersects(new_mask, cpu_active_mask)) { |
4786 | ret = -EINVAL; | 4786 | ret = -EINVAL; |
4787 | goto out; | 4787 | goto out; |
4788 | } | 4788 | } |
4789 | 4789 | ||
4790 | do_set_cpus_allowed(p, new_mask); | 4790 | do_set_cpus_allowed(p, new_mask); |
4791 | 4791 | ||
4792 | /* Can the task run on the task's current CPU? If so, we're done */ | 4792 | /* Can the task run on the task's current CPU? If so, we're done */ |
4793 | if (cpumask_test_cpu(task_cpu(p), new_mask)) | 4793 | if (cpumask_test_cpu(task_cpu(p), new_mask)) |
4794 | goto out; | 4794 | goto out; |
4795 | 4795 | ||
4796 | dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); | 4796 | dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); |
4797 | if (task_running(rq, p) || p->state == TASK_WAKING) { | 4797 | if (task_running(rq, p) || p->state == TASK_WAKING) { |
4798 | struct migration_arg arg = { p, dest_cpu }; | 4798 | struct migration_arg arg = { p, dest_cpu }; |
4799 | /* Need help from migration thread: drop lock and wait. */ | 4799 | /* Need help from migration thread: drop lock and wait. */ |
4800 | task_rq_unlock(rq, p, &flags); | 4800 | task_rq_unlock(rq, p, &flags); |
4801 | stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); | 4801 | stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); |
4802 | tlb_migrate_finish(p->mm); | 4802 | tlb_migrate_finish(p->mm); |
4803 | return 0; | 4803 | return 0; |
4804 | } else if (task_on_rq_queued(p)) | 4804 | } else if (task_on_rq_queued(p)) |
4805 | rq = move_queued_task(p, dest_cpu); | 4805 | rq = move_queued_task(p, dest_cpu); |
4806 | out: | 4806 | out: |
4807 | task_rq_unlock(rq, p, &flags); | 4807 | task_rq_unlock(rq, p, &flags); |
4808 | 4808 | ||
4809 | return ret; | 4809 | return ret; |
4810 | } | 4810 | } |
4811 | EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); | 4811 | EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); |
4812 | 4812 | ||
4813 | /* | 4813 | /* |
4814 | * Move (not current) task off this cpu, onto dest cpu. We're doing | 4814 | * Move (not current) task off this cpu, onto dest cpu. We're doing |
4815 | * this because either it can't run here any more (set_cpus_allowed() | 4815 | * this because either it can't run here any more (set_cpus_allowed() |
4816 | * away from this CPU, or CPU going down), or because we're | 4816 | * away from this CPU, or CPU going down), or because we're |
4817 | * attempting to rebalance this task on exec (sched_exec). | 4817 | * attempting to rebalance this task on exec (sched_exec). |
4818 | * | 4818 | * |
4819 | * So we race with normal scheduler movements, but that's OK, as long | 4819 | * So we race with normal scheduler movements, but that's OK, as long |
4820 | * as the task is no longer on this CPU. | 4820 | * as the task is no longer on this CPU. |
4821 | * | 4821 | * |
4822 | * Returns non-zero if task was successfully migrated. | 4822 | * Returns non-zero if task was successfully migrated. |
4823 | */ | 4823 | */ |
4824 | static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) | 4824 | static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) |
4825 | { | 4825 | { |
4826 | struct rq *rq; | 4826 | struct rq *rq; |
4827 | int ret = 0; | 4827 | int ret = 0; |
4828 | 4828 | ||
4829 | if (unlikely(!cpu_active(dest_cpu))) | 4829 | if (unlikely(!cpu_active(dest_cpu))) |
4830 | return ret; | 4830 | return ret; |
4831 | 4831 | ||
4832 | rq = cpu_rq(src_cpu); | 4832 | rq = cpu_rq(src_cpu); |
4833 | 4833 | ||
4834 | raw_spin_lock(&p->pi_lock); | 4834 | raw_spin_lock(&p->pi_lock); |
4835 | raw_spin_lock(&rq->lock); | 4835 | raw_spin_lock(&rq->lock); |
4836 | /* Already moved. */ | 4836 | /* Already moved. */ |
4837 | if (task_cpu(p) != src_cpu) | 4837 | if (task_cpu(p) != src_cpu) |
4838 | goto done; | 4838 | goto done; |
4839 | 4839 | ||
4840 | /* Affinity changed (again). */ | 4840 | /* Affinity changed (again). */ |
4841 | if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) | 4841 | if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) |
4842 | goto fail; | 4842 | goto fail; |
4843 | 4843 | ||
4844 | /* | 4844 | /* |
4845 | * If we're not on a rq, the next wake-up will ensure we're | 4845 | * If we're not on a rq, the next wake-up will ensure we're |
4846 | * placed properly. | 4846 | * placed properly. |
4847 | */ | 4847 | */ |
4848 | if (task_on_rq_queued(p)) | 4848 | if (task_on_rq_queued(p)) |
4849 | rq = move_queued_task(p, dest_cpu); | 4849 | rq = move_queued_task(p, dest_cpu); |
4850 | done: | 4850 | done: |
4851 | ret = 1; | 4851 | ret = 1; |
4852 | fail: | 4852 | fail: |
4853 | raw_spin_unlock(&rq->lock); | 4853 | raw_spin_unlock(&rq->lock); |
4854 | raw_spin_unlock(&p->pi_lock); | 4854 | raw_spin_unlock(&p->pi_lock); |
4855 | return ret; | 4855 | return ret; |
4856 | } | 4856 | } |
4857 | 4857 | ||
4858 | #ifdef CONFIG_NUMA_BALANCING | 4858 | #ifdef CONFIG_NUMA_BALANCING |
4859 | /* Migrate current task p to target_cpu */ | 4859 | /* Migrate current task p to target_cpu */ |
4860 | int migrate_task_to(struct task_struct *p, int target_cpu) | 4860 | int migrate_task_to(struct task_struct *p, int target_cpu) |
4861 | { | 4861 | { |
4862 | struct migration_arg arg = { p, target_cpu }; | 4862 | struct migration_arg arg = { p, target_cpu }; |
4863 | int curr_cpu = task_cpu(p); | 4863 | int curr_cpu = task_cpu(p); |
4864 | 4864 | ||
4865 | if (curr_cpu == target_cpu) | 4865 | if (curr_cpu == target_cpu) |
4866 | return 0; | 4866 | return 0; |
4867 | 4867 | ||
4868 | if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) | 4868 | if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) |
4869 | return -EINVAL; | 4869 | return -EINVAL; |
4870 | 4870 | ||
4871 | /* TODO: This is not properly updating schedstats */ | 4871 | /* TODO: This is not properly updating schedstats */ |
4872 | 4872 | ||
4873 | trace_sched_move_numa(p, curr_cpu, target_cpu); | 4873 | trace_sched_move_numa(p, curr_cpu, target_cpu); |
4874 | return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); | 4874 | return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); |
4875 | } | 4875 | } |
4876 | 4876 | ||
4877 | /* | 4877 | /* |
4878 | * Requeue a task on a given node and accurately track the number of NUMA | 4878 | * Requeue a task on a given node and accurately track the number of NUMA |
4879 | * tasks on the runqueues | 4879 | * tasks on the runqueues |
4880 | */ | 4880 | */ |
4881 | void sched_setnuma(struct task_struct *p, int nid) | 4881 | void sched_setnuma(struct task_struct *p, int nid) |
4882 | { | 4882 | { |
4883 | struct rq *rq; | 4883 | struct rq *rq; |
4884 | unsigned long flags; | 4884 | unsigned long flags; |
4885 | bool queued, running; | 4885 | bool queued, running; |
4886 | 4886 | ||
4887 | rq = task_rq_lock(p, &flags); | 4887 | rq = task_rq_lock(p, &flags); |
4888 | queued = task_on_rq_queued(p); | 4888 | queued = task_on_rq_queued(p); |
4889 | running = task_current(rq, p); | 4889 | running = task_current(rq, p); |
4890 | 4890 | ||
4891 | if (queued) | 4891 | if (queued) |
4892 | dequeue_task(rq, p, 0); | 4892 | dequeue_task(rq, p, 0); |
4893 | if (running) | 4893 | if (running) |
4894 | put_prev_task(rq, p); | 4894 | put_prev_task(rq, p); |
4895 | 4895 | ||
4896 | p->numa_preferred_nid = nid; | 4896 | p->numa_preferred_nid = nid; |
4897 | 4897 | ||
4898 | if (running) | 4898 | if (running) |
4899 | p->sched_class->set_curr_task(rq); | 4899 | p->sched_class->set_curr_task(rq); |
4900 | if (queued) | 4900 | if (queued) |
4901 | enqueue_task(rq, p, 0); | 4901 | enqueue_task(rq, p, 0); |
4902 | task_rq_unlock(rq, p, &flags); | 4902 | task_rq_unlock(rq, p, &flags); |
4903 | } | 4903 | } |
4904 | #endif | 4904 | #endif |
4905 | 4905 | ||
4906 | /* | 4906 | /* |
4907 | * migration_cpu_stop - this will be executed by a highprio stopper thread | 4907 | * migration_cpu_stop - this will be executed by a highprio stopper thread |
4908 | * and performs thread migration by bumping thread off CPU then | 4908 | * and performs thread migration by bumping thread off CPU then |
4909 | * 'pushing' onto another runqueue. | 4909 | * 'pushing' onto another runqueue. |
4910 | */ | 4910 | */ |
4911 | static int migration_cpu_stop(void *data) | 4911 | static int migration_cpu_stop(void *data) |
4912 | { | 4912 | { |
4913 | struct migration_arg *arg = data; | 4913 | struct migration_arg *arg = data; |
4914 | 4914 | ||
4915 | /* | 4915 | /* |
4916 | * The original target cpu might have gone down and we might | 4916 | * The original target cpu might have gone down and we might |
4917 | * be on another cpu but it doesn't matter. | 4917 | * be on another cpu but it doesn't matter. |
4918 | */ | 4918 | */ |
4919 | local_irq_disable(); | 4919 | local_irq_disable(); |
4920 | /* | 4920 | /* |
4921 | * We need to explicitly wake pending tasks before running | 4921 | * We need to explicitly wake pending tasks before running |
4922 | * __migrate_task() such that we will not miss enforcing cpus_allowed | 4922 | * __migrate_task() such that we will not miss enforcing cpus_allowed |
4923 | * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test. | 4923 | * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test. |
4924 | */ | 4924 | */ |
4925 | sched_ttwu_pending(); | 4925 | sched_ttwu_pending(); |
4926 | __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); | 4926 | __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); |
4927 | local_irq_enable(); | 4927 | local_irq_enable(); |
4928 | return 0; | 4928 | return 0; |
4929 | } | 4929 | } |
4930 | 4930 | ||
4931 | #ifdef CONFIG_HOTPLUG_CPU | 4931 | #ifdef CONFIG_HOTPLUG_CPU |
4932 | 4932 | ||
4933 | /* | 4933 | /* |
4934 | * Ensures that the idle task is using init_mm right before its cpu goes | 4934 | * Ensures that the idle task is using init_mm right before its cpu goes |
4935 | * offline. | 4935 | * offline. |
4936 | */ | 4936 | */ |
4937 | void idle_task_exit(void) | 4937 | void idle_task_exit(void) |
4938 | { | 4938 | { |
4939 | struct mm_struct *mm = current->active_mm; | 4939 | struct mm_struct *mm = current->active_mm; |
4940 | 4940 | ||
4941 | BUG_ON(cpu_online(smp_processor_id())); | 4941 | BUG_ON(cpu_online(smp_processor_id())); |
4942 | 4942 | ||
4943 | if (mm != &init_mm) { | 4943 | if (mm != &init_mm) { |
4944 | switch_mm(mm, &init_mm, current); | 4944 | switch_mm(mm, &init_mm, current); |
4945 | finish_arch_post_lock_switch(); | 4945 | finish_arch_post_lock_switch(); |
4946 | } | 4946 | } |
4947 | mmdrop(mm); | 4947 | mmdrop(mm); |
4948 | } | 4948 | } |
4949 | 4949 | ||
4950 | /* | 4950 | /* |
4951 | * Since this CPU is going 'away' for a while, fold any nr_active delta | 4951 | * Since this CPU is going 'away' for a while, fold any nr_active delta |
4952 | * we might have. Assumes we're called after migrate_tasks() so that the | 4952 | * we might have. Assumes we're called after migrate_tasks() so that the |
4953 | * nr_active count is stable. | 4953 | * nr_active count is stable. |
4954 | * | 4954 | * |
4955 | * Also see the comment "Global load-average calculations". | 4955 | * Also see the comment "Global load-average calculations". |
4956 | */ | 4956 | */ |
4957 | static void calc_load_migrate(struct rq *rq) | 4957 | static void calc_load_migrate(struct rq *rq) |
4958 | { | 4958 | { |
4959 | long delta = calc_load_fold_active(rq); | 4959 | long delta = calc_load_fold_active(rq); |
4960 | if (delta) | 4960 | if (delta) |
4961 | atomic_long_add(delta, &calc_load_tasks); | 4961 | atomic_long_add(delta, &calc_load_tasks); |
4962 | } | 4962 | } |
4963 | 4963 | ||
4964 | static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) | 4964 | static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) |
4965 | { | 4965 | { |
4966 | } | 4966 | } |
4967 | 4967 | ||
4968 | static const struct sched_class fake_sched_class = { | 4968 | static const struct sched_class fake_sched_class = { |
4969 | .put_prev_task = put_prev_task_fake, | 4969 | .put_prev_task = put_prev_task_fake, |
4970 | }; | 4970 | }; |
4971 | 4971 | ||
4972 | static struct task_struct fake_task = { | 4972 | static struct task_struct fake_task = { |
4973 | /* | 4973 | /* |
4974 | * Avoid pull_{rt,dl}_task() | 4974 | * Avoid pull_{rt,dl}_task() |
4975 | */ | 4975 | */ |
4976 | .prio = MAX_PRIO + 1, | 4976 | .prio = MAX_PRIO + 1, |
4977 | .sched_class = &fake_sched_class, | 4977 | .sched_class = &fake_sched_class, |
4978 | }; | 4978 | }; |
4979 | 4979 | ||
4980 | /* | 4980 | /* |
4981 | * Migrate all tasks from the rq, sleeping tasks will be migrated by | 4981 | * Migrate all tasks from the rq, sleeping tasks will be migrated by |
4982 | * try_to_wake_up()->select_task_rq(). | 4982 | * try_to_wake_up()->select_task_rq(). |
4983 | * | 4983 | * |
4984 | * Called with rq->lock held even though we'er in stop_machine() and | 4984 | * Called with rq->lock held even though we'er in stop_machine() and |
4985 | * there's no concurrency possible, we hold the required locks anyway | 4985 | * there's no concurrency possible, we hold the required locks anyway |
4986 | * because of lock validation efforts. | 4986 | * because of lock validation efforts. |
4987 | */ | 4987 | */ |
4988 | static void migrate_tasks(unsigned int dead_cpu) | 4988 | static void migrate_tasks(unsigned int dead_cpu) |
4989 | { | 4989 | { |
4990 | struct rq *rq = cpu_rq(dead_cpu); | 4990 | struct rq *rq = cpu_rq(dead_cpu); |
4991 | struct task_struct *next, *stop = rq->stop; | 4991 | struct task_struct *next, *stop = rq->stop; |
4992 | int dest_cpu; | 4992 | int dest_cpu; |
4993 | 4993 | ||
4994 | /* | 4994 | /* |
4995 | * Fudge the rq selection such that the below task selection loop | 4995 | * Fudge the rq selection such that the below task selection loop |
4996 | * doesn't get stuck on the currently eligible stop task. | 4996 | * doesn't get stuck on the currently eligible stop task. |
4997 | * | 4997 | * |
4998 | * We're currently inside stop_machine() and the rq is either stuck | 4998 | * We're currently inside stop_machine() and the rq is either stuck |
4999 | * in the stop_machine_cpu_stop() loop, or we're executing this code, | 4999 | * in the stop_machine_cpu_stop() loop, or we're executing this code, |
5000 | * either way we should never end up calling schedule() until we're | 5000 | * either way we should never end up calling schedule() until we're |
5001 | * done here. | 5001 | * done here. |
5002 | */ | 5002 | */ |
5003 | rq->stop = NULL; | 5003 | rq->stop = NULL; |
5004 | 5004 | ||
5005 | /* | 5005 | /* |
5006 | * put_prev_task() and pick_next_task() sched | 5006 | * put_prev_task() and pick_next_task() sched |
5007 | * class method both need to have an up-to-date | 5007 | * class method both need to have an up-to-date |
5008 | * value of rq->clock[_task] | 5008 | * value of rq->clock[_task] |
5009 | */ | 5009 | */ |
5010 | update_rq_clock(rq); | 5010 | update_rq_clock(rq); |
5011 | 5011 | ||
5012 | for ( ; ; ) { | 5012 | for ( ; ; ) { |
5013 | /* | 5013 | /* |
5014 | * There's this thread running, bail when that's the only | 5014 | * There's this thread running, bail when that's the only |
5015 | * remaining thread. | 5015 | * remaining thread. |
5016 | */ | 5016 | */ |
5017 | if (rq->nr_running == 1) | 5017 | if (rq->nr_running == 1) |
5018 | break; | 5018 | break; |
5019 | 5019 | ||
5020 | next = pick_next_task(rq, &fake_task); | 5020 | next = pick_next_task(rq, &fake_task); |
5021 | BUG_ON(!next); | 5021 | BUG_ON(!next); |
5022 | next->sched_class->put_prev_task(rq, next); | 5022 | next->sched_class->put_prev_task(rq, next); |
5023 | 5023 | ||
5024 | /* Find suitable destination for @next, with force if needed. */ | 5024 | /* Find suitable destination for @next, with force if needed. */ |
5025 | dest_cpu = select_fallback_rq(dead_cpu, next); | 5025 | dest_cpu = select_fallback_rq(dead_cpu, next); |
5026 | raw_spin_unlock(&rq->lock); | 5026 | raw_spin_unlock(&rq->lock); |
5027 | 5027 | ||
5028 | __migrate_task(next, dead_cpu, dest_cpu); | 5028 | __migrate_task(next, dead_cpu, dest_cpu); |
5029 | 5029 | ||
5030 | raw_spin_lock(&rq->lock); | 5030 | raw_spin_lock(&rq->lock); |
5031 | } | 5031 | } |
5032 | 5032 | ||
5033 | rq->stop = stop; | 5033 | rq->stop = stop; |
5034 | } | 5034 | } |
5035 | 5035 | ||
5036 | #endif /* CONFIG_HOTPLUG_CPU */ | 5036 | #endif /* CONFIG_HOTPLUG_CPU */ |
5037 | 5037 | ||
5038 | #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) | 5038 | #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) |
5039 | 5039 | ||
5040 | static struct ctl_table sd_ctl_dir[] = { | 5040 | static struct ctl_table sd_ctl_dir[] = { |
5041 | { | 5041 | { |
5042 | .procname = "sched_domain", | 5042 | .procname = "sched_domain", |
5043 | .mode = 0555, | 5043 | .mode = 0555, |
5044 | }, | 5044 | }, |
5045 | {} | 5045 | {} |
5046 | }; | 5046 | }; |
5047 | 5047 | ||
5048 | static struct ctl_table sd_ctl_root[] = { | 5048 | static struct ctl_table sd_ctl_root[] = { |
5049 | { | 5049 | { |
5050 | .procname = "kernel", | 5050 | .procname = "kernel", |
5051 | .mode = 0555, | 5051 | .mode = 0555, |
5052 | .child = sd_ctl_dir, | 5052 | .child = sd_ctl_dir, |
5053 | }, | 5053 | }, |
5054 | {} | 5054 | {} |
5055 | }; | 5055 | }; |
5056 | 5056 | ||
5057 | static struct ctl_table *sd_alloc_ctl_entry(int n) | 5057 | static struct ctl_table *sd_alloc_ctl_entry(int n) |
5058 | { | 5058 | { |
5059 | struct ctl_table *entry = | 5059 | struct ctl_table *entry = |
5060 | kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); | 5060 | kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); |
5061 | 5061 | ||
5062 | return entry; | 5062 | return entry; |
5063 | } | 5063 | } |
5064 | 5064 | ||
5065 | static void sd_free_ctl_entry(struct ctl_table **tablep) | 5065 | static void sd_free_ctl_entry(struct ctl_table **tablep) |
5066 | { | 5066 | { |
5067 | struct ctl_table *entry; | 5067 | struct ctl_table *entry; |
5068 | 5068 | ||
5069 | /* | 5069 | /* |
5070 | * In the intermediate directories, both the child directory and | 5070 | * In the intermediate directories, both the child directory and |
5071 | * procname are dynamically allocated and could fail but the mode | 5071 | * procname are dynamically allocated and could fail but the mode |
5072 | * will always be set. In the lowest directory the names are | 5072 | * will always be set. In the lowest directory the names are |
5073 | * static strings and all have proc handlers. | 5073 | * static strings and all have proc handlers. |
5074 | */ | 5074 | */ |
5075 | for (entry = *tablep; entry->mode; entry++) { | 5075 | for (entry = *tablep; entry->mode; entry++) { |
5076 | if (entry->child) | 5076 | if (entry->child) |
5077 | sd_free_ctl_entry(&entry->child); | 5077 | sd_free_ctl_entry(&entry->child); |
5078 | if (entry->proc_handler == NULL) | 5078 | if (entry->proc_handler == NULL) |
5079 | kfree(entry->procname); | 5079 | kfree(entry->procname); |
5080 | } | 5080 | } |
5081 | 5081 | ||
5082 | kfree(*tablep); | 5082 | kfree(*tablep); |
5083 | *tablep = NULL; | 5083 | *tablep = NULL; |
5084 | } | 5084 | } |
5085 | 5085 | ||
5086 | static int min_load_idx = 0; | 5086 | static int min_load_idx = 0; |
5087 | static int max_load_idx = CPU_LOAD_IDX_MAX-1; | 5087 | static int max_load_idx = CPU_LOAD_IDX_MAX-1; |
5088 | 5088 | ||
5089 | static void | 5089 | static void |
5090 | set_table_entry(struct ctl_table *entry, | 5090 | set_table_entry(struct ctl_table *entry, |
5091 | const char *procname, void *data, int maxlen, | 5091 | const char *procname, void *data, int maxlen, |
5092 | umode_t mode, proc_handler *proc_handler, | 5092 | umode_t mode, proc_handler *proc_handler, |
5093 | bool load_idx) | 5093 | bool load_idx) |
5094 | { | 5094 | { |
5095 | entry->procname = procname; | 5095 | entry->procname = procname; |
5096 | entry->data = data; | 5096 | entry->data = data; |
5097 | entry->maxlen = maxlen; | 5097 | entry->maxlen = maxlen; |
5098 | entry->mode = mode; | 5098 | entry->mode = mode; |
5099 | entry->proc_handler = proc_handler; | 5099 | entry->proc_handler = proc_handler; |
5100 | 5100 | ||
5101 | if (load_idx) { | 5101 | if (load_idx) { |
5102 | entry->extra1 = &min_load_idx; | 5102 | entry->extra1 = &min_load_idx; |
5103 | entry->extra2 = &max_load_idx; | 5103 | entry->extra2 = &max_load_idx; |
5104 | } | 5104 | } |
5105 | } | 5105 | } |
5106 | 5106 | ||
5107 | static struct ctl_table * | 5107 | static struct ctl_table * |
5108 | sd_alloc_ctl_domain_table(struct sched_domain *sd) | 5108 | sd_alloc_ctl_domain_table(struct sched_domain *sd) |
5109 | { | 5109 | { |
5110 | struct ctl_table *table = sd_alloc_ctl_entry(14); | 5110 | struct ctl_table *table = sd_alloc_ctl_entry(14); |
5111 | 5111 | ||
5112 | if (table == NULL) | 5112 | if (table == NULL) |
5113 | return NULL; | 5113 | return NULL; |
5114 | 5114 | ||
5115 | set_table_entry(&table[0], "min_interval", &sd->min_interval, | 5115 | set_table_entry(&table[0], "min_interval", &sd->min_interval, |
5116 | sizeof(long), 0644, proc_doulongvec_minmax, false); | 5116 | sizeof(long), 0644, proc_doulongvec_minmax, false); |
5117 | set_table_entry(&table[1], "max_interval", &sd->max_interval, | 5117 | set_table_entry(&table[1], "max_interval", &sd->max_interval, |
5118 | sizeof(long), 0644, proc_doulongvec_minmax, false); | 5118 | sizeof(long), 0644, proc_doulongvec_minmax, false); |
5119 | set_table_entry(&table[2], "busy_idx", &sd->busy_idx, | 5119 | set_table_entry(&table[2], "busy_idx", &sd->busy_idx, |
5120 | sizeof(int), 0644, proc_dointvec_minmax, true); | 5120 | sizeof(int), 0644, proc_dointvec_minmax, true); |
5121 | set_table_entry(&table[3], "idle_idx", &sd->idle_idx, | 5121 | set_table_entry(&table[3], "idle_idx", &sd->idle_idx, |
5122 | sizeof(int), 0644, proc_dointvec_minmax, true); | 5122 | sizeof(int), 0644, proc_dointvec_minmax, true); |
5123 | set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, | 5123 | set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, |
5124 | sizeof(int), 0644, proc_dointvec_minmax, true); | 5124 | sizeof(int), 0644, proc_dointvec_minmax, true); |
5125 | set_table_entry(&table[5], "wake_idx", &sd->wake_idx, | 5125 | set_table_entry(&table[5], "wake_idx", &sd->wake_idx, |
5126 | sizeof(int), 0644, proc_dointvec_minmax, true); | 5126 | sizeof(int), 0644, proc_dointvec_minmax, true); |
5127 | set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, | 5127 | set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, |
5128 | sizeof(int), 0644, proc_dointvec_minmax, true); | 5128 | sizeof(int), 0644, proc_dointvec_minmax, true); |
5129 | set_table_entry(&table[7], "busy_factor", &sd->busy_factor, | 5129 | set_table_entry(&table[7], "busy_factor", &sd->busy_factor, |
5130 | sizeof(int), 0644, proc_dointvec_minmax, false); | 5130 | sizeof(int), 0644, proc_dointvec_minmax, false); |
5131 | set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, | 5131 | set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, |
5132 | sizeof(int), 0644, proc_dointvec_minmax, false); | 5132 | sizeof(int), 0644, proc_dointvec_minmax, false); |
5133 | set_table_entry(&table[9], "cache_nice_tries", | 5133 | set_table_entry(&table[9], "cache_nice_tries", |
5134 | &sd->cache_nice_tries, | 5134 | &sd->cache_nice_tries, |
5135 | sizeof(int), 0644, proc_dointvec_minmax, false); | 5135 | sizeof(int), 0644, proc_dointvec_minmax, false); |
5136 | set_table_entry(&table[10], "flags", &sd->flags, | 5136 | set_table_entry(&table[10], "flags", &sd->flags, |
5137 | sizeof(int), 0644, proc_dointvec_minmax, false); | 5137 | sizeof(int), 0644, proc_dointvec_minmax, false); |
5138 | set_table_entry(&table[11], "max_newidle_lb_cost", | 5138 | set_table_entry(&table[11], "max_newidle_lb_cost", |
5139 | &sd->max_newidle_lb_cost, | 5139 | &sd->max_newidle_lb_cost, |
5140 | sizeof(long), 0644, proc_doulongvec_minmax, false); | 5140 | sizeof(long), 0644, proc_doulongvec_minmax, false); |
5141 | set_table_entry(&table[12], "name", sd->name, | 5141 | set_table_entry(&table[12], "name", sd->name, |
5142 | CORENAME_MAX_SIZE, 0444, proc_dostring, false); | 5142 | CORENAME_MAX_SIZE, 0444, proc_dostring, false); |
5143 | /* &table[13] is terminator */ | 5143 | /* &table[13] is terminator */ |
5144 | 5144 | ||
5145 | return table; | 5145 | return table; |
5146 | } | 5146 | } |
5147 | 5147 | ||
5148 | static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) | 5148 | static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) |
5149 | { | 5149 | { |
5150 | struct ctl_table *entry, *table; | 5150 | struct ctl_table *entry, *table; |
5151 | struct sched_domain *sd; | 5151 | struct sched_domain *sd; |
5152 | int domain_num = 0, i; | 5152 | int domain_num = 0, i; |
5153 | char buf[32]; | 5153 | char buf[32]; |
5154 | 5154 | ||
5155 | for_each_domain(cpu, sd) | 5155 | for_each_domain(cpu, sd) |
5156 | domain_num++; | 5156 | domain_num++; |
5157 | entry = table = sd_alloc_ctl_entry(domain_num + 1); | 5157 | entry = table = sd_alloc_ctl_entry(domain_num + 1); |
5158 | if (table == NULL) | 5158 | if (table == NULL) |
5159 | return NULL; | 5159 | return NULL; |
5160 | 5160 | ||
5161 | i = 0; | 5161 | i = 0; |
5162 | for_each_domain(cpu, sd) { | 5162 | for_each_domain(cpu, sd) { |
5163 | snprintf(buf, 32, "domain%d", i); | 5163 | snprintf(buf, 32, "domain%d", i); |
5164 | entry->procname = kstrdup(buf, GFP_KERNEL); | 5164 | entry->procname = kstrdup(buf, GFP_KERNEL); |
5165 | entry->mode = 0555; | 5165 | entry->mode = 0555; |
5166 | entry->child = sd_alloc_ctl_domain_table(sd); | 5166 | entry->child = sd_alloc_ctl_domain_table(sd); |
5167 | entry++; | 5167 | entry++; |
5168 | i++; | 5168 | i++; |
5169 | } | 5169 | } |
5170 | return table; | 5170 | return table; |
5171 | } | 5171 | } |
5172 | 5172 | ||
5173 | static struct ctl_table_header *sd_sysctl_header; | 5173 | static struct ctl_table_header *sd_sysctl_header; |
5174 | static void register_sched_domain_sysctl(void) | 5174 | static void register_sched_domain_sysctl(void) |
5175 | { | 5175 | { |
5176 | int i, cpu_num = num_possible_cpus(); | 5176 | int i, cpu_num = num_possible_cpus(); |
5177 | struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); | 5177 | struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); |
5178 | char buf[32]; | 5178 | char buf[32]; |
5179 | 5179 | ||
5180 | WARN_ON(sd_ctl_dir[0].child); | 5180 | WARN_ON(sd_ctl_dir[0].child); |
5181 | sd_ctl_dir[0].child = entry; | 5181 | sd_ctl_dir[0].child = entry; |
5182 | 5182 | ||
5183 | if (entry == NULL) | 5183 | if (entry == NULL) |
5184 | return; | 5184 | return; |
5185 | 5185 | ||
5186 | for_each_possible_cpu(i) { | 5186 | for_each_possible_cpu(i) { |
5187 | snprintf(buf, 32, "cpu%d", i); | 5187 | snprintf(buf, 32, "cpu%d", i); |
5188 | entry->procname = kstrdup(buf, GFP_KERNEL); | 5188 | entry->procname = kstrdup(buf, GFP_KERNEL); |
5189 | entry->mode = 0555; | 5189 | entry->mode = 0555; |
5190 | entry->child = sd_alloc_ctl_cpu_table(i); | 5190 | entry->child = sd_alloc_ctl_cpu_table(i); |
5191 | entry++; | 5191 | entry++; |
5192 | } | 5192 | } |
5193 | 5193 | ||
5194 | WARN_ON(sd_sysctl_header); | 5194 | WARN_ON(sd_sysctl_header); |
5195 | sd_sysctl_header = register_sysctl_table(sd_ctl_root); | 5195 | sd_sysctl_header = register_sysctl_table(sd_ctl_root); |
5196 | } | 5196 | } |
5197 | 5197 | ||
5198 | /* may be called multiple times per register */ | 5198 | /* may be called multiple times per register */ |
5199 | static void unregister_sched_domain_sysctl(void) | 5199 | static void unregister_sched_domain_sysctl(void) |
5200 | { | 5200 | { |
5201 | if (sd_sysctl_header) | 5201 | if (sd_sysctl_header) |
5202 | unregister_sysctl_table(sd_sysctl_header); | 5202 | unregister_sysctl_table(sd_sysctl_header); |
5203 | sd_sysctl_header = NULL; | 5203 | sd_sysctl_header = NULL; |
5204 | if (sd_ctl_dir[0].child) | 5204 | if (sd_ctl_dir[0].child) |
5205 | sd_free_ctl_entry(&sd_ctl_dir[0].child); | 5205 | sd_free_ctl_entry(&sd_ctl_dir[0].child); |
5206 | } | 5206 | } |
5207 | #else | 5207 | #else |
5208 | static void register_sched_domain_sysctl(void) | 5208 | static void register_sched_domain_sysctl(void) |
5209 | { | 5209 | { |
5210 | } | 5210 | } |
5211 | static void unregister_sched_domain_sysctl(void) | 5211 | static void unregister_sched_domain_sysctl(void) |
5212 | { | 5212 | { |
5213 | } | 5213 | } |
5214 | #endif | 5214 | #endif |
5215 | 5215 | ||
5216 | static void set_rq_online(struct rq *rq) | 5216 | static void set_rq_online(struct rq *rq) |
5217 | { | 5217 | { |
5218 | if (!rq->online) { | 5218 | if (!rq->online) { |
5219 | const struct sched_class *class; | 5219 | const struct sched_class *class; |
5220 | 5220 | ||
5221 | cpumask_set_cpu(rq->cpu, rq->rd->online); | 5221 | cpumask_set_cpu(rq->cpu, rq->rd->online); |
5222 | rq->online = 1; | 5222 | rq->online = 1; |
5223 | 5223 | ||
5224 | for_each_class(class) { | 5224 | for_each_class(class) { |
5225 | if (class->rq_online) | 5225 | if (class->rq_online) |
5226 | class->rq_online(rq); | 5226 | class->rq_online(rq); |
5227 | } | 5227 | } |
5228 | } | 5228 | } |
5229 | } | 5229 | } |
5230 | 5230 | ||
5231 | static void set_rq_offline(struct rq *rq) | 5231 | static void set_rq_offline(struct rq *rq) |
5232 | { | 5232 | { |
5233 | if (rq->online) { | 5233 | if (rq->online) { |
5234 | const struct sched_class *class; | 5234 | const struct sched_class *class; |
5235 | 5235 | ||
5236 | for_each_class(class) { | 5236 | for_each_class(class) { |
5237 | if (class->rq_offline) | 5237 | if (class->rq_offline) |
5238 | class->rq_offline(rq); | 5238 | class->rq_offline(rq); |
5239 | } | 5239 | } |
5240 | 5240 | ||
5241 | cpumask_clear_cpu(rq->cpu, rq->rd->online); | 5241 | cpumask_clear_cpu(rq->cpu, rq->rd->online); |
5242 | rq->online = 0; | 5242 | rq->online = 0; |
5243 | } | 5243 | } |
5244 | } | 5244 | } |
5245 | 5245 | ||
5246 | /* | 5246 | /* |
5247 | * migration_call - callback that gets triggered when a CPU is added. | 5247 | * migration_call - callback that gets triggered when a CPU is added. |
5248 | * Here we can start up the necessary migration thread for the new CPU. | 5248 | * Here we can start up the necessary migration thread for the new CPU. |
5249 | */ | 5249 | */ |
5250 | static int | 5250 | static int |
5251 | migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) | 5251 | migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) |
5252 | { | 5252 | { |
5253 | int cpu = (long)hcpu; | 5253 | int cpu = (long)hcpu; |
5254 | unsigned long flags; | 5254 | unsigned long flags; |
5255 | struct rq *rq = cpu_rq(cpu); | 5255 | struct rq *rq = cpu_rq(cpu); |
5256 | 5256 | ||
5257 | switch (action & ~CPU_TASKS_FROZEN) { | 5257 | switch (action & ~CPU_TASKS_FROZEN) { |
5258 | 5258 | ||
5259 | case CPU_UP_PREPARE: | 5259 | case CPU_UP_PREPARE: |
5260 | rq->calc_load_update = calc_load_update; | 5260 | rq->calc_load_update = calc_load_update; |
5261 | break; | 5261 | break; |
5262 | 5262 | ||
5263 | case CPU_ONLINE: | 5263 | case CPU_ONLINE: |
5264 | /* Update our root-domain */ | 5264 | /* Update our root-domain */ |
5265 | raw_spin_lock_irqsave(&rq->lock, flags); | 5265 | raw_spin_lock_irqsave(&rq->lock, flags); |
5266 | if (rq->rd) { | 5266 | if (rq->rd) { |
5267 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); | 5267 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); |
5268 | 5268 | ||
5269 | set_rq_online(rq); | 5269 | set_rq_online(rq); |
5270 | } | 5270 | } |
5271 | raw_spin_unlock_irqrestore(&rq->lock, flags); | 5271 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
5272 | break; | 5272 | break; |
5273 | 5273 | ||
5274 | #ifdef CONFIG_HOTPLUG_CPU | 5274 | #ifdef CONFIG_HOTPLUG_CPU |
5275 | case CPU_DYING: | 5275 | case CPU_DYING: |
5276 | sched_ttwu_pending(); | 5276 | sched_ttwu_pending(); |
5277 | /* Update our root-domain */ | 5277 | /* Update our root-domain */ |
5278 | raw_spin_lock_irqsave(&rq->lock, flags); | 5278 | raw_spin_lock_irqsave(&rq->lock, flags); |
5279 | if (rq->rd) { | 5279 | if (rq->rd) { |
5280 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); | 5280 | BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); |
5281 | set_rq_offline(rq); | 5281 | set_rq_offline(rq); |
5282 | } | 5282 | } |
5283 | migrate_tasks(cpu); | 5283 | migrate_tasks(cpu); |
5284 | BUG_ON(rq->nr_running != 1); /* the migration thread */ | 5284 | BUG_ON(rq->nr_running != 1); /* the migration thread */ |
5285 | raw_spin_unlock_irqrestore(&rq->lock, flags); | 5285 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
5286 | break; | 5286 | break; |
5287 | 5287 | ||
5288 | case CPU_DEAD: | 5288 | case CPU_DEAD: |
5289 | calc_load_migrate(rq); | 5289 | calc_load_migrate(rq); |
5290 | break; | 5290 | break; |
5291 | #endif | 5291 | #endif |
5292 | } | 5292 | } |
5293 | 5293 | ||
5294 | update_max_interval(); | 5294 | update_max_interval(); |
5295 | 5295 | ||
5296 | return NOTIFY_OK; | 5296 | return NOTIFY_OK; |
5297 | } | 5297 | } |
5298 | 5298 | ||
5299 | /* | 5299 | /* |
5300 | * Register at high priority so that task migration (migrate_all_tasks) | 5300 | * Register at high priority so that task migration (migrate_all_tasks) |
5301 | * happens before everything else. This has to be lower priority than | 5301 | * happens before everything else. This has to be lower priority than |
5302 | * the notifier in the perf_event subsystem, though. | 5302 | * the notifier in the perf_event subsystem, though. |
5303 | */ | 5303 | */ |
5304 | static struct notifier_block migration_notifier = { | 5304 | static struct notifier_block migration_notifier = { |
5305 | .notifier_call = migration_call, | 5305 | .notifier_call = migration_call, |
5306 | .priority = CPU_PRI_MIGRATION, | 5306 | .priority = CPU_PRI_MIGRATION, |
5307 | }; | 5307 | }; |
5308 | 5308 | ||
5309 | static void __cpuinit set_cpu_rq_start_time(void) | 5309 | static void __cpuinit set_cpu_rq_start_time(void) |
5310 | { | 5310 | { |
5311 | int cpu = smp_processor_id(); | 5311 | int cpu = smp_processor_id(); |
5312 | struct rq *rq = cpu_rq(cpu); | 5312 | struct rq *rq = cpu_rq(cpu); |
5313 | rq->age_stamp = sched_clock_cpu(cpu); | 5313 | rq->age_stamp = sched_clock_cpu(cpu); |
5314 | } | 5314 | } |
5315 | 5315 | ||
5316 | static int sched_cpu_active(struct notifier_block *nfb, | 5316 | static int sched_cpu_active(struct notifier_block *nfb, |
5317 | unsigned long action, void *hcpu) | 5317 | unsigned long action, void *hcpu) |
5318 | { | 5318 | { |
5319 | switch (action & ~CPU_TASKS_FROZEN) { | 5319 | switch (action & ~CPU_TASKS_FROZEN) { |
5320 | case CPU_STARTING: | 5320 | case CPU_STARTING: |
5321 | set_cpu_rq_start_time(); | 5321 | set_cpu_rq_start_time(); |
5322 | return NOTIFY_OK; | 5322 | return NOTIFY_OK; |
5323 | case CPU_DOWN_FAILED: | 5323 | case CPU_DOWN_FAILED: |
5324 | set_cpu_active((long)hcpu, true); | 5324 | set_cpu_active((long)hcpu, true); |
5325 | return NOTIFY_OK; | 5325 | return NOTIFY_OK; |
5326 | default: | 5326 | default: |
5327 | return NOTIFY_DONE; | 5327 | return NOTIFY_DONE; |
5328 | } | 5328 | } |
5329 | } | 5329 | } |
5330 | 5330 | ||
5331 | static int sched_cpu_inactive(struct notifier_block *nfb, | 5331 | static int sched_cpu_inactive(struct notifier_block *nfb, |
5332 | unsigned long action, void *hcpu) | 5332 | unsigned long action, void *hcpu) |
5333 | { | 5333 | { |
5334 | unsigned long flags; | 5334 | unsigned long flags; |
5335 | long cpu = (long)hcpu; | 5335 | long cpu = (long)hcpu; |
5336 | struct dl_bw *dl_b; | 5336 | struct dl_bw *dl_b; |
5337 | 5337 | ||
5338 | switch (action & ~CPU_TASKS_FROZEN) { | 5338 | switch (action & ~CPU_TASKS_FROZEN) { |
5339 | case CPU_DOWN_PREPARE: | 5339 | case CPU_DOWN_PREPARE: |
5340 | set_cpu_active(cpu, false); | 5340 | set_cpu_active(cpu, false); |
5341 | 5341 | ||
5342 | /* explicitly allow suspend */ | 5342 | /* explicitly allow suspend */ |
5343 | if (!(action & CPU_TASKS_FROZEN)) { | 5343 | if (!(action & CPU_TASKS_FROZEN)) { |
5344 | bool overflow; | 5344 | bool overflow; |
5345 | int cpus; | 5345 | int cpus; |
5346 | 5346 | ||
5347 | rcu_read_lock_sched(); | 5347 | rcu_read_lock_sched(); |
5348 | dl_b = dl_bw_of(cpu); | 5348 | dl_b = dl_bw_of(cpu); |
5349 | 5349 | ||
5350 | raw_spin_lock_irqsave(&dl_b->lock, flags); | 5350 | raw_spin_lock_irqsave(&dl_b->lock, flags); |
5351 | cpus = dl_bw_cpus(cpu); | 5351 | cpus = dl_bw_cpus(cpu); |
5352 | overflow = __dl_overflow(dl_b, cpus, 0, 0); | 5352 | overflow = __dl_overflow(dl_b, cpus, 0, 0); |
5353 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); | 5353 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); |
5354 | 5354 | ||
5355 | rcu_read_unlock_sched(); | 5355 | rcu_read_unlock_sched(); |
5356 | 5356 | ||
5357 | if (overflow) | 5357 | if (overflow) |
5358 | return notifier_from_errno(-EBUSY); | 5358 | return notifier_from_errno(-EBUSY); |
5359 | } | 5359 | } |
5360 | return NOTIFY_OK; | 5360 | return NOTIFY_OK; |
5361 | } | 5361 | } |
5362 | 5362 | ||
5363 | return NOTIFY_DONE; | 5363 | return NOTIFY_DONE; |
5364 | } | 5364 | } |
5365 | 5365 | ||
5366 | static int __init migration_init(void) | 5366 | static int __init migration_init(void) |
5367 | { | 5367 | { |
5368 | void *cpu = (void *)(long)smp_processor_id(); | 5368 | void *cpu = (void *)(long)smp_processor_id(); |
5369 | int err; | 5369 | int err; |
5370 | 5370 | ||
5371 | /* Initialize migration for the boot CPU */ | 5371 | /* Initialize migration for the boot CPU */ |
5372 | err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); | 5372 | err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); |
5373 | BUG_ON(err == NOTIFY_BAD); | 5373 | BUG_ON(err == NOTIFY_BAD); |
5374 | migration_call(&migration_notifier, CPU_ONLINE, cpu); | 5374 | migration_call(&migration_notifier, CPU_ONLINE, cpu); |
5375 | register_cpu_notifier(&migration_notifier); | 5375 | register_cpu_notifier(&migration_notifier); |
5376 | 5376 | ||
5377 | /* Register cpu active notifiers */ | 5377 | /* Register cpu active notifiers */ |
5378 | cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); | 5378 | cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); |
5379 | cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); | 5379 | cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); |
5380 | 5380 | ||
5381 | return 0; | 5381 | return 0; |
5382 | } | 5382 | } |
5383 | early_initcall(migration_init); | 5383 | early_initcall(migration_init); |
5384 | #endif | 5384 | #endif |
5385 | 5385 | ||
5386 | #ifdef CONFIG_SMP | 5386 | #ifdef CONFIG_SMP |
5387 | 5387 | ||
5388 | static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ | 5388 | static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ |
5389 | 5389 | ||
5390 | #ifdef CONFIG_SCHED_DEBUG | 5390 | #ifdef CONFIG_SCHED_DEBUG |
5391 | 5391 | ||
5392 | static __read_mostly int sched_debug_enabled; | 5392 | static __read_mostly int sched_debug_enabled; |
5393 | 5393 | ||
5394 | static int __init sched_debug_setup(char *str) | 5394 | static int __init sched_debug_setup(char *str) |
5395 | { | 5395 | { |
5396 | sched_debug_enabled = 1; | 5396 | sched_debug_enabled = 1; |
5397 | 5397 | ||
5398 | return 0; | 5398 | return 0; |
5399 | } | 5399 | } |
5400 | early_param("sched_debug", sched_debug_setup); | 5400 | early_param("sched_debug", sched_debug_setup); |
5401 | 5401 | ||
5402 | static inline bool sched_debug(void) | 5402 | static inline bool sched_debug(void) |
5403 | { | 5403 | { |
5404 | return sched_debug_enabled; | 5404 | return sched_debug_enabled; |
5405 | } | 5405 | } |
5406 | 5406 | ||
5407 | static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, | 5407 | static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, |
5408 | struct cpumask *groupmask) | 5408 | struct cpumask *groupmask) |
5409 | { | 5409 | { |
5410 | struct sched_group *group = sd->groups; | 5410 | struct sched_group *group = sd->groups; |
5411 | char str[256]; | 5411 | char str[256]; |
5412 | 5412 | ||
5413 | cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); | 5413 | cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); |
5414 | cpumask_clear(groupmask); | 5414 | cpumask_clear(groupmask); |
5415 | 5415 | ||
5416 | printk(KERN_DEBUG "%*s domain %d: ", level, "", level); | 5416 | printk(KERN_DEBUG "%*s domain %d: ", level, "", level); |
5417 | 5417 | ||
5418 | if (!(sd->flags & SD_LOAD_BALANCE)) { | 5418 | if (!(sd->flags & SD_LOAD_BALANCE)) { |
5419 | printk("does not load-balance\n"); | 5419 | printk("does not load-balance\n"); |
5420 | if (sd->parent) | 5420 | if (sd->parent) |
5421 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" | 5421 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" |
5422 | " has parent"); | 5422 | " has parent"); |
5423 | return -1; | 5423 | return -1; |
5424 | } | 5424 | } |
5425 | 5425 | ||
5426 | printk(KERN_CONT "span %s level %s\n", str, sd->name); | 5426 | printk(KERN_CONT "span %s level %s\n", str, sd->name); |
5427 | 5427 | ||
5428 | if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { | 5428 | if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { |
5429 | printk(KERN_ERR "ERROR: domain->span does not contain " | 5429 | printk(KERN_ERR "ERROR: domain->span does not contain " |
5430 | "CPU%d\n", cpu); | 5430 | "CPU%d\n", cpu); |
5431 | } | 5431 | } |
5432 | if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { | 5432 | if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { |
5433 | printk(KERN_ERR "ERROR: domain->groups does not contain" | 5433 | printk(KERN_ERR "ERROR: domain->groups does not contain" |
5434 | " CPU%d\n", cpu); | 5434 | " CPU%d\n", cpu); |
5435 | } | 5435 | } |
5436 | 5436 | ||
5437 | printk(KERN_DEBUG "%*s groups:", level + 1, ""); | 5437 | printk(KERN_DEBUG "%*s groups:", level + 1, ""); |
5438 | do { | 5438 | do { |
5439 | if (!group) { | 5439 | if (!group) { |
5440 | printk("\n"); | 5440 | printk("\n"); |
5441 | printk(KERN_ERR "ERROR: group is NULL\n"); | 5441 | printk(KERN_ERR "ERROR: group is NULL\n"); |
5442 | break; | 5442 | break; |
5443 | } | 5443 | } |
5444 | 5444 | ||
5445 | /* | 5445 | /* |
5446 | * Even though we initialize ->capacity to something semi-sane, | 5446 | * Even though we initialize ->capacity to something semi-sane, |
5447 | * we leave capacity_orig unset. This allows us to detect if | 5447 | * we leave capacity_orig unset. This allows us to detect if |
5448 | * domain iteration is still funny without causing /0 traps. | 5448 | * domain iteration is still funny without causing /0 traps. |
5449 | */ | 5449 | */ |
5450 | if (!group->sgc->capacity_orig) { | 5450 | if (!group->sgc->capacity_orig) { |
5451 | printk(KERN_CONT "\n"); | 5451 | printk(KERN_CONT "\n"); |
5452 | printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n"); | 5452 | printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n"); |
5453 | break; | 5453 | break; |
5454 | } | 5454 | } |
5455 | 5455 | ||
5456 | if (!cpumask_weight(sched_group_cpus(group))) { | 5456 | if (!cpumask_weight(sched_group_cpus(group))) { |
5457 | printk(KERN_CONT "\n"); | 5457 | printk(KERN_CONT "\n"); |
5458 | printk(KERN_ERR "ERROR: empty group\n"); | 5458 | printk(KERN_ERR "ERROR: empty group\n"); |
5459 | break; | 5459 | break; |
5460 | } | 5460 | } |
5461 | 5461 | ||
5462 | if (!(sd->flags & SD_OVERLAP) && | 5462 | if (!(sd->flags & SD_OVERLAP) && |
5463 | cpumask_intersects(groupmask, sched_group_cpus(group))) { | 5463 | cpumask_intersects(groupmask, sched_group_cpus(group))) { |
5464 | printk(KERN_CONT "\n"); | 5464 | printk(KERN_CONT "\n"); |
5465 | printk(KERN_ERR "ERROR: repeated CPUs\n"); | 5465 | printk(KERN_ERR "ERROR: repeated CPUs\n"); |
5466 | break; | 5466 | break; |
5467 | } | 5467 | } |
5468 | 5468 | ||
5469 | cpumask_or(groupmask, groupmask, sched_group_cpus(group)); | 5469 | cpumask_or(groupmask, groupmask, sched_group_cpus(group)); |
5470 | 5470 | ||
5471 | cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); | 5471 | cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); |
5472 | 5472 | ||
5473 | printk(KERN_CONT " %s", str); | 5473 | printk(KERN_CONT " %s", str); |
5474 | if (group->sgc->capacity != SCHED_CAPACITY_SCALE) { | 5474 | if (group->sgc->capacity != SCHED_CAPACITY_SCALE) { |
5475 | printk(KERN_CONT " (cpu_capacity = %d)", | 5475 | printk(KERN_CONT " (cpu_capacity = %d)", |
5476 | group->sgc->capacity); | 5476 | group->sgc->capacity); |
5477 | } | 5477 | } |
5478 | 5478 | ||
5479 | group = group->next; | 5479 | group = group->next; |
5480 | } while (group != sd->groups); | 5480 | } while (group != sd->groups); |
5481 | printk(KERN_CONT "\n"); | 5481 | printk(KERN_CONT "\n"); |
5482 | 5482 | ||
5483 | if (!cpumask_equal(sched_domain_span(sd), groupmask)) | 5483 | if (!cpumask_equal(sched_domain_span(sd), groupmask)) |
5484 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); | 5484 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); |
5485 | 5485 | ||
5486 | if (sd->parent && | 5486 | if (sd->parent && |
5487 | !cpumask_subset(groupmask, sched_domain_span(sd->parent))) | 5487 | !cpumask_subset(groupmask, sched_domain_span(sd->parent))) |
5488 | printk(KERN_ERR "ERROR: parent span is not a superset " | 5488 | printk(KERN_ERR "ERROR: parent span is not a superset " |
5489 | "of domain->span\n"); | 5489 | "of domain->span\n"); |
5490 | return 0; | 5490 | return 0; |
5491 | } | 5491 | } |
5492 | 5492 | ||
5493 | static void sched_domain_debug(struct sched_domain *sd, int cpu) | 5493 | static void sched_domain_debug(struct sched_domain *sd, int cpu) |
5494 | { | 5494 | { |
5495 | int level = 0; | 5495 | int level = 0; |
5496 | 5496 | ||
5497 | if (!sched_debug_enabled) | 5497 | if (!sched_debug_enabled) |
5498 | return; | 5498 | return; |
5499 | 5499 | ||
5500 | if (!sd) { | 5500 | if (!sd) { |
5501 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); | 5501 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); |
5502 | return; | 5502 | return; |
5503 | } | 5503 | } |
5504 | 5504 | ||
5505 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); | 5505 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
5506 | 5506 | ||
5507 | for (;;) { | 5507 | for (;;) { |
5508 | if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) | 5508 | if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) |
5509 | break; | 5509 | break; |
5510 | level++; | 5510 | level++; |
5511 | sd = sd->parent; | 5511 | sd = sd->parent; |
5512 | if (!sd) | 5512 | if (!sd) |
5513 | break; | 5513 | break; |
5514 | } | 5514 | } |
5515 | } | 5515 | } |
5516 | #else /* !CONFIG_SCHED_DEBUG */ | 5516 | #else /* !CONFIG_SCHED_DEBUG */ |
5517 | # define sched_domain_debug(sd, cpu) do { } while (0) | 5517 | # define sched_domain_debug(sd, cpu) do { } while (0) |
5518 | static inline bool sched_debug(void) | 5518 | static inline bool sched_debug(void) |
5519 | { | 5519 | { |
5520 | return false; | 5520 | return false; |
5521 | } | 5521 | } |
5522 | #endif /* CONFIG_SCHED_DEBUG */ | 5522 | #endif /* CONFIG_SCHED_DEBUG */ |
5523 | 5523 | ||
5524 | static int sd_degenerate(struct sched_domain *sd) | 5524 | static int sd_degenerate(struct sched_domain *sd) |
5525 | { | 5525 | { |
5526 | if (cpumask_weight(sched_domain_span(sd)) == 1) | 5526 | if (cpumask_weight(sched_domain_span(sd)) == 1) |
5527 | return 1; | 5527 | return 1; |
5528 | 5528 | ||
5529 | /* Following flags need at least 2 groups */ | 5529 | /* Following flags need at least 2 groups */ |
5530 | if (sd->flags & (SD_LOAD_BALANCE | | 5530 | if (sd->flags & (SD_LOAD_BALANCE | |
5531 | SD_BALANCE_NEWIDLE | | 5531 | SD_BALANCE_NEWIDLE | |
5532 | SD_BALANCE_FORK | | 5532 | SD_BALANCE_FORK | |
5533 | SD_BALANCE_EXEC | | 5533 | SD_BALANCE_EXEC | |
5534 | SD_SHARE_CPUCAPACITY | | 5534 | SD_SHARE_CPUCAPACITY | |
5535 | SD_SHARE_PKG_RESOURCES | | 5535 | SD_SHARE_PKG_RESOURCES | |
5536 | SD_SHARE_POWERDOMAIN)) { | 5536 | SD_SHARE_POWERDOMAIN)) { |
5537 | if (sd->groups != sd->groups->next) | 5537 | if (sd->groups != sd->groups->next) |
5538 | return 0; | 5538 | return 0; |
5539 | } | 5539 | } |
5540 | 5540 | ||
5541 | /* Following flags don't use groups */ | 5541 | /* Following flags don't use groups */ |
5542 | if (sd->flags & (SD_WAKE_AFFINE)) | 5542 | if (sd->flags & (SD_WAKE_AFFINE)) |
5543 | return 0; | 5543 | return 0; |
5544 | 5544 | ||
5545 | return 1; | 5545 | return 1; |
5546 | } | 5546 | } |
5547 | 5547 | ||
5548 | static int | 5548 | static int |
5549 | sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) | 5549 | sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) |
5550 | { | 5550 | { |
5551 | unsigned long cflags = sd->flags, pflags = parent->flags; | 5551 | unsigned long cflags = sd->flags, pflags = parent->flags; |
5552 | 5552 | ||
5553 | if (sd_degenerate(parent)) | 5553 | if (sd_degenerate(parent)) |
5554 | return 1; | 5554 | return 1; |
5555 | 5555 | ||
5556 | if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) | 5556 | if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) |
5557 | return 0; | 5557 | return 0; |
5558 | 5558 | ||
5559 | /* Flags needing groups don't count if only 1 group in parent */ | 5559 | /* Flags needing groups don't count if only 1 group in parent */ |
5560 | if (parent->groups == parent->groups->next) { | 5560 | if (parent->groups == parent->groups->next) { |
5561 | pflags &= ~(SD_LOAD_BALANCE | | 5561 | pflags &= ~(SD_LOAD_BALANCE | |
5562 | SD_BALANCE_NEWIDLE | | 5562 | SD_BALANCE_NEWIDLE | |
5563 | SD_BALANCE_FORK | | 5563 | SD_BALANCE_FORK | |
5564 | SD_BALANCE_EXEC | | 5564 | SD_BALANCE_EXEC | |
5565 | SD_SHARE_CPUCAPACITY | | 5565 | SD_SHARE_CPUCAPACITY | |
5566 | SD_SHARE_PKG_RESOURCES | | 5566 | SD_SHARE_PKG_RESOURCES | |
5567 | SD_PREFER_SIBLING | | 5567 | SD_PREFER_SIBLING | |
5568 | SD_SHARE_POWERDOMAIN); | 5568 | SD_SHARE_POWERDOMAIN); |
5569 | if (nr_node_ids == 1) | 5569 | if (nr_node_ids == 1) |
5570 | pflags &= ~SD_SERIALIZE; | 5570 | pflags &= ~SD_SERIALIZE; |
5571 | } | 5571 | } |
5572 | if (~cflags & pflags) | 5572 | if (~cflags & pflags) |
5573 | return 0; | 5573 | return 0; |
5574 | 5574 | ||
5575 | return 1; | 5575 | return 1; |
5576 | } | 5576 | } |
5577 | 5577 | ||
5578 | static void free_rootdomain(struct rcu_head *rcu) | 5578 | static void free_rootdomain(struct rcu_head *rcu) |
5579 | { | 5579 | { |
5580 | struct root_domain *rd = container_of(rcu, struct root_domain, rcu); | 5580 | struct root_domain *rd = container_of(rcu, struct root_domain, rcu); |
5581 | 5581 | ||
5582 | cpupri_cleanup(&rd->cpupri); | 5582 | cpupri_cleanup(&rd->cpupri); |
5583 | cpudl_cleanup(&rd->cpudl); | 5583 | cpudl_cleanup(&rd->cpudl); |
5584 | free_cpumask_var(rd->dlo_mask); | 5584 | free_cpumask_var(rd->dlo_mask); |
5585 | free_cpumask_var(rd->rto_mask); | 5585 | free_cpumask_var(rd->rto_mask); |
5586 | free_cpumask_var(rd->online); | 5586 | free_cpumask_var(rd->online); |
5587 | free_cpumask_var(rd->span); | 5587 | free_cpumask_var(rd->span); |
5588 | kfree(rd); | 5588 | kfree(rd); |
5589 | } | 5589 | } |
5590 | 5590 | ||
5591 | static void rq_attach_root(struct rq *rq, struct root_domain *rd) | 5591 | static void rq_attach_root(struct rq *rq, struct root_domain *rd) |
5592 | { | 5592 | { |
5593 | struct root_domain *old_rd = NULL; | 5593 | struct root_domain *old_rd = NULL; |
5594 | unsigned long flags; | 5594 | unsigned long flags; |
5595 | 5595 | ||
5596 | raw_spin_lock_irqsave(&rq->lock, flags); | 5596 | raw_spin_lock_irqsave(&rq->lock, flags); |
5597 | 5597 | ||
5598 | if (rq->rd) { | 5598 | if (rq->rd) { |
5599 | old_rd = rq->rd; | 5599 | old_rd = rq->rd; |
5600 | 5600 | ||
5601 | if (cpumask_test_cpu(rq->cpu, old_rd->online)) | 5601 | if (cpumask_test_cpu(rq->cpu, old_rd->online)) |
5602 | set_rq_offline(rq); | 5602 | set_rq_offline(rq); |
5603 | 5603 | ||
5604 | cpumask_clear_cpu(rq->cpu, old_rd->span); | 5604 | cpumask_clear_cpu(rq->cpu, old_rd->span); |
5605 | 5605 | ||
5606 | /* | 5606 | /* |
5607 | * If we dont want to free the old_rd yet then | 5607 | * If we dont want to free the old_rd yet then |
5608 | * set old_rd to NULL to skip the freeing later | 5608 | * set old_rd to NULL to skip the freeing later |
5609 | * in this function: | 5609 | * in this function: |
5610 | */ | 5610 | */ |
5611 | if (!atomic_dec_and_test(&old_rd->refcount)) | 5611 | if (!atomic_dec_and_test(&old_rd->refcount)) |
5612 | old_rd = NULL; | 5612 | old_rd = NULL; |
5613 | } | 5613 | } |
5614 | 5614 | ||
5615 | atomic_inc(&rd->refcount); | 5615 | atomic_inc(&rd->refcount); |
5616 | rq->rd = rd; | 5616 | rq->rd = rd; |
5617 | 5617 | ||
5618 | cpumask_set_cpu(rq->cpu, rd->span); | 5618 | cpumask_set_cpu(rq->cpu, rd->span); |
5619 | if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) | 5619 | if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) |
5620 | set_rq_online(rq); | 5620 | set_rq_online(rq); |
5621 | 5621 | ||
5622 | raw_spin_unlock_irqrestore(&rq->lock, flags); | 5622 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
5623 | 5623 | ||
5624 | if (old_rd) | 5624 | if (old_rd) |
5625 | call_rcu_sched(&old_rd->rcu, free_rootdomain); | 5625 | call_rcu_sched(&old_rd->rcu, free_rootdomain); |
5626 | } | 5626 | } |
5627 | 5627 | ||
5628 | static int init_rootdomain(struct root_domain *rd) | 5628 | static int init_rootdomain(struct root_domain *rd) |
5629 | { | 5629 | { |
5630 | memset(rd, 0, sizeof(*rd)); | 5630 | memset(rd, 0, sizeof(*rd)); |
5631 | 5631 | ||
5632 | if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) | 5632 | if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) |
5633 | goto out; | 5633 | goto out; |
5634 | if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) | 5634 | if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) |
5635 | goto free_span; | 5635 | goto free_span; |
5636 | if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) | 5636 | if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) |
5637 | goto free_online; | 5637 | goto free_online; |
5638 | if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) | 5638 | if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) |
5639 | goto free_dlo_mask; | 5639 | goto free_dlo_mask; |
5640 | 5640 | ||
5641 | init_dl_bw(&rd->dl_bw); | 5641 | init_dl_bw(&rd->dl_bw); |
5642 | if (cpudl_init(&rd->cpudl) != 0) | 5642 | if (cpudl_init(&rd->cpudl) != 0) |
5643 | goto free_dlo_mask; | 5643 | goto free_dlo_mask; |
5644 | 5644 | ||
5645 | if (cpupri_init(&rd->cpupri) != 0) | 5645 | if (cpupri_init(&rd->cpupri) != 0) |
5646 | goto free_rto_mask; | 5646 | goto free_rto_mask; |
5647 | return 0; | 5647 | return 0; |
5648 | 5648 | ||
5649 | free_rto_mask: | 5649 | free_rto_mask: |
5650 | free_cpumask_var(rd->rto_mask); | 5650 | free_cpumask_var(rd->rto_mask); |
5651 | free_dlo_mask: | 5651 | free_dlo_mask: |
5652 | free_cpumask_var(rd->dlo_mask); | 5652 | free_cpumask_var(rd->dlo_mask); |
5653 | free_online: | 5653 | free_online: |
5654 | free_cpumask_var(rd->online); | 5654 | free_cpumask_var(rd->online); |
5655 | free_span: | 5655 | free_span: |
5656 | free_cpumask_var(rd->span); | 5656 | free_cpumask_var(rd->span); |
5657 | out: | 5657 | out: |
5658 | return -ENOMEM; | 5658 | return -ENOMEM; |
5659 | } | 5659 | } |
5660 | 5660 | ||
5661 | /* | 5661 | /* |
5662 | * By default the system creates a single root-domain with all cpus as | 5662 | * By default the system creates a single root-domain with all cpus as |
5663 | * members (mimicking the global state we have today). | 5663 | * members (mimicking the global state we have today). |
5664 | */ | 5664 | */ |
5665 | struct root_domain def_root_domain; | 5665 | struct root_domain def_root_domain; |
5666 | 5666 | ||
5667 | static void init_defrootdomain(void) | 5667 | static void init_defrootdomain(void) |
5668 | { | 5668 | { |
5669 | init_rootdomain(&def_root_domain); | 5669 | init_rootdomain(&def_root_domain); |
5670 | 5670 | ||
5671 | atomic_set(&def_root_domain.refcount, 1); | 5671 | atomic_set(&def_root_domain.refcount, 1); |
5672 | } | 5672 | } |
5673 | 5673 | ||
5674 | static struct root_domain *alloc_rootdomain(void) | 5674 | static struct root_domain *alloc_rootdomain(void) |
5675 | { | 5675 | { |
5676 | struct root_domain *rd; | 5676 | struct root_domain *rd; |
5677 | 5677 | ||
5678 | rd = kmalloc(sizeof(*rd), GFP_KERNEL); | 5678 | rd = kmalloc(sizeof(*rd), GFP_KERNEL); |
5679 | if (!rd) | 5679 | if (!rd) |
5680 | return NULL; | 5680 | return NULL; |
5681 | 5681 | ||
5682 | if (init_rootdomain(rd) != 0) { | 5682 | if (init_rootdomain(rd) != 0) { |
5683 | kfree(rd); | 5683 | kfree(rd); |
5684 | return NULL; | 5684 | return NULL; |
5685 | } | 5685 | } |
5686 | 5686 | ||
5687 | return rd; | 5687 | return rd; |
5688 | } | 5688 | } |
5689 | 5689 | ||
5690 | static void free_sched_groups(struct sched_group *sg, int free_sgc) | 5690 | static void free_sched_groups(struct sched_group *sg, int free_sgc) |
5691 | { | 5691 | { |
5692 | struct sched_group *tmp, *first; | 5692 | struct sched_group *tmp, *first; |
5693 | 5693 | ||
5694 | if (!sg) | 5694 | if (!sg) |
5695 | return; | 5695 | return; |
5696 | 5696 | ||
5697 | first = sg; | 5697 | first = sg; |
5698 | do { | 5698 | do { |
5699 | tmp = sg->next; | 5699 | tmp = sg->next; |
5700 | 5700 | ||
5701 | if (free_sgc && atomic_dec_and_test(&sg->sgc->ref)) | 5701 | if (free_sgc && atomic_dec_and_test(&sg->sgc->ref)) |
5702 | kfree(sg->sgc); | 5702 | kfree(sg->sgc); |
5703 | 5703 | ||
5704 | kfree(sg); | 5704 | kfree(sg); |
5705 | sg = tmp; | 5705 | sg = tmp; |
5706 | } while (sg != first); | 5706 | } while (sg != first); |
5707 | } | 5707 | } |
5708 | 5708 | ||
5709 | static void free_sched_domain(struct rcu_head *rcu) | 5709 | static void free_sched_domain(struct rcu_head *rcu) |
5710 | { | 5710 | { |
5711 | struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); | 5711 | struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); |
5712 | 5712 | ||
5713 | /* | 5713 | /* |
5714 | * If its an overlapping domain it has private groups, iterate and | 5714 | * If its an overlapping domain it has private groups, iterate and |
5715 | * nuke them all. | 5715 | * nuke them all. |
5716 | */ | 5716 | */ |
5717 | if (sd->flags & SD_OVERLAP) { | 5717 | if (sd->flags & SD_OVERLAP) { |
5718 | free_sched_groups(sd->groups, 1); | 5718 | free_sched_groups(sd->groups, 1); |
5719 | } else if (atomic_dec_and_test(&sd->groups->ref)) { | 5719 | } else if (atomic_dec_and_test(&sd->groups->ref)) { |
5720 | kfree(sd->groups->sgc); | 5720 | kfree(sd->groups->sgc); |
5721 | kfree(sd->groups); | 5721 | kfree(sd->groups); |
5722 | } | 5722 | } |
5723 | kfree(sd); | 5723 | kfree(sd); |
5724 | } | 5724 | } |
5725 | 5725 | ||
5726 | static void destroy_sched_domain(struct sched_domain *sd, int cpu) | 5726 | static void destroy_sched_domain(struct sched_domain *sd, int cpu) |
5727 | { | 5727 | { |
5728 | call_rcu(&sd->rcu, free_sched_domain); | 5728 | call_rcu(&sd->rcu, free_sched_domain); |
5729 | } | 5729 | } |
5730 | 5730 | ||
5731 | static void destroy_sched_domains(struct sched_domain *sd, int cpu) | 5731 | static void destroy_sched_domains(struct sched_domain *sd, int cpu) |
5732 | { | 5732 | { |
5733 | for (; sd; sd = sd->parent) | 5733 | for (; sd; sd = sd->parent) |
5734 | destroy_sched_domain(sd, cpu); | 5734 | destroy_sched_domain(sd, cpu); |
5735 | } | 5735 | } |
5736 | 5736 | ||
5737 | /* | 5737 | /* |
5738 | * Keep a special pointer to the highest sched_domain that has | 5738 | * Keep a special pointer to the highest sched_domain that has |
5739 | * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this | 5739 | * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this |
5740 | * allows us to avoid some pointer chasing select_idle_sibling(). | 5740 | * allows us to avoid some pointer chasing select_idle_sibling(). |
5741 | * | 5741 | * |
5742 | * Also keep a unique ID per domain (we use the first cpu number in | 5742 | * Also keep a unique ID per domain (we use the first cpu number in |
5743 | * the cpumask of the domain), this allows us to quickly tell if | 5743 | * the cpumask of the domain), this allows us to quickly tell if |
5744 | * two cpus are in the same cache domain, see cpus_share_cache(). | 5744 | * two cpus are in the same cache domain, see cpus_share_cache(). |
5745 | */ | 5745 | */ |
5746 | DEFINE_PER_CPU(struct sched_domain *, sd_llc); | 5746 | DEFINE_PER_CPU(struct sched_domain *, sd_llc); |
5747 | DEFINE_PER_CPU(int, sd_llc_size); | 5747 | DEFINE_PER_CPU(int, sd_llc_size); |
5748 | DEFINE_PER_CPU(int, sd_llc_id); | 5748 | DEFINE_PER_CPU(int, sd_llc_id); |
5749 | DEFINE_PER_CPU(struct sched_domain *, sd_numa); | 5749 | DEFINE_PER_CPU(struct sched_domain *, sd_numa); |
5750 | DEFINE_PER_CPU(struct sched_domain *, sd_busy); | 5750 | DEFINE_PER_CPU(struct sched_domain *, sd_busy); |
5751 | DEFINE_PER_CPU(struct sched_domain *, sd_asym); | 5751 | DEFINE_PER_CPU(struct sched_domain *, sd_asym); |
5752 | 5752 | ||
5753 | static void update_top_cache_domain(int cpu) | 5753 | static void update_top_cache_domain(int cpu) |
5754 | { | 5754 | { |
5755 | struct sched_domain *sd; | 5755 | struct sched_domain *sd; |
5756 | struct sched_domain *busy_sd = NULL; | 5756 | struct sched_domain *busy_sd = NULL; |
5757 | int id = cpu; | 5757 | int id = cpu; |
5758 | int size = 1; | 5758 | int size = 1; |
5759 | 5759 | ||
5760 | sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); | 5760 | sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); |
5761 | if (sd) { | 5761 | if (sd) { |
5762 | id = cpumask_first(sched_domain_span(sd)); | 5762 | id = cpumask_first(sched_domain_span(sd)); |
5763 | size = cpumask_weight(sched_domain_span(sd)); | 5763 | size = cpumask_weight(sched_domain_span(sd)); |
5764 | busy_sd = sd->parent; /* sd_busy */ | 5764 | busy_sd = sd->parent; /* sd_busy */ |
5765 | } | 5765 | } |
5766 | rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); | 5766 | rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); |
5767 | 5767 | ||
5768 | rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); | 5768 | rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); |
5769 | per_cpu(sd_llc_size, cpu) = size; | 5769 | per_cpu(sd_llc_size, cpu) = size; |
5770 | per_cpu(sd_llc_id, cpu) = id; | 5770 | per_cpu(sd_llc_id, cpu) = id; |
5771 | 5771 | ||
5772 | sd = lowest_flag_domain(cpu, SD_NUMA); | 5772 | sd = lowest_flag_domain(cpu, SD_NUMA); |
5773 | rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); | 5773 | rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); |
5774 | 5774 | ||
5775 | sd = highest_flag_domain(cpu, SD_ASYM_PACKING); | 5775 | sd = highest_flag_domain(cpu, SD_ASYM_PACKING); |
5776 | rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); | 5776 | rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); |
5777 | } | 5777 | } |
5778 | 5778 | ||
5779 | /* | 5779 | /* |
5780 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must | 5780 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must |
5781 | * hold the hotplug lock. | 5781 | * hold the hotplug lock. |
5782 | */ | 5782 | */ |
5783 | static void | 5783 | static void |
5784 | cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) | 5784 | cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) |
5785 | { | 5785 | { |
5786 | struct rq *rq = cpu_rq(cpu); | 5786 | struct rq *rq = cpu_rq(cpu); |
5787 | struct sched_domain *tmp; | 5787 | struct sched_domain *tmp; |
5788 | 5788 | ||
5789 | /* Remove the sched domains which do not contribute to scheduling. */ | 5789 | /* Remove the sched domains which do not contribute to scheduling. */ |
5790 | for (tmp = sd; tmp; ) { | 5790 | for (tmp = sd; tmp; ) { |
5791 | struct sched_domain *parent = tmp->parent; | 5791 | struct sched_domain *parent = tmp->parent; |
5792 | if (!parent) | 5792 | if (!parent) |
5793 | break; | 5793 | break; |
5794 | 5794 | ||
5795 | if (sd_parent_degenerate(tmp, parent)) { | 5795 | if (sd_parent_degenerate(tmp, parent)) { |
5796 | tmp->parent = parent->parent; | 5796 | tmp->parent = parent->parent; |
5797 | if (parent->parent) | 5797 | if (parent->parent) |
5798 | parent->parent->child = tmp; | 5798 | parent->parent->child = tmp; |
5799 | /* | 5799 | /* |
5800 | * Transfer SD_PREFER_SIBLING down in case of a | 5800 | * Transfer SD_PREFER_SIBLING down in case of a |
5801 | * degenerate parent; the spans match for this | 5801 | * degenerate parent; the spans match for this |
5802 | * so the property transfers. | 5802 | * so the property transfers. |
5803 | */ | 5803 | */ |
5804 | if (parent->flags & SD_PREFER_SIBLING) | 5804 | if (parent->flags & SD_PREFER_SIBLING) |
5805 | tmp->flags |= SD_PREFER_SIBLING; | 5805 | tmp->flags |= SD_PREFER_SIBLING; |
5806 | destroy_sched_domain(parent, cpu); | 5806 | destroy_sched_domain(parent, cpu); |
5807 | } else | 5807 | } else |
5808 | tmp = tmp->parent; | 5808 | tmp = tmp->parent; |
5809 | } | 5809 | } |
5810 | 5810 | ||
5811 | if (sd && sd_degenerate(sd)) { | 5811 | if (sd && sd_degenerate(sd)) { |
5812 | tmp = sd; | 5812 | tmp = sd; |
5813 | sd = sd->parent; | 5813 | sd = sd->parent; |
5814 | destroy_sched_domain(tmp, cpu); | 5814 | destroy_sched_domain(tmp, cpu); |
5815 | if (sd) | 5815 | if (sd) |
5816 | sd->child = NULL; | 5816 | sd->child = NULL; |
5817 | } | 5817 | } |
5818 | 5818 | ||
5819 | sched_domain_debug(sd, cpu); | 5819 | sched_domain_debug(sd, cpu); |
5820 | 5820 | ||
5821 | rq_attach_root(rq, rd); | 5821 | rq_attach_root(rq, rd); |
5822 | tmp = rq->sd; | 5822 | tmp = rq->sd; |
5823 | rcu_assign_pointer(rq->sd, sd); | 5823 | rcu_assign_pointer(rq->sd, sd); |
5824 | destroy_sched_domains(tmp, cpu); | 5824 | destroy_sched_domains(tmp, cpu); |
5825 | 5825 | ||
5826 | update_top_cache_domain(cpu); | 5826 | update_top_cache_domain(cpu); |
5827 | } | 5827 | } |
5828 | 5828 | ||
5829 | /* cpus with isolated domains */ | 5829 | /* cpus with isolated domains */ |
5830 | static cpumask_var_t cpu_isolated_map; | 5830 | static cpumask_var_t cpu_isolated_map; |
5831 | 5831 | ||
5832 | /* Setup the mask of cpus configured for isolated domains */ | 5832 | /* Setup the mask of cpus configured for isolated domains */ |
5833 | static int __init isolated_cpu_setup(char *str) | 5833 | static int __init isolated_cpu_setup(char *str) |
5834 | { | 5834 | { |
5835 | alloc_bootmem_cpumask_var(&cpu_isolated_map); | 5835 | alloc_bootmem_cpumask_var(&cpu_isolated_map); |
5836 | cpulist_parse(str, cpu_isolated_map); | 5836 | cpulist_parse(str, cpu_isolated_map); |
5837 | return 1; | 5837 | return 1; |
5838 | } | 5838 | } |
5839 | 5839 | ||
5840 | __setup("isolcpus=", isolated_cpu_setup); | 5840 | __setup("isolcpus=", isolated_cpu_setup); |
5841 | 5841 | ||
5842 | struct s_data { | 5842 | struct s_data { |
5843 | struct sched_domain ** __percpu sd; | 5843 | struct sched_domain ** __percpu sd; |
5844 | struct root_domain *rd; | 5844 | struct root_domain *rd; |
5845 | }; | 5845 | }; |
5846 | 5846 | ||
5847 | enum s_alloc { | 5847 | enum s_alloc { |
5848 | sa_rootdomain, | 5848 | sa_rootdomain, |
5849 | sa_sd, | 5849 | sa_sd, |
5850 | sa_sd_storage, | 5850 | sa_sd_storage, |
5851 | sa_none, | 5851 | sa_none, |
5852 | }; | 5852 | }; |
5853 | 5853 | ||
5854 | /* | 5854 | /* |
5855 | * Build an iteration mask that can exclude certain CPUs from the upwards | 5855 | * Build an iteration mask that can exclude certain CPUs from the upwards |
5856 | * domain traversal. | 5856 | * domain traversal. |
5857 | * | 5857 | * |
5858 | * Asymmetric node setups can result in situations where the domain tree is of | 5858 | * Asymmetric node setups can result in situations where the domain tree is of |
5859 | * unequal depth, make sure to skip domains that already cover the entire | 5859 | * unequal depth, make sure to skip domains that already cover the entire |
5860 | * range. | 5860 | * range. |
5861 | * | 5861 | * |
5862 | * In that case build_sched_domains() will have terminated the iteration early | 5862 | * In that case build_sched_domains() will have terminated the iteration early |
5863 | * and our sibling sd spans will be empty. Domains should always include the | 5863 | * and our sibling sd spans will be empty. Domains should always include the |
5864 | * cpu they're built on, so check that. | 5864 | * cpu they're built on, so check that. |
5865 | * | 5865 | * |
5866 | */ | 5866 | */ |
5867 | static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) | 5867 | static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) |
5868 | { | 5868 | { |
5869 | const struct cpumask *span = sched_domain_span(sd); | 5869 | const struct cpumask *span = sched_domain_span(sd); |
5870 | struct sd_data *sdd = sd->private; | 5870 | struct sd_data *sdd = sd->private; |
5871 | struct sched_domain *sibling; | 5871 | struct sched_domain *sibling; |
5872 | int i; | 5872 | int i; |
5873 | 5873 | ||
5874 | for_each_cpu(i, span) { | 5874 | for_each_cpu(i, span) { |
5875 | sibling = *per_cpu_ptr(sdd->sd, i); | 5875 | sibling = *per_cpu_ptr(sdd->sd, i); |
5876 | if (!cpumask_test_cpu(i, sched_domain_span(sibling))) | 5876 | if (!cpumask_test_cpu(i, sched_domain_span(sibling))) |
5877 | continue; | 5877 | continue; |
5878 | 5878 | ||
5879 | cpumask_set_cpu(i, sched_group_mask(sg)); | 5879 | cpumask_set_cpu(i, sched_group_mask(sg)); |
5880 | } | 5880 | } |
5881 | } | 5881 | } |
5882 | 5882 | ||
5883 | /* | 5883 | /* |
5884 | * Return the canonical balance cpu for this group, this is the first cpu | 5884 | * Return the canonical balance cpu for this group, this is the first cpu |
5885 | * of this group that's also in the iteration mask. | 5885 | * of this group that's also in the iteration mask. |
5886 | */ | 5886 | */ |
5887 | int group_balance_cpu(struct sched_group *sg) | 5887 | int group_balance_cpu(struct sched_group *sg) |
5888 | { | 5888 | { |
5889 | return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); | 5889 | return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); |
5890 | } | 5890 | } |
5891 | 5891 | ||
5892 | static int | 5892 | static int |
5893 | build_overlap_sched_groups(struct sched_domain *sd, int cpu) | 5893 | build_overlap_sched_groups(struct sched_domain *sd, int cpu) |
5894 | { | 5894 | { |
5895 | struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; | 5895 | struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; |
5896 | const struct cpumask *span = sched_domain_span(sd); | 5896 | const struct cpumask *span = sched_domain_span(sd); |
5897 | struct cpumask *covered = sched_domains_tmpmask; | 5897 | struct cpumask *covered = sched_domains_tmpmask; |
5898 | struct sd_data *sdd = sd->private; | 5898 | struct sd_data *sdd = sd->private; |
5899 | struct sched_domain *sibling; | 5899 | struct sched_domain *sibling; |
5900 | int i; | 5900 | int i; |
5901 | 5901 | ||
5902 | cpumask_clear(covered); | 5902 | cpumask_clear(covered); |
5903 | 5903 | ||
5904 | for_each_cpu(i, span) { | 5904 | for_each_cpu(i, span) { |
5905 | struct cpumask *sg_span; | 5905 | struct cpumask *sg_span; |
5906 | 5906 | ||
5907 | if (cpumask_test_cpu(i, covered)) | 5907 | if (cpumask_test_cpu(i, covered)) |
5908 | continue; | 5908 | continue; |
5909 | 5909 | ||
5910 | sibling = *per_cpu_ptr(sdd->sd, i); | 5910 | sibling = *per_cpu_ptr(sdd->sd, i); |
5911 | 5911 | ||
5912 | /* See the comment near build_group_mask(). */ | 5912 | /* See the comment near build_group_mask(). */ |
5913 | if (!cpumask_test_cpu(i, sched_domain_span(sibling))) | 5913 | if (!cpumask_test_cpu(i, sched_domain_span(sibling))) |
5914 | continue; | 5914 | continue; |
5915 | 5915 | ||
5916 | sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), | 5916 | sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), |
5917 | GFP_KERNEL, cpu_to_node(cpu)); | 5917 | GFP_KERNEL, cpu_to_node(cpu)); |
5918 | 5918 | ||
5919 | if (!sg) | 5919 | if (!sg) |
5920 | goto fail; | 5920 | goto fail; |
5921 | 5921 | ||
5922 | sg_span = sched_group_cpus(sg); | 5922 | sg_span = sched_group_cpus(sg); |
5923 | if (sibling->child) | 5923 | if (sibling->child) |
5924 | cpumask_copy(sg_span, sched_domain_span(sibling->child)); | 5924 | cpumask_copy(sg_span, sched_domain_span(sibling->child)); |
5925 | else | 5925 | else |
5926 | cpumask_set_cpu(i, sg_span); | 5926 | cpumask_set_cpu(i, sg_span); |
5927 | 5927 | ||
5928 | cpumask_or(covered, covered, sg_span); | 5928 | cpumask_or(covered, covered, sg_span); |
5929 | 5929 | ||
5930 | sg->sgc = *per_cpu_ptr(sdd->sgc, i); | 5930 | sg->sgc = *per_cpu_ptr(sdd->sgc, i); |
5931 | if (atomic_inc_return(&sg->sgc->ref) == 1) | 5931 | if (atomic_inc_return(&sg->sgc->ref) == 1) |
5932 | build_group_mask(sd, sg); | 5932 | build_group_mask(sd, sg); |
5933 | 5933 | ||
5934 | /* | 5934 | /* |
5935 | * Initialize sgc->capacity such that even if we mess up the | 5935 | * Initialize sgc->capacity such that even if we mess up the |
5936 | * domains and no possible iteration will get us here, we won't | 5936 | * domains and no possible iteration will get us here, we won't |
5937 | * die on a /0 trap. | 5937 | * die on a /0 trap. |
5938 | */ | 5938 | */ |
5939 | sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span); | 5939 | sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span); |
5940 | sg->sgc->capacity_orig = sg->sgc->capacity; | 5940 | sg->sgc->capacity_orig = sg->sgc->capacity; |
5941 | 5941 | ||
5942 | /* | 5942 | /* |
5943 | * Make sure the first group of this domain contains the | 5943 | * Make sure the first group of this domain contains the |
5944 | * canonical balance cpu. Otherwise the sched_domain iteration | 5944 | * canonical balance cpu. Otherwise the sched_domain iteration |
5945 | * breaks. See update_sg_lb_stats(). | 5945 | * breaks. See update_sg_lb_stats(). |
5946 | */ | 5946 | */ |
5947 | if ((!groups && cpumask_test_cpu(cpu, sg_span)) || | 5947 | if ((!groups && cpumask_test_cpu(cpu, sg_span)) || |
5948 | group_balance_cpu(sg) == cpu) | 5948 | group_balance_cpu(sg) == cpu) |
5949 | groups = sg; | 5949 | groups = sg; |
5950 | 5950 | ||
5951 | if (!first) | 5951 | if (!first) |
5952 | first = sg; | 5952 | first = sg; |
5953 | if (last) | 5953 | if (last) |
5954 | last->next = sg; | 5954 | last->next = sg; |
5955 | last = sg; | 5955 | last = sg; |
5956 | last->next = first; | 5956 | last->next = first; |
5957 | } | 5957 | } |
5958 | sd->groups = groups; | 5958 | sd->groups = groups; |
5959 | 5959 | ||
5960 | return 0; | 5960 | return 0; |
5961 | 5961 | ||
5962 | fail: | 5962 | fail: |
5963 | free_sched_groups(first, 0); | 5963 | free_sched_groups(first, 0); |
5964 | 5964 | ||
5965 | return -ENOMEM; | 5965 | return -ENOMEM; |
5966 | } | 5966 | } |
5967 | 5967 | ||
5968 | static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) | 5968 | static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) |
5969 | { | 5969 | { |
5970 | struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); | 5970 | struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); |
5971 | struct sched_domain *child = sd->child; | 5971 | struct sched_domain *child = sd->child; |
5972 | 5972 | ||
5973 | if (child) | 5973 | if (child) |
5974 | cpu = cpumask_first(sched_domain_span(child)); | 5974 | cpu = cpumask_first(sched_domain_span(child)); |
5975 | 5975 | ||
5976 | if (sg) { | 5976 | if (sg) { |
5977 | *sg = *per_cpu_ptr(sdd->sg, cpu); | 5977 | *sg = *per_cpu_ptr(sdd->sg, cpu); |
5978 | (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu); | 5978 | (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu); |
5979 | atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */ | 5979 | atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */ |
5980 | } | 5980 | } |
5981 | 5981 | ||
5982 | return cpu; | 5982 | return cpu; |
5983 | } | 5983 | } |
5984 | 5984 | ||
5985 | /* | 5985 | /* |
5986 | * build_sched_groups will build a circular linked list of the groups | 5986 | * build_sched_groups will build a circular linked list of the groups |
5987 | * covered by the given span, and will set each group's ->cpumask correctly, | 5987 | * covered by the given span, and will set each group's ->cpumask correctly, |
5988 | * and ->cpu_capacity to 0. | 5988 | * and ->cpu_capacity to 0. |
5989 | * | 5989 | * |
5990 | * Assumes the sched_domain tree is fully constructed | 5990 | * Assumes the sched_domain tree is fully constructed |
5991 | */ | 5991 | */ |
5992 | static int | 5992 | static int |
5993 | build_sched_groups(struct sched_domain *sd, int cpu) | 5993 | build_sched_groups(struct sched_domain *sd, int cpu) |
5994 | { | 5994 | { |
5995 | struct sched_group *first = NULL, *last = NULL; | 5995 | struct sched_group *first = NULL, *last = NULL; |
5996 | struct sd_data *sdd = sd->private; | 5996 | struct sd_data *sdd = sd->private; |
5997 | const struct cpumask *span = sched_domain_span(sd); | 5997 | const struct cpumask *span = sched_domain_span(sd); |
5998 | struct cpumask *covered; | 5998 | struct cpumask *covered; |
5999 | int i; | 5999 | int i; |
6000 | 6000 | ||
6001 | get_group(cpu, sdd, &sd->groups); | 6001 | get_group(cpu, sdd, &sd->groups); |
6002 | atomic_inc(&sd->groups->ref); | 6002 | atomic_inc(&sd->groups->ref); |
6003 | 6003 | ||
6004 | if (cpu != cpumask_first(span)) | 6004 | if (cpu != cpumask_first(span)) |
6005 | return 0; | 6005 | return 0; |
6006 | 6006 | ||
6007 | lockdep_assert_held(&sched_domains_mutex); | 6007 | lockdep_assert_held(&sched_domains_mutex); |
6008 | covered = sched_domains_tmpmask; | 6008 | covered = sched_domains_tmpmask; |
6009 | 6009 | ||
6010 | cpumask_clear(covered); | 6010 | cpumask_clear(covered); |
6011 | 6011 | ||
6012 | for_each_cpu(i, span) { | 6012 | for_each_cpu(i, span) { |
6013 | struct sched_group *sg; | 6013 | struct sched_group *sg; |
6014 | int group, j; | 6014 | int group, j; |
6015 | 6015 | ||
6016 | if (cpumask_test_cpu(i, covered)) | 6016 | if (cpumask_test_cpu(i, covered)) |
6017 | continue; | 6017 | continue; |
6018 | 6018 | ||
6019 | group = get_group(i, sdd, &sg); | 6019 | group = get_group(i, sdd, &sg); |
6020 | cpumask_setall(sched_group_mask(sg)); | 6020 | cpumask_setall(sched_group_mask(sg)); |
6021 | 6021 | ||
6022 | for_each_cpu(j, span) { | 6022 | for_each_cpu(j, span) { |
6023 | if (get_group(j, sdd, NULL) != group) | 6023 | if (get_group(j, sdd, NULL) != group) |
6024 | continue; | 6024 | continue; |
6025 | 6025 | ||
6026 | cpumask_set_cpu(j, covered); | 6026 | cpumask_set_cpu(j, covered); |
6027 | cpumask_set_cpu(j, sched_group_cpus(sg)); | 6027 | cpumask_set_cpu(j, sched_group_cpus(sg)); |
6028 | } | 6028 | } |
6029 | 6029 | ||
6030 | if (!first) | 6030 | if (!first) |
6031 | first = sg; | 6031 | first = sg; |
6032 | if (last) | 6032 | if (last) |
6033 | last->next = sg; | 6033 | last->next = sg; |
6034 | last = sg; | 6034 | last = sg; |
6035 | } | 6035 | } |
6036 | last->next = first; | 6036 | last->next = first; |
6037 | 6037 | ||
6038 | return 0; | 6038 | return 0; |
6039 | } | 6039 | } |
6040 | 6040 | ||
6041 | /* | 6041 | /* |
6042 | * Initialize sched groups cpu_capacity. | 6042 | * Initialize sched groups cpu_capacity. |
6043 | * | 6043 | * |
6044 | * cpu_capacity indicates the capacity of sched group, which is used while | 6044 | * cpu_capacity indicates the capacity of sched group, which is used while |
6045 | * distributing the load between different sched groups in a sched domain. | 6045 | * distributing the load between different sched groups in a sched domain. |
6046 | * Typically cpu_capacity for all the groups in a sched domain will be same | 6046 | * Typically cpu_capacity for all the groups in a sched domain will be same |
6047 | * unless there are asymmetries in the topology. If there are asymmetries, | 6047 | * unless there are asymmetries in the topology. If there are asymmetries, |
6048 | * group having more cpu_capacity will pickup more load compared to the | 6048 | * group having more cpu_capacity will pickup more load compared to the |
6049 | * group having less cpu_capacity. | 6049 | * group having less cpu_capacity. |
6050 | */ | 6050 | */ |
6051 | static void init_sched_groups_capacity(int cpu, struct sched_domain *sd) | 6051 | static void init_sched_groups_capacity(int cpu, struct sched_domain *sd) |
6052 | { | 6052 | { |
6053 | struct sched_group *sg = sd->groups; | 6053 | struct sched_group *sg = sd->groups; |
6054 | 6054 | ||
6055 | WARN_ON(!sg); | 6055 | WARN_ON(!sg); |
6056 | 6056 | ||
6057 | do { | 6057 | do { |
6058 | sg->group_weight = cpumask_weight(sched_group_cpus(sg)); | 6058 | sg->group_weight = cpumask_weight(sched_group_cpus(sg)); |
6059 | sg = sg->next; | 6059 | sg = sg->next; |
6060 | } while (sg != sd->groups); | 6060 | } while (sg != sd->groups); |
6061 | 6061 | ||
6062 | if (cpu != group_balance_cpu(sg)) | 6062 | if (cpu != group_balance_cpu(sg)) |
6063 | return; | 6063 | return; |
6064 | 6064 | ||
6065 | update_group_capacity(sd, cpu); | 6065 | update_group_capacity(sd, cpu); |
6066 | atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight); | 6066 | atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight); |
6067 | } | 6067 | } |
6068 | 6068 | ||
6069 | /* | 6069 | /* |
6070 | * Initializers for schedule domains | 6070 | * Initializers for schedule domains |
6071 | * Non-inlined to reduce accumulated stack pressure in build_sched_domains() | 6071 | * Non-inlined to reduce accumulated stack pressure in build_sched_domains() |
6072 | */ | 6072 | */ |
6073 | 6073 | ||
6074 | static int default_relax_domain_level = -1; | 6074 | static int default_relax_domain_level = -1; |
6075 | int sched_domain_level_max; | 6075 | int sched_domain_level_max; |
6076 | 6076 | ||
6077 | static int __init setup_relax_domain_level(char *str) | 6077 | static int __init setup_relax_domain_level(char *str) |
6078 | { | 6078 | { |
6079 | if (kstrtoint(str, 0, &default_relax_domain_level)) | 6079 | if (kstrtoint(str, 0, &default_relax_domain_level)) |
6080 | pr_warn("Unable to set relax_domain_level\n"); | 6080 | pr_warn("Unable to set relax_domain_level\n"); |
6081 | 6081 | ||
6082 | return 1; | 6082 | return 1; |
6083 | } | 6083 | } |
6084 | __setup("relax_domain_level=", setup_relax_domain_level); | 6084 | __setup("relax_domain_level=", setup_relax_domain_level); |
6085 | 6085 | ||
6086 | static void set_domain_attribute(struct sched_domain *sd, | 6086 | static void set_domain_attribute(struct sched_domain *sd, |
6087 | struct sched_domain_attr *attr) | 6087 | struct sched_domain_attr *attr) |
6088 | { | 6088 | { |
6089 | int request; | 6089 | int request; |
6090 | 6090 | ||
6091 | if (!attr || attr->relax_domain_level < 0) { | 6091 | if (!attr || attr->relax_domain_level < 0) { |
6092 | if (default_relax_domain_level < 0) | 6092 | if (default_relax_domain_level < 0) |
6093 | return; | 6093 | return; |
6094 | else | 6094 | else |
6095 | request = default_relax_domain_level; | 6095 | request = default_relax_domain_level; |
6096 | } else | 6096 | } else |
6097 | request = attr->relax_domain_level; | 6097 | request = attr->relax_domain_level; |
6098 | if (request < sd->level) { | 6098 | if (request < sd->level) { |
6099 | /* turn off idle balance on this domain */ | 6099 | /* turn off idle balance on this domain */ |
6100 | sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); | 6100 | sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); |
6101 | } else { | 6101 | } else { |
6102 | /* turn on idle balance on this domain */ | 6102 | /* turn on idle balance on this domain */ |
6103 | sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); | 6103 | sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); |
6104 | } | 6104 | } |
6105 | } | 6105 | } |
6106 | 6106 | ||
6107 | static void __sdt_free(const struct cpumask *cpu_map); | 6107 | static void __sdt_free(const struct cpumask *cpu_map); |
6108 | static int __sdt_alloc(const struct cpumask *cpu_map); | 6108 | static int __sdt_alloc(const struct cpumask *cpu_map); |
6109 | 6109 | ||
6110 | static void __free_domain_allocs(struct s_data *d, enum s_alloc what, | 6110 | static void __free_domain_allocs(struct s_data *d, enum s_alloc what, |
6111 | const struct cpumask *cpu_map) | 6111 | const struct cpumask *cpu_map) |
6112 | { | 6112 | { |
6113 | switch (what) { | 6113 | switch (what) { |
6114 | case sa_rootdomain: | 6114 | case sa_rootdomain: |
6115 | if (!atomic_read(&d->rd->refcount)) | 6115 | if (!atomic_read(&d->rd->refcount)) |
6116 | free_rootdomain(&d->rd->rcu); /* fall through */ | 6116 | free_rootdomain(&d->rd->rcu); /* fall through */ |
6117 | case sa_sd: | 6117 | case sa_sd: |
6118 | free_percpu(d->sd); /* fall through */ | 6118 | free_percpu(d->sd); /* fall through */ |
6119 | case sa_sd_storage: | 6119 | case sa_sd_storage: |
6120 | __sdt_free(cpu_map); /* fall through */ | 6120 | __sdt_free(cpu_map); /* fall through */ |
6121 | case sa_none: | 6121 | case sa_none: |
6122 | break; | 6122 | break; |
6123 | } | 6123 | } |
6124 | } | 6124 | } |
6125 | 6125 | ||
6126 | static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, | 6126 | static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, |
6127 | const struct cpumask *cpu_map) | 6127 | const struct cpumask *cpu_map) |
6128 | { | 6128 | { |
6129 | memset(d, 0, sizeof(*d)); | 6129 | memset(d, 0, sizeof(*d)); |
6130 | 6130 | ||
6131 | if (__sdt_alloc(cpu_map)) | 6131 | if (__sdt_alloc(cpu_map)) |
6132 | return sa_sd_storage; | 6132 | return sa_sd_storage; |
6133 | d->sd = alloc_percpu(struct sched_domain *); | 6133 | d->sd = alloc_percpu(struct sched_domain *); |
6134 | if (!d->sd) | 6134 | if (!d->sd) |
6135 | return sa_sd_storage; | 6135 | return sa_sd_storage; |
6136 | d->rd = alloc_rootdomain(); | 6136 | d->rd = alloc_rootdomain(); |
6137 | if (!d->rd) | 6137 | if (!d->rd) |
6138 | return sa_sd; | 6138 | return sa_sd; |
6139 | return sa_rootdomain; | 6139 | return sa_rootdomain; |
6140 | } | 6140 | } |
6141 | 6141 | ||
6142 | /* | 6142 | /* |
6143 | * NULL the sd_data elements we've used to build the sched_domain and | 6143 | * NULL the sd_data elements we've used to build the sched_domain and |
6144 | * sched_group structure so that the subsequent __free_domain_allocs() | 6144 | * sched_group structure so that the subsequent __free_domain_allocs() |
6145 | * will not free the data we're using. | 6145 | * will not free the data we're using. |
6146 | */ | 6146 | */ |
6147 | static void claim_allocations(int cpu, struct sched_domain *sd) | 6147 | static void claim_allocations(int cpu, struct sched_domain *sd) |
6148 | { | 6148 | { |
6149 | struct sd_data *sdd = sd->private; | 6149 | struct sd_data *sdd = sd->private; |
6150 | 6150 | ||
6151 | WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); | 6151 | WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); |
6152 | *per_cpu_ptr(sdd->sd, cpu) = NULL; | 6152 | *per_cpu_ptr(sdd->sd, cpu) = NULL; |
6153 | 6153 | ||
6154 | if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) | 6154 | if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) |
6155 | *per_cpu_ptr(sdd->sg, cpu) = NULL; | 6155 | *per_cpu_ptr(sdd->sg, cpu) = NULL; |
6156 | 6156 | ||
6157 | if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref)) | 6157 | if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref)) |
6158 | *per_cpu_ptr(sdd->sgc, cpu) = NULL; | 6158 | *per_cpu_ptr(sdd->sgc, cpu) = NULL; |
6159 | } | 6159 | } |
6160 | 6160 | ||
6161 | #ifdef CONFIG_NUMA | 6161 | #ifdef CONFIG_NUMA |
6162 | static int sched_domains_numa_levels; | 6162 | static int sched_domains_numa_levels; |
6163 | enum numa_topology_type sched_numa_topology_type; | 6163 | enum numa_topology_type sched_numa_topology_type; |
6164 | static int *sched_domains_numa_distance; | 6164 | static int *sched_domains_numa_distance; |
6165 | int sched_max_numa_distance; | 6165 | int sched_max_numa_distance; |
6166 | static struct cpumask ***sched_domains_numa_masks; | 6166 | static struct cpumask ***sched_domains_numa_masks; |
6167 | static int sched_domains_curr_level; | 6167 | static int sched_domains_curr_level; |
6168 | #endif | 6168 | #endif |
6169 | 6169 | ||
6170 | /* | 6170 | /* |
6171 | * SD_flags allowed in topology descriptions. | 6171 | * SD_flags allowed in topology descriptions. |
6172 | * | 6172 | * |
6173 | * SD_SHARE_CPUCAPACITY - describes SMT topologies | 6173 | * SD_SHARE_CPUCAPACITY - describes SMT topologies |
6174 | * SD_SHARE_PKG_RESOURCES - describes shared caches | 6174 | * SD_SHARE_PKG_RESOURCES - describes shared caches |
6175 | * SD_NUMA - describes NUMA topologies | 6175 | * SD_NUMA - describes NUMA topologies |
6176 | * SD_SHARE_POWERDOMAIN - describes shared power domain | 6176 | * SD_SHARE_POWERDOMAIN - describes shared power domain |
6177 | * | 6177 | * |
6178 | * Odd one out: | 6178 | * Odd one out: |
6179 | * SD_ASYM_PACKING - describes SMT quirks | 6179 | * SD_ASYM_PACKING - describes SMT quirks |
6180 | */ | 6180 | */ |
6181 | #define TOPOLOGY_SD_FLAGS \ | 6181 | #define TOPOLOGY_SD_FLAGS \ |
6182 | (SD_SHARE_CPUCAPACITY | \ | 6182 | (SD_SHARE_CPUCAPACITY | \ |
6183 | SD_SHARE_PKG_RESOURCES | \ | 6183 | SD_SHARE_PKG_RESOURCES | \ |
6184 | SD_NUMA | \ | 6184 | SD_NUMA | \ |
6185 | SD_ASYM_PACKING | \ | 6185 | SD_ASYM_PACKING | \ |
6186 | SD_SHARE_POWERDOMAIN) | 6186 | SD_SHARE_POWERDOMAIN) |
6187 | 6187 | ||
6188 | static struct sched_domain * | 6188 | static struct sched_domain * |
6189 | sd_init(struct sched_domain_topology_level *tl, int cpu) | 6189 | sd_init(struct sched_domain_topology_level *tl, int cpu) |
6190 | { | 6190 | { |
6191 | struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); | 6191 | struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); |
6192 | int sd_weight, sd_flags = 0; | 6192 | int sd_weight, sd_flags = 0; |
6193 | 6193 | ||
6194 | #ifdef CONFIG_NUMA | 6194 | #ifdef CONFIG_NUMA |
6195 | /* | 6195 | /* |
6196 | * Ugly hack to pass state to sd_numa_mask()... | 6196 | * Ugly hack to pass state to sd_numa_mask()... |
6197 | */ | 6197 | */ |
6198 | sched_domains_curr_level = tl->numa_level; | 6198 | sched_domains_curr_level = tl->numa_level; |
6199 | #endif | 6199 | #endif |
6200 | 6200 | ||
6201 | sd_weight = cpumask_weight(tl->mask(cpu)); | 6201 | sd_weight = cpumask_weight(tl->mask(cpu)); |
6202 | 6202 | ||
6203 | if (tl->sd_flags) | 6203 | if (tl->sd_flags) |
6204 | sd_flags = (*tl->sd_flags)(); | 6204 | sd_flags = (*tl->sd_flags)(); |
6205 | if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS, | 6205 | if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS, |
6206 | "wrong sd_flags in topology description\n")) | 6206 | "wrong sd_flags in topology description\n")) |
6207 | sd_flags &= ~TOPOLOGY_SD_FLAGS; | 6207 | sd_flags &= ~TOPOLOGY_SD_FLAGS; |
6208 | 6208 | ||
6209 | *sd = (struct sched_domain){ | 6209 | *sd = (struct sched_domain){ |
6210 | .min_interval = sd_weight, | 6210 | .min_interval = sd_weight, |
6211 | .max_interval = 2*sd_weight, | 6211 | .max_interval = 2*sd_weight, |
6212 | .busy_factor = 32, | 6212 | .busy_factor = 32, |
6213 | .imbalance_pct = 125, | 6213 | .imbalance_pct = 125, |
6214 | 6214 | ||
6215 | .cache_nice_tries = 0, | 6215 | .cache_nice_tries = 0, |
6216 | .busy_idx = 0, | 6216 | .busy_idx = 0, |
6217 | .idle_idx = 0, | 6217 | .idle_idx = 0, |
6218 | .newidle_idx = 0, | 6218 | .newidle_idx = 0, |
6219 | .wake_idx = 0, | 6219 | .wake_idx = 0, |
6220 | .forkexec_idx = 0, | 6220 | .forkexec_idx = 0, |
6221 | 6221 | ||
6222 | .flags = 1*SD_LOAD_BALANCE | 6222 | .flags = 1*SD_LOAD_BALANCE |
6223 | | 1*SD_BALANCE_NEWIDLE | 6223 | | 1*SD_BALANCE_NEWIDLE |
6224 | | 1*SD_BALANCE_EXEC | 6224 | | 1*SD_BALANCE_EXEC |
6225 | | 1*SD_BALANCE_FORK | 6225 | | 1*SD_BALANCE_FORK |
6226 | | 0*SD_BALANCE_WAKE | 6226 | | 0*SD_BALANCE_WAKE |
6227 | | 1*SD_WAKE_AFFINE | 6227 | | 1*SD_WAKE_AFFINE |
6228 | | 0*SD_SHARE_CPUCAPACITY | 6228 | | 0*SD_SHARE_CPUCAPACITY |
6229 | | 0*SD_SHARE_PKG_RESOURCES | 6229 | | 0*SD_SHARE_PKG_RESOURCES |
6230 | | 0*SD_SERIALIZE | 6230 | | 0*SD_SERIALIZE |
6231 | | 0*SD_PREFER_SIBLING | 6231 | | 0*SD_PREFER_SIBLING |
6232 | | 0*SD_NUMA | 6232 | | 0*SD_NUMA |
6233 | | sd_flags | 6233 | | sd_flags |
6234 | , | 6234 | , |
6235 | 6235 | ||
6236 | .last_balance = jiffies, | 6236 | .last_balance = jiffies, |
6237 | .balance_interval = sd_weight, | 6237 | .balance_interval = sd_weight, |
6238 | .smt_gain = 0, | 6238 | .smt_gain = 0, |
6239 | .max_newidle_lb_cost = 0, | 6239 | .max_newidle_lb_cost = 0, |
6240 | .next_decay_max_lb_cost = jiffies, | 6240 | .next_decay_max_lb_cost = jiffies, |
6241 | #ifdef CONFIG_SCHED_DEBUG | 6241 | #ifdef CONFIG_SCHED_DEBUG |
6242 | .name = tl->name, | 6242 | .name = tl->name, |
6243 | #endif | 6243 | #endif |
6244 | }; | 6244 | }; |
6245 | 6245 | ||
6246 | /* | 6246 | /* |
6247 | * Convert topological properties into behaviour. | 6247 | * Convert topological properties into behaviour. |
6248 | */ | 6248 | */ |
6249 | 6249 | ||
6250 | if (sd->flags & SD_SHARE_CPUCAPACITY) { | 6250 | if (sd->flags & SD_SHARE_CPUCAPACITY) { |
6251 | sd->imbalance_pct = 110; | 6251 | sd->imbalance_pct = 110; |
6252 | sd->smt_gain = 1178; /* ~15% */ | 6252 | sd->smt_gain = 1178; /* ~15% */ |
6253 | 6253 | ||
6254 | } else if (sd->flags & SD_SHARE_PKG_RESOURCES) { | 6254 | } else if (sd->flags & SD_SHARE_PKG_RESOURCES) { |
6255 | sd->imbalance_pct = 117; | 6255 | sd->imbalance_pct = 117; |
6256 | sd->cache_nice_tries = 1; | 6256 | sd->cache_nice_tries = 1; |
6257 | sd->busy_idx = 2; | 6257 | sd->busy_idx = 2; |
6258 | 6258 | ||
6259 | #ifdef CONFIG_NUMA | 6259 | #ifdef CONFIG_NUMA |
6260 | } else if (sd->flags & SD_NUMA) { | 6260 | } else if (sd->flags & SD_NUMA) { |
6261 | sd->cache_nice_tries = 2; | 6261 | sd->cache_nice_tries = 2; |
6262 | sd->busy_idx = 3; | 6262 | sd->busy_idx = 3; |
6263 | sd->idle_idx = 2; | 6263 | sd->idle_idx = 2; |
6264 | 6264 | ||
6265 | sd->flags |= SD_SERIALIZE; | 6265 | sd->flags |= SD_SERIALIZE; |
6266 | if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) { | 6266 | if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) { |
6267 | sd->flags &= ~(SD_BALANCE_EXEC | | 6267 | sd->flags &= ~(SD_BALANCE_EXEC | |
6268 | SD_BALANCE_FORK | | 6268 | SD_BALANCE_FORK | |
6269 | SD_WAKE_AFFINE); | 6269 | SD_WAKE_AFFINE); |
6270 | } | 6270 | } |
6271 | 6271 | ||
6272 | #endif | 6272 | #endif |
6273 | } else { | 6273 | } else { |
6274 | sd->flags |= SD_PREFER_SIBLING; | 6274 | sd->flags |= SD_PREFER_SIBLING; |
6275 | sd->cache_nice_tries = 1; | 6275 | sd->cache_nice_tries = 1; |
6276 | sd->busy_idx = 2; | 6276 | sd->busy_idx = 2; |
6277 | sd->idle_idx = 1; | 6277 | sd->idle_idx = 1; |
6278 | } | 6278 | } |
6279 | 6279 | ||
6280 | sd->private = &tl->data; | 6280 | sd->private = &tl->data; |
6281 | 6281 | ||
6282 | return sd; | 6282 | return sd; |
6283 | } | 6283 | } |
6284 | 6284 | ||
6285 | /* | 6285 | /* |
6286 | * Topology list, bottom-up. | 6286 | * Topology list, bottom-up. |
6287 | */ | 6287 | */ |
6288 | static struct sched_domain_topology_level default_topology[] = { | 6288 | static struct sched_domain_topology_level default_topology[] = { |
6289 | #ifdef CONFIG_SCHED_SMT | 6289 | #ifdef CONFIG_SCHED_SMT |
6290 | { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) }, | 6290 | { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) }, |
6291 | #endif | 6291 | #endif |
6292 | #ifdef CONFIG_SCHED_MC | 6292 | #ifdef CONFIG_SCHED_MC |
6293 | { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) }, | 6293 | { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) }, |
6294 | #endif | 6294 | #endif |
6295 | { cpu_cpu_mask, SD_INIT_NAME(DIE) }, | 6295 | { cpu_cpu_mask, SD_INIT_NAME(DIE) }, |
6296 | { NULL, }, | 6296 | { NULL, }, |
6297 | }; | 6297 | }; |
6298 | 6298 | ||
6299 | struct sched_domain_topology_level *sched_domain_topology = default_topology; | 6299 | struct sched_domain_topology_level *sched_domain_topology = default_topology; |
6300 | 6300 | ||
6301 | #define for_each_sd_topology(tl) \ | 6301 | #define for_each_sd_topology(tl) \ |
6302 | for (tl = sched_domain_topology; tl->mask; tl++) | 6302 | for (tl = sched_domain_topology; tl->mask; tl++) |
6303 | 6303 | ||
6304 | void set_sched_topology(struct sched_domain_topology_level *tl) | 6304 | void set_sched_topology(struct sched_domain_topology_level *tl) |
6305 | { | 6305 | { |
6306 | sched_domain_topology = tl; | 6306 | sched_domain_topology = tl; |
6307 | } | 6307 | } |
6308 | 6308 | ||
6309 | #ifdef CONFIG_NUMA | 6309 | #ifdef CONFIG_NUMA |
6310 | 6310 | ||
6311 | static const struct cpumask *sd_numa_mask(int cpu) | 6311 | static const struct cpumask *sd_numa_mask(int cpu) |
6312 | { | 6312 | { |
6313 | return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; | 6313 | return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; |
6314 | } | 6314 | } |
6315 | 6315 | ||
6316 | static void sched_numa_warn(const char *str) | 6316 | static void sched_numa_warn(const char *str) |
6317 | { | 6317 | { |
6318 | static int done = false; | 6318 | static int done = false; |
6319 | int i,j; | 6319 | int i,j; |
6320 | 6320 | ||
6321 | if (done) | 6321 | if (done) |
6322 | return; | 6322 | return; |
6323 | 6323 | ||
6324 | done = true; | 6324 | done = true; |
6325 | 6325 | ||
6326 | printk(KERN_WARNING "ERROR: %s\n\n", str); | 6326 | printk(KERN_WARNING "ERROR: %s\n\n", str); |
6327 | 6327 | ||
6328 | for (i = 0; i < nr_node_ids; i++) { | 6328 | for (i = 0; i < nr_node_ids; i++) { |
6329 | printk(KERN_WARNING " "); | 6329 | printk(KERN_WARNING " "); |
6330 | for (j = 0; j < nr_node_ids; j++) | 6330 | for (j = 0; j < nr_node_ids; j++) |
6331 | printk(KERN_CONT "%02d ", node_distance(i,j)); | 6331 | printk(KERN_CONT "%02d ", node_distance(i,j)); |
6332 | printk(KERN_CONT "\n"); | 6332 | printk(KERN_CONT "\n"); |
6333 | } | 6333 | } |
6334 | printk(KERN_WARNING "\n"); | 6334 | printk(KERN_WARNING "\n"); |
6335 | } | 6335 | } |
6336 | 6336 | ||
6337 | bool find_numa_distance(int distance) | 6337 | bool find_numa_distance(int distance) |
6338 | { | 6338 | { |
6339 | int i; | 6339 | int i; |
6340 | 6340 | ||
6341 | if (distance == node_distance(0, 0)) | 6341 | if (distance == node_distance(0, 0)) |
6342 | return true; | 6342 | return true; |
6343 | 6343 | ||
6344 | for (i = 0; i < sched_domains_numa_levels; i++) { | 6344 | for (i = 0; i < sched_domains_numa_levels; i++) { |
6345 | if (sched_domains_numa_distance[i] == distance) | 6345 | if (sched_domains_numa_distance[i] == distance) |
6346 | return true; | 6346 | return true; |
6347 | } | 6347 | } |
6348 | 6348 | ||
6349 | return false; | 6349 | return false; |
6350 | } | 6350 | } |
6351 | 6351 | ||
6352 | /* | 6352 | /* |
6353 | * A system can have three types of NUMA topology: | 6353 | * A system can have three types of NUMA topology: |
6354 | * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system | 6354 | * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system |
6355 | * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes | 6355 | * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes |
6356 | * NUMA_BACKPLANE: nodes can reach other nodes through a backplane | 6356 | * NUMA_BACKPLANE: nodes can reach other nodes through a backplane |
6357 | * | 6357 | * |
6358 | * The difference between a glueless mesh topology and a backplane | 6358 | * The difference between a glueless mesh topology and a backplane |
6359 | * topology lies in whether communication between not directly | 6359 | * topology lies in whether communication between not directly |
6360 | * connected nodes goes through intermediary nodes (where programs | 6360 | * connected nodes goes through intermediary nodes (where programs |
6361 | * could run), or through backplane controllers. This affects | 6361 | * could run), or through backplane controllers. This affects |
6362 | * placement of programs. | 6362 | * placement of programs. |
6363 | * | 6363 | * |
6364 | * The type of topology can be discerned with the following tests: | 6364 | * The type of topology can be discerned with the following tests: |
6365 | * - If the maximum distance between any nodes is 1 hop, the system | 6365 | * - If the maximum distance between any nodes is 1 hop, the system |
6366 | * is directly connected. | 6366 | * is directly connected. |
6367 | * - If for two nodes A and B, located N > 1 hops away from each other, | 6367 | * - If for two nodes A and B, located N > 1 hops away from each other, |
6368 | * there is an intermediary node C, which is < N hops away from both | 6368 | * there is an intermediary node C, which is < N hops away from both |
6369 | * nodes A and B, the system is a glueless mesh. | 6369 | * nodes A and B, the system is a glueless mesh. |
6370 | */ | 6370 | */ |
6371 | static void init_numa_topology_type(void) | 6371 | static void init_numa_topology_type(void) |
6372 | { | 6372 | { |
6373 | int a, b, c, n; | 6373 | int a, b, c, n; |
6374 | 6374 | ||
6375 | n = sched_max_numa_distance; | 6375 | n = sched_max_numa_distance; |
6376 | 6376 | ||
6377 | if (n <= 1) | 6377 | if (n <= 1) |
6378 | sched_numa_topology_type = NUMA_DIRECT; | 6378 | sched_numa_topology_type = NUMA_DIRECT; |
6379 | 6379 | ||
6380 | for_each_online_node(a) { | 6380 | for_each_online_node(a) { |
6381 | for_each_online_node(b) { | 6381 | for_each_online_node(b) { |
6382 | /* Find two nodes furthest removed from each other. */ | 6382 | /* Find two nodes furthest removed from each other. */ |
6383 | if (node_distance(a, b) < n) | 6383 | if (node_distance(a, b) < n) |
6384 | continue; | 6384 | continue; |
6385 | 6385 | ||
6386 | /* Is there an intermediary node between a and b? */ | 6386 | /* Is there an intermediary node between a and b? */ |
6387 | for_each_online_node(c) { | 6387 | for_each_online_node(c) { |
6388 | if (node_distance(a, c) < n && | 6388 | if (node_distance(a, c) < n && |
6389 | node_distance(b, c) < n) { | 6389 | node_distance(b, c) < n) { |
6390 | sched_numa_topology_type = | 6390 | sched_numa_topology_type = |
6391 | NUMA_GLUELESS_MESH; | 6391 | NUMA_GLUELESS_MESH; |
6392 | return; | 6392 | return; |
6393 | } | 6393 | } |
6394 | } | 6394 | } |
6395 | 6395 | ||
6396 | sched_numa_topology_type = NUMA_BACKPLANE; | 6396 | sched_numa_topology_type = NUMA_BACKPLANE; |
6397 | return; | 6397 | return; |
6398 | } | 6398 | } |
6399 | } | 6399 | } |
6400 | } | 6400 | } |
6401 | 6401 | ||
6402 | static void sched_init_numa(void) | 6402 | static void sched_init_numa(void) |
6403 | { | 6403 | { |
6404 | int next_distance, curr_distance = node_distance(0, 0); | 6404 | int next_distance, curr_distance = node_distance(0, 0); |
6405 | struct sched_domain_topology_level *tl; | 6405 | struct sched_domain_topology_level *tl; |
6406 | int level = 0; | 6406 | int level = 0; |
6407 | int i, j, k; | 6407 | int i, j, k; |
6408 | 6408 | ||
6409 | sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); | 6409 | sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); |
6410 | if (!sched_domains_numa_distance) | 6410 | if (!sched_domains_numa_distance) |
6411 | return; | 6411 | return; |
6412 | 6412 | ||
6413 | /* | 6413 | /* |
6414 | * O(nr_nodes^2) deduplicating selection sort -- in order to find the | 6414 | * O(nr_nodes^2) deduplicating selection sort -- in order to find the |
6415 | * unique distances in the node_distance() table. | 6415 | * unique distances in the node_distance() table. |
6416 | * | 6416 | * |
6417 | * Assumes node_distance(0,j) includes all distances in | 6417 | * Assumes node_distance(0,j) includes all distances in |
6418 | * node_distance(i,j) in order to avoid cubic time. | 6418 | * node_distance(i,j) in order to avoid cubic time. |
6419 | */ | 6419 | */ |
6420 | next_distance = curr_distance; | 6420 | next_distance = curr_distance; |
6421 | for (i = 0; i < nr_node_ids; i++) { | 6421 | for (i = 0; i < nr_node_ids; i++) { |
6422 | for (j = 0; j < nr_node_ids; j++) { | 6422 | for (j = 0; j < nr_node_ids; j++) { |
6423 | for (k = 0; k < nr_node_ids; k++) { | 6423 | for (k = 0; k < nr_node_ids; k++) { |
6424 | int distance = node_distance(i, k); | 6424 | int distance = node_distance(i, k); |
6425 | 6425 | ||
6426 | if (distance > curr_distance && | 6426 | if (distance > curr_distance && |
6427 | (distance < next_distance || | 6427 | (distance < next_distance || |
6428 | next_distance == curr_distance)) | 6428 | next_distance == curr_distance)) |
6429 | next_distance = distance; | 6429 | next_distance = distance; |
6430 | 6430 | ||
6431 | /* | 6431 | /* |
6432 | * While not a strong assumption it would be nice to know | 6432 | * While not a strong assumption it would be nice to know |
6433 | * about cases where if node A is connected to B, B is not | 6433 | * about cases where if node A is connected to B, B is not |
6434 | * equally connected to A. | 6434 | * equally connected to A. |
6435 | */ | 6435 | */ |
6436 | if (sched_debug() && node_distance(k, i) != distance) | 6436 | if (sched_debug() && node_distance(k, i) != distance) |
6437 | sched_numa_warn("Node-distance not symmetric"); | 6437 | sched_numa_warn("Node-distance not symmetric"); |
6438 | 6438 | ||
6439 | if (sched_debug() && i && !find_numa_distance(distance)) | 6439 | if (sched_debug() && i && !find_numa_distance(distance)) |
6440 | sched_numa_warn("Node-0 not representative"); | 6440 | sched_numa_warn("Node-0 not representative"); |
6441 | } | 6441 | } |
6442 | if (next_distance != curr_distance) { | 6442 | if (next_distance != curr_distance) { |
6443 | sched_domains_numa_distance[level++] = next_distance; | 6443 | sched_domains_numa_distance[level++] = next_distance; |
6444 | sched_domains_numa_levels = level; | 6444 | sched_domains_numa_levels = level; |
6445 | curr_distance = next_distance; | 6445 | curr_distance = next_distance; |
6446 | } else break; | 6446 | } else break; |
6447 | } | 6447 | } |
6448 | 6448 | ||
6449 | /* | 6449 | /* |
6450 | * In case of sched_debug() we verify the above assumption. | 6450 | * In case of sched_debug() we verify the above assumption. |
6451 | */ | 6451 | */ |
6452 | if (!sched_debug()) | 6452 | if (!sched_debug()) |
6453 | break; | 6453 | break; |
6454 | } | 6454 | } |
6455 | 6455 | ||
6456 | if (!level) | 6456 | if (!level) |
6457 | return; | 6457 | return; |
6458 | 6458 | ||
6459 | /* | 6459 | /* |
6460 | * 'level' contains the number of unique distances, excluding the | 6460 | * 'level' contains the number of unique distances, excluding the |
6461 | * identity distance node_distance(i,i). | 6461 | * identity distance node_distance(i,i). |
6462 | * | 6462 | * |
6463 | * The sched_domains_numa_distance[] array includes the actual distance | 6463 | * The sched_domains_numa_distance[] array includes the actual distance |
6464 | * numbers. | 6464 | * numbers. |
6465 | */ | 6465 | */ |
6466 | 6466 | ||
6467 | /* | 6467 | /* |
6468 | * Here, we should temporarily reset sched_domains_numa_levels to 0. | 6468 | * Here, we should temporarily reset sched_domains_numa_levels to 0. |
6469 | * If it fails to allocate memory for array sched_domains_numa_masks[][], | 6469 | * If it fails to allocate memory for array sched_domains_numa_masks[][], |
6470 | * the array will contain less then 'level' members. This could be | 6470 | * the array will contain less then 'level' members. This could be |
6471 | * dangerous when we use it to iterate array sched_domains_numa_masks[][] | 6471 | * dangerous when we use it to iterate array sched_domains_numa_masks[][] |
6472 | * in other functions. | 6472 | * in other functions. |
6473 | * | 6473 | * |
6474 | * We reset it to 'level' at the end of this function. | 6474 | * We reset it to 'level' at the end of this function. |
6475 | */ | 6475 | */ |
6476 | sched_domains_numa_levels = 0; | 6476 | sched_domains_numa_levels = 0; |
6477 | 6477 | ||
6478 | sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); | 6478 | sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); |
6479 | if (!sched_domains_numa_masks) | 6479 | if (!sched_domains_numa_masks) |
6480 | return; | 6480 | return; |
6481 | 6481 | ||
6482 | /* | 6482 | /* |
6483 | * Now for each level, construct a mask per node which contains all | 6483 | * Now for each level, construct a mask per node which contains all |
6484 | * cpus of nodes that are that many hops away from us. | 6484 | * cpus of nodes that are that many hops away from us. |
6485 | */ | 6485 | */ |
6486 | for (i = 0; i < level; i++) { | 6486 | for (i = 0; i < level; i++) { |
6487 | sched_domains_numa_masks[i] = | 6487 | sched_domains_numa_masks[i] = |
6488 | kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); | 6488 | kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); |
6489 | if (!sched_domains_numa_masks[i]) | 6489 | if (!sched_domains_numa_masks[i]) |
6490 | return; | 6490 | return; |
6491 | 6491 | ||
6492 | for (j = 0; j < nr_node_ids; j++) { | 6492 | for (j = 0; j < nr_node_ids; j++) { |
6493 | struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); | 6493 | struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); |
6494 | if (!mask) | 6494 | if (!mask) |
6495 | return; | 6495 | return; |
6496 | 6496 | ||
6497 | sched_domains_numa_masks[i][j] = mask; | 6497 | sched_domains_numa_masks[i][j] = mask; |
6498 | 6498 | ||
6499 | for (k = 0; k < nr_node_ids; k++) { | 6499 | for (k = 0; k < nr_node_ids; k++) { |
6500 | if (node_distance(j, k) > sched_domains_numa_distance[i]) | 6500 | if (node_distance(j, k) > sched_domains_numa_distance[i]) |
6501 | continue; | 6501 | continue; |
6502 | 6502 | ||
6503 | cpumask_or(mask, mask, cpumask_of_node(k)); | 6503 | cpumask_or(mask, mask, cpumask_of_node(k)); |
6504 | } | 6504 | } |
6505 | } | 6505 | } |
6506 | } | 6506 | } |
6507 | 6507 | ||
6508 | /* Compute default topology size */ | 6508 | /* Compute default topology size */ |
6509 | for (i = 0; sched_domain_topology[i].mask; i++); | 6509 | for (i = 0; sched_domain_topology[i].mask; i++); |
6510 | 6510 | ||
6511 | tl = kzalloc((i + level + 1) * | 6511 | tl = kzalloc((i + level + 1) * |
6512 | sizeof(struct sched_domain_topology_level), GFP_KERNEL); | 6512 | sizeof(struct sched_domain_topology_level), GFP_KERNEL); |
6513 | if (!tl) | 6513 | if (!tl) |
6514 | return; | 6514 | return; |
6515 | 6515 | ||
6516 | /* | 6516 | /* |
6517 | * Copy the default topology bits.. | 6517 | * Copy the default topology bits.. |
6518 | */ | 6518 | */ |
6519 | for (i = 0; sched_domain_topology[i].mask; i++) | 6519 | for (i = 0; sched_domain_topology[i].mask; i++) |
6520 | tl[i] = sched_domain_topology[i]; | 6520 | tl[i] = sched_domain_topology[i]; |
6521 | 6521 | ||
6522 | /* | 6522 | /* |
6523 | * .. and append 'j' levels of NUMA goodness. | 6523 | * .. and append 'j' levels of NUMA goodness. |
6524 | */ | 6524 | */ |
6525 | for (j = 0; j < level; i++, j++) { | 6525 | for (j = 0; j < level; i++, j++) { |
6526 | tl[i] = (struct sched_domain_topology_level){ | 6526 | tl[i] = (struct sched_domain_topology_level){ |
6527 | .mask = sd_numa_mask, | 6527 | .mask = sd_numa_mask, |
6528 | .sd_flags = cpu_numa_flags, | 6528 | .sd_flags = cpu_numa_flags, |
6529 | .flags = SDTL_OVERLAP, | 6529 | .flags = SDTL_OVERLAP, |
6530 | .numa_level = j, | 6530 | .numa_level = j, |
6531 | SD_INIT_NAME(NUMA) | 6531 | SD_INIT_NAME(NUMA) |
6532 | }; | 6532 | }; |
6533 | } | 6533 | } |
6534 | 6534 | ||
6535 | sched_domain_topology = tl; | 6535 | sched_domain_topology = tl; |
6536 | 6536 | ||
6537 | sched_domains_numa_levels = level; | 6537 | sched_domains_numa_levels = level; |
6538 | sched_max_numa_distance = sched_domains_numa_distance[level - 1]; | 6538 | sched_max_numa_distance = sched_domains_numa_distance[level - 1]; |
6539 | 6539 | ||
6540 | init_numa_topology_type(); | 6540 | init_numa_topology_type(); |
6541 | } | 6541 | } |
6542 | 6542 | ||
6543 | static void sched_domains_numa_masks_set(int cpu) | 6543 | static void sched_domains_numa_masks_set(int cpu) |
6544 | { | 6544 | { |
6545 | int i, j; | 6545 | int i, j; |
6546 | int node = cpu_to_node(cpu); | 6546 | int node = cpu_to_node(cpu); |
6547 | 6547 | ||
6548 | for (i = 0; i < sched_domains_numa_levels; i++) { | 6548 | for (i = 0; i < sched_domains_numa_levels; i++) { |
6549 | for (j = 0; j < nr_node_ids; j++) { | 6549 | for (j = 0; j < nr_node_ids; j++) { |
6550 | if (node_distance(j, node) <= sched_domains_numa_distance[i]) | 6550 | if (node_distance(j, node) <= sched_domains_numa_distance[i]) |
6551 | cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); | 6551 | cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); |
6552 | } | 6552 | } |
6553 | } | 6553 | } |
6554 | } | 6554 | } |
6555 | 6555 | ||
6556 | static void sched_domains_numa_masks_clear(int cpu) | 6556 | static void sched_domains_numa_masks_clear(int cpu) |
6557 | { | 6557 | { |
6558 | int i, j; | 6558 | int i, j; |
6559 | for (i = 0; i < sched_domains_numa_levels; i++) { | 6559 | for (i = 0; i < sched_domains_numa_levels; i++) { |
6560 | for (j = 0; j < nr_node_ids; j++) | 6560 | for (j = 0; j < nr_node_ids; j++) |
6561 | cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); | 6561 | cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); |
6562 | } | 6562 | } |
6563 | } | 6563 | } |
6564 | 6564 | ||
6565 | /* | 6565 | /* |
6566 | * Update sched_domains_numa_masks[level][node] array when new cpus | 6566 | * Update sched_domains_numa_masks[level][node] array when new cpus |
6567 | * are onlined. | 6567 | * are onlined. |
6568 | */ | 6568 | */ |
6569 | static int sched_domains_numa_masks_update(struct notifier_block *nfb, | 6569 | static int sched_domains_numa_masks_update(struct notifier_block *nfb, |
6570 | unsigned long action, | 6570 | unsigned long action, |
6571 | void *hcpu) | 6571 | void *hcpu) |
6572 | { | 6572 | { |
6573 | int cpu = (long)hcpu; | 6573 | int cpu = (long)hcpu; |
6574 | 6574 | ||
6575 | switch (action & ~CPU_TASKS_FROZEN) { | 6575 | switch (action & ~CPU_TASKS_FROZEN) { |
6576 | case CPU_ONLINE: | 6576 | case CPU_ONLINE: |
6577 | sched_domains_numa_masks_set(cpu); | 6577 | sched_domains_numa_masks_set(cpu); |
6578 | break; | 6578 | break; |
6579 | 6579 | ||
6580 | case CPU_DEAD: | 6580 | case CPU_DEAD: |
6581 | sched_domains_numa_masks_clear(cpu); | 6581 | sched_domains_numa_masks_clear(cpu); |
6582 | break; | 6582 | break; |
6583 | 6583 | ||
6584 | default: | 6584 | default: |
6585 | return NOTIFY_DONE; | 6585 | return NOTIFY_DONE; |
6586 | } | 6586 | } |
6587 | 6587 | ||
6588 | return NOTIFY_OK; | 6588 | return NOTIFY_OK; |
6589 | } | 6589 | } |
6590 | #else | 6590 | #else |
6591 | static inline void sched_init_numa(void) | 6591 | static inline void sched_init_numa(void) |
6592 | { | 6592 | { |
6593 | } | 6593 | } |
6594 | 6594 | ||
6595 | static int sched_domains_numa_masks_update(struct notifier_block *nfb, | 6595 | static int sched_domains_numa_masks_update(struct notifier_block *nfb, |
6596 | unsigned long action, | 6596 | unsigned long action, |
6597 | void *hcpu) | 6597 | void *hcpu) |
6598 | { | 6598 | { |
6599 | return 0; | 6599 | return 0; |
6600 | } | 6600 | } |
6601 | #endif /* CONFIG_NUMA */ | 6601 | #endif /* CONFIG_NUMA */ |
6602 | 6602 | ||
6603 | static int __sdt_alloc(const struct cpumask *cpu_map) | 6603 | static int __sdt_alloc(const struct cpumask *cpu_map) |
6604 | { | 6604 | { |
6605 | struct sched_domain_topology_level *tl; | 6605 | struct sched_domain_topology_level *tl; |
6606 | int j; | 6606 | int j; |
6607 | 6607 | ||
6608 | for_each_sd_topology(tl) { | 6608 | for_each_sd_topology(tl) { |
6609 | struct sd_data *sdd = &tl->data; | 6609 | struct sd_data *sdd = &tl->data; |
6610 | 6610 | ||
6611 | sdd->sd = alloc_percpu(struct sched_domain *); | 6611 | sdd->sd = alloc_percpu(struct sched_domain *); |
6612 | if (!sdd->sd) | 6612 | if (!sdd->sd) |
6613 | return -ENOMEM; | 6613 | return -ENOMEM; |
6614 | 6614 | ||
6615 | sdd->sg = alloc_percpu(struct sched_group *); | 6615 | sdd->sg = alloc_percpu(struct sched_group *); |
6616 | if (!sdd->sg) | 6616 | if (!sdd->sg) |
6617 | return -ENOMEM; | 6617 | return -ENOMEM; |
6618 | 6618 | ||
6619 | sdd->sgc = alloc_percpu(struct sched_group_capacity *); | 6619 | sdd->sgc = alloc_percpu(struct sched_group_capacity *); |
6620 | if (!sdd->sgc) | 6620 | if (!sdd->sgc) |
6621 | return -ENOMEM; | 6621 | return -ENOMEM; |
6622 | 6622 | ||
6623 | for_each_cpu(j, cpu_map) { | 6623 | for_each_cpu(j, cpu_map) { |
6624 | struct sched_domain *sd; | 6624 | struct sched_domain *sd; |
6625 | struct sched_group *sg; | 6625 | struct sched_group *sg; |
6626 | struct sched_group_capacity *sgc; | 6626 | struct sched_group_capacity *sgc; |
6627 | 6627 | ||
6628 | sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), | 6628 | sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), |
6629 | GFP_KERNEL, cpu_to_node(j)); | 6629 | GFP_KERNEL, cpu_to_node(j)); |
6630 | if (!sd) | 6630 | if (!sd) |
6631 | return -ENOMEM; | 6631 | return -ENOMEM; |
6632 | 6632 | ||
6633 | *per_cpu_ptr(sdd->sd, j) = sd; | 6633 | *per_cpu_ptr(sdd->sd, j) = sd; |
6634 | 6634 | ||
6635 | sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), | 6635 | sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), |
6636 | GFP_KERNEL, cpu_to_node(j)); | 6636 | GFP_KERNEL, cpu_to_node(j)); |
6637 | if (!sg) | 6637 | if (!sg) |
6638 | return -ENOMEM; | 6638 | return -ENOMEM; |
6639 | 6639 | ||
6640 | sg->next = sg; | 6640 | sg->next = sg; |
6641 | 6641 | ||
6642 | *per_cpu_ptr(sdd->sg, j) = sg; | 6642 | *per_cpu_ptr(sdd->sg, j) = sg; |
6643 | 6643 | ||
6644 | sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(), | 6644 | sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(), |
6645 | GFP_KERNEL, cpu_to_node(j)); | 6645 | GFP_KERNEL, cpu_to_node(j)); |
6646 | if (!sgc) | 6646 | if (!sgc) |
6647 | return -ENOMEM; | 6647 | return -ENOMEM; |
6648 | 6648 | ||
6649 | *per_cpu_ptr(sdd->sgc, j) = sgc; | 6649 | *per_cpu_ptr(sdd->sgc, j) = sgc; |
6650 | } | 6650 | } |
6651 | } | 6651 | } |
6652 | 6652 | ||
6653 | return 0; | 6653 | return 0; |
6654 | } | 6654 | } |
6655 | 6655 | ||
6656 | static void __sdt_free(const struct cpumask *cpu_map) | 6656 | static void __sdt_free(const struct cpumask *cpu_map) |
6657 | { | 6657 | { |
6658 | struct sched_domain_topology_level *tl; | 6658 | struct sched_domain_topology_level *tl; |
6659 | int j; | 6659 | int j; |
6660 | 6660 | ||
6661 | for_each_sd_topology(tl) { | 6661 | for_each_sd_topology(tl) { |
6662 | struct sd_data *sdd = &tl->data; | 6662 | struct sd_data *sdd = &tl->data; |
6663 | 6663 | ||
6664 | for_each_cpu(j, cpu_map) { | 6664 | for_each_cpu(j, cpu_map) { |
6665 | struct sched_domain *sd; | 6665 | struct sched_domain *sd; |
6666 | 6666 | ||
6667 | if (sdd->sd) { | 6667 | if (sdd->sd) { |
6668 | sd = *per_cpu_ptr(sdd->sd, j); | 6668 | sd = *per_cpu_ptr(sdd->sd, j); |
6669 | if (sd && (sd->flags & SD_OVERLAP)) | 6669 | if (sd && (sd->flags & SD_OVERLAP)) |
6670 | free_sched_groups(sd->groups, 0); | 6670 | free_sched_groups(sd->groups, 0); |
6671 | kfree(*per_cpu_ptr(sdd->sd, j)); | 6671 | kfree(*per_cpu_ptr(sdd->sd, j)); |
6672 | } | 6672 | } |
6673 | 6673 | ||
6674 | if (sdd->sg) | 6674 | if (sdd->sg) |
6675 | kfree(*per_cpu_ptr(sdd->sg, j)); | 6675 | kfree(*per_cpu_ptr(sdd->sg, j)); |
6676 | if (sdd->sgc) | 6676 | if (sdd->sgc) |
6677 | kfree(*per_cpu_ptr(sdd->sgc, j)); | 6677 | kfree(*per_cpu_ptr(sdd->sgc, j)); |
6678 | } | 6678 | } |
6679 | free_percpu(sdd->sd); | 6679 | free_percpu(sdd->sd); |
6680 | sdd->sd = NULL; | 6680 | sdd->sd = NULL; |
6681 | free_percpu(sdd->sg); | 6681 | free_percpu(sdd->sg); |
6682 | sdd->sg = NULL; | 6682 | sdd->sg = NULL; |
6683 | free_percpu(sdd->sgc); | 6683 | free_percpu(sdd->sgc); |
6684 | sdd->sgc = NULL; | 6684 | sdd->sgc = NULL; |
6685 | } | 6685 | } |
6686 | } | 6686 | } |
6687 | 6687 | ||
6688 | struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, | 6688 | struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, |
6689 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, | 6689 | const struct cpumask *cpu_map, struct sched_domain_attr *attr, |
6690 | struct sched_domain *child, int cpu) | 6690 | struct sched_domain *child, int cpu) |
6691 | { | 6691 | { |
6692 | struct sched_domain *sd = sd_init(tl, cpu); | 6692 | struct sched_domain *sd = sd_init(tl, cpu); |
6693 | if (!sd) | 6693 | if (!sd) |
6694 | return child; | 6694 | return child; |
6695 | 6695 | ||
6696 | cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); | 6696 | cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); |
6697 | if (child) { | 6697 | if (child) { |
6698 | sd->level = child->level + 1; | 6698 | sd->level = child->level + 1; |
6699 | sched_domain_level_max = max(sched_domain_level_max, sd->level); | 6699 | sched_domain_level_max = max(sched_domain_level_max, sd->level); |
6700 | child->parent = sd; | 6700 | child->parent = sd; |
6701 | sd->child = child; | 6701 | sd->child = child; |
6702 | 6702 | ||
6703 | if (!cpumask_subset(sched_domain_span(child), | 6703 | if (!cpumask_subset(sched_domain_span(child), |
6704 | sched_domain_span(sd))) { | 6704 | sched_domain_span(sd))) { |
6705 | pr_err("BUG: arch topology borken\n"); | 6705 | pr_err("BUG: arch topology borken\n"); |
6706 | #ifdef CONFIG_SCHED_DEBUG | 6706 | #ifdef CONFIG_SCHED_DEBUG |
6707 | pr_err(" the %s domain not a subset of the %s domain\n", | 6707 | pr_err(" the %s domain not a subset of the %s domain\n", |
6708 | child->name, sd->name); | 6708 | child->name, sd->name); |
6709 | #endif | 6709 | #endif |
6710 | /* Fixup, ensure @sd has at least @child cpus. */ | 6710 | /* Fixup, ensure @sd has at least @child cpus. */ |
6711 | cpumask_or(sched_domain_span(sd), | 6711 | cpumask_or(sched_domain_span(sd), |
6712 | sched_domain_span(sd), | 6712 | sched_domain_span(sd), |
6713 | sched_domain_span(child)); | 6713 | sched_domain_span(child)); |
6714 | } | 6714 | } |
6715 | 6715 | ||
6716 | } | 6716 | } |
6717 | set_domain_attribute(sd, attr); | 6717 | set_domain_attribute(sd, attr); |
6718 | 6718 | ||
6719 | return sd; | 6719 | return sd; |
6720 | } | 6720 | } |
6721 | 6721 | ||
6722 | /* | 6722 | /* |
6723 | * Build sched domains for a given set of cpus and attach the sched domains | 6723 | * Build sched domains for a given set of cpus and attach the sched domains |
6724 | * to the individual cpus | 6724 | * to the individual cpus |
6725 | */ | 6725 | */ |
6726 | static int build_sched_domains(const struct cpumask *cpu_map, | 6726 | static int build_sched_domains(const struct cpumask *cpu_map, |
6727 | struct sched_domain_attr *attr) | 6727 | struct sched_domain_attr *attr) |
6728 | { | 6728 | { |
6729 | enum s_alloc alloc_state; | 6729 | enum s_alloc alloc_state; |
6730 | struct sched_domain *sd; | 6730 | struct sched_domain *sd; |
6731 | struct s_data d; | 6731 | struct s_data d; |
6732 | int i, ret = -ENOMEM; | 6732 | int i, ret = -ENOMEM; |
6733 | 6733 | ||
6734 | alloc_state = __visit_domain_allocation_hell(&d, cpu_map); | 6734 | alloc_state = __visit_domain_allocation_hell(&d, cpu_map); |
6735 | if (alloc_state != sa_rootdomain) | 6735 | if (alloc_state != sa_rootdomain) |
6736 | goto error; | 6736 | goto error; |
6737 | 6737 | ||
6738 | /* Set up domains for cpus specified by the cpu_map. */ | 6738 | /* Set up domains for cpus specified by the cpu_map. */ |
6739 | for_each_cpu(i, cpu_map) { | 6739 | for_each_cpu(i, cpu_map) { |
6740 | struct sched_domain_topology_level *tl; | 6740 | struct sched_domain_topology_level *tl; |
6741 | 6741 | ||
6742 | sd = NULL; | 6742 | sd = NULL; |
6743 | for_each_sd_topology(tl) { | 6743 | for_each_sd_topology(tl) { |
6744 | sd = build_sched_domain(tl, cpu_map, attr, sd, i); | 6744 | sd = build_sched_domain(tl, cpu_map, attr, sd, i); |
6745 | if (tl == sched_domain_topology) | 6745 | if (tl == sched_domain_topology) |
6746 | *per_cpu_ptr(d.sd, i) = sd; | 6746 | *per_cpu_ptr(d.sd, i) = sd; |
6747 | if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) | 6747 | if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) |
6748 | sd->flags |= SD_OVERLAP; | 6748 | sd->flags |= SD_OVERLAP; |
6749 | if (cpumask_equal(cpu_map, sched_domain_span(sd))) | 6749 | if (cpumask_equal(cpu_map, sched_domain_span(sd))) |
6750 | break; | 6750 | break; |
6751 | } | 6751 | } |
6752 | } | 6752 | } |
6753 | 6753 | ||
6754 | /* Build the groups for the domains */ | 6754 | /* Build the groups for the domains */ |
6755 | for_each_cpu(i, cpu_map) { | 6755 | for_each_cpu(i, cpu_map) { |
6756 | for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { | 6756 | for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { |
6757 | sd->span_weight = cpumask_weight(sched_domain_span(sd)); | 6757 | sd->span_weight = cpumask_weight(sched_domain_span(sd)); |
6758 | if (sd->flags & SD_OVERLAP) { | 6758 | if (sd->flags & SD_OVERLAP) { |
6759 | if (build_overlap_sched_groups(sd, i)) | 6759 | if (build_overlap_sched_groups(sd, i)) |
6760 | goto error; | 6760 | goto error; |
6761 | } else { | 6761 | } else { |
6762 | if (build_sched_groups(sd, i)) | 6762 | if (build_sched_groups(sd, i)) |
6763 | goto error; | 6763 | goto error; |
6764 | } | 6764 | } |
6765 | } | 6765 | } |
6766 | } | 6766 | } |
6767 | 6767 | ||
6768 | /* Calculate CPU capacity for physical packages and nodes */ | 6768 | /* Calculate CPU capacity for physical packages and nodes */ |
6769 | for (i = nr_cpumask_bits-1; i >= 0; i--) { | 6769 | for (i = nr_cpumask_bits-1; i >= 0; i--) { |
6770 | if (!cpumask_test_cpu(i, cpu_map)) | 6770 | if (!cpumask_test_cpu(i, cpu_map)) |
6771 | continue; | 6771 | continue; |
6772 | 6772 | ||
6773 | for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { | 6773 | for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { |
6774 | claim_allocations(i, sd); | 6774 | claim_allocations(i, sd); |
6775 | init_sched_groups_capacity(i, sd); | 6775 | init_sched_groups_capacity(i, sd); |
6776 | } | 6776 | } |
6777 | } | 6777 | } |
6778 | 6778 | ||
6779 | /* Attach the domains */ | 6779 | /* Attach the domains */ |
6780 | rcu_read_lock(); | 6780 | rcu_read_lock(); |
6781 | for_each_cpu(i, cpu_map) { | 6781 | for_each_cpu(i, cpu_map) { |
6782 | sd = *per_cpu_ptr(d.sd, i); | 6782 | sd = *per_cpu_ptr(d.sd, i); |
6783 | cpu_attach_domain(sd, d.rd, i); | 6783 | cpu_attach_domain(sd, d.rd, i); |
6784 | } | 6784 | } |
6785 | rcu_read_unlock(); | 6785 | rcu_read_unlock(); |
6786 | 6786 | ||
6787 | ret = 0; | 6787 | ret = 0; |
6788 | error: | 6788 | error: |
6789 | __free_domain_allocs(&d, alloc_state, cpu_map); | 6789 | __free_domain_allocs(&d, alloc_state, cpu_map); |
6790 | return ret; | 6790 | return ret; |
6791 | } | 6791 | } |
6792 | 6792 | ||
6793 | static cpumask_var_t *doms_cur; /* current sched domains */ | 6793 | static cpumask_var_t *doms_cur; /* current sched domains */ |
6794 | static int ndoms_cur; /* number of sched domains in 'doms_cur' */ | 6794 | static int ndoms_cur; /* number of sched domains in 'doms_cur' */ |
6795 | static struct sched_domain_attr *dattr_cur; | 6795 | static struct sched_domain_attr *dattr_cur; |
6796 | /* attribues of custom domains in 'doms_cur' */ | 6796 | /* attribues of custom domains in 'doms_cur' */ |
6797 | 6797 | ||
6798 | /* | 6798 | /* |
6799 | * Special case: If a kmalloc of a doms_cur partition (array of | 6799 | * Special case: If a kmalloc of a doms_cur partition (array of |
6800 | * cpumask) fails, then fallback to a single sched domain, | 6800 | * cpumask) fails, then fallback to a single sched domain, |
6801 | * as determined by the single cpumask fallback_doms. | 6801 | * as determined by the single cpumask fallback_doms. |
6802 | */ | 6802 | */ |
6803 | static cpumask_var_t fallback_doms; | 6803 | static cpumask_var_t fallback_doms; |
6804 | 6804 | ||
6805 | /* | 6805 | /* |
6806 | * arch_update_cpu_topology lets virtualized architectures update the | 6806 | * arch_update_cpu_topology lets virtualized architectures update the |
6807 | * cpu core maps. It is supposed to return 1 if the topology changed | 6807 | * cpu core maps. It is supposed to return 1 if the topology changed |
6808 | * or 0 if it stayed the same. | 6808 | * or 0 if it stayed the same. |
6809 | */ | 6809 | */ |
6810 | int __weak arch_update_cpu_topology(void) | 6810 | int __weak arch_update_cpu_topology(void) |
6811 | { | 6811 | { |
6812 | return 0; | 6812 | return 0; |
6813 | } | 6813 | } |
6814 | 6814 | ||
6815 | cpumask_var_t *alloc_sched_domains(unsigned int ndoms) | 6815 | cpumask_var_t *alloc_sched_domains(unsigned int ndoms) |
6816 | { | 6816 | { |
6817 | int i; | 6817 | int i; |
6818 | cpumask_var_t *doms; | 6818 | cpumask_var_t *doms; |
6819 | 6819 | ||
6820 | doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); | 6820 | doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); |
6821 | if (!doms) | 6821 | if (!doms) |
6822 | return NULL; | 6822 | return NULL; |
6823 | for (i = 0; i < ndoms; i++) { | 6823 | for (i = 0; i < ndoms; i++) { |
6824 | if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { | 6824 | if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { |
6825 | free_sched_domains(doms, i); | 6825 | free_sched_domains(doms, i); |
6826 | return NULL; | 6826 | return NULL; |
6827 | } | 6827 | } |
6828 | } | 6828 | } |
6829 | return doms; | 6829 | return doms; |
6830 | } | 6830 | } |
6831 | 6831 | ||
6832 | void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) | 6832 | void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) |
6833 | { | 6833 | { |
6834 | unsigned int i; | 6834 | unsigned int i; |
6835 | for (i = 0; i < ndoms; i++) | 6835 | for (i = 0; i < ndoms; i++) |
6836 | free_cpumask_var(doms[i]); | 6836 | free_cpumask_var(doms[i]); |
6837 | kfree(doms); | 6837 | kfree(doms); |
6838 | } | 6838 | } |
6839 | 6839 | ||
6840 | /* | 6840 | /* |
6841 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. | 6841 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. |
6842 | * For now this just excludes isolated cpus, but could be used to | 6842 | * For now this just excludes isolated cpus, but could be used to |
6843 | * exclude other special cases in the future. | 6843 | * exclude other special cases in the future. |
6844 | */ | 6844 | */ |
6845 | static int init_sched_domains(const struct cpumask *cpu_map) | 6845 | static int init_sched_domains(const struct cpumask *cpu_map) |
6846 | { | 6846 | { |
6847 | int err; | 6847 | int err; |
6848 | 6848 | ||
6849 | arch_update_cpu_topology(); | 6849 | arch_update_cpu_topology(); |
6850 | ndoms_cur = 1; | 6850 | ndoms_cur = 1; |
6851 | doms_cur = alloc_sched_domains(ndoms_cur); | 6851 | doms_cur = alloc_sched_domains(ndoms_cur); |
6852 | if (!doms_cur) | 6852 | if (!doms_cur) |
6853 | doms_cur = &fallback_doms; | 6853 | doms_cur = &fallback_doms; |
6854 | cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); | 6854 | cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); |
6855 | err = build_sched_domains(doms_cur[0], NULL); | 6855 | err = build_sched_domains(doms_cur[0], NULL); |
6856 | register_sched_domain_sysctl(); | 6856 | register_sched_domain_sysctl(); |
6857 | 6857 | ||
6858 | return err; | 6858 | return err; |
6859 | } | 6859 | } |
6860 | 6860 | ||
6861 | /* | 6861 | /* |
6862 | * Detach sched domains from a group of cpus specified in cpu_map | 6862 | * Detach sched domains from a group of cpus specified in cpu_map |
6863 | * These cpus will now be attached to the NULL domain | 6863 | * These cpus will now be attached to the NULL domain |
6864 | */ | 6864 | */ |
6865 | static void detach_destroy_domains(const struct cpumask *cpu_map) | 6865 | static void detach_destroy_domains(const struct cpumask *cpu_map) |
6866 | { | 6866 | { |
6867 | int i; | 6867 | int i; |
6868 | 6868 | ||
6869 | rcu_read_lock(); | 6869 | rcu_read_lock(); |
6870 | for_each_cpu(i, cpu_map) | 6870 | for_each_cpu(i, cpu_map) |
6871 | cpu_attach_domain(NULL, &def_root_domain, i); | 6871 | cpu_attach_domain(NULL, &def_root_domain, i); |
6872 | rcu_read_unlock(); | 6872 | rcu_read_unlock(); |
6873 | } | 6873 | } |
6874 | 6874 | ||
6875 | /* handle null as "default" */ | 6875 | /* handle null as "default" */ |
6876 | static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, | 6876 | static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, |
6877 | struct sched_domain_attr *new, int idx_new) | 6877 | struct sched_domain_attr *new, int idx_new) |
6878 | { | 6878 | { |
6879 | struct sched_domain_attr tmp; | 6879 | struct sched_domain_attr tmp; |
6880 | 6880 | ||
6881 | /* fast path */ | 6881 | /* fast path */ |
6882 | if (!new && !cur) | 6882 | if (!new && !cur) |
6883 | return 1; | 6883 | return 1; |
6884 | 6884 | ||
6885 | tmp = SD_ATTR_INIT; | 6885 | tmp = SD_ATTR_INIT; |
6886 | return !memcmp(cur ? (cur + idx_cur) : &tmp, | 6886 | return !memcmp(cur ? (cur + idx_cur) : &tmp, |
6887 | new ? (new + idx_new) : &tmp, | 6887 | new ? (new + idx_new) : &tmp, |
6888 | sizeof(struct sched_domain_attr)); | 6888 | sizeof(struct sched_domain_attr)); |
6889 | } | 6889 | } |
6890 | 6890 | ||
6891 | /* | 6891 | /* |
6892 | * Partition sched domains as specified by the 'ndoms_new' | 6892 | * Partition sched domains as specified by the 'ndoms_new' |
6893 | * cpumasks in the array doms_new[] of cpumasks. This compares | 6893 | * cpumasks in the array doms_new[] of cpumasks. This compares |
6894 | * doms_new[] to the current sched domain partitioning, doms_cur[]. | 6894 | * doms_new[] to the current sched domain partitioning, doms_cur[]. |
6895 | * It destroys each deleted domain and builds each new domain. | 6895 | * It destroys each deleted domain and builds each new domain. |
6896 | * | 6896 | * |
6897 | * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. | 6897 | * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. |
6898 | * The masks don't intersect (don't overlap.) We should setup one | 6898 | * The masks don't intersect (don't overlap.) We should setup one |
6899 | * sched domain for each mask. CPUs not in any of the cpumasks will | 6899 | * sched domain for each mask. CPUs not in any of the cpumasks will |
6900 | * not be load balanced. If the same cpumask appears both in the | 6900 | * not be load balanced. If the same cpumask appears both in the |
6901 | * current 'doms_cur' domains and in the new 'doms_new', we can leave | 6901 | * current 'doms_cur' domains and in the new 'doms_new', we can leave |
6902 | * it as it is. | 6902 | * it as it is. |
6903 | * | 6903 | * |
6904 | * The passed in 'doms_new' should be allocated using | 6904 | * The passed in 'doms_new' should be allocated using |
6905 | * alloc_sched_domains. This routine takes ownership of it and will | 6905 | * alloc_sched_domains. This routine takes ownership of it and will |
6906 | * free_sched_domains it when done with it. If the caller failed the | 6906 | * free_sched_domains it when done with it. If the caller failed the |
6907 | * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, | 6907 | * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, |
6908 | * and partition_sched_domains() will fallback to the single partition | 6908 | * and partition_sched_domains() will fallback to the single partition |
6909 | * 'fallback_doms', it also forces the domains to be rebuilt. | 6909 | * 'fallback_doms', it also forces the domains to be rebuilt. |
6910 | * | 6910 | * |
6911 | * If doms_new == NULL it will be replaced with cpu_online_mask. | 6911 | * If doms_new == NULL it will be replaced with cpu_online_mask. |
6912 | * ndoms_new == 0 is a special case for destroying existing domains, | 6912 | * ndoms_new == 0 is a special case for destroying existing domains, |
6913 | * and it will not create the default domain. | 6913 | * and it will not create the default domain. |
6914 | * | 6914 | * |
6915 | * Call with hotplug lock held | 6915 | * Call with hotplug lock held |
6916 | */ | 6916 | */ |
6917 | void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], | 6917 | void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], |
6918 | struct sched_domain_attr *dattr_new) | 6918 | struct sched_domain_attr *dattr_new) |
6919 | { | 6919 | { |
6920 | int i, j, n; | 6920 | int i, j, n; |
6921 | int new_topology; | 6921 | int new_topology; |
6922 | 6922 | ||
6923 | mutex_lock(&sched_domains_mutex); | 6923 | mutex_lock(&sched_domains_mutex); |
6924 | 6924 | ||
6925 | /* always unregister in case we don't destroy any domains */ | 6925 | /* always unregister in case we don't destroy any domains */ |
6926 | unregister_sched_domain_sysctl(); | 6926 | unregister_sched_domain_sysctl(); |
6927 | 6927 | ||
6928 | /* Let architecture update cpu core mappings. */ | 6928 | /* Let architecture update cpu core mappings. */ |
6929 | new_topology = arch_update_cpu_topology(); | 6929 | new_topology = arch_update_cpu_topology(); |
6930 | 6930 | ||
6931 | n = doms_new ? ndoms_new : 0; | 6931 | n = doms_new ? ndoms_new : 0; |
6932 | 6932 | ||
6933 | /* Destroy deleted domains */ | 6933 | /* Destroy deleted domains */ |
6934 | for (i = 0; i < ndoms_cur; i++) { | 6934 | for (i = 0; i < ndoms_cur; i++) { |
6935 | for (j = 0; j < n && !new_topology; j++) { | 6935 | for (j = 0; j < n && !new_topology; j++) { |
6936 | if (cpumask_equal(doms_cur[i], doms_new[j]) | 6936 | if (cpumask_equal(doms_cur[i], doms_new[j]) |
6937 | && dattrs_equal(dattr_cur, i, dattr_new, j)) | 6937 | && dattrs_equal(dattr_cur, i, dattr_new, j)) |
6938 | goto match1; | 6938 | goto match1; |
6939 | } | 6939 | } |
6940 | /* no match - a current sched domain not in new doms_new[] */ | 6940 | /* no match - a current sched domain not in new doms_new[] */ |
6941 | detach_destroy_domains(doms_cur[i]); | 6941 | detach_destroy_domains(doms_cur[i]); |
6942 | match1: | 6942 | match1: |
6943 | ; | 6943 | ; |
6944 | } | 6944 | } |
6945 | 6945 | ||
6946 | n = ndoms_cur; | 6946 | n = ndoms_cur; |
6947 | if (doms_new == NULL) { | 6947 | if (doms_new == NULL) { |
6948 | n = 0; | 6948 | n = 0; |
6949 | doms_new = &fallback_doms; | 6949 | doms_new = &fallback_doms; |
6950 | cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); | 6950 | cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); |
6951 | WARN_ON_ONCE(dattr_new); | 6951 | WARN_ON_ONCE(dattr_new); |
6952 | } | 6952 | } |
6953 | 6953 | ||
6954 | /* Build new domains */ | 6954 | /* Build new domains */ |
6955 | for (i = 0; i < ndoms_new; i++) { | 6955 | for (i = 0; i < ndoms_new; i++) { |
6956 | for (j = 0; j < n && !new_topology; j++) { | 6956 | for (j = 0; j < n && !new_topology; j++) { |
6957 | if (cpumask_equal(doms_new[i], doms_cur[j]) | 6957 | if (cpumask_equal(doms_new[i], doms_cur[j]) |
6958 | && dattrs_equal(dattr_new, i, dattr_cur, j)) | 6958 | && dattrs_equal(dattr_new, i, dattr_cur, j)) |
6959 | goto match2; | 6959 | goto match2; |
6960 | } | 6960 | } |
6961 | /* no match - add a new doms_new */ | 6961 | /* no match - add a new doms_new */ |
6962 | build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); | 6962 | build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); |
6963 | match2: | 6963 | match2: |
6964 | ; | 6964 | ; |
6965 | } | 6965 | } |
6966 | 6966 | ||
6967 | /* Remember the new sched domains */ | 6967 | /* Remember the new sched domains */ |
6968 | if (doms_cur != &fallback_doms) | 6968 | if (doms_cur != &fallback_doms) |
6969 | free_sched_domains(doms_cur, ndoms_cur); | 6969 | free_sched_domains(doms_cur, ndoms_cur); |
6970 | kfree(dattr_cur); /* kfree(NULL) is safe */ | 6970 | kfree(dattr_cur); /* kfree(NULL) is safe */ |
6971 | doms_cur = doms_new; | 6971 | doms_cur = doms_new; |
6972 | dattr_cur = dattr_new; | 6972 | dattr_cur = dattr_new; |
6973 | ndoms_cur = ndoms_new; | 6973 | ndoms_cur = ndoms_new; |
6974 | 6974 | ||
6975 | register_sched_domain_sysctl(); | 6975 | register_sched_domain_sysctl(); |
6976 | 6976 | ||
6977 | mutex_unlock(&sched_domains_mutex); | 6977 | mutex_unlock(&sched_domains_mutex); |
6978 | } | 6978 | } |
6979 | 6979 | ||
6980 | static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ | 6980 | static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ |
6981 | 6981 | ||
6982 | /* | 6982 | /* |
6983 | * Update cpusets according to cpu_active mask. If cpusets are | 6983 | * Update cpusets according to cpu_active mask. If cpusets are |
6984 | * disabled, cpuset_update_active_cpus() becomes a simple wrapper | 6984 | * disabled, cpuset_update_active_cpus() becomes a simple wrapper |
6985 | * around partition_sched_domains(). | 6985 | * around partition_sched_domains(). |
6986 | * | 6986 | * |
6987 | * If we come here as part of a suspend/resume, don't touch cpusets because we | 6987 | * If we come here as part of a suspend/resume, don't touch cpusets because we |
6988 | * want to restore it back to its original state upon resume anyway. | 6988 | * want to restore it back to its original state upon resume anyway. |
6989 | */ | 6989 | */ |
6990 | static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, | 6990 | static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, |
6991 | void *hcpu) | 6991 | void *hcpu) |
6992 | { | 6992 | { |
6993 | switch (action) { | 6993 | switch (action) { |
6994 | case CPU_ONLINE_FROZEN: | 6994 | case CPU_ONLINE_FROZEN: |
6995 | case CPU_DOWN_FAILED_FROZEN: | 6995 | case CPU_DOWN_FAILED_FROZEN: |
6996 | 6996 | ||
6997 | /* | 6997 | /* |
6998 | * num_cpus_frozen tracks how many CPUs are involved in suspend | 6998 | * num_cpus_frozen tracks how many CPUs are involved in suspend |
6999 | * resume sequence. As long as this is not the last online | 6999 | * resume sequence. As long as this is not the last online |
7000 | * operation in the resume sequence, just build a single sched | 7000 | * operation in the resume sequence, just build a single sched |
7001 | * domain, ignoring cpusets. | 7001 | * domain, ignoring cpusets. |
7002 | */ | 7002 | */ |
7003 | num_cpus_frozen--; | 7003 | num_cpus_frozen--; |
7004 | if (likely(num_cpus_frozen)) { | 7004 | if (likely(num_cpus_frozen)) { |
7005 | partition_sched_domains(1, NULL, NULL); | 7005 | partition_sched_domains(1, NULL, NULL); |
7006 | break; | 7006 | break; |
7007 | } | 7007 | } |
7008 | 7008 | ||
7009 | /* | 7009 | /* |
7010 | * This is the last CPU online operation. So fall through and | 7010 | * This is the last CPU online operation. So fall through and |
7011 | * restore the original sched domains by considering the | 7011 | * restore the original sched domains by considering the |
7012 | * cpuset configurations. | 7012 | * cpuset configurations. |
7013 | */ | 7013 | */ |
7014 | 7014 | ||
7015 | case CPU_ONLINE: | 7015 | case CPU_ONLINE: |
7016 | case CPU_DOWN_FAILED: | 7016 | case CPU_DOWN_FAILED: |
7017 | cpuset_update_active_cpus(true); | 7017 | cpuset_update_active_cpus(true); |
7018 | break; | 7018 | break; |
7019 | default: | 7019 | default: |
7020 | return NOTIFY_DONE; | 7020 | return NOTIFY_DONE; |
7021 | } | 7021 | } |
7022 | return NOTIFY_OK; | 7022 | return NOTIFY_OK; |
7023 | } | 7023 | } |
7024 | 7024 | ||
7025 | static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, | 7025 | static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, |
7026 | void *hcpu) | 7026 | void *hcpu) |
7027 | { | 7027 | { |
7028 | switch (action) { | 7028 | switch (action) { |
7029 | case CPU_DOWN_PREPARE: | 7029 | case CPU_DOWN_PREPARE: |
7030 | cpuset_update_active_cpus(false); | 7030 | cpuset_update_active_cpus(false); |
7031 | break; | 7031 | break; |
7032 | case CPU_DOWN_PREPARE_FROZEN: | 7032 | case CPU_DOWN_PREPARE_FROZEN: |
7033 | num_cpus_frozen++; | 7033 | num_cpus_frozen++; |
7034 | partition_sched_domains(1, NULL, NULL); | 7034 | partition_sched_domains(1, NULL, NULL); |
7035 | break; | 7035 | break; |
7036 | default: | 7036 | default: |
7037 | return NOTIFY_DONE; | 7037 | return NOTIFY_DONE; |
7038 | } | 7038 | } |
7039 | return NOTIFY_OK; | 7039 | return NOTIFY_OK; |
7040 | } | 7040 | } |
7041 | 7041 | ||
7042 | void __init sched_init_smp(void) | 7042 | void __init sched_init_smp(void) |
7043 | { | 7043 | { |
7044 | cpumask_var_t non_isolated_cpus; | 7044 | cpumask_var_t non_isolated_cpus; |
7045 | 7045 | ||
7046 | alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); | 7046 | alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); |
7047 | alloc_cpumask_var(&fallback_doms, GFP_KERNEL); | 7047 | alloc_cpumask_var(&fallback_doms, GFP_KERNEL); |
7048 | 7048 | ||
7049 | sched_init_numa(); | 7049 | sched_init_numa(); |
7050 | 7050 | ||
7051 | /* | 7051 | /* |
7052 | * There's no userspace yet to cause hotplug operations; hence all the | 7052 | * There's no userspace yet to cause hotplug operations; hence all the |
7053 | * cpu masks are stable and all blatant races in the below code cannot | 7053 | * cpu masks are stable and all blatant races in the below code cannot |
7054 | * happen. | 7054 | * happen. |
7055 | */ | 7055 | */ |
7056 | mutex_lock(&sched_domains_mutex); | 7056 | mutex_lock(&sched_domains_mutex); |
7057 | init_sched_domains(cpu_active_mask); | 7057 | init_sched_domains(cpu_active_mask); |
7058 | cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); | 7058 | cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); |
7059 | if (cpumask_empty(non_isolated_cpus)) | 7059 | if (cpumask_empty(non_isolated_cpus)) |
7060 | cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); | 7060 | cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); |
7061 | mutex_unlock(&sched_domains_mutex); | 7061 | mutex_unlock(&sched_domains_mutex); |
7062 | 7062 | ||
7063 | hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); | 7063 | hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); |
7064 | hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); | 7064 | hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); |
7065 | hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); | 7065 | hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); |
7066 | 7066 | ||
7067 | init_hrtick(); | 7067 | init_hrtick(); |
7068 | 7068 | ||
7069 | /* Move init over to a non-isolated CPU */ | 7069 | /* Move init over to a non-isolated CPU */ |
7070 | if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) | 7070 | if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) |
7071 | BUG(); | 7071 | BUG(); |
7072 | sched_init_granularity(); | 7072 | sched_init_granularity(); |
7073 | free_cpumask_var(non_isolated_cpus); | 7073 | free_cpumask_var(non_isolated_cpus); |
7074 | 7074 | ||
7075 | init_sched_rt_class(); | 7075 | init_sched_rt_class(); |
7076 | init_sched_dl_class(); | 7076 | init_sched_dl_class(); |
7077 | } | 7077 | } |
7078 | #else | 7078 | #else |
7079 | void __init sched_init_smp(void) | 7079 | void __init sched_init_smp(void) |
7080 | { | 7080 | { |
7081 | sched_init_granularity(); | 7081 | sched_init_granularity(); |
7082 | } | 7082 | } |
7083 | #endif /* CONFIG_SMP */ | 7083 | #endif /* CONFIG_SMP */ |
7084 | 7084 | ||
7085 | const_debug unsigned int sysctl_timer_migration = 1; | 7085 | const_debug unsigned int sysctl_timer_migration = 1; |
7086 | 7086 | ||
7087 | int in_sched_functions(unsigned long addr) | 7087 | int in_sched_functions(unsigned long addr) |
7088 | { | 7088 | { |
7089 | return in_lock_functions(addr) || | 7089 | return in_lock_functions(addr) || |
7090 | (addr >= (unsigned long)__sched_text_start | 7090 | (addr >= (unsigned long)__sched_text_start |
7091 | && addr < (unsigned long)__sched_text_end); | 7091 | && addr < (unsigned long)__sched_text_end); |
7092 | } | 7092 | } |
7093 | 7093 | ||
7094 | #ifdef CONFIG_CGROUP_SCHED | 7094 | #ifdef CONFIG_CGROUP_SCHED |
7095 | /* | 7095 | /* |
7096 | * Default task group. | 7096 | * Default task group. |
7097 | * Every task in system belongs to this group at bootup. | 7097 | * Every task in system belongs to this group at bootup. |
7098 | */ | 7098 | */ |
7099 | struct task_group root_task_group; | 7099 | struct task_group root_task_group; |
7100 | LIST_HEAD(task_groups); | 7100 | LIST_HEAD(task_groups); |
7101 | #endif | 7101 | #endif |
7102 | 7102 | ||
7103 | DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); | 7103 | DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); |
7104 | 7104 | ||
7105 | void __init sched_init(void) | 7105 | void __init sched_init(void) |
7106 | { | 7106 | { |
7107 | int i, j; | 7107 | int i, j; |
7108 | unsigned long alloc_size = 0, ptr; | 7108 | unsigned long alloc_size = 0, ptr; |
7109 | 7109 | ||
7110 | #ifdef CONFIG_FAIR_GROUP_SCHED | 7110 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7111 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); | 7111 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); |
7112 | #endif | 7112 | #endif |
7113 | #ifdef CONFIG_RT_GROUP_SCHED | 7113 | #ifdef CONFIG_RT_GROUP_SCHED |
7114 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); | 7114 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); |
7115 | #endif | 7115 | #endif |
7116 | #ifdef CONFIG_CPUMASK_OFFSTACK | ||
7117 | alloc_size += num_possible_cpus() * cpumask_size(); | ||
7118 | #endif | ||
7119 | if (alloc_size) { | 7116 | if (alloc_size) { |
7120 | ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); | 7117 | ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); |
7121 | 7118 | ||
7122 | #ifdef CONFIG_FAIR_GROUP_SCHED | 7119 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7123 | root_task_group.se = (struct sched_entity **)ptr; | 7120 | root_task_group.se = (struct sched_entity **)ptr; |
7124 | ptr += nr_cpu_ids * sizeof(void **); | 7121 | ptr += nr_cpu_ids * sizeof(void **); |
7125 | 7122 | ||
7126 | root_task_group.cfs_rq = (struct cfs_rq **)ptr; | 7123 | root_task_group.cfs_rq = (struct cfs_rq **)ptr; |
7127 | ptr += nr_cpu_ids * sizeof(void **); | 7124 | ptr += nr_cpu_ids * sizeof(void **); |
7128 | 7125 | ||
7129 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | 7126 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
7130 | #ifdef CONFIG_RT_GROUP_SCHED | 7127 | #ifdef CONFIG_RT_GROUP_SCHED |
7131 | root_task_group.rt_se = (struct sched_rt_entity **)ptr; | 7128 | root_task_group.rt_se = (struct sched_rt_entity **)ptr; |
7132 | ptr += nr_cpu_ids * sizeof(void **); | 7129 | ptr += nr_cpu_ids * sizeof(void **); |
7133 | 7130 | ||
7134 | root_task_group.rt_rq = (struct rt_rq **)ptr; | 7131 | root_task_group.rt_rq = (struct rt_rq **)ptr; |
7135 | ptr += nr_cpu_ids * sizeof(void **); | 7132 | ptr += nr_cpu_ids * sizeof(void **); |
7136 | 7133 | ||
7137 | #endif /* CONFIG_RT_GROUP_SCHED */ | 7134 | #endif /* CONFIG_RT_GROUP_SCHED */ |
7135 | } | ||
7138 | #ifdef CONFIG_CPUMASK_OFFSTACK | 7136 | #ifdef CONFIG_CPUMASK_OFFSTACK |
7139 | for_each_possible_cpu(i) { | 7137 | for_each_possible_cpu(i) { |
7140 | per_cpu(load_balance_mask, i) = (void *)ptr; | 7138 | per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node( |
7141 | ptr += cpumask_size(); | 7139 | cpumask_size(), GFP_KERNEL, cpu_to_node(i)); |
7142 | } | ||
7143 | #endif /* CONFIG_CPUMASK_OFFSTACK */ | ||
7144 | } | 7140 | } |
7141 | #endif /* CONFIG_CPUMASK_OFFSTACK */ | ||
7145 | 7142 | ||
7146 | init_rt_bandwidth(&def_rt_bandwidth, | 7143 | init_rt_bandwidth(&def_rt_bandwidth, |
7147 | global_rt_period(), global_rt_runtime()); | 7144 | global_rt_period(), global_rt_runtime()); |
7148 | init_dl_bandwidth(&def_dl_bandwidth, | 7145 | init_dl_bandwidth(&def_dl_bandwidth, |
7149 | global_rt_period(), global_rt_runtime()); | 7146 | global_rt_period(), global_rt_runtime()); |
7150 | 7147 | ||
7151 | #ifdef CONFIG_SMP | 7148 | #ifdef CONFIG_SMP |
7152 | init_defrootdomain(); | 7149 | init_defrootdomain(); |
7153 | #endif | 7150 | #endif |
7154 | 7151 | ||
7155 | #ifdef CONFIG_RT_GROUP_SCHED | 7152 | #ifdef CONFIG_RT_GROUP_SCHED |
7156 | init_rt_bandwidth(&root_task_group.rt_bandwidth, | 7153 | init_rt_bandwidth(&root_task_group.rt_bandwidth, |
7157 | global_rt_period(), global_rt_runtime()); | 7154 | global_rt_period(), global_rt_runtime()); |
7158 | #endif /* CONFIG_RT_GROUP_SCHED */ | 7155 | #endif /* CONFIG_RT_GROUP_SCHED */ |
7159 | 7156 | ||
7160 | #ifdef CONFIG_CGROUP_SCHED | 7157 | #ifdef CONFIG_CGROUP_SCHED |
7161 | list_add(&root_task_group.list, &task_groups); | 7158 | list_add(&root_task_group.list, &task_groups); |
7162 | INIT_LIST_HEAD(&root_task_group.children); | 7159 | INIT_LIST_HEAD(&root_task_group.children); |
7163 | INIT_LIST_HEAD(&root_task_group.siblings); | 7160 | INIT_LIST_HEAD(&root_task_group.siblings); |
7164 | autogroup_init(&init_task); | 7161 | autogroup_init(&init_task); |
7165 | 7162 | ||
7166 | #endif /* CONFIG_CGROUP_SCHED */ | 7163 | #endif /* CONFIG_CGROUP_SCHED */ |
7167 | 7164 | ||
7168 | for_each_possible_cpu(i) { | 7165 | for_each_possible_cpu(i) { |
7169 | struct rq *rq; | 7166 | struct rq *rq; |
7170 | 7167 | ||
7171 | rq = cpu_rq(i); | 7168 | rq = cpu_rq(i); |
7172 | raw_spin_lock_init(&rq->lock); | 7169 | raw_spin_lock_init(&rq->lock); |
7173 | rq->nr_running = 0; | 7170 | rq->nr_running = 0; |
7174 | rq->calc_load_active = 0; | 7171 | rq->calc_load_active = 0; |
7175 | rq->calc_load_update = jiffies + LOAD_FREQ; | 7172 | rq->calc_load_update = jiffies + LOAD_FREQ; |
7176 | init_cfs_rq(&rq->cfs); | 7173 | init_cfs_rq(&rq->cfs); |
7177 | init_rt_rq(&rq->rt, rq); | 7174 | init_rt_rq(&rq->rt, rq); |
7178 | init_dl_rq(&rq->dl, rq); | 7175 | init_dl_rq(&rq->dl, rq); |
7179 | #ifdef CONFIG_FAIR_GROUP_SCHED | 7176 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7180 | root_task_group.shares = ROOT_TASK_GROUP_LOAD; | 7177 | root_task_group.shares = ROOT_TASK_GROUP_LOAD; |
7181 | INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); | 7178 | INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); |
7182 | /* | 7179 | /* |
7183 | * How much cpu bandwidth does root_task_group get? | 7180 | * How much cpu bandwidth does root_task_group get? |
7184 | * | 7181 | * |
7185 | * In case of task-groups formed thr' the cgroup filesystem, it | 7182 | * In case of task-groups formed thr' the cgroup filesystem, it |
7186 | * gets 100% of the cpu resources in the system. This overall | 7183 | * gets 100% of the cpu resources in the system. This overall |
7187 | * system cpu resource is divided among the tasks of | 7184 | * system cpu resource is divided among the tasks of |
7188 | * root_task_group and its child task-groups in a fair manner, | 7185 | * root_task_group and its child task-groups in a fair manner, |
7189 | * based on each entity's (task or task-group's) weight | 7186 | * based on each entity's (task or task-group's) weight |
7190 | * (se->load.weight). | 7187 | * (se->load.weight). |
7191 | * | 7188 | * |
7192 | * In other words, if root_task_group has 10 tasks of weight | 7189 | * In other words, if root_task_group has 10 tasks of weight |
7193 | * 1024) and two child groups A0 and A1 (of weight 1024 each), | 7190 | * 1024) and two child groups A0 and A1 (of weight 1024 each), |
7194 | * then A0's share of the cpu resource is: | 7191 | * then A0's share of the cpu resource is: |
7195 | * | 7192 | * |
7196 | * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% | 7193 | * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% |
7197 | * | 7194 | * |
7198 | * We achieve this by letting root_task_group's tasks sit | 7195 | * We achieve this by letting root_task_group's tasks sit |
7199 | * directly in rq->cfs (i.e root_task_group->se[] = NULL). | 7196 | * directly in rq->cfs (i.e root_task_group->se[] = NULL). |
7200 | */ | 7197 | */ |
7201 | init_cfs_bandwidth(&root_task_group.cfs_bandwidth); | 7198 | init_cfs_bandwidth(&root_task_group.cfs_bandwidth); |
7202 | init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); | 7199 | init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); |
7203 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | 7200 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
7204 | 7201 | ||
7205 | rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; | 7202 | rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; |
7206 | #ifdef CONFIG_RT_GROUP_SCHED | 7203 | #ifdef CONFIG_RT_GROUP_SCHED |
7207 | init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); | 7204 | init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); |
7208 | #endif | 7205 | #endif |
7209 | 7206 | ||
7210 | for (j = 0; j < CPU_LOAD_IDX_MAX; j++) | 7207 | for (j = 0; j < CPU_LOAD_IDX_MAX; j++) |
7211 | rq->cpu_load[j] = 0; | 7208 | rq->cpu_load[j] = 0; |
7212 | 7209 | ||
7213 | rq->last_load_update_tick = jiffies; | 7210 | rq->last_load_update_tick = jiffies; |
7214 | 7211 | ||
7215 | #ifdef CONFIG_SMP | 7212 | #ifdef CONFIG_SMP |
7216 | rq->sd = NULL; | 7213 | rq->sd = NULL; |
7217 | rq->rd = NULL; | 7214 | rq->rd = NULL; |
7218 | rq->cpu_capacity = SCHED_CAPACITY_SCALE; | 7215 | rq->cpu_capacity = SCHED_CAPACITY_SCALE; |
7219 | rq->post_schedule = 0; | 7216 | rq->post_schedule = 0; |
7220 | rq->active_balance = 0; | 7217 | rq->active_balance = 0; |
7221 | rq->next_balance = jiffies; | 7218 | rq->next_balance = jiffies; |
7222 | rq->push_cpu = 0; | 7219 | rq->push_cpu = 0; |
7223 | rq->cpu = i; | 7220 | rq->cpu = i; |
7224 | rq->online = 0; | 7221 | rq->online = 0; |
7225 | rq->idle_stamp = 0; | 7222 | rq->idle_stamp = 0; |
7226 | rq->avg_idle = 2*sysctl_sched_migration_cost; | 7223 | rq->avg_idle = 2*sysctl_sched_migration_cost; |
7227 | rq->max_idle_balance_cost = sysctl_sched_migration_cost; | 7224 | rq->max_idle_balance_cost = sysctl_sched_migration_cost; |
7228 | 7225 | ||
7229 | INIT_LIST_HEAD(&rq->cfs_tasks); | 7226 | INIT_LIST_HEAD(&rq->cfs_tasks); |
7230 | 7227 | ||
7231 | rq_attach_root(rq, &def_root_domain); | 7228 | rq_attach_root(rq, &def_root_domain); |
7232 | #ifdef CONFIG_NO_HZ_COMMON | 7229 | #ifdef CONFIG_NO_HZ_COMMON |
7233 | rq->nohz_flags = 0; | 7230 | rq->nohz_flags = 0; |
7234 | #endif | 7231 | #endif |
7235 | #ifdef CONFIG_NO_HZ_FULL | 7232 | #ifdef CONFIG_NO_HZ_FULL |
7236 | rq->last_sched_tick = 0; | 7233 | rq->last_sched_tick = 0; |
7237 | #endif | 7234 | #endif |
7238 | #endif | 7235 | #endif |
7239 | init_rq_hrtick(rq); | 7236 | init_rq_hrtick(rq); |
7240 | atomic_set(&rq->nr_iowait, 0); | 7237 | atomic_set(&rq->nr_iowait, 0); |
7241 | } | 7238 | } |
7242 | 7239 | ||
7243 | set_load_weight(&init_task); | 7240 | set_load_weight(&init_task); |
7244 | 7241 | ||
7245 | #ifdef CONFIG_PREEMPT_NOTIFIERS | 7242 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
7246 | INIT_HLIST_HEAD(&init_task.preempt_notifiers); | 7243 | INIT_HLIST_HEAD(&init_task.preempt_notifiers); |
7247 | #endif | 7244 | #endif |
7248 | 7245 | ||
7249 | /* | 7246 | /* |
7250 | * The boot idle thread does lazy MMU switching as well: | 7247 | * The boot idle thread does lazy MMU switching as well: |
7251 | */ | 7248 | */ |
7252 | atomic_inc(&init_mm.mm_count); | 7249 | atomic_inc(&init_mm.mm_count); |
7253 | enter_lazy_tlb(&init_mm, current); | 7250 | enter_lazy_tlb(&init_mm, current); |
7254 | 7251 | ||
7255 | /* | 7252 | /* |
7256 | * Make us the idle thread. Technically, schedule() should not be | 7253 | * Make us the idle thread. Technically, schedule() should not be |
7257 | * called from this thread, however somewhere below it might be, | 7254 | * called from this thread, however somewhere below it might be, |
7258 | * but because we are the idle thread, we just pick up running again | 7255 | * but because we are the idle thread, we just pick up running again |
7259 | * when this runqueue becomes "idle". | 7256 | * when this runqueue becomes "idle". |
7260 | */ | 7257 | */ |
7261 | init_idle(current, smp_processor_id()); | 7258 | init_idle(current, smp_processor_id()); |
7262 | 7259 | ||
7263 | calc_load_update = jiffies + LOAD_FREQ; | 7260 | calc_load_update = jiffies + LOAD_FREQ; |
7264 | 7261 | ||
7265 | /* | 7262 | /* |
7266 | * During early bootup we pretend to be a normal task: | 7263 | * During early bootup we pretend to be a normal task: |
7267 | */ | 7264 | */ |
7268 | current->sched_class = &fair_sched_class; | 7265 | current->sched_class = &fair_sched_class; |
7269 | 7266 | ||
7270 | #ifdef CONFIG_SMP | 7267 | #ifdef CONFIG_SMP |
7271 | zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); | 7268 | zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); |
7272 | /* May be allocated at isolcpus cmdline parse time */ | 7269 | /* May be allocated at isolcpus cmdline parse time */ |
7273 | if (cpu_isolated_map == NULL) | 7270 | if (cpu_isolated_map == NULL) |
7274 | zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); | 7271 | zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); |
7275 | idle_thread_set_boot_cpu(); | 7272 | idle_thread_set_boot_cpu(); |
7276 | set_cpu_rq_start_time(); | 7273 | set_cpu_rq_start_time(); |
7277 | #endif | 7274 | #endif |
7278 | init_sched_fair_class(); | 7275 | init_sched_fair_class(); |
7279 | 7276 | ||
7280 | scheduler_running = 1; | 7277 | scheduler_running = 1; |
7281 | } | 7278 | } |
7282 | 7279 | ||
7283 | #ifdef CONFIG_DEBUG_ATOMIC_SLEEP | 7280 | #ifdef CONFIG_DEBUG_ATOMIC_SLEEP |
7284 | static inline int preempt_count_equals(int preempt_offset) | 7281 | static inline int preempt_count_equals(int preempt_offset) |
7285 | { | 7282 | { |
7286 | int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); | 7283 | int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); |
7287 | 7284 | ||
7288 | return (nested == preempt_offset); | 7285 | return (nested == preempt_offset); |
7289 | } | 7286 | } |
7290 | 7287 | ||
7291 | void __might_sleep(const char *file, int line, int preempt_offset) | 7288 | void __might_sleep(const char *file, int line, int preempt_offset) |
7292 | { | 7289 | { |
7293 | /* | 7290 | /* |
7294 | * Blocking primitives will set (and therefore destroy) current->state, | 7291 | * Blocking primitives will set (and therefore destroy) current->state, |
7295 | * since we will exit with TASK_RUNNING make sure we enter with it, | 7292 | * since we will exit with TASK_RUNNING make sure we enter with it, |
7296 | * otherwise we will destroy state. | 7293 | * otherwise we will destroy state. |
7297 | */ | 7294 | */ |
7298 | if (WARN_ONCE(current->state != TASK_RUNNING, | 7295 | if (WARN_ONCE(current->state != TASK_RUNNING, |
7299 | "do not call blocking ops when !TASK_RUNNING; " | 7296 | "do not call blocking ops when !TASK_RUNNING; " |
7300 | "state=%lx set at [<%p>] %pS\n", | 7297 | "state=%lx set at [<%p>] %pS\n", |
7301 | current->state, | 7298 | current->state, |
7302 | (void *)current->task_state_change, | 7299 | (void *)current->task_state_change, |
7303 | (void *)current->task_state_change)) | 7300 | (void *)current->task_state_change)) |
7304 | __set_current_state(TASK_RUNNING); | 7301 | __set_current_state(TASK_RUNNING); |
7305 | 7302 | ||
7306 | ___might_sleep(file, line, preempt_offset); | 7303 | ___might_sleep(file, line, preempt_offset); |
7307 | } | 7304 | } |
7308 | EXPORT_SYMBOL(__might_sleep); | 7305 | EXPORT_SYMBOL(__might_sleep); |
7309 | 7306 | ||
7310 | void ___might_sleep(const char *file, int line, int preempt_offset) | 7307 | void ___might_sleep(const char *file, int line, int preempt_offset) |
7311 | { | 7308 | { |
7312 | static unsigned long prev_jiffy; /* ratelimiting */ | 7309 | static unsigned long prev_jiffy; /* ratelimiting */ |
7313 | 7310 | ||
7314 | rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ | 7311 | rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ |
7315 | if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && | 7312 | if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && |
7316 | !is_idle_task(current)) || | 7313 | !is_idle_task(current)) || |
7317 | system_state != SYSTEM_RUNNING || oops_in_progress) | 7314 | system_state != SYSTEM_RUNNING || oops_in_progress) |
7318 | return; | 7315 | return; |
7319 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) | 7316 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) |
7320 | return; | 7317 | return; |
7321 | prev_jiffy = jiffies; | 7318 | prev_jiffy = jiffies; |
7322 | 7319 | ||
7323 | printk(KERN_ERR | 7320 | printk(KERN_ERR |
7324 | "BUG: sleeping function called from invalid context at %s:%d\n", | 7321 | "BUG: sleeping function called from invalid context at %s:%d\n", |
7325 | file, line); | 7322 | file, line); |
7326 | printk(KERN_ERR | 7323 | printk(KERN_ERR |
7327 | "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", | 7324 | "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", |
7328 | in_atomic(), irqs_disabled(), | 7325 | in_atomic(), irqs_disabled(), |
7329 | current->pid, current->comm); | 7326 | current->pid, current->comm); |
7330 | 7327 | ||
7331 | debug_show_held_locks(current); | 7328 | debug_show_held_locks(current); |
7332 | if (irqs_disabled()) | 7329 | if (irqs_disabled()) |
7333 | print_irqtrace_events(current); | 7330 | print_irqtrace_events(current); |
7334 | #ifdef CONFIG_DEBUG_PREEMPT | 7331 | #ifdef CONFIG_DEBUG_PREEMPT |
7335 | if (!preempt_count_equals(preempt_offset)) { | 7332 | if (!preempt_count_equals(preempt_offset)) { |
7336 | pr_err("Preemption disabled at:"); | 7333 | pr_err("Preemption disabled at:"); |
7337 | print_ip_sym(current->preempt_disable_ip); | 7334 | print_ip_sym(current->preempt_disable_ip); |
7338 | pr_cont("\n"); | 7335 | pr_cont("\n"); |
7339 | } | 7336 | } |
7340 | #endif | 7337 | #endif |
7341 | dump_stack(); | 7338 | dump_stack(); |
7342 | } | 7339 | } |
7343 | EXPORT_SYMBOL(___might_sleep); | 7340 | EXPORT_SYMBOL(___might_sleep); |
7344 | #endif | 7341 | #endif |
7345 | 7342 | ||
7346 | #ifdef CONFIG_MAGIC_SYSRQ | 7343 | #ifdef CONFIG_MAGIC_SYSRQ |
7347 | static void normalize_task(struct rq *rq, struct task_struct *p) | 7344 | static void normalize_task(struct rq *rq, struct task_struct *p) |
7348 | { | 7345 | { |
7349 | const struct sched_class *prev_class = p->sched_class; | 7346 | const struct sched_class *prev_class = p->sched_class; |
7350 | struct sched_attr attr = { | 7347 | struct sched_attr attr = { |
7351 | .sched_policy = SCHED_NORMAL, | 7348 | .sched_policy = SCHED_NORMAL, |
7352 | }; | 7349 | }; |
7353 | int old_prio = p->prio; | 7350 | int old_prio = p->prio; |
7354 | int queued; | 7351 | int queued; |
7355 | 7352 | ||
7356 | queued = task_on_rq_queued(p); | 7353 | queued = task_on_rq_queued(p); |
7357 | if (queued) | 7354 | if (queued) |
7358 | dequeue_task(rq, p, 0); | 7355 | dequeue_task(rq, p, 0); |
7359 | __setscheduler(rq, p, &attr); | 7356 | __setscheduler(rq, p, &attr); |
7360 | if (queued) { | 7357 | if (queued) { |
7361 | enqueue_task(rq, p, 0); | 7358 | enqueue_task(rq, p, 0); |
7362 | resched_curr(rq); | 7359 | resched_curr(rq); |
7363 | } | 7360 | } |
7364 | 7361 | ||
7365 | check_class_changed(rq, p, prev_class, old_prio); | 7362 | check_class_changed(rq, p, prev_class, old_prio); |
7366 | } | 7363 | } |
7367 | 7364 | ||
7368 | void normalize_rt_tasks(void) | 7365 | void normalize_rt_tasks(void) |
7369 | { | 7366 | { |
7370 | struct task_struct *g, *p; | 7367 | struct task_struct *g, *p; |
7371 | unsigned long flags; | 7368 | unsigned long flags; |
7372 | struct rq *rq; | 7369 | struct rq *rq; |
7373 | 7370 | ||
7374 | read_lock(&tasklist_lock); | 7371 | read_lock(&tasklist_lock); |
7375 | for_each_process_thread(g, p) { | 7372 | for_each_process_thread(g, p) { |
7376 | /* | 7373 | /* |
7377 | * Only normalize user tasks: | 7374 | * Only normalize user tasks: |
7378 | */ | 7375 | */ |
7379 | if (p->flags & PF_KTHREAD) | 7376 | if (p->flags & PF_KTHREAD) |
7380 | continue; | 7377 | continue; |
7381 | 7378 | ||
7382 | p->se.exec_start = 0; | 7379 | p->se.exec_start = 0; |
7383 | #ifdef CONFIG_SCHEDSTATS | 7380 | #ifdef CONFIG_SCHEDSTATS |
7384 | p->se.statistics.wait_start = 0; | 7381 | p->se.statistics.wait_start = 0; |
7385 | p->se.statistics.sleep_start = 0; | 7382 | p->se.statistics.sleep_start = 0; |
7386 | p->se.statistics.block_start = 0; | 7383 | p->se.statistics.block_start = 0; |
7387 | #endif | 7384 | #endif |
7388 | 7385 | ||
7389 | if (!dl_task(p) && !rt_task(p)) { | 7386 | if (!dl_task(p) && !rt_task(p)) { |
7390 | /* | 7387 | /* |
7391 | * Renice negative nice level userspace | 7388 | * Renice negative nice level userspace |
7392 | * tasks back to 0: | 7389 | * tasks back to 0: |
7393 | */ | 7390 | */ |
7394 | if (task_nice(p) < 0) | 7391 | if (task_nice(p) < 0) |
7395 | set_user_nice(p, 0); | 7392 | set_user_nice(p, 0); |
7396 | continue; | 7393 | continue; |
7397 | } | 7394 | } |
7398 | 7395 | ||
7399 | rq = task_rq_lock(p, &flags); | 7396 | rq = task_rq_lock(p, &flags); |
7400 | normalize_task(rq, p); | 7397 | normalize_task(rq, p); |
7401 | task_rq_unlock(rq, p, &flags); | 7398 | task_rq_unlock(rq, p, &flags); |
7402 | } | 7399 | } |
7403 | read_unlock(&tasklist_lock); | 7400 | read_unlock(&tasklist_lock); |
7404 | } | 7401 | } |
7405 | 7402 | ||
7406 | #endif /* CONFIG_MAGIC_SYSRQ */ | 7403 | #endif /* CONFIG_MAGIC_SYSRQ */ |
7407 | 7404 | ||
7408 | #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) | 7405 | #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) |
7409 | /* | 7406 | /* |
7410 | * These functions are only useful for the IA64 MCA handling, or kdb. | 7407 | * These functions are only useful for the IA64 MCA handling, or kdb. |
7411 | * | 7408 | * |
7412 | * They can only be called when the whole system has been | 7409 | * They can only be called when the whole system has been |
7413 | * stopped - every CPU needs to be quiescent, and no scheduling | 7410 | * stopped - every CPU needs to be quiescent, and no scheduling |
7414 | * activity can take place. Using them for anything else would | 7411 | * activity can take place. Using them for anything else would |
7415 | * be a serious bug, and as a result, they aren't even visible | 7412 | * be a serious bug, and as a result, they aren't even visible |
7416 | * under any other configuration. | 7413 | * under any other configuration. |
7417 | */ | 7414 | */ |
7418 | 7415 | ||
7419 | /** | 7416 | /** |
7420 | * curr_task - return the current task for a given cpu. | 7417 | * curr_task - return the current task for a given cpu. |
7421 | * @cpu: the processor in question. | 7418 | * @cpu: the processor in question. |
7422 | * | 7419 | * |
7423 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | 7420 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! |
7424 | * | 7421 | * |
7425 | * Return: The current task for @cpu. | 7422 | * Return: The current task for @cpu. |
7426 | */ | 7423 | */ |
7427 | struct task_struct *curr_task(int cpu) | 7424 | struct task_struct *curr_task(int cpu) |
7428 | { | 7425 | { |
7429 | return cpu_curr(cpu); | 7426 | return cpu_curr(cpu); |
7430 | } | 7427 | } |
7431 | 7428 | ||
7432 | #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ | 7429 | #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ |
7433 | 7430 | ||
7434 | #ifdef CONFIG_IA64 | 7431 | #ifdef CONFIG_IA64 |
7435 | /** | 7432 | /** |
7436 | * set_curr_task - set the current task for a given cpu. | 7433 | * set_curr_task - set the current task for a given cpu. |
7437 | * @cpu: the processor in question. | 7434 | * @cpu: the processor in question. |
7438 | * @p: the task pointer to set. | 7435 | * @p: the task pointer to set. |
7439 | * | 7436 | * |
7440 | * Description: This function must only be used when non-maskable interrupts | 7437 | * Description: This function must only be used when non-maskable interrupts |
7441 | * are serviced on a separate stack. It allows the architecture to switch the | 7438 | * are serviced on a separate stack. It allows the architecture to switch the |
7442 | * notion of the current task on a cpu in a non-blocking manner. This function | 7439 | * notion of the current task on a cpu in a non-blocking manner. This function |
7443 | * must be called with all CPU's synchronized, and interrupts disabled, the | 7440 | * must be called with all CPU's synchronized, and interrupts disabled, the |
7444 | * and caller must save the original value of the current task (see | 7441 | * and caller must save the original value of the current task (see |
7445 | * curr_task() above) and restore that value before reenabling interrupts and | 7442 | * curr_task() above) and restore that value before reenabling interrupts and |
7446 | * re-starting the system. | 7443 | * re-starting the system. |
7447 | * | 7444 | * |
7448 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | 7445 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! |
7449 | */ | 7446 | */ |
7450 | void set_curr_task(int cpu, struct task_struct *p) | 7447 | void set_curr_task(int cpu, struct task_struct *p) |
7451 | { | 7448 | { |
7452 | cpu_curr(cpu) = p; | 7449 | cpu_curr(cpu) = p; |
7453 | } | 7450 | } |
7454 | 7451 | ||
7455 | #endif | 7452 | #endif |
7456 | 7453 | ||
7457 | #ifdef CONFIG_CGROUP_SCHED | 7454 | #ifdef CONFIG_CGROUP_SCHED |
7458 | /* task_group_lock serializes the addition/removal of task groups */ | 7455 | /* task_group_lock serializes the addition/removal of task groups */ |
7459 | static DEFINE_SPINLOCK(task_group_lock); | 7456 | static DEFINE_SPINLOCK(task_group_lock); |
7460 | 7457 | ||
7461 | static void free_sched_group(struct task_group *tg) | 7458 | static void free_sched_group(struct task_group *tg) |
7462 | { | 7459 | { |
7463 | free_fair_sched_group(tg); | 7460 | free_fair_sched_group(tg); |
7464 | free_rt_sched_group(tg); | 7461 | free_rt_sched_group(tg); |
7465 | autogroup_free(tg); | 7462 | autogroup_free(tg); |
7466 | kfree(tg); | 7463 | kfree(tg); |
7467 | } | 7464 | } |
7468 | 7465 | ||
7469 | /* allocate runqueue etc for a new task group */ | 7466 | /* allocate runqueue etc for a new task group */ |
7470 | struct task_group *sched_create_group(struct task_group *parent) | 7467 | struct task_group *sched_create_group(struct task_group *parent) |
7471 | { | 7468 | { |
7472 | struct task_group *tg; | 7469 | struct task_group *tg; |
7473 | 7470 | ||
7474 | tg = kzalloc(sizeof(*tg), GFP_KERNEL); | 7471 | tg = kzalloc(sizeof(*tg), GFP_KERNEL); |
7475 | if (!tg) | 7472 | if (!tg) |
7476 | return ERR_PTR(-ENOMEM); | 7473 | return ERR_PTR(-ENOMEM); |
7477 | 7474 | ||
7478 | if (!alloc_fair_sched_group(tg, parent)) | 7475 | if (!alloc_fair_sched_group(tg, parent)) |
7479 | goto err; | 7476 | goto err; |
7480 | 7477 | ||
7481 | if (!alloc_rt_sched_group(tg, parent)) | 7478 | if (!alloc_rt_sched_group(tg, parent)) |
7482 | goto err; | 7479 | goto err; |
7483 | 7480 | ||
7484 | return tg; | 7481 | return tg; |
7485 | 7482 | ||
7486 | err: | 7483 | err: |
7487 | free_sched_group(tg); | 7484 | free_sched_group(tg); |
7488 | return ERR_PTR(-ENOMEM); | 7485 | return ERR_PTR(-ENOMEM); |
7489 | } | 7486 | } |
7490 | 7487 | ||
7491 | void sched_online_group(struct task_group *tg, struct task_group *parent) | 7488 | void sched_online_group(struct task_group *tg, struct task_group *parent) |
7492 | { | 7489 | { |
7493 | unsigned long flags; | 7490 | unsigned long flags; |
7494 | 7491 | ||
7495 | spin_lock_irqsave(&task_group_lock, flags); | 7492 | spin_lock_irqsave(&task_group_lock, flags); |
7496 | list_add_rcu(&tg->list, &task_groups); | 7493 | list_add_rcu(&tg->list, &task_groups); |
7497 | 7494 | ||
7498 | WARN_ON(!parent); /* root should already exist */ | 7495 | WARN_ON(!parent); /* root should already exist */ |
7499 | 7496 | ||
7500 | tg->parent = parent; | 7497 | tg->parent = parent; |
7501 | INIT_LIST_HEAD(&tg->children); | 7498 | INIT_LIST_HEAD(&tg->children); |
7502 | list_add_rcu(&tg->siblings, &parent->children); | 7499 | list_add_rcu(&tg->siblings, &parent->children); |
7503 | spin_unlock_irqrestore(&task_group_lock, flags); | 7500 | spin_unlock_irqrestore(&task_group_lock, flags); |
7504 | } | 7501 | } |
7505 | 7502 | ||
7506 | /* rcu callback to free various structures associated with a task group */ | 7503 | /* rcu callback to free various structures associated with a task group */ |
7507 | static void free_sched_group_rcu(struct rcu_head *rhp) | 7504 | static void free_sched_group_rcu(struct rcu_head *rhp) |
7508 | { | 7505 | { |
7509 | /* now it should be safe to free those cfs_rqs */ | 7506 | /* now it should be safe to free those cfs_rqs */ |
7510 | free_sched_group(container_of(rhp, struct task_group, rcu)); | 7507 | free_sched_group(container_of(rhp, struct task_group, rcu)); |
7511 | } | 7508 | } |
7512 | 7509 | ||
7513 | /* Destroy runqueue etc associated with a task group */ | 7510 | /* Destroy runqueue etc associated with a task group */ |
7514 | void sched_destroy_group(struct task_group *tg) | 7511 | void sched_destroy_group(struct task_group *tg) |
7515 | { | 7512 | { |
7516 | /* wait for possible concurrent references to cfs_rqs complete */ | 7513 | /* wait for possible concurrent references to cfs_rqs complete */ |
7517 | call_rcu(&tg->rcu, free_sched_group_rcu); | 7514 | call_rcu(&tg->rcu, free_sched_group_rcu); |
7518 | } | 7515 | } |
7519 | 7516 | ||
7520 | void sched_offline_group(struct task_group *tg) | 7517 | void sched_offline_group(struct task_group *tg) |
7521 | { | 7518 | { |
7522 | unsigned long flags; | 7519 | unsigned long flags; |
7523 | int i; | 7520 | int i; |
7524 | 7521 | ||
7525 | /* end participation in shares distribution */ | 7522 | /* end participation in shares distribution */ |
7526 | for_each_possible_cpu(i) | 7523 | for_each_possible_cpu(i) |
7527 | unregister_fair_sched_group(tg, i); | 7524 | unregister_fair_sched_group(tg, i); |
7528 | 7525 | ||
7529 | spin_lock_irqsave(&task_group_lock, flags); | 7526 | spin_lock_irqsave(&task_group_lock, flags); |
7530 | list_del_rcu(&tg->list); | 7527 | list_del_rcu(&tg->list); |
7531 | list_del_rcu(&tg->siblings); | 7528 | list_del_rcu(&tg->siblings); |
7532 | spin_unlock_irqrestore(&task_group_lock, flags); | 7529 | spin_unlock_irqrestore(&task_group_lock, flags); |
7533 | } | 7530 | } |
7534 | 7531 | ||
7535 | /* change task's runqueue when it moves between groups. | 7532 | /* change task's runqueue when it moves between groups. |
7536 | * The caller of this function should have put the task in its new group | 7533 | * The caller of this function should have put the task in its new group |
7537 | * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to | 7534 | * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to |
7538 | * reflect its new group. | 7535 | * reflect its new group. |
7539 | */ | 7536 | */ |
7540 | void sched_move_task(struct task_struct *tsk) | 7537 | void sched_move_task(struct task_struct *tsk) |
7541 | { | 7538 | { |
7542 | struct task_group *tg; | 7539 | struct task_group *tg; |
7543 | int queued, running; | 7540 | int queued, running; |
7544 | unsigned long flags; | 7541 | unsigned long flags; |
7545 | struct rq *rq; | 7542 | struct rq *rq; |
7546 | 7543 | ||
7547 | rq = task_rq_lock(tsk, &flags); | 7544 | rq = task_rq_lock(tsk, &flags); |
7548 | 7545 | ||
7549 | running = task_current(rq, tsk); | 7546 | running = task_current(rq, tsk); |
7550 | queued = task_on_rq_queued(tsk); | 7547 | queued = task_on_rq_queued(tsk); |
7551 | 7548 | ||
7552 | if (queued) | 7549 | if (queued) |
7553 | dequeue_task(rq, tsk, 0); | 7550 | dequeue_task(rq, tsk, 0); |
7554 | if (unlikely(running)) | 7551 | if (unlikely(running)) |
7555 | put_prev_task(rq, tsk); | 7552 | put_prev_task(rq, tsk); |
7556 | 7553 | ||
7557 | /* | 7554 | /* |
7558 | * All callers are synchronized by task_rq_lock(); we do not use RCU | 7555 | * All callers are synchronized by task_rq_lock(); we do not use RCU |
7559 | * which is pointless here. Thus, we pass "true" to task_css_check() | 7556 | * which is pointless here. Thus, we pass "true" to task_css_check() |
7560 | * to prevent lockdep warnings. | 7557 | * to prevent lockdep warnings. |
7561 | */ | 7558 | */ |
7562 | tg = container_of(task_css_check(tsk, cpu_cgrp_id, true), | 7559 | tg = container_of(task_css_check(tsk, cpu_cgrp_id, true), |
7563 | struct task_group, css); | 7560 | struct task_group, css); |
7564 | tg = autogroup_task_group(tsk, tg); | 7561 | tg = autogroup_task_group(tsk, tg); |
7565 | tsk->sched_task_group = tg; | 7562 | tsk->sched_task_group = tg; |
7566 | 7563 | ||
7567 | #ifdef CONFIG_FAIR_GROUP_SCHED | 7564 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7568 | if (tsk->sched_class->task_move_group) | 7565 | if (tsk->sched_class->task_move_group) |
7569 | tsk->sched_class->task_move_group(tsk, queued); | 7566 | tsk->sched_class->task_move_group(tsk, queued); |
7570 | else | 7567 | else |
7571 | #endif | 7568 | #endif |
7572 | set_task_rq(tsk, task_cpu(tsk)); | 7569 | set_task_rq(tsk, task_cpu(tsk)); |
7573 | 7570 | ||
7574 | if (unlikely(running)) | 7571 | if (unlikely(running)) |
7575 | tsk->sched_class->set_curr_task(rq); | 7572 | tsk->sched_class->set_curr_task(rq); |
7576 | if (queued) | 7573 | if (queued) |
7577 | enqueue_task(rq, tsk, 0); | 7574 | enqueue_task(rq, tsk, 0); |
7578 | 7575 | ||
7579 | task_rq_unlock(rq, tsk, &flags); | 7576 | task_rq_unlock(rq, tsk, &flags); |
7580 | } | 7577 | } |
7581 | #endif /* CONFIG_CGROUP_SCHED */ | 7578 | #endif /* CONFIG_CGROUP_SCHED */ |
7582 | 7579 | ||
7583 | #ifdef CONFIG_RT_GROUP_SCHED | 7580 | #ifdef CONFIG_RT_GROUP_SCHED |
7584 | /* | 7581 | /* |
7585 | * Ensure that the real time constraints are schedulable. | 7582 | * Ensure that the real time constraints are schedulable. |
7586 | */ | 7583 | */ |
7587 | static DEFINE_MUTEX(rt_constraints_mutex); | 7584 | static DEFINE_MUTEX(rt_constraints_mutex); |
7588 | 7585 | ||
7589 | /* Must be called with tasklist_lock held */ | 7586 | /* Must be called with tasklist_lock held */ |
7590 | static inline int tg_has_rt_tasks(struct task_group *tg) | 7587 | static inline int tg_has_rt_tasks(struct task_group *tg) |
7591 | { | 7588 | { |
7592 | struct task_struct *g, *p; | 7589 | struct task_struct *g, *p; |
7593 | 7590 | ||
7594 | for_each_process_thread(g, p) { | 7591 | for_each_process_thread(g, p) { |
7595 | if (rt_task(p) && task_group(p) == tg) | 7592 | if (rt_task(p) && task_group(p) == tg) |
7596 | return 1; | 7593 | return 1; |
7597 | } | 7594 | } |
7598 | 7595 | ||
7599 | return 0; | 7596 | return 0; |
7600 | } | 7597 | } |
7601 | 7598 | ||
7602 | struct rt_schedulable_data { | 7599 | struct rt_schedulable_data { |
7603 | struct task_group *tg; | 7600 | struct task_group *tg; |
7604 | u64 rt_period; | 7601 | u64 rt_period; |
7605 | u64 rt_runtime; | 7602 | u64 rt_runtime; |
7606 | }; | 7603 | }; |
7607 | 7604 | ||
7608 | static int tg_rt_schedulable(struct task_group *tg, void *data) | 7605 | static int tg_rt_schedulable(struct task_group *tg, void *data) |
7609 | { | 7606 | { |
7610 | struct rt_schedulable_data *d = data; | 7607 | struct rt_schedulable_data *d = data; |
7611 | struct task_group *child; | 7608 | struct task_group *child; |
7612 | unsigned long total, sum = 0; | 7609 | unsigned long total, sum = 0; |
7613 | u64 period, runtime; | 7610 | u64 period, runtime; |
7614 | 7611 | ||
7615 | period = ktime_to_ns(tg->rt_bandwidth.rt_period); | 7612 | period = ktime_to_ns(tg->rt_bandwidth.rt_period); |
7616 | runtime = tg->rt_bandwidth.rt_runtime; | 7613 | runtime = tg->rt_bandwidth.rt_runtime; |
7617 | 7614 | ||
7618 | if (tg == d->tg) { | 7615 | if (tg == d->tg) { |
7619 | period = d->rt_period; | 7616 | period = d->rt_period; |
7620 | runtime = d->rt_runtime; | 7617 | runtime = d->rt_runtime; |
7621 | } | 7618 | } |
7622 | 7619 | ||
7623 | /* | 7620 | /* |
7624 | * Cannot have more runtime than the period. | 7621 | * Cannot have more runtime than the period. |
7625 | */ | 7622 | */ |
7626 | if (runtime > period && runtime != RUNTIME_INF) | 7623 | if (runtime > period && runtime != RUNTIME_INF) |
7627 | return -EINVAL; | 7624 | return -EINVAL; |
7628 | 7625 | ||
7629 | /* | 7626 | /* |
7630 | * Ensure we don't starve existing RT tasks. | 7627 | * Ensure we don't starve existing RT tasks. |
7631 | */ | 7628 | */ |
7632 | if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) | 7629 | if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) |
7633 | return -EBUSY; | 7630 | return -EBUSY; |
7634 | 7631 | ||
7635 | total = to_ratio(period, runtime); | 7632 | total = to_ratio(period, runtime); |
7636 | 7633 | ||
7637 | /* | 7634 | /* |
7638 | * Nobody can have more than the global setting allows. | 7635 | * Nobody can have more than the global setting allows. |
7639 | */ | 7636 | */ |
7640 | if (total > to_ratio(global_rt_period(), global_rt_runtime())) | 7637 | if (total > to_ratio(global_rt_period(), global_rt_runtime())) |
7641 | return -EINVAL; | 7638 | return -EINVAL; |
7642 | 7639 | ||
7643 | /* | 7640 | /* |
7644 | * The sum of our children's runtime should not exceed our own. | 7641 | * The sum of our children's runtime should not exceed our own. |
7645 | */ | 7642 | */ |
7646 | list_for_each_entry_rcu(child, &tg->children, siblings) { | 7643 | list_for_each_entry_rcu(child, &tg->children, siblings) { |
7647 | period = ktime_to_ns(child->rt_bandwidth.rt_period); | 7644 | period = ktime_to_ns(child->rt_bandwidth.rt_period); |
7648 | runtime = child->rt_bandwidth.rt_runtime; | 7645 | runtime = child->rt_bandwidth.rt_runtime; |
7649 | 7646 | ||
7650 | if (child == d->tg) { | 7647 | if (child == d->tg) { |
7651 | period = d->rt_period; | 7648 | period = d->rt_period; |
7652 | runtime = d->rt_runtime; | 7649 | runtime = d->rt_runtime; |
7653 | } | 7650 | } |
7654 | 7651 | ||
7655 | sum += to_ratio(period, runtime); | 7652 | sum += to_ratio(period, runtime); |
7656 | } | 7653 | } |
7657 | 7654 | ||
7658 | if (sum > total) | 7655 | if (sum > total) |
7659 | return -EINVAL; | 7656 | return -EINVAL; |
7660 | 7657 | ||
7661 | return 0; | 7658 | return 0; |
7662 | } | 7659 | } |
7663 | 7660 | ||
7664 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) | 7661 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) |
7665 | { | 7662 | { |
7666 | int ret; | 7663 | int ret; |
7667 | 7664 | ||
7668 | struct rt_schedulable_data data = { | 7665 | struct rt_schedulable_data data = { |
7669 | .tg = tg, | 7666 | .tg = tg, |
7670 | .rt_period = period, | 7667 | .rt_period = period, |
7671 | .rt_runtime = runtime, | 7668 | .rt_runtime = runtime, |
7672 | }; | 7669 | }; |
7673 | 7670 | ||
7674 | rcu_read_lock(); | 7671 | rcu_read_lock(); |
7675 | ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); | 7672 | ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); |
7676 | rcu_read_unlock(); | 7673 | rcu_read_unlock(); |
7677 | 7674 | ||
7678 | return ret; | 7675 | return ret; |
7679 | } | 7676 | } |
7680 | 7677 | ||
7681 | static int tg_set_rt_bandwidth(struct task_group *tg, | 7678 | static int tg_set_rt_bandwidth(struct task_group *tg, |
7682 | u64 rt_period, u64 rt_runtime) | 7679 | u64 rt_period, u64 rt_runtime) |
7683 | { | 7680 | { |
7684 | int i, err = 0; | 7681 | int i, err = 0; |
7685 | 7682 | ||
7686 | mutex_lock(&rt_constraints_mutex); | 7683 | mutex_lock(&rt_constraints_mutex); |
7687 | read_lock(&tasklist_lock); | 7684 | read_lock(&tasklist_lock); |
7688 | err = __rt_schedulable(tg, rt_period, rt_runtime); | 7685 | err = __rt_schedulable(tg, rt_period, rt_runtime); |
7689 | if (err) | 7686 | if (err) |
7690 | goto unlock; | 7687 | goto unlock; |
7691 | 7688 | ||
7692 | raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); | 7689 | raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); |
7693 | tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); | 7690 | tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); |
7694 | tg->rt_bandwidth.rt_runtime = rt_runtime; | 7691 | tg->rt_bandwidth.rt_runtime = rt_runtime; |
7695 | 7692 | ||
7696 | for_each_possible_cpu(i) { | 7693 | for_each_possible_cpu(i) { |
7697 | struct rt_rq *rt_rq = tg->rt_rq[i]; | 7694 | struct rt_rq *rt_rq = tg->rt_rq[i]; |
7698 | 7695 | ||
7699 | raw_spin_lock(&rt_rq->rt_runtime_lock); | 7696 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
7700 | rt_rq->rt_runtime = rt_runtime; | 7697 | rt_rq->rt_runtime = rt_runtime; |
7701 | raw_spin_unlock(&rt_rq->rt_runtime_lock); | 7698 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
7702 | } | 7699 | } |
7703 | raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); | 7700 | raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); |
7704 | unlock: | 7701 | unlock: |
7705 | read_unlock(&tasklist_lock); | 7702 | read_unlock(&tasklist_lock); |
7706 | mutex_unlock(&rt_constraints_mutex); | 7703 | mutex_unlock(&rt_constraints_mutex); |
7707 | 7704 | ||
7708 | return err; | 7705 | return err; |
7709 | } | 7706 | } |
7710 | 7707 | ||
7711 | static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) | 7708 | static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) |
7712 | { | 7709 | { |
7713 | u64 rt_runtime, rt_period; | 7710 | u64 rt_runtime, rt_period; |
7714 | 7711 | ||
7715 | rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); | 7712 | rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); |
7716 | rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; | 7713 | rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; |
7717 | if (rt_runtime_us < 0) | 7714 | if (rt_runtime_us < 0) |
7718 | rt_runtime = RUNTIME_INF; | 7715 | rt_runtime = RUNTIME_INF; |
7719 | 7716 | ||
7720 | return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); | 7717 | return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); |
7721 | } | 7718 | } |
7722 | 7719 | ||
7723 | static long sched_group_rt_runtime(struct task_group *tg) | 7720 | static long sched_group_rt_runtime(struct task_group *tg) |
7724 | { | 7721 | { |
7725 | u64 rt_runtime_us; | 7722 | u64 rt_runtime_us; |
7726 | 7723 | ||
7727 | if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) | 7724 | if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) |
7728 | return -1; | 7725 | return -1; |
7729 | 7726 | ||
7730 | rt_runtime_us = tg->rt_bandwidth.rt_runtime; | 7727 | rt_runtime_us = tg->rt_bandwidth.rt_runtime; |
7731 | do_div(rt_runtime_us, NSEC_PER_USEC); | 7728 | do_div(rt_runtime_us, NSEC_PER_USEC); |
7732 | return rt_runtime_us; | 7729 | return rt_runtime_us; |
7733 | } | 7730 | } |
7734 | 7731 | ||
7735 | static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) | 7732 | static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) |
7736 | { | 7733 | { |
7737 | u64 rt_runtime, rt_period; | 7734 | u64 rt_runtime, rt_period; |
7738 | 7735 | ||
7739 | rt_period = (u64)rt_period_us * NSEC_PER_USEC; | 7736 | rt_period = (u64)rt_period_us * NSEC_PER_USEC; |
7740 | rt_runtime = tg->rt_bandwidth.rt_runtime; | 7737 | rt_runtime = tg->rt_bandwidth.rt_runtime; |
7741 | 7738 | ||
7742 | if (rt_period == 0) | 7739 | if (rt_period == 0) |
7743 | return -EINVAL; | 7740 | return -EINVAL; |
7744 | 7741 | ||
7745 | return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); | 7742 | return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); |
7746 | } | 7743 | } |
7747 | 7744 | ||
7748 | static long sched_group_rt_period(struct task_group *tg) | 7745 | static long sched_group_rt_period(struct task_group *tg) |
7749 | { | 7746 | { |
7750 | u64 rt_period_us; | 7747 | u64 rt_period_us; |
7751 | 7748 | ||
7752 | rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); | 7749 | rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); |
7753 | do_div(rt_period_us, NSEC_PER_USEC); | 7750 | do_div(rt_period_us, NSEC_PER_USEC); |
7754 | return rt_period_us; | 7751 | return rt_period_us; |
7755 | } | 7752 | } |
7756 | #endif /* CONFIG_RT_GROUP_SCHED */ | 7753 | #endif /* CONFIG_RT_GROUP_SCHED */ |
7757 | 7754 | ||
7758 | #ifdef CONFIG_RT_GROUP_SCHED | 7755 | #ifdef CONFIG_RT_GROUP_SCHED |
7759 | static int sched_rt_global_constraints(void) | 7756 | static int sched_rt_global_constraints(void) |
7760 | { | 7757 | { |
7761 | int ret = 0; | 7758 | int ret = 0; |
7762 | 7759 | ||
7763 | mutex_lock(&rt_constraints_mutex); | 7760 | mutex_lock(&rt_constraints_mutex); |
7764 | read_lock(&tasklist_lock); | 7761 | read_lock(&tasklist_lock); |
7765 | ret = __rt_schedulable(NULL, 0, 0); | 7762 | ret = __rt_schedulable(NULL, 0, 0); |
7766 | read_unlock(&tasklist_lock); | 7763 | read_unlock(&tasklist_lock); |
7767 | mutex_unlock(&rt_constraints_mutex); | 7764 | mutex_unlock(&rt_constraints_mutex); |
7768 | 7765 | ||
7769 | return ret; | 7766 | return ret; |
7770 | } | 7767 | } |
7771 | 7768 | ||
7772 | static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) | 7769 | static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) |
7773 | { | 7770 | { |
7774 | /* Don't accept realtime tasks when there is no way for them to run */ | 7771 | /* Don't accept realtime tasks when there is no way for them to run */ |
7775 | if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) | 7772 | if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) |
7776 | return 0; | 7773 | return 0; |
7777 | 7774 | ||
7778 | return 1; | 7775 | return 1; |
7779 | } | 7776 | } |
7780 | 7777 | ||
7781 | #else /* !CONFIG_RT_GROUP_SCHED */ | 7778 | #else /* !CONFIG_RT_GROUP_SCHED */ |
7782 | static int sched_rt_global_constraints(void) | 7779 | static int sched_rt_global_constraints(void) |
7783 | { | 7780 | { |
7784 | unsigned long flags; | 7781 | unsigned long flags; |
7785 | int i, ret = 0; | 7782 | int i, ret = 0; |
7786 | 7783 | ||
7787 | raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); | 7784 | raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); |
7788 | for_each_possible_cpu(i) { | 7785 | for_each_possible_cpu(i) { |
7789 | struct rt_rq *rt_rq = &cpu_rq(i)->rt; | 7786 | struct rt_rq *rt_rq = &cpu_rq(i)->rt; |
7790 | 7787 | ||
7791 | raw_spin_lock(&rt_rq->rt_runtime_lock); | 7788 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
7792 | rt_rq->rt_runtime = global_rt_runtime(); | 7789 | rt_rq->rt_runtime = global_rt_runtime(); |
7793 | raw_spin_unlock(&rt_rq->rt_runtime_lock); | 7790 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
7794 | } | 7791 | } |
7795 | raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); | 7792 | raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); |
7796 | 7793 | ||
7797 | return ret; | 7794 | return ret; |
7798 | } | 7795 | } |
7799 | #endif /* CONFIG_RT_GROUP_SCHED */ | 7796 | #endif /* CONFIG_RT_GROUP_SCHED */ |
7800 | 7797 | ||
7801 | static int sched_dl_global_constraints(void) | 7798 | static int sched_dl_global_constraints(void) |
7802 | { | 7799 | { |
7803 | u64 runtime = global_rt_runtime(); | 7800 | u64 runtime = global_rt_runtime(); |
7804 | u64 period = global_rt_period(); | 7801 | u64 period = global_rt_period(); |
7805 | u64 new_bw = to_ratio(period, runtime); | 7802 | u64 new_bw = to_ratio(period, runtime); |
7806 | struct dl_bw *dl_b; | 7803 | struct dl_bw *dl_b; |
7807 | int cpu, ret = 0; | 7804 | int cpu, ret = 0; |
7808 | unsigned long flags; | 7805 | unsigned long flags; |
7809 | 7806 | ||
7810 | /* | 7807 | /* |
7811 | * Here we want to check the bandwidth not being set to some | 7808 | * Here we want to check the bandwidth not being set to some |
7812 | * value smaller than the currently allocated bandwidth in | 7809 | * value smaller than the currently allocated bandwidth in |
7813 | * any of the root_domains. | 7810 | * any of the root_domains. |
7814 | * | 7811 | * |
7815 | * FIXME: Cycling on all the CPUs is overdoing, but simpler than | 7812 | * FIXME: Cycling on all the CPUs is overdoing, but simpler than |
7816 | * cycling on root_domains... Discussion on different/better | 7813 | * cycling on root_domains... Discussion on different/better |
7817 | * solutions is welcome! | 7814 | * solutions is welcome! |
7818 | */ | 7815 | */ |
7819 | for_each_possible_cpu(cpu) { | 7816 | for_each_possible_cpu(cpu) { |
7820 | rcu_read_lock_sched(); | 7817 | rcu_read_lock_sched(); |
7821 | dl_b = dl_bw_of(cpu); | 7818 | dl_b = dl_bw_of(cpu); |
7822 | 7819 | ||
7823 | raw_spin_lock_irqsave(&dl_b->lock, flags); | 7820 | raw_spin_lock_irqsave(&dl_b->lock, flags); |
7824 | if (new_bw < dl_b->total_bw) | 7821 | if (new_bw < dl_b->total_bw) |
7825 | ret = -EBUSY; | 7822 | ret = -EBUSY; |
7826 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); | 7823 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); |
7827 | 7824 | ||
7828 | rcu_read_unlock_sched(); | 7825 | rcu_read_unlock_sched(); |
7829 | 7826 | ||
7830 | if (ret) | 7827 | if (ret) |
7831 | break; | 7828 | break; |
7832 | } | 7829 | } |
7833 | 7830 | ||
7834 | return ret; | 7831 | return ret; |
7835 | } | 7832 | } |
7836 | 7833 | ||
7837 | static void sched_dl_do_global(void) | 7834 | static void sched_dl_do_global(void) |
7838 | { | 7835 | { |
7839 | u64 new_bw = -1; | 7836 | u64 new_bw = -1; |
7840 | struct dl_bw *dl_b; | 7837 | struct dl_bw *dl_b; |
7841 | int cpu; | 7838 | int cpu; |
7842 | unsigned long flags; | 7839 | unsigned long flags; |
7843 | 7840 | ||
7844 | def_dl_bandwidth.dl_period = global_rt_period(); | 7841 | def_dl_bandwidth.dl_period = global_rt_period(); |
7845 | def_dl_bandwidth.dl_runtime = global_rt_runtime(); | 7842 | def_dl_bandwidth.dl_runtime = global_rt_runtime(); |
7846 | 7843 | ||
7847 | if (global_rt_runtime() != RUNTIME_INF) | 7844 | if (global_rt_runtime() != RUNTIME_INF) |
7848 | new_bw = to_ratio(global_rt_period(), global_rt_runtime()); | 7845 | new_bw = to_ratio(global_rt_period(), global_rt_runtime()); |
7849 | 7846 | ||
7850 | /* | 7847 | /* |
7851 | * FIXME: As above... | 7848 | * FIXME: As above... |
7852 | */ | 7849 | */ |
7853 | for_each_possible_cpu(cpu) { | 7850 | for_each_possible_cpu(cpu) { |
7854 | rcu_read_lock_sched(); | 7851 | rcu_read_lock_sched(); |
7855 | dl_b = dl_bw_of(cpu); | 7852 | dl_b = dl_bw_of(cpu); |
7856 | 7853 | ||
7857 | raw_spin_lock_irqsave(&dl_b->lock, flags); | 7854 | raw_spin_lock_irqsave(&dl_b->lock, flags); |
7858 | dl_b->bw = new_bw; | 7855 | dl_b->bw = new_bw; |
7859 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); | 7856 | raw_spin_unlock_irqrestore(&dl_b->lock, flags); |
7860 | 7857 | ||
7861 | rcu_read_unlock_sched(); | 7858 | rcu_read_unlock_sched(); |
7862 | } | 7859 | } |
7863 | } | 7860 | } |
7864 | 7861 | ||
7865 | static int sched_rt_global_validate(void) | 7862 | static int sched_rt_global_validate(void) |
7866 | { | 7863 | { |
7867 | if (sysctl_sched_rt_period <= 0) | 7864 | if (sysctl_sched_rt_period <= 0) |
7868 | return -EINVAL; | 7865 | return -EINVAL; |
7869 | 7866 | ||
7870 | if ((sysctl_sched_rt_runtime != RUNTIME_INF) && | 7867 | if ((sysctl_sched_rt_runtime != RUNTIME_INF) && |
7871 | (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) | 7868 | (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) |
7872 | return -EINVAL; | 7869 | return -EINVAL; |
7873 | 7870 | ||
7874 | return 0; | 7871 | return 0; |
7875 | } | 7872 | } |
7876 | 7873 | ||
7877 | static void sched_rt_do_global(void) | 7874 | static void sched_rt_do_global(void) |
7878 | { | 7875 | { |
7879 | def_rt_bandwidth.rt_runtime = global_rt_runtime(); | 7876 | def_rt_bandwidth.rt_runtime = global_rt_runtime(); |
7880 | def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); | 7877 | def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); |
7881 | } | 7878 | } |
7882 | 7879 | ||
7883 | int sched_rt_handler(struct ctl_table *table, int write, | 7880 | int sched_rt_handler(struct ctl_table *table, int write, |
7884 | void __user *buffer, size_t *lenp, | 7881 | void __user *buffer, size_t *lenp, |
7885 | loff_t *ppos) | 7882 | loff_t *ppos) |
7886 | { | 7883 | { |
7887 | int old_period, old_runtime; | 7884 | int old_period, old_runtime; |
7888 | static DEFINE_MUTEX(mutex); | 7885 | static DEFINE_MUTEX(mutex); |
7889 | int ret; | 7886 | int ret; |
7890 | 7887 | ||
7891 | mutex_lock(&mutex); | 7888 | mutex_lock(&mutex); |
7892 | old_period = sysctl_sched_rt_period; | 7889 | old_period = sysctl_sched_rt_period; |
7893 | old_runtime = sysctl_sched_rt_runtime; | 7890 | old_runtime = sysctl_sched_rt_runtime; |
7894 | 7891 | ||
7895 | ret = proc_dointvec(table, write, buffer, lenp, ppos); | 7892 | ret = proc_dointvec(table, write, buffer, lenp, ppos); |
7896 | 7893 | ||
7897 | if (!ret && write) { | 7894 | if (!ret && write) { |
7898 | ret = sched_rt_global_validate(); | 7895 | ret = sched_rt_global_validate(); |
7899 | if (ret) | 7896 | if (ret) |
7900 | goto undo; | 7897 | goto undo; |
7901 | 7898 | ||
7902 | ret = sched_rt_global_constraints(); | 7899 | ret = sched_rt_global_constraints(); |
7903 | if (ret) | 7900 | if (ret) |
7904 | goto undo; | 7901 | goto undo; |
7905 | 7902 | ||
7906 | ret = sched_dl_global_constraints(); | 7903 | ret = sched_dl_global_constraints(); |
7907 | if (ret) | 7904 | if (ret) |
7908 | goto undo; | 7905 | goto undo; |
7909 | 7906 | ||
7910 | sched_rt_do_global(); | 7907 | sched_rt_do_global(); |
7911 | sched_dl_do_global(); | 7908 | sched_dl_do_global(); |
7912 | } | 7909 | } |
7913 | if (0) { | 7910 | if (0) { |
7914 | undo: | 7911 | undo: |
7915 | sysctl_sched_rt_period = old_period; | 7912 | sysctl_sched_rt_period = old_period; |
7916 | sysctl_sched_rt_runtime = old_runtime; | 7913 | sysctl_sched_rt_runtime = old_runtime; |
7917 | } | 7914 | } |
7918 | mutex_unlock(&mutex); | 7915 | mutex_unlock(&mutex); |
7919 | 7916 | ||
7920 | return ret; | 7917 | return ret; |
7921 | } | 7918 | } |
7922 | 7919 | ||
7923 | int sched_rr_handler(struct ctl_table *table, int write, | 7920 | int sched_rr_handler(struct ctl_table *table, int write, |
7924 | void __user *buffer, size_t *lenp, | 7921 | void __user *buffer, size_t *lenp, |
7925 | loff_t *ppos) | 7922 | loff_t *ppos) |
7926 | { | 7923 | { |
7927 | int ret; | 7924 | int ret; |
7928 | static DEFINE_MUTEX(mutex); | 7925 | static DEFINE_MUTEX(mutex); |
7929 | 7926 | ||
7930 | mutex_lock(&mutex); | 7927 | mutex_lock(&mutex); |
7931 | ret = proc_dointvec(table, write, buffer, lenp, ppos); | 7928 | ret = proc_dointvec(table, write, buffer, lenp, ppos); |
7932 | /* make sure that internally we keep jiffies */ | 7929 | /* make sure that internally we keep jiffies */ |
7933 | /* also, writing zero resets timeslice to default */ | 7930 | /* also, writing zero resets timeslice to default */ |
7934 | if (!ret && write) { | 7931 | if (!ret && write) { |
7935 | sched_rr_timeslice = sched_rr_timeslice <= 0 ? | 7932 | sched_rr_timeslice = sched_rr_timeslice <= 0 ? |
7936 | RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); | 7933 | RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); |
7937 | } | 7934 | } |
7938 | mutex_unlock(&mutex); | 7935 | mutex_unlock(&mutex); |
7939 | return ret; | 7936 | return ret; |
7940 | } | 7937 | } |
7941 | 7938 | ||
7942 | #ifdef CONFIG_CGROUP_SCHED | 7939 | #ifdef CONFIG_CGROUP_SCHED |
7943 | 7940 | ||
7944 | static inline struct task_group *css_tg(struct cgroup_subsys_state *css) | 7941 | static inline struct task_group *css_tg(struct cgroup_subsys_state *css) |
7945 | { | 7942 | { |
7946 | return css ? container_of(css, struct task_group, css) : NULL; | 7943 | return css ? container_of(css, struct task_group, css) : NULL; |
7947 | } | 7944 | } |
7948 | 7945 | ||
7949 | static struct cgroup_subsys_state * | 7946 | static struct cgroup_subsys_state * |
7950 | cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) | 7947 | cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) |
7951 | { | 7948 | { |
7952 | struct task_group *parent = css_tg(parent_css); | 7949 | struct task_group *parent = css_tg(parent_css); |
7953 | struct task_group *tg; | 7950 | struct task_group *tg; |
7954 | 7951 | ||
7955 | if (!parent) { | 7952 | if (!parent) { |
7956 | /* This is early initialization for the top cgroup */ | 7953 | /* This is early initialization for the top cgroup */ |
7957 | return &root_task_group.css; | 7954 | return &root_task_group.css; |
7958 | } | 7955 | } |
7959 | 7956 | ||
7960 | tg = sched_create_group(parent); | 7957 | tg = sched_create_group(parent); |
7961 | if (IS_ERR(tg)) | 7958 | if (IS_ERR(tg)) |
7962 | return ERR_PTR(-ENOMEM); | 7959 | return ERR_PTR(-ENOMEM); |
7963 | 7960 | ||
7964 | return &tg->css; | 7961 | return &tg->css; |
7965 | } | 7962 | } |
7966 | 7963 | ||
7967 | static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) | 7964 | static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) |
7968 | { | 7965 | { |
7969 | struct task_group *tg = css_tg(css); | 7966 | struct task_group *tg = css_tg(css); |
7970 | struct task_group *parent = css_tg(css->parent); | 7967 | struct task_group *parent = css_tg(css->parent); |
7971 | 7968 | ||
7972 | if (parent) | 7969 | if (parent) |
7973 | sched_online_group(tg, parent); | 7970 | sched_online_group(tg, parent); |
7974 | return 0; | 7971 | return 0; |
7975 | } | 7972 | } |
7976 | 7973 | ||
7977 | static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) | 7974 | static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) |
7978 | { | 7975 | { |
7979 | struct task_group *tg = css_tg(css); | 7976 | struct task_group *tg = css_tg(css); |
7980 | 7977 | ||
7981 | sched_destroy_group(tg); | 7978 | sched_destroy_group(tg); |
7982 | } | 7979 | } |
7983 | 7980 | ||
7984 | static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) | 7981 | static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) |
7985 | { | 7982 | { |
7986 | struct task_group *tg = css_tg(css); | 7983 | struct task_group *tg = css_tg(css); |
7987 | 7984 | ||
7988 | sched_offline_group(tg); | 7985 | sched_offline_group(tg); |
7989 | } | 7986 | } |
7990 | 7987 | ||
7991 | static void cpu_cgroup_fork(struct task_struct *task) | 7988 | static void cpu_cgroup_fork(struct task_struct *task) |
7992 | { | 7989 | { |
7993 | sched_move_task(task); | 7990 | sched_move_task(task); |
7994 | } | 7991 | } |
7995 | 7992 | ||
7996 | static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, | 7993 | static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, |
7997 | struct cgroup_taskset *tset) | 7994 | struct cgroup_taskset *tset) |
7998 | { | 7995 | { |
7999 | struct task_struct *task; | 7996 | struct task_struct *task; |
8000 | 7997 | ||
8001 | cgroup_taskset_for_each(task, tset) { | 7998 | cgroup_taskset_for_each(task, tset) { |
8002 | #ifdef CONFIG_RT_GROUP_SCHED | 7999 | #ifdef CONFIG_RT_GROUP_SCHED |
8003 | if (!sched_rt_can_attach(css_tg(css), task)) | 8000 | if (!sched_rt_can_attach(css_tg(css), task)) |
8004 | return -EINVAL; | 8001 | return -EINVAL; |
8005 | #else | 8002 | #else |
8006 | /* We don't support RT-tasks being in separate groups */ | 8003 | /* We don't support RT-tasks being in separate groups */ |
8007 | if (task->sched_class != &fair_sched_class) | 8004 | if (task->sched_class != &fair_sched_class) |
8008 | return -EINVAL; | 8005 | return -EINVAL; |
8009 | #endif | 8006 | #endif |
8010 | } | 8007 | } |
8011 | return 0; | 8008 | return 0; |
8012 | } | 8009 | } |
8013 | 8010 | ||
8014 | static void cpu_cgroup_attach(struct cgroup_subsys_state *css, | 8011 | static void cpu_cgroup_attach(struct cgroup_subsys_state *css, |
8015 | struct cgroup_taskset *tset) | 8012 | struct cgroup_taskset *tset) |
8016 | { | 8013 | { |
8017 | struct task_struct *task; | 8014 | struct task_struct *task; |
8018 | 8015 | ||
8019 | cgroup_taskset_for_each(task, tset) | 8016 | cgroup_taskset_for_each(task, tset) |
8020 | sched_move_task(task); | 8017 | sched_move_task(task); |
8021 | } | 8018 | } |
8022 | 8019 | ||
8023 | static void cpu_cgroup_exit(struct cgroup_subsys_state *css, | 8020 | static void cpu_cgroup_exit(struct cgroup_subsys_state *css, |
8024 | struct cgroup_subsys_state *old_css, | 8021 | struct cgroup_subsys_state *old_css, |
8025 | struct task_struct *task) | 8022 | struct task_struct *task) |
8026 | { | 8023 | { |
8027 | /* | 8024 | /* |
8028 | * cgroup_exit() is called in the copy_process() failure path. | 8025 | * cgroup_exit() is called in the copy_process() failure path. |
8029 | * Ignore this case since the task hasn't ran yet, this avoids | 8026 | * Ignore this case since the task hasn't ran yet, this avoids |
8030 | * trying to poke a half freed task state from generic code. | 8027 | * trying to poke a half freed task state from generic code. |
8031 | */ | 8028 | */ |
8032 | if (!(task->flags & PF_EXITING)) | 8029 | if (!(task->flags & PF_EXITING)) |
8033 | return; | 8030 | return; |
8034 | 8031 | ||
8035 | sched_move_task(task); | 8032 | sched_move_task(task); |
8036 | } | 8033 | } |
8037 | 8034 | ||
8038 | #ifdef CONFIG_FAIR_GROUP_SCHED | 8035 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8039 | static int cpu_shares_write_u64(struct cgroup_subsys_state *css, | 8036 | static int cpu_shares_write_u64(struct cgroup_subsys_state *css, |
8040 | struct cftype *cftype, u64 shareval) | 8037 | struct cftype *cftype, u64 shareval) |
8041 | { | 8038 | { |
8042 | return sched_group_set_shares(css_tg(css), scale_load(shareval)); | 8039 | return sched_group_set_shares(css_tg(css), scale_load(shareval)); |
8043 | } | 8040 | } |
8044 | 8041 | ||
8045 | static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, | 8042 | static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, |
8046 | struct cftype *cft) | 8043 | struct cftype *cft) |
8047 | { | 8044 | { |
8048 | struct task_group *tg = css_tg(css); | 8045 | struct task_group *tg = css_tg(css); |
8049 | 8046 | ||
8050 | return (u64) scale_load_down(tg->shares); | 8047 | return (u64) scale_load_down(tg->shares); |
8051 | } | 8048 | } |
8052 | 8049 | ||
8053 | #ifdef CONFIG_CFS_BANDWIDTH | 8050 | #ifdef CONFIG_CFS_BANDWIDTH |
8054 | static DEFINE_MUTEX(cfs_constraints_mutex); | 8051 | static DEFINE_MUTEX(cfs_constraints_mutex); |
8055 | 8052 | ||
8056 | const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ | 8053 | const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ |
8057 | const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ | 8054 | const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ |
8058 | 8055 | ||
8059 | static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); | 8056 | static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); |
8060 | 8057 | ||
8061 | static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) | 8058 | static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) |
8062 | { | 8059 | { |
8063 | int i, ret = 0, runtime_enabled, runtime_was_enabled; | 8060 | int i, ret = 0, runtime_enabled, runtime_was_enabled; |
8064 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; | 8061 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; |
8065 | 8062 | ||
8066 | if (tg == &root_task_group) | 8063 | if (tg == &root_task_group) |
8067 | return -EINVAL; | 8064 | return -EINVAL; |
8068 | 8065 | ||
8069 | /* | 8066 | /* |
8070 | * Ensure we have at some amount of bandwidth every period. This is | 8067 | * Ensure we have at some amount of bandwidth every period. This is |
8071 | * to prevent reaching a state of large arrears when throttled via | 8068 | * to prevent reaching a state of large arrears when throttled via |
8072 | * entity_tick() resulting in prolonged exit starvation. | 8069 | * entity_tick() resulting in prolonged exit starvation. |
8073 | */ | 8070 | */ |
8074 | if (quota < min_cfs_quota_period || period < min_cfs_quota_period) | 8071 | if (quota < min_cfs_quota_period || period < min_cfs_quota_period) |
8075 | return -EINVAL; | 8072 | return -EINVAL; |
8076 | 8073 | ||
8077 | /* | 8074 | /* |
8078 | * Likewise, bound things on the otherside by preventing insane quota | 8075 | * Likewise, bound things on the otherside by preventing insane quota |
8079 | * periods. This also allows us to normalize in computing quota | 8076 | * periods. This also allows us to normalize in computing quota |
8080 | * feasibility. | 8077 | * feasibility. |
8081 | */ | 8078 | */ |
8082 | if (period > max_cfs_quota_period) | 8079 | if (period > max_cfs_quota_period) |
8083 | return -EINVAL; | 8080 | return -EINVAL; |
8084 | 8081 | ||
8085 | /* | 8082 | /* |
8086 | * Prevent race between setting of cfs_rq->runtime_enabled and | 8083 | * Prevent race between setting of cfs_rq->runtime_enabled and |
8087 | * unthrottle_offline_cfs_rqs(). | 8084 | * unthrottle_offline_cfs_rqs(). |
8088 | */ | 8085 | */ |
8089 | get_online_cpus(); | 8086 | get_online_cpus(); |
8090 | mutex_lock(&cfs_constraints_mutex); | 8087 | mutex_lock(&cfs_constraints_mutex); |
8091 | ret = __cfs_schedulable(tg, period, quota); | 8088 | ret = __cfs_schedulable(tg, period, quota); |
8092 | if (ret) | 8089 | if (ret) |
8093 | goto out_unlock; | 8090 | goto out_unlock; |
8094 | 8091 | ||
8095 | runtime_enabled = quota != RUNTIME_INF; | 8092 | runtime_enabled = quota != RUNTIME_INF; |
8096 | runtime_was_enabled = cfs_b->quota != RUNTIME_INF; | 8093 | runtime_was_enabled = cfs_b->quota != RUNTIME_INF; |
8097 | /* | 8094 | /* |
8098 | * If we need to toggle cfs_bandwidth_used, off->on must occur | 8095 | * If we need to toggle cfs_bandwidth_used, off->on must occur |
8099 | * before making related changes, and on->off must occur afterwards | 8096 | * before making related changes, and on->off must occur afterwards |
8100 | */ | 8097 | */ |
8101 | if (runtime_enabled && !runtime_was_enabled) | 8098 | if (runtime_enabled && !runtime_was_enabled) |
8102 | cfs_bandwidth_usage_inc(); | 8099 | cfs_bandwidth_usage_inc(); |
8103 | raw_spin_lock_irq(&cfs_b->lock); | 8100 | raw_spin_lock_irq(&cfs_b->lock); |
8104 | cfs_b->period = ns_to_ktime(period); | 8101 | cfs_b->period = ns_to_ktime(period); |
8105 | cfs_b->quota = quota; | 8102 | cfs_b->quota = quota; |
8106 | 8103 | ||
8107 | __refill_cfs_bandwidth_runtime(cfs_b); | 8104 | __refill_cfs_bandwidth_runtime(cfs_b); |
8108 | /* restart the period timer (if active) to handle new period expiry */ | 8105 | /* restart the period timer (if active) to handle new period expiry */ |
8109 | if (runtime_enabled && cfs_b->timer_active) { | 8106 | if (runtime_enabled && cfs_b->timer_active) { |
8110 | /* force a reprogram */ | 8107 | /* force a reprogram */ |
8111 | __start_cfs_bandwidth(cfs_b, true); | 8108 | __start_cfs_bandwidth(cfs_b, true); |
8112 | } | 8109 | } |
8113 | raw_spin_unlock_irq(&cfs_b->lock); | 8110 | raw_spin_unlock_irq(&cfs_b->lock); |
8114 | 8111 | ||
8115 | for_each_online_cpu(i) { | 8112 | for_each_online_cpu(i) { |
8116 | struct cfs_rq *cfs_rq = tg->cfs_rq[i]; | 8113 | struct cfs_rq *cfs_rq = tg->cfs_rq[i]; |
8117 | struct rq *rq = cfs_rq->rq; | 8114 | struct rq *rq = cfs_rq->rq; |
8118 | 8115 | ||
8119 | raw_spin_lock_irq(&rq->lock); | 8116 | raw_spin_lock_irq(&rq->lock); |
8120 | cfs_rq->runtime_enabled = runtime_enabled; | 8117 | cfs_rq->runtime_enabled = runtime_enabled; |
8121 | cfs_rq->runtime_remaining = 0; | 8118 | cfs_rq->runtime_remaining = 0; |
8122 | 8119 | ||
8123 | if (cfs_rq->throttled) | 8120 | if (cfs_rq->throttled) |
8124 | unthrottle_cfs_rq(cfs_rq); | 8121 | unthrottle_cfs_rq(cfs_rq); |
8125 | raw_spin_unlock_irq(&rq->lock); | 8122 | raw_spin_unlock_irq(&rq->lock); |
8126 | } | 8123 | } |
8127 | if (runtime_was_enabled && !runtime_enabled) | 8124 | if (runtime_was_enabled && !runtime_enabled) |
8128 | cfs_bandwidth_usage_dec(); | 8125 | cfs_bandwidth_usage_dec(); |
8129 | out_unlock: | 8126 | out_unlock: |
8130 | mutex_unlock(&cfs_constraints_mutex); | 8127 | mutex_unlock(&cfs_constraints_mutex); |
8131 | put_online_cpus(); | 8128 | put_online_cpus(); |
8132 | 8129 | ||
8133 | return ret; | 8130 | return ret; |
8134 | } | 8131 | } |
8135 | 8132 | ||
8136 | int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) | 8133 | int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) |
8137 | { | 8134 | { |
8138 | u64 quota, period; | 8135 | u64 quota, period; |
8139 | 8136 | ||
8140 | period = ktime_to_ns(tg->cfs_bandwidth.period); | 8137 | period = ktime_to_ns(tg->cfs_bandwidth.period); |
8141 | if (cfs_quota_us < 0) | 8138 | if (cfs_quota_us < 0) |
8142 | quota = RUNTIME_INF; | 8139 | quota = RUNTIME_INF; |
8143 | else | 8140 | else |
8144 | quota = (u64)cfs_quota_us * NSEC_PER_USEC; | 8141 | quota = (u64)cfs_quota_us * NSEC_PER_USEC; |
8145 | 8142 | ||
8146 | return tg_set_cfs_bandwidth(tg, period, quota); | 8143 | return tg_set_cfs_bandwidth(tg, period, quota); |
8147 | } | 8144 | } |
8148 | 8145 | ||
8149 | long tg_get_cfs_quota(struct task_group *tg) | 8146 | long tg_get_cfs_quota(struct task_group *tg) |
8150 | { | 8147 | { |
8151 | u64 quota_us; | 8148 | u64 quota_us; |
8152 | 8149 | ||
8153 | if (tg->cfs_bandwidth.quota == RUNTIME_INF) | 8150 | if (tg->cfs_bandwidth.quota == RUNTIME_INF) |
8154 | return -1; | 8151 | return -1; |
8155 | 8152 | ||
8156 | quota_us = tg->cfs_bandwidth.quota; | 8153 | quota_us = tg->cfs_bandwidth.quota; |
8157 | do_div(quota_us, NSEC_PER_USEC); | 8154 | do_div(quota_us, NSEC_PER_USEC); |
8158 | 8155 | ||
8159 | return quota_us; | 8156 | return quota_us; |
8160 | } | 8157 | } |
8161 | 8158 | ||
8162 | int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) | 8159 | int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) |
8163 | { | 8160 | { |
8164 | u64 quota, period; | 8161 | u64 quota, period; |
8165 | 8162 | ||
8166 | period = (u64)cfs_period_us * NSEC_PER_USEC; | 8163 | period = (u64)cfs_period_us * NSEC_PER_USEC; |
8167 | quota = tg->cfs_bandwidth.quota; | 8164 | quota = tg->cfs_bandwidth.quota; |
8168 | 8165 | ||
8169 | return tg_set_cfs_bandwidth(tg, period, quota); | 8166 | return tg_set_cfs_bandwidth(tg, period, quota); |
8170 | } | 8167 | } |
8171 | 8168 | ||
8172 | long tg_get_cfs_period(struct task_group *tg) | 8169 | long tg_get_cfs_period(struct task_group *tg) |
8173 | { | 8170 | { |
8174 | u64 cfs_period_us; | 8171 | u64 cfs_period_us; |
8175 | 8172 | ||
8176 | cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); | 8173 | cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); |
8177 | do_div(cfs_period_us, NSEC_PER_USEC); | 8174 | do_div(cfs_period_us, NSEC_PER_USEC); |
8178 | 8175 | ||
8179 | return cfs_period_us; | 8176 | return cfs_period_us; |
8180 | } | 8177 | } |
8181 | 8178 | ||
8182 | static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, | 8179 | static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, |
8183 | struct cftype *cft) | 8180 | struct cftype *cft) |
8184 | { | 8181 | { |
8185 | return tg_get_cfs_quota(css_tg(css)); | 8182 | return tg_get_cfs_quota(css_tg(css)); |
8186 | } | 8183 | } |
8187 | 8184 | ||
8188 | static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, | 8185 | static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, |
8189 | struct cftype *cftype, s64 cfs_quota_us) | 8186 | struct cftype *cftype, s64 cfs_quota_us) |
8190 | { | 8187 | { |
8191 | return tg_set_cfs_quota(css_tg(css), cfs_quota_us); | 8188 | return tg_set_cfs_quota(css_tg(css), cfs_quota_us); |
8192 | } | 8189 | } |
8193 | 8190 | ||
8194 | static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, | 8191 | static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, |
8195 | struct cftype *cft) | 8192 | struct cftype *cft) |
8196 | { | 8193 | { |
8197 | return tg_get_cfs_period(css_tg(css)); | 8194 | return tg_get_cfs_period(css_tg(css)); |
8198 | } | 8195 | } |
8199 | 8196 | ||
8200 | static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, | 8197 | static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, |
8201 | struct cftype *cftype, u64 cfs_period_us) | 8198 | struct cftype *cftype, u64 cfs_period_us) |
8202 | { | 8199 | { |
8203 | return tg_set_cfs_period(css_tg(css), cfs_period_us); | 8200 | return tg_set_cfs_period(css_tg(css), cfs_period_us); |
8204 | } | 8201 | } |
8205 | 8202 | ||
8206 | struct cfs_schedulable_data { | 8203 | struct cfs_schedulable_data { |
8207 | struct task_group *tg; | 8204 | struct task_group *tg; |
8208 | u64 period, quota; | 8205 | u64 period, quota; |
8209 | }; | 8206 | }; |
8210 | 8207 | ||
8211 | /* | 8208 | /* |
8212 | * normalize group quota/period to be quota/max_period | 8209 | * normalize group quota/period to be quota/max_period |
8213 | * note: units are usecs | 8210 | * note: units are usecs |
8214 | */ | 8211 | */ |
8215 | static u64 normalize_cfs_quota(struct task_group *tg, | 8212 | static u64 normalize_cfs_quota(struct task_group *tg, |
8216 | struct cfs_schedulable_data *d) | 8213 | struct cfs_schedulable_data *d) |
8217 | { | 8214 | { |
8218 | u64 quota, period; | 8215 | u64 quota, period; |
8219 | 8216 | ||
8220 | if (tg == d->tg) { | 8217 | if (tg == d->tg) { |
8221 | period = d->period; | 8218 | period = d->period; |
8222 | quota = d->quota; | 8219 | quota = d->quota; |
8223 | } else { | 8220 | } else { |
8224 | period = tg_get_cfs_period(tg); | 8221 | period = tg_get_cfs_period(tg); |
8225 | quota = tg_get_cfs_quota(tg); | 8222 | quota = tg_get_cfs_quota(tg); |
8226 | } | 8223 | } |
8227 | 8224 | ||
8228 | /* note: these should typically be equivalent */ | 8225 | /* note: these should typically be equivalent */ |
8229 | if (quota == RUNTIME_INF || quota == -1) | 8226 | if (quota == RUNTIME_INF || quota == -1) |
8230 | return RUNTIME_INF; | 8227 | return RUNTIME_INF; |
8231 | 8228 | ||
8232 | return to_ratio(period, quota); | 8229 | return to_ratio(period, quota); |
8233 | } | 8230 | } |
8234 | 8231 | ||
8235 | static int tg_cfs_schedulable_down(struct task_group *tg, void *data) | 8232 | static int tg_cfs_schedulable_down(struct task_group *tg, void *data) |
8236 | { | 8233 | { |
8237 | struct cfs_schedulable_data *d = data; | 8234 | struct cfs_schedulable_data *d = data; |
8238 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; | 8235 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; |
8239 | s64 quota = 0, parent_quota = -1; | 8236 | s64 quota = 0, parent_quota = -1; |
8240 | 8237 | ||
8241 | if (!tg->parent) { | 8238 | if (!tg->parent) { |
8242 | quota = RUNTIME_INF; | 8239 | quota = RUNTIME_INF; |
8243 | } else { | 8240 | } else { |
8244 | struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; | 8241 | struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; |
8245 | 8242 | ||
8246 | quota = normalize_cfs_quota(tg, d); | 8243 | quota = normalize_cfs_quota(tg, d); |
8247 | parent_quota = parent_b->hierarchical_quota; | 8244 | parent_quota = parent_b->hierarchical_quota; |
8248 | 8245 | ||
8249 | /* | 8246 | /* |
8250 | * ensure max(child_quota) <= parent_quota, inherit when no | 8247 | * ensure max(child_quota) <= parent_quota, inherit when no |
8251 | * limit is set | 8248 | * limit is set |
8252 | */ | 8249 | */ |
8253 | if (quota == RUNTIME_INF) | 8250 | if (quota == RUNTIME_INF) |
8254 | quota = parent_quota; | 8251 | quota = parent_quota; |
8255 | else if (parent_quota != RUNTIME_INF && quota > parent_quota) | 8252 | else if (parent_quota != RUNTIME_INF && quota > parent_quota) |
8256 | return -EINVAL; | 8253 | return -EINVAL; |
8257 | } | 8254 | } |
8258 | cfs_b->hierarchical_quota = quota; | 8255 | cfs_b->hierarchical_quota = quota; |
8259 | 8256 | ||
8260 | return 0; | 8257 | return 0; |
8261 | } | 8258 | } |
8262 | 8259 | ||
8263 | static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) | 8260 | static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) |
8264 | { | 8261 | { |
8265 | int ret; | 8262 | int ret; |
8266 | struct cfs_schedulable_data data = { | 8263 | struct cfs_schedulable_data data = { |
8267 | .tg = tg, | 8264 | .tg = tg, |
8268 | .period = period, | 8265 | .period = period, |
8269 | .quota = quota, | 8266 | .quota = quota, |
8270 | }; | 8267 | }; |
8271 | 8268 | ||
8272 | if (quota != RUNTIME_INF) { | 8269 | if (quota != RUNTIME_INF) { |
8273 | do_div(data.period, NSEC_PER_USEC); | 8270 | do_div(data.period, NSEC_PER_USEC); |
8274 | do_div(data.quota, NSEC_PER_USEC); | 8271 | do_div(data.quota, NSEC_PER_USEC); |
8275 | } | 8272 | } |
8276 | 8273 | ||
8277 | rcu_read_lock(); | 8274 | rcu_read_lock(); |
8278 | ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); | 8275 | ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); |
8279 | rcu_read_unlock(); | 8276 | rcu_read_unlock(); |
8280 | 8277 | ||
8281 | return ret; | 8278 | return ret; |
8282 | } | 8279 | } |
8283 | 8280 | ||
8284 | static int cpu_stats_show(struct seq_file *sf, void *v) | 8281 | static int cpu_stats_show(struct seq_file *sf, void *v) |
8285 | { | 8282 | { |
8286 | struct task_group *tg = css_tg(seq_css(sf)); | 8283 | struct task_group *tg = css_tg(seq_css(sf)); |
8287 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; | 8284 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; |
8288 | 8285 | ||
8289 | seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); | 8286 | seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); |
8290 | seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); | 8287 | seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); |
8291 | seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); | 8288 | seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); |
8292 | 8289 | ||
8293 | return 0; | 8290 | return 0; |
8294 | } | 8291 | } |
8295 | #endif /* CONFIG_CFS_BANDWIDTH */ | 8292 | #endif /* CONFIG_CFS_BANDWIDTH */ |
8296 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | 8293 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
8297 | 8294 | ||
8298 | #ifdef CONFIG_RT_GROUP_SCHED | 8295 | #ifdef CONFIG_RT_GROUP_SCHED |
8299 | static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, | 8296 | static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, |
8300 | struct cftype *cft, s64 val) | 8297 | struct cftype *cft, s64 val) |
8301 | { | 8298 | { |
8302 | return sched_group_set_rt_runtime(css_tg(css), val); | 8299 | return sched_group_set_rt_runtime(css_tg(css), val); |
8303 | } | 8300 | } |
8304 | 8301 | ||
8305 | static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, | 8302 | static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, |
8306 | struct cftype *cft) | 8303 | struct cftype *cft) |
8307 | { | 8304 | { |
8308 | return sched_group_rt_runtime(css_tg(css)); | 8305 | return sched_group_rt_runtime(css_tg(css)); |
8309 | } | 8306 | } |
8310 | 8307 | ||
8311 | static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, | 8308 | static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, |
8312 | struct cftype *cftype, u64 rt_period_us) | 8309 | struct cftype *cftype, u64 rt_period_us) |
8313 | { | 8310 | { |
8314 | return sched_group_set_rt_period(css_tg(css), rt_period_us); | 8311 | return sched_group_set_rt_period(css_tg(css), rt_period_us); |
8315 | } | 8312 | } |
8316 | 8313 | ||
8317 | static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, | 8314 | static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, |
8318 | struct cftype *cft) | 8315 | struct cftype *cft) |
8319 | { | 8316 | { |
8320 | return sched_group_rt_period(css_tg(css)); | 8317 | return sched_group_rt_period(css_tg(css)); |
8321 | } | 8318 | } |
8322 | #endif /* CONFIG_RT_GROUP_SCHED */ | 8319 | #endif /* CONFIG_RT_GROUP_SCHED */ |
8323 | 8320 | ||
8324 | static struct cftype cpu_files[] = { | 8321 | static struct cftype cpu_files[] = { |
8325 | #ifdef CONFIG_FAIR_GROUP_SCHED | 8322 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8326 | { | 8323 | { |
8327 | .name = "shares", | 8324 | .name = "shares", |
8328 | .read_u64 = cpu_shares_read_u64, | 8325 | .read_u64 = cpu_shares_read_u64, |
8329 | .write_u64 = cpu_shares_write_u64, | 8326 | .write_u64 = cpu_shares_write_u64, |
8330 | }, | 8327 | }, |
8331 | #endif | 8328 | #endif |
8332 | #ifdef CONFIG_CFS_BANDWIDTH | 8329 | #ifdef CONFIG_CFS_BANDWIDTH |
8333 | { | 8330 | { |
8334 | .name = "cfs_quota_us", | 8331 | .name = "cfs_quota_us", |
8335 | .read_s64 = cpu_cfs_quota_read_s64, | 8332 | .read_s64 = cpu_cfs_quota_read_s64, |
8336 | .write_s64 = cpu_cfs_quota_write_s64, | 8333 | .write_s64 = cpu_cfs_quota_write_s64, |
8337 | }, | 8334 | }, |
8338 | { | 8335 | { |
8339 | .name = "cfs_period_us", | 8336 | .name = "cfs_period_us", |
8340 | .read_u64 = cpu_cfs_period_read_u64, | 8337 | .read_u64 = cpu_cfs_period_read_u64, |
8341 | .write_u64 = cpu_cfs_period_write_u64, | 8338 | .write_u64 = cpu_cfs_period_write_u64, |
8342 | }, | 8339 | }, |
8343 | { | 8340 | { |
8344 | .name = "stat", | 8341 | .name = "stat", |
8345 | .seq_show = cpu_stats_show, | 8342 | .seq_show = cpu_stats_show, |
8346 | }, | 8343 | }, |
8347 | #endif | 8344 | #endif |
8348 | #ifdef CONFIG_RT_GROUP_SCHED | 8345 | #ifdef CONFIG_RT_GROUP_SCHED |
8349 | { | 8346 | { |
8350 | .name = "rt_runtime_us", | 8347 | .name = "rt_runtime_us", |
8351 | .read_s64 = cpu_rt_runtime_read, | 8348 | .read_s64 = cpu_rt_runtime_read, |
8352 | .write_s64 = cpu_rt_runtime_write, | 8349 | .write_s64 = cpu_rt_runtime_write, |
8353 | }, | 8350 | }, |
8354 | { | 8351 | { |
8355 | .name = "rt_period_us", | 8352 | .name = "rt_period_us", |
8356 | .read_u64 = cpu_rt_period_read_uint, | 8353 | .read_u64 = cpu_rt_period_read_uint, |
8357 | .write_u64 = cpu_rt_period_write_uint, | 8354 | .write_u64 = cpu_rt_period_write_uint, |
8358 | }, | 8355 | }, |
8359 | #endif | 8356 | #endif |
8360 | { } /* terminate */ | 8357 | { } /* terminate */ |
8361 | }; | 8358 | }; |
8362 | 8359 | ||
8363 | struct cgroup_subsys cpu_cgrp_subsys = { | 8360 | struct cgroup_subsys cpu_cgrp_subsys = { |
8364 | .css_alloc = cpu_cgroup_css_alloc, | 8361 | .css_alloc = cpu_cgroup_css_alloc, |
8365 | .css_free = cpu_cgroup_css_free, | 8362 | .css_free = cpu_cgroup_css_free, |
8366 | .css_online = cpu_cgroup_css_online, | 8363 | .css_online = cpu_cgroup_css_online, |
8367 | .css_offline = cpu_cgroup_css_offline, | 8364 | .css_offline = cpu_cgroup_css_offline, |
8368 | .fork = cpu_cgroup_fork, | 8365 | .fork = cpu_cgroup_fork, |
8369 | .can_attach = cpu_cgroup_can_attach, | 8366 | .can_attach = cpu_cgroup_can_attach, |
8370 | .attach = cpu_cgroup_attach, | 8367 | .attach = cpu_cgroup_attach, |
8371 | .exit = cpu_cgroup_exit, | 8368 | .exit = cpu_cgroup_exit, |
8372 | .legacy_cftypes = cpu_files, | 8369 | .legacy_cftypes = cpu_files, |
8373 | .early_init = 1, | 8370 | .early_init = 1, |
8374 | }; | 8371 | }; |
8375 | 8372 | ||
8376 | #endif /* CONFIG_CGROUP_SCHED */ | 8373 | #endif /* CONFIG_CGROUP_SCHED */ |
8377 | 8374 | ||
8378 | void dump_cpu_task(int cpu) | 8375 | void dump_cpu_task(int cpu) |
8379 | { | 8376 | { |
8380 | pr_info("Task dump for CPU %d:\n", cpu); | 8377 | pr_info("Task dump for CPU %d:\n", cpu); |
8381 | sched_show_task(cpu_curr(cpu)); | 8378 | sched_show_task(cpu_curr(cpu)); |
kernel/sched/deadline.c
1 | /* | 1 | /* |
2 | * Deadline Scheduling Class (SCHED_DEADLINE) | 2 | * Deadline Scheduling Class (SCHED_DEADLINE) |
3 | * | 3 | * |
4 | * Earliest Deadline First (EDF) + Constant Bandwidth Server (CBS). | 4 | * Earliest Deadline First (EDF) + Constant Bandwidth Server (CBS). |
5 | * | 5 | * |
6 | * Tasks that periodically executes their instances for less than their | 6 | * Tasks that periodically executes their instances for less than their |
7 | * runtime won't miss any of their deadlines. | 7 | * runtime won't miss any of their deadlines. |
8 | * Tasks that are not periodic or sporadic or that tries to execute more | 8 | * Tasks that are not periodic or sporadic or that tries to execute more |
9 | * than their reserved bandwidth will be slowed down (and may potentially | 9 | * than their reserved bandwidth will be slowed down (and may potentially |
10 | * miss some of their deadlines), and won't affect any other task. | 10 | * miss some of their deadlines), and won't affect any other task. |
11 | * | 11 | * |
12 | * Copyright (C) 2012 Dario Faggioli <raistlin@linux.it>, | 12 | * Copyright (C) 2012 Dario Faggioli <raistlin@linux.it>, |
13 | * Juri Lelli <juri.lelli@gmail.com>, | 13 | * Juri Lelli <juri.lelli@gmail.com>, |
14 | * Michael Trimarchi <michael@amarulasolutions.com>, | 14 | * Michael Trimarchi <michael@amarulasolutions.com>, |
15 | * Fabio Checconi <fchecconi@gmail.com> | 15 | * Fabio Checconi <fchecconi@gmail.com> |
16 | */ | 16 | */ |
17 | #include "sched.h" | 17 | #include "sched.h" |
18 | 18 | ||
19 | #include <linux/slab.h> | 19 | #include <linux/slab.h> |
20 | 20 | ||
21 | struct dl_bandwidth def_dl_bandwidth; | 21 | struct dl_bandwidth def_dl_bandwidth; |
22 | 22 | ||
23 | static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se) | 23 | static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se) |
24 | { | 24 | { |
25 | return container_of(dl_se, struct task_struct, dl); | 25 | return container_of(dl_se, struct task_struct, dl); |
26 | } | 26 | } |
27 | 27 | ||
28 | static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq) | 28 | static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq) |
29 | { | 29 | { |
30 | return container_of(dl_rq, struct rq, dl); | 30 | return container_of(dl_rq, struct rq, dl); |
31 | } | 31 | } |
32 | 32 | ||
33 | static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se) | 33 | static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se) |
34 | { | 34 | { |
35 | struct task_struct *p = dl_task_of(dl_se); | 35 | struct task_struct *p = dl_task_of(dl_se); |
36 | struct rq *rq = task_rq(p); | 36 | struct rq *rq = task_rq(p); |
37 | 37 | ||
38 | return &rq->dl; | 38 | return &rq->dl; |
39 | } | 39 | } |
40 | 40 | ||
41 | static inline int on_dl_rq(struct sched_dl_entity *dl_se) | 41 | static inline int on_dl_rq(struct sched_dl_entity *dl_se) |
42 | { | 42 | { |
43 | return !RB_EMPTY_NODE(&dl_se->rb_node); | 43 | return !RB_EMPTY_NODE(&dl_se->rb_node); |
44 | } | 44 | } |
45 | 45 | ||
46 | static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq) | 46 | static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq) |
47 | { | 47 | { |
48 | struct sched_dl_entity *dl_se = &p->dl; | 48 | struct sched_dl_entity *dl_se = &p->dl; |
49 | 49 | ||
50 | return dl_rq->rb_leftmost == &dl_se->rb_node; | 50 | return dl_rq->rb_leftmost == &dl_se->rb_node; |
51 | } | 51 | } |
52 | 52 | ||
53 | void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime) | 53 | void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime) |
54 | { | 54 | { |
55 | raw_spin_lock_init(&dl_b->dl_runtime_lock); | 55 | raw_spin_lock_init(&dl_b->dl_runtime_lock); |
56 | dl_b->dl_period = period; | 56 | dl_b->dl_period = period; |
57 | dl_b->dl_runtime = runtime; | 57 | dl_b->dl_runtime = runtime; |
58 | } | 58 | } |
59 | 59 | ||
60 | void init_dl_bw(struct dl_bw *dl_b) | 60 | void init_dl_bw(struct dl_bw *dl_b) |
61 | { | 61 | { |
62 | raw_spin_lock_init(&dl_b->lock); | 62 | raw_spin_lock_init(&dl_b->lock); |
63 | raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock); | 63 | raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock); |
64 | if (global_rt_runtime() == RUNTIME_INF) | 64 | if (global_rt_runtime() == RUNTIME_INF) |
65 | dl_b->bw = -1; | 65 | dl_b->bw = -1; |
66 | else | 66 | else |
67 | dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime()); | 67 | dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime()); |
68 | raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock); | 68 | raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock); |
69 | dl_b->total_bw = 0; | 69 | dl_b->total_bw = 0; |
70 | } | 70 | } |
71 | 71 | ||
72 | void init_dl_rq(struct dl_rq *dl_rq, struct rq *rq) | 72 | void init_dl_rq(struct dl_rq *dl_rq, struct rq *rq) |
73 | { | 73 | { |
74 | dl_rq->rb_root = RB_ROOT; | 74 | dl_rq->rb_root = RB_ROOT; |
75 | 75 | ||
76 | #ifdef CONFIG_SMP | 76 | #ifdef CONFIG_SMP |
77 | /* zero means no -deadline tasks */ | 77 | /* zero means no -deadline tasks */ |
78 | dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0; | 78 | dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0; |
79 | 79 | ||
80 | dl_rq->dl_nr_migratory = 0; | 80 | dl_rq->dl_nr_migratory = 0; |
81 | dl_rq->overloaded = 0; | 81 | dl_rq->overloaded = 0; |
82 | dl_rq->pushable_dl_tasks_root = RB_ROOT; | 82 | dl_rq->pushable_dl_tasks_root = RB_ROOT; |
83 | #else | 83 | #else |
84 | init_dl_bw(&dl_rq->dl_bw); | 84 | init_dl_bw(&dl_rq->dl_bw); |
85 | #endif | 85 | #endif |
86 | } | 86 | } |
87 | 87 | ||
88 | #ifdef CONFIG_SMP | 88 | #ifdef CONFIG_SMP |
89 | 89 | ||
90 | static inline int dl_overloaded(struct rq *rq) | 90 | static inline int dl_overloaded(struct rq *rq) |
91 | { | 91 | { |
92 | return atomic_read(&rq->rd->dlo_count); | 92 | return atomic_read(&rq->rd->dlo_count); |
93 | } | 93 | } |
94 | 94 | ||
95 | static inline void dl_set_overload(struct rq *rq) | 95 | static inline void dl_set_overload(struct rq *rq) |
96 | { | 96 | { |
97 | if (!rq->online) | 97 | if (!rq->online) |
98 | return; | 98 | return; |
99 | 99 | ||
100 | cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask); | 100 | cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask); |
101 | /* | 101 | /* |
102 | * Must be visible before the overload count is | 102 | * Must be visible before the overload count is |
103 | * set (as in sched_rt.c). | 103 | * set (as in sched_rt.c). |
104 | * | 104 | * |
105 | * Matched by the barrier in pull_dl_task(). | 105 | * Matched by the barrier in pull_dl_task(). |
106 | */ | 106 | */ |
107 | smp_wmb(); | 107 | smp_wmb(); |
108 | atomic_inc(&rq->rd->dlo_count); | 108 | atomic_inc(&rq->rd->dlo_count); |
109 | } | 109 | } |
110 | 110 | ||
111 | static inline void dl_clear_overload(struct rq *rq) | 111 | static inline void dl_clear_overload(struct rq *rq) |
112 | { | 112 | { |
113 | if (!rq->online) | 113 | if (!rq->online) |
114 | return; | 114 | return; |
115 | 115 | ||
116 | atomic_dec(&rq->rd->dlo_count); | 116 | atomic_dec(&rq->rd->dlo_count); |
117 | cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask); | 117 | cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask); |
118 | } | 118 | } |
119 | 119 | ||
120 | static void update_dl_migration(struct dl_rq *dl_rq) | 120 | static void update_dl_migration(struct dl_rq *dl_rq) |
121 | { | 121 | { |
122 | if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) { | 122 | if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) { |
123 | if (!dl_rq->overloaded) { | 123 | if (!dl_rq->overloaded) { |
124 | dl_set_overload(rq_of_dl_rq(dl_rq)); | 124 | dl_set_overload(rq_of_dl_rq(dl_rq)); |
125 | dl_rq->overloaded = 1; | 125 | dl_rq->overloaded = 1; |
126 | } | 126 | } |
127 | } else if (dl_rq->overloaded) { | 127 | } else if (dl_rq->overloaded) { |
128 | dl_clear_overload(rq_of_dl_rq(dl_rq)); | 128 | dl_clear_overload(rq_of_dl_rq(dl_rq)); |
129 | dl_rq->overloaded = 0; | 129 | dl_rq->overloaded = 0; |
130 | } | 130 | } |
131 | } | 131 | } |
132 | 132 | ||
133 | static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | 133 | static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) |
134 | { | 134 | { |
135 | struct task_struct *p = dl_task_of(dl_se); | 135 | struct task_struct *p = dl_task_of(dl_se); |
136 | 136 | ||
137 | if (p->nr_cpus_allowed > 1) | 137 | if (p->nr_cpus_allowed > 1) |
138 | dl_rq->dl_nr_migratory++; | 138 | dl_rq->dl_nr_migratory++; |
139 | 139 | ||
140 | update_dl_migration(dl_rq); | 140 | update_dl_migration(dl_rq); |
141 | } | 141 | } |
142 | 142 | ||
143 | static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | 143 | static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) |
144 | { | 144 | { |
145 | struct task_struct *p = dl_task_of(dl_se); | 145 | struct task_struct *p = dl_task_of(dl_se); |
146 | 146 | ||
147 | if (p->nr_cpus_allowed > 1) | 147 | if (p->nr_cpus_allowed > 1) |
148 | dl_rq->dl_nr_migratory--; | 148 | dl_rq->dl_nr_migratory--; |
149 | 149 | ||
150 | update_dl_migration(dl_rq); | 150 | update_dl_migration(dl_rq); |
151 | } | 151 | } |
152 | 152 | ||
153 | /* | 153 | /* |
154 | * The list of pushable -deadline task is not a plist, like in | 154 | * The list of pushable -deadline task is not a plist, like in |
155 | * sched_rt.c, it is an rb-tree with tasks ordered by deadline. | 155 | * sched_rt.c, it is an rb-tree with tasks ordered by deadline. |
156 | */ | 156 | */ |
157 | static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) | 157 | static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) |
158 | { | 158 | { |
159 | struct dl_rq *dl_rq = &rq->dl; | 159 | struct dl_rq *dl_rq = &rq->dl; |
160 | struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_node; | 160 | struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_node; |
161 | struct rb_node *parent = NULL; | 161 | struct rb_node *parent = NULL; |
162 | struct task_struct *entry; | 162 | struct task_struct *entry; |
163 | int leftmost = 1; | 163 | int leftmost = 1; |
164 | 164 | ||
165 | BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks)); | 165 | BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks)); |
166 | 166 | ||
167 | while (*link) { | 167 | while (*link) { |
168 | parent = *link; | 168 | parent = *link; |
169 | entry = rb_entry(parent, struct task_struct, | 169 | entry = rb_entry(parent, struct task_struct, |
170 | pushable_dl_tasks); | 170 | pushable_dl_tasks); |
171 | if (dl_entity_preempt(&p->dl, &entry->dl)) | 171 | if (dl_entity_preempt(&p->dl, &entry->dl)) |
172 | link = &parent->rb_left; | 172 | link = &parent->rb_left; |
173 | else { | 173 | else { |
174 | link = &parent->rb_right; | 174 | link = &parent->rb_right; |
175 | leftmost = 0; | 175 | leftmost = 0; |
176 | } | 176 | } |
177 | } | 177 | } |
178 | 178 | ||
179 | if (leftmost) | 179 | if (leftmost) |
180 | dl_rq->pushable_dl_tasks_leftmost = &p->pushable_dl_tasks; | 180 | dl_rq->pushable_dl_tasks_leftmost = &p->pushable_dl_tasks; |
181 | 181 | ||
182 | rb_link_node(&p->pushable_dl_tasks, parent, link); | 182 | rb_link_node(&p->pushable_dl_tasks, parent, link); |
183 | rb_insert_color(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root); | 183 | rb_insert_color(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root); |
184 | } | 184 | } |
185 | 185 | ||
186 | static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) | 186 | static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) |
187 | { | 187 | { |
188 | struct dl_rq *dl_rq = &rq->dl; | 188 | struct dl_rq *dl_rq = &rq->dl; |
189 | 189 | ||
190 | if (RB_EMPTY_NODE(&p->pushable_dl_tasks)) | 190 | if (RB_EMPTY_NODE(&p->pushable_dl_tasks)) |
191 | return; | 191 | return; |
192 | 192 | ||
193 | if (dl_rq->pushable_dl_tasks_leftmost == &p->pushable_dl_tasks) { | 193 | if (dl_rq->pushable_dl_tasks_leftmost == &p->pushable_dl_tasks) { |
194 | struct rb_node *next_node; | 194 | struct rb_node *next_node; |
195 | 195 | ||
196 | next_node = rb_next(&p->pushable_dl_tasks); | 196 | next_node = rb_next(&p->pushable_dl_tasks); |
197 | dl_rq->pushable_dl_tasks_leftmost = next_node; | 197 | dl_rq->pushable_dl_tasks_leftmost = next_node; |
198 | } | 198 | } |
199 | 199 | ||
200 | rb_erase(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root); | 200 | rb_erase(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root); |
201 | RB_CLEAR_NODE(&p->pushable_dl_tasks); | 201 | RB_CLEAR_NODE(&p->pushable_dl_tasks); |
202 | } | 202 | } |
203 | 203 | ||
204 | static inline int has_pushable_dl_tasks(struct rq *rq) | 204 | static inline int has_pushable_dl_tasks(struct rq *rq) |
205 | { | 205 | { |
206 | return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root); | 206 | return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root); |
207 | } | 207 | } |
208 | 208 | ||
209 | static int push_dl_task(struct rq *rq); | 209 | static int push_dl_task(struct rq *rq); |
210 | 210 | ||
211 | static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev) | 211 | static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev) |
212 | { | 212 | { |
213 | return dl_task(prev); | 213 | return dl_task(prev); |
214 | } | 214 | } |
215 | 215 | ||
216 | static inline void set_post_schedule(struct rq *rq) | 216 | static inline void set_post_schedule(struct rq *rq) |
217 | { | 217 | { |
218 | rq->post_schedule = has_pushable_dl_tasks(rq); | 218 | rq->post_schedule = has_pushable_dl_tasks(rq); |
219 | } | 219 | } |
220 | 220 | ||
221 | #else | 221 | #else |
222 | 222 | ||
223 | static inline | 223 | static inline |
224 | void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) | 224 | void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) |
225 | { | 225 | { |
226 | } | 226 | } |
227 | 227 | ||
228 | static inline | 228 | static inline |
229 | void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) | 229 | void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) |
230 | { | 230 | { |
231 | } | 231 | } |
232 | 232 | ||
233 | static inline | 233 | static inline |
234 | void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | 234 | void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) |
235 | { | 235 | { |
236 | } | 236 | } |
237 | 237 | ||
238 | static inline | 238 | static inline |
239 | void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | 239 | void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) |
240 | { | 240 | { |
241 | } | 241 | } |
242 | 242 | ||
243 | static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev) | 243 | static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev) |
244 | { | 244 | { |
245 | return false; | 245 | return false; |
246 | } | 246 | } |
247 | 247 | ||
248 | static inline int pull_dl_task(struct rq *rq) | 248 | static inline int pull_dl_task(struct rq *rq) |
249 | { | 249 | { |
250 | return 0; | 250 | return 0; |
251 | } | 251 | } |
252 | 252 | ||
253 | static inline void set_post_schedule(struct rq *rq) | 253 | static inline void set_post_schedule(struct rq *rq) |
254 | { | 254 | { |
255 | } | 255 | } |
256 | #endif /* CONFIG_SMP */ | 256 | #endif /* CONFIG_SMP */ |
257 | 257 | ||
258 | static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags); | 258 | static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags); |
259 | static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags); | 259 | static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags); |
260 | static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, | 260 | static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, |
261 | int flags); | 261 | int flags); |
262 | 262 | ||
263 | /* | 263 | /* |
264 | * We are being explicitly informed that a new instance is starting, | 264 | * We are being explicitly informed that a new instance is starting, |
265 | * and this means that: | 265 | * and this means that: |
266 | * - the absolute deadline of the entity has to be placed at | 266 | * - the absolute deadline of the entity has to be placed at |
267 | * current time + relative deadline; | 267 | * current time + relative deadline; |
268 | * - the runtime of the entity has to be set to the maximum value. | 268 | * - the runtime of the entity has to be set to the maximum value. |
269 | * | 269 | * |
270 | * The capability of specifying such event is useful whenever a -deadline | 270 | * The capability of specifying such event is useful whenever a -deadline |
271 | * entity wants to (try to!) synchronize its behaviour with the scheduler's | 271 | * entity wants to (try to!) synchronize its behaviour with the scheduler's |
272 | * one, and to (try to!) reconcile itself with its own scheduling | 272 | * one, and to (try to!) reconcile itself with its own scheduling |
273 | * parameters. | 273 | * parameters. |
274 | */ | 274 | */ |
275 | static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se, | 275 | static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se, |
276 | struct sched_dl_entity *pi_se) | 276 | struct sched_dl_entity *pi_se) |
277 | { | 277 | { |
278 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | 278 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); |
279 | struct rq *rq = rq_of_dl_rq(dl_rq); | 279 | struct rq *rq = rq_of_dl_rq(dl_rq); |
280 | 280 | ||
281 | WARN_ON(!dl_se->dl_new || dl_se->dl_throttled); | 281 | WARN_ON(!dl_se->dl_new || dl_se->dl_throttled); |
282 | 282 | ||
283 | /* | 283 | /* |
284 | * We use the regular wall clock time to set deadlines in the | 284 | * We use the regular wall clock time to set deadlines in the |
285 | * future; in fact, we must consider execution overheads (time | 285 | * future; in fact, we must consider execution overheads (time |
286 | * spent on hardirq context, etc.). | 286 | * spent on hardirq context, etc.). |
287 | */ | 287 | */ |
288 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; | 288 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; |
289 | dl_se->runtime = pi_se->dl_runtime; | 289 | dl_se->runtime = pi_se->dl_runtime; |
290 | dl_se->dl_new = 0; | 290 | dl_se->dl_new = 0; |
291 | } | 291 | } |
292 | 292 | ||
293 | /* | 293 | /* |
294 | * Pure Earliest Deadline First (EDF) scheduling does not deal with the | 294 | * Pure Earliest Deadline First (EDF) scheduling does not deal with the |
295 | * possibility of a entity lasting more than what it declared, and thus | 295 | * possibility of a entity lasting more than what it declared, and thus |
296 | * exhausting its runtime. | 296 | * exhausting its runtime. |
297 | * | 297 | * |
298 | * Here we are interested in making runtime overrun possible, but we do | 298 | * Here we are interested in making runtime overrun possible, but we do |
299 | * not want a entity which is misbehaving to affect the scheduling of all | 299 | * not want a entity which is misbehaving to affect the scheduling of all |
300 | * other entities. | 300 | * other entities. |
301 | * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS) | 301 | * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS) |
302 | * is used, in order to confine each entity within its own bandwidth. | 302 | * is used, in order to confine each entity within its own bandwidth. |
303 | * | 303 | * |
304 | * This function deals exactly with that, and ensures that when the runtime | 304 | * This function deals exactly with that, and ensures that when the runtime |
305 | * of a entity is replenished, its deadline is also postponed. That ensures | 305 | * of a entity is replenished, its deadline is also postponed. That ensures |
306 | * the overrunning entity can't interfere with other entity in the system and | 306 | * the overrunning entity can't interfere with other entity in the system and |
307 | * can't make them miss their deadlines. Reasons why this kind of overruns | 307 | * can't make them miss their deadlines. Reasons why this kind of overruns |
308 | * could happen are, typically, a entity voluntarily trying to overcome its | 308 | * could happen are, typically, a entity voluntarily trying to overcome its |
309 | * runtime, or it just underestimated it during sched_setattr(). | 309 | * runtime, or it just underestimated it during sched_setattr(). |
310 | */ | 310 | */ |
311 | static void replenish_dl_entity(struct sched_dl_entity *dl_se, | 311 | static void replenish_dl_entity(struct sched_dl_entity *dl_se, |
312 | struct sched_dl_entity *pi_se) | 312 | struct sched_dl_entity *pi_se) |
313 | { | 313 | { |
314 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | 314 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); |
315 | struct rq *rq = rq_of_dl_rq(dl_rq); | 315 | struct rq *rq = rq_of_dl_rq(dl_rq); |
316 | 316 | ||
317 | BUG_ON(pi_se->dl_runtime <= 0); | 317 | BUG_ON(pi_se->dl_runtime <= 0); |
318 | 318 | ||
319 | /* | 319 | /* |
320 | * This could be the case for a !-dl task that is boosted. | 320 | * This could be the case for a !-dl task that is boosted. |
321 | * Just go with full inherited parameters. | 321 | * Just go with full inherited parameters. |
322 | */ | 322 | */ |
323 | if (dl_se->dl_deadline == 0) { | 323 | if (dl_se->dl_deadline == 0) { |
324 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; | 324 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; |
325 | dl_se->runtime = pi_se->dl_runtime; | 325 | dl_se->runtime = pi_se->dl_runtime; |
326 | } | 326 | } |
327 | 327 | ||
328 | /* | 328 | /* |
329 | * We keep moving the deadline away until we get some | 329 | * We keep moving the deadline away until we get some |
330 | * available runtime for the entity. This ensures correct | 330 | * available runtime for the entity. This ensures correct |
331 | * handling of situations where the runtime overrun is | 331 | * handling of situations where the runtime overrun is |
332 | * arbitrary large. | 332 | * arbitrary large. |
333 | */ | 333 | */ |
334 | while (dl_se->runtime <= 0) { | 334 | while (dl_se->runtime <= 0) { |
335 | dl_se->deadline += pi_se->dl_period; | 335 | dl_se->deadline += pi_se->dl_period; |
336 | dl_se->runtime += pi_se->dl_runtime; | 336 | dl_se->runtime += pi_se->dl_runtime; |
337 | } | 337 | } |
338 | 338 | ||
339 | /* | 339 | /* |
340 | * At this point, the deadline really should be "in | 340 | * At this point, the deadline really should be "in |
341 | * the future" with respect to rq->clock. If it's | 341 | * the future" with respect to rq->clock. If it's |
342 | * not, we are, for some reason, lagging too much! | 342 | * not, we are, for some reason, lagging too much! |
343 | * Anyway, after having warn userspace abut that, | 343 | * Anyway, after having warn userspace abut that, |
344 | * we still try to keep the things running by | 344 | * we still try to keep the things running by |
345 | * resetting the deadline and the budget of the | 345 | * resetting the deadline and the budget of the |
346 | * entity. | 346 | * entity. |
347 | */ | 347 | */ |
348 | if (dl_time_before(dl_se->deadline, rq_clock(rq))) { | 348 | if (dl_time_before(dl_se->deadline, rq_clock(rq))) { |
349 | printk_deferred_once("sched: DL replenish lagged to much\n"); | 349 | printk_deferred_once("sched: DL replenish lagged to much\n"); |
350 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; | 350 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; |
351 | dl_se->runtime = pi_se->dl_runtime; | 351 | dl_se->runtime = pi_se->dl_runtime; |
352 | } | 352 | } |
353 | } | 353 | } |
354 | 354 | ||
355 | /* | 355 | /* |
356 | * Here we check if --at time t-- an entity (which is probably being | 356 | * Here we check if --at time t-- an entity (which is probably being |
357 | * [re]activated or, in general, enqueued) can use its remaining runtime | 357 | * [re]activated or, in general, enqueued) can use its remaining runtime |
358 | * and its current deadline _without_ exceeding the bandwidth it is | 358 | * and its current deadline _without_ exceeding the bandwidth it is |
359 | * assigned (function returns true if it can't). We are in fact applying | 359 | * assigned (function returns true if it can't). We are in fact applying |
360 | * one of the CBS rules: when a task wakes up, if the residual runtime | 360 | * one of the CBS rules: when a task wakes up, if the residual runtime |
361 | * over residual deadline fits within the allocated bandwidth, then we | 361 | * over residual deadline fits within the allocated bandwidth, then we |
362 | * can keep the current (absolute) deadline and residual budget without | 362 | * can keep the current (absolute) deadline and residual budget without |
363 | * disrupting the schedulability of the system. Otherwise, we should | 363 | * disrupting the schedulability of the system. Otherwise, we should |
364 | * refill the runtime and set the deadline a period in the future, | 364 | * refill the runtime and set the deadline a period in the future, |
365 | * because keeping the current (absolute) deadline of the task would | 365 | * because keeping the current (absolute) deadline of the task would |
366 | * result in breaking guarantees promised to other tasks (refer to | 366 | * result in breaking guarantees promised to other tasks (refer to |
367 | * Documentation/scheduler/sched-deadline.txt for more informations). | 367 | * Documentation/scheduler/sched-deadline.txt for more informations). |
368 | * | 368 | * |
369 | * This function returns true if: | 369 | * This function returns true if: |
370 | * | 370 | * |
371 | * runtime / (deadline - t) > dl_runtime / dl_period , | 371 | * runtime / (deadline - t) > dl_runtime / dl_period , |
372 | * | 372 | * |
373 | * IOW we can't recycle current parameters. | 373 | * IOW we can't recycle current parameters. |
374 | * | 374 | * |
375 | * Notice that the bandwidth check is done against the period. For | 375 | * Notice that the bandwidth check is done against the period. For |
376 | * task with deadline equal to period this is the same of using | 376 | * task with deadline equal to period this is the same of using |
377 | * dl_deadline instead of dl_period in the equation above. | 377 | * dl_deadline instead of dl_period in the equation above. |
378 | */ | 378 | */ |
379 | static bool dl_entity_overflow(struct sched_dl_entity *dl_se, | 379 | static bool dl_entity_overflow(struct sched_dl_entity *dl_se, |
380 | struct sched_dl_entity *pi_se, u64 t) | 380 | struct sched_dl_entity *pi_se, u64 t) |
381 | { | 381 | { |
382 | u64 left, right; | 382 | u64 left, right; |
383 | 383 | ||
384 | /* | 384 | /* |
385 | * left and right are the two sides of the equation above, | 385 | * left and right are the two sides of the equation above, |
386 | * after a bit of shuffling to use multiplications instead | 386 | * after a bit of shuffling to use multiplications instead |
387 | * of divisions. | 387 | * of divisions. |
388 | * | 388 | * |
389 | * Note that none of the time values involved in the two | 389 | * Note that none of the time values involved in the two |
390 | * multiplications are absolute: dl_deadline and dl_runtime | 390 | * multiplications are absolute: dl_deadline and dl_runtime |
391 | * are the relative deadline and the maximum runtime of each | 391 | * are the relative deadline and the maximum runtime of each |
392 | * instance, runtime is the runtime left for the last instance | 392 | * instance, runtime is the runtime left for the last instance |
393 | * and (deadline - t), since t is rq->clock, is the time left | 393 | * and (deadline - t), since t is rq->clock, is the time left |
394 | * to the (absolute) deadline. Even if overflowing the u64 type | 394 | * to the (absolute) deadline. Even if overflowing the u64 type |
395 | * is very unlikely to occur in both cases, here we scale down | 395 | * is very unlikely to occur in both cases, here we scale down |
396 | * as we want to avoid that risk at all. Scaling down by 10 | 396 | * as we want to avoid that risk at all. Scaling down by 10 |
397 | * means that we reduce granularity to 1us. We are fine with it, | 397 | * means that we reduce granularity to 1us. We are fine with it, |
398 | * since this is only a true/false check and, anyway, thinking | 398 | * since this is only a true/false check and, anyway, thinking |
399 | * of anything below microseconds resolution is actually fiction | 399 | * of anything below microseconds resolution is actually fiction |
400 | * (but still we want to give the user that illusion >;). | 400 | * (but still we want to give the user that illusion >;). |
401 | */ | 401 | */ |
402 | left = (pi_se->dl_period >> DL_SCALE) * (dl_se->runtime >> DL_SCALE); | 402 | left = (pi_se->dl_period >> DL_SCALE) * (dl_se->runtime >> DL_SCALE); |
403 | right = ((dl_se->deadline - t) >> DL_SCALE) * | 403 | right = ((dl_se->deadline - t) >> DL_SCALE) * |
404 | (pi_se->dl_runtime >> DL_SCALE); | 404 | (pi_se->dl_runtime >> DL_SCALE); |
405 | 405 | ||
406 | return dl_time_before(right, left); | 406 | return dl_time_before(right, left); |
407 | } | 407 | } |
408 | 408 | ||
409 | /* | 409 | /* |
410 | * When a -deadline entity is queued back on the runqueue, its runtime and | 410 | * When a -deadline entity is queued back on the runqueue, its runtime and |
411 | * deadline might need updating. | 411 | * deadline might need updating. |
412 | * | 412 | * |
413 | * The policy here is that we update the deadline of the entity only if: | 413 | * The policy here is that we update the deadline of the entity only if: |
414 | * - the current deadline is in the past, | 414 | * - the current deadline is in the past, |
415 | * - using the remaining runtime with the current deadline would make | 415 | * - using the remaining runtime with the current deadline would make |
416 | * the entity exceed its bandwidth. | 416 | * the entity exceed its bandwidth. |
417 | */ | 417 | */ |
418 | static void update_dl_entity(struct sched_dl_entity *dl_se, | 418 | static void update_dl_entity(struct sched_dl_entity *dl_se, |
419 | struct sched_dl_entity *pi_se) | 419 | struct sched_dl_entity *pi_se) |
420 | { | 420 | { |
421 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | 421 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); |
422 | struct rq *rq = rq_of_dl_rq(dl_rq); | 422 | struct rq *rq = rq_of_dl_rq(dl_rq); |
423 | 423 | ||
424 | /* | 424 | /* |
425 | * The arrival of a new instance needs special treatment, i.e., | 425 | * The arrival of a new instance needs special treatment, i.e., |
426 | * the actual scheduling parameters have to be "renewed". | 426 | * the actual scheduling parameters have to be "renewed". |
427 | */ | 427 | */ |
428 | if (dl_se->dl_new) { | 428 | if (dl_se->dl_new) { |
429 | setup_new_dl_entity(dl_se, pi_se); | 429 | setup_new_dl_entity(dl_se, pi_se); |
430 | return; | 430 | return; |
431 | } | 431 | } |
432 | 432 | ||
433 | if (dl_time_before(dl_se->deadline, rq_clock(rq)) || | 433 | if (dl_time_before(dl_se->deadline, rq_clock(rq)) || |
434 | dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) { | 434 | dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) { |
435 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; | 435 | dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; |
436 | dl_se->runtime = pi_se->dl_runtime; | 436 | dl_se->runtime = pi_se->dl_runtime; |
437 | } | 437 | } |
438 | } | 438 | } |
439 | 439 | ||
440 | /* | 440 | /* |
441 | * If the entity depleted all its runtime, and if we want it to sleep | 441 | * If the entity depleted all its runtime, and if we want it to sleep |
442 | * while waiting for some new execution time to become available, we | 442 | * while waiting for some new execution time to become available, we |
443 | * set the bandwidth enforcement timer to the replenishment instant | 443 | * set the bandwidth enforcement timer to the replenishment instant |
444 | * and try to activate it. | 444 | * and try to activate it. |
445 | * | 445 | * |
446 | * Notice that it is important for the caller to know if the timer | 446 | * Notice that it is important for the caller to know if the timer |
447 | * actually started or not (i.e., the replenishment instant is in | 447 | * actually started or not (i.e., the replenishment instant is in |
448 | * the future or in the past). | 448 | * the future or in the past). |
449 | */ | 449 | */ |
450 | static int start_dl_timer(struct sched_dl_entity *dl_se, bool boosted) | 450 | static int start_dl_timer(struct sched_dl_entity *dl_se, bool boosted) |
451 | { | 451 | { |
452 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | 452 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); |
453 | struct rq *rq = rq_of_dl_rq(dl_rq); | 453 | struct rq *rq = rq_of_dl_rq(dl_rq); |
454 | ktime_t now, act; | 454 | ktime_t now, act; |
455 | ktime_t soft, hard; | 455 | ktime_t soft, hard; |
456 | unsigned long range; | 456 | unsigned long range; |
457 | s64 delta; | 457 | s64 delta; |
458 | 458 | ||
459 | if (boosted) | 459 | if (boosted) |
460 | return 0; | 460 | return 0; |
461 | /* | 461 | /* |
462 | * We want the timer to fire at the deadline, but considering | 462 | * We want the timer to fire at the deadline, but considering |
463 | * that it is actually coming from rq->clock and not from | 463 | * that it is actually coming from rq->clock and not from |
464 | * hrtimer's time base reading. | 464 | * hrtimer's time base reading. |
465 | */ | 465 | */ |
466 | act = ns_to_ktime(dl_se->deadline); | 466 | act = ns_to_ktime(dl_se->deadline); |
467 | now = hrtimer_cb_get_time(&dl_se->dl_timer); | 467 | now = hrtimer_cb_get_time(&dl_se->dl_timer); |
468 | delta = ktime_to_ns(now) - rq_clock(rq); | 468 | delta = ktime_to_ns(now) - rq_clock(rq); |
469 | act = ktime_add_ns(act, delta); | 469 | act = ktime_add_ns(act, delta); |
470 | 470 | ||
471 | /* | 471 | /* |
472 | * If the expiry time already passed, e.g., because the value | 472 | * If the expiry time already passed, e.g., because the value |
473 | * chosen as the deadline is too small, don't even try to | 473 | * chosen as the deadline is too small, don't even try to |
474 | * start the timer in the past! | 474 | * start the timer in the past! |
475 | */ | 475 | */ |
476 | if (ktime_us_delta(act, now) < 0) | 476 | if (ktime_us_delta(act, now) < 0) |
477 | return 0; | 477 | return 0; |
478 | 478 | ||
479 | hrtimer_set_expires(&dl_se->dl_timer, act); | 479 | hrtimer_set_expires(&dl_se->dl_timer, act); |
480 | 480 | ||
481 | soft = hrtimer_get_softexpires(&dl_se->dl_timer); | 481 | soft = hrtimer_get_softexpires(&dl_se->dl_timer); |
482 | hard = hrtimer_get_expires(&dl_se->dl_timer); | 482 | hard = hrtimer_get_expires(&dl_se->dl_timer); |
483 | range = ktime_to_ns(ktime_sub(hard, soft)); | 483 | range = ktime_to_ns(ktime_sub(hard, soft)); |
484 | __hrtimer_start_range_ns(&dl_se->dl_timer, soft, | 484 | __hrtimer_start_range_ns(&dl_se->dl_timer, soft, |
485 | range, HRTIMER_MODE_ABS, 0); | 485 | range, HRTIMER_MODE_ABS, 0); |
486 | 486 | ||
487 | return hrtimer_active(&dl_se->dl_timer); | 487 | return hrtimer_active(&dl_se->dl_timer); |
488 | } | 488 | } |
489 | 489 | ||
490 | /* | 490 | /* |
491 | * This is the bandwidth enforcement timer callback. If here, we know | 491 | * This is the bandwidth enforcement timer callback. If here, we know |
492 | * a task is not on its dl_rq, since the fact that the timer was running | 492 | * a task is not on its dl_rq, since the fact that the timer was running |
493 | * means the task is throttled and needs a runtime replenishment. | 493 | * means the task is throttled and needs a runtime replenishment. |
494 | * | 494 | * |
495 | * However, what we actually do depends on the fact the task is active, | 495 | * However, what we actually do depends on the fact the task is active, |
496 | * (it is on its rq) or has been removed from there by a call to | 496 | * (it is on its rq) or has been removed from there by a call to |
497 | * dequeue_task_dl(). In the former case we must issue the runtime | 497 | * dequeue_task_dl(). In the former case we must issue the runtime |
498 | * replenishment and add the task back to the dl_rq; in the latter, we just | 498 | * replenishment and add the task back to the dl_rq; in the latter, we just |
499 | * do nothing but clearing dl_throttled, so that runtime and deadline | 499 | * do nothing but clearing dl_throttled, so that runtime and deadline |
500 | * updating (and the queueing back to dl_rq) will be done by the | 500 | * updating (and the queueing back to dl_rq) will be done by the |
501 | * next call to enqueue_task_dl(). | 501 | * next call to enqueue_task_dl(). |
502 | */ | 502 | */ |
503 | static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) | 503 | static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) |
504 | { | 504 | { |
505 | struct sched_dl_entity *dl_se = container_of(timer, | 505 | struct sched_dl_entity *dl_se = container_of(timer, |
506 | struct sched_dl_entity, | 506 | struct sched_dl_entity, |
507 | dl_timer); | 507 | dl_timer); |
508 | struct task_struct *p = dl_task_of(dl_se); | 508 | struct task_struct *p = dl_task_of(dl_se); |
509 | struct rq *rq; | 509 | struct rq *rq; |
510 | again: | 510 | again: |
511 | rq = task_rq(p); | 511 | rq = task_rq(p); |
512 | raw_spin_lock(&rq->lock); | 512 | raw_spin_lock(&rq->lock); |
513 | 513 | ||
514 | if (rq != task_rq(p)) { | 514 | if (rq != task_rq(p)) { |
515 | /* Task was moved, retrying. */ | 515 | /* Task was moved, retrying. */ |
516 | raw_spin_unlock(&rq->lock); | 516 | raw_spin_unlock(&rq->lock); |
517 | goto again; | 517 | goto again; |
518 | } | 518 | } |
519 | 519 | ||
520 | /* | 520 | /* |
521 | * We need to take care of several possible races here: | 521 | * We need to take care of several possible races here: |
522 | * | 522 | * |
523 | * - the task might have changed its scheduling policy | 523 | * - the task might have changed its scheduling policy |
524 | * to something different than SCHED_DEADLINE | 524 | * to something different than SCHED_DEADLINE |
525 | * - the task might have changed its reservation parameters | 525 | * - the task might have changed its reservation parameters |
526 | * (through sched_setattr()) | 526 | * (through sched_setattr()) |
527 | * - the task might have been boosted by someone else and | 527 | * - the task might have been boosted by someone else and |
528 | * might be in the boosting/deboosting path | 528 | * might be in the boosting/deboosting path |
529 | * | 529 | * |
530 | * In all this cases we bail out, as the task is already | 530 | * In all this cases we bail out, as the task is already |
531 | * in the runqueue or is going to be enqueued back anyway. | 531 | * in the runqueue or is going to be enqueued back anyway. |
532 | */ | 532 | */ |
533 | if (!dl_task(p) || dl_se->dl_new || | 533 | if (!dl_task(p) || dl_se->dl_new || |
534 | dl_se->dl_boosted || !dl_se->dl_throttled) | 534 | dl_se->dl_boosted || !dl_se->dl_throttled) |
535 | goto unlock; | 535 | goto unlock; |
536 | 536 | ||
537 | sched_clock_tick(); | 537 | sched_clock_tick(); |
538 | update_rq_clock(rq); | 538 | update_rq_clock(rq); |
539 | dl_se->dl_throttled = 0; | 539 | dl_se->dl_throttled = 0; |
540 | dl_se->dl_yielded = 0; | 540 | dl_se->dl_yielded = 0; |
541 | if (task_on_rq_queued(p)) { | 541 | if (task_on_rq_queued(p)) { |
542 | enqueue_task_dl(rq, p, ENQUEUE_REPLENISH); | 542 | enqueue_task_dl(rq, p, ENQUEUE_REPLENISH); |
543 | if (dl_task(rq->curr)) | 543 | if (dl_task(rq->curr)) |
544 | check_preempt_curr_dl(rq, p, 0); | 544 | check_preempt_curr_dl(rq, p, 0); |
545 | else | 545 | else |
546 | resched_curr(rq); | 546 | resched_curr(rq); |
547 | #ifdef CONFIG_SMP | 547 | #ifdef CONFIG_SMP |
548 | /* | 548 | /* |
549 | * Queueing this task back might have overloaded rq, | 549 | * Queueing this task back might have overloaded rq, |
550 | * check if we need to kick someone away. | 550 | * check if we need to kick someone away. |
551 | */ | 551 | */ |
552 | if (has_pushable_dl_tasks(rq)) | 552 | if (has_pushable_dl_tasks(rq)) |
553 | push_dl_task(rq); | 553 | push_dl_task(rq); |
554 | #endif | 554 | #endif |
555 | } | 555 | } |
556 | unlock: | 556 | unlock: |
557 | raw_spin_unlock(&rq->lock); | 557 | raw_spin_unlock(&rq->lock); |
558 | 558 | ||
559 | return HRTIMER_NORESTART; | 559 | return HRTIMER_NORESTART; |
560 | } | 560 | } |
561 | 561 | ||
562 | void init_dl_task_timer(struct sched_dl_entity *dl_se) | 562 | void init_dl_task_timer(struct sched_dl_entity *dl_se) |
563 | { | 563 | { |
564 | struct hrtimer *timer = &dl_se->dl_timer; | 564 | struct hrtimer *timer = &dl_se->dl_timer; |
565 | 565 | ||
566 | hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | 566 | hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
567 | timer->function = dl_task_timer; | 567 | timer->function = dl_task_timer; |
568 | } | 568 | } |
569 | 569 | ||
570 | static | 570 | static |
571 | int dl_runtime_exceeded(struct rq *rq, struct sched_dl_entity *dl_se) | 571 | int dl_runtime_exceeded(struct rq *rq, struct sched_dl_entity *dl_se) |
572 | { | 572 | { |
573 | int dmiss = dl_time_before(dl_se->deadline, rq_clock(rq)); | 573 | return (dl_se->runtime <= 0); |
574 | int rorun = dl_se->runtime <= 0; | ||
575 | |||
576 | if (!rorun && !dmiss) | ||
577 | return 0; | ||
578 | |||
579 | /* | ||
580 | * If we are beyond our current deadline and we are still | ||
581 | * executing, then we have already used some of the runtime of | ||
582 | * the next instance. Thus, if we do not account that, we are | ||
583 | * stealing bandwidth from the system at each deadline miss! | ||
584 | */ | ||
585 | if (dmiss) { | ||
586 | dl_se->runtime = rorun ? dl_se->runtime : 0; | ||
587 | dl_se->runtime -= rq_clock(rq) - dl_se->deadline; | ||
588 | } | ||
589 | |||
590 | return 1; | ||
591 | } | 574 | } |
592 | 575 | ||
593 | extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq); | 576 | extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq); |
594 | 577 | ||
595 | /* | 578 | /* |
596 | * Update the current task's runtime statistics (provided it is still | 579 | * Update the current task's runtime statistics (provided it is still |
597 | * a -deadline task and has not been removed from the dl_rq). | 580 | * a -deadline task and has not been removed from the dl_rq). |
598 | */ | 581 | */ |
599 | static void update_curr_dl(struct rq *rq) | 582 | static void update_curr_dl(struct rq *rq) |
600 | { | 583 | { |
601 | struct task_struct *curr = rq->curr; | 584 | struct task_struct *curr = rq->curr; |
602 | struct sched_dl_entity *dl_se = &curr->dl; | 585 | struct sched_dl_entity *dl_se = &curr->dl; |
603 | u64 delta_exec; | 586 | u64 delta_exec; |
604 | 587 | ||
605 | if (!dl_task(curr) || !on_dl_rq(dl_se)) | 588 | if (!dl_task(curr) || !on_dl_rq(dl_se)) |
606 | return; | 589 | return; |
607 | 590 | ||
608 | /* | 591 | /* |
609 | * Consumed budget is computed considering the time as | 592 | * Consumed budget is computed considering the time as |
610 | * observed by schedulable tasks (excluding time spent | 593 | * observed by schedulable tasks (excluding time spent |
611 | * in hardirq context, etc.). Deadlines are instead | 594 | * in hardirq context, etc.). Deadlines are instead |
612 | * computed using hard walltime. This seems to be the more | 595 | * computed using hard walltime. This seems to be the more |
613 | * natural solution, but the full ramifications of this | 596 | * natural solution, but the full ramifications of this |
614 | * approach need further study. | 597 | * approach need further study. |
615 | */ | 598 | */ |
616 | delta_exec = rq_clock_task(rq) - curr->se.exec_start; | 599 | delta_exec = rq_clock_task(rq) - curr->se.exec_start; |
617 | if (unlikely((s64)delta_exec <= 0)) | 600 | if (unlikely((s64)delta_exec <= 0)) |
618 | return; | 601 | return; |
619 | 602 | ||
620 | schedstat_set(curr->se.statistics.exec_max, | 603 | schedstat_set(curr->se.statistics.exec_max, |
621 | max(curr->se.statistics.exec_max, delta_exec)); | 604 | max(curr->se.statistics.exec_max, delta_exec)); |
622 | 605 | ||
623 | curr->se.sum_exec_runtime += delta_exec; | 606 | curr->se.sum_exec_runtime += delta_exec; |
624 | account_group_exec_runtime(curr, delta_exec); | 607 | account_group_exec_runtime(curr, delta_exec); |
625 | 608 | ||
626 | curr->se.exec_start = rq_clock_task(rq); | 609 | curr->se.exec_start = rq_clock_task(rq); |
627 | cpuacct_charge(curr, delta_exec); | 610 | cpuacct_charge(curr, delta_exec); |
628 | 611 | ||
629 | sched_rt_avg_update(rq, delta_exec); | 612 | sched_rt_avg_update(rq, delta_exec); |
630 | 613 | ||
631 | dl_se->runtime -= dl_se->dl_yielded ? 0 : delta_exec; | 614 | dl_se->runtime -= dl_se->dl_yielded ? 0 : delta_exec; |
632 | if (dl_runtime_exceeded(rq, dl_se)) { | 615 | if (dl_runtime_exceeded(rq, dl_se)) { |
633 | __dequeue_task_dl(rq, curr, 0); | 616 | __dequeue_task_dl(rq, curr, 0); |
634 | if (likely(start_dl_timer(dl_se, curr->dl.dl_boosted))) | 617 | if (likely(start_dl_timer(dl_se, curr->dl.dl_boosted))) |
635 | dl_se->dl_throttled = 1; | 618 | dl_se->dl_throttled = 1; |
636 | else | 619 | else |
637 | enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH); | 620 | enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH); |
638 | 621 | ||
639 | if (!is_leftmost(curr, &rq->dl)) | 622 | if (!is_leftmost(curr, &rq->dl)) |
640 | resched_curr(rq); | 623 | resched_curr(rq); |
641 | } | 624 | } |
642 | 625 | ||
643 | /* | 626 | /* |
644 | * Because -- for now -- we share the rt bandwidth, we need to | 627 | * Because -- for now -- we share the rt bandwidth, we need to |
645 | * account our runtime there too, otherwise actual rt tasks | 628 | * account our runtime there too, otherwise actual rt tasks |
646 | * would be able to exceed the shared quota. | 629 | * would be able to exceed the shared quota. |
647 | * | 630 | * |
648 | * Account to the root rt group for now. | 631 | * Account to the root rt group for now. |
649 | * | 632 | * |
650 | * The solution we're working towards is having the RT groups scheduled | 633 | * The solution we're working towards is having the RT groups scheduled |
651 | * using deadline servers -- however there's a few nasties to figure | 634 | * using deadline servers -- however there's a few nasties to figure |
652 | * out before that can happen. | 635 | * out before that can happen. |
653 | */ | 636 | */ |
654 | if (rt_bandwidth_enabled()) { | 637 | if (rt_bandwidth_enabled()) { |
655 | struct rt_rq *rt_rq = &rq->rt; | 638 | struct rt_rq *rt_rq = &rq->rt; |
656 | 639 | ||
657 | raw_spin_lock(&rt_rq->rt_runtime_lock); | 640 | raw_spin_lock(&rt_rq->rt_runtime_lock); |
658 | /* | 641 | /* |
659 | * We'll let actual RT tasks worry about the overflow here, we | 642 | * We'll let actual RT tasks worry about the overflow here, we |
660 | * have our own CBS to keep us inline; only account when RT | 643 | * have our own CBS to keep us inline; only account when RT |
661 | * bandwidth is relevant. | 644 | * bandwidth is relevant. |
662 | */ | 645 | */ |
663 | if (sched_rt_bandwidth_account(rt_rq)) | 646 | if (sched_rt_bandwidth_account(rt_rq)) |
664 | rt_rq->rt_time += delta_exec; | 647 | rt_rq->rt_time += delta_exec; |
665 | raw_spin_unlock(&rt_rq->rt_runtime_lock); | 648 | raw_spin_unlock(&rt_rq->rt_runtime_lock); |
666 | } | 649 | } |
667 | } | 650 | } |
668 | 651 | ||
669 | #ifdef CONFIG_SMP | 652 | #ifdef CONFIG_SMP |
670 | 653 | ||
671 | static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu); | 654 | static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu); |
672 | 655 | ||
673 | static inline u64 next_deadline(struct rq *rq) | 656 | static inline u64 next_deadline(struct rq *rq) |
674 | { | 657 | { |
675 | struct task_struct *next = pick_next_earliest_dl_task(rq, rq->cpu); | 658 | struct task_struct *next = pick_next_earliest_dl_task(rq, rq->cpu); |
676 | 659 | ||
677 | if (next && dl_prio(next->prio)) | 660 | if (next && dl_prio(next->prio)) |
678 | return next->dl.deadline; | 661 | return next->dl.deadline; |
679 | else | 662 | else |
680 | return 0; | 663 | return 0; |
681 | } | 664 | } |
682 | 665 | ||
683 | static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) | 666 | static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) |
684 | { | 667 | { |
685 | struct rq *rq = rq_of_dl_rq(dl_rq); | 668 | struct rq *rq = rq_of_dl_rq(dl_rq); |
686 | 669 | ||
687 | if (dl_rq->earliest_dl.curr == 0 || | 670 | if (dl_rq->earliest_dl.curr == 0 || |
688 | dl_time_before(deadline, dl_rq->earliest_dl.curr)) { | 671 | dl_time_before(deadline, dl_rq->earliest_dl.curr)) { |
689 | /* | 672 | /* |
690 | * If the dl_rq had no -deadline tasks, or if the new task | 673 | * If the dl_rq had no -deadline tasks, or if the new task |
691 | * has shorter deadline than the current one on dl_rq, we | 674 | * has shorter deadline than the current one on dl_rq, we |
692 | * know that the previous earliest becomes our next earliest, | 675 | * know that the previous earliest becomes our next earliest, |
693 | * as the new task becomes the earliest itself. | 676 | * as the new task becomes the earliest itself. |
694 | */ | 677 | */ |
695 | dl_rq->earliest_dl.next = dl_rq->earliest_dl.curr; | 678 | dl_rq->earliest_dl.next = dl_rq->earliest_dl.curr; |
696 | dl_rq->earliest_dl.curr = deadline; | 679 | dl_rq->earliest_dl.curr = deadline; |
697 | cpudl_set(&rq->rd->cpudl, rq->cpu, deadline, 1); | 680 | cpudl_set(&rq->rd->cpudl, rq->cpu, deadline, 1); |
698 | } else if (dl_rq->earliest_dl.next == 0 || | 681 | } else if (dl_rq->earliest_dl.next == 0 || |
699 | dl_time_before(deadline, dl_rq->earliest_dl.next)) { | 682 | dl_time_before(deadline, dl_rq->earliest_dl.next)) { |
700 | /* | 683 | /* |
701 | * On the other hand, if the new -deadline task has a | 684 | * On the other hand, if the new -deadline task has a |
702 | * a later deadline than the earliest one on dl_rq, but | 685 | * a later deadline than the earliest one on dl_rq, but |
703 | * it is earlier than the next (if any), we must | 686 | * it is earlier than the next (if any), we must |
704 | * recompute the next-earliest. | 687 | * recompute the next-earliest. |
705 | */ | 688 | */ |
706 | dl_rq->earliest_dl.next = next_deadline(rq); | 689 | dl_rq->earliest_dl.next = next_deadline(rq); |
707 | } | 690 | } |
708 | } | 691 | } |
709 | 692 | ||
710 | static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) | 693 | static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) |
711 | { | 694 | { |
712 | struct rq *rq = rq_of_dl_rq(dl_rq); | 695 | struct rq *rq = rq_of_dl_rq(dl_rq); |
713 | 696 | ||
714 | /* | 697 | /* |
715 | * Since we may have removed our earliest (and/or next earliest) | 698 | * Since we may have removed our earliest (and/or next earliest) |
716 | * task we must recompute them. | 699 | * task we must recompute them. |
717 | */ | 700 | */ |
718 | if (!dl_rq->dl_nr_running) { | 701 | if (!dl_rq->dl_nr_running) { |
719 | dl_rq->earliest_dl.curr = 0; | 702 | dl_rq->earliest_dl.curr = 0; |
720 | dl_rq->earliest_dl.next = 0; | 703 | dl_rq->earliest_dl.next = 0; |
721 | cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0); | 704 | cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0); |
722 | } else { | 705 | } else { |
723 | struct rb_node *leftmost = dl_rq->rb_leftmost; | 706 | struct rb_node *leftmost = dl_rq->rb_leftmost; |
724 | struct sched_dl_entity *entry; | 707 | struct sched_dl_entity *entry; |
725 | 708 | ||
726 | entry = rb_entry(leftmost, struct sched_dl_entity, rb_node); | 709 | entry = rb_entry(leftmost, struct sched_dl_entity, rb_node); |
727 | dl_rq->earliest_dl.curr = entry->deadline; | 710 | dl_rq->earliest_dl.curr = entry->deadline; |
728 | dl_rq->earliest_dl.next = next_deadline(rq); | 711 | dl_rq->earliest_dl.next = next_deadline(rq); |
729 | cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline, 1); | 712 | cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline, 1); |
730 | } | 713 | } |
731 | } | 714 | } |
732 | 715 | ||
733 | #else | 716 | #else |
734 | 717 | ||
735 | static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} | 718 | static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} |
736 | static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} | 719 | static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} |
737 | 720 | ||
738 | #endif /* CONFIG_SMP */ | 721 | #endif /* CONFIG_SMP */ |
739 | 722 | ||
740 | static inline | 723 | static inline |
741 | void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | 724 | void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) |
742 | { | 725 | { |
743 | int prio = dl_task_of(dl_se)->prio; | 726 | int prio = dl_task_of(dl_se)->prio; |
744 | u64 deadline = dl_se->deadline; | 727 | u64 deadline = dl_se->deadline; |
745 | 728 | ||
746 | WARN_ON(!dl_prio(prio)); | 729 | WARN_ON(!dl_prio(prio)); |
747 | dl_rq->dl_nr_running++; | 730 | dl_rq->dl_nr_running++; |
748 | add_nr_running(rq_of_dl_rq(dl_rq), 1); | 731 | add_nr_running(rq_of_dl_rq(dl_rq), 1); |
749 | 732 | ||
750 | inc_dl_deadline(dl_rq, deadline); | 733 | inc_dl_deadline(dl_rq, deadline); |
751 | inc_dl_migration(dl_se, dl_rq); | 734 | inc_dl_migration(dl_se, dl_rq); |
752 | } | 735 | } |
753 | 736 | ||
754 | static inline | 737 | static inline |
755 | void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) | 738 | void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) |
756 | { | 739 | { |
757 | int prio = dl_task_of(dl_se)->prio; | 740 | int prio = dl_task_of(dl_se)->prio; |
758 | 741 | ||
759 | WARN_ON(!dl_prio(prio)); | 742 | WARN_ON(!dl_prio(prio)); |
760 | WARN_ON(!dl_rq->dl_nr_running); | 743 | WARN_ON(!dl_rq->dl_nr_running); |
761 | dl_rq->dl_nr_running--; | 744 | dl_rq->dl_nr_running--; |
762 | sub_nr_running(rq_of_dl_rq(dl_rq), 1); | 745 | sub_nr_running(rq_of_dl_rq(dl_rq), 1); |
763 | 746 | ||
764 | dec_dl_deadline(dl_rq, dl_se->deadline); | 747 | dec_dl_deadline(dl_rq, dl_se->deadline); |
765 | dec_dl_migration(dl_se, dl_rq); | 748 | dec_dl_migration(dl_se, dl_rq); |
766 | } | 749 | } |
767 | 750 | ||
768 | static void __enqueue_dl_entity(struct sched_dl_entity *dl_se) | 751 | static void __enqueue_dl_entity(struct sched_dl_entity *dl_se) |
769 | { | 752 | { |
770 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | 753 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); |
771 | struct rb_node **link = &dl_rq->rb_root.rb_node; | 754 | struct rb_node **link = &dl_rq->rb_root.rb_node; |
772 | struct rb_node *parent = NULL; | 755 | struct rb_node *parent = NULL; |
773 | struct sched_dl_entity *entry; | 756 | struct sched_dl_entity *entry; |
774 | int leftmost = 1; | 757 | int leftmost = 1; |
775 | 758 | ||
776 | BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node)); | 759 | BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node)); |
777 | 760 | ||
778 | while (*link) { | 761 | while (*link) { |
779 | parent = *link; | 762 | parent = *link; |
780 | entry = rb_entry(parent, struct sched_dl_entity, rb_node); | 763 | entry = rb_entry(parent, struct sched_dl_entity, rb_node); |
781 | if (dl_time_before(dl_se->deadline, entry->deadline)) | 764 | if (dl_time_before(dl_se->deadline, entry->deadline)) |
782 | link = &parent->rb_left; | 765 | link = &parent->rb_left; |
783 | else { | 766 | else { |
784 | link = &parent->rb_right; | 767 | link = &parent->rb_right; |
785 | leftmost = 0; | 768 | leftmost = 0; |
786 | } | 769 | } |
787 | } | 770 | } |
788 | 771 | ||
789 | if (leftmost) | 772 | if (leftmost) |
790 | dl_rq->rb_leftmost = &dl_se->rb_node; | 773 | dl_rq->rb_leftmost = &dl_se->rb_node; |
791 | 774 | ||
792 | rb_link_node(&dl_se->rb_node, parent, link); | 775 | rb_link_node(&dl_se->rb_node, parent, link); |
793 | rb_insert_color(&dl_se->rb_node, &dl_rq->rb_root); | 776 | rb_insert_color(&dl_se->rb_node, &dl_rq->rb_root); |
794 | 777 | ||
795 | inc_dl_tasks(dl_se, dl_rq); | 778 | inc_dl_tasks(dl_se, dl_rq); |
796 | } | 779 | } |
797 | 780 | ||
798 | static void __dequeue_dl_entity(struct sched_dl_entity *dl_se) | 781 | static void __dequeue_dl_entity(struct sched_dl_entity *dl_se) |
799 | { | 782 | { |
800 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); | 783 | struct dl_rq *dl_rq = dl_rq_of_se(dl_se); |
801 | 784 | ||
802 | if (RB_EMPTY_NODE(&dl_se->rb_node)) | 785 | if (RB_EMPTY_NODE(&dl_se->rb_node)) |
803 | return; | 786 | return; |
804 | 787 | ||
805 | if (dl_rq->rb_leftmost == &dl_se->rb_node) { | 788 | if (dl_rq->rb_leftmost == &dl_se->rb_node) { |
806 | struct rb_node *next_node; | 789 | struct rb_node *next_node; |
807 | 790 | ||
808 | next_node = rb_next(&dl_se->rb_node); | 791 | next_node = rb_next(&dl_se->rb_node); |
809 | dl_rq->rb_leftmost = next_node; | 792 | dl_rq->rb_leftmost = next_node; |
810 | } | 793 | } |
811 | 794 | ||
812 | rb_erase(&dl_se->rb_node, &dl_rq->rb_root); | 795 | rb_erase(&dl_se->rb_node, &dl_rq->rb_root); |
813 | RB_CLEAR_NODE(&dl_se->rb_node); | 796 | RB_CLEAR_NODE(&dl_se->rb_node); |
814 | 797 | ||
815 | dec_dl_tasks(dl_se, dl_rq); | 798 | dec_dl_tasks(dl_se, dl_rq); |
816 | } | 799 | } |
817 | 800 | ||
818 | static void | 801 | static void |
819 | enqueue_dl_entity(struct sched_dl_entity *dl_se, | 802 | enqueue_dl_entity(struct sched_dl_entity *dl_se, |
820 | struct sched_dl_entity *pi_se, int flags) | 803 | struct sched_dl_entity *pi_se, int flags) |
821 | { | 804 | { |
822 | BUG_ON(on_dl_rq(dl_se)); | 805 | BUG_ON(on_dl_rq(dl_se)); |
823 | 806 | ||
824 | /* | 807 | /* |
825 | * If this is a wakeup or a new instance, the scheduling | 808 | * If this is a wakeup or a new instance, the scheduling |
826 | * parameters of the task might need updating. Otherwise, | 809 | * parameters of the task might need updating. Otherwise, |
827 | * we want a replenishment of its runtime. | 810 | * we want a replenishment of its runtime. |
828 | */ | 811 | */ |
829 | if (!dl_se->dl_new && flags & ENQUEUE_REPLENISH) | 812 | if (dl_se->dl_new || flags & ENQUEUE_WAKEUP) |
830 | replenish_dl_entity(dl_se, pi_se); | ||
831 | else | ||
832 | update_dl_entity(dl_se, pi_se); | 813 | update_dl_entity(dl_se, pi_se); |
814 | else if (flags & ENQUEUE_REPLENISH) | ||
815 | replenish_dl_entity(dl_se, pi_se); | ||
833 | 816 | ||
834 | __enqueue_dl_entity(dl_se); | 817 | __enqueue_dl_entity(dl_se); |
835 | } | 818 | } |
836 | 819 | ||
837 | static void dequeue_dl_entity(struct sched_dl_entity *dl_se) | 820 | static void dequeue_dl_entity(struct sched_dl_entity *dl_se) |
838 | { | 821 | { |
839 | __dequeue_dl_entity(dl_se); | 822 | __dequeue_dl_entity(dl_se); |
840 | } | 823 | } |
841 | 824 | ||
842 | static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags) | 825 | static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags) |
843 | { | 826 | { |
844 | struct task_struct *pi_task = rt_mutex_get_top_task(p); | 827 | struct task_struct *pi_task = rt_mutex_get_top_task(p); |
845 | struct sched_dl_entity *pi_se = &p->dl; | 828 | struct sched_dl_entity *pi_se = &p->dl; |
846 | 829 | ||
847 | /* | 830 | /* |
848 | * Use the scheduling parameters of the top pi-waiter | 831 | * Use the scheduling parameters of the top pi-waiter |
849 | * task if we have one and its (relative) deadline is | 832 | * task if we have one and its (relative) deadline is |
850 | * smaller than our one... OTW we keep our runtime and | 833 | * smaller than our one... OTW we keep our runtime and |
851 | * deadline. | 834 | * deadline. |
852 | */ | 835 | */ |
853 | if (pi_task && p->dl.dl_boosted && dl_prio(pi_task->normal_prio)) { | 836 | if (pi_task && p->dl.dl_boosted && dl_prio(pi_task->normal_prio)) { |
854 | pi_se = &pi_task->dl; | 837 | pi_se = &pi_task->dl; |
855 | } else if (!dl_prio(p->normal_prio)) { | 838 | } else if (!dl_prio(p->normal_prio)) { |
856 | /* | 839 | /* |
857 | * Special case in which we have a !SCHED_DEADLINE task | 840 | * Special case in which we have a !SCHED_DEADLINE task |
858 | * that is going to be deboosted, but exceedes its | 841 | * that is going to be deboosted, but exceedes its |
859 | * runtime while doing so. No point in replenishing | 842 | * runtime while doing so. No point in replenishing |
860 | * it, as it's going to return back to its original | 843 | * it, as it's going to return back to its original |
861 | * scheduling class after this. | 844 | * scheduling class after this. |
862 | */ | 845 | */ |
863 | BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH); | 846 | BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH); |
864 | return; | 847 | return; |
865 | } | 848 | } |
866 | 849 | ||
867 | /* | 850 | /* |
868 | * If p is throttled, we do nothing. In fact, if it exhausted | 851 | * If p is throttled, we do nothing. In fact, if it exhausted |
869 | * its budget it needs a replenishment and, since it now is on | 852 | * its budget it needs a replenishment and, since it now is on |
870 | * its rq, the bandwidth timer callback (which clearly has not | 853 | * its rq, the bandwidth timer callback (which clearly has not |
871 | * run yet) will take care of this. | 854 | * run yet) will take care of this. |
872 | */ | 855 | */ |
873 | if (p->dl.dl_throttled) | 856 | if (p->dl.dl_throttled) |
874 | return; | 857 | return; |
875 | 858 | ||
876 | enqueue_dl_entity(&p->dl, pi_se, flags); | 859 | enqueue_dl_entity(&p->dl, pi_se, flags); |
877 | 860 | ||
878 | if (!task_current(rq, p) && p->nr_cpus_allowed > 1) | 861 | if (!task_current(rq, p) && p->nr_cpus_allowed > 1) |
879 | enqueue_pushable_dl_task(rq, p); | 862 | enqueue_pushable_dl_task(rq, p); |
880 | } | 863 | } |
881 | 864 | ||
882 | static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) | 865 | static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) |
883 | { | 866 | { |
884 | dequeue_dl_entity(&p->dl); | 867 | dequeue_dl_entity(&p->dl); |
885 | dequeue_pushable_dl_task(rq, p); | 868 | dequeue_pushable_dl_task(rq, p); |
886 | } | 869 | } |
887 | 870 | ||
888 | static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) | 871 | static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) |
889 | { | 872 | { |
890 | update_curr_dl(rq); | 873 | update_curr_dl(rq); |
891 | __dequeue_task_dl(rq, p, flags); | 874 | __dequeue_task_dl(rq, p, flags); |
892 | } | 875 | } |
893 | 876 | ||
894 | /* | 877 | /* |
895 | * Yield task semantic for -deadline tasks is: | 878 | * Yield task semantic for -deadline tasks is: |
896 | * | 879 | * |
897 | * get off from the CPU until our next instance, with | 880 | * get off from the CPU until our next instance, with |
898 | * a new runtime. This is of little use now, since we | 881 | * a new runtime. This is of little use now, since we |
899 | * don't have a bandwidth reclaiming mechanism. Anyway, | 882 | * don't have a bandwidth reclaiming mechanism. Anyway, |
900 | * bandwidth reclaiming is planned for the future, and | 883 | * bandwidth reclaiming is planned for the future, and |
901 | * yield_task_dl will indicate that some spare budget | 884 | * yield_task_dl will indicate that some spare budget |
902 | * is available for other task instances to use it. | 885 | * is available for other task instances to use it. |
903 | */ | 886 | */ |
904 | static void yield_task_dl(struct rq *rq) | 887 | static void yield_task_dl(struct rq *rq) |
905 | { | 888 | { |
906 | struct task_struct *p = rq->curr; | 889 | struct task_struct *p = rq->curr; |
907 | 890 | ||
908 | /* | 891 | /* |
909 | * We make the task go to sleep until its current deadline by | 892 | * We make the task go to sleep until its current deadline by |
910 | * forcing its runtime to zero. This way, update_curr_dl() stops | 893 | * forcing its runtime to zero. This way, update_curr_dl() stops |
911 | * it and the bandwidth timer will wake it up and will give it | 894 | * it and the bandwidth timer will wake it up and will give it |
912 | * new scheduling parameters (thanks to dl_yielded=1). | 895 | * new scheduling parameters (thanks to dl_yielded=1). |
913 | */ | 896 | */ |
914 | if (p->dl.runtime > 0) { | 897 | if (p->dl.runtime > 0) { |
915 | rq->curr->dl.dl_yielded = 1; | 898 | rq->curr->dl.dl_yielded = 1; |
916 | p->dl.runtime = 0; | 899 | p->dl.runtime = 0; |
917 | } | 900 | } |
918 | update_curr_dl(rq); | 901 | update_curr_dl(rq); |
919 | } | 902 | } |
920 | 903 | ||
921 | #ifdef CONFIG_SMP | 904 | #ifdef CONFIG_SMP |
922 | 905 | ||
923 | static int find_later_rq(struct task_struct *task); | 906 | static int find_later_rq(struct task_struct *task); |
924 | 907 | ||
925 | static int | 908 | static int |
926 | select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags) | 909 | select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags) |
927 | { | 910 | { |
928 | struct task_struct *curr; | 911 | struct task_struct *curr; |
929 | struct rq *rq; | 912 | struct rq *rq; |
930 | 913 | ||
931 | if (sd_flag != SD_BALANCE_WAKE) | 914 | if (sd_flag != SD_BALANCE_WAKE) |
932 | goto out; | 915 | goto out; |
933 | 916 | ||
934 | rq = cpu_rq(cpu); | 917 | rq = cpu_rq(cpu); |
935 | 918 | ||
936 | rcu_read_lock(); | 919 | rcu_read_lock(); |
937 | curr = ACCESS_ONCE(rq->curr); /* unlocked access */ | 920 | curr = ACCESS_ONCE(rq->curr); /* unlocked access */ |
938 | 921 | ||
939 | /* | 922 | /* |
940 | * If we are dealing with a -deadline task, we must | 923 | * If we are dealing with a -deadline task, we must |
941 | * decide where to wake it up. | 924 | * decide where to wake it up. |
942 | * If it has a later deadline and the current task | 925 | * If it has a later deadline and the current task |
943 | * on this rq can't move (provided the waking task | 926 | * on this rq can't move (provided the waking task |
944 | * can!) we prefer to send it somewhere else. On the | 927 | * can!) we prefer to send it somewhere else. On the |
945 | * other hand, if it has a shorter deadline, we | 928 | * other hand, if it has a shorter deadline, we |
946 | * try to make it stay here, it might be important. | 929 | * try to make it stay here, it might be important. |
947 | */ | 930 | */ |
948 | if (unlikely(dl_task(curr)) && | 931 | if (unlikely(dl_task(curr)) && |
949 | (curr->nr_cpus_allowed < 2 || | 932 | (curr->nr_cpus_allowed < 2 || |
950 | !dl_entity_preempt(&p->dl, &curr->dl)) && | 933 | !dl_entity_preempt(&p->dl, &curr->dl)) && |
951 | (p->nr_cpus_allowed > 1)) { | 934 | (p->nr_cpus_allowed > 1)) { |
952 | int target = find_later_rq(p); | 935 | int target = find_later_rq(p); |
953 | 936 | ||
954 | if (target != -1) | 937 | if (target != -1) |
955 | cpu = target; | 938 | cpu = target; |
956 | } | 939 | } |
957 | rcu_read_unlock(); | 940 | rcu_read_unlock(); |
958 | 941 | ||
959 | out: | 942 | out: |
960 | return cpu; | 943 | return cpu; |
961 | } | 944 | } |
962 | 945 | ||
963 | static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p) | 946 | static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p) |
964 | { | 947 | { |
965 | /* | 948 | /* |
966 | * Current can't be migrated, useless to reschedule, | 949 | * Current can't be migrated, useless to reschedule, |
967 | * let's hope p can move out. | 950 | * let's hope p can move out. |
968 | */ | 951 | */ |
969 | if (rq->curr->nr_cpus_allowed == 1 || | 952 | if (rq->curr->nr_cpus_allowed == 1 || |
970 | cpudl_find(&rq->rd->cpudl, rq->curr, NULL) == -1) | 953 | cpudl_find(&rq->rd->cpudl, rq->curr, NULL) == -1) |
971 | return; | 954 | return; |
972 | 955 | ||
973 | /* | 956 | /* |
974 | * p is migratable, so let's not schedule it and | 957 | * p is migratable, so let's not schedule it and |
975 | * see if it is pushed or pulled somewhere else. | 958 | * see if it is pushed or pulled somewhere else. |
976 | */ | 959 | */ |
977 | if (p->nr_cpus_allowed != 1 && | 960 | if (p->nr_cpus_allowed != 1 && |
978 | cpudl_find(&rq->rd->cpudl, p, NULL) != -1) | 961 | cpudl_find(&rq->rd->cpudl, p, NULL) != -1) |
979 | return; | 962 | return; |
980 | 963 | ||
981 | resched_curr(rq); | 964 | resched_curr(rq); |
982 | } | 965 | } |
983 | 966 | ||
984 | static int pull_dl_task(struct rq *this_rq); | 967 | static int pull_dl_task(struct rq *this_rq); |
985 | 968 | ||
986 | #endif /* CONFIG_SMP */ | 969 | #endif /* CONFIG_SMP */ |
987 | 970 | ||
988 | /* | 971 | /* |
989 | * Only called when both the current and waking task are -deadline | 972 | * Only called when both the current and waking task are -deadline |
990 | * tasks. | 973 | * tasks. |
991 | */ | 974 | */ |
992 | static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, | 975 | static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, |
993 | int flags) | 976 | int flags) |
994 | { | 977 | { |
995 | if (dl_entity_preempt(&p->dl, &rq->curr->dl)) { | 978 | if (dl_entity_preempt(&p->dl, &rq->curr->dl)) { |
996 | resched_curr(rq); | 979 | resched_curr(rq); |
997 | return; | 980 | return; |
998 | } | 981 | } |
999 | 982 | ||
1000 | #ifdef CONFIG_SMP | 983 | #ifdef CONFIG_SMP |
1001 | /* | 984 | /* |
1002 | * In the unlikely case current and p have the same deadline | 985 | * In the unlikely case current and p have the same deadline |
1003 | * let us try to decide what's the best thing to do... | 986 | * let us try to decide what's the best thing to do... |
1004 | */ | 987 | */ |
1005 | if ((p->dl.deadline == rq->curr->dl.deadline) && | 988 | if ((p->dl.deadline == rq->curr->dl.deadline) && |
1006 | !test_tsk_need_resched(rq->curr)) | 989 | !test_tsk_need_resched(rq->curr)) |
1007 | check_preempt_equal_dl(rq, p); | 990 | check_preempt_equal_dl(rq, p); |
1008 | #endif /* CONFIG_SMP */ | 991 | #endif /* CONFIG_SMP */ |
1009 | } | 992 | } |
1010 | 993 | ||
1011 | #ifdef CONFIG_SCHED_HRTICK | 994 | #ifdef CONFIG_SCHED_HRTICK |
1012 | static void start_hrtick_dl(struct rq *rq, struct task_struct *p) | 995 | static void start_hrtick_dl(struct rq *rq, struct task_struct *p) |
1013 | { | 996 | { |
1014 | hrtick_start(rq, p->dl.runtime); | 997 | hrtick_start(rq, p->dl.runtime); |
1015 | } | 998 | } |
1016 | #else /* !CONFIG_SCHED_HRTICK */ | 999 | #else /* !CONFIG_SCHED_HRTICK */ |
1017 | static void start_hrtick_dl(struct rq *rq, struct task_struct *p) | 1000 | static void start_hrtick_dl(struct rq *rq, struct task_struct *p) |
1018 | { | 1001 | { |
1019 | } | 1002 | } |
1020 | #endif | 1003 | #endif |
1021 | 1004 | ||
1022 | static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq, | 1005 | static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq, |
1023 | struct dl_rq *dl_rq) | 1006 | struct dl_rq *dl_rq) |
1024 | { | 1007 | { |
1025 | struct rb_node *left = dl_rq->rb_leftmost; | 1008 | struct rb_node *left = dl_rq->rb_leftmost; |
1026 | 1009 | ||
1027 | if (!left) | 1010 | if (!left) |
1028 | return NULL; | 1011 | return NULL; |
1029 | 1012 | ||
1030 | return rb_entry(left, struct sched_dl_entity, rb_node); | 1013 | return rb_entry(left, struct sched_dl_entity, rb_node); |
1031 | } | 1014 | } |
1032 | 1015 | ||
1033 | struct task_struct *pick_next_task_dl(struct rq *rq, struct task_struct *prev) | 1016 | struct task_struct *pick_next_task_dl(struct rq *rq, struct task_struct *prev) |
1034 | { | 1017 | { |
1035 | struct sched_dl_entity *dl_se; | 1018 | struct sched_dl_entity *dl_se; |
1036 | struct task_struct *p; | 1019 | struct task_struct *p; |
1037 | struct dl_rq *dl_rq; | 1020 | struct dl_rq *dl_rq; |
1038 | 1021 | ||
1039 | dl_rq = &rq->dl; | 1022 | dl_rq = &rq->dl; |
1040 | 1023 | ||
1041 | if (need_pull_dl_task(rq, prev)) { | 1024 | if (need_pull_dl_task(rq, prev)) { |
1042 | pull_dl_task(rq); | 1025 | pull_dl_task(rq); |
1043 | /* | 1026 | /* |
1044 | * pull_rt_task() can drop (and re-acquire) rq->lock; this | 1027 | * pull_rt_task() can drop (and re-acquire) rq->lock; this |
1045 | * means a stop task can slip in, in which case we need to | 1028 | * means a stop task can slip in, in which case we need to |
1046 | * re-start task selection. | 1029 | * re-start task selection. |
1047 | */ | 1030 | */ |
1048 | if (rq->stop && task_on_rq_queued(rq->stop)) | 1031 | if (rq->stop && task_on_rq_queued(rq->stop)) |
1049 | return RETRY_TASK; | 1032 | return RETRY_TASK; |
1050 | } | 1033 | } |
1051 | 1034 | ||
1052 | /* | 1035 | /* |
1053 | * When prev is DL, we may throttle it in put_prev_task(). | 1036 | * When prev is DL, we may throttle it in put_prev_task(). |
1054 | * So, we update time before we check for dl_nr_running. | 1037 | * So, we update time before we check for dl_nr_running. |
1055 | */ | 1038 | */ |
1056 | if (prev->sched_class == &dl_sched_class) | 1039 | if (prev->sched_class == &dl_sched_class) |
1057 | update_curr_dl(rq); | 1040 | update_curr_dl(rq); |
1058 | 1041 | ||
1059 | if (unlikely(!dl_rq->dl_nr_running)) | 1042 | if (unlikely(!dl_rq->dl_nr_running)) |
1060 | return NULL; | 1043 | return NULL; |
1061 | 1044 | ||
1062 | put_prev_task(rq, prev); | 1045 | put_prev_task(rq, prev); |
1063 | 1046 | ||
1064 | dl_se = pick_next_dl_entity(rq, dl_rq); | 1047 | dl_se = pick_next_dl_entity(rq, dl_rq); |
1065 | BUG_ON(!dl_se); | 1048 | BUG_ON(!dl_se); |
1066 | 1049 | ||
1067 | p = dl_task_of(dl_se); | 1050 | p = dl_task_of(dl_se); |
1068 | p->se.exec_start = rq_clock_task(rq); | 1051 | p->se.exec_start = rq_clock_task(rq); |
1069 | 1052 | ||
1070 | /* Running task will never be pushed. */ | 1053 | /* Running task will never be pushed. */ |
1071 | dequeue_pushable_dl_task(rq, p); | 1054 | dequeue_pushable_dl_task(rq, p); |
1072 | 1055 | ||
1073 | if (hrtick_enabled(rq)) | 1056 | if (hrtick_enabled(rq)) |
1074 | start_hrtick_dl(rq, p); | 1057 | start_hrtick_dl(rq, p); |
1075 | 1058 | ||
1076 | set_post_schedule(rq); | 1059 | set_post_schedule(rq); |
1077 | 1060 | ||
1078 | return p; | 1061 | return p; |
1079 | } | 1062 | } |
1080 | 1063 | ||
1081 | static void put_prev_task_dl(struct rq *rq, struct task_struct *p) | 1064 | static void put_prev_task_dl(struct rq *rq, struct task_struct *p) |
1082 | { | 1065 | { |
1083 | update_curr_dl(rq); | 1066 | update_curr_dl(rq); |
1084 | 1067 | ||
1085 | if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1) | 1068 | if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1) |
1086 | enqueue_pushable_dl_task(rq, p); | 1069 | enqueue_pushable_dl_task(rq, p); |
1087 | } | 1070 | } |
1088 | 1071 | ||
1089 | static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued) | 1072 | static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued) |
1090 | { | 1073 | { |
1091 | update_curr_dl(rq); | 1074 | update_curr_dl(rq); |
1092 | 1075 | ||
1093 | if (hrtick_enabled(rq) && queued && p->dl.runtime > 0) | 1076 | if (hrtick_enabled(rq) && queued && p->dl.runtime > 0) |
1094 | start_hrtick_dl(rq, p); | 1077 | start_hrtick_dl(rq, p); |
1095 | } | 1078 | } |
1096 | 1079 | ||
1097 | static void task_fork_dl(struct task_struct *p) | 1080 | static void task_fork_dl(struct task_struct *p) |
1098 | { | 1081 | { |
1099 | /* | 1082 | /* |
1100 | * SCHED_DEADLINE tasks cannot fork and this is achieved through | 1083 | * SCHED_DEADLINE tasks cannot fork and this is achieved through |
1101 | * sched_fork() | 1084 | * sched_fork() |
1102 | */ | 1085 | */ |
1103 | } | 1086 | } |
1104 | 1087 | ||
1105 | static void task_dead_dl(struct task_struct *p) | 1088 | static void task_dead_dl(struct task_struct *p) |
1106 | { | 1089 | { |
1107 | struct hrtimer *timer = &p->dl.dl_timer; | 1090 | struct hrtimer *timer = &p->dl.dl_timer; |
1108 | struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); | 1091 | struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); |
1109 | 1092 | ||
1110 | /* | 1093 | /* |
1111 | * Since we are TASK_DEAD we won't slip out of the domain! | 1094 | * Since we are TASK_DEAD we won't slip out of the domain! |
1112 | */ | 1095 | */ |
1113 | raw_spin_lock_irq(&dl_b->lock); | 1096 | raw_spin_lock_irq(&dl_b->lock); |
1114 | dl_b->total_bw -= p->dl.dl_bw; | 1097 | dl_b->total_bw -= p->dl.dl_bw; |
1115 | raw_spin_unlock_irq(&dl_b->lock); | 1098 | raw_spin_unlock_irq(&dl_b->lock); |
1116 | 1099 | ||
1117 | hrtimer_cancel(timer); | 1100 | hrtimer_cancel(timer); |
1118 | } | 1101 | } |
1119 | 1102 | ||
1120 | static void set_curr_task_dl(struct rq *rq) | 1103 | static void set_curr_task_dl(struct rq *rq) |
1121 | { | 1104 | { |
1122 | struct task_struct *p = rq->curr; | 1105 | struct task_struct *p = rq->curr; |
1123 | 1106 | ||
1124 | p->se.exec_start = rq_clock_task(rq); | 1107 | p->se.exec_start = rq_clock_task(rq); |
1125 | 1108 | ||
1126 | /* You can't push away the running task */ | 1109 | /* You can't push away the running task */ |
1127 | dequeue_pushable_dl_task(rq, p); | 1110 | dequeue_pushable_dl_task(rq, p); |
1128 | } | 1111 | } |
1129 | 1112 | ||
1130 | #ifdef CONFIG_SMP | 1113 | #ifdef CONFIG_SMP |
1131 | 1114 | ||
1132 | /* Only try algorithms three times */ | 1115 | /* Only try algorithms three times */ |
1133 | #define DL_MAX_TRIES 3 | 1116 | #define DL_MAX_TRIES 3 |
1134 | 1117 | ||
1135 | static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu) | 1118 | static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu) |
1136 | { | 1119 | { |
1137 | if (!task_running(rq, p) && | 1120 | if (!task_running(rq, p) && |
1138 | cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) | 1121 | cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) |
1139 | return 1; | 1122 | return 1; |
1140 | return 0; | 1123 | return 0; |
1141 | } | 1124 | } |
1142 | 1125 | ||
1143 | /* Returns the second earliest -deadline task, NULL otherwise */ | 1126 | /* Returns the second earliest -deadline task, NULL otherwise */ |
1144 | static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu) | 1127 | static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu) |
1145 | { | 1128 | { |
1146 | struct rb_node *next_node = rq->dl.rb_leftmost; | 1129 | struct rb_node *next_node = rq->dl.rb_leftmost; |
1147 | struct sched_dl_entity *dl_se; | 1130 | struct sched_dl_entity *dl_se; |
1148 | struct task_struct *p = NULL; | 1131 | struct task_struct *p = NULL; |
1149 | 1132 | ||
1150 | next_node: | 1133 | next_node: |
1151 | next_node = rb_next(next_node); | 1134 | next_node = rb_next(next_node); |
1152 | if (next_node) { | 1135 | if (next_node) { |
1153 | dl_se = rb_entry(next_node, struct sched_dl_entity, rb_node); | 1136 | dl_se = rb_entry(next_node, struct sched_dl_entity, rb_node); |
1154 | p = dl_task_of(dl_se); | 1137 | p = dl_task_of(dl_se); |
1155 | 1138 | ||
1156 | if (pick_dl_task(rq, p, cpu)) | 1139 | if (pick_dl_task(rq, p, cpu)) |
1157 | return p; | 1140 | return p; |
1158 | 1141 | ||
1159 | goto next_node; | 1142 | goto next_node; |
1160 | } | 1143 | } |
1161 | 1144 | ||
1162 | return NULL; | 1145 | return NULL; |
1163 | } | 1146 | } |
1164 | 1147 | ||
1165 | static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl); | 1148 | static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl); |
1166 | 1149 | ||
1167 | static int find_later_rq(struct task_struct *task) | 1150 | static int find_later_rq(struct task_struct *task) |
1168 | { | 1151 | { |
1169 | struct sched_domain *sd; | 1152 | struct sched_domain *sd; |
1170 | struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl); | 1153 | struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl); |
1171 | int this_cpu = smp_processor_id(); | 1154 | int this_cpu = smp_processor_id(); |
1172 | int best_cpu, cpu = task_cpu(task); | 1155 | int best_cpu, cpu = task_cpu(task); |
1173 | 1156 | ||
1174 | /* Make sure the mask is initialized first */ | 1157 | /* Make sure the mask is initialized first */ |
1175 | if (unlikely(!later_mask)) | 1158 | if (unlikely(!later_mask)) |
1176 | return -1; | 1159 | return -1; |
1177 | 1160 | ||
1178 | if (task->nr_cpus_allowed == 1) | 1161 | if (task->nr_cpus_allowed == 1) |
1179 | return -1; | 1162 | return -1; |
1180 | 1163 | ||
1181 | /* | 1164 | /* |
1182 | * We have to consider system topology and task affinity | 1165 | * We have to consider system topology and task affinity |
1183 | * first, then we can look for a suitable cpu. | 1166 | * first, then we can look for a suitable cpu. |
1184 | */ | 1167 | */ |
1185 | cpumask_copy(later_mask, task_rq(task)->rd->span); | 1168 | cpumask_copy(later_mask, task_rq(task)->rd->span); |
1186 | cpumask_and(later_mask, later_mask, cpu_active_mask); | 1169 | cpumask_and(later_mask, later_mask, cpu_active_mask); |
1187 | cpumask_and(later_mask, later_mask, &task->cpus_allowed); | 1170 | cpumask_and(later_mask, later_mask, &task->cpus_allowed); |
1188 | best_cpu = cpudl_find(&task_rq(task)->rd->cpudl, | 1171 | best_cpu = cpudl_find(&task_rq(task)->rd->cpudl, |
1189 | task, later_mask); | 1172 | task, later_mask); |
1190 | if (best_cpu == -1) | 1173 | if (best_cpu == -1) |
1191 | return -1; | 1174 | return -1; |
1192 | 1175 | ||
1193 | /* | 1176 | /* |
1194 | * If we are here, some target has been found, | 1177 | * If we are here, some target has been found, |
1195 | * the most suitable of which is cached in best_cpu. | 1178 | * the most suitable of which is cached in best_cpu. |
1196 | * This is, among the runqueues where the current tasks | 1179 | * This is, among the runqueues where the current tasks |
1197 | * have later deadlines than the task's one, the rq | 1180 | * have later deadlines than the task's one, the rq |
1198 | * with the latest possible one. | 1181 | * with the latest possible one. |
1199 | * | 1182 | * |
1200 | * Now we check how well this matches with task's | 1183 | * Now we check how well this matches with task's |
1201 | * affinity and system topology. | 1184 | * affinity and system topology. |
1202 | * | 1185 | * |
1203 | * The last cpu where the task run is our first | 1186 | * The last cpu where the task run is our first |
1204 | * guess, since it is most likely cache-hot there. | 1187 | * guess, since it is most likely cache-hot there. |
1205 | */ | 1188 | */ |
1206 | if (cpumask_test_cpu(cpu, later_mask)) | 1189 | if (cpumask_test_cpu(cpu, later_mask)) |
1207 | return cpu; | 1190 | return cpu; |
1208 | /* | 1191 | /* |
1209 | * Check if this_cpu is to be skipped (i.e., it is | 1192 | * Check if this_cpu is to be skipped (i.e., it is |
1210 | * not in the mask) or not. | 1193 | * not in the mask) or not. |
1211 | */ | 1194 | */ |
1212 | if (!cpumask_test_cpu(this_cpu, later_mask)) | 1195 | if (!cpumask_test_cpu(this_cpu, later_mask)) |
1213 | this_cpu = -1; | 1196 | this_cpu = -1; |
1214 | 1197 | ||
1215 | rcu_read_lock(); | 1198 | rcu_read_lock(); |
1216 | for_each_domain(cpu, sd) { | 1199 | for_each_domain(cpu, sd) { |
1217 | if (sd->flags & SD_WAKE_AFFINE) { | 1200 | if (sd->flags & SD_WAKE_AFFINE) { |
1218 | 1201 | ||
1219 | /* | 1202 | /* |
1220 | * If possible, preempting this_cpu is | 1203 | * If possible, preempting this_cpu is |
1221 | * cheaper than migrating. | 1204 | * cheaper than migrating. |
1222 | */ | 1205 | */ |
1223 | if (this_cpu != -1 && | 1206 | if (this_cpu != -1 && |
1224 | cpumask_test_cpu(this_cpu, sched_domain_span(sd))) { | 1207 | cpumask_test_cpu(this_cpu, sched_domain_span(sd))) { |
1225 | rcu_read_unlock(); | 1208 | rcu_read_unlock(); |
1226 | return this_cpu; | 1209 | return this_cpu; |
1227 | } | 1210 | } |
1228 | 1211 | ||
1229 | /* | 1212 | /* |
1230 | * Last chance: if best_cpu is valid and is | 1213 | * Last chance: if best_cpu is valid and is |
1231 | * in the mask, that becomes our choice. | 1214 | * in the mask, that becomes our choice. |
1232 | */ | 1215 | */ |
1233 | if (best_cpu < nr_cpu_ids && | 1216 | if (best_cpu < nr_cpu_ids && |
1234 | cpumask_test_cpu(best_cpu, sched_domain_span(sd))) { | 1217 | cpumask_test_cpu(best_cpu, sched_domain_span(sd))) { |
1235 | rcu_read_unlock(); | 1218 | rcu_read_unlock(); |
1236 | return best_cpu; | 1219 | return best_cpu; |
1237 | } | 1220 | } |
1238 | } | 1221 | } |
1239 | } | 1222 | } |
1240 | rcu_read_unlock(); | 1223 | rcu_read_unlock(); |
1241 | 1224 | ||
1242 | /* | 1225 | /* |
1243 | * At this point, all our guesses failed, we just return | 1226 | * At this point, all our guesses failed, we just return |
1244 | * 'something', and let the caller sort the things out. | 1227 | * 'something', and let the caller sort the things out. |
1245 | */ | 1228 | */ |
1246 | if (this_cpu != -1) | 1229 | if (this_cpu != -1) |
1247 | return this_cpu; | 1230 | return this_cpu; |
1248 | 1231 | ||
1249 | cpu = cpumask_any(later_mask); | 1232 | cpu = cpumask_any(later_mask); |
1250 | if (cpu < nr_cpu_ids) | 1233 | if (cpu < nr_cpu_ids) |
1251 | return cpu; | 1234 | return cpu; |
1252 | 1235 | ||
1253 | return -1; | 1236 | return -1; |
1254 | } | 1237 | } |
1255 | 1238 | ||
1256 | /* Locks the rq it finds */ | 1239 | /* Locks the rq it finds */ |
1257 | static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq) | 1240 | static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq) |
1258 | { | 1241 | { |
1259 | struct rq *later_rq = NULL; | 1242 | struct rq *later_rq = NULL; |
1260 | int tries; | 1243 | int tries; |
1261 | int cpu; | 1244 | int cpu; |
1262 | 1245 | ||
1263 | for (tries = 0; tries < DL_MAX_TRIES; tries++) { | 1246 | for (tries = 0; tries < DL_MAX_TRIES; tries++) { |
1264 | cpu = find_later_rq(task); | 1247 | cpu = find_later_rq(task); |
1265 | 1248 | ||
1266 | if ((cpu == -1) || (cpu == rq->cpu)) | 1249 | if ((cpu == -1) || (cpu == rq->cpu)) |
1267 | break; | 1250 | break; |
1268 | 1251 | ||
1269 | later_rq = cpu_rq(cpu); | 1252 | later_rq = cpu_rq(cpu); |
1270 | 1253 | ||
1271 | /* Retry if something changed. */ | 1254 | /* Retry if something changed. */ |
1272 | if (double_lock_balance(rq, later_rq)) { | 1255 | if (double_lock_balance(rq, later_rq)) { |
1273 | if (unlikely(task_rq(task) != rq || | 1256 | if (unlikely(task_rq(task) != rq || |
1274 | !cpumask_test_cpu(later_rq->cpu, | 1257 | !cpumask_test_cpu(later_rq->cpu, |
1275 | &task->cpus_allowed) || | 1258 | &task->cpus_allowed) || |
1276 | task_running(rq, task) || | 1259 | task_running(rq, task) || |
1277 | !task_on_rq_queued(task))) { | 1260 | !task_on_rq_queued(task))) { |
1278 | double_unlock_balance(rq, later_rq); | 1261 | double_unlock_balance(rq, later_rq); |
1279 | later_rq = NULL; | 1262 | later_rq = NULL; |
1280 | break; | 1263 | break; |
1281 | } | 1264 | } |
1282 | } | 1265 | } |
1283 | 1266 | ||
1284 | /* | 1267 | /* |
1285 | * If the rq we found has no -deadline task, or | 1268 | * If the rq we found has no -deadline task, or |
1286 | * its earliest one has a later deadline than our | 1269 | * its earliest one has a later deadline than our |
1287 | * task, the rq is a good one. | 1270 | * task, the rq is a good one. |
1288 | */ | 1271 | */ |
1289 | if (!later_rq->dl.dl_nr_running || | 1272 | if (!later_rq->dl.dl_nr_running || |
1290 | dl_time_before(task->dl.deadline, | 1273 | dl_time_before(task->dl.deadline, |
1291 | later_rq->dl.earliest_dl.curr)) | 1274 | later_rq->dl.earliest_dl.curr)) |
1292 | break; | 1275 | break; |
1293 | 1276 | ||
1294 | /* Otherwise we try again. */ | 1277 | /* Otherwise we try again. */ |
1295 | double_unlock_balance(rq, later_rq); | 1278 | double_unlock_balance(rq, later_rq); |
1296 | later_rq = NULL; | 1279 | later_rq = NULL; |
1297 | } | 1280 | } |
1298 | 1281 | ||
1299 | return later_rq; | 1282 | return later_rq; |
1300 | } | 1283 | } |
1301 | 1284 | ||
1302 | static struct task_struct *pick_next_pushable_dl_task(struct rq *rq) | 1285 | static struct task_struct *pick_next_pushable_dl_task(struct rq *rq) |
1303 | { | 1286 | { |
1304 | struct task_struct *p; | 1287 | struct task_struct *p; |
1305 | 1288 | ||
1306 | if (!has_pushable_dl_tasks(rq)) | 1289 | if (!has_pushable_dl_tasks(rq)) |
1307 | return NULL; | 1290 | return NULL; |
1308 | 1291 | ||
1309 | p = rb_entry(rq->dl.pushable_dl_tasks_leftmost, | 1292 | p = rb_entry(rq->dl.pushable_dl_tasks_leftmost, |
1310 | struct task_struct, pushable_dl_tasks); | 1293 | struct task_struct, pushable_dl_tasks); |
1311 | 1294 | ||
1312 | BUG_ON(rq->cpu != task_cpu(p)); | 1295 | BUG_ON(rq->cpu != task_cpu(p)); |
1313 | BUG_ON(task_current(rq, p)); | 1296 | BUG_ON(task_current(rq, p)); |
1314 | BUG_ON(p->nr_cpus_allowed <= 1); | 1297 | BUG_ON(p->nr_cpus_allowed <= 1); |
1315 | 1298 | ||
1316 | BUG_ON(!task_on_rq_queued(p)); | 1299 | BUG_ON(!task_on_rq_queued(p)); |
1317 | BUG_ON(!dl_task(p)); | 1300 | BUG_ON(!dl_task(p)); |
1318 | 1301 | ||
1319 | return p; | 1302 | return p; |
1320 | } | 1303 | } |
1321 | 1304 | ||
1322 | /* | 1305 | /* |
1323 | * See if the non running -deadline tasks on this rq | 1306 | * See if the non running -deadline tasks on this rq |
1324 | * can be sent to some other CPU where they can preempt | 1307 | * can be sent to some other CPU where they can preempt |
1325 | * and start executing. | 1308 | * and start executing. |
1326 | */ | 1309 | */ |
1327 | static int push_dl_task(struct rq *rq) | 1310 | static int push_dl_task(struct rq *rq) |
1328 | { | 1311 | { |
1329 | struct task_struct *next_task; | 1312 | struct task_struct *next_task; |
1330 | struct rq *later_rq; | 1313 | struct rq *later_rq; |
1331 | int ret = 0; | 1314 | int ret = 0; |
1332 | 1315 | ||
1333 | if (!rq->dl.overloaded) | 1316 | if (!rq->dl.overloaded) |
1334 | return 0; | 1317 | return 0; |
1335 | 1318 | ||
1336 | next_task = pick_next_pushable_dl_task(rq); | 1319 | next_task = pick_next_pushable_dl_task(rq); |
1337 | if (!next_task) | 1320 | if (!next_task) |
1338 | return 0; | 1321 | return 0; |
1339 | 1322 | ||
1340 | retry: | 1323 | retry: |
1341 | if (unlikely(next_task == rq->curr)) { | 1324 | if (unlikely(next_task == rq->curr)) { |
1342 | WARN_ON(1); | 1325 | WARN_ON(1); |
1343 | return 0; | 1326 | return 0; |
1344 | } | 1327 | } |
1345 | 1328 | ||
1346 | /* | 1329 | /* |
1347 | * If next_task preempts rq->curr, and rq->curr | 1330 | * If next_task preempts rq->curr, and rq->curr |
1348 | * can move away, it makes sense to just reschedule | 1331 | * can move away, it makes sense to just reschedule |
1349 | * without going further in pushing next_task. | 1332 | * without going further in pushing next_task. |
1350 | */ | 1333 | */ |
1351 | if (dl_task(rq->curr) && | 1334 | if (dl_task(rq->curr) && |
1352 | dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) && | 1335 | dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) && |
1353 | rq->curr->nr_cpus_allowed > 1) { | 1336 | rq->curr->nr_cpus_allowed > 1) { |
1354 | resched_curr(rq); | 1337 | resched_curr(rq); |
1355 | return 0; | 1338 | return 0; |
1356 | } | 1339 | } |
1357 | 1340 | ||
1358 | /* We might release rq lock */ | 1341 | /* We might release rq lock */ |
1359 | get_task_struct(next_task); | 1342 | get_task_struct(next_task); |
1360 | 1343 | ||
1361 | /* Will lock the rq it'll find */ | 1344 | /* Will lock the rq it'll find */ |
1362 | later_rq = find_lock_later_rq(next_task, rq); | 1345 | later_rq = find_lock_later_rq(next_task, rq); |
1363 | if (!later_rq) { | 1346 | if (!later_rq) { |
1364 | struct task_struct *task; | 1347 | struct task_struct *task; |
1365 | 1348 | ||
1366 | /* | 1349 | /* |
1367 | * We must check all this again, since | 1350 | * We must check all this again, since |
1368 | * find_lock_later_rq releases rq->lock and it is | 1351 | * find_lock_later_rq releases rq->lock and it is |
1369 | * then possible that next_task has migrated. | 1352 | * then possible that next_task has migrated. |
1370 | */ | 1353 | */ |
1371 | task = pick_next_pushable_dl_task(rq); | 1354 | task = pick_next_pushable_dl_task(rq); |
1372 | if (task_cpu(next_task) == rq->cpu && task == next_task) { | 1355 | if (task_cpu(next_task) == rq->cpu && task == next_task) { |
1373 | /* | 1356 | /* |
1374 | * The task is still there. We don't try | 1357 | * The task is still there. We don't try |
1375 | * again, some other cpu will pull it when ready. | 1358 | * again, some other cpu will pull it when ready. |
1376 | */ | 1359 | */ |
1377 | goto out; | 1360 | goto out; |
1378 | } | 1361 | } |
1379 | 1362 | ||
1380 | if (!task) | 1363 | if (!task) |
1381 | /* No more tasks */ | 1364 | /* No more tasks */ |
1382 | goto out; | 1365 | goto out; |
1383 | 1366 | ||
1384 | put_task_struct(next_task); | 1367 | put_task_struct(next_task); |
1385 | next_task = task; | 1368 | next_task = task; |
1386 | goto retry; | 1369 | goto retry; |
1387 | } | 1370 | } |
1388 | 1371 | ||
1389 | deactivate_task(rq, next_task, 0); | 1372 | deactivate_task(rq, next_task, 0); |
1390 | set_task_cpu(next_task, later_rq->cpu); | 1373 | set_task_cpu(next_task, later_rq->cpu); |
1391 | activate_task(later_rq, next_task, 0); | 1374 | activate_task(later_rq, next_task, 0); |
1392 | ret = 1; | 1375 | ret = 1; |
1393 | 1376 | ||
1394 | resched_curr(later_rq); | 1377 | resched_curr(later_rq); |
1395 | 1378 | ||
1396 | double_unlock_balance(rq, later_rq); | 1379 | double_unlock_balance(rq, later_rq); |
1397 | 1380 | ||
1398 | out: | 1381 | out: |
1399 | put_task_struct(next_task); | 1382 | put_task_struct(next_task); |
1400 | 1383 | ||
1401 | return ret; | 1384 | return ret; |
1402 | } | 1385 | } |
1403 | 1386 | ||
1404 | static void push_dl_tasks(struct rq *rq) | 1387 | static void push_dl_tasks(struct rq *rq) |
1405 | { | 1388 | { |
1406 | /* Terminates as it moves a -deadline task */ | 1389 | /* Terminates as it moves a -deadline task */ |
1407 | while (push_dl_task(rq)) | 1390 | while (push_dl_task(rq)) |
1408 | ; | 1391 | ; |
1409 | } | 1392 | } |
1410 | 1393 | ||
1411 | static int pull_dl_task(struct rq *this_rq) | 1394 | static int pull_dl_task(struct rq *this_rq) |
1412 | { | 1395 | { |
1413 | int this_cpu = this_rq->cpu, ret = 0, cpu; | 1396 | int this_cpu = this_rq->cpu, ret = 0, cpu; |
1414 | struct task_struct *p; | 1397 | struct task_struct *p; |
1415 | struct rq *src_rq; | 1398 | struct rq *src_rq; |
1416 | u64 dmin = LONG_MAX; | 1399 | u64 dmin = LONG_MAX; |
1417 | 1400 | ||
1418 | if (likely(!dl_overloaded(this_rq))) | 1401 | if (likely(!dl_overloaded(this_rq))) |
1419 | return 0; | 1402 | return 0; |
1420 | 1403 | ||
1421 | /* | 1404 | /* |
1422 | * Match the barrier from dl_set_overloaded; this guarantees that if we | 1405 | * Match the barrier from dl_set_overloaded; this guarantees that if we |
1423 | * see overloaded we must also see the dlo_mask bit. | 1406 | * see overloaded we must also see the dlo_mask bit. |
1424 | */ | 1407 | */ |
1425 | smp_rmb(); | 1408 | smp_rmb(); |
1426 | 1409 | ||
1427 | for_each_cpu(cpu, this_rq->rd->dlo_mask) { | 1410 | for_each_cpu(cpu, this_rq->rd->dlo_mask) { |
1428 | if (this_cpu == cpu) | 1411 | if (this_cpu == cpu) |
1429 | continue; | 1412 | continue; |
1430 | 1413 | ||
1431 | src_rq = cpu_rq(cpu); | 1414 | src_rq = cpu_rq(cpu); |
1432 | 1415 | ||
1433 | /* | 1416 | /* |
1434 | * It looks racy, abd it is! However, as in sched_rt.c, | 1417 | * It looks racy, abd it is! However, as in sched_rt.c, |
1435 | * we are fine with this. | 1418 | * we are fine with this. |
1436 | */ | 1419 | */ |
1437 | if (this_rq->dl.dl_nr_running && | 1420 | if (this_rq->dl.dl_nr_running && |
1438 | dl_time_before(this_rq->dl.earliest_dl.curr, | 1421 | dl_time_before(this_rq->dl.earliest_dl.curr, |
1439 | src_rq->dl.earliest_dl.next)) | 1422 | src_rq->dl.earliest_dl.next)) |
1440 | continue; | 1423 | continue; |
1441 | 1424 | ||
1442 | /* Might drop this_rq->lock */ | 1425 | /* Might drop this_rq->lock */ |
1443 | double_lock_balance(this_rq, src_rq); | 1426 | double_lock_balance(this_rq, src_rq); |
1444 | 1427 | ||
1445 | /* | 1428 | /* |
1446 | * If there are no more pullable tasks on the | 1429 | * If there are no more pullable tasks on the |
1447 | * rq, we're done with it. | 1430 | * rq, we're done with it. |
1448 | */ | 1431 | */ |
1449 | if (src_rq->dl.dl_nr_running <= 1) | 1432 | if (src_rq->dl.dl_nr_running <= 1) |
1450 | goto skip; | 1433 | goto skip; |
1451 | 1434 | ||
1452 | p = pick_next_earliest_dl_task(src_rq, this_cpu); | 1435 | p = pick_next_earliest_dl_task(src_rq, this_cpu); |
1453 | 1436 | ||
1454 | /* | 1437 | /* |
1455 | * We found a task to be pulled if: | 1438 | * We found a task to be pulled if: |
1456 | * - it preempts our current (if there's one), | 1439 | * - it preempts our current (if there's one), |
1457 | * - it will preempt the last one we pulled (if any). | 1440 | * - it will preempt the last one we pulled (if any). |
1458 | */ | 1441 | */ |
1459 | if (p && dl_time_before(p->dl.deadline, dmin) && | 1442 | if (p && dl_time_before(p->dl.deadline, dmin) && |
1460 | (!this_rq->dl.dl_nr_running || | 1443 | (!this_rq->dl.dl_nr_running || |
1461 | dl_time_before(p->dl.deadline, | 1444 | dl_time_before(p->dl.deadline, |
1462 | this_rq->dl.earliest_dl.curr))) { | 1445 | this_rq->dl.earliest_dl.curr))) { |
1463 | WARN_ON(p == src_rq->curr); | 1446 | WARN_ON(p == src_rq->curr); |
1464 | WARN_ON(!task_on_rq_queued(p)); | 1447 | WARN_ON(!task_on_rq_queued(p)); |
1465 | 1448 | ||
1466 | /* | 1449 | /* |
1467 | * Then we pull iff p has actually an earlier | 1450 | * Then we pull iff p has actually an earlier |
1468 | * deadline than the current task of its runqueue. | 1451 | * deadline than the current task of its runqueue. |
1469 | */ | 1452 | */ |
1470 | if (dl_time_before(p->dl.deadline, | 1453 | if (dl_time_before(p->dl.deadline, |
1471 | src_rq->curr->dl.deadline)) | 1454 | src_rq->curr->dl.deadline)) |
1472 | goto skip; | 1455 | goto skip; |
1473 | 1456 | ||
1474 | ret = 1; | 1457 | ret = 1; |
1475 | 1458 | ||
1476 | deactivate_task(src_rq, p, 0); | 1459 | deactivate_task(src_rq, p, 0); |
1477 | set_task_cpu(p, this_cpu); | 1460 | set_task_cpu(p, this_cpu); |
1478 | activate_task(this_rq, p, 0); | 1461 | activate_task(this_rq, p, 0); |
1479 | dmin = p->dl.deadline; | 1462 | dmin = p->dl.deadline; |
1480 | 1463 | ||
1481 | /* Is there any other task even earlier? */ | 1464 | /* Is there any other task even earlier? */ |
1482 | } | 1465 | } |
1483 | skip: | 1466 | skip: |
1484 | double_unlock_balance(this_rq, src_rq); | 1467 | double_unlock_balance(this_rq, src_rq); |
1485 | } | 1468 | } |
1486 | 1469 | ||
1487 | return ret; | 1470 | return ret; |
1488 | } | 1471 | } |
1489 | 1472 | ||
1490 | static void post_schedule_dl(struct rq *rq) | 1473 | static void post_schedule_dl(struct rq *rq) |
1491 | { | 1474 | { |
1492 | push_dl_tasks(rq); | 1475 | push_dl_tasks(rq); |
1493 | } | 1476 | } |
1494 | 1477 | ||
1495 | /* | 1478 | /* |
1496 | * Since the task is not running and a reschedule is not going to happen | 1479 | * Since the task is not running and a reschedule is not going to happen |
1497 | * anytime soon on its runqueue, we try pushing it away now. | 1480 | * anytime soon on its runqueue, we try pushing it away now. |
1498 | */ | 1481 | */ |
1499 | static void task_woken_dl(struct rq *rq, struct task_struct *p) | 1482 | static void task_woken_dl(struct rq *rq, struct task_struct *p) |
1500 | { | 1483 | { |
1501 | if (!task_running(rq, p) && | 1484 | if (!task_running(rq, p) && |
1502 | !test_tsk_need_resched(rq->curr) && | 1485 | !test_tsk_need_resched(rq->curr) && |
1503 | has_pushable_dl_tasks(rq) && | 1486 | has_pushable_dl_tasks(rq) && |
1504 | p->nr_cpus_allowed > 1 && | 1487 | p->nr_cpus_allowed > 1 && |
1505 | dl_task(rq->curr) && | 1488 | dl_task(rq->curr) && |
1506 | (rq->curr->nr_cpus_allowed < 2 || | 1489 | (rq->curr->nr_cpus_allowed < 2 || |
1507 | !dl_entity_preempt(&p->dl, &rq->curr->dl))) { | 1490 | !dl_entity_preempt(&p->dl, &rq->curr->dl))) { |
1508 | push_dl_tasks(rq); | 1491 | push_dl_tasks(rq); |
1509 | } | 1492 | } |
1510 | } | 1493 | } |
1511 | 1494 | ||
1512 | static void set_cpus_allowed_dl(struct task_struct *p, | 1495 | static void set_cpus_allowed_dl(struct task_struct *p, |
1513 | const struct cpumask *new_mask) | 1496 | const struct cpumask *new_mask) |
1514 | { | 1497 | { |
1515 | struct rq *rq; | 1498 | struct rq *rq; |
1516 | struct root_domain *src_rd; | 1499 | struct root_domain *src_rd; |
1517 | int weight; | 1500 | int weight; |
1518 | 1501 | ||
1519 | BUG_ON(!dl_task(p)); | 1502 | BUG_ON(!dl_task(p)); |
1520 | 1503 | ||
1521 | rq = task_rq(p); | 1504 | rq = task_rq(p); |
1522 | src_rd = rq->rd; | 1505 | src_rd = rq->rd; |
1523 | /* | 1506 | /* |
1524 | * Migrating a SCHED_DEADLINE task between exclusive | 1507 | * Migrating a SCHED_DEADLINE task between exclusive |
1525 | * cpusets (different root_domains) entails a bandwidth | 1508 | * cpusets (different root_domains) entails a bandwidth |
1526 | * update. We already made space for us in the destination | 1509 | * update. We already made space for us in the destination |
1527 | * domain (see cpuset_can_attach()). | 1510 | * domain (see cpuset_can_attach()). |
1528 | */ | 1511 | */ |
1529 | if (!cpumask_intersects(src_rd->span, new_mask)) { | 1512 | if (!cpumask_intersects(src_rd->span, new_mask)) { |
1530 | struct dl_bw *src_dl_b; | 1513 | struct dl_bw *src_dl_b; |
1531 | 1514 | ||
1532 | src_dl_b = dl_bw_of(cpu_of(rq)); | 1515 | src_dl_b = dl_bw_of(cpu_of(rq)); |
1533 | /* | 1516 | /* |
1534 | * We now free resources of the root_domain we are migrating | 1517 | * We now free resources of the root_domain we are migrating |
1535 | * off. In the worst case, sched_setattr() may temporary fail | 1518 | * off. In the worst case, sched_setattr() may temporary fail |
1536 | * until we complete the update. | 1519 | * until we complete the update. |
1537 | */ | 1520 | */ |
1538 | raw_spin_lock(&src_dl_b->lock); | 1521 | raw_spin_lock(&src_dl_b->lock); |
1539 | __dl_clear(src_dl_b, p->dl.dl_bw); | 1522 | __dl_clear(src_dl_b, p->dl.dl_bw); |
1540 | raw_spin_unlock(&src_dl_b->lock); | 1523 | raw_spin_unlock(&src_dl_b->lock); |
1541 | } | 1524 | } |
1542 | 1525 | ||
1543 | /* | 1526 | /* |
1544 | * Update only if the task is actually running (i.e., | 1527 | * Update only if the task is actually running (i.e., |
1545 | * it is on the rq AND it is not throttled). | 1528 | * it is on the rq AND it is not throttled). |
1546 | */ | 1529 | */ |
1547 | if (!on_dl_rq(&p->dl)) | 1530 | if (!on_dl_rq(&p->dl)) |
1548 | return; | 1531 | return; |
1549 | 1532 | ||
1550 | weight = cpumask_weight(new_mask); | 1533 | weight = cpumask_weight(new_mask); |
1551 | 1534 | ||
1552 | /* | 1535 | /* |
1553 | * Only update if the process changes its state from whether it | 1536 | * Only update if the process changes its state from whether it |
1554 | * can migrate or not. | 1537 | * can migrate or not. |
1555 | */ | 1538 | */ |
1556 | if ((p->nr_cpus_allowed > 1) == (weight > 1)) | 1539 | if ((p->nr_cpus_allowed > 1) == (weight > 1)) |
1557 | return; | 1540 | return; |
1558 | 1541 | ||
1559 | /* | 1542 | /* |
1560 | * The process used to be able to migrate OR it can now migrate | 1543 | * The process used to be able to migrate OR it can now migrate |
1561 | */ | 1544 | */ |
1562 | if (weight <= 1) { | 1545 | if (weight <= 1) { |
1563 | if (!task_current(rq, p)) | 1546 | if (!task_current(rq, p)) |
1564 | dequeue_pushable_dl_task(rq, p); | 1547 | dequeue_pushable_dl_task(rq, p); |
1565 | BUG_ON(!rq->dl.dl_nr_migratory); | 1548 | BUG_ON(!rq->dl.dl_nr_migratory); |
1566 | rq->dl.dl_nr_migratory--; | 1549 | rq->dl.dl_nr_migratory--; |
1567 | } else { | 1550 | } else { |
1568 | if (!task_current(rq, p)) | 1551 | if (!task_current(rq, p)) |
1569 | enqueue_pushable_dl_task(rq, p); | 1552 | enqueue_pushable_dl_task(rq, p); |
1570 | rq->dl.dl_nr_migratory++; | 1553 | rq->dl.dl_nr_migratory++; |
1571 | } | 1554 | } |
1572 | 1555 | ||
1573 | update_dl_migration(&rq->dl); | 1556 | update_dl_migration(&rq->dl); |
1574 | } | 1557 | } |
1575 | 1558 | ||
1576 | /* Assumes rq->lock is held */ | 1559 | /* Assumes rq->lock is held */ |
1577 | static void rq_online_dl(struct rq *rq) | 1560 | static void rq_online_dl(struct rq *rq) |
1578 | { | 1561 | { |
1579 | if (rq->dl.overloaded) | 1562 | if (rq->dl.overloaded) |
1580 | dl_set_overload(rq); | 1563 | dl_set_overload(rq); |
1581 | 1564 | ||
1582 | if (rq->dl.dl_nr_running > 0) | 1565 | if (rq->dl.dl_nr_running > 0) |
1583 | cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr, 1); | 1566 | cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr, 1); |
1584 | } | 1567 | } |
1585 | 1568 | ||
1586 | /* Assumes rq->lock is held */ | 1569 | /* Assumes rq->lock is held */ |
1587 | static void rq_offline_dl(struct rq *rq) | 1570 | static void rq_offline_dl(struct rq *rq) |
1588 | { | 1571 | { |
1589 | if (rq->dl.overloaded) | 1572 | if (rq->dl.overloaded) |
1590 | dl_clear_overload(rq); | 1573 | dl_clear_overload(rq); |
1591 | 1574 | ||
1592 | cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0); | 1575 | cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0); |
1593 | } | 1576 | } |
1594 | 1577 | ||
1595 | void init_sched_dl_class(void) | 1578 | void init_sched_dl_class(void) |
1596 | { | 1579 | { |
1597 | unsigned int i; | 1580 | unsigned int i; |
1598 | 1581 | ||
1599 | for_each_possible_cpu(i) | 1582 | for_each_possible_cpu(i) |
1600 | zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i), | 1583 | zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i), |
1601 | GFP_KERNEL, cpu_to_node(i)); | 1584 | GFP_KERNEL, cpu_to_node(i)); |
1602 | } | 1585 | } |
1603 | 1586 | ||
1604 | #endif /* CONFIG_SMP */ | 1587 | #endif /* CONFIG_SMP */ |
1605 | 1588 | ||
1606 | /* | 1589 | /* |
1607 | * Ensure p's dl_timer is cancelled. May drop rq->lock for a while. | 1590 | * Ensure p's dl_timer is cancelled. May drop rq->lock for a while. |
1608 | */ | 1591 | */ |
1609 | static void cancel_dl_timer(struct rq *rq, struct task_struct *p) | 1592 | static void cancel_dl_timer(struct rq *rq, struct task_struct *p) |
1610 | { | 1593 | { |
1611 | struct hrtimer *dl_timer = &p->dl.dl_timer; | 1594 | struct hrtimer *dl_timer = &p->dl.dl_timer; |
1612 | 1595 | ||
1613 | /* Nobody will change task's class if pi_lock is held */ | 1596 | /* Nobody will change task's class if pi_lock is held */ |
1614 | lockdep_assert_held(&p->pi_lock); | 1597 | lockdep_assert_held(&p->pi_lock); |
1615 | 1598 | ||
1616 | if (hrtimer_active(dl_timer)) { | 1599 | if (hrtimer_active(dl_timer)) { |
1617 | int ret = hrtimer_try_to_cancel(dl_timer); | 1600 | int ret = hrtimer_try_to_cancel(dl_timer); |
1618 | 1601 | ||
1619 | if (unlikely(ret == -1)) { | 1602 | if (unlikely(ret == -1)) { |
1620 | /* | 1603 | /* |
1621 | * Note, p may migrate OR new deadline tasks | 1604 | * Note, p may migrate OR new deadline tasks |
1622 | * may appear in rq when we are unlocking it. | 1605 | * may appear in rq when we are unlocking it. |
1623 | * A caller of us must be fine with that. | 1606 | * A caller of us must be fine with that. |
1624 | */ | 1607 | */ |
1625 | raw_spin_unlock(&rq->lock); | 1608 | raw_spin_unlock(&rq->lock); |
1626 | hrtimer_cancel(dl_timer); | 1609 | hrtimer_cancel(dl_timer); |
1627 | raw_spin_lock(&rq->lock); | 1610 | raw_spin_lock(&rq->lock); |
1628 | } | 1611 | } |
1629 | } | 1612 | } |
1630 | } | 1613 | } |
1631 | 1614 | ||
1632 | static void switched_from_dl(struct rq *rq, struct task_struct *p) | 1615 | static void switched_from_dl(struct rq *rq, struct task_struct *p) |
1633 | { | 1616 | { |
1634 | cancel_dl_timer(rq, p); | 1617 | cancel_dl_timer(rq, p); |
1635 | 1618 | ||
1636 | __dl_clear_params(p); | 1619 | __dl_clear_params(p); |
1637 | 1620 | ||
1638 | /* | 1621 | /* |
1639 | * Since this might be the only -deadline task on the rq, | 1622 | * Since this might be the only -deadline task on the rq, |
1640 | * this is the right place to try to pull some other one | 1623 | * this is the right place to try to pull some other one |
1641 | * from an overloaded cpu, if any. | 1624 | * from an overloaded cpu, if any. |
1642 | */ | 1625 | */ |
1643 | if (!task_on_rq_queued(p) || rq->dl.dl_nr_running) | 1626 | if (!task_on_rq_queued(p) || rq->dl.dl_nr_running) |
1644 | return; | 1627 | return; |
1645 | 1628 | ||
1646 | if (pull_dl_task(rq)) | 1629 | if (pull_dl_task(rq)) |
1647 | resched_curr(rq); | 1630 | resched_curr(rq); |
1648 | } | 1631 | } |
1649 | 1632 | ||
1650 | /* | 1633 | /* |
1651 | * When switching to -deadline, we may overload the rq, then | 1634 | * When switching to -deadline, we may overload the rq, then |
1652 | * we try to push someone off, if possible. | 1635 | * we try to push someone off, if possible. |
1653 | */ | 1636 | */ |
1654 | static void switched_to_dl(struct rq *rq, struct task_struct *p) | 1637 | static void switched_to_dl(struct rq *rq, struct task_struct *p) |
1655 | { | 1638 | { |
1656 | int check_resched = 1; | 1639 | int check_resched = 1; |
1657 | 1640 | ||
1658 | /* | 1641 | /* |
1659 | * If p is throttled, don't consider the possibility | 1642 | * If p is throttled, don't consider the possibility |
1660 | * of preempting rq->curr, the check will be done right | 1643 | * of preempting rq->curr, the check will be done right |
1661 | * after its runtime will get replenished. | 1644 | * after its runtime will get replenished. |
1662 | */ | 1645 | */ |
1663 | if (unlikely(p->dl.dl_throttled)) | 1646 | if (unlikely(p->dl.dl_throttled)) |
1664 | return; | 1647 | return; |
1665 | 1648 | ||
1666 | if (task_on_rq_queued(p) && rq->curr != p) { | 1649 | if (task_on_rq_queued(p) && rq->curr != p) { |
1667 | #ifdef CONFIG_SMP | 1650 | #ifdef CONFIG_SMP |
1668 | if (p->nr_cpus_allowed > 1 && rq->dl.overloaded && | 1651 | if (p->nr_cpus_allowed > 1 && rq->dl.overloaded && |
1669 | push_dl_task(rq) && rq != task_rq(p)) | 1652 | push_dl_task(rq) && rq != task_rq(p)) |
1670 | /* Only reschedule if pushing failed */ | 1653 | /* Only reschedule if pushing failed */ |
1671 | check_resched = 0; | 1654 | check_resched = 0; |
1672 | #endif /* CONFIG_SMP */ | 1655 | #endif /* CONFIG_SMP */ |
1673 | if (check_resched) { | 1656 | if (check_resched) { |
1674 | if (dl_task(rq->curr)) | 1657 | if (dl_task(rq->curr)) |
1675 | check_preempt_curr_dl(rq, p, 0); | 1658 | check_preempt_curr_dl(rq, p, 0); |
1676 | else | 1659 | else |
1677 | resched_curr(rq); | 1660 | resched_curr(rq); |
1678 | } | 1661 | } |
1679 | } | 1662 | } |
1680 | } | 1663 | } |
1681 | 1664 | ||
1682 | /* | 1665 | /* |
1683 | * If the scheduling parameters of a -deadline task changed, | 1666 | * If the scheduling parameters of a -deadline task changed, |
1684 | * a push or pull operation might be needed. | 1667 | * a push or pull operation might be needed. |
1685 | */ | 1668 | */ |
1686 | static void prio_changed_dl(struct rq *rq, struct task_struct *p, | 1669 | static void prio_changed_dl(struct rq *rq, struct task_struct *p, |
1687 | int oldprio) | 1670 | int oldprio) |
1688 | { | 1671 | { |
1689 | if (task_on_rq_queued(p) || rq->curr == p) { | 1672 | if (task_on_rq_queued(p) || rq->curr == p) { |
1690 | #ifdef CONFIG_SMP | 1673 | #ifdef CONFIG_SMP |
1691 | /* | 1674 | /* |
1692 | * This might be too much, but unfortunately | 1675 | * This might be too much, but unfortunately |
1693 | * we don't have the old deadline value, and | 1676 | * we don't have the old deadline value, and |
1694 | * we can't argue if the task is increasing | 1677 | * we can't argue if the task is increasing |
1695 | * or lowering its prio, so... | 1678 | * or lowering its prio, so... |
1696 | */ | 1679 | */ |
1697 | if (!rq->dl.overloaded) | 1680 | if (!rq->dl.overloaded) |
1698 | pull_dl_task(rq); | 1681 | pull_dl_task(rq); |
1699 | 1682 | ||
1700 | /* | 1683 | /* |
1701 | * If we now have a earlier deadline task than p, | 1684 | * If we now have a earlier deadline task than p, |
1702 | * then reschedule, provided p is still on this | 1685 | * then reschedule, provided p is still on this |
1703 | * runqueue. | 1686 | * runqueue. |
1704 | */ | 1687 | */ |
1705 | if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline) && | 1688 | if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline) && |
1706 | rq->curr == p) | 1689 | rq->curr == p) |
1707 | resched_curr(rq); | 1690 | resched_curr(rq); |
1708 | #else | 1691 | #else |
1709 | /* | 1692 | /* |
1710 | * Again, we don't know if p has a earlier | 1693 | * Again, we don't know if p has a earlier |
1711 | * or later deadline, so let's blindly set a | 1694 | * or later deadline, so let's blindly set a |
1712 | * (maybe not needed) rescheduling point. | 1695 | * (maybe not needed) rescheduling point. |
1713 | */ | 1696 | */ |
1714 | resched_curr(rq); | 1697 | resched_curr(rq); |
1715 | #endif /* CONFIG_SMP */ | 1698 | #endif /* CONFIG_SMP */ |
1716 | } else | 1699 | } else |
1717 | switched_to_dl(rq, p); | 1700 | switched_to_dl(rq, p); |
1718 | } | 1701 | } |
1719 | 1702 | ||
1720 | const struct sched_class dl_sched_class = { | 1703 | const struct sched_class dl_sched_class = { |
1721 | .next = &rt_sched_class, | 1704 | .next = &rt_sched_class, |
1722 | .enqueue_task = enqueue_task_dl, | 1705 | .enqueue_task = enqueue_task_dl, |
1723 | .dequeue_task = dequeue_task_dl, | 1706 | .dequeue_task = dequeue_task_dl, |
1724 | .yield_task = yield_task_dl, | 1707 | .yield_task = yield_task_dl, |
1725 | 1708 | ||
1726 | .check_preempt_curr = check_preempt_curr_dl, | 1709 | .check_preempt_curr = check_preempt_curr_dl, |
1727 | 1710 | ||
1728 | .pick_next_task = pick_next_task_dl, | 1711 | .pick_next_task = pick_next_task_dl, |
1729 | .put_prev_task = put_prev_task_dl, | 1712 | .put_prev_task = put_prev_task_dl, |
1730 | 1713 | ||
1731 | #ifdef CONFIG_SMP | 1714 | #ifdef CONFIG_SMP |
1732 | .select_task_rq = select_task_rq_dl, | 1715 | .select_task_rq = select_task_rq_dl, |
1733 | .set_cpus_allowed = set_cpus_allowed_dl, | 1716 | .set_cpus_allowed = set_cpus_allowed_dl, |
1734 | .rq_online = rq_online_dl, | 1717 | .rq_online = rq_online_dl, |
1735 | .rq_offline = rq_offline_dl, | 1718 | .rq_offline = rq_offline_dl, |
1736 | .post_schedule = post_schedule_dl, | 1719 | .post_schedule = post_schedule_dl, |
1737 | .task_woken = task_woken_dl, | 1720 | .task_woken = task_woken_dl, |
1738 | #endif | 1721 | #endif |
1739 | 1722 | ||
1740 | .set_curr_task = set_curr_task_dl, | 1723 | .set_curr_task = set_curr_task_dl, |
1741 | .task_tick = task_tick_dl, | 1724 | .task_tick = task_tick_dl, |
1742 | .task_fork = task_fork_dl, | 1725 | .task_fork = task_fork_dl, |
1743 | .task_dead = task_dead_dl, | 1726 | .task_dead = task_dead_dl, |
1744 | 1727 | ||
1745 | .prio_changed = prio_changed_dl, | 1728 | .prio_changed = prio_changed_dl, |
1746 | .switched_from = switched_from_dl, | 1729 | .switched_from = switched_from_dl, |
1747 | .switched_to = switched_to_dl, | 1730 | .switched_to = switched_to_dl, |
1748 | 1731 | ||
1749 | .update_curr = update_curr_dl, | 1732 | .update_curr = update_curr_dl, |
1750 | }; | 1733 | }; |
1751 | 1734 | ||
1752 | #ifdef CONFIG_SCHED_DEBUG | 1735 | #ifdef CONFIG_SCHED_DEBUG |
1753 | extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq); | 1736 | extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq); |
1754 | 1737 | ||
1755 | void print_dl_stats(struct seq_file *m, int cpu) | 1738 | void print_dl_stats(struct seq_file *m, int cpu) |
1756 | { | 1739 | { |
1757 | print_dl_rq(m, cpu, &cpu_rq(cpu)->dl); | 1740 | print_dl_rq(m, cpu, &cpu_rq(cpu)->dl); |
1758 | } | 1741 | } |
kernel/sched/fair.c
1 | /* | 1 | /* |
2 | * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH) | 2 | * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH) |
3 | * | 3 | * |
4 | * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> | 4 | * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> |
5 | * | 5 | * |
6 | * Interactivity improvements by Mike Galbraith | 6 | * Interactivity improvements by Mike Galbraith |
7 | * (C) 2007 Mike Galbraith <efault@gmx.de> | 7 | * (C) 2007 Mike Galbraith <efault@gmx.de> |
8 | * | 8 | * |
9 | * Various enhancements by Dmitry Adamushko. | 9 | * Various enhancements by Dmitry Adamushko. |
10 | * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com> | 10 | * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com> |
11 | * | 11 | * |
12 | * Group scheduling enhancements by Srivatsa Vaddagiri | 12 | * Group scheduling enhancements by Srivatsa Vaddagiri |
13 | * Copyright IBM Corporation, 2007 | 13 | * Copyright IBM Corporation, 2007 |
14 | * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> | 14 | * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> |
15 | * | 15 | * |
16 | * Scaled math optimizations by Thomas Gleixner | 16 | * Scaled math optimizations by Thomas Gleixner |
17 | * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de> | 17 | * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de> |
18 | * | 18 | * |
19 | * Adaptive scheduling granularity, math enhancements by Peter Zijlstra | 19 | * Adaptive scheduling granularity, math enhancements by Peter Zijlstra |
20 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | 20 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> |
21 | */ | 21 | */ |
22 | 22 | ||
23 | #include <linux/latencytop.h> | 23 | #include <linux/latencytop.h> |
24 | #include <linux/sched.h> | 24 | #include <linux/sched.h> |
25 | #include <linux/cpumask.h> | 25 | #include <linux/cpumask.h> |
26 | #include <linux/cpuidle.h> | 26 | #include <linux/cpuidle.h> |
27 | #include <linux/slab.h> | 27 | #include <linux/slab.h> |
28 | #include <linux/profile.h> | 28 | #include <linux/profile.h> |
29 | #include <linux/interrupt.h> | 29 | #include <linux/interrupt.h> |
30 | #include <linux/mempolicy.h> | 30 | #include <linux/mempolicy.h> |
31 | #include <linux/migrate.h> | 31 | #include <linux/migrate.h> |
32 | #include <linux/task_work.h> | 32 | #include <linux/task_work.h> |
33 | 33 | ||
34 | #include <trace/events/sched.h> | 34 | #include <trace/events/sched.h> |
35 | 35 | ||
36 | #include "sched.h" | 36 | #include "sched.h" |
37 | 37 | ||
38 | /* | 38 | /* |
39 | * Targeted preemption latency for CPU-bound tasks: | 39 | * Targeted preemption latency for CPU-bound tasks: |
40 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) | 40 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) |
41 | * | 41 | * |
42 | * NOTE: this latency value is not the same as the concept of | 42 | * NOTE: this latency value is not the same as the concept of |
43 | * 'timeslice length' - timeslices in CFS are of variable length | 43 | * 'timeslice length' - timeslices in CFS are of variable length |
44 | * and have no persistent notion like in traditional, time-slice | 44 | * and have no persistent notion like in traditional, time-slice |
45 | * based scheduling concepts. | 45 | * based scheduling concepts. |
46 | * | 46 | * |
47 | * (to see the precise effective timeslice length of your workload, | 47 | * (to see the precise effective timeslice length of your workload, |
48 | * run vmstat and monitor the context-switches (cs) field) | 48 | * run vmstat and monitor the context-switches (cs) field) |
49 | */ | 49 | */ |
50 | unsigned int sysctl_sched_latency = 6000000ULL; | 50 | unsigned int sysctl_sched_latency = 6000000ULL; |
51 | unsigned int normalized_sysctl_sched_latency = 6000000ULL; | 51 | unsigned int normalized_sysctl_sched_latency = 6000000ULL; |
52 | 52 | ||
53 | /* | 53 | /* |
54 | * The initial- and re-scaling of tunables is configurable | 54 | * The initial- and re-scaling of tunables is configurable |
55 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | 55 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) |
56 | * | 56 | * |
57 | * Options are: | 57 | * Options are: |
58 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | 58 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 |
59 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | 59 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) |
60 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | 60 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus |
61 | */ | 61 | */ |
62 | enum sched_tunable_scaling sysctl_sched_tunable_scaling | 62 | enum sched_tunable_scaling sysctl_sched_tunable_scaling |
63 | = SCHED_TUNABLESCALING_LOG; | 63 | = SCHED_TUNABLESCALING_LOG; |
64 | 64 | ||
65 | /* | 65 | /* |
66 | * Minimal preemption granularity for CPU-bound tasks: | 66 | * Minimal preemption granularity for CPU-bound tasks: |
67 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) | 67 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
68 | */ | 68 | */ |
69 | unsigned int sysctl_sched_min_granularity = 750000ULL; | 69 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
70 | unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | 70 | unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; |
71 | 71 | ||
72 | /* | 72 | /* |
73 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity | 73 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity |
74 | */ | 74 | */ |
75 | static unsigned int sched_nr_latency = 8; | 75 | static unsigned int sched_nr_latency = 8; |
76 | 76 | ||
77 | /* | 77 | /* |
78 | * After fork, child runs first. If set to 0 (default) then | 78 | * After fork, child runs first. If set to 0 (default) then |
79 | * parent will (try to) run first. | 79 | * parent will (try to) run first. |
80 | */ | 80 | */ |
81 | unsigned int sysctl_sched_child_runs_first __read_mostly; | 81 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
82 | 82 | ||
83 | /* | 83 | /* |
84 | * SCHED_OTHER wake-up granularity. | 84 | * SCHED_OTHER wake-up granularity. |
85 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) | 85 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) |
86 | * | 86 | * |
87 | * This option delays the preemption effects of decoupled workloads | 87 | * This option delays the preemption effects of decoupled workloads |
88 | * and reduces their over-scheduling. Synchronous workloads will still | 88 | * and reduces their over-scheduling. Synchronous workloads will still |
89 | * have immediate wakeup/sleep latencies. | 89 | * have immediate wakeup/sleep latencies. |
90 | */ | 90 | */ |
91 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; | 91 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
92 | unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; | 92 | unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; |
93 | 93 | ||
94 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; | 94 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
95 | 95 | ||
96 | /* | 96 | /* |
97 | * The exponential sliding window over which load is averaged for shares | 97 | * The exponential sliding window over which load is averaged for shares |
98 | * distribution. | 98 | * distribution. |
99 | * (default: 10msec) | 99 | * (default: 10msec) |
100 | */ | 100 | */ |
101 | unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL; | 101 | unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL; |
102 | 102 | ||
103 | #ifdef CONFIG_CFS_BANDWIDTH | 103 | #ifdef CONFIG_CFS_BANDWIDTH |
104 | /* | 104 | /* |
105 | * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool | 105 | * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool |
106 | * each time a cfs_rq requests quota. | 106 | * each time a cfs_rq requests quota. |
107 | * | 107 | * |
108 | * Note: in the case that the slice exceeds the runtime remaining (either due | 108 | * Note: in the case that the slice exceeds the runtime remaining (either due |
109 | * to consumption or the quota being specified to be smaller than the slice) | 109 | * to consumption or the quota being specified to be smaller than the slice) |
110 | * we will always only issue the remaining available time. | 110 | * we will always only issue the remaining available time. |
111 | * | 111 | * |
112 | * default: 5 msec, units: microseconds | 112 | * default: 5 msec, units: microseconds |
113 | */ | 113 | */ |
114 | unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; | 114 | unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; |
115 | #endif | 115 | #endif |
116 | 116 | ||
117 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) | 117 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
118 | { | 118 | { |
119 | lw->weight += inc; | 119 | lw->weight += inc; |
120 | lw->inv_weight = 0; | 120 | lw->inv_weight = 0; |
121 | } | 121 | } |
122 | 122 | ||
123 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) | 123 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) |
124 | { | 124 | { |
125 | lw->weight -= dec; | 125 | lw->weight -= dec; |
126 | lw->inv_weight = 0; | 126 | lw->inv_weight = 0; |
127 | } | 127 | } |
128 | 128 | ||
129 | static inline void update_load_set(struct load_weight *lw, unsigned long w) | 129 | static inline void update_load_set(struct load_weight *lw, unsigned long w) |
130 | { | 130 | { |
131 | lw->weight = w; | 131 | lw->weight = w; |
132 | lw->inv_weight = 0; | 132 | lw->inv_weight = 0; |
133 | } | 133 | } |
134 | 134 | ||
135 | /* | 135 | /* |
136 | * Increase the granularity value when there are more CPUs, | 136 | * Increase the granularity value when there are more CPUs, |
137 | * because with more CPUs the 'effective latency' as visible | 137 | * because with more CPUs the 'effective latency' as visible |
138 | * to users decreases. But the relationship is not linear, | 138 | * to users decreases. But the relationship is not linear, |
139 | * so pick a second-best guess by going with the log2 of the | 139 | * so pick a second-best guess by going with the log2 of the |
140 | * number of CPUs. | 140 | * number of CPUs. |
141 | * | 141 | * |
142 | * This idea comes from the SD scheduler of Con Kolivas: | 142 | * This idea comes from the SD scheduler of Con Kolivas: |
143 | */ | 143 | */ |
144 | static int get_update_sysctl_factor(void) | 144 | static int get_update_sysctl_factor(void) |
145 | { | 145 | { |
146 | unsigned int cpus = min_t(int, num_online_cpus(), 8); | 146 | unsigned int cpus = min_t(int, num_online_cpus(), 8); |
147 | unsigned int factor; | 147 | unsigned int factor; |
148 | 148 | ||
149 | switch (sysctl_sched_tunable_scaling) { | 149 | switch (sysctl_sched_tunable_scaling) { |
150 | case SCHED_TUNABLESCALING_NONE: | 150 | case SCHED_TUNABLESCALING_NONE: |
151 | factor = 1; | 151 | factor = 1; |
152 | break; | 152 | break; |
153 | case SCHED_TUNABLESCALING_LINEAR: | 153 | case SCHED_TUNABLESCALING_LINEAR: |
154 | factor = cpus; | 154 | factor = cpus; |
155 | break; | 155 | break; |
156 | case SCHED_TUNABLESCALING_LOG: | 156 | case SCHED_TUNABLESCALING_LOG: |
157 | default: | 157 | default: |
158 | factor = 1 + ilog2(cpus); | 158 | factor = 1 + ilog2(cpus); |
159 | break; | 159 | break; |
160 | } | 160 | } |
161 | 161 | ||
162 | return factor; | 162 | return factor; |
163 | } | 163 | } |
164 | 164 | ||
165 | static void update_sysctl(void) | 165 | static void update_sysctl(void) |
166 | { | 166 | { |
167 | unsigned int factor = get_update_sysctl_factor(); | 167 | unsigned int factor = get_update_sysctl_factor(); |
168 | 168 | ||
169 | #define SET_SYSCTL(name) \ | 169 | #define SET_SYSCTL(name) \ |
170 | (sysctl_##name = (factor) * normalized_sysctl_##name) | 170 | (sysctl_##name = (factor) * normalized_sysctl_##name) |
171 | SET_SYSCTL(sched_min_granularity); | 171 | SET_SYSCTL(sched_min_granularity); |
172 | SET_SYSCTL(sched_latency); | 172 | SET_SYSCTL(sched_latency); |
173 | SET_SYSCTL(sched_wakeup_granularity); | 173 | SET_SYSCTL(sched_wakeup_granularity); |
174 | #undef SET_SYSCTL | 174 | #undef SET_SYSCTL |
175 | } | 175 | } |
176 | 176 | ||
177 | void sched_init_granularity(void) | 177 | void sched_init_granularity(void) |
178 | { | 178 | { |
179 | update_sysctl(); | 179 | update_sysctl(); |
180 | } | 180 | } |
181 | 181 | ||
182 | #define WMULT_CONST (~0U) | 182 | #define WMULT_CONST (~0U) |
183 | #define WMULT_SHIFT 32 | 183 | #define WMULT_SHIFT 32 |
184 | 184 | ||
185 | static void __update_inv_weight(struct load_weight *lw) | 185 | static void __update_inv_weight(struct load_weight *lw) |
186 | { | 186 | { |
187 | unsigned long w; | 187 | unsigned long w; |
188 | 188 | ||
189 | if (likely(lw->inv_weight)) | 189 | if (likely(lw->inv_weight)) |
190 | return; | 190 | return; |
191 | 191 | ||
192 | w = scale_load_down(lw->weight); | 192 | w = scale_load_down(lw->weight); |
193 | 193 | ||
194 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | 194 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) |
195 | lw->inv_weight = 1; | 195 | lw->inv_weight = 1; |
196 | else if (unlikely(!w)) | 196 | else if (unlikely(!w)) |
197 | lw->inv_weight = WMULT_CONST; | 197 | lw->inv_weight = WMULT_CONST; |
198 | else | 198 | else |
199 | lw->inv_weight = WMULT_CONST / w; | 199 | lw->inv_weight = WMULT_CONST / w; |
200 | } | 200 | } |
201 | 201 | ||
202 | /* | 202 | /* |
203 | * delta_exec * weight / lw.weight | 203 | * delta_exec * weight / lw.weight |
204 | * OR | 204 | * OR |
205 | * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT | 205 | * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT |
206 | * | 206 | * |
207 | * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case | 207 | * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case |
208 | * we're guaranteed shift stays positive because inv_weight is guaranteed to | 208 | * we're guaranteed shift stays positive because inv_weight is guaranteed to |
209 | * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22. | 209 | * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22. |
210 | * | 210 | * |
211 | * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus | 211 | * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus |
212 | * weight/lw.weight <= 1, and therefore our shift will also be positive. | 212 | * weight/lw.weight <= 1, and therefore our shift will also be positive. |
213 | */ | 213 | */ |
214 | static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw) | 214 | static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw) |
215 | { | 215 | { |
216 | u64 fact = scale_load_down(weight); | 216 | u64 fact = scale_load_down(weight); |
217 | int shift = WMULT_SHIFT; | 217 | int shift = WMULT_SHIFT; |
218 | 218 | ||
219 | __update_inv_weight(lw); | 219 | __update_inv_weight(lw); |
220 | 220 | ||
221 | if (unlikely(fact >> 32)) { | 221 | if (unlikely(fact >> 32)) { |
222 | while (fact >> 32) { | 222 | while (fact >> 32) { |
223 | fact >>= 1; | 223 | fact >>= 1; |
224 | shift--; | 224 | shift--; |
225 | } | 225 | } |
226 | } | 226 | } |
227 | 227 | ||
228 | /* hint to use a 32x32->64 mul */ | 228 | /* hint to use a 32x32->64 mul */ |
229 | fact = (u64)(u32)fact * lw->inv_weight; | 229 | fact = (u64)(u32)fact * lw->inv_weight; |
230 | 230 | ||
231 | while (fact >> 32) { | 231 | while (fact >> 32) { |
232 | fact >>= 1; | 232 | fact >>= 1; |
233 | shift--; | 233 | shift--; |
234 | } | 234 | } |
235 | 235 | ||
236 | return mul_u64_u32_shr(delta_exec, fact, shift); | 236 | return mul_u64_u32_shr(delta_exec, fact, shift); |
237 | } | 237 | } |
238 | 238 | ||
239 | 239 | ||
240 | const struct sched_class fair_sched_class; | 240 | const struct sched_class fair_sched_class; |
241 | 241 | ||
242 | /************************************************************** | 242 | /************************************************************** |
243 | * CFS operations on generic schedulable entities: | 243 | * CFS operations on generic schedulable entities: |
244 | */ | 244 | */ |
245 | 245 | ||
246 | #ifdef CONFIG_FAIR_GROUP_SCHED | 246 | #ifdef CONFIG_FAIR_GROUP_SCHED |
247 | 247 | ||
248 | /* cpu runqueue to which this cfs_rq is attached */ | 248 | /* cpu runqueue to which this cfs_rq is attached */ |
249 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) | 249 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
250 | { | 250 | { |
251 | return cfs_rq->rq; | 251 | return cfs_rq->rq; |
252 | } | 252 | } |
253 | 253 | ||
254 | /* An entity is a task if it doesn't "own" a runqueue */ | 254 | /* An entity is a task if it doesn't "own" a runqueue */ |
255 | #define entity_is_task(se) (!se->my_q) | 255 | #define entity_is_task(se) (!se->my_q) |
256 | 256 | ||
257 | static inline struct task_struct *task_of(struct sched_entity *se) | 257 | static inline struct task_struct *task_of(struct sched_entity *se) |
258 | { | 258 | { |
259 | #ifdef CONFIG_SCHED_DEBUG | 259 | #ifdef CONFIG_SCHED_DEBUG |
260 | WARN_ON_ONCE(!entity_is_task(se)); | 260 | WARN_ON_ONCE(!entity_is_task(se)); |
261 | #endif | 261 | #endif |
262 | return container_of(se, struct task_struct, se); | 262 | return container_of(se, struct task_struct, se); |
263 | } | 263 | } |
264 | 264 | ||
265 | /* Walk up scheduling entities hierarchy */ | 265 | /* Walk up scheduling entities hierarchy */ |
266 | #define for_each_sched_entity(se) \ | 266 | #define for_each_sched_entity(se) \ |
267 | for (; se; se = se->parent) | 267 | for (; se; se = se->parent) |
268 | 268 | ||
269 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | 269 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
270 | { | 270 | { |
271 | return p->se.cfs_rq; | 271 | return p->se.cfs_rq; |
272 | } | 272 | } |
273 | 273 | ||
274 | /* runqueue on which this entity is (to be) queued */ | 274 | /* runqueue on which this entity is (to be) queued */ |
275 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | 275 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
276 | { | 276 | { |
277 | return se->cfs_rq; | 277 | return se->cfs_rq; |
278 | } | 278 | } |
279 | 279 | ||
280 | /* runqueue "owned" by this group */ | 280 | /* runqueue "owned" by this group */ |
281 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | 281 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) |
282 | { | 282 | { |
283 | return grp->my_q; | 283 | return grp->my_q; |
284 | } | 284 | } |
285 | 285 | ||
286 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, | 286 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
287 | int force_update); | 287 | int force_update); |
288 | 288 | ||
289 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) | 289 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
290 | { | 290 | { |
291 | if (!cfs_rq->on_list) { | 291 | if (!cfs_rq->on_list) { |
292 | /* | 292 | /* |
293 | * Ensure we either appear before our parent (if already | 293 | * Ensure we either appear before our parent (if already |
294 | * enqueued) or force our parent to appear after us when it is | 294 | * enqueued) or force our parent to appear after us when it is |
295 | * enqueued. The fact that we always enqueue bottom-up | 295 | * enqueued. The fact that we always enqueue bottom-up |
296 | * reduces this to two cases. | 296 | * reduces this to two cases. |
297 | */ | 297 | */ |
298 | if (cfs_rq->tg->parent && | 298 | if (cfs_rq->tg->parent && |
299 | cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { | 299 | cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { |
300 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | 300 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, |
301 | &rq_of(cfs_rq)->leaf_cfs_rq_list); | 301 | &rq_of(cfs_rq)->leaf_cfs_rq_list); |
302 | } else { | 302 | } else { |
303 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | 303 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, |
304 | &rq_of(cfs_rq)->leaf_cfs_rq_list); | 304 | &rq_of(cfs_rq)->leaf_cfs_rq_list); |
305 | } | 305 | } |
306 | 306 | ||
307 | cfs_rq->on_list = 1; | 307 | cfs_rq->on_list = 1; |
308 | /* We should have no load, but we need to update last_decay. */ | 308 | /* We should have no load, but we need to update last_decay. */ |
309 | update_cfs_rq_blocked_load(cfs_rq, 0); | 309 | update_cfs_rq_blocked_load(cfs_rq, 0); |
310 | } | 310 | } |
311 | } | 311 | } |
312 | 312 | ||
313 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | 313 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
314 | { | 314 | { |
315 | if (cfs_rq->on_list) { | 315 | if (cfs_rq->on_list) { |
316 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | 316 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); |
317 | cfs_rq->on_list = 0; | 317 | cfs_rq->on_list = 0; |
318 | } | 318 | } |
319 | } | 319 | } |
320 | 320 | ||
321 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ | 321 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
322 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | 322 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
323 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | 323 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) |
324 | 324 | ||
325 | /* Do the two (enqueued) entities belong to the same group ? */ | 325 | /* Do the two (enqueued) entities belong to the same group ? */ |
326 | static inline struct cfs_rq * | 326 | static inline struct cfs_rq * |
327 | is_same_group(struct sched_entity *se, struct sched_entity *pse) | 327 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
328 | { | 328 | { |
329 | if (se->cfs_rq == pse->cfs_rq) | 329 | if (se->cfs_rq == pse->cfs_rq) |
330 | return se->cfs_rq; | 330 | return se->cfs_rq; |
331 | 331 | ||
332 | return NULL; | 332 | return NULL; |
333 | } | 333 | } |
334 | 334 | ||
335 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | 335 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
336 | { | 336 | { |
337 | return se->parent; | 337 | return se->parent; |
338 | } | 338 | } |
339 | 339 | ||
340 | static void | 340 | static void |
341 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | 341 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) |
342 | { | 342 | { |
343 | int se_depth, pse_depth; | 343 | int se_depth, pse_depth; |
344 | 344 | ||
345 | /* | 345 | /* |
346 | * preemption test can be made between sibling entities who are in the | 346 | * preemption test can be made between sibling entities who are in the |
347 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | 347 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of |
348 | * both tasks until we find their ancestors who are siblings of common | 348 | * both tasks until we find their ancestors who are siblings of common |
349 | * parent. | 349 | * parent. |
350 | */ | 350 | */ |
351 | 351 | ||
352 | /* First walk up until both entities are at same depth */ | 352 | /* First walk up until both entities are at same depth */ |
353 | se_depth = (*se)->depth; | 353 | se_depth = (*se)->depth; |
354 | pse_depth = (*pse)->depth; | 354 | pse_depth = (*pse)->depth; |
355 | 355 | ||
356 | while (se_depth > pse_depth) { | 356 | while (se_depth > pse_depth) { |
357 | se_depth--; | 357 | se_depth--; |
358 | *se = parent_entity(*se); | 358 | *se = parent_entity(*se); |
359 | } | 359 | } |
360 | 360 | ||
361 | while (pse_depth > se_depth) { | 361 | while (pse_depth > se_depth) { |
362 | pse_depth--; | 362 | pse_depth--; |
363 | *pse = parent_entity(*pse); | 363 | *pse = parent_entity(*pse); |
364 | } | 364 | } |
365 | 365 | ||
366 | while (!is_same_group(*se, *pse)) { | 366 | while (!is_same_group(*se, *pse)) { |
367 | *se = parent_entity(*se); | 367 | *se = parent_entity(*se); |
368 | *pse = parent_entity(*pse); | 368 | *pse = parent_entity(*pse); |
369 | } | 369 | } |
370 | } | 370 | } |
371 | 371 | ||
372 | #else /* !CONFIG_FAIR_GROUP_SCHED */ | 372 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
373 | 373 | ||
374 | static inline struct task_struct *task_of(struct sched_entity *se) | 374 | static inline struct task_struct *task_of(struct sched_entity *se) |
375 | { | 375 | { |
376 | return container_of(se, struct task_struct, se); | 376 | return container_of(se, struct task_struct, se); |
377 | } | 377 | } |
378 | 378 | ||
379 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) | 379 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
380 | { | 380 | { |
381 | return container_of(cfs_rq, struct rq, cfs); | 381 | return container_of(cfs_rq, struct rq, cfs); |
382 | } | 382 | } |
383 | 383 | ||
384 | #define entity_is_task(se) 1 | 384 | #define entity_is_task(se) 1 |
385 | 385 | ||
386 | #define for_each_sched_entity(se) \ | 386 | #define for_each_sched_entity(se) \ |
387 | for (; se; se = NULL) | 387 | for (; se; se = NULL) |
388 | 388 | ||
389 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | 389 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
390 | { | 390 | { |
391 | return &task_rq(p)->cfs; | 391 | return &task_rq(p)->cfs; |
392 | } | 392 | } |
393 | 393 | ||
394 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | 394 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
395 | { | 395 | { |
396 | struct task_struct *p = task_of(se); | 396 | struct task_struct *p = task_of(se); |
397 | struct rq *rq = task_rq(p); | 397 | struct rq *rq = task_rq(p); |
398 | 398 | ||
399 | return &rq->cfs; | 399 | return &rq->cfs; |
400 | } | 400 | } |
401 | 401 | ||
402 | /* runqueue "owned" by this group */ | 402 | /* runqueue "owned" by this group */ |
403 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | 403 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) |
404 | { | 404 | { |
405 | return NULL; | 405 | return NULL; |
406 | } | 406 | } |
407 | 407 | ||
408 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) | 408 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
409 | { | 409 | { |
410 | } | 410 | } |
411 | 411 | ||
412 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | 412 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
413 | { | 413 | { |
414 | } | 414 | } |
415 | 415 | ||
416 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | 416 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
417 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | 417 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) |
418 | 418 | ||
419 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | 419 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
420 | { | 420 | { |
421 | return NULL; | 421 | return NULL; |
422 | } | 422 | } |
423 | 423 | ||
424 | static inline void | 424 | static inline void |
425 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | 425 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) |
426 | { | 426 | { |
427 | } | 427 | } |
428 | 428 | ||
429 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | 429 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
430 | 430 | ||
431 | static __always_inline | 431 | static __always_inline |
432 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); | 432 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); |
433 | 433 | ||
434 | /************************************************************** | 434 | /************************************************************** |
435 | * Scheduling class tree data structure manipulation methods: | 435 | * Scheduling class tree data structure manipulation methods: |
436 | */ | 436 | */ |
437 | 437 | ||
438 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) | 438 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
439 | { | 439 | { |
440 | s64 delta = (s64)(vruntime - max_vruntime); | 440 | s64 delta = (s64)(vruntime - max_vruntime); |
441 | if (delta > 0) | 441 | if (delta > 0) |
442 | max_vruntime = vruntime; | 442 | max_vruntime = vruntime; |
443 | 443 | ||
444 | return max_vruntime; | 444 | return max_vruntime; |
445 | } | 445 | } |
446 | 446 | ||
447 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) | 447 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
448 | { | 448 | { |
449 | s64 delta = (s64)(vruntime - min_vruntime); | 449 | s64 delta = (s64)(vruntime - min_vruntime); |
450 | if (delta < 0) | 450 | if (delta < 0) |
451 | min_vruntime = vruntime; | 451 | min_vruntime = vruntime; |
452 | 452 | ||
453 | return min_vruntime; | 453 | return min_vruntime; |
454 | } | 454 | } |
455 | 455 | ||
456 | static inline int entity_before(struct sched_entity *a, | 456 | static inline int entity_before(struct sched_entity *a, |
457 | struct sched_entity *b) | 457 | struct sched_entity *b) |
458 | { | 458 | { |
459 | return (s64)(a->vruntime - b->vruntime) < 0; | 459 | return (s64)(a->vruntime - b->vruntime) < 0; |
460 | } | 460 | } |
461 | 461 | ||
462 | static void update_min_vruntime(struct cfs_rq *cfs_rq) | 462 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
463 | { | 463 | { |
464 | u64 vruntime = cfs_rq->min_vruntime; | 464 | u64 vruntime = cfs_rq->min_vruntime; |
465 | 465 | ||
466 | if (cfs_rq->curr) | 466 | if (cfs_rq->curr) |
467 | vruntime = cfs_rq->curr->vruntime; | 467 | vruntime = cfs_rq->curr->vruntime; |
468 | 468 | ||
469 | if (cfs_rq->rb_leftmost) { | 469 | if (cfs_rq->rb_leftmost) { |
470 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | 470 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, |
471 | struct sched_entity, | 471 | struct sched_entity, |
472 | run_node); | 472 | run_node); |
473 | 473 | ||
474 | if (!cfs_rq->curr) | 474 | if (!cfs_rq->curr) |
475 | vruntime = se->vruntime; | 475 | vruntime = se->vruntime; |
476 | else | 476 | else |
477 | vruntime = min_vruntime(vruntime, se->vruntime); | 477 | vruntime = min_vruntime(vruntime, se->vruntime); |
478 | } | 478 | } |
479 | 479 | ||
480 | /* ensure we never gain time by being placed backwards. */ | 480 | /* ensure we never gain time by being placed backwards. */ |
481 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); | 481 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); |
482 | #ifndef CONFIG_64BIT | 482 | #ifndef CONFIG_64BIT |
483 | smp_wmb(); | 483 | smp_wmb(); |
484 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | 484 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; |
485 | #endif | 485 | #endif |
486 | } | 486 | } |
487 | 487 | ||
488 | /* | 488 | /* |
489 | * Enqueue an entity into the rb-tree: | 489 | * Enqueue an entity into the rb-tree: |
490 | */ | 490 | */ |
491 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) | 491 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
492 | { | 492 | { |
493 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | 493 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; |
494 | struct rb_node *parent = NULL; | 494 | struct rb_node *parent = NULL; |
495 | struct sched_entity *entry; | 495 | struct sched_entity *entry; |
496 | int leftmost = 1; | 496 | int leftmost = 1; |
497 | 497 | ||
498 | /* | 498 | /* |
499 | * Find the right place in the rbtree: | 499 | * Find the right place in the rbtree: |
500 | */ | 500 | */ |
501 | while (*link) { | 501 | while (*link) { |
502 | parent = *link; | 502 | parent = *link; |
503 | entry = rb_entry(parent, struct sched_entity, run_node); | 503 | entry = rb_entry(parent, struct sched_entity, run_node); |
504 | /* | 504 | /* |
505 | * We dont care about collisions. Nodes with | 505 | * We dont care about collisions. Nodes with |
506 | * the same key stay together. | 506 | * the same key stay together. |
507 | */ | 507 | */ |
508 | if (entity_before(se, entry)) { | 508 | if (entity_before(se, entry)) { |
509 | link = &parent->rb_left; | 509 | link = &parent->rb_left; |
510 | } else { | 510 | } else { |
511 | link = &parent->rb_right; | 511 | link = &parent->rb_right; |
512 | leftmost = 0; | 512 | leftmost = 0; |
513 | } | 513 | } |
514 | } | 514 | } |
515 | 515 | ||
516 | /* | 516 | /* |
517 | * Maintain a cache of leftmost tree entries (it is frequently | 517 | * Maintain a cache of leftmost tree entries (it is frequently |
518 | * used): | 518 | * used): |
519 | */ | 519 | */ |
520 | if (leftmost) | 520 | if (leftmost) |
521 | cfs_rq->rb_leftmost = &se->run_node; | 521 | cfs_rq->rb_leftmost = &se->run_node; |
522 | 522 | ||
523 | rb_link_node(&se->run_node, parent, link); | 523 | rb_link_node(&se->run_node, parent, link); |
524 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | 524 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); |
525 | } | 525 | } |
526 | 526 | ||
527 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) | 527 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
528 | { | 528 | { |
529 | if (cfs_rq->rb_leftmost == &se->run_node) { | 529 | if (cfs_rq->rb_leftmost == &se->run_node) { |
530 | struct rb_node *next_node; | 530 | struct rb_node *next_node; |
531 | 531 | ||
532 | next_node = rb_next(&se->run_node); | 532 | next_node = rb_next(&se->run_node); |
533 | cfs_rq->rb_leftmost = next_node; | 533 | cfs_rq->rb_leftmost = next_node; |
534 | } | 534 | } |
535 | 535 | ||
536 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); | 536 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
537 | } | 537 | } |
538 | 538 | ||
539 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) | 539 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
540 | { | 540 | { |
541 | struct rb_node *left = cfs_rq->rb_leftmost; | 541 | struct rb_node *left = cfs_rq->rb_leftmost; |
542 | 542 | ||
543 | if (!left) | 543 | if (!left) |
544 | return NULL; | 544 | return NULL; |
545 | 545 | ||
546 | return rb_entry(left, struct sched_entity, run_node); | 546 | return rb_entry(left, struct sched_entity, run_node); |
547 | } | 547 | } |
548 | 548 | ||
549 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) | 549 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
550 | { | 550 | { |
551 | struct rb_node *next = rb_next(&se->run_node); | 551 | struct rb_node *next = rb_next(&se->run_node); |
552 | 552 | ||
553 | if (!next) | 553 | if (!next) |
554 | return NULL; | 554 | return NULL; |
555 | 555 | ||
556 | return rb_entry(next, struct sched_entity, run_node); | 556 | return rb_entry(next, struct sched_entity, run_node); |
557 | } | 557 | } |
558 | 558 | ||
559 | #ifdef CONFIG_SCHED_DEBUG | 559 | #ifdef CONFIG_SCHED_DEBUG |
560 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) | 560 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
561 | { | 561 | { |
562 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); | 562 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
563 | 563 | ||
564 | if (!last) | 564 | if (!last) |
565 | return NULL; | 565 | return NULL; |
566 | 566 | ||
567 | return rb_entry(last, struct sched_entity, run_node); | 567 | return rb_entry(last, struct sched_entity, run_node); |
568 | } | 568 | } |
569 | 569 | ||
570 | /************************************************************** | 570 | /************************************************************** |
571 | * Scheduling class statistics methods: | 571 | * Scheduling class statistics methods: |
572 | */ | 572 | */ |
573 | 573 | ||
574 | int sched_proc_update_handler(struct ctl_table *table, int write, | 574 | int sched_proc_update_handler(struct ctl_table *table, int write, |
575 | void __user *buffer, size_t *lenp, | 575 | void __user *buffer, size_t *lenp, |
576 | loff_t *ppos) | 576 | loff_t *ppos) |
577 | { | 577 | { |
578 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | 578 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
579 | int factor = get_update_sysctl_factor(); | 579 | int factor = get_update_sysctl_factor(); |
580 | 580 | ||
581 | if (ret || !write) | 581 | if (ret || !write) |
582 | return ret; | 582 | return ret; |
583 | 583 | ||
584 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | 584 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, |
585 | sysctl_sched_min_granularity); | 585 | sysctl_sched_min_granularity); |
586 | 586 | ||
587 | #define WRT_SYSCTL(name) \ | 587 | #define WRT_SYSCTL(name) \ |
588 | (normalized_sysctl_##name = sysctl_##name / (factor)) | 588 | (normalized_sysctl_##name = sysctl_##name / (factor)) |
589 | WRT_SYSCTL(sched_min_granularity); | 589 | WRT_SYSCTL(sched_min_granularity); |
590 | WRT_SYSCTL(sched_latency); | 590 | WRT_SYSCTL(sched_latency); |
591 | WRT_SYSCTL(sched_wakeup_granularity); | 591 | WRT_SYSCTL(sched_wakeup_granularity); |
592 | #undef WRT_SYSCTL | 592 | #undef WRT_SYSCTL |
593 | 593 | ||
594 | return 0; | 594 | return 0; |
595 | } | 595 | } |
596 | #endif | 596 | #endif |
597 | 597 | ||
598 | /* | 598 | /* |
599 | * delta /= w | 599 | * delta /= w |
600 | */ | 600 | */ |
601 | static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) | 601 | static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) |
602 | { | 602 | { |
603 | if (unlikely(se->load.weight != NICE_0_LOAD)) | 603 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
604 | delta = __calc_delta(delta, NICE_0_LOAD, &se->load); | 604 | delta = __calc_delta(delta, NICE_0_LOAD, &se->load); |
605 | 605 | ||
606 | return delta; | 606 | return delta; |
607 | } | 607 | } |
608 | 608 | ||
609 | /* | 609 | /* |
610 | * The idea is to set a period in which each task runs once. | 610 | * The idea is to set a period in which each task runs once. |
611 | * | 611 | * |
612 | * When there are too many tasks (sched_nr_latency) we have to stretch | 612 | * When there are too many tasks (sched_nr_latency) we have to stretch |
613 | * this period because otherwise the slices get too small. | 613 | * this period because otherwise the slices get too small. |
614 | * | 614 | * |
615 | * p = (nr <= nl) ? l : l*nr/nl | 615 | * p = (nr <= nl) ? l : l*nr/nl |
616 | */ | 616 | */ |
617 | static u64 __sched_period(unsigned long nr_running) | 617 | static u64 __sched_period(unsigned long nr_running) |
618 | { | 618 | { |
619 | u64 period = sysctl_sched_latency; | 619 | u64 period = sysctl_sched_latency; |
620 | unsigned long nr_latency = sched_nr_latency; | 620 | unsigned long nr_latency = sched_nr_latency; |
621 | 621 | ||
622 | if (unlikely(nr_running > nr_latency)) { | 622 | if (unlikely(nr_running > nr_latency)) { |
623 | period = sysctl_sched_min_granularity; | 623 | period = sysctl_sched_min_granularity; |
624 | period *= nr_running; | 624 | period *= nr_running; |
625 | } | 625 | } |
626 | 626 | ||
627 | return period; | 627 | return period; |
628 | } | 628 | } |
629 | 629 | ||
630 | /* | 630 | /* |
631 | * We calculate the wall-time slice from the period by taking a part | 631 | * We calculate the wall-time slice from the period by taking a part |
632 | * proportional to the weight. | 632 | * proportional to the weight. |
633 | * | 633 | * |
634 | * s = p*P[w/rw] | 634 | * s = p*P[w/rw] |
635 | */ | 635 | */ |
636 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) | 636 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
637 | { | 637 | { |
638 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); | 638 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
639 | 639 | ||
640 | for_each_sched_entity(se) { | 640 | for_each_sched_entity(se) { |
641 | struct load_weight *load; | 641 | struct load_weight *load; |
642 | struct load_weight lw; | 642 | struct load_weight lw; |
643 | 643 | ||
644 | cfs_rq = cfs_rq_of(se); | 644 | cfs_rq = cfs_rq_of(se); |
645 | load = &cfs_rq->load; | 645 | load = &cfs_rq->load; |
646 | 646 | ||
647 | if (unlikely(!se->on_rq)) { | 647 | if (unlikely(!se->on_rq)) { |
648 | lw = cfs_rq->load; | 648 | lw = cfs_rq->load; |
649 | 649 | ||
650 | update_load_add(&lw, se->load.weight); | 650 | update_load_add(&lw, se->load.weight); |
651 | load = &lw; | 651 | load = &lw; |
652 | } | 652 | } |
653 | slice = __calc_delta(slice, se->load.weight, load); | 653 | slice = __calc_delta(slice, se->load.weight, load); |
654 | } | 654 | } |
655 | return slice; | 655 | return slice; |
656 | } | 656 | } |
657 | 657 | ||
658 | /* | 658 | /* |
659 | * We calculate the vruntime slice of a to-be-inserted task. | 659 | * We calculate the vruntime slice of a to-be-inserted task. |
660 | * | 660 | * |
661 | * vs = s/w | 661 | * vs = s/w |
662 | */ | 662 | */ |
663 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) | 663 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
664 | { | 664 | { |
665 | return calc_delta_fair(sched_slice(cfs_rq, se), se); | 665 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
666 | } | 666 | } |
667 | 667 | ||
668 | #ifdef CONFIG_SMP | 668 | #ifdef CONFIG_SMP |
669 | static int select_idle_sibling(struct task_struct *p, int cpu); | 669 | static int select_idle_sibling(struct task_struct *p, int cpu); |
670 | static unsigned long task_h_load(struct task_struct *p); | 670 | static unsigned long task_h_load(struct task_struct *p); |
671 | 671 | ||
672 | static inline void __update_task_entity_contrib(struct sched_entity *se); | 672 | static inline void __update_task_entity_contrib(struct sched_entity *se); |
673 | 673 | ||
674 | /* Give new task start runnable values to heavy its load in infant time */ | 674 | /* Give new task start runnable values to heavy its load in infant time */ |
675 | void init_task_runnable_average(struct task_struct *p) | 675 | void init_task_runnable_average(struct task_struct *p) |
676 | { | 676 | { |
677 | u32 slice; | 677 | u32 slice; |
678 | 678 | ||
679 | p->se.avg.decay_count = 0; | 679 | p->se.avg.decay_count = 0; |
680 | slice = sched_slice(task_cfs_rq(p), &p->se) >> 10; | 680 | slice = sched_slice(task_cfs_rq(p), &p->se) >> 10; |
681 | p->se.avg.runnable_avg_sum = slice; | 681 | p->se.avg.runnable_avg_sum = slice; |
682 | p->se.avg.runnable_avg_period = slice; | 682 | p->se.avg.runnable_avg_period = slice; |
683 | __update_task_entity_contrib(&p->se); | 683 | __update_task_entity_contrib(&p->se); |
684 | } | 684 | } |
685 | #else | 685 | #else |
686 | void init_task_runnable_average(struct task_struct *p) | 686 | void init_task_runnable_average(struct task_struct *p) |
687 | { | 687 | { |
688 | } | 688 | } |
689 | #endif | 689 | #endif |
690 | 690 | ||
691 | /* | 691 | /* |
692 | * Update the current task's runtime statistics. | 692 | * Update the current task's runtime statistics. |
693 | */ | 693 | */ |
694 | static void update_curr(struct cfs_rq *cfs_rq) | 694 | static void update_curr(struct cfs_rq *cfs_rq) |
695 | { | 695 | { |
696 | struct sched_entity *curr = cfs_rq->curr; | 696 | struct sched_entity *curr = cfs_rq->curr; |
697 | u64 now = rq_clock_task(rq_of(cfs_rq)); | 697 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
698 | u64 delta_exec; | 698 | u64 delta_exec; |
699 | 699 | ||
700 | if (unlikely(!curr)) | 700 | if (unlikely(!curr)) |
701 | return; | 701 | return; |
702 | 702 | ||
703 | delta_exec = now - curr->exec_start; | 703 | delta_exec = now - curr->exec_start; |
704 | if (unlikely((s64)delta_exec <= 0)) | 704 | if (unlikely((s64)delta_exec <= 0)) |
705 | return; | 705 | return; |
706 | 706 | ||
707 | curr->exec_start = now; | 707 | curr->exec_start = now; |
708 | 708 | ||
709 | schedstat_set(curr->statistics.exec_max, | 709 | schedstat_set(curr->statistics.exec_max, |
710 | max(delta_exec, curr->statistics.exec_max)); | 710 | max(delta_exec, curr->statistics.exec_max)); |
711 | 711 | ||
712 | curr->sum_exec_runtime += delta_exec; | 712 | curr->sum_exec_runtime += delta_exec; |
713 | schedstat_add(cfs_rq, exec_clock, delta_exec); | 713 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
714 | 714 | ||
715 | curr->vruntime += calc_delta_fair(delta_exec, curr); | 715 | curr->vruntime += calc_delta_fair(delta_exec, curr); |
716 | update_min_vruntime(cfs_rq); | 716 | update_min_vruntime(cfs_rq); |
717 | 717 | ||
718 | if (entity_is_task(curr)) { | 718 | if (entity_is_task(curr)) { |
719 | struct task_struct *curtask = task_of(curr); | 719 | struct task_struct *curtask = task_of(curr); |
720 | 720 | ||
721 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); | 721 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
722 | cpuacct_charge(curtask, delta_exec); | 722 | cpuacct_charge(curtask, delta_exec); |
723 | account_group_exec_runtime(curtask, delta_exec); | 723 | account_group_exec_runtime(curtask, delta_exec); |
724 | } | 724 | } |
725 | 725 | ||
726 | account_cfs_rq_runtime(cfs_rq, delta_exec); | 726 | account_cfs_rq_runtime(cfs_rq, delta_exec); |
727 | } | 727 | } |
728 | 728 | ||
729 | static void update_curr_fair(struct rq *rq) | 729 | static void update_curr_fair(struct rq *rq) |
730 | { | 730 | { |
731 | update_curr(cfs_rq_of(&rq->curr->se)); | 731 | update_curr(cfs_rq_of(&rq->curr->se)); |
732 | } | 732 | } |
733 | 733 | ||
734 | static inline void | 734 | static inline void |
735 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) | 735 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
736 | { | 736 | { |
737 | schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq))); | 737 | schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq))); |
738 | } | 738 | } |
739 | 739 | ||
740 | /* | 740 | /* |
741 | * Task is being enqueued - update stats: | 741 | * Task is being enqueued - update stats: |
742 | */ | 742 | */ |
743 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | 743 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
744 | { | 744 | { |
745 | /* | 745 | /* |
746 | * Are we enqueueing a waiting task? (for current tasks | 746 | * Are we enqueueing a waiting task? (for current tasks |
747 | * a dequeue/enqueue event is a NOP) | 747 | * a dequeue/enqueue event is a NOP) |
748 | */ | 748 | */ |
749 | if (se != cfs_rq->curr) | 749 | if (se != cfs_rq->curr) |
750 | update_stats_wait_start(cfs_rq, se); | 750 | update_stats_wait_start(cfs_rq, se); |
751 | } | 751 | } |
752 | 752 | ||
753 | static void | 753 | static void |
754 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) | 754 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
755 | { | 755 | { |
756 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, | 756 | schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, |
757 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start)); | 757 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start)); |
758 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); | 758 | schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); |
759 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + | 759 | schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + |
760 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); | 760 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); |
761 | #ifdef CONFIG_SCHEDSTATS | 761 | #ifdef CONFIG_SCHEDSTATS |
762 | if (entity_is_task(se)) { | 762 | if (entity_is_task(se)) { |
763 | trace_sched_stat_wait(task_of(se), | 763 | trace_sched_stat_wait(task_of(se), |
764 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); | 764 | rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); |
765 | } | 765 | } |
766 | #endif | 766 | #endif |
767 | schedstat_set(se->statistics.wait_start, 0); | 767 | schedstat_set(se->statistics.wait_start, 0); |
768 | } | 768 | } |
769 | 769 | ||
770 | static inline void | 770 | static inline void |
771 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | 771 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
772 | { | 772 | { |
773 | /* | 773 | /* |
774 | * Mark the end of the wait period if dequeueing a | 774 | * Mark the end of the wait period if dequeueing a |
775 | * waiting task: | 775 | * waiting task: |
776 | */ | 776 | */ |
777 | if (se != cfs_rq->curr) | 777 | if (se != cfs_rq->curr) |
778 | update_stats_wait_end(cfs_rq, se); | 778 | update_stats_wait_end(cfs_rq, se); |
779 | } | 779 | } |
780 | 780 | ||
781 | /* | 781 | /* |
782 | * We are picking a new current task - update its stats: | 782 | * We are picking a new current task - update its stats: |
783 | */ | 783 | */ |
784 | static inline void | 784 | static inline void |
785 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) | 785 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
786 | { | 786 | { |
787 | /* | 787 | /* |
788 | * We are starting a new run period: | 788 | * We are starting a new run period: |
789 | */ | 789 | */ |
790 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); | 790 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
791 | } | 791 | } |
792 | 792 | ||
793 | /************************************************** | 793 | /************************************************** |
794 | * Scheduling class queueing methods: | 794 | * Scheduling class queueing methods: |
795 | */ | 795 | */ |
796 | 796 | ||
797 | #ifdef CONFIG_NUMA_BALANCING | 797 | #ifdef CONFIG_NUMA_BALANCING |
798 | /* | 798 | /* |
799 | * Approximate time to scan a full NUMA task in ms. The task scan period is | 799 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
800 | * calculated based on the tasks virtual memory size and | 800 | * calculated based on the tasks virtual memory size and |
801 | * numa_balancing_scan_size. | 801 | * numa_balancing_scan_size. |
802 | */ | 802 | */ |
803 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; | 803 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
804 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | 804 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; |
805 | 805 | ||
806 | /* Portion of address space to scan in MB */ | 806 | /* Portion of address space to scan in MB */ |
807 | unsigned int sysctl_numa_balancing_scan_size = 256; | 807 | unsigned int sysctl_numa_balancing_scan_size = 256; |
808 | 808 | ||
809 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ | 809 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
810 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | 810 | unsigned int sysctl_numa_balancing_scan_delay = 1000; |
811 | 811 | ||
812 | static unsigned int task_nr_scan_windows(struct task_struct *p) | 812 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
813 | { | 813 | { |
814 | unsigned long rss = 0; | 814 | unsigned long rss = 0; |
815 | unsigned long nr_scan_pages; | 815 | unsigned long nr_scan_pages; |
816 | 816 | ||
817 | /* | 817 | /* |
818 | * Calculations based on RSS as non-present and empty pages are skipped | 818 | * Calculations based on RSS as non-present and empty pages are skipped |
819 | * by the PTE scanner and NUMA hinting faults should be trapped based | 819 | * by the PTE scanner and NUMA hinting faults should be trapped based |
820 | * on resident pages | 820 | * on resident pages |
821 | */ | 821 | */ |
822 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | 822 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); |
823 | rss = get_mm_rss(p->mm); | 823 | rss = get_mm_rss(p->mm); |
824 | if (!rss) | 824 | if (!rss) |
825 | rss = nr_scan_pages; | 825 | rss = nr_scan_pages; |
826 | 826 | ||
827 | rss = round_up(rss, nr_scan_pages); | 827 | rss = round_up(rss, nr_scan_pages); |
828 | return rss / nr_scan_pages; | 828 | return rss / nr_scan_pages; |
829 | } | 829 | } |
830 | 830 | ||
831 | /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ | 831 | /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ |
832 | #define MAX_SCAN_WINDOW 2560 | 832 | #define MAX_SCAN_WINDOW 2560 |
833 | 833 | ||
834 | static unsigned int task_scan_min(struct task_struct *p) | 834 | static unsigned int task_scan_min(struct task_struct *p) |
835 | { | 835 | { |
836 | unsigned int scan_size = ACCESS_ONCE(sysctl_numa_balancing_scan_size); | 836 | unsigned int scan_size = ACCESS_ONCE(sysctl_numa_balancing_scan_size); |
837 | unsigned int scan, floor; | 837 | unsigned int scan, floor; |
838 | unsigned int windows = 1; | 838 | unsigned int windows = 1; |
839 | 839 | ||
840 | if (scan_size < MAX_SCAN_WINDOW) | 840 | if (scan_size < MAX_SCAN_WINDOW) |
841 | windows = MAX_SCAN_WINDOW / scan_size; | 841 | windows = MAX_SCAN_WINDOW / scan_size; |
842 | floor = 1000 / windows; | 842 | floor = 1000 / windows; |
843 | 843 | ||
844 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | 844 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); |
845 | return max_t(unsigned int, floor, scan); | 845 | return max_t(unsigned int, floor, scan); |
846 | } | 846 | } |
847 | 847 | ||
848 | static unsigned int task_scan_max(struct task_struct *p) | 848 | static unsigned int task_scan_max(struct task_struct *p) |
849 | { | 849 | { |
850 | unsigned int smin = task_scan_min(p); | 850 | unsigned int smin = task_scan_min(p); |
851 | unsigned int smax; | 851 | unsigned int smax; |
852 | 852 | ||
853 | /* Watch for min being lower than max due to floor calculations */ | 853 | /* Watch for min being lower than max due to floor calculations */ |
854 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | 854 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); |
855 | return max(smin, smax); | 855 | return max(smin, smax); |
856 | } | 856 | } |
857 | 857 | ||
858 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) | 858 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
859 | { | 859 | { |
860 | rq->nr_numa_running += (p->numa_preferred_nid != -1); | 860 | rq->nr_numa_running += (p->numa_preferred_nid != -1); |
861 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); | 861 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); |
862 | } | 862 | } |
863 | 863 | ||
864 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) | 864 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) |
865 | { | 865 | { |
866 | rq->nr_numa_running -= (p->numa_preferred_nid != -1); | 866 | rq->nr_numa_running -= (p->numa_preferred_nid != -1); |
867 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); | 867 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); |
868 | } | 868 | } |
869 | 869 | ||
870 | struct numa_group { | 870 | struct numa_group { |
871 | atomic_t refcount; | 871 | atomic_t refcount; |
872 | 872 | ||
873 | spinlock_t lock; /* nr_tasks, tasks */ | 873 | spinlock_t lock; /* nr_tasks, tasks */ |
874 | int nr_tasks; | 874 | int nr_tasks; |
875 | pid_t gid; | 875 | pid_t gid; |
876 | 876 | ||
877 | struct rcu_head rcu; | 877 | struct rcu_head rcu; |
878 | nodemask_t active_nodes; | 878 | nodemask_t active_nodes; |
879 | unsigned long total_faults; | 879 | unsigned long total_faults; |
880 | /* | 880 | /* |
881 | * Faults_cpu is used to decide whether memory should move | 881 | * Faults_cpu is used to decide whether memory should move |
882 | * towards the CPU. As a consequence, these stats are weighted | 882 | * towards the CPU. As a consequence, these stats are weighted |
883 | * more by CPU use than by memory faults. | 883 | * more by CPU use than by memory faults. |
884 | */ | 884 | */ |
885 | unsigned long *faults_cpu; | 885 | unsigned long *faults_cpu; |
886 | unsigned long faults[0]; | 886 | unsigned long faults[0]; |
887 | }; | 887 | }; |
888 | 888 | ||
889 | /* Shared or private faults. */ | 889 | /* Shared or private faults. */ |
890 | #define NR_NUMA_HINT_FAULT_TYPES 2 | 890 | #define NR_NUMA_HINT_FAULT_TYPES 2 |
891 | 891 | ||
892 | /* Memory and CPU locality */ | 892 | /* Memory and CPU locality */ |
893 | #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) | 893 | #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) |
894 | 894 | ||
895 | /* Averaged statistics, and temporary buffers. */ | 895 | /* Averaged statistics, and temporary buffers. */ |
896 | #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) | 896 | #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) |
897 | 897 | ||
898 | pid_t task_numa_group_id(struct task_struct *p) | 898 | pid_t task_numa_group_id(struct task_struct *p) |
899 | { | 899 | { |
900 | return p->numa_group ? p->numa_group->gid : 0; | 900 | return p->numa_group ? p->numa_group->gid : 0; |
901 | } | 901 | } |
902 | 902 | ||
903 | /* | 903 | /* |
904 | * The averaged statistics, shared & private, memory & cpu, | 904 | * The averaged statistics, shared & private, memory & cpu, |
905 | * occupy the first half of the array. The second half of the | 905 | * occupy the first half of the array. The second half of the |
906 | * array is for current counters, which are averaged into the | 906 | * array is for current counters, which are averaged into the |
907 | * first set by task_numa_placement. | 907 | * first set by task_numa_placement. |
908 | */ | 908 | */ |
909 | static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) | 909 | static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) |
910 | { | 910 | { |
911 | return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; | 911 | return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; |
912 | } | 912 | } |
913 | 913 | ||
914 | static inline unsigned long task_faults(struct task_struct *p, int nid) | 914 | static inline unsigned long task_faults(struct task_struct *p, int nid) |
915 | { | 915 | { |
916 | if (!p->numa_faults) | 916 | if (!p->numa_faults) |
917 | return 0; | 917 | return 0; |
918 | 918 | ||
919 | return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + | 919 | return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
920 | p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; | 920 | p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; |
921 | } | 921 | } |
922 | 922 | ||
923 | static inline unsigned long group_faults(struct task_struct *p, int nid) | 923 | static inline unsigned long group_faults(struct task_struct *p, int nid) |
924 | { | 924 | { |
925 | if (!p->numa_group) | 925 | if (!p->numa_group) |
926 | return 0; | 926 | return 0; |
927 | 927 | ||
928 | return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] + | 928 | return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
929 | p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)]; | 929 | p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)]; |
930 | } | 930 | } |
931 | 931 | ||
932 | static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) | 932 | static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) |
933 | { | 933 | { |
934 | return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] + | 934 | return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] + |
935 | group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)]; | 935 | group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)]; |
936 | } | 936 | } |
937 | 937 | ||
938 | /* Handle placement on systems where not all nodes are directly connected. */ | 938 | /* Handle placement on systems where not all nodes are directly connected. */ |
939 | static unsigned long score_nearby_nodes(struct task_struct *p, int nid, | 939 | static unsigned long score_nearby_nodes(struct task_struct *p, int nid, |
940 | int maxdist, bool task) | 940 | int maxdist, bool task) |
941 | { | 941 | { |
942 | unsigned long score = 0; | 942 | unsigned long score = 0; |
943 | int node; | 943 | int node; |
944 | 944 | ||
945 | /* | 945 | /* |
946 | * All nodes are directly connected, and the same distance | 946 | * All nodes are directly connected, and the same distance |
947 | * from each other. No need for fancy placement algorithms. | 947 | * from each other. No need for fancy placement algorithms. |
948 | */ | 948 | */ |
949 | if (sched_numa_topology_type == NUMA_DIRECT) | 949 | if (sched_numa_topology_type == NUMA_DIRECT) |
950 | return 0; | 950 | return 0; |
951 | 951 | ||
952 | /* | 952 | /* |
953 | * This code is called for each node, introducing N^2 complexity, | 953 | * This code is called for each node, introducing N^2 complexity, |
954 | * which should be ok given the number of nodes rarely exceeds 8. | 954 | * which should be ok given the number of nodes rarely exceeds 8. |
955 | */ | 955 | */ |
956 | for_each_online_node(node) { | 956 | for_each_online_node(node) { |
957 | unsigned long faults; | 957 | unsigned long faults; |
958 | int dist = node_distance(nid, node); | 958 | int dist = node_distance(nid, node); |
959 | 959 | ||
960 | /* | 960 | /* |
961 | * The furthest away nodes in the system are not interesting | 961 | * The furthest away nodes in the system are not interesting |
962 | * for placement; nid was already counted. | 962 | * for placement; nid was already counted. |
963 | */ | 963 | */ |
964 | if (dist == sched_max_numa_distance || node == nid) | 964 | if (dist == sched_max_numa_distance || node == nid) |
965 | continue; | 965 | continue; |
966 | 966 | ||
967 | /* | 967 | /* |
968 | * On systems with a backplane NUMA topology, compare groups | 968 | * On systems with a backplane NUMA topology, compare groups |
969 | * of nodes, and move tasks towards the group with the most | 969 | * of nodes, and move tasks towards the group with the most |
970 | * memory accesses. When comparing two nodes at distance | 970 | * memory accesses. When comparing two nodes at distance |
971 | * "hoplimit", only nodes closer by than "hoplimit" are part | 971 | * "hoplimit", only nodes closer by than "hoplimit" are part |
972 | * of each group. Skip other nodes. | 972 | * of each group. Skip other nodes. |
973 | */ | 973 | */ |
974 | if (sched_numa_topology_type == NUMA_BACKPLANE && | 974 | if (sched_numa_topology_type == NUMA_BACKPLANE && |
975 | dist > maxdist) | 975 | dist > maxdist) |
976 | continue; | 976 | continue; |
977 | 977 | ||
978 | /* Add up the faults from nearby nodes. */ | 978 | /* Add up the faults from nearby nodes. */ |
979 | if (task) | 979 | if (task) |
980 | faults = task_faults(p, node); | 980 | faults = task_faults(p, node); |
981 | else | 981 | else |
982 | faults = group_faults(p, node); | 982 | faults = group_faults(p, node); |
983 | 983 | ||
984 | /* | 984 | /* |
985 | * On systems with a glueless mesh NUMA topology, there are | 985 | * On systems with a glueless mesh NUMA topology, there are |
986 | * no fixed "groups of nodes". Instead, nodes that are not | 986 | * no fixed "groups of nodes". Instead, nodes that are not |
987 | * directly connected bounce traffic through intermediate | 987 | * directly connected bounce traffic through intermediate |
988 | * nodes; a numa_group can occupy any set of nodes. | 988 | * nodes; a numa_group can occupy any set of nodes. |
989 | * The further away a node is, the less the faults count. | 989 | * The further away a node is, the less the faults count. |
990 | * This seems to result in good task placement. | 990 | * This seems to result in good task placement. |
991 | */ | 991 | */ |
992 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | 992 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { |
993 | faults *= (sched_max_numa_distance - dist); | 993 | faults *= (sched_max_numa_distance - dist); |
994 | faults /= (sched_max_numa_distance - LOCAL_DISTANCE); | 994 | faults /= (sched_max_numa_distance - LOCAL_DISTANCE); |
995 | } | 995 | } |
996 | 996 | ||
997 | score += faults; | 997 | score += faults; |
998 | } | 998 | } |
999 | 999 | ||
1000 | return score; | 1000 | return score; |
1001 | } | 1001 | } |
1002 | 1002 | ||
1003 | /* | 1003 | /* |
1004 | * These return the fraction of accesses done by a particular task, or | 1004 | * These return the fraction of accesses done by a particular task, or |
1005 | * task group, on a particular numa node. The group weight is given a | 1005 | * task group, on a particular numa node. The group weight is given a |
1006 | * larger multiplier, in order to group tasks together that are almost | 1006 | * larger multiplier, in order to group tasks together that are almost |
1007 | * evenly spread out between numa nodes. | 1007 | * evenly spread out between numa nodes. |
1008 | */ | 1008 | */ |
1009 | static inline unsigned long task_weight(struct task_struct *p, int nid, | 1009 | static inline unsigned long task_weight(struct task_struct *p, int nid, |
1010 | int dist) | 1010 | int dist) |
1011 | { | 1011 | { |
1012 | unsigned long faults, total_faults; | 1012 | unsigned long faults, total_faults; |
1013 | 1013 | ||
1014 | if (!p->numa_faults) | 1014 | if (!p->numa_faults) |
1015 | return 0; | 1015 | return 0; |
1016 | 1016 | ||
1017 | total_faults = p->total_numa_faults; | 1017 | total_faults = p->total_numa_faults; |
1018 | 1018 | ||
1019 | if (!total_faults) | 1019 | if (!total_faults) |
1020 | return 0; | 1020 | return 0; |
1021 | 1021 | ||
1022 | faults = task_faults(p, nid); | 1022 | faults = task_faults(p, nid); |
1023 | faults += score_nearby_nodes(p, nid, dist, true); | 1023 | faults += score_nearby_nodes(p, nid, dist, true); |
1024 | 1024 | ||
1025 | return 1000 * faults / total_faults; | 1025 | return 1000 * faults / total_faults; |
1026 | } | 1026 | } |
1027 | 1027 | ||
1028 | static inline unsigned long group_weight(struct task_struct *p, int nid, | 1028 | static inline unsigned long group_weight(struct task_struct *p, int nid, |
1029 | int dist) | 1029 | int dist) |
1030 | { | 1030 | { |
1031 | unsigned long faults, total_faults; | 1031 | unsigned long faults, total_faults; |
1032 | 1032 | ||
1033 | if (!p->numa_group) | 1033 | if (!p->numa_group) |
1034 | return 0; | 1034 | return 0; |
1035 | 1035 | ||
1036 | total_faults = p->numa_group->total_faults; | 1036 | total_faults = p->numa_group->total_faults; |
1037 | 1037 | ||
1038 | if (!total_faults) | 1038 | if (!total_faults) |
1039 | return 0; | 1039 | return 0; |
1040 | 1040 | ||
1041 | faults = group_faults(p, nid); | 1041 | faults = group_faults(p, nid); |
1042 | faults += score_nearby_nodes(p, nid, dist, false); | 1042 | faults += score_nearby_nodes(p, nid, dist, false); |
1043 | 1043 | ||
1044 | return 1000 * faults / total_faults; | 1044 | return 1000 * faults / total_faults; |
1045 | } | 1045 | } |
1046 | 1046 | ||
1047 | bool should_numa_migrate_memory(struct task_struct *p, struct page * page, | 1047 | bool should_numa_migrate_memory(struct task_struct *p, struct page * page, |
1048 | int src_nid, int dst_cpu) | 1048 | int src_nid, int dst_cpu) |
1049 | { | 1049 | { |
1050 | struct numa_group *ng = p->numa_group; | 1050 | struct numa_group *ng = p->numa_group; |
1051 | int dst_nid = cpu_to_node(dst_cpu); | 1051 | int dst_nid = cpu_to_node(dst_cpu); |
1052 | int last_cpupid, this_cpupid; | 1052 | int last_cpupid, this_cpupid; |
1053 | 1053 | ||
1054 | this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); | 1054 | this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); |
1055 | 1055 | ||
1056 | /* | 1056 | /* |
1057 | * Multi-stage node selection is used in conjunction with a periodic | 1057 | * Multi-stage node selection is used in conjunction with a periodic |
1058 | * migration fault to build a temporal task<->page relation. By using | 1058 | * migration fault to build a temporal task<->page relation. By using |
1059 | * a two-stage filter we remove short/unlikely relations. | 1059 | * a two-stage filter we remove short/unlikely relations. |
1060 | * | 1060 | * |
1061 | * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate | 1061 | * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate |
1062 | * a task's usage of a particular page (n_p) per total usage of this | 1062 | * a task's usage of a particular page (n_p) per total usage of this |
1063 | * page (n_t) (in a given time-span) to a probability. | 1063 | * page (n_t) (in a given time-span) to a probability. |
1064 | * | 1064 | * |
1065 | * Our periodic faults will sample this probability and getting the | 1065 | * Our periodic faults will sample this probability and getting the |
1066 | * same result twice in a row, given these samples are fully | 1066 | * same result twice in a row, given these samples are fully |
1067 | * independent, is then given by P(n)^2, provided our sample period | 1067 | * independent, is then given by P(n)^2, provided our sample period |
1068 | * is sufficiently short compared to the usage pattern. | 1068 | * is sufficiently short compared to the usage pattern. |
1069 | * | 1069 | * |
1070 | * This quadric squishes small probabilities, making it less likely we | 1070 | * This quadric squishes small probabilities, making it less likely we |
1071 | * act on an unlikely task<->page relation. | 1071 | * act on an unlikely task<->page relation. |
1072 | */ | 1072 | */ |
1073 | last_cpupid = page_cpupid_xchg_last(page, this_cpupid); | 1073 | last_cpupid = page_cpupid_xchg_last(page, this_cpupid); |
1074 | if (!cpupid_pid_unset(last_cpupid) && | 1074 | if (!cpupid_pid_unset(last_cpupid) && |
1075 | cpupid_to_nid(last_cpupid) != dst_nid) | 1075 | cpupid_to_nid(last_cpupid) != dst_nid) |
1076 | return false; | 1076 | return false; |
1077 | 1077 | ||
1078 | /* Always allow migrate on private faults */ | 1078 | /* Always allow migrate on private faults */ |
1079 | if (cpupid_match_pid(p, last_cpupid)) | 1079 | if (cpupid_match_pid(p, last_cpupid)) |
1080 | return true; | 1080 | return true; |
1081 | 1081 | ||
1082 | /* A shared fault, but p->numa_group has not been set up yet. */ | 1082 | /* A shared fault, but p->numa_group has not been set up yet. */ |
1083 | if (!ng) | 1083 | if (!ng) |
1084 | return true; | 1084 | return true; |
1085 | 1085 | ||
1086 | /* | 1086 | /* |
1087 | * Do not migrate if the destination is not a node that | 1087 | * Do not migrate if the destination is not a node that |
1088 | * is actively used by this numa group. | 1088 | * is actively used by this numa group. |
1089 | */ | 1089 | */ |
1090 | if (!node_isset(dst_nid, ng->active_nodes)) | 1090 | if (!node_isset(dst_nid, ng->active_nodes)) |
1091 | return false; | 1091 | return false; |
1092 | 1092 | ||
1093 | /* | 1093 | /* |
1094 | * Source is a node that is not actively used by this | 1094 | * Source is a node that is not actively used by this |
1095 | * numa group, while the destination is. Migrate. | 1095 | * numa group, while the destination is. Migrate. |
1096 | */ | 1096 | */ |
1097 | if (!node_isset(src_nid, ng->active_nodes)) | 1097 | if (!node_isset(src_nid, ng->active_nodes)) |
1098 | return true; | 1098 | return true; |
1099 | 1099 | ||
1100 | /* | 1100 | /* |
1101 | * Both source and destination are nodes in active | 1101 | * Both source and destination are nodes in active |
1102 | * use by this numa group. Maximize memory bandwidth | 1102 | * use by this numa group. Maximize memory bandwidth |
1103 | * by migrating from more heavily used groups, to less | 1103 | * by migrating from more heavily used groups, to less |
1104 | * heavily used ones, spreading the load around. | 1104 | * heavily used ones, spreading the load around. |
1105 | * Use a 1/4 hysteresis to avoid spurious page movement. | 1105 | * Use a 1/4 hysteresis to avoid spurious page movement. |
1106 | */ | 1106 | */ |
1107 | return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4); | 1107 | return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4); |
1108 | } | 1108 | } |
1109 | 1109 | ||
1110 | static unsigned long weighted_cpuload(const int cpu); | 1110 | static unsigned long weighted_cpuload(const int cpu); |
1111 | static unsigned long source_load(int cpu, int type); | 1111 | static unsigned long source_load(int cpu, int type); |
1112 | static unsigned long target_load(int cpu, int type); | 1112 | static unsigned long target_load(int cpu, int type); |
1113 | static unsigned long capacity_of(int cpu); | 1113 | static unsigned long capacity_of(int cpu); |
1114 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg); | 1114 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg); |
1115 | 1115 | ||
1116 | /* Cached statistics for all CPUs within a node */ | 1116 | /* Cached statistics for all CPUs within a node */ |
1117 | struct numa_stats { | 1117 | struct numa_stats { |
1118 | unsigned long nr_running; | 1118 | unsigned long nr_running; |
1119 | unsigned long load; | 1119 | unsigned long load; |
1120 | 1120 | ||
1121 | /* Total compute capacity of CPUs on a node */ | 1121 | /* Total compute capacity of CPUs on a node */ |
1122 | unsigned long compute_capacity; | 1122 | unsigned long compute_capacity; |
1123 | 1123 | ||
1124 | /* Approximate capacity in terms of runnable tasks on a node */ | 1124 | /* Approximate capacity in terms of runnable tasks on a node */ |
1125 | unsigned long task_capacity; | 1125 | unsigned long task_capacity; |
1126 | int has_free_capacity; | 1126 | int has_free_capacity; |
1127 | }; | 1127 | }; |
1128 | 1128 | ||
1129 | /* | 1129 | /* |
1130 | * XXX borrowed from update_sg_lb_stats | 1130 | * XXX borrowed from update_sg_lb_stats |
1131 | */ | 1131 | */ |
1132 | static void update_numa_stats(struct numa_stats *ns, int nid) | 1132 | static void update_numa_stats(struct numa_stats *ns, int nid) |
1133 | { | 1133 | { |
1134 | int smt, cpu, cpus = 0; | 1134 | int smt, cpu, cpus = 0; |
1135 | unsigned long capacity; | 1135 | unsigned long capacity; |
1136 | 1136 | ||
1137 | memset(ns, 0, sizeof(*ns)); | 1137 | memset(ns, 0, sizeof(*ns)); |
1138 | for_each_cpu(cpu, cpumask_of_node(nid)) { | 1138 | for_each_cpu(cpu, cpumask_of_node(nid)) { |
1139 | struct rq *rq = cpu_rq(cpu); | 1139 | struct rq *rq = cpu_rq(cpu); |
1140 | 1140 | ||
1141 | ns->nr_running += rq->nr_running; | 1141 | ns->nr_running += rq->nr_running; |
1142 | ns->load += weighted_cpuload(cpu); | 1142 | ns->load += weighted_cpuload(cpu); |
1143 | ns->compute_capacity += capacity_of(cpu); | 1143 | ns->compute_capacity += capacity_of(cpu); |
1144 | 1144 | ||
1145 | cpus++; | 1145 | cpus++; |
1146 | } | 1146 | } |
1147 | 1147 | ||
1148 | /* | 1148 | /* |
1149 | * If we raced with hotplug and there are no CPUs left in our mask | 1149 | * If we raced with hotplug and there are no CPUs left in our mask |
1150 | * the @ns structure is NULL'ed and task_numa_compare() will | 1150 | * the @ns structure is NULL'ed and task_numa_compare() will |
1151 | * not find this node attractive. | 1151 | * not find this node attractive. |
1152 | * | 1152 | * |
1153 | * We'll either bail at !has_free_capacity, or we'll detect a huge | 1153 | * We'll either bail at !has_free_capacity, or we'll detect a huge |
1154 | * imbalance and bail there. | 1154 | * imbalance and bail there. |
1155 | */ | 1155 | */ |
1156 | if (!cpus) | 1156 | if (!cpus) |
1157 | return; | 1157 | return; |
1158 | 1158 | ||
1159 | /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */ | 1159 | /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */ |
1160 | smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity); | 1160 | smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity); |
1161 | capacity = cpus / smt; /* cores */ | 1161 | capacity = cpus / smt; /* cores */ |
1162 | 1162 | ||
1163 | ns->task_capacity = min_t(unsigned, capacity, | 1163 | ns->task_capacity = min_t(unsigned, capacity, |
1164 | DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE)); | 1164 | DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE)); |
1165 | ns->has_free_capacity = (ns->nr_running < ns->task_capacity); | 1165 | ns->has_free_capacity = (ns->nr_running < ns->task_capacity); |
1166 | } | 1166 | } |
1167 | 1167 | ||
1168 | struct task_numa_env { | 1168 | struct task_numa_env { |
1169 | struct task_struct *p; | 1169 | struct task_struct *p; |
1170 | 1170 | ||
1171 | int src_cpu, src_nid; | 1171 | int src_cpu, src_nid; |
1172 | int dst_cpu, dst_nid; | 1172 | int dst_cpu, dst_nid; |
1173 | 1173 | ||
1174 | struct numa_stats src_stats, dst_stats; | 1174 | struct numa_stats src_stats, dst_stats; |
1175 | 1175 | ||
1176 | int imbalance_pct; | 1176 | int imbalance_pct; |
1177 | int dist; | 1177 | int dist; |
1178 | 1178 | ||
1179 | struct task_struct *best_task; | 1179 | struct task_struct *best_task; |
1180 | long best_imp; | 1180 | long best_imp; |
1181 | int best_cpu; | 1181 | int best_cpu; |
1182 | }; | 1182 | }; |
1183 | 1183 | ||
1184 | static void task_numa_assign(struct task_numa_env *env, | 1184 | static void task_numa_assign(struct task_numa_env *env, |
1185 | struct task_struct *p, long imp) | 1185 | struct task_struct *p, long imp) |
1186 | { | 1186 | { |
1187 | if (env->best_task) | 1187 | if (env->best_task) |
1188 | put_task_struct(env->best_task); | 1188 | put_task_struct(env->best_task); |
1189 | if (p) | 1189 | if (p) |
1190 | get_task_struct(p); | 1190 | get_task_struct(p); |
1191 | 1191 | ||
1192 | env->best_task = p; | 1192 | env->best_task = p; |
1193 | env->best_imp = imp; | 1193 | env->best_imp = imp; |
1194 | env->best_cpu = env->dst_cpu; | 1194 | env->best_cpu = env->dst_cpu; |
1195 | } | 1195 | } |
1196 | 1196 | ||
1197 | static bool load_too_imbalanced(long src_load, long dst_load, | 1197 | static bool load_too_imbalanced(long src_load, long dst_load, |
1198 | struct task_numa_env *env) | 1198 | struct task_numa_env *env) |
1199 | { | 1199 | { |
1200 | long imb, old_imb; | 1200 | long imb, old_imb; |
1201 | long orig_src_load, orig_dst_load; | 1201 | long orig_src_load, orig_dst_load; |
1202 | long src_capacity, dst_capacity; | 1202 | long src_capacity, dst_capacity; |
1203 | 1203 | ||
1204 | /* | 1204 | /* |
1205 | * The load is corrected for the CPU capacity available on each node. | 1205 | * The load is corrected for the CPU capacity available on each node. |
1206 | * | 1206 | * |
1207 | * src_load dst_load | 1207 | * src_load dst_load |
1208 | * ------------ vs --------- | 1208 | * ------------ vs --------- |
1209 | * src_capacity dst_capacity | 1209 | * src_capacity dst_capacity |
1210 | */ | 1210 | */ |
1211 | src_capacity = env->src_stats.compute_capacity; | 1211 | src_capacity = env->src_stats.compute_capacity; |
1212 | dst_capacity = env->dst_stats.compute_capacity; | 1212 | dst_capacity = env->dst_stats.compute_capacity; |
1213 | 1213 | ||
1214 | /* We care about the slope of the imbalance, not the direction. */ | 1214 | /* We care about the slope of the imbalance, not the direction. */ |
1215 | if (dst_load < src_load) | 1215 | if (dst_load < src_load) |
1216 | swap(dst_load, src_load); | 1216 | swap(dst_load, src_load); |
1217 | 1217 | ||
1218 | /* Is the difference below the threshold? */ | 1218 | /* Is the difference below the threshold? */ |
1219 | imb = dst_load * src_capacity * 100 - | 1219 | imb = dst_load * src_capacity * 100 - |
1220 | src_load * dst_capacity * env->imbalance_pct; | 1220 | src_load * dst_capacity * env->imbalance_pct; |
1221 | if (imb <= 0) | 1221 | if (imb <= 0) |
1222 | return false; | 1222 | return false; |
1223 | 1223 | ||
1224 | /* | 1224 | /* |
1225 | * The imbalance is above the allowed threshold. | 1225 | * The imbalance is above the allowed threshold. |
1226 | * Compare it with the old imbalance. | 1226 | * Compare it with the old imbalance. |
1227 | */ | 1227 | */ |
1228 | orig_src_load = env->src_stats.load; | 1228 | orig_src_load = env->src_stats.load; |
1229 | orig_dst_load = env->dst_stats.load; | 1229 | orig_dst_load = env->dst_stats.load; |
1230 | 1230 | ||
1231 | if (orig_dst_load < orig_src_load) | 1231 | if (orig_dst_load < orig_src_load) |
1232 | swap(orig_dst_load, orig_src_load); | 1232 | swap(orig_dst_load, orig_src_load); |
1233 | 1233 | ||
1234 | old_imb = orig_dst_load * src_capacity * 100 - | 1234 | old_imb = orig_dst_load * src_capacity * 100 - |
1235 | orig_src_load * dst_capacity * env->imbalance_pct; | 1235 | orig_src_load * dst_capacity * env->imbalance_pct; |
1236 | 1236 | ||
1237 | /* Would this change make things worse? */ | 1237 | /* Would this change make things worse? */ |
1238 | return (imb > old_imb); | 1238 | return (imb > old_imb); |
1239 | } | 1239 | } |
1240 | 1240 | ||
1241 | /* | 1241 | /* |
1242 | * This checks if the overall compute and NUMA accesses of the system would | 1242 | * This checks if the overall compute and NUMA accesses of the system would |
1243 | * be improved if the source tasks was migrated to the target dst_cpu taking | 1243 | * be improved if the source tasks was migrated to the target dst_cpu taking |
1244 | * into account that it might be best if task running on the dst_cpu should | 1244 | * into account that it might be best if task running on the dst_cpu should |
1245 | * be exchanged with the source task | 1245 | * be exchanged with the source task |
1246 | */ | 1246 | */ |
1247 | static void task_numa_compare(struct task_numa_env *env, | 1247 | static void task_numa_compare(struct task_numa_env *env, |
1248 | long taskimp, long groupimp) | 1248 | long taskimp, long groupimp) |
1249 | { | 1249 | { |
1250 | struct rq *src_rq = cpu_rq(env->src_cpu); | 1250 | struct rq *src_rq = cpu_rq(env->src_cpu); |
1251 | struct rq *dst_rq = cpu_rq(env->dst_cpu); | 1251 | struct rq *dst_rq = cpu_rq(env->dst_cpu); |
1252 | struct task_struct *cur; | 1252 | struct task_struct *cur; |
1253 | long src_load, dst_load; | 1253 | long src_load, dst_load; |
1254 | long load; | 1254 | long load; |
1255 | long imp = env->p->numa_group ? groupimp : taskimp; | 1255 | long imp = env->p->numa_group ? groupimp : taskimp; |
1256 | long moveimp = imp; | 1256 | long moveimp = imp; |
1257 | int dist = env->dist; | 1257 | int dist = env->dist; |
1258 | 1258 | ||
1259 | rcu_read_lock(); | 1259 | rcu_read_lock(); |
1260 | 1260 | ||
1261 | raw_spin_lock_irq(&dst_rq->lock); | 1261 | raw_spin_lock_irq(&dst_rq->lock); |
1262 | cur = dst_rq->curr; | 1262 | cur = dst_rq->curr; |
1263 | /* | 1263 | /* |
1264 | * No need to move the exiting task, and this ensures that ->curr | 1264 | * No need to move the exiting task, and this ensures that ->curr |
1265 | * wasn't reaped and thus get_task_struct() in task_numa_assign() | 1265 | * wasn't reaped and thus get_task_struct() in task_numa_assign() |
1266 | * is safe under RCU read lock. | 1266 | * is safe under RCU read lock. |
1267 | * Note that rcu_read_lock() itself can't protect from the final | 1267 | * Note that rcu_read_lock() itself can't protect from the final |
1268 | * put_task_struct() after the last schedule(). | 1268 | * put_task_struct() after the last schedule(). |
1269 | */ | 1269 | */ |
1270 | if ((cur->flags & PF_EXITING) || is_idle_task(cur)) | 1270 | if ((cur->flags & PF_EXITING) || is_idle_task(cur)) |
1271 | cur = NULL; | 1271 | cur = NULL; |
1272 | raw_spin_unlock_irq(&dst_rq->lock); | 1272 | raw_spin_unlock_irq(&dst_rq->lock); |
1273 | 1273 | ||
1274 | /* | 1274 | /* |
1275 | * Because we have preemption enabled we can get migrated around and | 1275 | * Because we have preemption enabled we can get migrated around and |
1276 | * end try selecting ourselves (current == env->p) as a swap candidate. | 1276 | * end try selecting ourselves (current == env->p) as a swap candidate. |
1277 | */ | 1277 | */ |
1278 | if (cur == env->p) | 1278 | if (cur == env->p) |
1279 | goto unlock; | 1279 | goto unlock; |
1280 | 1280 | ||
1281 | /* | 1281 | /* |
1282 | * "imp" is the fault differential for the source task between the | 1282 | * "imp" is the fault differential for the source task between the |
1283 | * source and destination node. Calculate the total differential for | 1283 | * source and destination node. Calculate the total differential for |
1284 | * the source task and potential destination task. The more negative | 1284 | * the source task and potential destination task. The more negative |
1285 | * the value is, the more rmeote accesses that would be expected to | 1285 | * the value is, the more rmeote accesses that would be expected to |
1286 | * be incurred if the tasks were swapped. | 1286 | * be incurred if the tasks were swapped. |
1287 | */ | 1287 | */ |
1288 | if (cur) { | 1288 | if (cur) { |
1289 | /* Skip this swap candidate if cannot move to the source cpu */ | 1289 | /* Skip this swap candidate if cannot move to the source cpu */ |
1290 | if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur))) | 1290 | if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur))) |
1291 | goto unlock; | 1291 | goto unlock; |
1292 | 1292 | ||
1293 | /* | 1293 | /* |
1294 | * If dst and source tasks are in the same NUMA group, or not | 1294 | * If dst and source tasks are in the same NUMA group, or not |
1295 | * in any group then look only at task weights. | 1295 | * in any group then look only at task weights. |
1296 | */ | 1296 | */ |
1297 | if (cur->numa_group == env->p->numa_group) { | 1297 | if (cur->numa_group == env->p->numa_group) { |
1298 | imp = taskimp + task_weight(cur, env->src_nid, dist) - | 1298 | imp = taskimp + task_weight(cur, env->src_nid, dist) - |
1299 | task_weight(cur, env->dst_nid, dist); | 1299 | task_weight(cur, env->dst_nid, dist); |
1300 | /* | 1300 | /* |
1301 | * Add some hysteresis to prevent swapping the | 1301 | * Add some hysteresis to prevent swapping the |
1302 | * tasks within a group over tiny differences. | 1302 | * tasks within a group over tiny differences. |
1303 | */ | 1303 | */ |
1304 | if (cur->numa_group) | 1304 | if (cur->numa_group) |
1305 | imp -= imp/16; | 1305 | imp -= imp/16; |
1306 | } else { | 1306 | } else { |
1307 | /* | 1307 | /* |
1308 | * Compare the group weights. If a task is all by | 1308 | * Compare the group weights. If a task is all by |
1309 | * itself (not part of a group), use the task weight | 1309 | * itself (not part of a group), use the task weight |
1310 | * instead. | 1310 | * instead. |
1311 | */ | 1311 | */ |
1312 | if (cur->numa_group) | 1312 | if (cur->numa_group) |
1313 | imp += group_weight(cur, env->src_nid, dist) - | 1313 | imp += group_weight(cur, env->src_nid, dist) - |
1314 | group_weight(cur, env->dst_nid, dist); | 1314 | group_weight(cur, env->dst_nid, dist); |
1315 | else | 1315 | else |
1316 | imp += task_weight(cur, env->src_nid, dist) - | 1316 | imp += task_weight(cur, env->src_nid, dist) - |
1317 | task_weight(cur, env->dst_nid, dist); | 1317 | task_weight(cur, env->dst_nid, dist); |
1318 | } | 1318 | } |
1319 | } | 1319 | } |
1320 | 1320 | ||
1321 | if (imp <= env->best_imp && moveimp <= env->best_imp) | 1321 | if (imp <= env->best_imp && moveimp <= env->best_imp) |
1322 | goto unlock; | 1322 | goto unlock; |
1323 | 1323 | ||
1324 | if (!cur) { | 1324 | if (!cur) { |
1325 | /* Is there capacity at our destination? */ | 1325 | /* Is there capacity at our destination? */ |
1326 | if (env->src_stats.nr_running <= env->src_stats.task_capacity && | 1326 | if (env->src_stats.nr_running <= env->src_stats.task_capacity && |
1327 | !env->dst_stats.has_free_capacity) | 1327 | !env->dst_stats.has_free_capacity) |
1328 | goto unlock; | 1328 | goto unlock; |
1329 | 1329 | ||
1330 | goto balance; | 1330 | goto balance; |
1331 | } | 1331 | } |
1332 | 1332 | ||
1333 | /* Balance doesn't matter much if we're running a task per cpu */ | 1333 | /* Balance doesn't matter much if we're running a task per cpu */ |
1334 | if (imp > env->best_imp && src_rq->nr_running == 1 && | 1334 | if (imp > env->best_imp && src_rq->nr_running == 1 && |
1335 | dst_rq->nr_running == 1) | 1335 | dst_rq->nr_running == 1) |
1336 | goto assign; | 1336 | goto assign; |
1337 | 1337 | ||
1338 | /* | 1338 | /* |
1339 | * In the overloaded case, try and keep the load balanced. | 1339 | * In the overloaded case, try and keep the load balanced. |
1340 | */ | 1340 | */ |
1341 | balance: | 1341 | balance: |
1342 | load = task_h_load(env->p); | 1342 | load = task_h_load(env->p); |
1343 | dst_load = env->dst_stats.load + load; | 1343 | dst_load = env->dst_stats.load + load; |
1344 | src_load = env->src_stats.load - load; | 1344 | src_load = env->src_stats.load - load; |
1345 | 1345 | ||
1346 | if (moveimp > imp && moveimp > env->best_imp) { | 1346 | if (moveimp > imp && moveimp > env->best_imp) { |
1347 | /* | 1347 | /* |
1348 | * If the improvement from just moving env->p direction is | 1348 | * If the improvement from just moving env->p direction is |
1349 | * better than swapping tasks around, check if a move is | 1349 | * better than swapping tasks around, check if a move is |
1350 | * possible. Store a slightly smaller score than moveimp, | 1350 | * possible. Store a slightly smaller score than moveimp, |
1351 | * so an actually idle CPU will win. | 1351 | * so an actually idle CPU will win. |
1352 | */ | 1352 | */ |
1353 | if (!load_too_imbalanced(src_load, dst_load, env)) { | 1353 | if (!load_too_imbalanced(src_load, dst_load, env)) { |
1354 | imp = moveimp - 1; | 1354 | imp = moveimp - 1; |
1355 | cur = NULL; | 1355 | cur = NULL; |
1356 | goto assign; | 1356 | goto assign; |
1357 | } | 1357 | } |
1358 | } | 1358 | } |
1359 | 1359 | ||
1360 | if (imp <= env->best_imp) | 1360 | if (imp <= env->best_imp) |
1361 | goto unlock; | 1361 | goto unlock; |
1362 | 1362 | ||
1363 | if (cur) { | 1363 | if (cur) { |
1364 | load = task_h_load(cur); | 1364 | load = task_h_load(cur); |
1365 | dst_load -= load; | 1365 | dst_load -= load; |
1366 | src_load += load; | 1366 | src_load += load; |
1367 | } | 1367 | } |
1368 | 1368 | ||
1369 | if (load_too_imbalanced(src_load, dst_load, env)) | 1369 | if (load_too_imbalanced(src_load, dst_load, env)) |
1370 | goto unlock; | 1370 | goto unlock; |
1371 | 1371 | ||
1372 | /* | 1372 | /* |
1373 | * One idle CPU per node is evaluated for a task numa move. | 1373 | * One idle CPU per node is evaluated for a task numa move. |
1374 | * Call select_idle_sibling to maybe find a better one. | 1374 | * Call select_idle_sibling to maybe find a better one. |
1375 | */ | 1375 | */ |
1376 | if (!cur) | 1376 | if (!cur) |
1377 | env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu); | 1377 | env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu); |
1378 | 1378 | ||
1379 | assign: | 1379 | assign: |
1380 | task_numa_assign(env, cur, imp); | 1380 | task_numa_assign(env, cur, imp); |
1381 | unlock: | 1381 | unlock: |
1382 | rcu_read_unlock(); | 1382 | rcu_read_unlock(); |
1383 | } | 1383 | } |
1384 | 1384 | ||
1385 | static void task_numa_find_cpu(struct task_numa_env *env, | 1385 | static void task_numa_find_cpu(struct task_numa_env *env, |
1386 | long taskimp, long groupimp) | 1386 | long taskimp, long groupimp) |
1387 | { | 1387 | { |
1388 | int cpu; | 1388 | int cpu; |
1389 | 1389 | ||
1390 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { | 1390 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { |
1391 | /* Skip this CPU if the source task cannot migrate */ | 1391 | /* Skip this CPU if the source task cannot migrate */ |
1392 | if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p))) | 1392 | if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p))) |
1393 | continue; | 1393 | continue; |
1394 | 1394 | ||
1395 | env->dst_cpu = cpu; | 1395 | env->dst_cpu = cpu; |
1396 | task_numa_compare(env, taskimp, groupimp); | 1396 | task_numa_compare(env, taskimp, groupimp); |
1397 | } | 1397 | } |
1398 | } | 1398 | } |
1399 | 1399 | ||
1400 | static int task_numa_migrate(struct task_struct *p) | 1400 | static int task_numa_migrate(struct task_struct *p) |
1401 | { | 1401 | { |
1402 | struct task_numa_env env = { | 1402 | struct task_numa_env env = { |
1403 | .p = p, | 1403 | .p = p, |
1404 | 1404 | ||
1405 | .src_cpu = task_cpu(p), | 1405 | .src_cpu = task_cpu(p), |
1406 | .src_nid = task_node(p), | 1406 | .src_nid = task_node(p), |
1407 | 1407 | ||
1408 | .imbalance_pct = 112, | 1408 | .imbalance_pct = 112, |
1409 | 1409 | ||
1410 | .best_task = NULL, | 1410 | .best_task = NULL, |
1411 | .best_imp = 0, | 1411 | .best_imp = 0, |
1412 | .best_cpu = -1 | 1412 | .best_cpu = -1 |
1413 | }; | 1413 | }; |
1414 | struct sched_domain *sd; | 1414 | struct sched_domain *sd; |
1415 | unsigned long taskweight, groupweight; | 1415 | unsigned long taskweight, groupweight; |
1416 | int nid, ret, dist; | 1416 | int nid, ret, dist; |
1417 | long taskimp, groupimp; | 1417 | long taskimp, groupimp; |
1418 | 1418 | ||
1419 | /* | 1419 | /* |
1420 | * Pick the lowest SD_NUMA domain, as that would have the smallest | 1420 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
1421 | * imbalance and would be the first to start moving tasks about. | 1421 | * imbalance and would be the first to start moving tasks about. |
1422 | * | 1422 | * |
1423 | * And we want to avoid any moving of tasks about, as that would create | 1423 | * And we want to avoid any moving of tasks about, as that would create |
1424 | * random movement of tasks -- counter the numa conditions we're trying | 1424 | * random movement of tasks -- counter the numa conditions we're trying |
1425 | * to satisfy here. | 1425 | * to satisfy here. |
1426 | */ | 1426 | */ |
1427 | rcu_read_lock(); | 1427 | rcu_read_lock(); |
1428 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); | 1428 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
1429 | if (sd) | 1429 | if (sd) |
1430 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; | 1430 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; |
1431 | rcu_read_unlock(); | 1431 | rcu_read_unlock(); |
1432 | 1432 | ||
1433 | /* | 1433 | /* |
1434 | * Cpusets can break the scheduler domain tree into smaller | 1434 | * Cpusets can break the scheduler domain tree into smaller |
1435 | * balance domains, some of which do not cross NUMA boundaries. | 1435 | * balance domains, some of which do not cross NUMA boundaries. |
1436 | * Tasks that are "trapped" in such domains cannot be migrated | 1436 | * Tasks that are "trapped" in such domains cannot be migrated |
1437 | * elsewhere, so there is no point in (re)trying. | 1437 | * elsewhere, so there is no point in (re)trying. |
1438 | */ | 1438 | */ |
1439 | if (unlikely(!sd)) { | 1439 | if (unlikely(!sd)) { |
1440 | p->numa_preferred_nid = task_node(p); | 1440 | p->numa_preferred_nid = task_node(p); |
1441 | return -EINVAL; | 1441 | return -EINVAL; |
1442 | } | 1442 | } |
1443 | 1443 | ||
1444 | env.dst_nid = p->numa_preferred_nid; | 1444 | env.dst_nid = p->numa_preferred_nid; |
1445 | dist = env.dist = node_distance(env.src_nid, env.dst_nid); | 1445 | dist = env.dist = node_distance(env.src_nid, env.dst_nid); |
1446 | taskweight = task_weight(p, env.src_nid, dist); | 1446 | taskweight = task_weight(p, env.src_nid, dist); |
1447 | groupweight = group_weight(p, env.src_nid, dist); | 1447 | groupweight = group_weight(p, env.src_nid, dist); |
1448 | update_numa_stats(&env.src_stats, env.src_nid); | 1448 | update_numa_stats(&env.src_stats, env.src_nid); |
1449 | taskimp = task_weight(p, env.dst_nid, dist) - taskweight; | 1449 | taskimp = task_weight(p, env.dst_nid, dist) - taskweight; |
1450 | groupimp = group_weight(p, env.dst_nid, dist) - groupweight; | 1450 | groupimp = group_weight(p, env.dst_nid, dist) - groupweight; |
1451 | update_numa_stats(&env.dst_stats, env.dst_nid); | 1451 | update_numa_stats(&env.dst_stats, env.dst_nid); |
1452 | 1452 | ||
1453 | /* Try to find a spot on the preferred nid. */ | 1453 | /* Try to find a spot on the preferred nid. */ |
1454 | task_numa_find_cpu(&env, taskimp, groupimp); | 1454 | task_numa_find_cpu(&env, taskimp, groupimp); |
1455 | 1455 | ||
1456 | /* | 1456 | /* |
1457 | * Look at other nodes in these cases: | 1457 | * Look at other nodes in these cases: |
1458 | * - there is no space available on the preferred_nid | 1458 | * - there is no space available on the preferred_nid |
1459 | * - the task is part of a numa_group that is interleaved across | 1459 | * - the task is part of a numa_group that is interleaved across |
1460 | * multiple NUMA nodes; in order to better consolidate the group, | 1460 | * multiple NUMA nodes; in order to better consolidate the group, |
1461 | * we need to check other locations. | 1461 | * we need to check other locations. |
1462 | */ | 1462 | */ |
1463 | if (env.best_cpu == -1 || (p->numa_group && | 1463 | if (env.best_cpu == -1 || (p->numa_group && |
1464 | nodes_weight(p->numa_group->active_nodes) > 1)) { | 1464 | nodes_weight(p->numa_group->active_nodes) > 1)) { |
1465 | for_each_online_node(nid) { | 1465 | for_each_online_node(nid) { |
1466 | if (nid == env.src_nid || nid == p->numa_preferred_nid) | 1466 | if (nid == env.src_nid || nid == p->numa_preferred_nid) |
1467 | continue; | 1467 | continue; |
1468 | 1468 | ||
1469 | dist = node_distance(env.src_nid, env.dst_nid); | 1469 | dist = node_distance(env.src_nid, env.dst_nid); |
1470 | if (sched_numa_topology_type == NUMA_BACKPLANE && | 1470 | if (sched_numa_topology_type == NUMA_BACKPLANE && |
1471 | dist != env.dist) { | 1471 | dist != env.dist) { |
1472 | taskweight = task_weight(p, env.src_nid, dist); | 1472 | taskweight = task_weight(p, env.src_nid, dist); |
1473 | groupweight = group_weight(p, env.src_nid, dist); | 1473 | groupweight = group_weight(p, env.src_nid, dist); |
1474 | } | 1474 | } |
1475 | 1475 | ||
1476 | /* Only consider nodes where both task and groups benefit */ | 1476 | /* Only consider nodes where both task and groups benefit */ |
1477 | taskimp = task_weight(p, nid, dist) - taskweight; | 1477 | taskimp = task_weight(p, nid, dist) - taskweight; |
1478 | groupimp = group_weight(p, nid, dist) - groupweight; | 1478 | groupimp = group_weight(p, nid, dist) - groupweight; |
1479 | if (taskimp < 0 && groupimp < 0) | 1479 | if (taskimp < 0 && groupimp < 0) |
1480 | continue; | 1480 | continue; |
1481 | 1481 | ||
1482 | env.dist = dist; | 1482 | env.dist = dist; |
1483 | env.dst_nid = nid; | 1483 | env.dst_nid = nid; |
1484 | update_numa_stats(&env.dst_stats, env.dst_nid); | 1484 | update_numa_stats(&env.dst_stats, env.dst_nid); |
1485 | task_numa_find_cpu(&env, taskimp, groupimp); | 1485 | task_numa_find_cpu(&env, taskimp, groupimp); |
1486 | } | 1486 | } |
1487 | } | 1487 | } |
1488 | 1488 | ||
1489 | /* | 1489 | /* |
1490 | * If the task is part of a workload that spans multiple NUMA nodes, | 1490 | * If the task is part of a workload that spans multiple NUMA nodes, |
1491 | * and is migrating into one of the workload's active nodes, remember | 1491 | * and is migrating into one of the workload's active nodes, remember |
1492 | * this node as the task's preferred numa node, so the workload can | 1492 | * this node as the task's preferred numa node, so the workload can |
1493 | * settle down. | 1493 | * settle down. |
1494 | * A task that migrated to a second choice node will be better off | 1494 | * A task that migrated to a second choice node will be better off |
1495 | * trying for a better one later. Do not set the preferred node here. | 1495 | * trying for a better one later. Do not set the preferred node here. |
1496 | */ | 1496 | */ |
1497 | if (p->numa_group) { | 1497 | if (p->numa_group) { |
1498 | if (env.best_cpu == -1) | 1498 | if (env.best_cpu == -1) |
1499 | nid = env.src_nid; | 1499 | nid = env.src_nid; |
1500 | else | 1500 | else |
1501 | nid = env.dst_nid; | 1501 | nid = env.dst_nid; |
1502 | 1502 | ||
1503 | if (node_isset(nid, p->numa_group->active_nodes)) | 1503 | if (node_isset(nid, p->numa_group->active_nodes)) |
1504 | sched_setnuma(p, env.dst_nid); | 1504 | sched_setnuma(p, env.dst_nid); |
1505 | } | 1505 | } |
1506 | 1506 | ||
1507 | /* No better CPU than the current one was found. */ | 1507 | /* No better CPU than the current one was found. */ |
1508 | if (env.best_cpu == -1) | 1508 | if (env.best_cpu == -1) |
1509 | return -EAGAIN; | 1509 | return -EAGAIN; |
1510 | 1510 | ||
1511 | /* | 1511 | /* |
1512 | * Reset the scan period if the task is being rescheduled on an | 1512 | * Reset the scan period if the task is being rescheduled on an |
1513 | * alternative node to recheck if the tasks is now properly placed. | 1513 | * alternative node to recheck if the tasks is now properly placed. |
1514 | */ | 1514 | */ |
1515 | p->numa_scan_period = task_scan_min(p); | 1515 | p->numa_scan_period = task_scan_min(p); |
1516 | 1516 | ||
1517 | if (env.best_task == NULL) { | 1517 | if (env.best_task == NULL) { |
1518 | ret = migrate_task_to(p, env.best_cpu); | 1518 | ret = migrate_task_to(p, env.best_cpu); |
1519 | if (ret != 0) | 1519 | if (ret != 0) |
1520 | trace_sched_stick_numa(p, env.src_cpu, env.best_cpu); | 1520 | trace_sched_stick_numa(p, env.src_cpu, env.best_cpu); |
1521 | return ret; | 1521 | return ret; |
1522 | } | 1522 | } |
1523 | 1523 | ||
1524 | ret = migrate_swap(p, env.best_task); | 1524 | ret = migrate_swap(p, env.best_task); |
1525 | if (ret != 0) | 1525 | if (ret != 0) |
1526 | trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task)); | 1526 | trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task)); |
1527 | put_task_struct(env.best_task); | 1527 | put_task_struct(env.best_task); |
1528 | return ret; | 1528 | return ret; |
1529 | } | 1529 | } |
1530 | 1530 | ||
1531 | /* Attempt to migrate a task to a CPU on the preferred node. */ | 1531 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
1532 | static void numa_migrate_preferred(struct task_struct *p) | 1532 | static void numa_migrate_preferred(struct task_struct *p) |
1533 | { | 1533 | { |
1534 | unsigned long interval = HZ; | 1534 | unsigned long interval = HZ; |
1535 | 1535 | ||
1536 | /* This task has no NUMA fault statistics yet */ | 1536 | /* This task has no NUMA fault statistics yet */ |
1537 | if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults)) | 1537 | if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults)) |
1538 | return; | 1538 | return; |
1539 | 1539 | ||
1540 | /* Periodically retry migrating the task to the preferred node */ | 1540 | /* Periodically retry migrating the task to the preferred node */ |
1541 | interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); | 1541 | interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); |
1542 | p->numa_migrate_retry = jiffies + interval; | 1542 | p->numa_migrate_retry = jiffies + interval; |
1543 | 1543 | ||
1544 | /* Success if task is already running on preferred CPU */ | 1544 | /* Success if task is already running on preferred CPU */ |
1545 | if (task_node(p) == p->numa_preferred_nid) | 1545 | if (task_node(p) == p->numa_preferred_nid) |
1546 | return; | 1546 | return; |
1547 | 1547 | ||
1548 | /* Otherwise, try migrate to a CPU on the preferred node */ | 1548 | /* Otherwise, try migrate to a CPU on the preferred node */ |
1549 | task_numa_migrate(p); | 1549 | task_numa_migrate(p); |
1550 | } | 1550 | } |
1551 | 1551 | ||
1552 | /* | 1552 | /* |
1553 | * Find the nodes on which the workload is actively running. We do this by | 1553 | * Find the nodes on which the workload is actively running. We do this by |
1554 | * tracking the nodes from which NUMA hinting faults are triggered. This can | 1554 | * tracking the nodes from which NUMA hinting faults are triggered. This can |
1555 | * be different from the set of nodes where the workload's memory is currently | 1555 | * be different from the set of nodes where the workload's memory is currently |
1556 | * located. | 1556 | * located. |
1557 | * | 1557 | * |
1558 | * The bitmask is used to make smarter decisions on when to do NUMA page | 1558 | * The bitmask is used to make smarter decisions on when to do NUMA page |
1559 | * migrations, To prevent flip-flopping, and excessive page migrations, nodes | 1559 | * migrations, To prevent flip-flopping, and excessive page migrations, nodes |
1560 | * are added when they cause over 6/16 of the maximum number of faults, but | 1560 | * are added when they cause over 6/16 of the maximum number of faults, but |
1561 | * only removed when they drop below 3/16. | 1561 | * only removed when they drop below 3/16. |
1562 | */ | 1562 | */ |
1563 | static void update_numa_active_node_mask(struct numa_group *numa_group) | 1563 | static void update_numa_active_node_mask(struct numa_group *numa_group) |
1564 | { | 1564 | { |
1565 | unsigned long faults, max_faults = 0; | 1565 | unsigned long faults, max_faults = 0; |
1566 | int nid; | 1566 | int nid; |
1567 | 1567 | ||
1568 | for_each_online_node(nid) { | 1568 | for_each_online_node(nid) { |
1569 | faults = group_faults_cpu(numa_group, nid); | 1569 | faults = group_faults_cpu(numa_group, nid); |
1570 | if (faults > max_faults) | 1570 | if (faults > max_faults) |
1571 | max_faults = faults; | 1571 | max_faults = faults; |
1572 | } | 1572 | } |
1573 | 1573 | ||
1574 | for_each_online_node(nid) { | 1574 | for_each_online_node(nid) { |
1575 | faults = group_faults_cpu(numa_group, nid); | 1575 | faults = group_faults_cpu(numa_group, nid); |
1576 | if (!node_isset(nid, numa_group->active_nodes)) { | 1576 | if (!node_isset(nid, numa_group->active_nodes)) { |
1577 | if (faults > max_faults * 6 / 16) | 1577 | if (faults > max_faults * 6 / 16) |
1578 | node_set(nid, numa_group->active_nodes); | 1578 | node_set(nid, numa_group->active_nodes); |
1579 | } else if (faults < max_faults * 3 / 16) | 1579 | } else if (faults < max_faults * 3 / 16) |
1580 | node_clear(nid, numa_group->active_nodes); | 1580 | node_clear(nid, numa_group->active_nodes); |
1581 | } | 1581 | } |
1582 | } | 1582 | } |
1583 | 1583 | ||
1584 | /* | 1584 | /* |
1585 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS | 1585 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS |
1586 | * increments. The more local the fault statistics are, the higher the scan | 1586 | * increments. The more local the fault statistics are, the higher the scan |
1587 | * period will be for the next scan window. If local/(local+remote) ratio is | 1587 | * period will be for the next scan window. If local/(local+remote) ratio is |
1588 | * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) | 1588 | * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) |
1589 | * the scan period will decrease. Aim for 70% local accesses. | 1589 | * the scan period will decrease. Aim for 70% local accesses. |
1590 | */ | 1590 | */ |
1591 | #define NUMA_PERIOD_SLOTS 10 | 1591 | #define NUMA_PERIOD_SLOTS 10 |
1592 | #define NUMA_PERIOD_THRESHOLD 7 | 1592 | #define NUMA_PERIOD_THRESHOLD 7 |
1593 | 1593 | ||
1594 | /* | 1594 | /* |
1595 | * Increase the scan period (slow down scanning) if the majority of | 1595 | * Increase the scan period (slow down scanning) if the majority of |
1596 | * our memory is already on our local node, or if the majority of | 1596 | * our memory is already on our local node, or if the majority of |
1597 | * the page accesses are shared with other processes. | 1597 | * the page accesses are shared with other processes. |
1598 | * Otherwise, decrease the scan period. | 1598 | * Otherwise, decrease the scan period. |
1599 | */ | 1599 | */ |
1600 | static void update_task_scan_period(struct task_struct *p, | 1600 | static void update_task_scan_period(struct task_struct *p, |
1601 | unsigned long shared, unsigned long private) | 1601 | unsigned long shared, unsigned long private) |
1602 | { | 1602 | { |
1603 | unsigned int period_slot; | 1603 | unsigned int period_slot; |
1604 | int ratio; | 1604 | int ratio; |
1605 | int diff; | 1605 | int diff; |
1606 | 1606 | ||
1607 | unsigned long remote = p->numa_faults_locality[0]; | 1607 | unsigned long remote = p->numa_faults_locality[0]; |
1608 | unsigned long local = p->numa_faults_locality[1]; | 1608 | unsigned long local = p->numa_faults_locality[1]; |
1609 | 1609 | ||
1610 | /* | 1610 | /* |
1611 | * If there were no record hinting faults then either the task is | 1611 | * If there were no record hinting faults then either the task is |
1612 | * completely idle or all activity is areas that are not of interest | 1612 | * completely idle or all activity is areas that are not of interest |
1613 | * to automatic numa balancing. Scan slower | 1613 | * to automatic numa balancing. Scan slower |
1614 | */ | 1614 | */ |
1615 | if (local + shared == 0) { | 1615 | if (local + shared == 0) { |
1616 | p->numa_scan_period = min(p->numa_scan_period_max, | 1616 | p->numa_scan_period = min(p->numa_scan_period_max, |
1617 | p->numa_scan_period << 1); | 1617 | p->numa_scan_period << 1); |
1618 | 1618 | ||
1619 | p->mm->numa_next_scan = jiffies + | 1619 | p->mm->numa_next_scan = jiffies + |
1620 | msecs_to_jiffies(p->numa_scan_period); | 1620 | msecs_to_jiffies(p->numa_scan_period); |
1621 | 1621 | ||
1622 | return; | 1622 | return; |
1623 | } | 1623 | } |
1624 | 1624 | ||
1625 | /* | 1625 | /* |
1626 | * Prepare to scale scan period relative to the current period. | 1626 | * Prepare to scale scan period relative to the current period. |
1627 | * == NUMA_PERIOD_THRESHOLD scan period stays the same | 1627 | * == NUMA_PERIOD_THRESHOLD scan period stays the same |
1628 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) | 1628 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) |
1629 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) | 1629 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) |
1630 | */ | 1630 | */ |
1631 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); | 1631 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); |
1632 | ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); | 1632 | ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); |
1633 | if (ratio >= NUMA_PERIOD_THRESHOLD) { | 1633 | if (ratio >= NUMA_PERIOD_THRESHOLD) { |
1634 | int slot = ratio - NUMA_PERIOD_THRESHOLD; | 1634 | int slot = ratio - NUMA_PERIOD_THRESHOLD; |
1635 | if (!slot) | 1635 | if (!slot) |
1636 | slot = 1; | 1636 | slot = 1; |
1637 | diff = slot * period_slot; | 1637 | diff = slot * period_slot; |
1638 | } else { | 1638 | } else { |
1639 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; | 1639 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; |
1640 | 1640 | ||
1641 | /* | 1641 | /* |
1642 | * Scale scan rate increases based on sharing. There is an | 1642 | * Scale scan rate increases based on sharing. There is an |
1643 | * inverse relationship between the degree of sharing and | 1643 | * inverse relationship between the degree of sharing and |
1644 | * the adjustment made to the scanning period. Broadly | 1644 | * the adjustment made to the scanning period. Broadly |
1645 | * speaking the intent is that there is little point | 1645 | * speaking the intent is that there is little point |
1646 | * scanning faster if shared accesses dominate as it may | 1646 | * scanning faster if shared accesses dominate as it may |
1647 | * simply bounce migrations uselessly | 1647 | * simply bounce migrations uselessly |
1648 | */ | 1648 | */ |
1649 | ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1)); | 1649 | ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1)); |
1650 | diff = (diff * ratio) / NUMA_PERIOD_SLOTS; | 1650 | diff = (diff * ratio) / NUMA_PERIOD_SLOTS; |
1651 | } | 1651 | } |
1652 | 1652 | ||
1653 | p->numa_scan_period = clamp(p->numa_scan_period + diff, | 1653 | p->numa_scan_period = clamp(p->numa_scan_period + diff, |
1654 | task_scan_min(p), task_scan_max(p)); | 1654 | task_scan_min(p), task_scan_max(p)); |
1655 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | 1655 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
1656 | } | 1656 | } |
1657 | 1657 | ||
1658 | /* | 1658 | /* |
1659 | * Get the fraction of time the task has been running since the last | 1659 | * Get the fraction of time the task has been running since the last |
1660 | * NUMA placement cycle. The scheduler keeps similar statistics, but | 1660 | * NUMA placement cycle. The scheduler keeps similar statistics, but |
1661 | * decays those on a 32ms period, which is orders of magnitude off | 1661 | * decays those on a 32ms period, which is orders of magnitude off |
1662 | * from the dozens-of-seconds NUMA balancing period. Use the scheduler | 1662 | * from the dozens-of-seconds NUMA balancing period. Use the scheduler |
1663 | * stats only if the task is so new there are no NUMA statistics yet. | 1663 | * stats only if the task is so new there are no NUMA statistics yet. |
1664 | */ | 1664 | */ |
1665 | static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) | 1665 | static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) |
1666 | { | 1666 | { |
1667 | u64 runtime, delta, now; | 1667 | u64 runtime, delta, now; |
1668 | /* Use the start of this time slice to avoid calculations. */ | 1668 | /* Use the start of this time slice to avoid calculations. */ |
1669 | now = p->se.exec_start; | 1669 | now = p->se.exec_start; |
1670 | runtime = p->se.sum_exec_runtime; | 1670 | runtime = p->se.sum_exec_runtime; |
1671 | 1671 | ||
1672 | if (p->last_task_numa_placement) { | 1672 | if (p->last_task_numa_placement) { |
1673 | delta = runtime - p->last_sum_exec_runtime; | 1673 | delta = runtime - p->last_sum_exec_runtime; |
1674 | *period = now - p->last_task_numa_placement; | 1674 | *period = now - p->last_task_numa_placement; |
1675 | } else { | 1675 | } else { |
1676 | delta = p->se.avg.runnable_avg_sum; | 1676 | delta = p->se.avg.runnable_avg_sum; |
1677 | *period = p->se.avg.runnable_avg_period; | 1677 | *period = p->se.avg.runnable_avg_period; |
1678 | } | 1678 | } |
1679 | 1679 | ||
1680 | p->last_sum_exec_runtime = runtime; | 1680 | p->last_sum_exec_runtime = runtime; |
1681 | p->last_task_numa_placement = now; | 1681 | p->last_task_numa_placement = now; |
1682 | 1682 | ||
1683 | return delta; | 1683 | return delta; |
1684 | } | 1684 | } |
1685 | 1685 | ||
1686 | /* | 1686 | /* |
1687 | * Determine the preferred nid for a task in a numa_group. This needs to | 1687 | * Determine the preferred nid for a task in a numa_group. This needs to |
1688 | * be done in a way that produces consistent results with group_weight, | 1688 | * be done in a way that produces consistent results with group_weight, |
1689 | * otherwise workloads might not converge. | 1689 | * otherwise workloads might not converge. |
1690 | */ | 1690 | */ |
1691 | static int preferred_group_nid(struct task_struct *p, int nid) | 1691 | static int preferred_group_nid(struct task_struct *p, int nid) |
1692 | { | 1692 | { |
1693 | nodemask_t nodes; | 1693 | nodemask_t nodes; |
1694 | int dist; | 1694 | int dist; |
1695 | 1695 | ||
1696 | /* Direct connections between all NUMA nodes. */ | 1696 | /* Direct connections between all NUMA nodes. */ |
1697 | if (sched_numa_topology_type == NUMA_DIRECT) | 1697 | if (sched_numa_topology_type == NUMA_DIRECT) |
1698 | return nid; | 1698 | return nid; |
1699 | 1699 | ||
1700 | /* | 1700 | /* |
1701 | * On a system with glueless mesh NUMA topology, group_weight | 1701 | * On a system with glueless mesh NUMA topology, group_weight |
1702 | * scores nodes according to the number of NUMA hinting faults on | 1702 | * scores nodes according to the number of NUMA hinting faults on |
1703 | * both the node itself, and on nearby nodes. | 1703 | * both the node itself, and on nearby nodes. |
1704 | */ | 1704 | */ |
1705 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | 1705 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { |
1706 | unsigned long score, max_score = 0; | 1706 | unsigned long score, max_score = 0; |
1707 | int node, max_node = nid; | 1707 | int node, max_node = nid; |
1708 | 1708 | ||
1709 | dist = sched_max_numa_distance; | 1709 | dist = sched_max_numa_distance; |
1710 | 1710 | ||
1711 | for_each_online_node(node) { | 1711 | for_each_online_node(node) { |
1712 | score = group_weight(p, node, dist); | 1712 | score = group_weight(p, node, dist); |
1713 | if (score > max_score) { | 1713 | if (score > max_score) { |
1714 | max_score = score; | 1714 | max_score = score; |
1715 | max_node = node; | 1715 | max_node = node; |
1716 | } | 1716 | } |
1717 | } | 1717 | } |
1718 | return max_node; | 1718 | return max_node; |
1719 | } | 1719 | } |
1720 | 1720 | ||
1721 | /* | 1721 | /* |
1722 | * Finding the preferred nid in a system with NUMA backplane | 1722 | * Finding the preferred nid in a system with NUMA backplane |
1723 | * interconnect topology is more involved. The goal is to locate | 1723 | * interconnect topology is more involved. The goal is to locate |
1724 | * tasks from numa_groups near each other in the system, and | 1724 | * tasks from numa_groups near each other in the system, and |
1725 | * untangle workloads from different sides of the system. This requires | 1725 | * untangle workloads from different sides of the system. This requires |
1726 | * searching down the hierarchy of node groups, recursively searching | 1726 | * searching down the hierarchy of node groups, recursively searching |
1727 | * inside the highest scoring group of nodes. The nodemask tricks | 1727 | * inside the highest scoring group of nodes. The nodemask tricks |
1728 | * keep the complexity of the search down. | 1728 | * keep the complexity of the search down. |
1729 | */ | 1729 | */ |
1730 | nodes = node_online_map; | 1730 | nodes = node_online_map; |
1731 | for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { | 1731 | for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { |
1732 | unsigned long max_faults = 0; | 1732 | unsigned long max_faults = 0; |
1733 | nodemask_t max_group; | 1733 | nodemask_t max_group; |
1734 | int a, b; | 1734 | int a, b; |
1735 | 1735 | ||
1736 | /* Are there nodes at this distance from each other? */ | 1736 | /* Are there nodes at this distance from each other? */ |
1737 | if (!find_numa_distance(dist)) | 1737 | if (!find_numa_distance(dist)) |
1738 | continue; | 1738 | continue; |
1739 | 1739 | ||
1740 | for_each_node_mask(a, nodes) { | 1740 | for_each_node_mask(a, nodes) { |
1741 | unsigned long faults = 0; | 1741 | unsigned long faults = 0; |
1742 | nodemask_t this_group; | 1742 | nodemask_t this_group; |
1743 | nodes_clear(this_group); | 1743 | nodes_clear(this_group); |
1744 | 1744 | ||
1745 | /* Sum group's NUMA faults; includes a==b case. */ | 1745 | /* Sum group's NUMA faults; includes a==b case. */ |
1746 | for_each_node_mask(b, nodes) { | 1746 | for_each_node_mask(b, nodes) { |
1747 | if (node_distance(a, b) < dist) { | 1747 | if (node_distance(a, b) < dist) { |
1748 | faults += group_faults(p, b); | 1748 | faults += group_faults(p, b); |
1749 | node_set(b, this_group); | 1749 | node_set(b, this_group); |
1750 | node_clear(b, nodes); | 1750 | node_clear(b, nodes); |
1751 | } | 1751 | } |
1752 | } | 1752 | } |
1753 | 1753 | ||
1754 | /* Remember the top group. */ | 1754 | /* Remember the top group. */ |
1755 | if (faults > max_faults) { | 1755 | if (faults > max_faults) { |
1756 | max_faults = faults; | 1756 | max_faults = faults; |
1757 | max_group = this_group; | 1757 | max_group = this_group; |
1758 | /* | 1758 | /* |
1759 | * subtle: at the smallest distance there is | 1759 | * subtle: at the smallest distance there is |
1760 | * just one node left in each "group", the | 1760 | * just one node left in each "group", the |
1761 | * winner is the preferred nid. | 1761 | * winner is the preferred nid. |
1762 | */ | 1762 | */ |
1763 | nid = a; | 1763 | nid = a; |
1764 | } | 1764 | } |
1765 | } | 1765 | } |
1766 | /* Next round, evaluate the nodes within max_group. */ | 1766 | /* Next round, evaluate the nodes within max_group. */ |
1767 | nodes = max_group; | 1767 | nodes = max_group; |
1768 | } | 1768 | } |
1769 | return nid; | 1769 | return nid; |
1770 | } | 1770 | } |
1771 | 1771 | ||
1772 | static void task_numa_placement(struct task_struct *p) | 1772 | static void task_numa_placement(struct task_struct *p) |
1773 | { | 1773 | { |
1774 | int seq, nid, max_nid = -1, max_group_nid = -1; | 1774 | int seq, nid, max_nid = -1, max_group_nid = -1; |
1775 | unsigned long max_faults = 0, max_group_faults = 0; | 1775 | unsigned long max_faults = 0, max_group_faults = 0; |
1776 | unsigned long fault_types[2] = { 0, 0 }; | 1776 | unsigned long fault_types[2] = { 0, 0 }; |
1777 | unsigned long total_faults; | 1777 | unsigned long total_faults; |
1778 | u64 runtime, period; | 1778 | u64 runtime, period; |
1779 | spinlock_t *group_lock = NULL; | 1779 | spinlock_t *group_lock = NULL; |
1780 | 1780 | ||
1781 | seq = ACCESS_ONCE(p->mm->numa_scan_seq); | 1781 | seq = ACCESS_ONCE(p->mm->numa_scan_seq); |
1782 | if (p->numa_scan_seq == seq) | 1782 | if (p->numa_scan_seq == seq) |
1783 | return; | 1783 | return; |
1784 | p->numa_scan_seq = seq; | 1784 | p->numa_scan_seq = seq; |
1785 | p->numa_scan_period_max = task_scan_max(p); | 1785 | p->numa_scan_period_max = task_scan_max(p); |
1786 | 1786 | ||
1787 | total_faults = p->numa_faults_locality[0] + | 1787 | total_faults = p->numa_faults_locality[0] + |
1788 | p->numa_faults_locality[1]; | 1788 | p->numa_faults_locality[1]; |
1789 | runtime = numa_get_avg_runtime(p, &period); | 1789 | runtime = numa_get_avg_runtime(p, &period); |
1790 | 1790 | ||
1791 | /* If the task is part of a group prevent parallel updates to group stats */ | 1791 | /* If the task is part of a group prevent parallel updates to group stats */ |
1792 | if (p->numa_group) { | 1792 | if (p->numa_group) { |
1793 | group_lock = &p->numa_group->lock; | 1793 | group_lock = &p->numa_group->lock; |
1794 | spin_lock_irq(group_lock); | 1794 | spin_lock_irq(group_lock); |
1795 | } | 1795 | } |
1796 | 1796 | ||
1797 | /* Find the node with the highest number of faults */ | 1797 | /* Find the node with the highest number of faults */ |
1798 | for_each_online_node(nid) { | 1798 | for_each_online_node(nid) { |
1799 | /* Keep track of the offsets in numa_faults array */ | 1799 | /* Keep track of the offsets in numa_faults array */ |
1800 | int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; | 1800 | int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; |
1801 | unsigned long faults = 0, group_faults = 0; | 1801 | unsigned long faults = 0, group_faults = 0; |
1802 | int priv; | 1802 | int priv; |
1803 | 1803 | ||
1804 | for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { | 1804 | for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { |
1805 | long diff, f_diff, f_weight; | 1805 | long diff, f_diff, f_weight; |
1806 | 1806 | ||
1807 | mem_idx = task_faults_idx(NUMA_MEM, nid, priv); | 1807 | mem_idx = task_faults_idx(NUMA_MEM, nid, priv); |
1808 | membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); | 1808 | membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); |
1809 | cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); | 1809 | cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); |
1810 | cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); | 1810 | cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); |
1811 | 1811 | ||
1812 | /* Decay existing window, copy faults since last scan */ | 1812 | /* Decay existing window, copy faults since last scan */ |
1813 | diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; | 1813 | diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; |
1814 | fault_types[priv] += p->numa_faults[membuf_idx]; | 1814 | fault_types[priv] += p->numa_faults[membuf_idx]; |
1815 | p->numa_faults[membuf_idx] = 0; | 1815 | p->numa_faults[membuf_idx] = 0; |
1816 | 1816 | ||
1817 | /* | 1817 | /* |
1818 | * Normalize the faults_from, so all tasks in a group | 1818 | * Normalize the faults_from, so all tasks in a group |
1819 | * count according to CPU use, instead of by the raw | 1819 | * count according to CPU use, instead of by the raw |
1820 | * number of faults. Tasks with little runtime have | 1820 | * number of faults. Tasks with little runtime have |
1821 | * little over-all impact on throughput, and thus their | 1821 | * little over-all impact on throughput, and thus their |
1822 | * faults are less important. | 1822 | * faults are less important. |
1823 | */ | 1823 | */ |
1824 | f_weight = div64_u64(runtime << 16, period + 1); | 1824 | f_weight = div64_u64(runtime << 16, period + 1); |
1825 | f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / | 1825 | f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / |
1826 | (total_faults + 1); | 1826 | (total_faults + 1); |
1827 | f_diff = f_weight - p->numa_faults[cpu_idx] / 2; | 1827 | f_diff = f_weight - p->numa_faults[cpu_idx] / 2; |
1828 | p->numa_faults[cpubuf_idx] = 0; | 1828 | p->numa_faults[cpubuf_idx] = 0; |
1829 | 1829 | ||
1830 | p->numa_faults[mem_idx] += diff; | 1830 | p->numa_faults[mem_idx] += diff; |
1831 | p->numa_faults[cpu_idx] += f_diff; | 1831 | p->numa_faults[cpu_idx] += f_diff; |
1832 | faults += p->numa_faults[mem_idx]; | 1832 | faults += p->numa_faults[mem_idx]; |
1833 | p->total_numa_faults += diff; | 1833 | p->total_numa_faults += diff; |
1834 | if (p->numa_group) { | 1834 | if (p->numa_group) { |
1835 | /* | 1835 | /* |
1836 | * safe because we can only change our own group | 1836 | * safe because we can only change our own group |
1837 | * | 1837 | * |
1838 | * mem_idx represents the offset for a given | 1838 | * mem_idx represents the offset for a given |
1839 | * nid and priv in a specific region because it | 1839 | * nid and priv in a specific region because it |
1840 | * is at the beginning of the numa_faults array. | 1840 | * is at the beginning of the numa_faults array. |
1841 | */ | 1841 | */ |
1842 | p->numa_group->faults[mem_idx] += diff; | 1842 | p->numa_group->faults[mem_idx] += diff; |
1843 | p->numa_group->faults_cpu[mem_idx] += f_diff; | 1843 | p->numa_group->faults_cpu[mem_idx] += f_diff; |
1844 | p->numa_group->total_faults += diff; | 1844 | p->numa_group->total_faults += diff; |
1845 | group_faults += p->numa_group->faults[mem_idx]; | 1845 | group_faults += p->numa_group->faults[mem_idx]; |
1846 | } | 1846 | } |
1847 | } | 1847 | } |
1848 | 1848 | ||
1849 | if (faults > max_faults) { | 1849 | if (faults > max_faults) { |
1850 | max_faults = faults; | 1850 | max_faults = faults; |
1851 | max_nid = nid; | 1851 | max_nid = nid; |
1852 | } | 1852 | } |
1853 | 1853 | ||
1854 | if (group_faults > max_group_faults) { | 1854 | if (group_faults > max_group_faults) { |
1855 | max_group_faults = group_faults; | 1855 | max_group_faults = group_faults; |
1856 | max_group_nid = nid; | 1856 | max_group_nid = nid; |
1857 | } | 1857 | } |
1858 | } | 1858 | } |
1859 | 1859 | ||
1860 | update_task_scan_period(p, fault_types[0], fault_types[1]); | 1860 | update_task_scan_period(p, fault_types[0], fault_types[1]); |
1861 | 1861 | ||
1862 | if (p->numa_group) { | 1862 | if (p->numa_group) { |
1863 | update_numa_active_node_mask(p->numa_group); | 1863 | update_numa_active_node_mask(p->numa_group); |
1864 | spin_unlock_irq(group_lock); | 1864 | spin_unlock_irq(group_lock); |
1865 | max_nid = preferred_group_nid(p, max_group_nid); | 1865 | max_nid = preferred_group_nid(p, max_group_nid); |
1866 | } | 1866 | } |
1867 | 1867 | ||
1868 | if (max_faults) { | 1868 | if (max_faults) { |
1869 | /* Set the new preferred node */ | 1869 | /* Set the new preferred node */ |
1870 | if (max_nid != p->numa_preferred_nid) | 1870 | if (max_nid != p->numa_preferred_nid) |
1871 | sched_setnuma(p, max_nid); | 1871 | sched_setnuma(p, max_nid); |
1872 | 1872 | ||
1873 | if (task_node(p) != p->numa_preferred_nid) | 1873 | if (task_node(p) != p->numa_preferred_nid) |
1874 | numa_migrate_preferred(p); | 1874 | numa_migrate_preferred(p); |
1875 | } | 1875 | } |
1876 | } | 1876 | } |
1877 | 1877 | ||
1878 | static inline int get_numa_group(struct numa_group *grp) | 1878 | static inline int get_numa_group(struct numa_group *grp) |
1879 | { | 1879 | { |
1880 | return atomic_inc_not_zero(&grp->refcount); | 1880 | return atomic_inc_not_zero(&grp->refcount); |
1881 | } | 1881 | } |
1882 | 1882 | ||
1883 | static inline void put_numa_group(struct numa_group *grp) | 1883 | static inline void put_numa_group(struct numa_group *grp) |
1884 | { | 1884 | { |
1885 | if (atomic_dec_and_test(&grp->refcount)) | 1885 | if (atomic_dec_and_test(&grp->refcount)) |
1886 | kfree_rcu(grp, rcu); | 1886 | kfree_rcu(grp, rcu); |
1887 | } | 1887 | } |
1888 | 1888 | ||
1889 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, | 1889 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, |
1890 | int *priv) | 1890 | int *priv) |
1891 | { | 1891 | { |
1892 | struct numa_group *grp, *my_grp; | 1892 | struct numa_group *grp, *my_grp; |
1893 | struct task_struct *tsk; | 1893 | struct task_struct *tsk; |
1894 | bool join = false; | 1894 | bool join = false; |
1895 | int cpu = cpupid_to_cpu(cpupid); | 1895 | int cpu = cpupid_to_cpu(cpupid); |
1896 | int i; | 1896 | int i; |
1897 | 1897 | ||
1898 | if (unlikely(!p->numa_group)) { | 1898 | if (unlikely(!p->numa_group)) { |
1899 | unsigned int size = sizeof(struct numa_group) + | 1899 | unsigned int size = sizeof(struct numa_group) + |
1900 | 4*nr_node_ids*sizeof(unsigned long); | 1900 | 4*nr_node_ids*sizeof(unsigned long); |
1901 | 1901 | ||
1902 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | 1902 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); |
1903 | if (!grp) | 1903 | if (!grp) |
1904 | return; | 1904 | return; |
1905 | 1905 | ||
1906 | atomic_set(&grp->refcount, 1); | 1906 | atomic_set(&grp->refcount, 1); |
1907 | spin_lock_init(&grp->lock); | 1907 | spin_lock_init(&grp->lock); |
1908 | grp->gid = p->pid; | 1908 | grp->gid = p->pid; |
1909 | /* Second half of the array tracks nids where faults happen */ | 1909 | /* Second half of the array tracks nids where faults happen */ |
1910 | grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * | 1910 | grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * |
1911 | nr_node_ids; | 1911 | nr_node_ids; |
1912 | 1912 | ||
1913 | node_set(task_node(current), grp->active_nodes); | 1913 | node_set(task_node(current), grp->active_nodes); |
1914 | 1914 | ||
1915 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) | 1915 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
1916 | grp->faults[i] = p->numa_faults[i]; | 1916 | grp->faults[i] = p->numa_faults[i]; |
1917 | 1917 | ||
1918 | grp->total_faults = p->total_numa_faults; | 1918 | grp->total_faults = p->total_numa_faults; |
1919 | 1919 | ||
1920 | grp->nr_tasks++; | 1920 | grp->nr_tasks++; |
1921 | rcu_assign_pointer(p->numa_group, grp); | 1921 | rcu_assign_pointer(p->numa_group, grp); |
1922 | } | 1922 | } |
1923 | 1923 | ||
1924 | rcu_read_lock(); | 1924 | rcu_read_lock(); |
1925 | tsk = ACCESS_ONCE(cpu_rq(cpu)->curr); | 1925 | tsk = ACCESS_ONCE(cpu_rq(cpu)->curr); |
1926 | 1926 | ||
1927 | if (!cpupid_match_pid(tsk, cpupid)) | 1927 | if (!cpupid_match_pid(tsk, cpupid)) |
1928 | goto no_join; | 1928 | goto no_join; |
1929 | 1929 | ||
1930 | grp = rcu_dereference(tsk->numa_group); | 1930 | grp = rcu_dereference(tsk->numa_group); |
1931 | if (!grp) | 1931 | if (!grp) |
1932 | goto no_join; | 1932 | goto no_join; |
1933 | 1933 | ||
1934 | my_grp = p->numa_group; | 1934 | my_grp = p->numa_group; |
1935 | if (grp == my_grp) | 1935 | if (grp == my_grp) |
1936 | goto no_join; | 1936 | goto no_join; |
1937 | 1937 | ||
1938 | /* | 1938 | /* |
1939 | * Only join the other group if its bigger; if we're the bigger group, | 1939 | * Only join the other group if its bigger; if we're the bigger group, |
1940 | * the other task will join us. | 1940 | * the other task will join us. |
1941 | */ | 1941 | */ |
1942 | if (my_grp->nr_tasks > grp->nr_tasks) | 1942 | if (my_grp->nr_tasks > grp->nr_tasks) |
1943 | goto no_join; | 1943 | goto no_join; |
1944 | 1944 | ||
1945 | /* | 1945 | /* |
1946 | * Tie-break on the grp address. | 1946 | * Tie-break on the grp address. |
1947 | */ | 1947 | */ |
1948 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | 1948 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) |
1949 | goto no_join; | 1949 | goto no_join; |
1950 | 1950 | ||
1951 | /* Always join threads in the same process. */ | 1951 | /* Always join threads in the same process. */ |
1952 | if (tsk->mm == current->mm) | 1952 | if (tsk->mm == current->mm) |
1953 | join = true; | 1953 | join = true; |
1954 | 1954 | ||
1955 | /* Simple filter to avoid false positives due to PID collisions */ | 1955 | /* Simple filter to avoid false positives due to PID collisions */ |
1956 | if (flags & TNF_SHARED) | 1956 | if (flags & TNF_SHARED) |
1957 | join = true; | 1957 | join = true; |
1958 | 1958 | ||
1959 | /* Update priv based on whether false sharing was detected */ | 1959 | /* Update priv based on whether false sharing was detected */ |
1960 | *priv = !join; | 1960 | *priv = !join; |
1961 | 1961 | ||
1962 | if (join && !get_numa_group(grp)) | 1962 | if (join && !get_numa_group(grp)) |
1963 | goto no_join; | 1963 | goto no_join; |
1964 | 1964 | ||
1965 | rcu_read_unlock(); | 1965 | rcu_read_unlock(); |
1966 | 1966 | ||
1967 | if (!join) | 1967 | if (!join) |
1968 | return; | 1968 | return; |
1969 | 1969 | ||
1970 | BUG_ON(irqs_disabled()); | 1970 | BUG_ON(irqs_disabled()); |
1971 | double_lock_irq(&my_grp->lock, &grp->lock); | 1971 | double_lock_irq(&my_grp->lock, &grp->lock); |
1972 | 1972 | ||
1973 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { | 1973 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { |
1974 | my_grp->faults[i] -= p->numa_faults[i]; | 1974 | my_grp->faults[i] -= p->numa_faults[i]; |
1975 | grp->faults[i] += p->numa_faults[i]; | 1975 | grp->faults[i] += p->numa_faults[i]; |
1976 | } | 1976 | } |
1977 | my_grp->total_faults -= p->total_numa_faults; | 1977 | my_grp->total_faults -= p->total_numa_faults; |
1978 | grp->total_faults += p->total_numa_faults; | 1978 | grp->total_faults += p->total_numa_faults; |
1979 | 1979 | ||
1980 | my_grp->nr_tasks--; | 1980 | my_grp->nr_tasks--; |
1981 | grp->nr_tasks++; | 1981 | grp->nr_tasks++; |
1982 | 1982 | ||
1983 | spin_unlock(&my_grp->lock); | 1983 | spin_unlock(&my_grp->lock); |
1984 | spin_unlock_irq(&grp->lock); | 1984 | spin_unlock_irq(&grp->lock); |
1985 | 1985 | ||
1986 | rcu_assign_pointer(p->numa_group, grp); | 1986 | rcu_assign_pointer(p->numa_group, grp); |
1987 | 1987 | ||
1988 | put_numa_group(my_grp); | 1988 | put_numa_group(my_grp); |
1989 | return; | 1989 | return; |
1990 | 1990 | ||
1991 | no_join: | 1991 | no_join: |
1992 | rcu_read_unlock(); | 1992 | rcu_read_unlock(); |
1993 | return; | 1993 | return; |
1994 | } | 1994 | } |
1995 | 1995 | ||
1996 | void task_numa_free(struct task_struct *p) | 1996 | void task_numa_free(struct task_struct *p) |
1997 | { | 1997 | { |
1998 | struct numa_group *grp = p->numa_group; | 1998 | struct numa_group *grp = p->numa_group; |
1999 | void *numa_faults = p->numa_faults; | 1999 | void *numa_faults = p->numa_faults; |
2000 | unsigned long flags; | 2000 | unsigned long flags; |
2001 | int i; | 2001 | int i; |
2002 | 2002 | ||
2003 | if (grp) { | 2003 | if (grp) { |
2004 | spin_lock_irqsave(&grp->lock, flags); | 2004 | spin_lock_irqsave(&grp->lock, flags); |
2005 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) | 2005 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
2006 | grp->faults[i] -= p->numa_faults[i]; | 2006 | grp->faults[i] -= p->numa_faults[i]; |
2007 | grp->total_faults -= p->total_numa_faults; | 2007 | grp->total_faults -= p->total_numa_faults; |
2008 | 2008 | ||
2009 | grp->nr_tasks--; | 2009 | grp->nr_tasks--; |
2010 | spin_unlock_irqrestore(&grp->lock, flags); | 2010 | spin_unlock_irqrestore(&grp->lock, flags); |
2011 | RCU_INIT_POINTER(p->numa_group, NULL); | 2011 | RCU_INIT_POINTER(p->numa_group, NULL); |
2012 | put_numa_group(grp); | 2012 | put_numa_group(grp); |
2013 | } | 2013 | } |
2014 | 2014 | ||
2015 | p->numa_faults = NULL; | 2015 | p->numa_faults = NULL; |
2016 | kfree(numa_faults); | 2016 | kfree(numa_faults); |
2017 | } | 2017 | } |
2018 | 2018 | ||
2019 | /* | 2019 | /* |
2020 | * Got a PROT_NONE fault for a page on @node. | 2020 | * Got a PROT_NONE fault for a page on @node. |
2021 | */ | 2021 | */ |
2022 | void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) | 2022 | void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) |
2023 | { | 2023 | { |
2024 | struct task_struct *p = current; | 2024 | struct task_struct *p = current; |
2025 | bool migrated = flags & TNF_MIGRATED; | 2025 | bool migrated = flags & TNF_MIGRATED; |
2026 | int cpu_node = task_node(current); | 2026 | int cpu_node = task_node(current); |
2027 | int local = !!(flags & TNF_FAULT_LOCAL); | 2027 | int local = !!(flags & TNF_FAULT_LOCAL); |
2028 | int priv; | 2028 | int priv; |
2029 | 2029 | ||
2030 | if (!numabalancing_enabled) | 2030 | if (!numabalancing_enabled) |
2031 | return; | 2031 | return; |
2032 | 2032 | ||
2033 | /* for example, ksmd faulting in a user's mm */ | 2033 | /* for example, ksmd faulting in a user's mm */ |
2034 | if (!p->mm) | 2034 | if (!p->mm) |
2035 | return; | 2035 | return; |
2036 | 2036 | ||
2037 | /* Allocate buffer to track faults on a per-node basis */ | 2037 | /* Allocate buffer to track faults on a per-node basis */ |
2038 | if (unlikely(!p->numa_faults)) { | 2038 | if (unlikely(!p->numa_faults)) { |
2039 | int size = sizeof(*p->numa_faults) * | 2039 | int size = sizeof(*p->numa_faults) * |
2040 | NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; | 2040 | NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; |
2041 | 2041 | ||
2042 | p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); | 2042 | p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); |
2043 | if (!p->numa_faults) | 2043 | if (!p->numa_faults) |
2044 | return; | 2044 | return; |
2045 | 2045 | ||
2046 | p->total_numa_faults = 0; | 2046 | p->total_numa_faults = 0; |
2047 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | 2047 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
2048 | } | 2048 | } |
2049 | 2049 | ||
2050 | /* | 2050 | /* |
2051 | * First accesses are treated as private, otherwise consider accesses | 2051 | * First accesses are treated as private, otherwise consider accesses |
2052 | * to be private if the accessing pid has not changed | 2052 | * to be private if the accessing pid has not changed |
2053 | */ | 2053 | */ |
2054 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | 2054 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { |
2055 | priv = 1; | 2055 | priv = 1; |
2056 | } else { | 2056 | } else { |
2057 | priv = cpupid_match_pid(p, last_cpupid); | 2057 | priv = cpupid_match_pid(p, last_cpupid); |
2058 | if (!priv && !(flags & TNF_NO_GROUP)) | 2058 | if (!priv && !(flags & TNF_NO_GROUP)) |
2059 | task_numa_group(p, last_cpupid, flags, &priv); | 2059 | task_numa_group(p, last_cpupid, flags, &priv); |
2060 | } | 2060 | } |
2061 | 2061 | ||
2062 | /* | 2062 | /* |
2063 | * If a workload spans multiple NUMA nodes, a shared fault that | 2063 | * If a workload spans multiple NUMA nodes, a shared fault that |
2064 | * occurs wholly within the set of nodes that the workload is | 2064 | * occurs wholly within the set of nodes that the workload is |
2065 | * actively using should be counted as local. This allows the | 2065 | * actively using should be counted as local. This allows the |
2066 | * scan rate to slow down when a workload has settled down. | 2066 | * scan rate to slow down when a workload has settled down. |
2067 | */ | 2067 | */ |
2068 | if (!priv && !local && p->numa_group && | 2068 | if (!priv && !local && p->numa_group && |
2069 | node_isset(cpu_node, p->numa_group->active_nodes) && | 2069 | node_isset(cpu_node, p->numa_group->active_nodes) && |
2070 | node_isset(mem_node, p->numa_group->active_nodes)) | 2070 | node_isset(mem_node, p->numa_group->active_nodes)) |
2071 | local = 1; | 2071 | local = 1; |
2072 | 2072 | ||
2073 | task_numa_placement(p); | 2073 | task_numa_placement(p); |
2074 | 2074 | ||
2075 | /* | 2075 | /* |
2076 | * Retry task to preferred node migration periodically, in case it | 2076 | * Retry task to preferred node migration periodically, in case it |
2077 | * case it previously failed, or the scheduler moved us. | 2077 | * case it previously failed, or the scheduler moved us. |
2078 | */ | 2078 | */ |
2079 | if (time_after(jiffies, p->numa_migrate_retry)) | 2079 | if (time_after(jiffies, p->numa_migrate_retry)) |
2080 | numa_migrate_preferred(p); | 2080 | numa_migrate_preferred(p); |
2081 | 2081 | ||
2082 | if (migrated) | 2082 | if (migrated) |
2083 | p->numa_pages_migrated += pages; | 2083 | p->numa_pages_migrated += pages; |
2084 | 2084 | ||
2085 | p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; | 2085 | p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; |
2086 | p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; | 2086 | p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; |
2087 | p->numa_faults_locality[local] += pages; | 2087 | p->numa_faults_locality[local] += pages; |
2088 | } | 2088 | } |
2089 | 2089 | ||
2090 | static void reset_ptenuma_scan(struct task_struct *p) | 2090 | static void reset_ptenuma_scan(struct task_struct *p) |
2091 | { | 2091 | { |
2092 | ACCESS_ONCE(p->mm->numa_scan_seq)++; | 2092 | ACCESS_ONCE(p->mm->numa_scan_seq)++; |
2093 | p->mm->numa_scan_offset = 0; | 2093 | p->mm->numa_scan_offset = 0; |
2094 | } | 2094 | } |
2095 | 2095 | ||
2096 | /* | 2096 | /* |
2097 | * The expensive part of numa migration is done from task_work context. | 2097 | * The expensive part of numa migration is done from task_work context. |
2098 | * Triggered from task_tick_numa(). | 2098 | * Triggered from task_tick_numa(). |
2099 | */ | 2099 | */ |
2100 | void task_numa_work(struct callback_head *work) | 2100 | void task_numa_work(struct callback_head *work) |
2101 | { | 2101 | { |
2102 | unsigned long migrate, next_scan, now = jiffies; | 2102 | unsigned long migrate, next_scan, now = jiffies; |
2103 | struct task_struct *p = current; | 2103 | struct task_struct *p = current; |
2104 | struct mm_struct *mm = p->mm; | 2104 | struct mm_struct *mm = p->mm; |
2105 | struct vm_area_struct *vma; | 2105 | struct vm_area_struct *vma; |
2106 | unsigned long start, end; | 2106 | unsigned long start, end; |
2107 | unsigned long nr_pte_updates = 0; | 2107 | unsigned long nr_pte_updates = 0; |
2108 | long pages; | 2108 | long pages; |
2109 | 2109 | ||
2110 | WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); | 2110 | WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); |
2111 | 2111 | ||
2112 | work->next = work; /* protect against double add */ | 2112 | work->next = work; /* protect against double add */ |
2113 | /* | 2113 | /* |
2114 | * Who cares about NUMA placement when they're dying. | 2114 | * Who cares about NUMA placement when they're dying. |
2115 | * | 2115 | * |
2116 | * NOTE: make sure not to dereference p->mm before this check, | 2116 | * NOTE: make sure not to dereference p->mm before this check, |
2117 | * exit_task_work() happens _after_ exit_mm() so we could be called | 2117 | * exit_task_work() happens _after_ exit_mm() so we could be called |
2118 | * without p->mm even though we still had it when we enqueued this | 2118 | * without p->mm even though we still had it when we enqueued this |
2119 | * work. | 2119 | * work. |
2120 | */ | 2120 | */ |
2121 | if (p->flags & PF_EXITING) | 2121 | if (p->flags & PF_EXITING) |
2122 | return; | 2122 | return; |
2123 | 2123 | ||
2124 | if (!mm->numa_next_scan) { | 2124 | if (!mm->numa_next_scan) { |
2125 | mm->numa_next_scan = now + | 2125 | mm->numa_next_scan = now + |
2126 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | 2126 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); |
2127 | } | 2127 | } |
2128 | 2128 | ||
2129 | /* | 2129 | /* |
2130 | * Enforce maximal scan/migration frequency.. | 2130 | * Enforce maximal scan/migration frequency.. |
2131 | */ | 2131 | */ |
2132 | migrate = mm->numa_next_scan; | 2132 | migrate = mm->numa_next_scan; |
2133 | if (time_before(now, migrate)) | 2133 | if (time_before(now, migrate)) |
2134 | return; | 2134 | return; |
2135 | 2135 | ||
2136 | if (p->numa_scan_period == 0) { | 2136 | if (p->numa_scan_period == 0) { |
2137 | p->numa_scan_period_max = task_scan_max(p); | 2137 | p->numa_scan_period_max = task_scan_max(p); |
2138 | p->numa_scan_period = task_scan_min(p); | 2138 | p->numa_scan_period = task_scan_min(p); |
2139 | } | 2139 | } |
2140 | 2140 | ||
2141 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); | 2141 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
2142 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) | 2142 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
2143 | return; | 2143 | return; |
2144 | 2144 | ||
2145 | /* | 2145 | /* |
2146 | * Delay this task enough that another task of this mm will likely win | 2146 | * Delay this task enough that another task of this mm will likely win |
2147 | * the next time around. | 2147 | * the next time around. |
2148 | */ | 2148 | */ |
2149 | p->node_stamp += 2 * TICK_NSEC; | 2149 | p->node_stamp += 2 * TICK_NSEC; |
2150 | 2150 | ||
2151 | start = mm->numa_scan_offset; | 2151 | start = mm->numa_scan_offset; |
2152 | pages = sysctl_numa_balancing_scan_size; | 2152 | pages = sysctl_numa_balancing_scan_size; |
2153 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | 2153 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ |
2154 | if (!pages) | 2154 | if (!pages) |
2155 | return; | 2155 | return; |
2156 | 2156 | ||
2157 | down_read(&mm->mmap_sem); | 2157 | down_read(&mm->mmap_sem); |
2158 | vma = find_vma(mm, start); | 2158 | vma = find_vma(mm, start); |
2159 | if (!vma) { | 2159 | if (!vma) { |
2160 | reset_ptenuma_scan(p); | 2160 | reset_ptenuma_scan(p); |
2161 | start = 0; | 2161 | start = 0; |
2162 | vma = mm->mmap; | 2162 | vma = mm->mmap; |
2163 | } | 2163 | } |
2164 | for (; vma; vma = vma->vm_next) { | 2164 | for (; vma; vma = vma->vm_next) { |
2165 | if (!vma_migratable(vma) || !vma_policy_mof(vma)) | 2165 | if (!vma_migratable(vma) || !vma_policy_mof(vma)) |
2166 | continue; | 2166 | continue; |
2167 | 2167 | ||
2168 | /* | 2168 | /* |
2169 | * Shared library pages mapped by multiple processes are not | 2169 | * Shared library pages mapped by multiple processes are not |
2170 | * migrated as it is expected they are cache replicated. Avoid | 2170 | * migrated as it is expected they are cache replicated. Avoid |
2171 | * hinting faults in read-only file-backed mappings or the vdso | 2171 | * hinting faults in read-only file-backed mappings or the vdso |
2172 | * as migrating the pages will be of marginal benefit. | 2172 | * as migrating the pages will be of marginal benefit. |
2173 | */ | 2173 | */ |
2174 | if (!vma->vm_mm || | 2174 | if (!vma->vm_mm || |
2175 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | 2175 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) |
2176 | continue; | 2176 | continue; |
2177 | 2177 | ||
2178 | /* | 2178 | /* |
2179 | * Skip inaccessible VMAs to avoid any confusion between | 2179 | * Skip inaccessible VMAs to avoid any confusion between |
2180 | * PROT_NONE and NUMA hinting ptes | 2180 | * PROT_NONE and NUMA hinting ptes |
2181 | */ | 2181 | */ |
2182 | if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))) | 2182 | if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))) |
2183 | continue; | 2183 | continue; |
2184 | 2184 | ||
2185 | do { | 2185 | do { |
2186 | start = max(start, vma->vm_start); | 2186 | start = max(start, vma->vm_start); |
2187 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | 2187 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); |
2188 | end = min(end, vma->vm_end); | 2188 | end = min(end, vma->vm_end); |
2189 | nr_pte_updates += change_prot_numa(vma, start, end); | 2189 | nr_pte_updates += change_prot_numa(vma, start, end); |
2190 | 2190 | ||
2191 | /* | 2191 | /* |
2192 | * Scan sysctl_numa_balancing_scan_size but ensure that | 2192 | * Scan sysctl_numa_balancing_scan_size but ensure that |
2193 | * at least one PTE is updated so that unused virtual | 2193 | * at least one PTE is updated so that unused virtual |
2194 | * address space is quickly skipped. | 2194 | * address space is quickly skipped. |
2195 | */ | 2195 | */ |
2196 | if (nr_pte_updates) | 2196 | if (nr_pte_updates) |
2197 | pages -= (end - start) >> PAGE_SHIFT; | 2197 | pages -= (end - start) >> PAGE_SHIFT; |
2198 | 2198 | ||
2199 | start = end; | 2199 | start = end; |
2200 | if (pages <= 0) | 2200 | if (pages <= 0) |
2201 | goto out; | 2201 | goto out; |
2202 | 2202 | ||
2203 | cond_resched(); | 2203 | cond_resched(); |
2204 | } while (end != vma->vm_end); | 2204 | } while (end != vma->vm_end); |
2205 | } | 2205 | } |
2206 | 2206 | ||
2207 | out: | 2207 | out: |
2208 | /* | 2208 | /* |
2209 | * It is possible to reach the end of the VMA list but the last few | 2209 | * It is possible to reach the end of the VMA list but the last few |
2210 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | 2210 | * VMAs are not guaranteed to the vma_migratable. If they are not, we |
2211 | * would find the !migratable VMA on the next scan but not reset the | 2211 | * would find the !migratable VMA on the next scan but not reset the |
2212 | * scanner to the start so check it now. | 2212 | * scanner to the start so check it now. |
2213 | */ | 2213 | */ |
2214 | if (vma) | 2214 | if (vma) |
2215 | mm->numa_scan_offset = start; | 2215 | mm->numa_scan_offset = start; |
2216 | else | 2216 | else |
2217 | reset_ptenuma_scan(p); | 2217 | reset_ptenuma_scan(p); |
2218 | up_read(&mm->mmap_sem); | 2218 | up_read(&mm->mmap_sem); |
2219 | } | 2219 | } |
2220 | 2220 | ||
2221 | /* | 2221 | /* |
2222 | * Drive the periodic memory faults.. | 2222 | * Drive the periodic memory faults.. |
2223 | */ | 2223 | */ |
2224 | void task_tick_numa(struct rq *rq, struct task_struct *curr) | 2224 | void task_tick_numa(struct rq *rq, struct task_struct *curr) |
2225 | { | 2225 | { |
2226 | struct callback_head *work = &curr->numa_work; | 2226 | struct callback_head *work = &curr->numa_work; |
2227 | u64 period, now; | 2227 | u64 period, now; |
2228 | 2228 | ||
2229 | /* | 2229 | /* |
2230 | * We don't care about NUMA placement if we don't have memory. | 2230 | * We don't care about NUMA placement if we don't have memory. |
2231 | */ | 2231 | */ |
2232 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | 2232 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) |
2233 | return; | 2233 | return; |
2234 | 2234 | ||
2235 | /* | 2235 | /* |
2236 | * Using runtime rather than walltime has the dual advantage that | 2236 | * Using runtime rather than walltime has the dual advantage that |
2237 | * we (mostly) drive the selection from busy threads and that the | 2237 | * we (mostly) drive the selection from busy threads and that the |
2238 | * task needs to have done some actual work before we bother with | 2238 | * task needs to have done some actual work before we bother with |
2239 | * NUMA placement. | 2239 | * NUMA placement. |
2240 | */ | 2240 | */ |
2241 | now = curr->se.sum_exec_runtime; | 2241 | now = curr->se.sum_exec_runtime; |
2242 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | 2242 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; |
2243 | 2243 | ||
2244 | if (now - curr->node_stamp > period) { | 2244 | if (now - curr->node_stamp > period) { |
2245 | if (!curr->node_stamp) | 2245 | if (!curr->node_stamp) |
2246 | curr->numa_scan_period = task_scan_min(curr); | 2246 | curr->numa_scan_period = task_scan_min(curr); |
2247 | curr->node_stamp += period; | 2247 | curr->node_stamp += period; |
2248 | 2248 | ||
2249 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { | 2249 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { |
2250 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ | 2250 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ |
2251 | task_work_add(curr, work, true); | 2251 | task_work_add(curr, work, true); |
2252 | } | 2252 | } |
2253 | } | 2253 | } |
2254 | } | 2254 | } |
2255 | #else | 2255 | #else |
2256 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | 2256 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) |
2257 | { | 2257 | { |
2258 | } | 2258 | } |
2259 | 2259 | ||
2260 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) | 2260 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
2261 | { | 2261 | { |
2262 | } | 2262 | } |
2263 | 2263 | ||
2264 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) | 2264 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) |
2265 | { | 2265 | { |
2266 | } | 2266 | } |
2267 | #endif /* CONFIG_NUMA_BALANCING */ | 2267 | #endif /* CONFIG_NUMA_BALANCING */ |
2268 | 2268 | ||
2269 | static void | 2269 | static void |
2270 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | 2270 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2271 | { | 2271 | { |
2272 | update_load_add(&cfs_rq->load, se->load.weight); | 2272 | update_load_add(&cfs_rq->load, se->load.weight); |
2273 | if (!parent_entity(se)) | 2273 | if (!parent_entity(se)) |
2274 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); | 2274 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
2275 | #ifdef CONFIG_SMP | 2275 | #ifdef CONFIG_SMP |
2276 | if (entity_is_task(se)) { | 2276 | if (entity_is_task(se)) { |
2277 | struct rq *rq = rq_of(cfs_rq); | 2277 | struct rq *rq = rq_of(cfs_rq); |
2278 | 2278 | ||
2279 | account_numa_enqueue(rq, task_of(se)); | 2279 | account_numa_enqueue(rq, task_of(se)); |
2280 | list_add(&se->group_node, &rq->cfs_tasks); | 2280 | list_add(&se->group_node, &rq->cfs_tasks); |
2281 | } | 2281 | } |
2282 | #endif | 2282 | #endif |
2283 | cfs_rq->nr_running++; | 2283 | cfs_rq->nr_running++; |
2284 | } | 2284 | } |
2285 | 2285 | ||
2286 | static void | 2286 | static void |
2287 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | 2287 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2288 | { | 2288 | { |
2289 | update_load_sub(&cfs_rq->load, se->load.weight); | 2289 | update_load_sub(&cfs_rq->load, se->load.weight); |
2290 | if (!parent_entity(se)) | 2290 | if (!parent_entity(se)) |
2291 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); | 2291 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
2292 | if (entity_is_task(se)) { | 2292 | if (entity_is_task(se)) { |
2293 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); | 2293 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); |
2294 | list_del_init(&se->group_node); | 2294 | list_del_init(&se->group_node); |
2295 | } | 2295 | } |
2296 | cfs_rq->nr_running--; | 2296 | cfs_rq->nr_running--; |
2297 | } | 2297 | } |
2298 | 2298 | ||
2299 | #ifdef CONFIG_FAIR_GROUP_SCHED | 2299 | #ifdef CONFIG_FAIR_GROUP_SCHED |
2300 | # ifdef CONFIG_SMP | 2300 | # ifdef CONFIG_SMP |
2301 | static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) | 2301 | static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) |
2302 | { | 2302 | { |
2303 | long tg_weight; | 2303 | long tg_weight; |
2304 | 2304 | ||
2305 | /* | 2305 | /* |
2306 | * Use this CPU's actual weight instead of the last load_contribution | 2306 | * Use this CPU's actual weight instead of the last load_contribution |
2307 | * to gain a more accurate current total weight. See | 2307 | * to gain a more accurate current total weight. See |
2308 | * update_cfs_rq_load_contribution(). | 2308 | * update_cfs_rq_load_contribution(). |
2309 | */ | 2309 | */ |
2310 | tg_weight = atomic_long_read(&tg->load_avg); | 2310 | tg_weight = atomic_long_read(&tg->load_avg); |
2311 | tg_weight -= cfs_rq->tg_load_contrib; | 2311 | tg_weight -= cfs_rq->tg_load_contrib; |
2312 | tg_weight += cfs_rq->load.weight; | 2312 | tg_weight += cfs_rq->load.weight; |
2313 | 2313 | ||
2314 | return tg_weight; | 2314 | return tg_weight; |
2315 | } | 2315 | } |
2316 | 2316 | ||
2317 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) | 2317 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
2318 | { | 2318 | { |
2319 | long tg_weight, load, shares; | 2319 | long tg_weight, load, shares; |
2320 | 2320 | ||
2321 | tg_weight = calc_tg_weight(tg, cfs_rq); | 2321 | tg_weight = calc_tg_weight(tg, cfs_rq); |
2322 | load = cfs_rq->load.weight; | 2322 | load = cfs_rq->load.weight; |
2323 | 2323 | ||
2324 | shares = (tg->shares * load); | 2324 | shares = (tg->shares * load); |
2325 | if (tg_weight) | 2325 | if (tg_weight) |
2326 | shares /= tg_weight; | 2326 | shares /= tg_weight; |
2327 | 2327 | ||
2328 | if (shares < MIN_SHARES) | 2328 | if (shares < MIN_SHARES) |
2329 | shares = MIN_SHARES; | 2329 | shares = MIN_SHARES; |
2330 | if (shares > tg->shares) | 2330 | if (shares > tg->shares) |
2331 | shares = tg->shares; | 2331 | shares = tg->shares; |
2332 | 2332 | ||
2333 | return shares; | 2333 | return shares; |
2334 | } | 2334 | } |
2335 | # else /* CONFIG_SMP */ | 2335 | # else /* CONFIG_SMP */ |
2336 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) | 2336 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
2337 | { | 2337 | { |
2338 | return tg->shares; | 2338 | return tg->shares; |
2339 | } | 2339 | } |
2340 | # endif /* CONFIG_SMP */ | 2340 | # endif /* CONFIG_SMP */ |
2341 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, | 2341 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
2342 | unsigned long weight) | 2342 | unsigned long weight) |
2343 | { | 2343 | { |
2344 | if (se->on_rq) { | 2344 | if (se->on_rq) { |
2345 | /* commit outstanding execution time */ | 2345 | /* commit outstanding execution time */ |
2346 | if (cfs_rq->curr == se) | 2346 | if (cfs_rq->curr == se) |
2347 | update_curr(cfs_rq); | 2347 | update_curr(cfs_rq); |
2348 | account_entity_dequeue(cfs_rq, se); | 2348 | account_entity_dequeue(cfs_rq, se); |
2349 | } | 2349 | } |
2350 | 2350 | ||
2351 | update_load_set(&se->load, weight); | 2351 | update_load_set(&se->load, weight); |
2352 | 2352 | ||
2353 | if (se->on_rq) | 2353 | if (se->on_rq) |
2354 | account_entity_enqueue(cfs_rq, se); | 2354 | account_entity_enqueue(cfs_rq, se); |
2355 | } | 2355 | } |
2356 | 2356 | ||
2357 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); | 2357 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
2358 | 2358 | ||
2359 | static void update_cfs_shares(struct cfs_rq *cfs_rq) | 2359 | static void update_cfs_shares(struct cfs_rq *cfs_rq) |
2360 | { | 2360 | { |
2361 | struct task_group *tg; | 2361 | struct task_group *tg; |
2362 | struct sched_entity *se; | 2362 | struct sched_entity *se; |
2363 | long shares; | 2363 | long shares; |
2364 | 2364 | ||
2365 | tg = cfs_rq->tg; | 2365 | tg = cfs_rq->tg; |
2366 | se = tg->se[cpu_of(rq_of(cfs_rq))]; | 2366 | se = tg->se[cpu_of(rq_of(cfs_rq))]; |
2367 | if (!se || throttled_hierarchy(cfs_rq)) | 2367 | if (!se || throttled_hierarchy(cfs_rq)) |
2368 | return; | 2368 | return; |
2369 | #ifndef CONFIG_SMP | 2369 | #ifndef CONFIG_SMP |
2370 | if (likely(se->load.weight == tg->shares)) | 2370 | if (likely(se->load.weight == tg->shares)) |
2371 | return; | 2371 | return; |
2372 | #endif | 2372 | #endif |
2373 | shares = calc_cfs_shares(cfs_rq, tg); | 2373 | shares = calc_cfs_shares(cfs_rq, tg); |
2374 | 2374 | ||
2375 | reweight_entity(cfs_rq_of(se), se, shares); | 2375 | reweight_entity(cfs_rq_of(se), se, shares); |
2376 | } | 2376 | } |
2377 | #else /* CONFIG_FAIR_GROUP_SCHED */ | 2377 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
2378 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) | 2378 | static inline void update_cfs_shares(struct cfs_rq *cfs_rq) |
2379 | { | 2379 | { |
2380 | } | 2380 | } |
2381 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | 2381 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
2382 | 2382 | ||
2383 | #ifdef CONFIG_SMP | 2383 | #ifdef CONFIG_SMP |
2384 | /* | 2384 | /* |
2385 | * We choose a half-life close to 1 scheduling period. | 2385 | * We choose a half-life close to 1 scheduling period. |
2386 | * Note: The tables below are dependent on this value. | 2386 | * Note: The tables below are dependent on this value. |
2387 | */ | 2387 | */ |
2388 | #define LOAD_AVG_PERIOD 32 | 2388 | #define LOAD_AVG_PERIOD 32 |
2389 | #define LOAD_AVG_MAX 47742 /* maximum possible load avg */ | 2389 | #define LOAD_AVG_MAX 47742 /* maximum possible load avg */ |
2390 | #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */ | 2390 | #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */ |
2391 | 2391 | ||
2392 | /* Precomputed fixed inverse multiplies for multiplication by y^n */ | 2392 | /* Precomputed fixed inverse multiplies for multiplication by y^n */ |
2393 | static const u32 runnable_avg_yN_inv[] = { | 2393 | static const u32 runnable_avg_yN_inv[] = { |
2394 | 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, | 2394 | 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, |
2395 | 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, | 2395 | 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, |
2396 | 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, | 2396 | 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, |
2397 | 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, | 2397 | 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, |
2398 | 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, | 2398 | 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, |
2399 | 0x85aac367, 0x82cd8698, | 2399 | 0x85aac367, 0x82cd8698, |
2400 | }; | 2400 | }; |
2401 | 2401 | ||
2402 | /* | 2402 | /* |
2403 | * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent | 2403 | * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent |
2404 | * over-estimates when re-combining. | 2404 | * over-estimates when re-combining. |
2405 | */ | 2405 | */ |
2406 | static const u32 runnable_avg_yN_sum[] = { | 2406 | static const u32 runnable_avg_yN_sum[] = { |
2407 | 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103, | 2407 | 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103, |
2408 | 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082, | 2408 | 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082, |
2409 | 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371, | 2409 | 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371, |
2410 | }; | 2410 | }; |
2411 | 2411 | ||
2412 | /* | 2412 | /* |
2413 | * Approximate: | 2413 | * Approximate: |
2414 | * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) | 2414 | * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) |
2415 | */ | 2415 | */ |
2416 | static __always_inline u64 decay_load(u64 val, u64 n) | 2416 | static __always_inline u64 decay_load(u64 val, u64 n) |
2417 | { | 2417 | { |
2418 | unsigned int local_n; | 2418 | unsigned int local_n; |
2419 | 2419 | ||
2420 | if (!n) | 2420 | if (!n) |
2421 | return val; | 2421 | return val; |
2422 | else if (unlikely(n > LOAD_AVG_PERIOD * 63)) | 2422 | else if (unlikely(n > LOAD_AVG_PERIOD * 63)) |
2423 | return 0; | 2423 | return 0; |
2424 | 2424 | ||
2425 | /* after bounds checking we can collapse to 32-bit */ | 2425 | /* after bounds checking we can collapse to 32-bit */ |
2426 | local_n = n; | 2426 | local_n = n; |
2427 | 2427 | ||
2428 | /* | 2428 | /* |
2429 | * As y^PERIOD = 1/2, we can combine | 2429 | * As y^PERIOD = 1/2, we can combine |
2430 | * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD) | 2430 | * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD) |
2431 | * With a look-up table which covers y^n (n<PERIOD) | 2431 | * With a look-up table which covers y^n (n<PERIOD) |
2432 | * | 2432 | * |
2433 | * To achieve constant time decay_load. | 2433 | * To achieve constant time decay_load. |
2434 | */ | 2434 | */ |
2435 | if (unlikely(local_n >= LOAD_AVG_PERIOD)) { | 2435 | if (unlikely(local_n >= LOAD_AVG_PERIOD)) { |
2436 | val >>= local_n / LOAD_AVG_PERIOD; | 2436 | val >>= local_n / LOAD_AVG_PERIOD; |
2437 | local_n %= LOAD_AVG_PERIOD; | 2437 | local_n %= LOAD_AVG_PERIOD; |
2438 | } | 2438 | } |
2439 | 2439 | ||
2440 | val *= runnable_avg_yN_inv[local_n]; | 2440 | val *= runnable_avg_yN_inv[local_n]; |
2441 | /* We don't use SRR here since we always want to round down. */ | 2441 | /* We don't use SRR here since we always want to round down. */ |
2442 | return val >> 32; | 2442 | return val >> 32; |
2443 | } | 2443 | } |
2444 | 2444 | ||
2445 | /* | 2445 | /* |
2446 | * For updates fully spanning n periods, the contribution to runnable | 2446 | * For updates fully spanning n periods, the contribution to runnable |
2447 | * average will be: \Sum 1024*y^n | 2447 | * average will be: \Sum 1024*y^n |
2448 | * | 2448 | * |
2449 | * We can compute this reasonably efficiently by combining: | 2449 | * We can compute this reasonably efficiently by combining: |
2450 | * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD} | 2450 | * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD} |
2451 | */ | 2451 | */ |
2452 | static u32 __compute_runnable_contrib(u64 n) | 2452 | static u32 __compute_runnable_contrib(u64 n) |
2453 | { | 2453 | { |
2454 | u32 contrib = 0; | 2454 | u32 contrib = 0; |
2455 | 2455 | ||
2456 | if (likely(n <= LOAD_AVG_PERIOD)) | 2456 | if (likely(n <= LOAD_AVG_PERIOD)) |
2457 | return runnable_avg_yN_sum[n]; | 2457 | return runnable_avg_yN_sum[n]; |
2458 | else if (unlikely(n >= LOAD_AVG_MAX_N)) | 2458 | else if (unlikely(n >= LOAD_AVG_MAX_N)) |
2459 | return LOAD_AVG_MAX; | 2459 | return LOAD_AVG_MAX; |
2460 | 2460 | ||
2461 | /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */ | 2461 | /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */ |
2462 | do { | 2462 | do { |
2463 | contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */ | 2463 | contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */ |
2464 | contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD]; | 2464 | contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD]; |
2465 | 2465 | ||
2466 | n -= LOAD_AVG_PERIOD; | 2466 | n -= LOAD_AVG_PERIOD; |
2467 | } while (n > LOAD_AVG_PERIOD); | 2467 | } while (n > LOAD_AVG_PERIOD); |
2468 | 2468 | ||
2469 | contrib = decay_load(contrib, n); | 2469 | contrib = decay_load(contrib, n); |
2470 | return contrib + runnable_avg_yN_sum[n]; | 2470 | return contrib + runnable_avg_yN_sum[n]; |
2471 | } | 2471 | } |
2472 | 2472 | ||
2473 | /* | 2473 | /* |
2474 | * We can represent the historical contribution to runnable average as the | 2474 | * We can represent the historical contribution to runnable average as the |
2475 | * coefficients of a geometric series. To do this we sub-divide our runnable | 2475 | * coefficients of a geometric series. To do this we sub-divide our runnable |
2476 | * history into segments of approximately 1ms (1024us); label the segment that | 2476 | * history into segments of approximately 1ms (1024us); label the segment that |
2477 | * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. | 2477 | * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. |
2478 | * | 2478 | * |
2479 | * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... | 2479 | * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... |
2480 | * p0 p1 p2 | 2480 | * p0 p1 p2 |
2481 | * (now) (~1ms ago) (~2ms ago) | 2481 | * (now) (~1ms ago) (~2ms ago) |
2482 | * | 2482 | * |
2483 | * Let u_i denote the fraction of p_i that the entity was runnable. | 2483 | * Let u_i denote the fraction of p_i that the entity was runnable. |
2484 | * | 2484 | * |
2485 | * We then designate the fractions u_i as our co-efficients, yielding the | 2485 | * We then designate the fractions u_i as our co-efficients, yielding the |
2486 | * following representation of historical load: | 2486 | * following representation of historical load: |
2487 | * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... | 2487 | * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... |
2488 | * | 2488 | * |
2489 | * We choose y based on the with of a reasonably scheduling period, fixing: | 2489 | * We choose y based on the with of a reasonably scheduling period, fixing: |
2490 | * y^32 = 0.5 | 2490 | * y^32 = 0.5 |
2491 | * | 2491 | * |
2492 | * This means that the contribution to load ~32ms ago (u_32) will be weighted | 2492 | * This means that the contribution to load ~32ms ago (u_32) will be weighted |
2493 | * approximately half as much as the contribution to load within the last ms | 2493 | * approximately half as much as the contribution to load within the last ms |
2494 | * (u_0). | 2494 | * (u_0). |
2495 | * | 2495 | * |
2496 | * When a period "rolls over" and we have new u_0`, multiplying the previous | 2496 | * When a period "rolls over" and we have new u_0`, multiplying the previous |
2497 | * sum again by y is sufficient to update: | 2497 | * sum again by y is sufficient to update: |
2498 | * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) | 2498 | * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) |
2499 | * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] | 2499 | * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] |
2500 | */ | 2500 | */ |
2501 | static __always_inline int __update_entity_runnable_avg(u64 now, | 2501 | static __always_inline int __update_entity_runnable_avg(u64 now, |
2502 | struct sched_avg *sa, | 2502 | struct sched_avg *sa, |
2503 | int runnable) | 2503 | int runnable) |
2504 | { | 2504 | { |
2505 | u64 delta, periods; | 2505 | u64 delta, periods; |
2506 | u32 runnable_contrib; | 2506 | u32 runnable_contrib; |
2507 | int delta_w, decayed = 0; | 2507 | int delta_w, decayed = 0; |
2508 | 2508 | ||
2509 | delta = now - sa->last_runnable_update; | 2509 | delta = now - sa->last_runnable_update; |
2510 | /* | 2510 | /* |
2511 | * This should only happen when time goes backwards, which it | 2511 | * This should only happen when time goes backwards, which it |
2512 | * unfortunately does during sched clock init when we swap over to TSC. | 2512 | * unfortunately does during sched clock init when we swap over to TSC. |
2513 | */ | 2513 | */ |
2514 | if ((s64)delta < 0) { | 2514 | if ((s64)delta < 0) { |
2515 | sa->last_runnable_update = now; | 2515 | sa->last_runnable_update = now; |
2516 | return 0; | 2516 | return 0; |
2517 | } | 2517 | } |
2518 | 2518 | ||
2519 | /* | 2519 | /* |
2520 | * Use 1024ns as the unit of measurement since it's a reasonable | 2520 | * Use 1024ns as the unit of measurement since it's a reasonable |
2521 | * approximation of 1us and fast to compute. | 2521 | * approximation of 1us and fast to compute. |
2522 | */ | 2522 | */ |
2523 | delta >>= 10; | 2523 | delta >>= 10; |
2524 | if (!delta) | 2524 | if (!delta) |
2525 | return 0; | 2525 | return 0; |
2526 | sa->last_runnable_update = now; | 2526 | sa->last_runnable_update = now; |
2527 | 2527 | ||
2528 | /* delta_w is the amount already accumulated against our next period */ | 2528 | /* delta_w is the amount already accumulated against our next period */ |
2529 | delta_w = sa->runnable_avg_period % 1024; | 2529 | delta_w = sa->runnable_avg_period % 1024; |
2530 | if (delta + delta_w >= 1024) { | 2530 | if (delta + delta_w >= 1024) { |
2531 | /* period roll-over */ | 2531 | /* period roll-over */ |
2532 | decayed = 1; | 2532 | decayed = 1; |
2533 | 2533 | ||
2534 | /* | 2534 | /* |
2535 | * Now that we know we're crossing a period boundary, figure | 2535 | * Now that we know we're crossing a period boundary, figure |
2536 | * out how much from delta we need to complete the current | 2536 | * out how much from delta we need to complete the current |
2537 | * period and accrue it. | 2537 | * period and accrue it. |
2538 | */ | 2538 | */ |
2539 | delta_w = 1024 - delta_w; | 2539 | delta_w = 1024 - delta_w; |
2540 | if (runnable) | 2540 | if (runnable) |
2541 | sa->runnable_avg_sum += delta_w; | 2541 | sa->runnable_avg_sum += delta_w; |
2542 | sa->runnable_avg_period += delta_w; | 2542 | sa->runnable_avg_period += delta_w; |
2543 | 2543 | ||
2544 | delta -= delta_w; | 2544 | delta -= delta_w; |
2545 | 2545 | ||
2546 | /* Figure out how many additional periods this update spans */ | 2546 | /* Figure out how many additional periods this update spans */ |
2547 | periods = delta / 1024; | 2547 | periods = delta / 1024; |
2548 | delta %= 1024; | 2548 | delta %= 1024; |
2549 | 2549 | ||
2550 | sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum, | 2550 | sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum, |
2551 | periods + 1); | 2551 | periods + 1); |
2552 | sa->runnable_avg_period = decay_load(sa->runnable_avg_period, | 2552 | sa->runnable_avg_period = decay_load(sa->runnable_avg_period, |
2553 | periods + 1); | 2553 | periods + 1); |
2554 | 2554 | ||
2555 | /* Efficiently calculate \sum (1..n_period) 1024*y^i */ | 2555 | /* Efficiently calculate \sum (1..n_period) 1024*y^i */ |
2556 | runnable_contrib = __compute_runnable_contrib(periods); | 2556 | runnable_contrib = __compute_runnable_contrib(periods); |
2557 | if (runnable) | 2557 | if (runnable) |
2558 | sa->runnable_avg_sum += runnable_contrib; | 2558 | sa->runnable_avg_sum += runnable_contrib; |
2559 | sa->runnable_avg_period += runnable_contrib; | 2559 | sa->runnable_avg_period += runnable_contrib; |
2560 | } | 2560 | } |
2561 | 2561 | ||
2562 | /* Remainder of delta accrued against u_0` */ | 2562 | /* Remainder of delta accrued against u_0` */ |
2563 | if (runnable) | 2563 | if (runnable) |
2564 | sa->runnable_avg_sum += delta; | 2564 | sa->runnable_avg_sum += delta; |
2565 | sa->runnable_avg_period += delta; | 2565 | sa->runnable_avg_period += delta; |
2566 | 2566 | ||
2567 | return decayed; | 2567 | return decayed; |
2568 | } | 2568 | } |
2569 | 2569 | ||
2570 | /* Synchronize an entity's decay with its parenting cfs_rq.*/ | 2570 | /* Synchronize an entity's decay with its parenting cfs_rq.*/ |
2571 | static inline u64 __synchronize_entity_decay(struct sched_entity *se) | 2571 | static inline u64 __synchronize_entity_decay(struct sched_entity *se) |
2572 | { | 2572 | { |
2573 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | 2573 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
2574 | u64 decays = atomic64_read(&cfs_rq->decay_counter); | 2574 | u64 decays = atomic64_read(&cfs_rq->decay_counter); |
2575 | 2575 | ||
2576 | decays -= se->avg.decay_count; | 2576 | decays -= se->avg.decay_count; |
2577 | if (!decays) | 2577 | if (!decays) |
2578 | return 0; | 2578 | return 0; |
2579 | 2579 | ||
2580 | se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays); | 2580 | se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays); |
2581 | se->avg.decay_count = 0; | 2581 | se->avg.decay_count = 0; |
2582 | 2582 | ||
2583 | return decays; | 2583 | return decays; |
2584 | } | 2584 | } |
2585 | 2585 | ||
2586 | #ifdef CONFIG_FAIR_GROUP_SCHED | 2586 | #ifdef CONFIG_FAIR_GROUP_SCHED |
2587 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | 2587 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, |
2588 | int force_update) | 2588 | int force_update) |
2589 | { | 2589 | { |
2590 | struct task_group *tg = cfs_rq->tg; | 2590 | struct task_group *tg = cfs_rq->tg; |
2591 | long tg_contrib; | 2591 | long tg_contrib; |
2592 | 2592 | ||
2593 | tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg; | 2593 | tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg; |
2594 | tg_contrib -= cfs_rq->tg_load_contrib; | 2594 | tg_contrib -= cfs_rq->tg_load_contrib; |
2595 | 2595 | ||
2596 | if (!tg_contrib) | 2596 | if (!tg_contrib) |
2597 | return; | 2597 | return; |
2598 | 2598 | ||
2599 | if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) { | 2599 | if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) { |
2600 | atomic_long_add(tg_contrib, &tg->load_avg); | 2600 | atomic_long_add(tg_contrib, &tg->load_avg); |
2601 | cfs_rq->tg_load_contrib += tg_contrib; | 2601 | cfs_rq->tg_load_contrib += tg_contrib; |
2602 | } | 2602 | } |
2603 | } | 2603 | } |
2604 | 2604 | ||
2605 | /* | 2605 | /* |
2606 | * Aggregate cfs_rq runnable averages into an equivalent task_group | 2606 | * Aggregate cfs_rq runnable averages into an equivalent task_group |
2607 | * representation for computing load contributions. | 2607 | * representation for computing load contributions. |
2608 | */ | 2608 | */ |
2609 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, | 2609 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, |
2610 | struct cfs_rq *cfs_rq) | 2610 | struct cfs_rq *cfs_rq) |
2611 | { | 2611 | { |
2612 | struct task_group *tg = cfs_rq->tg; | 2612 | struct task_group *tg = cfs_rq->tg; |
2613 | long contrib; | 2613 | long contrib; |
2614 | 2614 | ||
2615 | /* The fraction of a cpu used by this cfs_rq */ | 2615 | /* The fraction of a cpu used by this cfs_rq */ |
2616 | contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT, | 2616 | contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT, |
2617 | sa->runnable_avg_period + 1); | 2617 | sa->runnable_avg_period + 1); |
2618 | contrib -= cfs_rq->tg_runnable_contrib; | 2618 | contrib -= cfs_rq->tg_runnable_contrib; |
2619 | 2619 | ||
2620 | if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) { | 2620 | if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) { |
2621 | atomic_add(contrib, &tg->runnable_avg); | 2621 | atomic_add(contrib, &tg->runnable_avg); |
2622 | cfs_rq->tg_runnable_contrib += contrib; | 2622 | cfs_rq->tg_runnable_contrib += contrib; |
2623 | } | 2623 | } |
2624 | } | 2624 | } |
2625 | 2625 | ||
2626 | static inline void __update_group_entity_contrib(struct sched_entity *se) | 2626 | static inline void __update_group_entity_contrib(struct sched_entity *se) |
2627 | { | 2627 | { |
2628 | struct cfs_rq *cfs_rq = group_cfs_rq(se); | 2628 | struct cfs_rq *cfs_rq = group_cfs_rq(se); |
2629 | struct task_group *tg = cfs_rq->tg; | 2629 | struct task_group *tg = cfs_rq->tg; |
2630 | int runnable_avg; | 2630 | int runnable_avg; |
2631 | 2631 | ||
2632 | u64 contrib; | 2632 | u64 contrib; |
2633 | 2633 | ||
2634 | contrib = cfs_rq->tg_load_contrib * tg->shares; | 2634 | contrib = cfs_rq->tg_load_contrib * tg->shares; |
2635 | se->avg.load_avg_contrib = div_u64(contrib, | 2635 | se->avg.load_avg_contrib = div_u64(contrib, |
2636 | atomic_long_read(&tg->load_avg) + 1); | 2636 | atomic_long_read(&tg->load_avg) + 1); |
2637 | 2637 | ||
2638 | /* | 2638 | /* |
2639 | * For group entities we need to compute a correction term in the case | 2639 | * For group entities we need to compute a correction term in the case |
2640 | * that they are consuming <1 cpu so that we would contribute the same | 2640 | * that they are consuming <1 cpu so that we would contribute the same |
2641 | * load as a task of equal weight. | 2641 | * load as a task of equal weight. |
2642 | * | 2642 | * |
2643 | * Explicitly co-ordinating this measurement would be expensive, but | 2643 | * Explicitly co-ordinating this measurement would be expensive, but |
2644 | * fortunately the sum of each cpus contribution forms a usable | 2644 | * fortunately the sum of each cpus contribution forms a usable |
2645 | * lower-bound on the true value. | 2645 | * lower-bound on the true value. |
2646 | * | 2646 | * |
2647 | * Consider the aggregate of 2 contributions. Either they are disjoint | 2647 | * Consider the aggregate of 2 contributions. Either they are disjoint |
2648 | * (and the sum represents true value) or they are disjoint and we are | 2648 | * (and the sum represents true value) or they are disjoint and we are |
2649 | * understating by the aggregate of their overlap. | 2649 | * understating by the aggregate of their overlap. |
2650 | * | 2650 | * |
2651 | * Extending this to N cpus, for a given overlap, the maximum amount we | 2651 | * Extending this to N cpus, for a given overlap, the maximum amount we |
2652 | * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of | 2652 | * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of |
2653 | * cpus that overlap for this interval and w_i is the interval width. | 2653 | * cpus that overlap for this interval and w_i is the interval width. |
2654 | * | 2654 | * |
2655 | * On a small machine; the first term is well-bounded which bounds the | 2655 | * On a small machine; the first term is well-bounded which bounds the |
2656 | * total error since w_i is a subset of the period. Whereas on a | 2656 | * total error since w_i is a subset of the period. Whereas on a |
2657 | * larger machine, while this first term can be larger, if w_i is the | 2657 | * larger machine, while this first term can be larger, if w_i is the |
2658 | * of consequential size guaranteed to see n_i*w_i quickly converge to | 2658 | * of consequential size guaranteed to see n_i*w_i quickly converge to |
2659 | * our upper bound of 1-cpu. | 2659 | * our upper bound of 1-cpu. |
2660 | */ | 2660 | */ |
2661 | runnable_avg = atomic_read(&tg->runnable_avg); | 2661 | runnable_avg = atomic_read(&tg->runnable_avg); |
2662 | if (runnable_avg < NICE_0_LOAD) { | 2662 | if (runnable_avg < NICE_0_LOAD) { |
2663 | se->avg.load_avg_contrib *= runnable_avg; | 2663 | se->avg.load_avg_contrib *= runnable_avg; |
2664 | se->avg.load_avg_contrib >>= NICE_0_SHIFT; | 2664 | se->avg.load_avg_contrib >>= NICE_0_SHIFT; |
2665 | } | 2665 | } |
2666 | } | 2666 | } |
2667 | 2667 | ||
2668 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) | 2668 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) |
2669 | { | 2669 | { |
2670 | __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable); | 2670 | __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable); |
2671 | __update_tg_runnable_avg(&rq->avg, &rq->cfs); | 2671 | __update_tg_runnable_avg(&rq->avg, &rq->cfs); |
2672 | } | 2672 | } |
2673 | #else /* CONFIG_FAIR_GROUP_SCHED */ | 2673 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
2674 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, | 2674 | static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, |
2675 | int force_update) {} | 2675 | int force_update) {} |
2676 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, | 2676 | static inline void __update_tg_runnable_avg(struct sched_avg *sa, |
2677 | struct cfs_rq *cfs_rq) {} | 2677 | struct cfs_rq *cfs_rq) {} |
2678 | static inline void __update_group_entity_contrib(struct sched_entity *se) {} | 2678 | static inline void __update_group_entity_contrib(struct sched_entity *se) {} |
2679 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} | 2679 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} |
2680 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | 2680 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
2681 | 2681 | ||
2682 | static inline void __update_task_entity_contrib(struct sched_entity *se) | 2682 | static inline void __update_task_entity_contrib(struct sched_entity *se) |
2683 | { | 2683 | { |
2684 | u32 contrib; | 2684 | u32 contrib; |
2685 | 2685 | ||
2686 | /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */ | 2686 | /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */ |
2687 | contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight); | 2687 | contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight); |
2688 | contrib /= (se->avg.runnable_avg_period + 1); | 2688 | contrib /= (se->avg.runnable_avg_period + 1); |
2689 | se->avg.load_avg_contrib = scale_load(contrib); | 2689 | se->avg.load_avg_contrib = scale_load(contrib); |
2690 | } | 2690 | } |
2691 | 2691 | ||
2692 | /* Compute the current contribution to load_avg by se, return any delta */ | 2692 | /* Compute the current contribution to load_avg by se, return any delta */ |
2693 | static long __update_entity_load_avg_contrib(struct sched_entity *se) | 2693 | static long __update_entity_load_avg_contrib(struct sched_entity *se) |
2694 | { | 2694 | { |
2695 | long old_contrib = se->avg.load_avg_contrib; | 2695 | long old_contrib = se->avg.load_avg_contrib; |
2696 | 2696 | ||
2697 | if (entity_is_task(se)) { | 2697 | if (entity_is_task(se)) { |
2698 | __update_task_entity_contrib(se); | 2698 | __update_task_entity_contrib(se); |
2699 | } else { | 2699 | } else { |
2700 | __update_tg_runnable_avg(&se->avg, group_cfs_rq(se)); | 2700 | __update_tg_runnable_avg(&se->avg, group_cfs_rq(se)); |
2701 | __update_group_entity_contrib(se); | 2701 | __update_group_entity_contrib(se); |
2702 | } | 2702 | } |
2703 | 2703 | ||
2704 | return se->avg.load_avg_contrib - old_contrib; | 2704 | return se->avg.load_avg_contrib - old_contrib; |
2705 | } | 2705 | } |
2706 | 2706 | ||
2707 | static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq, | 2707 | static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq, |
2708 | long load_contrib) | 2708 | long load_contrib) |
2709 | { | 2709 | { |
2710 | if (likely(load_contrib < cfs_rq->blocked_load_avg)) | 2710 | if (likely(load_contrib < cfs_rq->blocked_load_avg)) |
2711 | cfs_rq->blocked_load_avg -= load_contrib; | 2711 | cfs_rq->blocked_load_avg -= load_contrib; |
2712 | else | 2712 | else |
2713 | cfs_rq->blocked_load_avg = 0; | 2713 | cfs_rq->blocked_load_avg = 0; |
2714 | } | 2714 | } |
2715 | 2715 | ||
2716 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); | 2716 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); |
2717 | 2717 | ||
2718 | /* Update a sched_entity's runnable average */ | 2718 | /* Update a sched_entity's runnable average */ |
2719 | static inline void update_entity_load_avg(struct sched_entity *se, | 2719 | static inline void update_entity_load_avg(struct sched_entity *se, |
2720 | int update_cfs_rq) | 2720 | int update_cfs_rq) |
2721 | { | 2721 | { |
2722 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | 2722 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
2723 | long contrib_delta; | 2723 | long contrib_delta; |
2724 | u64 now; | 2724 | u64 now; |
2725 | 2725 | ||
2726 | /* | 2726 | /* |
2727 | * For a group entity we need to use their owned cfs_rq_clock_task() in | 2727 | * For a group entity we need to use their owned cfs_rq_clock_task() in |
2728 | * case they are the parent of a throttled hierarchy. | 2728 | * case they are the parent of a throttled hierarchy. |
2729 | */ | 2729 | */ |
2730 | if (entity_is_task(se)) | 2730 | if (entity_is_task(se)) |
2731 | now = cfs_rq_clock_task(cfs_rq); | 2731 | now = cfs_rq_clock_task(cfs_rq); |
2732 | else | 2732 | else |
2733 | now = cfs_rq_clock_task(group_cfs_rq(se)); | 2733 | now = cfs_rq_clock_task(group_cfs_rq(se)); |
2734 | 2734 | ||
2735 | if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq)) | 2735 | if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq)) |
2736 | return; | 2736 | return; |
2737 | 2737 | ||
2738 | contrib_delta = __update_entity_load_avg_contrib(se); | 2738 | contrib_delta = __update_entity_load_avg_contrib(se); |
2739 | 2739 | ||
2740 | if (!update_cfs_rq) | 2740 | if (!update_cfs_rq) |
2741 | return; | 2741 | return; |
2742 | 2742 | ||
2743 | if (se->on_rq) | 2743 | if (se->on_rq) |
2744 | cfs_rq->runnable_load_avg += contrib_delta; | 2744 | cfs_rq->runnable_load_avg += contrib_delta; |
2745 | else | 2745 | else |
2746 | subtract_blocked_load_contrib(cfs_rq, -contrib_delta); | 2746 | subtract_blocked_load_contrib(cfs_rq, -contrib_delta); |
2747 | } | 2747 | } |
2748 | 2748 | ||
2749 | /* | 2749 | /* |
2750 | * Decay the load contributed by all blocked children and account this so that | 2750 | * Decay the load contributed by all blocked children and account this so that |
2751 | * their contribution may appropriately discounted when they wake up. | 2751 | * their contribution may appropriately discounted when they wake up. |
2752 | */ | 2752 | */ |
2753 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update) | 2753 | static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update) |
2754 | { | 2754 | { |
2755 | u64 now = cfs_rq_clock_task(cfs_rq) >> 20; | 2755 | u64 now = cfs_rq_clock_task(cfs_rq) >> 20; |
2756 | u64 decays; | 2756 | u64 decays; |
2757 | 2757 | ||
2758 | decays = now - cfs_rq->last_decay; | 2758 | decays = now - cfs_rq->last_decay; |
2759 | if (!decays && !force_update) | 2759 | if (!decays && !force_update) |
2760 | return; | 2760 | return; |
2761 | 2761 | ||
2762 | if (atomic_long_read(&cfs_rq->removed_load)) { | 2762 | if (atomic_long_read(&cfs_rq->removed_load)) { |
2763 | unsigned long removed_load; | 2763 | unsigned long removed_load; |
2764 | removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0); | 2764 | removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0); |
2765 | subtract_blocked_load_contrib(cfs_rq, removed_load); | 2765 | subtract_blocked_load_contrib(cfs_rq, removed_load); |
2766 | } | 2766 | } |
2767 | 2767 | ||
2768 | if (decays) { | 2768 | if (decays) { |
2769 | cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg, | 2769 | cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg, |
2770 | decays); | 2770 | decays); |
2771 | atomic64_add(decays, &cfs_rq->decay_counter); | 2771 | atomic64_add(decays, &cfs_rq->decay_counter); |
2772 | cfs_rq->last_decay = now; | 2772 | cfs_rq->last_decay = now; |
2773 | } | 2773 | } |
2774 | 2774 | ||
2775 | __update_cfs_rq_tg_load_contrib(cfs_rq, force_update); | 2775 | __update_cfs_rq_tg_load_contrib(cfs_rq, force_update); |
2776 | } | 2776 | } |
2777 | 2777 | ||
2778 | /* Add the load generated by se into cfs_rq's child load-average */ | 2778 | /* Add the load generated by se into cfs_rq's child load-average */ |
2779 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, | 2779 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, |
2780 | struct sched_entity *se, | 2780 | struct sched_entity *se, |
2781 | int wakeup) | 2781 | int wakeup) |
2782 | { | 2782 | { |
2783 | /* | 2783 | /* |
2784 | * We track migrations using entity decay_count <= 0, on a wake-up | 2784 | * We track migrations using entity decay_count <= 0, on a wake-up |
2785 | * migration we use a negative decay count to track the remote decays | 2785 | * migration we use a negative decay count to track the remote decays |
2786 | * accumulated while sleeping. | 2786 | * accumulated while sleeping. |
2787 | * | 2787 | * |
2788 | * Newly forked tasks are enqueued with se->avg.decay_count == 0, they | 2788 | * Newly forked tasks are enqueued with se->avg.decay_count == 0, they |
2789 | * are seen by enqueue_entity_load_avg() as a migration with an already | 2789 | * are seen by enqueue_entity_load_avg() as a migration with an already |
2790 | * constructed load_avg_contrib. | 2790 | * constructed load_avg_contrib. |
2791 | */ | 2791 | */ |
2792 | if (unlikely(se->avg.decay_count <= 0)) { | 2792 | if (unlikely(se->avg.decay_count <= 0)) { |
2793 | se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq)); | 2793 | se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq)); |
2794 | if (se->avg.decay_count) { | 2794 | if (se->avg.decay_count) { |
2795 | /* | 2795 | /* |
2796 | * In a wake-up migration we have to approximate the | 2796 | * In a wake-up migration we have to approximate the |
2797 | * time sleeping. This is because we can't synchronize | 2797 | * time sleeping. This is because we can't synchronize |
2798 | * clock_task between the two cpus, and it is not | 2798 | * clock_task between the two cpus, and it is not |
2799 | * guaranteed to be read-safe. Instead, we can | 2799 | * guaranteed to be read-safe. Instead, we can |
2800 | * approximate this using our carried decays, which are | 2800 | * approximate this using our carried decays, which are |
2801 | * explicitly atomically readable. | 2801 | * explicitly atomically readable. |
2802 | */ | 2802 | */ |
2803 | se->avg.last_runnable_update -= (-se->avg.decay_count) | 2803 | se->avg.last_runnable_update -= (-se->avg.decay_count) |
2804 | << 20; | 2804 | << 20; |
2805 | update_entity_load_avg(se, 0); | 2805 | update_entity_load_avg(se, 0); |
2806 | /* Indicate that we're now synchronized and on-rq */ | 2806 | /* Indicate that we're now synchronized and on-rq */ |
2807 | se->avg.decay_count = 0; | 2807 | se->avg.decay_count = 0; |
2808 | } | 2808 | } |
2809 | wakeup = 0; | 2809 | wakeup = 0; |
2810 | } else { | 2810 | } else { |
2811 | __synchronize_entity_decay(se); | 2811 | __synchronize_entity_decay(se); |
2812 | } | 2812 | } |
2813 | 2813 | ||
2814 | /* migrated tasks did not contribute to our blocked load */ | 2814 | /* migrated tasks did not contribute to our blocked load */ |
2815 | if (wakeup) { | 2815 | if (wakeup) { |
2816 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); | 2816 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); |
2817 | update_entity_load_avg(se, 0); | 2817 | update_entity_load_avg(se, 0); |
2818 | } | 2818 | } |
2819 | 2819 | ||
2820 | cfs_rq->runnable_load_avg += se->avg.load_avg_contrib; | 2820 | cfs_rq->runnable_load_avg += se->avg.load_avg_contrib; |
2821 | /* we force update consideration on load-balancer moves */ | 2821 | /* we force update consideration on load-balancer moves */ |
2822 | update_cfs_rq_blocked_load(cfs_rq, !wakeup); | 2822 | update_cfs_rq_blocked_load(cfs_rq, !wakeup); |
2823 | } | 2823 | } |
2824 | 2824 | ||
2825 | /* | 2825 | /* |
2826 | * Remove se's load from this cfs_rq child load-average, if the entity is | 2826 | * Remove se's load from this cfs_rq child load-average, if the entity is |
2827 | * transitioning to a blocked state we track its projected decay using | 2827 | * transitioning to a blocked state we track its projected decay using |
2828 | * blocked_load_avg. | 2828 | * blocked_load_avg. |
2829 | */ | 2829 | */ |
2830 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, | 2830 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
2831 | struct sched_entity *se, | 2831 | struct sched_entity *se, |
2832 | int sleep) | 2832 | int sleep) |
2833 | { | 2833 | { |
2834 | update_entity_load_avg(se, 1); | 2834 | update_entity_load_avg(se, 1); |
2835 | /* we force update consideration on load-balancer moves */ | 2835 | /* we force update consideration on load-balancer moves */ |
2836 | update_cfs_rq_blocked_load(cfs_rq, !sleep); | 2836 | update_cfs_rq_blocked_load(cfs_rq, !sleep); |
2837 | 2837 | ||
2838 | cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib; | 2838 | cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib; |
2839 | if (sleep) { | 2839 | if (sleep) { |
2840 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; | 2840 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; |
2841 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | 2841 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); |
2842 | } /* migrations, e.g. sleep=0 leave decay_count == 0 */ | 2842 | } /* migrations, e.g. sleep=0 leave decay_count == 0 */ |
2843 | } | 2843 | } |
2844 | 2844 | ||
2845 | /* | 2845 | /* |
2846 | * Update the rq's load with the elapsed running time before entering | 2846 | * Update the rq's load with the elapsed running time before entering |
2847 | * idle. if the last scheduled task is not a CFS task, idle_enter will | 2847 | * idle. if the last scheduled task is not a CFS task, idle_enter will |
2848 | * be the only way to update the runnable statistic. | 2848 | * be the only way to update the runnable statistic. |
2849 | */ | 2849 | */ |
2850 | void idle_enter_fair(struct rq *this_rq) | 2850 | void idle_enter_fair(struct rq *this_rq) |
2851 | { | 2851 | { |
2852 | update_rq_runnable_avg(this_rq, 1); | 2852 | update_rq_runnable_avg(this_rq, 1); |
2853 | } | 2853 | } |
2854 | 2854 | ||
2855 | /* | 2855 | /* |
2856 | * Update the rq's load with the elapsed idle time before a task is | 2856 | * Update the rq's load with the elapsed idle time before a task is |
2857 | * scheduled. if the newly scheduled task is not a CFS task, idle_exit will | 2857 | * scheduled. if the newly scheduled task is not a CFS task, idle_exit will |
2858 | * be the only way to update the runnable statistic. | 2858 | * be the only way to update the runnable statistic. |
2859 | */ | 2859 | */ |
2860 | void idle_exit_fair(struct rq *this_rq) | 2860 | void idle_exit_fair(struct rq *this_rq) |
2861 | { | 2861 | { |
2862 | update_rq_runnable_avg(this_rq, 0); | 2862 | update_rq_runnable_avg(this_rq, 0); |
2863 | } | 2863 | } |
2864 | 2864 | ||
2865 | static int idle_balance(struct rq *this_rq); | 2865 | static int idle_balance(struct rq *this_rq); |
2866 | 2866 | ||
2867 | #else /* CONFIG_SMP */ | 2867 | #else /* CONFIG_SMP */ |
2868 | 2868 | ||
2869 | static inline void update_entity_load_avg(struct sched_entity *se, | 2869 | static inline void update_entity_load_avg(struct sched_entity *se, |
2870 | int update_cfs_rq) {} | 2870 | int update_cfs_rq) {} |
2871 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} | 2871 | static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} |
2872 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, | 2872 | static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, |
2873 | struct sched_entity *se, | 2873 | struct sched_entity *se, |
2874 | int wakeup) {} | 2874 | int wakeup) {} |
2875 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, | 2875 | static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, |
2876 | struct sched_entity *se, | 2876 | struct sched_entity *se, |
2877 | int sleep) {} | 2877 | int sleep) {} |
2878 | static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, | 2878 | static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, |
2879 | int force_update) {} | 2879 | int force_update) {} |
2880 | 2880 | ||
2881 | static inline int idle_balance(struct rq *rq) | 2881 | static inline int idle_balance(struct rq *rq) |
2882 | { | 2882 | { |
2883 | return 0; | 2883 | return 0; |
2884 | } | 2884 | } |
2885 | 2885 | ||
2886 | #endif /* CONFIG_SMP */ | 2886 | #endif /* CONFIG_SMP */ |
2887 | 2887 | ||
2888 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) | 2888 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2889 | { | 2889 | { |
2890 | #ifdef CONFIG_SCHEDSTATS | 2890 | #ifdef CONFIG_SCHEDSTATS |
2891 | struct task_struct *tsk = NULL; | 2891 | struct task_struct *tsk = NULL; |
2892 | 2892 | ||
2893 | if (entity_is_task(se)) | 2893 | if (entity_is_task(se)) |
2894 | tsk = task_of(se); | 2894 | tsk = task_of(se); |
2895 | 2895 | ||
2896 | if (se->statistics.sleep_start) { | 2896 | if (se->statistics.sleep_start) { |
2897 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start; | 2897 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start; |
2898 | 2898 | ||
2899 | if ((s64)delta < 0) | 2899 | if ((s64)delta < 0) |
2900 | delta = 0; | 2900 | delta = 0; |
2901 | 2901 | ||
2902 | if (unlikely(delta > se->statistics.sleep_max)) | 2902 | if (unlikely(delta > se->statistics.sleep_max)) |
2903 | se->statistics.sleep_max = delta; | 2903 | se->statistics.sleep_max = delta; |
2904 | 2904 | ||
2905 | se->statistics.sleep_start = 0; | 2905 | se->statistics.sleep_start = 0; |
2906 | se->statistics.sum_sleep_runtime += delta; | 2906 | se->statistics.sum_sleep_runtime += delta; |
2907 | 2907 | ||
2908 | if (tsk) { | 2908 | if (tsk) { |
2909 | account_scheduler_latency(tsk, delta >> 10, 1); | 2909 | account_scheduler_latency(tsk, delta >> 10, 1); |
2910 | trace_sched_stat_sleep(tsk, delta); | 2910 | trace_sched_stat_sleep(tsk, delta); |
2911 | } | 2911 | } |
2912 | } | 2912 | } |
2913 | if (se->statistics.block_start) { | 2913 | if (se->statistics.block_start) { |
2914 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start; | 2914 | u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start; |
2915 | 2915 | ||
2916 | if ((s64)delta < 0) | 2916 | if ((s64)delta < 0) |
2917 | delta = 0; | 2917 | delta = 0; |
2918 | 2918 | ||
2919 | if (unlikely(delta > se->statistics.block_max)) | 2919 | if (unlikely(delta > se->statistics.block_max)) |
2920 | se->statistics.block_max = delta; | 2920 | se->statistics.block_max = delta; |
2921 | 2921 | ||
2922 | se->statistics.block_start = 0; | 2922 | se->statistics.block_start = 0; |
2923 | se->statistics.sum_sleep_runtime += delta; | 2923 | se->statistics.sum_sleep_runtime += delta; |
2924 | 2924 | ||
2925 | if (tsk) { | 2925 | if (tsk) { |
2926 | if (tsk->in_iowait) { | 2926 | if (tsk->in_iowait) { |
2927 | se->statistics.iowait_sum += delta; | 2927 | se->statistics.iowait_sum += delta; |
2928 | se->statistics.iowait_count++; | 2928 | se->statistics.iowait_count++; |
2929 | trace_sched_stat_iowait(tsk, delta); | 2929 | trace_sched_stat_iowait(tsk, delta); |
2930 | } | 2930 | } |
2931 | 2931 | ||
2932 | trace_sched_stat_blocked(tsk, delta); | 2932 | trace_sched_stat_blocked(tsk, delta); |
2933 | 2933 | ||
2934 | /* | 2934 | /* |
2935 | * Blocking time is in units of nanosecs, so shift by | 2935 | * Blocking time is in units of nanosecs, so shift by |
2936 | * 20 to get a milliseconds-range estimation of the | 2936 | * 20 to get a milliseconds-range estimation of the |
2937 | * amount of time that the task spent sleeping: | 2937 | * amount of time that the task spent sleeping: |
2938 | */ | 2938 | */ |
2939 | if (unlikely(prof_on == SLEEP_PROFILING)) { | 2939 | if (unlikely(prof_on == SLEEP_PROFILING)) { |
2940 | profile_hits(SLEEP_PROFILING, | 2940 | profile_hits(SLEEP_PROFILING, |
2941 | (void *)get_wchan(tsk), | 2941 | (void *)get_wchan(tsk), |
2942 | delta >> 20); | 2942 | delta >> 20); |
2943 | } | 2943 | } |
2944 | account_scheduler_latency(tsk, delta >> 10, 0); | 2944 | account_scheduler_latency(tsk, delta >> 10, 0); |
2945 | } | 2945 | } |
2946 | } | 2946 | } |
2947 | #endif | 2947 | #endif |
2948 | } | 2948 | } |
2949 | 2949 | ||
2950 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) | 2950 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
2951 | { | 2951 | { |
2952 | #ifdef CONFIG_SCHED_DEBUG | 2952 | #ifdef CONFIG_SCHED_DEBUG |
2953 | s64 d = se->vruntime - cfs_rq->min_vruntime; | 2953 | s64 d = se->vruntime - cfs_rq->min_vruntime; |
2954 | 2954 | ||
2955 | if (d < 0) | 2955 | if (d < 0) |
2956 | d = -d; | 2956 | d = -d; |
2957 | 2957 | ||
2958 | if (d > 3*sysctl_sched_latency) | 2958 | if (d > 3*sysctl_sched_latency) |
2959 | schedstat_inc(cfs_rq, nr_spread_over); | 2959 | schedstat_inc(cfs_rq, nr_spread_over); |
2960 | #endif | 2960 | #endif |
2961 | } | 2961 | } |
2962 | 2962 | ||
2963 | static void | 2963 | static void |
2964 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | 2964 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) |
2965 | { | 2965 | { |
2966 | u64 vruntime = cfs_rq->min_vruntime; | 2966 | u64 vruntime = cfs_rq->min_vruntime; |
2967 | 2967 | ||
2968 | /* | 2968 | /* |
2969 | * The 'current' period is already promised to the current tasks, | 2969 | * The 'current' period is already promised to the current tasks, |
2970 | * however the extra weight of the new task will slow them down a | 2970 | * however the extra weight of the new task will slow them down a |
2971 | * little, place the new task so that it fits in the slot that | 2971 | * little, place the new task so that it fits in the slot that |
2972 | * stays open at the end. | 2972 | * stays open at the end. |
2973 | */ | 2973 | */ |
2974 | if (initial && sched_feat(START_DEBIT)) | 2974 | if (initial && sched_feat(START_DEBIT)) |
2975 | vruntime += sched_vslice(cfs_rq, se); | 2975 | vruntime += sched_vslice(cfs_rq, se); |
2976 | 2976 | ||
2977 | /* sleeps up to a single latency don't count. */ | 2977 | /* sleeps up to a single latency don't count. */ |
2978 | if (!initial) { | 2978 | if (!initial) { |
2979 | unsigned long thresh = sysctl_sched_latency; | 2979 | unsigned long thresh = sysctl_sched_latency; |
2980 | 2980 | ||
2981 | /* | 2981 | /* |
2982 | * Halve their sleep time's effect, to allow | 2982 | * Halve their sleep time's effect, to allow |
2983 | * for a gentler effect of sleepers: | 2983 | * for a gentler effect of sleepers: |
2984 | */ | 2984 | */ |
2985 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | 2985 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) |
2986 | thresh >>= 1; | 2986 | thresh >>= 1; |
2987 | 2987 | ||
2988 | vruntime -= thresh; | 2988 | vruntime -= thresh; |
2989 | } | 2989 | } |
2990 | 2990 | ||
2991 | /* ensure we never gain time by being placed backwards. */ | 2991 | /* ensure we never gain time by being placed backwards. */ |
2992 | se->vruntime = max_vruntime(se->vruntime, vruntime); | 2992 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
2993 | } | 2993 | } |
2994 | 2994 | ||
2995 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); | 2995 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
2996 | 2996 | ||
2997 | static void | 2997 | static void |
2998 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) | 2998 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
2999 | { | 2999 | { |
3000 | /* | 3000 | /* |
3001 | * Update the normalized vruntime before updating min_vruntime | 3001 | * Update the normalized vruntime before updating min_vruntime |
3002 | * through calling update_curr(). | 3002 | * through calling update_curr(). |
3003 | */ | 3003 | */ |
3004 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) | 3004 | if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) |
3005 | se->vruntime += cfs_rq->min_vruntime; | 3005 | se->vruntime += cfs_rq->min_vruntime; |
3006 | 3006 | ||
3007 | /* | 3007 | /* |
3008 | * Update run-time statistics of the 'current'. | 3008 | * Update run-time statistics of the 'current'. |
3009 | */ | 3009 | */ |
3010 | update_curr(cfs_rq); | 3010 | update_curr(cfs_rq); |
3011 | enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP); | 3011 | enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP); |
3012 | account_entity_enqueue(cfs_rq, se); | 3012 | account_entity_enqueue(cfs_rq, se); |
3013 | update_cfs_shares(cfs_rq); | 3013 | update_cfs_shares(cfs_rq); |
3014 | 3014 | ||
3015 | if (flags & ENQUEUE_WAKEUP) { | 3015 | if (flags & ENQUEUE_WAKEUP) { |
3016 | place_entity(cfs_rq, se, 0); | 3016 | place_entity(cfs_rq, se, 0); |
3017 | enqueue_sleeper(cfs_rq, se); | 3017 | enqueue_sleeper(cfs_rq, se); |
3018 | } | 3018 | } |
3019 | 3019 | ||
3020 | update_stats_enqueue(cfs_rq, se); | 3020 | update_stats_enqueue(cfs_rq, se); |
3021 | check_spread(cfs_rq, se); | 3021 | check_spread(cfs_rq, se); |
3022 | if (se != cfs_rq->curr) | 3022 | if (se != cfs_rq->curr) |
3023 | __enqueue_entity(cfs_rq, se); | 3023 | __enqueue_entity(cfs_rq, se); |
3024 | se->on_rq = 1; | 3024 | se->on_rq = 1; |
3025 | 3025 | ||
3026 | if (cfs_rq->nr_running == 1) { | 3026 | if (cfs_rq->nr_running == 1) { |
3027 | list_add_leaf_cfs_rq(cfs_rq); | 3027 | list_add_leaf_cfs_rq(cfs_rq); |
3028 | check_enqueue_throttle(cfs_rq); | 3028 | check_enqueue_throttle(cfs_rq); |
3029 | } | 3029 | } |
3030 | } | 3030 | } |
3031 | 3031 | ||
3032 | static void __clear_buddies_last(struct sched_entity *se) | 3032 | static void __clear_buddies_last(struct sched_entity *se) |
3033 | { | 3033 | { |
3034 | for_each_sched_entity(se) { | 3034 | for_each_sched_entity(se) { |
3035 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | 3035 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
3036 | if (cfs_rq->last != se) | 3036 | if (cfs_rq->last != se) |
3037 | break; | 3037 | break; |
3038 | 3038 | ||
3039 | cfs_rq->last = NULL; | 3039 | cfs_rq->last = NULL; |
3040 | } | 3040 | } |
3041 | } | 3041 | } |
3042 | 3042 | ||
3043 | static void __clear_buddies_next(struct sched_entity *se) | 3043 | static void __clear_buddies_next(struct sched_entity *se) |
3044 | { | 3044 | { |
3045 | for_each_sched_entity(se) { | 3045 | for_each_sched_entity(se) { |
3046 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | 3046 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
3047 | if (cfs_rq->next != se) | 3047 | if (cfs_rq->next != se) |
3048 | break; | 3048 | break; |
3049 | 3049 | ||
3050 | cfs_rq->next = NULL; | 3050 | cfs_rq->next = NULL; |
3051 | } | 3051 | } |
3052 | } | 3052 | } |
3053 | 3053 | ||
3054 | static void __clear_buddies_skip(struct sched_entity *se) | 3054 | static void __clear_buddies_skip(struct sched_entity *se) |
3055 | { | 3055 | { |
3056 | for_each_sched_entity(se) { | 3056 | for_each_sched_entity(se) { |
3057 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | 3057 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
3058 | if (cfs_rq->skip != se) | 3058 | if (cfs_rq->skip != se) |
3059 | break; | 3059 | break; |
3060 | 3060 | ||
3061 | cfs_rq->skip = NULL; | 3061 | cfs_rq->skip = NULL; |
3062 | } | 3062 | } |
3063 | } | 3063 | } |
3064 | 3064 | ||
3065 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) | 3065 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3066 | { | 3066 | { |
3067 | if (cfs_rq->last == se) | 3067 | if (cfs_rq->last == se) |
3068 | __clear_buddies_last(se); | 3068 | __clear_buddies_last(se); |
3069 | 3069 | ||
3070 | if (cfs_rq->next == se) | 3070 | if (cfs_rq->next == se) |
3071 | __clear_buddies_next(se); | 3071 | __clear_buddies_next(se); |
3072 | 3072 | ||
3073 | if (cfs_rq->skip == se) | 3073 | if (cfs_rq->skip == se) |
3074 | __clear_buddies_skip(se); | 3074 | __clear_buddies_skip(se); |
3075 | } | 3075 | } |
3076 | 3076 | ||
3077 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); | 3077 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
3078 | 3078 | ||
3079 | static void | 3079 | static void |
3080 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) | 3080 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
3081 | { | 3081 | { |
3082 | /* | 3082 | /* |
3083 | * Update run-time statistics of the 'current'. | 3083 | * Update run-time statistics of the 'current'. |
3084 | */ | 3084 | */ |
3085 | update_curr(cfs_rq); | 3085 | update_curr(cfs_rq); |
3086 | dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP); | 3086 | dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP); |
3087 | 3087 | ||
3088 | update_stats_dequeue(cfs_rq, se); | 3088 | update_stats_dequeue(cfs_rq, se); |
3089 | if (flags & DEQUEUE_SLEEP) { | 3089 | if (flags & DEQUEUE_SLEEP) { |
3090 | #ifdef CONFIG_SCHEDSTATS | 3090 | #ifdef CONFIG_SCHEDSTATS |
3091 | if (entity_is_task(se)) { | 3091 | if (entity_is_task(se)) { |
3092 | struct task_struct *tsk = task_of(se); | 3092 | struct task_struct *tsk = task_of(se); |
3093 | 3093 | ||
3094 | if (tsk->state & TASK_INTERRUPTIBLE) | 3094 | if (tsk->state & TASK_INTERRUPTIBLE) |
3095 | se->statistics.sleep_start = rq_clock(rq_of(cfs_rq)); | 3095 | se->statistics.sleep_start = rq_clock(rq_of(cfs_rq)); |
3096 | if (tsk->state & TASK_UNINTERRUPTIBLE) | 3096 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
3097 | se->statistics.block_start = rq_clock(rq_of(cfs_rq)); | 3097 | se->statistics.block_start = rq_clock(rq_of(cfs_rq)); |
3098 | } | 3098 | } |
3099 | #endif | 3099 | #endif |
3100 | } | 3100 | } |
3101 | 3101 | ||
3102 | clear_buddies(cfs_rq, se); | 3102 | clear_buddies(cfs_rq, se); |
3103 | 3103 | ||
3104 | if (se != cfs_rq->curr) | 3104 | if (se != cfs_rq->curr) |
3105 | __dequeue_entity(cfs_rq, se); | 3105 | __dequeue_entity(cfs_rq, se); |
3106 | se->on_rq = 0; | 3106 | se->on_rq = 0; |
3107 | account_entity_dequeue(cfs_rq, se); | 3107 | account_entity_dequeue(cfs_rq, se); |
3108 | 3108 | ||
3109 | /* | 3109 | /* |
3110 | * Normalize the entity after updating the min_vruntime because the | 3110 | * Normalize the entity after updating the min_vruntime because the |
3111 | * update can refer to the ->curr item and we need to reflect this | 3111 | * update can refer to the ->curr item and we need to reflect this |
3112 | * movement in our normalized position. | 3112 | * movement in our normalized position. |
3113 | */ | 3113 | */ |
3114 | if (!(flags & DEQUEUE_SLEEP)) | 3114 | if (!(flags & DEQUEUE_SLEEP)) |
3115 | se->vruntime -= cfs_rq->min_vruntime; | 3115 | se->vruntime -= cfs_rq->min_vruntime; |
3116 | 3116 | ||
3117 | /* return excess runtime on last dequeue */ | 3117 | /* return excess runtime on last dequeue */ |
3118 | return_cfs_rq_runtime(cfs_rq); | 3118 | return_cfs_rq_runtime(cfs_rq); |
3119 | 3119 | ||
3120 | update_min_vruntime(cfs_rq); | 3120 | update_min_vruntime(cfs_rq); |
3121 | update_cfs_shares(cfs_rq); | 3121 | update_cfs_shares(cfs_rq); |
3122 | } | 3122 | } |
3123 | 3123 | ||
3124 | /* | 3124 | /* |
3125 | * Preempt the current task with a newly woken task if needed: | 3125 | * Preempt the current task with a newly woken task if needed: |
3126 | */ | 3126 | */ |
3127 | static void | 3127 | static void |
3128 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) | 3128 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
3129 | { | 3129 | { |
3130 | unsigned long ideal_runtime, delta_exec; | 3130 | unsigned long ideal_runtime, delta_exec; |
3131 | struct sched_entity *se; | 3131 | struct sched_entity *se; |
3132 | s64 delta; | 3132 | s64 delta; |
3133 | 3133 | ||
3134 | ideal_runtime = sched_slice(cfs_rq, curr); | 3134 | ideal_runtime = sched_slice(cfs_rq, curr); |
3135 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; | 3135 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
3136 | if (delta_exec > ideal_runtime) { | 3136 | if (delta_exec > ideal_runtime) { |
3137 | resched_curr(rq_of(cfs_rq)); | 3137 | resched_curr(rq_of(cfs_rq)); |
3138 | /* | 3138 | /* |
3139 | * The current task ran long enough, ensure it doesn't get | 3139 | * The current task ran long enough, ensure it doesn't get |
3140 | * re-elected due to buddy favours. | 3140 | * re-elected due to buddy favours. |
3141 | */ | 3141 | */ |
3142 | clear_buddies(cfs_rq, curr); | 3142 | clear_buddies(cfs_rq, curr); |
3143 | return; | 3143 | return; |
3144 | } | 3144 | } |
3145 | 3145 | ||
3146 | /* | 3146 | /* |
3147 | * Ensure that a task that missed wakeup preemption by a | 3147 | * Ensure that a task that missed wakeup preemption by a |
3148 | * narrow margin doesn't have to wait for a full slice. | 3148 | * narrow margin doesn't have to wait for a full slice. |
3149 | * This also mitigates buddy induced latencies under load. | 3149 | * This also mitigates buddy induced latencies under load. |
3150 | */ | 3150 | */ |
3151 | if (delta_exec < sysctl_sched_min_granularity) | 3151 | if (delta_exec < sysctl_sched_min_granularity) |
3152 | return; | 3152 | return; |
3153 | 3153 | ||
3154 | se = __pick_first_entity(cfs_rq); | 3154 | se = __pick_first_entity(cfs_rq); |
3155 | delta = curr->vruntime - se->vruntime; | 3155 | delta = curr->vruntime - se->vruntime; |
3156 | 3156 | ||
3157 | if (delta < 0) | 3157 | if (delta < 0) |
3158 | return; | 3158 | return; |
3159 | 3159 | ||
3160 | if (delta > ideal_runtime) | 3160 | if (delta > ideal_runtime) |
3161 | resched_curr(rq_of(cfs_rq)); | 3161 | resched_curr(rq_of(cfs_rq)); |
3162 | } | 3162 | } |
3163 | 3163 | ||
3164 | static void | 3164 | static void |
3165 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) | 3165 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3166 | { | 3166 | { |
3167 | /* 'current' is not kept within the tree. */ | 3167 | /* 'current' is not kept within the tree. */ |
3168 | if (se->on_rq) { | 3168 | if (se->on_rq) { |
3169 | /* | 3169 | /* |
3170 | * Any task has to be enqueued before it get to execute on | 3170 | * Any task has to be enqueued before it get to execute on |
3171 | * a CPU. So account for the time it spent waiting on the | 3171 | * a CPU. So account for the time it spent waiting on the |
3172 | * runqueue. | 3172 | * runqueue. |
3173 | */ | 3173 | */ |
3174 | update_stats_wait_end(cfs_rq, se); | 3174 | update_stats_wait_end(cfs_rq, se); |
3175 | __dequeue_entity(cfs_rq, se); | 3175 | __dequeue_entity(cfs_rq, se); |
3176 | } | 3176 | } |
3177 | 3177 | ||
3178 | update_stats_curr_start(cfs_rq, se); | 3178 | update_stats_curr_start(cfs_rq, se); |
3179 | cfs_rq->curr = se; | 3179 | cfs_rq->curr = se; |
3180 | #ifdef CONFIG_SCHEDSTATS | 3180 | #ifdef CONFIG_SCHEDSTATS |
3181 | /* | 3181 | /* |
3182 | * Track our maximum slice length, if the CPU's load is at | 3182 | * Track our maximum slice length, if the CPU's load is at |
3183 | * least twice that of our own weight (i.e. dont track it | 3183 | * least twice that of our own weight (i.e. dont track it |
3184 | * when there are only lesser-weight tasks around): | 3184 | * when there are only lesser-weight tasks around): |
3185 | */ | 3185 | */ |
3186 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { | 3186 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
3187 | se->statistics.slice_max = max(se->statistics.slice_max, | 3187 | se->statistics.slice_max = max(se->statistics.slice_max, |
3188 | se->sum_exec_runtime - se->prev_sum_exec_runtime); | 3188 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
3189 | } | 3189 | } |
3190 | #endif | 3190 | #endif |
3191 | se->prev_sum_exec_runtime = se->sum_exec_runtime; | 3191 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
3192 | } | 3192 | } |
3193 | 3193 | ||
3194 | static int | 3194 | static int |
3195 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | 3195 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); |
3196 | 3196 | ||
3197 | /* | 3197 | /* |
3198 | * Pick the next process, keeping these things in mind, in this order: | 3198 | * Pick the next process, keeping these things in mind, in this order: |
3199 | * 1) keep things fair between processes/task groups | 3199 | * 1) keep things fair between processes/task groups |
3200 | * 2) pick the "next" process, since someone really wants that to run | 3200 | * 2) pick the "next" process, since someone really wants that to run |
3201 | * 3) pick the "last" process, for cache locality | 3201 | * 3) pick the "last" process, for cache locality |
3202 | * 4) do not run the "skip" process, if something else is available | 3202 | * 4) do not run the "skip" process, if something else is available |
3203 | */ | 3203 | */ |
3204 | static struct sched_entity * | 3204 | static struct sched_entity * |
3205 | pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) | 3205 | pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
3206 | { | 3206 | { |
3207 | struct sched_entity *left = __pick_first_entity(cfs_rq); | 3207 | struct sched_entity *left = __pick_first_entity(cfs_rq); |
3208 | struct sched_entity *se; | 3208 | struct sched_entity *se; |
3209 | 3209 | ||
3210 | /* | 3210 | /* |
3211 | * If curr is set we have to see if its left of the leftmost entity | 3211 | * If curr is set we have to see if its left of the leftmost entity |
3212 | * still in the tree, provided there was anything in the tree at all. | 3212 | * still in the tree, provided there was anything in the tree at all. |
3213 | */ | 3213 | */ |
3214 | if (!left || (curr && entity_before(curr, left))) | 3214 | if (!left || (curr && entity_before(curr, left))) |
3215 | left = curr; | 3215 | left = curr; |
3216 | 3216 | ||
3217 | se = left; /* ideally we run the leftmost entity */ | 3217 | se = left; /* ideally we run the leftmost entity */ |
3218 | 3218 | ||
3219 | /* | 3219 | /* |
3220 | * Avoid running the skip buddy, if running something else can | 3220 | * Avoid running the skip buddy, if running something else can |
3221 | * be done without getting too unfair. | 3221 | * be done without getting too unfair. |
3222 | */ | 3222 | */ |
3223 | if (cfs_rq->skip == se) { | 3223 | if (cfs_rq->skip == se) { |
3224 | struct sched_entity *second; | 3224 | struct sched_entity *second; |
3225 | 3225 | ||
3226 | if (se == curr) { | 3226 | if (se == curr) { |
3227 | second = __pick_first_entity(cfs_rq); | 3227 | second = __pick_first_entity(cfs_rq); |
3228 | } else { | 3228 | } else { |
3229 | second = __pick_next_entity(se); | 3229 | second = __pick_next_entity(se); |
3230 | if (!second || (curr && entity_before(curr, second))) | 3230 | if (!second || (curr && entity_before(curr, second))) |
3231 | second = curr; | 3231 | second = curr; |
3232 | } | 3232 | } |
3233 | 3233 | ||
3234 | if (second && wakeup_preempt_entity(second, left) < 1) | 3234 | if (second && wakeup_preempt_entity(second, left) < 1) |
3235 | se = second; | 3235 | se = second; |
3236 | } | 3236 | } |
3237 | 3237 | ||
3238 | /* | 3238 | /* |
3239 | * Prefer last buddy, try to return the CPU to a preempted task. | 3239 | * Prefer last buddy, try to return the CPU to a preempted task. |
3240 | */ | 3240 | */ |
3241 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | 3241 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) |
3242 | se = cfs_rq->last; | 3242 | se = cfs_rq->last; |
3243 | 3243 | ||
3244 | /* | 3244 | /* |
3245 | * Someone really wants this to run. If it's not unfair, run it. | 3245 | * Someone really wants this to run. If it's not unfair, run it. |
3246 | */ | 3246 | */ |
3247 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | 3247 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) |
3248 | se = cfs_rq->next; | 3248 | se = cfs_rq->next; |
3249 | 3249 | ||
3250 | clear_buddies(cfs_rq, se); | 3250 | clear_buddies(cfs_rq, se); |
3251 | 3251 | ||
3252 | return se; | 3252 | return se; |
3253 | } | 3253 | } |
3254 | 3254 | ||
3255 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); | 3255 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
3256 | 3256 | ||
3257 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) | 3257 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
3258 | { | 3258 | { |
3259 | /* | 3259 | /* |
3260 | * If still on the runqueue then deactivate_task() | 3260 | * If still on the runqueue then deactivate_task() |
3261 | * was not called and update_curr() has to be done: | 3261 | * was not called and update_curr() has to be done: |
3262 | */ | 3262 | */ |
3263 | if (prev->on_rq) | 3263 | if (prev->on_rq) |
3264 | update_curr(cfs_rq); | 3264 | update_curr(cfs_rq); |
3265 | 3265 | ||
3266 | /* throttle cfs_rqs exceeding runtime */ | 3266 | /* throttle cfs_rqs exceeding runtime */ |
3267 | check_cfs_rq_runtime(cfs_rq); | 3267 | check_cfs_rq_runtime(cfs_rq); |
3268 | 3268 | ||
3269 | check_spread(cfs_rq, prev); | 3269 | check_spread(cfs_rq, prev); |
3270 | if (prev->on_rq) { | 3270 | if (prev->on_rq) { |
3271 | update_stats_wait_start(cfs_rq, prev); | 3271 | update_stats_wait_start(cfs_rq, prev); |
3272 | /* Put 'current' back into the tree. */ | 3272 | /* Put 'current' back into the tree. */ |
3273 | __enqueue_entity(cfs_rq, prev); | 3273 | __enqueue_entity(cfs_rq, prev); |
3274 | /* in !on_rq case, update occurred at dequeue */ | 3274 | /* in !on_rq case, update occurred at dequeue */ |
3275 | update_entity_load_avg(prev, 1); | 3275 | update_entity_load_avg(prev, 1); |
3276 | } | 3276 | } |
3277 | cfs_rq->curr = NULL; | 3277 | cfs_rq->curr = NULL; |
3278 | } | 3278 | } |
3279 | 3279 | ||
3280 | static void | 3280 | static void |
3281 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | 3281 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) |
3282 | { | 3282 | { |
3283 | /* | 3283 | /* |
3284 | * Update run-time statistics of the 'current'. | 3284 | * Update run-time statistics of the 'current'. |
3285 | */ | 3285 | */ |
3286 | update_curr(cfs_rq); | 3286 | update_curr(cfs_rq); |
3287 | 3287 | ||
3288 | /* | 3288 | /* |
3289 | * Ensure that runnable average is periodically updated. | 3289 | * Ensure that runnable average is periodically updated. |
3290 | */ | 3290 | */ |
3291 | update_entity_load_avg(curr, 1); | 3291 | update_entity_load_avg(curr, 1); |
3292 | update_cfs_rq_blocked_load(cfs_rq, 1); | 3292 | update_cfs_rq_blocked_load(cfs_rq, 1); |
3293 | update_cfs_shares(cfs_rq); | 3293 | update_cfs_shares(cfs_rq); |
3294 | 3294 | ||
3295 | #ifdef CONFIG_SCHED_HRTICK | 3295 | #ifdef CONFIG_SCHED_HRTICK |
3296 | /* | 3296 | /* |
3297 | * queued ticks are scheduled to match the slice, so don't bother | 3297 | * queued ticks are scheduled to match the slice, so don't bother |
3298 | * validating it and just reschedule. | 3298 | * validating it and just reschedule. |
3299 | */ | 3299 | */ |
3300 | if (queued) { | 3300 | if (queued) { |
3301 | resched_curr(rq_of(cfs_rq)); | 3301 | resched_curr(rq_of(cfs_rq)); |
3302 | return; | 3302 | return; |
3303 | } | 3303 | } |
3304 | /* | 3304 | /* |
3305 | * don't let the period tick interfere with the hrtick preemption | 3305 | * don't let the period tick interfere with the hrtick preemption |
3306 | */ | 3306 | */ |
3307 | if (!sched_feat(DOUBLE_TICK) && | 3307 | if (!sched_feat(DOUBLE_TICK) && |
3308 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | 3308 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) |
3309 | return; | 3309 | return; |
3310 | #endif | 3310 | #endif |
3311 | 3311 | ||
3312 | if (cfs_rq->nr_running > 1) | 3312 | if (cfs_rq->nr_running > 1) |
3313 | check_preempt_tick(cfs_rq, curr); | 3313 | check_preempt_tick(cfs_rq, curr); |
3314 | } | 3314 | } |
3315 | 3315 | ||
3316 | 3316 | ||
3317 | /************************************************** | 3317 | /************************************************** |
3318 | * CFS bandwidth control machinery | 3318 | * CFS bandwidth control machinery |
3319 | */ | 3319 | */ |
3320 | 3320 | ||
3321 | #ifdef CONFIG_CFS_BANDWIDTH | 3321 | #ifdef CONFIG_CFS_BANDWIDTH |
3322 | 3322 | ||
3323 | #ifdef HAVE_JUMP_LABEL | 3323 | #ifdef HAVE_JUMP_LABEL |
3324 | static struct static_key __cfs_bandwidth_used; | 3324 | static struct static_key __cfs_bandwidth_used; |
3325 | 3325 | ||
3326 | static inline bool cfs_bandwidth_used(void) | 3326 | static inline bool cfs_bandwidth_used(void) |
3327 | { | 3327 | { |
3328 | return static_key_false(&__cfs_bandwidth_used); | 3328 | return static_key_false(&__cfs_bandwidth_used); |
3329 | } | 3329 | } |
3330 | 3330 | ||
3331 | void cfs_bandwidth_usage_inc(void) | 3331 | void cfs_bandwidth_usage_inc(void) |
3332 | { | 3332 | { |
3333 | static_key_slow_inc(&__cfs_bandwidth_used); | 3333 | static_key_slow_inc(&__cfs_bandwidth_used); |
3334 | } | 3334 | } |
3335 | 3335 | ||
3336 | void cfs_bandwidth_usage_dec(void) | 3336 | void cfs_bandwidth_usage_dec(void) |
3337 | { | 3337 | { |
3338 | static_key_slow_dec(&__cfs_bandwidth_used); | 3338 | static_key_slow_dec(&__cfs_bandwidth_used); |
3339 | } | 3339 | } |
3340 | #else /* HAVE_JUMP_LABEL */ | 3340 | #else /* HAVE_JUMP_LABEL */ |
3341 | static bool cfs_bandwidth_used(void) | 3341 | static bool cfs_bandwidth_used(void) |
3342 | { | 3342 | { |
3343 | return true; | 3343 | return true; |
3344 | } | 3344 | } |
3345 | 3345 | ||
3346 | void cfs_bandwidth_usage_inc(void) {} | 3346 | void cfs_bandwidth_usage_inc(void) {} |
3347 | void cfs_bandwidth_usage_dec(void) {} | 3347 | void cfs_bandwidth_usage_dec(void) {} |
3348 | #endif /* HAVE_JUMP_LABEL */ | 3348 | #endif /* HAVE_JUMP_LABEL */ |
3349 | 3349 | ||
3350 | /* | 3350 | /* |
3351 | * default period for cfs group bandwidth. | 3351 | * default period for cfs group bandwidth. |
3352 | * default: 0.1s, units: nanoseconds | 3352 | * default: 0.1s, units: nanoseconds |
3353 | */ | 3353 | */ |
3354 | static inline u64 default_cfs_period(void) | 3354 | static inline u64 default_cfs_period(void) |
3355 | { | 3355 | { |
3356 | return 100000000ULL; | 3356 | return 100000000ULL; |
3357 | } | 3357 | } |
3358 | 3358 | ||
3359 | static inline u64 sched_cfs_bandwidth_slice(void) | 3359 | static inline u64 sched_cfs_bandwidth_slice(void) |
3360 | { | 3360 | { |
3361 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | 3361 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; |
3362 | } | 3362 | } |
3363 | 3363 | ||
3364 | /* | 3364 | /* |
3365 | * Replenish runtime according to assigned quota and update expiration time. | 3365 | * Replenish runtime according to assigned quota and update expiration time. |
3366 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | 3366 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding |
3367 | * additional synchronization around rq->lock. | 3367 | * additional synchronization around rq->lock. |
3368 | * | 3368 | * |
3369 | * requires cfs_b->lock | 3369 | * requires cfs_b->lock |
3370 | */ | 3370 | */ |
3371 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) | 3371 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
3372 | { | 3372 | { |
3373 | u64 now; | 3373 | u64 now; |
3374 | 3374 | ||
3375 | if (cfs_b->quota == RUNTIME_INF) | 3375 | if (cfs_b->quota == RUNTIME_INF) |
3376 | return; | 3376 | return; |
3377 | 3377 | ||
3378 | now = sched_clock_cpu(smp_processor_id()); | 3378 | now = sched_clock_cpu(smp_processor_id()); |
3379 | cfs_b->runtime = cfs_b->quota; | 3379 | cfs_b->runtime = cfs_b->quota; |
3380 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | 3380 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); |
3381 | } | 3381 | } |
3382 | 3382 | ||
3383 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) | 3383 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
3384 | { | 3384 | { |
3385 | return &tg->cfs_bandwidth; | 3385 | return &tg->cfs_bandwidth; |
3386 | } | 3386 | } |
3387 | 3387 | ||
3388 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ | 3388 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ |
3389 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | 3389 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
3390 | { | 3390 | { |
3391 | if (unlikely(cfs_rq->throttle_count)) | 3391 | if (unlikely(cfs_rq->throttle_count)) |
3392 | return cfs_rq->throttled_clock_task; | 3392 | return cfs_rq->throttled_clock_task; |
3393 | 3393 | ||
3394 | return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; | 3394 | return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; |
3395 | } | 3395 | } |
3396 | 3396 | ||
3397 | /* returns 0 on failure to allocate runtime */ | 3397 | /* returns 0 on failure to allocate runtime */ |
3398 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | 3398 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
3399 | { | 3399 | { |
3400 | struct task_group *tg = cfs_rq->tg; | 3400 | struct task_group *tg = cfs_rq->tg; |
3401 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | 3401 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); |
3402 | u64 amount = 0, min_amount, expires; | 3402 | u64 amount = 0, min_amount, expires; |
3403 | 3403 | ||
3404 | /* note: this is a positive sum as runtime_remaining <= 0 */ | 3404 | /* note: this is a positive sum as runtime_remaining <= 0 */ |
3405 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | 3405 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; |
3406 | 3406 | ||
3407 | raw_spin_lock(&cfs_b->lock); | 3407 | raw_spin_lock(&cfs_b->lock); |
3408 | if (cfs_b->quota == RUNTIME_INF) | 3408 | if (cfs_b->quota == RUNTIME_INF) |
3409 | amount = min_amount; | 3409 | amount = min_amount; |
3410 | else { | 3410 | else { |
3411 | /* | 3411 | /* |
3412 | * If the bandwidth pool has become inactive, then at least one | 3412 | * If the bandwidth pool has become inactive, then at least one |
3413 | * period must have elapsed since the last consumption. | 3413 | * period must have elapsed since the last consumption. |
3414 | * Refresh the global state and ensure bandwidth timer becomes | 3414 | * Refresh the global state and ensure bandwidth timer becomes |
3415 | * active. | 3415 | * active. |
3416 | */ | 3416 | */ |
3417 | if (!cfs_b->timer_active) { | 3417 | if (!cfs_b->timer_active) { |
3418 | __refill_cfs_bandwidth_runtime(cfs_b); | 3418 | __refill_cfs_bandwidth_runtime(cfs_b); |
3419 | __start_cfs_bandwidth(cfs_b, false); | 3419 | __start_cfs_bandwidth(cfs_b, false); |
3420 | } | 3420 | } |
3421 | 3421 | ||
3422 | if (cfs_b->runtime > 0) { | 3422 | if (cfs_b->runtime > 0) { |
3423 | amount = min(cfs_b->runtime, min_amount); | 3423 | amount = min(cfs_b->runtime, min_amount); |
3424 | cfs_b->runtime -= amount; | 3424 | cfs_b->runtime -= amount; |
3425 | cfs_b->idle = 0; | 3425 | cfs_b->idle = 0; |
3426 | } | 3426 | } |
3427 | } | 3427 | } |
3428 | expires = cfs_b->runtime_expires; | 3428 | expires = cfs_b->runtime_expires; |
3429 | raw_spin_unlock(&cfs_b->lock); | 3429 | raw_spin_unlock(&cfs_b->lock); |
3430 | 3430 | ||
3431 | cfs_rq->runtime_remaining += amount; | 3431 | cfs_rq->runtime_remaining += amount; |
3432 | /* | 3432 | /* |
3433 | * we may have advanced our local expiration to account for allowed | 3433 | * we may have advanced our local expiration to account for allowed |
3434 | * spread between our sched_clock and the one on which runtime was | 3434 | * spread between our sched_clock and the one on which runtime was |
3435 | * issued. | 3435 | * issued. |
3436 | */ | 3436 | */ |
3437 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | 3437 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) |
3438 | cfs_rq->runtime_expires = expires; | 3438 | cfs_rq->runtime_expires = expires; |
3439 | 3439 | ||
3440 | return cfs_rq->runtime_remaining > 0; | 3440 | return cfs_rq->runtime_remaining > 0; |
3441 | } | 3441 | } |
3442 | 3442 | ||
3443 | /* | 3443 | /* |
3444 | * Note: This depends on the synchronization provided by sched_clock and the | 3444 | * Note: This depends on the synchronization provided by sched_clock and the |
3445 | * fact that rq->clock snapshots this value. | 3445 | * fact that rq->clock snapshots this value. |
3446 | */ | 3446 | */ |
3447 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | 3447 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
3448 | { | 3448 | { |
3449 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | 3449 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
3450 | 3450 | ||
3451 | /* if the deadline is ahead of our clock, nothing to do */ | 3451 | /* if the deadline is ahead of our clock, nothing to do */ |
3452 | if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) | 3452 | if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) |
3453 | return; | 3453 | return; |
3454 | 3454 | ||
3455 | if (cfs_rq->runtime_remaining < 0) | 3455 | if (cfs_rq->runtime_remaining < 0) |
3456 | return; | 3456 | return; |
3457 | 3457 | ||
3458 | /* | 3458 | /* |
3459 | * If the local deadline has passed we have to consider the | 3459 | * If the local deadline has passed we have to consider the |
3460 | * possibility that our sched_clock is 'fast' and the global deadline | 3460 | * possibility that our sched_clock is 'fast' and the global deadline |
3461 | * has not truly expired. | 3461 | * has not truly expired. |
3462 | * | 3462 | * |
3463 | * Fortunately we can check determine whether this the case by checking | 3463 | * Fortunately we can check determine whether this the case by checking |
3464 | * whether the global deadline has advanced. It is valid to compare | 3464 | * whether the global deadline has advanced. It is valid to compare |
3465 | * cfs_b->runtime_expires without any locks since we only care about | 3465 | * cfs_b->runtime_expires without any locks since we only care about |
3466 | * exact equality, so a partial write will still work. | 3466 | * exact equality, so a partial write will still work. |
3467 | */ | 3467 | */ |
3468 | 3468 | ||
3469 | if (cfs_rq->runtime_expires != cfs_b->runtime_expires) { | 3469 | if (cfs_rq->runtime_expires != cfs_b->runtime_expires) { |
3470 | /* extend local deadline, drift is bounded above by 2 ticks */ | 3470 | /* extend local deadline, drift is bounded above by 2 ticks */ |
3471 | cfs_rq->runtime_expires += TICK_NSEC; | 3471 | cfs_rq->runtime_expires += TICK_NSEC; |
3472 | } else { | 3472 | } else { |
3473 | /* global deadline is ahead, expiration has passed */ | 3473 | /* global deadline is ahead, expiration has passed */ |
3474 | cfs_rq->runtime_remaining = 0; | 3474 | cfs_rq->runtime_remaining = 0; |
3475 | } | 3475 | } |
3476 | } | 3476 | } |
3477 | 3477 | ||
3478 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) | 3478 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
3479 | { | 3479 | { |
3480 | /* dock delta_exec before expiring quota (as it could span periods) */ | 3480 | /* dock delta_exec before expiring quota (as it could span periods) */ |
3481 | cfs_rq->runtime_remaining -= delta_exec; | 3481 | cfs_rq->runtime_remaining -= delta_exec; |
3482 | expire_cfs_rq_runtime(cfs_rq); | 3482 | expire_cfs_rq_runtime(cfs_rq); |
3483 | 3483 | ||
3484 | if (likely(cfs_rq->runtime_remaining > 0)) | 3484 | if (likely(cfs_rq->runtime_remaining > 0)) |
3485 | return; | 3485 | return; |
3486 | 3486 | ||
3487 | /* | 3487 | /* |
3488 | * if we're unable to extend our runtime we resched so that the active | 3488 | * if we're unable to extend our runtime we resched so that the active |
3489 | * hierarchy can be throttled | 3489 | * hierarchy can be throttled |
3490 | */ | 3490 | */ |
3491 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | 3491 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) |
3492 | resched_curr(rq_of(cfs_rq)); | 3492 | resched_curr(rq_of(cfs_rq)); |
3493 | } | 3493 | } |
3494 | 3494 | ||
3495 | static __always_inline | 3495 | static __always_inline |
3496 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) | 3496 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
3497 | { | 3497 | { |
3498 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) | 3498 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
3499 | return; | 3499 | return; |
3500 | 3500 | ||
3501 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | 3501 | __account_cfs_rq_runtime(cfs_rq, delta_exec); |
3502 | } | 3502 | } |
3503 | 3503 | ||
3504 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | 3504 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
3505 | { | 3505 | { |
3506 | return cfs_bandwidth_used() && cfs_rq->throttled; | 3506 | return cfs_bandwidth_used() && cfs_rq->throttled; |
3507 | } | 3507 | } |
3508 | 3508 | ||
3509 | /* check whether cfs_rq, or any parent, is throttled */ | 3509 | /* check whether cfs_rq, or any parent, is throttled */ |
3510 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | 3510 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) |
3511 | { | 3511 | { |
3512 | return cfs_bandwidth_used() && cfs_rq->throttle_count; | 3512 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
3513 | } | 3513 | } |
3514 | 3514 | ||
3515 | /* | 3515 | /* |
3516 | * Ensure that neither of the group entities corresponding to src_cpu or | 3516 | * Ensure that neither of the group entities corresponding to src_cpu or |
3517 | * dest_cpu are members of a throttled hierarchy when performing group | 3517 | * dest_cpu are members of a throttled hierarchy when performing group |
3518 | * load-balance operations. | 3518 | * load-balance operations. |
3519 | */ | 3519 | */ |
3520 | static inline int throttled_lb_pair(struct task_group *tg, | 3520 | static inline int throttled_lb_pair(struct task_group *tg, |
3521 | int src_cpu, int dest_cpu) | 3521 | int src_cpu, int dest_cpu) |
3522 | { | 3522 | { |
3523 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | 3523 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; |
3524 | 3524 | ||
3525 | src_cfs_rq = tg->cfs_rq[src_cpu]; | 3525 | src_cfs_rq = tg->cfs_rq[src_cpu]; |
3526 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | 3526 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; |
3527 | 3527 | ||
3528 | return throttled_hierarchy(src_cfs_rq) || | 3528 | return throttled_hierarchy(src_cfs_rq) || |
3529 | throttled_hierarchy(dest_cfs_rq); | 3529 | throttled_hierarchy(dest_cfs_rq); |
3530 | } | 3530 | } |
3531 | 3531 | ||
3532 | /* updated child weight may affect parent so we have to do this bottom up */ | 3532 | /* updated child weight may affect parent so we have to do this bottom up */ |
3533 | static int tg_unthrottle_up(struct task_group *tg, void *data) | 3533 | static int tg_unthrottle_up(struct task_group *tg, void *data) |
3534 | { | 3534 | { |
3535 | struct rq *rq = data; | 3535 | struct rq *rq = data; |
3536 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | 3536 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; |
3537 | 3537 | ||
3538 | cfs_rq->throttle_count--; | 3538 | cfs_rq->throttle_count--; |
3539 | #ifdef CONFIG_SMP | 3539 | #ifdef CONFIG_SMP |
3540 | if (!cfs_rq->throttle_count) { | 3540 | if (!cfs_rq->throttle_count) { |
3541 | /* adjust cfs_rq_clock_task() */ | 3541 | /* adjust cfs_rq_clock_task() */ |
3542 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - | 3542 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
3543 | cfs_rq->throttled_clock_task; | 3543 | cfs_rq->throttled_clock_task; |
3544 | } | 3544 | } |
3545 | #endif | 3545 | #endif |
3546 | 3546 | ||
3547 | return 0; | 3547 | return 0; |
3548 | } | 3548 | } |
3549 | 3549 | ||
3550 | static int tg_throttle_down(struct task_group *tg, void *data) | 3550 | static int tg_throttle_down(struct task_group *tg, void *data) |
3551 | { | 3551 | { |
3552 | struct rq *rq = data; | 3552 | struct rq *rq = data; |
3553 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | 3553 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; |
3554 | 3554 | ||
3555 | /* group is entering throttled state, stop time */ | 3555 | /* group is entering throttled state, stop time */ |
3556 | if (!cfs_rq->throttle_count) | 3556 | if (!cfs_rq->throttle_count) |
3557 | cfs_rq->throttled_clock_task = rq_clock_task(rq); | 3557 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
3558 | cfs_rq->throttle_count++; | 3558 | cfs_rq->throttle_count++; |
3559 | 3559 | ||
3560 | return 0; | 3560 | return 0; |
3561 | } | 3561 | } |
3562 | 3562 | ||
3563 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) | 3563 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
3564 | { | 3564 | { |
3565 | struct rq *rq = rq_of(cfs_rq); | 3565 | struct rq *rq = rq_of(cfs_rq); |
3566 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | 3566 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
3567 | struct sched_entity *se; | 3567 | struct sched_entity *se; |
3568 | long task_delta, dequeue = 1; | 3568 | long task_delta, dequeue = 1; |
3569 | 3569 | ||
3570 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | 3570 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; |
3571 | 3571 | ||
3572 | /* freeze hierarchy runnable averages while throttled */ | 3572 | /* freeze hierarchy runnable averages while throttled */ |
3573 | rcu_read_lock(); | 3573 | rcu_read_lock(); |
3574 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | 3574 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); |
3575 | rcu_read_unlock(); | 3575 | rcu_read_unlock(); |
3576 | 3576 | ||
3577 | task_delta = cfs_rq->h_nr_running; | 3577 | task_delta = cfs_rq->h_nr_running; |
3578 | for_each_sched_entity(se) { | 3578 | for_each_sched_entity(se) { |
3579 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | 3579 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); |
3580 | /* throttled entity or throttle-on-deactivate */ | 3580 | /* throttled entity or throttle-on-deactivate */ |
3581 | if (!se->on_rq) | 3581 | if (!se->on_rq) |
3582 | break; | 3582 | break; |
3583 | 3583 | ||
3584 | if (dequeue) | 3584 | if (dequeue) |
3585 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | 3585 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); |
3586 | qcfs_rq->h_nr_running -= task_delta; | 3586 | qcfs_rq->h_nr_running -= task_delta; |
3587 | 3587 | ||
3588 | if (qcfs_rq->load.weight) | 3588 | if (qcfs_rq->load.weight) |
3589 | dequeue = 0; | 3589 | dequeue = 0; |
3590 | } | 3590 | } |
3591 | 3591 | ||
3592 | if (!se) | 3592 | if (!se) |
3593 | sub_nr_running(rq, task_delta); | 3593 | sub_nr_running(rq, task_delta); |
3594 | 3594 | ||
3595 | cfs_rq->throttled = 1; | 3595 | cfs_rq->throttled = 1; |
3596 | cfs_rq->throttled_clock = rq_clock(rq); | 3596 | cfs_rq->throttled_clock = rq_clock(rq); |
3597 | raw_spin_lock(&cfs_b->lock); | 3597 | raw_spin_lock(&cfs_b->lock); |
3598 | /* | 3598 | /* |
3599 | * Add to the _head_ of the list, so that an already-started | 3599 | * Add to the _head_ of the list, so that an already-started |
3600 | * distribute_cfs_runtime will not see us | 3600 | * distribute_cfs_runtime will not see us |
3601 | */ | 3601 | */ |
3602 | list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | 3602 | list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); |
3603 | if (!cfs_b->timer_active) | 3603 | if (!cfs_b->timer_active) |
3604 | __start_cfs_bandwidth(cfs_b, false); | 3604 | __start_cfs_bandwidth(cfs_b, false); |
3605 | raw_spin_unlock(&cfs_b->lock); | 3605 | raw_spin_unlock(&cfs_b->lock); |
3606 | } | 3606 | } |
3607 | 3607 | ||
3608 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) | 3608 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
3609 | { | 3609 | { |
3610 | struct rq *rq = rq_of(cfs_rq); | 3610 | struct rq *rq = rq_of(cfs_rq); |
3611 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | 3611 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
3612 | struct sched_entity *se; | 3612 | struct sched_entity *se; |
3613 | int enqueue = 1; | 3613 | int enqueue = 1; |
3614 | long task_delta; | 3614 | long task_delta; |
3615 | 3615 | ||
3616 | se = cfs_rq->tg->se[cpu_of(rq)]; | 3616 | se = cfs_rq->tg->se[cpu_of(rq)]; |
3617 | 3617 | ||
3618 | cfs_rq->throttled = 0; | 3618 | cfs_rq->throttled = 0; |
3619 | 3619 | ||
3620 | update_rq_clock(rq); | 3620 | update_rq_clock(rq); |
3621 | 3621 | ||
3622 | raw_spin_lock(&cfs_b->lock); | 3622 | raw_spin_lock(&cfs_b->lock); |
3623 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; | 3623 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
3624 | list_del_rcu(&cfs_rq->throttled_list); | 3624 | list_del_rcu(&cfs_rq->throttled_list); |
3625 | raw_spin_unlock(&cfs_b->lock); | 3625 | raw_spin_unlock(&cfs_b->lock); |
3626 | 3626 | ||
3627 | /* update hierarchical throttle state */ | 3627 | /* update hierarchical throttle state */ |
3628 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | 3628 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); |
3629 | 3629 | ||
3630 | if (!cfs_rq->load.weight) | 3630 | if (!cfs_rq->load.weight) |
3631 | return; | 3631 | return; |
3632 | 3632 | ||
3633 | task_delta = cfs_rq->h_nr_running; | 3633 | task_delta = cfs_rq->h_nr_running; |
3634 | for_each_sched_entity(se) { | 3634 | for_each_sched_entity(se) { |
3635 | if (se->on_rq) | 3635 | if (se->on_rq) |
3636 | enqueue = 0; | 3636 | enqueue = 0; |
3637 | 3637 | ||
3638 | cfs_rq = cfs_rq_of(se); | 3638 | cfs_rq = cfs_rq_of(se); |
3639 | if (enqueue) | 3639 | if (enqueue) |
3640 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | 3640 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); |
3641 | cfs_rq->h_nr_running += task_delta; | 3641 | cfs_rq->h_nr_running += task_delta; |
3642 | 3642 | ||
3643 | if (cfs_rq_throttled(cfs_rq)) | 3643 | if (cfs_rq_throttled(cfs_rq)) |
3644 | break; | 3644 | break; |
3645 | } | 3645 | } |
3646 | 3646 | ||
3647 | if (!se) | 3647 | if (!se) |
3648 | add_nr_running(rq, task_delta); | 3648 | add_nr_running(rq, task_delta); |
3649 | 3649 | ||
3650 | /* determine whether we need to wake up potentially idle cpu */ | 3650 | /* determine whether we need to wake up potentially idle cpu */ |
3651 | if (rq->curr == rq->idle && rq->cfs.nr_running) | 3651 | if (rq->curr == rq->idle && rq->cfs.nr_running) |
3652 | resched_curr(rq); | 3652 | resched_curr(rq); |
3653 | } | 3653 | } |
3654 | 3654 | ||
3655 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | 3655 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, |
3656 | u64 remaining, u64 expires) | 3656 | u64 remaining, u64 expires) |
3657 | { | 3657 | { |
3658 | struct cfs_rq *cfs_rq; | 3658 | struct cfs_rq *cfs_rq; |
3659 | u64 runtime; | 3659 | u64 runtime; |
3660 | u64 starting_runtime = remaining; | 3660 | u64 starting_runtime = remaining; |
3661 | 3661 | ||
3662 | rcu_read_lock(); | 3662 | rcu_read_lock(); |
3663 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | 3663 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, |
3664 | throttled_list) { | 3664 | throttled_list) { |
3665 | struct rq *rq = rq_of(cfs_rq); | 3665 | struct rq *rq = rq_of(cfs_rq); |
3666 | 3666 | ||
3667 | raw_spin_lock(&rq->lock); | 3667 | raw_spin_lock(&rq->lock); |
3668 | if (!cfs_rq_throttled(cfs_rq)) | 3668 | if (!cfs_rq_throttled(cfs_rq)) |
3669 | goto next; | 3669 | goto next; |
3670 | 3670 | ||
3671 | runtime = -cfs_rq->runtime_remaining + 1; | 3671 | runtime = -cfs_rq->runtime_remaining + 1; |
3672 | if (runtime > remaining) | 3672 | if (runtime > remaining) |
3673 | runtime = remaining; | 3673 | runtime = remaining; |
3674 | remaining -= runtime; | 3674 | remaining -= runtime; |
3675 | 3675 | ||
3676 | cfs_rq->runtime_remaining += runtime; | 3676 | cfs_rq->runtime_remaining += runtime; |
3677 | cfs_rq->runtime_expires = expires; | 3677 | cfs_rq->runtime_expires = expires; |
3678 | 3678 | ||
3679 | /* we check whether we're throttled above */ | 3679 | /* we check whether we're throttled above */ |
3680 | if (cfs_rq->runtime_remaining > 0) | 3680 | if (cfs_rq->runtime_remaining > 0) |
3681 | unthrottle_cfs_rq(cfs_rq); | 3681 | unthrottle_cfs_rq(cfs_rq); |
3682 | 3682 | ||
3683 | next: | 3683 | next: |
3684 | raw_spin_unlock(&rq->lock); | 3684 | raw_spin_unlock(&rq->lock); |
3685 | 3685 | ||
3686 | if (!remaining) | 3686 | if (!remaining) |
3687 | break; | 3687 | break; |
3688 | } | 3688 | } |
3689 | rcu_read_unlock(); | 3689 | rcu_read_unlock(); |
3690 | 3690 | ||
3691 | return starting_runtime - remaining; | 3691 | return starting_runtime - remaining; |
3692 | } | 3692 | } |
3693 | 3693 | ||
3694 | /* | 3694 | /* |
3695 | * Responsible for refilling a task_group's bandwidth and unthrottling its | 3695 | * Responsible for refilling a task_group's bandwidth and unthrottling its |
3696 | * cfs_rqs as appropriate. If there has been no activity within the last | 3696 | * cfs_rqs as appropriate. If there has been no activity within the last |
3697 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | 3697 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is |
3698 | * used to track this state. | 3698 | * used to track this state. |
3699 | */ | 3699 | */ |
3700 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | 3700 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) |
3701 | { | 3701 | { |
3702 | u64 runtime, runtime_expires; | 3702 | u64 runtime, runtime_expires; |
3703 | int throttled; | 3703 | int throttled; |
3704 | 3704 | ||
3705 | /* no need to continue the timer with no bandwidth constraint */ | 3705 | /* no need to continue the timer with no bandwidth constraint */ |
3706 | if (cfs_b->quota == RUNTIME_INF) | 3706 | if (cfs_b->quota == RUNTIME_INF) |
3707 | goto out_deactivate; | 3707 | goto out_deactivate; |
3708 | 3708 | ||
3709 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | 3709 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
3710 | cfs_b->nr_periods += overrun; | 3710 | cfs_b->nr_periods += overrun; |
3711 | 3711 | ||
3712 | /* | 3712 | /* |
3713 | * idle depends on !throttled (for the case of a large deficit), and if | 3713 | * idle depends on !throttled (for the case of a large deficit), and if |
3714 | * we're going inactive then everything else can be deferred | 3714 | * we're going inactive then everything else can be deferred |
3715 | */ | 3715 | */ |
3716 | if (cfs_b->idle && !throttled) | 3716 | if (cfs_b->idle && !throttled) |
3717 | goto out_deactivate; | 3717 | goto out_deactivate; |
3718 | 3718 | ||
3719 | /* | 3719 | /* |
3720 | * if we have relooped after returning idle once, we need to update our | 3720 | * if we have relooped after returning idle once, we need to update our |
3721 | * status as actually running, so that other cpus doing | 3721 | * status as actually running, so that other cpus doing |
3722 | * __start_cfs_bandwidth will stop trying to cancel us. | 3722 | * __start_cfs_bandwidth will stop trying to cancel us. |
3723 | */ | 3723 | */ |
3724 | cfs_b->timer_active = 1; | 3724 | cfs_b->timer_active = 1; |
3725 | 3725 | ||
3726 | __refill_cfs_bandwidth_runtime(cfs_b); | 3726 | __refill_cfs_bandwidth_runtime(cfs_b); |
3727 | 3727 | ||
3728 | if (!throttled) { | 3728 | if (!throttled) { |
3729 | /* mark as potentially idle for the upcoming period */ | 3729 | /* mark as potentially idle for the upcoming period */ |
3730 | cfs_b->idle = 1; | 3730 | cfs_b->idle = 1; |
3731 | return 0; | 3731 | return 0; |
3732 | } | 3732 | } |
3733 | 3733 | ||
3734 | /* account preceding periods in which throttling occurred */ | 3734 | /* account preceding periods in which throttling occurred */ |
3735 | cfs_b->nr_throttled += overrun; | 3735 | cfs_b->nr_throttled += overrun; |
3736 | 3736 | ||
3737 | runtime_expires = cfs_b->runtime_expires; | 3737 | runtime_expires = cfs_b->runtime_expires; |
3738 | 3738 | ||
3739 | /* | 3739 | /* |
3740 | * This check is repeated as we are holding onto the new bandwidth while | 3740 | * This check is repeated as we are holding onto the new bandwidth while |
3741 | * we unthrottle. This can potentially race with an unthrottled group | 3741 | * we unthrottle. This can potentially race with an unthrottled group |
3742 | * trying to acquire new bandwidth from the global pool. This can result | 3742 | * trying to acquire new bandwidth from the global pool. This can result |
3743 | * in us over-using our runtime if it is all used during this loop, but | 3743 | * in us over-using our runtime if it is all used during this loop, but |
3744 | * only by limited amounts in that extreme case. | 3744 | * only by limited amounts in that extreme case. |
3745 | */ | 3745 | */ |
3746 | while (throttled && cfs_b->runtime > 0) { | 3746 | while (throttled && cfs_b->runtime > 0) { |
3747 | runtime = cfs_b->runtime; | 3747 | runtime = cfs_b->runtime; |
3748 | raw_spin_unlock(&cfs_b->lock); | 3748 | raw_spin_unlock(&cfs_b->lock); |
3749 | /* we can't nest cfs_b->lock while distributing bandwidth */ | 3749 | /* we can't nest cfs_b->lock while distributing bandwidth */ |
3750 | runtime = distribute_cfs_runtime(cfs_b, runtime, | 3750 | runtime = distribute_cfs_runtime(cfs_b, runtime, |
3751 | runtime_expires); | 3751 | runtime_expires); |
3752 | raw_spin_lock(&cfs_b->lock); | 3752 | raw_spin_lock(&cfs_b->lock); |
3753 | 3753 | ||
3754 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | 3754 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
3755 | 3755 | ||
3756 | cfs_b->runtime -= min(runtime, cfs_b->runtime); | 3756 | cfs_b->runtime -= min(runtime, cfs_b->runtime); |
3757 | } | 3757 | } |
3758 | 3758 | ||
3759 | /* | 3759 | /* |
3760 | * While we are ensured activity in the period following an | 3760 | * While we are ensured activity in the period following an |
3761 | * unthrottle, this also covers the case in which the new bandwidth is | 3761 | * unthrottle, this also covers the case in which the new bandwidth is |
3762 | * insufficient to cover the existing bandwidth deficit. (Forcing the | 3762 | * insufficient to cover the existing bandwidth deficit. (Forcing the |
3763 | * timer to remain active while there are any throttled entities.) | 3763 | * timer to remain active while there are any throttled entities.) |
3764 | */ | 3764 | */ |
3765 | cfs_b->idle = 0; | 3765 | cfs_b->idle = 0; |
3766 | 3766 | ||
3767 | return 0; | 3767 | return 0; |
3768 | 3768 | ||
3769 | out_deactivate: | 3769 | out_deactivate: |
3770 | cfs_b->timer_active = 0; | 3770 | cfs_b->timer_active = 0; |
3771 | return 1; | 3771 | return 1; |
3772 | } | 3772 | } |
3773 | 3773 | ||
3774 | /* a cfs_rq won't donate quota below this amount */ | 3774 | /* a cfs_rq won't donate quota below this amount */ |
3775 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | 3775 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; |
3776 | /* minimum remaining period time to redistribute slack quota */ | 3776 | /* minimum remaining period time to redistribute slack quota */ |
3777 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | 3777 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; |
3778 | /* how long we wait to gather additional slack before distributing */ | 3778 | /* how long we wait to gather additional slack before distributing */ |
3779 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | 3779 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; |
3780 | 3780 | ||
3781 | /* | 3781 | /* |
3782 | * Are we near the end of the current quota period? | 3782 | * Are we near the end of the current quota period? |
3783 | * | 3783 | * |
3784 | * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the | 3784 | * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the |
3785 | * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of | 3785 | * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of |
3786 | * migrate_hrtimers, base is never cleared, so we are fine. | 3786 | * migrate_hrtimers, base is never cleared, so we are fine. |
3787 | */ | 3787 | */ |
3788 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) | 3788 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) |
3789 | { | 3789 | { |
3790 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | 3790 | struct hrtimer *refresh_timer = &cfs_b->period_timer; |
3791 | u64 remaining; | 3791 | u64 remaining; |
3792 | 3792 | ||
3793 | /* if the call-back is running a quota refresh is already occurring */ | 3793 | /* if the call-back is running a quota refresh is already occurring */ |
3794 | if (hrtimer_callback_running(refresh_timer)) | 3794 | if (hrtimer_callback_running(refresh_timer)) |
3795 | return 1; | 3795 | return 1; |
3796 | 3796 | ||
3797 | /* is a quota refresh about to occur? */ | 3797 | /* is a quota refresh about to occur? */ |
3798 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | 3798 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); |
3799 | if (remaining < min_expire) | 3799 | if (remaining < min_expire) |
3800 | return 1; | 3800 | return 1; |
3801 | 3801 | ||
3802 | return 0; | 3802 | return 0; |
3803 | } | 3803 | } |
3804 | 3804 | ||
3805 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | 3805 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) |
3806 | { | 3806 | { |
3807 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | 3807 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; |
3808 | 3808 | ||
3809 | /* if there's a quota refresh soon don't bother with slack */ | 3809 | /* if there's a quota refresh soon don't bother with slack */ |
3810 | if (runtime_refresh_within(cfs_b, min_left)) | 3810 | if (runtime_refresh_within(cfs_b, min_left)) |
3811 | return; | 3811 | return; |
3812 | 3812 | ||
3813 | start_bandwidth_timer(&cfs_b->slack_timer, | 3813 | start_bandwidth_timer(&cfs_b->slack_timer, |
3814 | ns_to_ktime(cfs_bandwidth_slack_period)); | 3814 | ns_to_ktime(cfs_bandwidth_slack_period)); |
3815 | } | 3815 | } |
3816 | 3816 | ||
3817 | /* we know any runtime found here is valid as update_curr() precedes return */ | 3817 | /* we know any runtime found here is valid as update_curr() precedes return */ |
3818 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | 3818 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
3819 | { | 3819 | { |
3820 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | 3820 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
3821 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | 3821 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; |
3822 | 3822 | ||
3823 | if (slack_runtime <= 0) | 3823 | if (slack_runtime <= 0) |
3824 | return; | 3824 | return; |
3825 | 3825 | ||
3826 | raw_spin_lock(&cfs_b->lock); | 3826 | raw_spin_lock(&cfs_b->lock); |
3827 | if (cfs_b->quota != RUNTIME_INF && | 3827 | if (cfs_b->quota != RUNTIME_INF && |
3828 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | 3828 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { |
3829 | cfs_b->runtime += slack_runtime; | 3829 | cfs_b->runtime += slack_runtime; |
3830 | 3830 | ||
3831 | /* we are under rq->lock, defer unthrottling using a timer */ | 3831 | /* we are under rq->lock, defer unthrottling using a timer */ |
3832 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | 3832 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && |
3833 | !list_empty(&cfs_b->throttled_cfs_rq)) | 3833 | !list_empty(&cfs_b->throttled_cfs_rq)) |
3834 | start_cfs_slack_bandwidth(cfs_b); | 3834 | start_cfs_slack_bandwidth(cfs_b); |
3835 | } | 3835 | } |
3836 | raw_spin_unlock(&cfs_b->lock); | 3836 | raw_spin_unlock(&cfs_b->lock); |
3837 | 3837 | ||
3838 | /* even if it's not valid for return we don't want to try again */ | 3838 | /* even if it's not valid for return we don't want to try again */ |
3839 | cfs_rq->runtime_remaining -= slack_runtime; | 3839 | cfs_rq->runtime_remaining -= slack_runtime; |
3840 | } | 3840 | } |
3841 | 3841 | ||
3842 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | 3842 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
3843 | { | 3843 | { |
3844 | if (!cfs_bandwidth_used()) | 3844 | if (!cfs_bandwidth_used()) |
3845 | return; | 3845 | return; |
3846 | 3846 | ||
3847 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) | 3847 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
3848 | return; | 3848 | return; |
3849 | 3849 | ||
3850 | __return_cfs_rq_runtime(cfs_rq); | 3850 | __return_cfs_rq_runtime(cfs_rq); |
3851 | } | 3851 | } |
3852 | 3852 | ||
3853 | /* | 3853 | /* |
3854 | * This is done with a timer (instead of inline with bandwidth return) since | 3854 | * This is done with a timer (instead of inline with bandwidth return) since |
3855 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | 3855 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. |
3856 | */ | 3856 | */ |
3857 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | 3857 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) |
3858 | { | 3858 | { |
3859 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | 3859 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); |
3860 | u64 expires; | 3860 | u64 expires; |
3861 | 3861 | ||
3862 | /* confirm we're still not at a refresh boundary */ | 3862 | /* confirm we're still not at a refresh boundary */ |
3863 | raw_spin_lock(&cfs_b->lock); | 3863 | raw_spin_lock(&cfs_b->lock); |
3864 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { | 3864 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { |
3865 | raw_spin_unlock(&cfs_b->lock); | 3865 | raw_spin_unlock(&cfs_b->lock); |
3866 | return; | 3866 | return; |
3867 | } | 3867 | } |
3868 | 3868 | ||
3869 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) | 3869 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) |
3870 | runtime = cfs_b->runtime; | 3870 | runtime = cfs_b->runtime; |
3871 | 3871 | ||
3872 | expires = cfs_b->runtime_expires; | 3872 | expires = cfs_b->runtime_expires; |
3873 | raw_spin_unlock(&cfs_b->lock); | 3873 | raw_spin_unlock(&cfs_b->lock); |
3874 | 3874 | ||
3875 | if (!runtime) | 3875 | if (!runtime) |
3876 | return; | 3876 | return; |
3877 | 3877 | ||
3878 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | 3878 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); |
3879 | 3879 | ||
3880 | raw_spin_lock(&cfs_b->lock); | 3880 | raw_spin_lock(&cfs_b->lock); |
3881 | if (expires == cfs_b->runtime_expires) | 3881 | if (expires == cfs_b->runtime_expires) |
3882 | cfs_b->runtime -= min(runtime, cfs_b->runtime); | 3882 | cfs_b->runtime -= min(runtime, cfs_b->runtime); |
3883 | raw_spin_unlock(&cfs_b->lock); | 3883 | raw_spin_unlock(&cfs_b->lock); |
3884 | } | 3884 | } |
3885 | 3885 | ||
3886 | /* | 3886 | /* |
3887 | * When a group wakes up we want to make sure that its quota is not already | 3887 | * When a group wakes up we want to make sure that its quota is not already |
3888 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | 3888 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of |
3889 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | 3889 | * runtime as update_curr() throttling can not not trigger until it's on-rq. |
3890 | */ | 3890 | */ |
3891 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | 3891 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) |
3892 | { | 3892 | { |
3893 | if (!cfs_bandwidth_used()) | 3893 | if (!cfs_bandwidth_used()) |
3894 | return; | 3894 | return; |
3895 | 3895 | ||
3896 | /* an active group must be handled by the update_curr()->put() path */ | 3896 | /* an active group must be handled by the update_curr()->put() path */ |
3897 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | 3897 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) |
3898 | return; | 3898 | return; |
3899 | 3899 | ||
3900 | /* ensure the group is not already throttled */ | 3900 | /* ensure the group is not already throttled */ |
3901 | if (cfs_rq_throttled(cfs_rq)) | 3901 | if (cfs_rq_throttled(cfs_rq)) |
3902 | return; | 3902 | return; |
3903 | 3903 | ||
3904 | /* update runtime allocation */ | 3904 | /* update runtime allocation */ |
3905 | account_cfs_rq_runtime(cfs_rq, 0); | 3905 | account_cfs_rq_runtime(cfs_rq, 0); |
3906 | if (cfs_rq->runtime_remaining <= 0) | 3906 | if (cfs_rq->runtime_remaining <= 0) |
3907 | throttle_cfs_rq(cfs_rq); | 3907 | throttle_cfs_rq(cfs_rq); |
3908 | } | 3908 | } |
3909 | 3909 | ||
3910 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ | 3910 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ |
3911 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) | 3911 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
3912 | { | 3912 | { |
3913 | if (!cfs_bandwidth_used()) | 3913 | if (!cfs_bandwidth_used()) |
3914 | return false; | 3914 | return false; |
3915 | 3915 | ||
3916 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) | 3916 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
3917 | return false; | 3917 | return false; |
3918 | 3918 | ||
3919 | /* | 3919 | /* |
3920 | * it's possible for a throttled entity to be forced into a running | 3920 | * it's possible for a throttled entity to be forced into a running |
3921 | * state (e.g. set_curr_task), in this case we're finished. | 3921 | * state (e.g. set_curr_task), in this case we're finished. |
3922 | */ | 3922 | */ |
3923 | if (cfs_rq_throttled(cfs_rq)) | 3923 | if (cfs_rq_throttled(cfs_rq)) |
3924 | return true; | 3924 | return true; |
3925 | 3925 | ||
3926 | throttle_cfs_rq(cfs_rq); | 3926 | throttle_cfs_rq(cfs_rq); |
3927 | return true; | 3927 | return true; |
3928 | } | 3928 | } |
3929 | 3929 | ||
3930 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) | 3930 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
3931 | { | 3931 | { |
3932 | struct cfs_bandwidth *cfs_b = | 3932 | struct cfs_bandwidth *cfs_b = |
3933 | container_of(timer, struct cfs_bandwidth, slack_timer); | 3933 | container_of(timer, struct cfs_bandwidth, slack_timer); |
3934 | do_sched_cfs_slack_timer(cfs_b); | 3934 | do_sched_cfs_slack_timer(cfs_b); |
3935 | 3935 | ||
3936 | return HRTIMER_NORESTART; | 3936 | return HRTIMER_NORESTART; |
3937 | } | 3937 | } |
3938 | 3938 | ||
3939 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | 3939 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) |
3940 | { | 3940 | { |
3941 | struct cfs_bandwidth *cfs_b = | 3941 | struct cfs_bandwidth *cfs_b = |
3942 | container_of(timer, struct cfs_bandwidth, period_timer); | 3942 | container_of(timer, struct cfs_bandwidth, period_timer); |
3943 | ktime_t now; | 3943 | ktime_t now; |
3944 | int overrun; | 3944 | int overrun; |
3945 | int idle = 0; | 3945 | int idle = 0; |
3946 | 3946 | ||
3947 | raw_spin_lock(&cfs_b->lock); | 3947 | raw_spin_lock(&cfs_b->lock); |
3948 | for (;;) { | 3948 | for (;;) { |
3949 | now = hrtimer_cb_get_time(timer); | 3949 | now = hrtimer_cb_get_time(timer); |
3950 | overrun = hrtimer_forward(timer, now, cfs_b->period); | 3950 | overrun = hrtimer_forward(timer, now, cfs_b->period); |
3951 | 3951 | ||
3952 | if (!overrun) | 3952 | if (!overrun) |
3953 | break; | 3953 | break; |
3954 | 3954 | ||
3955 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | 3955 | idle = do_sched_cfs_period_timer(cfs_b, overrun); |
3956 | } | 3956 | } |
3957 | raw_spin_unlock(&cfs_b->lock); | 3957 | raw_spin_unlock(&cfs_b->lock); |
3958 | 3958 | ||
3959 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | 3959 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; |
3960 | } | 3960 | } |
3961 | 3961 | ||
3962 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | 3962 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) |
3963 | { | 3963 | { |
3964 | raw_spin_lock_init(&cfs_b->lock); | 3964 | raw_spin_lock_init(&cfs_b->lock); |
3965 | cfs_b->runtime = 0; | 3965 | cfs_b->runtime = 0; |
3966 | cfs_b->quota = RUNTIME_INF; | 3966 | cfs_b->quota = RUNTIME_INF; |
3967 | cfs_b->period = ns_to_ktime(default_cfs_period()); | 3967 | cfs_b->period = ns_to_ktime(default_cfs_period()); |
3968 | 3968 | ||
3969 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | 3969 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); |
3970 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | 3970 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
3971 | cfs_b->period_timer.function = sched_cfs_period_timer; | 3971 | cfs_b->period_timer.function = sched_cfs_period_timer; |
3972 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | 3972 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
3973 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | 3973 | cfs_b->slack_timer.function = sched_cfs_slack_timer; |
3974 | } | 3974 | } |
3975 | 3975 | ||
3976 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | 3976 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
3977 | { | 3977 | { |
3978 | cfs_rq->runtime_enabled = 0; | 3978 | cfs_rq->runtime_enabled = 0; |
3979 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | 3979 | INIT_LIST_HEAD(&cfs_rq->throttled_list); |
3980 | } | 3980 | } |
3981 | 3981 | ||
3982 | /* requires cfs_b->lock, may release to reprogram timer */ | 3982 | /* requires cfs_b->lock, may release to reprogram timer */ |
3983 | void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force) | 3983 | void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force) |
3984 | { | 3984 | { |
3985 | /* | 3985 | /* |
3986 | * The timer may be active because we're trying to set a new bandwidth | 3986 | * The timer may be active because we're trying to set a new bandwidth |
3987 | * period or because we're racing with the tear-down path | 3987 | * period or because we're racing with the tear-down path |
3988 | * (timer_active==0 becomes visible before the hrtimer call-back | 3988 | * (timer_active==0 becomes visible before the hrtimer call-back |
3989 | * terminates). In either case we ensure that it's re-programmed | 3989 | * terminates). In either case we ensure that it's re-programmed |
3990 | */ | 3990 | */ |
3991 | while (unlikely(hrtimer_active(&cfs_b->period_timer)) && | 3991 | while (unlikely(hrtimer_active(&cfs_b->period_timer)) && |
3992 | hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) { | 3992 | hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) { |
3993 | /* bounce the lock to allow do_sched_cfs_period_timer to run */ | 3993 | /* bounce the lock to allow do_sched_cfs_period_timer to run */ |
3994 | raw_spin_unlock(&cfs_b->lock); | 3994 | raw_spin_unlock(&cfs_b->lock); |
3995 | cpu_relax(); | 3995 | cpu_relax(); |
3996 | raw_spin_lock(&cfs_b->lock); | 3996 | raw_spin_lock(&cfs_b->lock); |
3997 | /* if someone else restarted the timer then we're done */ | 3997 | /* if someone else restarted the timer then we're done */ |
3998 | if (!force && cfs_b->timer_active) | 3998 | if (!force && cfs_b->timer_active) |
3999 | return; | 3999 | return; |
4000 | } | 4000 | } |
4001 | 4001 | ||
4002 | cfs_b->timer_active = 1; | 4002 | cfs_b->timer_active = 1; |
4003 | start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); | 4003 | start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); |
4004 | } | 4004 | } |
4005 | 4005 | ||
4006 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | 4006 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) |
4007 | { | 4007 | { |
4008 | /* init_cfs_bandwidth() was not called */ | ||
4009 | if (!cfs_b->throttled_cfs_rq.next) | ||
4010 | return; | ||
4011 | |||
4008 | hrtimer_cancel(&cfs_b->period_timer); | 4012 | hrtimer_cancel(&cfs_b->period_timer); |
4009 | hrtimer_cancel(&cfs_b->slack_timer); | 4013 | hrtimer_cancel(&cfs_b->slack_timer); |
4010 | } | 4014 | } |
4011 | 4015 | ||
4012 | static void __maybe_unused update_runtime_enabled(struct rq *rq) | 4016 | static void __maybe_unused update_runtime_enabled(struct rq *rq) |
4013 | { | 4017 | { |
4014 | struct cfs_rq *cfs_rq; | 4018 | struct cfs_rq *cfs_rq; |
4015 | 4019 | ||
4016 | for_each_leaf_cfs_rq(rq, cfs_rq) { | 4020 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
4017 | struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth; | 4021 | struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth; |
4018 | 4022 | ||
4019 | raw_spin_lock(&cfs_b->lock); | 4023 | raw_spin_lock(&cfs_b->lock); |
4020 | cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; | 4024 | cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; |
4021 | raw_spin_unlock(&cfs_b->lock); | 4025 | raw_spin_unlock(&cfs_b->lock); |
4022 | } | 4026 | } |
4023 | } | 4027 | } |
4024 | 4028 | ||
4025 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) | 4029 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
4026 | { | 4030 | { |
4027 | struct cfs_rq *cfs_rq; | 4031 | struct cfs_rq *cfs_rq; |
4028 | 4032 | ||
4029 | for_each_leaf_cfs_rq(rq, cfs_rq) { | 4033 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
4030 | if (!cfs_rq->runtime_enabled) | 4034 | if (!cfs_rq->runtime_enabled) |
4031 | continue; | 4035 | continue; |
4032 | 4036 | ||
4033 | /* | 4037 | /* |
4034 | * clock_task is not advancing so we just need to make sure | 4038 | * clock_task is not advancing so we just need to make sure |
4035 | * there's some valid quota amount | 4039 | * there's some valid quota amount |
4036 | */ | 4040 | */ |
4037 | cfs_rq->runtime_remaining = 1; | 4041 | cfs_rq->runtime_remaining = 1; |
4038 | /* | 4042 | /* |
4039 | * Offline rq is schedulable till cpu is completely disabled | 4043 | * Offline rq is schedulable till cpu is completely disabled |
4040 | * in take_cpu_down(), so we prevent new cfs throttling here. | 4044 | * in take_cpu_down(), so we prevent new cfs throttling here. |
4041 | */ | 4045 | */ |
4042 | cfs_rq->runtime_enabled = 0; | 4046 | cfs_rq->runtime_enabled = 0; |
4043 | 4047 | ||
4044 | if (cfs_rq_throttled(cfs_rq)) | 4048 | if (cfs_rq_throttled(cfs_rq)) |
4045 | unthrottle_cfs_rq(cfs_rq); | 4049 | unthrottle_cfs_rq(cfs_rq); |
4046 | } | 4050 | } |
4047 | } | 4051 | } |
4048 | 4052 | ||
4049 | #else /* CONFIG_CFS_BANDWIDTH */ | 4053 | #else /* CONFIG_CFS_BANDWIDTH */ |
4050 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | 4054 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
4051 | { | 4055 | { |
4052 | return rq_clock_task(rq_of(cfs_rq)); | 4056 | return rq_clock_task(rq_of(cfs_rq)); |
4053 | } | 4057 | } |
4054 | 4058 | ||
4055 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} | 4059 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} |
4056 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } | 4060 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } |
4057 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} | 4061 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} |
4058 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | 4062 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
4059 | 4063 | ||
4060 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | 4064 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
4061 | { | 4065 | { |
4062 | return 0; | 4066 | return 0; |
4063 | } | 4067 | } |
4064 | 4068 | ||
4065 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | 4069 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) |
4066 | { | 4070 | { |
4067 | return 0; | 4071 | return 0; |
4068 | } | 4072 | } |
4069 | 4073 | ||
4070 | static inline int throttled_lb_pair(struct task_group *tg, | 4074 | static inline int throttled_lb_pair(struct task_group *tg, |
4071 | int src_cpu, int dest_cpu) | 4075 | int src_cpu, int dest_cpu) |
4072 | { | 4076 | { |
4073 | return 0; | 4077 | return 0; |
4074 | } | 4078 | } |
4075 | 4079 | ||
4076 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | 4080 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} |
4077 | 4081 | ||
4078 | #ifdef CONFIG_FAIR_GROUP_SCHED | 4082 | #ifdef CONFIG_FAIR_GROUP_SCHED |
4079 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | 4083 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
4080 | #endif | 4084 | #endif |
4081 | 4085 | ||
4082 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) | 4086 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
4083 | { | 4087 | { |
4084 | return NULL; | 4088 | return NULL; |
4085 | } | 4089 | } |
4086 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | 4090 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} |
4087 | static inline void update_runtime_enabled(struct rq *rq) {} | 4091 | static inline void update_runtime_enabled(struct rq *rq) {} |
4088 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} | 4092 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
4089 | 4093 | ||
4090 | #endif /* CONFIG_CFS_BANDWIDTH */ | 4094 | #endif /* CONFIG_CFS_BANDWIDTH */ |
4091 | 4095 | ||
4092 | /************************************************** | 4096 | /************************************************** |
4093 | * CFS operations on tasks: | 4097 | * CFS operations on tasks: |
4094 | */ | 4098 | */ |
4095 | 4099 | ||
4096 | #ifdef CONFIG_SCHED_HRTICK | 4100 | #ifdef CONFIG_SCHED_HRTICK |
4097 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | 4101 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) |
4098 | { | 4102 | { |
4099 | struct sched_entity *se = &p->se; | 4103 | struct sched_entity *se = &p->se; |
4100 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | 4104 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
4101 | 4105 | ||
4102 | WARN_ON(task_rq(p) != rq); | 4106 | WARN_ON(task_rq(p) != rq); |
4103 | 4107 | ||
4104 | if (cfs_rq->nr_running > 1) { | 4108 | if (cfs_rq->nr_running > 1) { |
4105 | u64 slice = sched_slice(cfs_rq, se); | 4109 | u64 slice = sched_slice(cfs_rq, se); |
4106 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | 4110 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; |
4107 | s64 delta = slice - ran; | 4111 | s64 delta = slice - ran; |
4108 | 4112 | ||
4109 | if (delta < 0) { | 4113 | if (delta < 0) { |
4110 | if (rq->curr == p) | 4114 | if (rq->curr == p) |
4111 | resched_curr(rq); | 4115 | resched_curr(rq); |
4112 | return; | 4116 | return; |
4113 | } | 4117 | } |
4114 | hrtick_start(rq, delta); | 4118 | hrtick_start(rq, delta); |
4115 | } | 4119 | } |
4116 | } | 4120 | } |
4117 | 4121 | ||
4118 | /* | 4122 | /* |
4119 | * called from enqueue/dequeue and updates the hrtick when the | 4123 | * called from enqueue/dequeue and updates the hrtick when the |
4120 | * current task is from our class and nr_running is low enough | 4124 | * current task is from our class and nr_running is low enough |
4121 | * to matter. | 4125 | * to matter. |
4122 | */ | 4126 | */ |
4123 | static void hrtick_update(struct rq *rq) | 4127 | static void hrtick_update(struct rq *rq) |
4124 | { | 4128 | { |
4125 | struct task_struct *curr = rq->curr; | 4129 | struct task_struct *curr = rq->curr; |
4126 | 4130 | ||
4127 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) | 4131 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
4128 | return; | 4132 | return; |
4129 | 4133 | ||
4130 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | 4134 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) |
4131 | hrtick_start_fair(rq, curr); | 4135 | hrtick_start_fair(rq, curr); |
4132 | } | 4136 | } |
4133 | #else /* !CONFIG_SCHED_HRTICK */ | 4137 | #else /* !CONFIG_SCHED_HRTICK */ |
4134 | static inline void | 4138 | static inline void |
4135 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | 4139 | hrtick_start_fair(struct rq *rq, struct task_struct *p) |
4136 | { | 4140 | { |
4137 | } | 4141 | } |
4138 | 4142 | ||
4139 | static inline void hrtick_update(struct rq *rq) | 4143 | static inline void hrtick_update(struct rq *rq) |
4140 | { | 4144 | { |
4141 | } | 4145 | } |
4142 | #endif | 4146 | #endif |
4143 | 4147 | ||
4144 | /* | 4148 | /* |
4145 | * The enqueue_task method is called before nr_running is | 4149 | * The enqueue_task method is called before nr_running is |
4146 | * increased. Here we update the fair scheduling stats and | 4150 | * increased. Here we update the fair scheduling stats and |
4147 | * then put the task into the rbtree: | 4151 | * then put the task into the rbtree: |
4148 | */ | 4152 | */ |
4149 | static void | 4153 | static void |
4150 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) | 4154 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
4151 | { | 4155 | { |
4152 | struct cfs_rq *cfs_rq; | 4156 | struct cfs_rq *cfs_rq; |
4153 | struct sched_entity *se = &p->se; | 4157 | struct sched_entity *se = &p->se; |
4154 | 4158 | ||
4155 | for_each_sched_entity(se) { | 4159 | for_each_sched_entity(se) { |
4156 | if (se->on_rq) | 4160 | if (se->on_rq) |
4157 | break; | 4161 | break; |
4158 | cfs_rq = cfs_rq_of(se); | 4162 | cfs_rq = cfs_rq_of(se); |
4159 | enqueue_entity(cfs_rq, se, flags); | 4163 | enqueue_entity(cfs_rq, se, flags); |
4160 | 4164 | ||
4161 | /* | 4165 | /* |
4162 | * end evaluation on encountering a throttled cfs_rq | 4166 | * end evaluation on encountering a throttled cfs_rq |
4163 | * | 4167 | * |
4164 | * note: in the case of encountering a throttled cfs_rq we will | 4168 | * note: in the case of encountering a throttled cfs_rq we will |
4165 | * post the final h_nr_running increment below. | 4169 | * post the final h_nr_running increment below. |
4166 | */ | 4170 | */ |
4167 | if (cfs_rq_throttled(cfs_rq)) | 4171 | if (cfs_rq_throttled(cfs_rq)) |
4168 | break; | 4172 | break; |
4169 | cfs_rq->h_nr_running++; | 4173 | cfs_rq->h_nr_running++; |
4170 | 4174 | ||
4171 | flags = ENQUEUE_WAKEUP; | 4175 | flags = ENQUEUE_WAKEUP; |
4172 | } | 4176 | } |
4173 | 4177 | ||
4174 | for_each_sched_entity(se) { | 4178 | for_each_sched_entity(se) { |
4175 | cfs_rq = cfs_rq_of(se); | 4179 | cfs_rq = cfs_rq_of(se); |
4176 | cfs_rq->h_nr_running++; | 4180 | cfs_rq->h_nr_running++; |
4177 | 4181 | ||
4178 | if (cfs_rq_throttled(cfs_rq)) | 4182 | if (cfs_rq_throttled(cfs_rq)) |
4179 | break; | 4183 | break; |
4180 | 4184 | ||
4181 | update_cfs_shares(cfs_rq); | 4185 | update_cfs_shares(cfs_rq); |
4182 | update_entity_load_avg(se, 1); | 4186 | update_entity_load_avg(se, 1); |
4183 | } | 4187 | } |
4184 | 4188 | ||
4185 | if (!se) { | 4189 | if (!se) { |
4186 | update_rq_runnable_avg(rq, rq->nr_running); | 4190 | update_rq_runnable_avg(rq, rq->nr_running); |
4187 | add_nr_running(rq, 1); | 4191 | add_nr_running(rq, 1); |
4188 | } | 4192 | } |
4189 | hrtick_update(rq); | 4193 | hrtick_update(rq); |
4190 | } | 4194 | } |
4191 | 4195 | ||
4192 | static void set_next_buddy(struct sched_entity *se); | 4196 | static void set_next_buddy(struct sched_entity *se); |
4193 | 4197 | ||
4194 | /* | 4198 | /* |
4195 | * The dequeue_task method is called before nr_running is | 4199 | * The dequeue_task method is called before nr_running is |
4196 | * decreased. We remove the task from the rbtree and | 4200 | * decreased. We remove the task from the rbtree and |
4197 | * update the fair scheduling stats: | 4201 | * update the fair scheduling stats: |
4198 | */ | 4202 | */ |
4199 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) | 4203 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
4200 | { | 4204 | { |
4201 | struct cfs_rq *cfs_rq; | 4205 | struct cfs_rq *cfs_rq; |
4202 | struct sched_entity *se = &p->se; | 4206 | struct sched_entity *se = &p->se; |
4203 | int task_sleep = flags & DEQUEUE_SLEEP; | 4207 | int task_sleep = flags & DEQUEUE_SLEEP; |
4204 | 4208 | ||
4205 | for_each_sched_entity(se) { | 4209 | for_each_sched_entity(se) { |
4206 | cfs_rq = cfs_rq_of(se); | 4210 | cfs_rq = cfs_rq_of(se); |
4207 | dequeue_entity(cfs_rq, se, flags); | 4211 | dequeue_entity(cfs_rq, se, flags); |
4208 | 4212 | ||
4209 | /* | 4213 | /* |
4210 | * end evaluation on encountering a throttled cfs_rq | 4214 | * end evaluation on encountering a throttled cfs_rq |
4211 | * | 4215 | * |
4212 | * note: in the case of encountering a throttled cfs_rq we will | 4216 | * note: in the case of encountering a throttled cfs_rq we will |
4213 | * post the final h_nr_running decrement below. | 4217 | * post the final h_nr_running decrement below. |
4214 | */ | 4218 | */ |
4215 | if (cfs_rq_throttled(cfs_rq)) | 4219 | if (cfs_rq_throttled(cfs_rq)) |
4216 | break; | 4220 | break; |
4217 | cfs_rq->h_nr_running--; | 4221 | cfs_rq->h_nr_running--; |
4218 | 4222 | ||
4219 | /* Don't dequeue parent if it has other entities besides us */ | 4223 | /* Don't dequeue parent if it has other entities besides us */ |
4220 | if (cfs_rq->load.weight) { | 4224 | if (cfs_rq->load.weight) { |
4221 | /* | 4225 | /* |
4222 | * Bias pick_next to pick a task from this cfs_rq, as | 4226 | * Bias pick_next to pick a task from this cfs_rq, as |
4223 | * p is sleeping when it is within its sched_slice. | 4227 | * p is sleeping when it is within its sched_slice. |
4224 | */ | 4228 | */ |
4225 | if (task_sleep && parent_entity(se)) | 4229 | if (task_sleep && parent_entity(se)) |
4226 | set_next_buddy(parent_entity(se)); | 4230 | set_next_buddy(parent_entity(se)); |
4227 | 4231 | ||
4228 | /* avoid re-evaluating load for this entity */ | 4232 | /* avoid re-evaluating load for this entity */ |
4229 | se = parent_entity(se); | 4233 | se = parent_entity(se); |
4230 | break; | 4234 | break; |
4231 | } | 4235 | } |
4232 | flags |= DEQUEUE_SLEEP; | 4236 | flags |= DEQUEUE_SLEEP; |
4233 | } | 4237 | } |
4234 | 4238 | ||
4235 | for_each_sched_entity(se) { | 4239 | for_each_sched_entity(se) { |
4236 | cfs_rq = cfs_rq_of(se); | 4240 | cfs_rq = cfs_rq_of(se); |
4237 | cfs_rq->h_nr_running--; | 4241 | cfs_rq->h_nr_running--; |
4238 | 4242 | ||
4239 | if (cfs_rq_throttled(cfs_rq)) | 4243 | if (cfs_rq_throttled(cfs_rq)) |
4240 | break; | 4244 | break; |
4241 | 4245 | ||
4242 | update_cfs_shares(cfs_rq); | 4246 | update_cfs_shares(cfs_rq); |
4243 | update_entity_load_avg(se, 1); | 4247 | update_entity_load_avg(se, 1); |
4244 | } | 4248 | } |
4245 | 4249 | ||
4246 | if (!se) { | 4250 | if (!se) { |
4247 | sub_nr_running(rq, 1); | 4251 | sub_nr_running(rq, 1); |
4248 | update_rq_runnable_avg(rq, 1); | 4252 | update_rq_runnable_avg(rq, 1); |
4249 | } | 4253 | } |
4250 | hrtick_update(rq); | 4254 | hrtick_update(rq); |
4251 | } | 4255 | } |
4252 | 4256 | ||
4253 | #ifdef CONFIG_SMP | 4257 | #ifdef CONFIG_SMP |
4254 | /* Used instead of source_load when we know the type == 0 */ | 4258 | /* Used instead of source_load when we know the type == 0 */ |
4255 | static unsigned long weighted_cpuload(const int cpu) | 4259 | static unsigned long weighted_cpuload(const int cpu) |
4256 | { | 4260 | { |
4257 | return cpu_rq(cpu)->cfs.runnable_load_avg; | 4261 | return cpu_rq(cpu)->cfs.runnable_load_avg; |
4258 | } | 4262 | } |
4259 | 4263 | ||
4260 | /* | 4264 | /* |
4261 | * Return a low guess at the load of a migration-source cpu weighted | 4265 | * Return a low guess at the load of a migration-source cpu weighted |
4262 | * according to the scheduling class and "nice" value. | 4266 | * according to the scheduling class and "nice" value. |
4263 | * | 4267 | * |
4264 | * We want to under-estimate the load of migration sources, to | 4268 | * We want to under-estimate the load of migration sources, to |
4265 | * balance conservatively. | 4269 | * balance conservatively. |
4266 | */ | 4270 | */ |
4267 | static unsigned long source_load(int cpu, int type) | 4271 | static unsigned long source_load(int cpu, int type) |
4268 | { | 4272 | { |
4269 | struct rq *rq = cpu_rq(cpu); | 4273 | struct rq *rq = cpu_rq(cpu); |
4270 | unsigned long total = weighted_cpuload(cpu); | 4274 | unsigned long total = weighted_cpuload(cpu); |
4271 | 4275 | ||
4272 | if (type == 0 || !sched_feat(LB_BIAS)) | 4276 | if (type == 0 || !sched_feat(LB_BIAS)) |
4273 | return total; | 4277 | return total; |
4274 | 4278 | ||
4275 | return min(rq->cpu_load[type-1], total); | 4279 | return min(rq->cpu_load[type-1], total); |
4276 | } | 4280 | } |
4277 | 4281 | ||
4278 | /* | 4282 | /* |
4279 | * Return a high guess at the load of a migration-target cpu weighted | 4283 | * Return a high guess at the load of a migration-target cpu weighted |
4280 | * according to the scheduling class and "nice" value. | 4284 | * according to the scheduling class and "nice" value. |
4281 | */ | 4285 | */ |
4282 | static unsigned long target_load(int cpu, int type) | 4286 | static unsigned long target_load(int cpu, int type) |
4283 | { | 4287 | { |
4284 | struct rq *rq = cpu_rq(cpu); | 4288 | struct rq *rq = cpu_rq(cpu); |
4285 | unsigned long total = weighted_cpuload(cpu); | 4289 | unsigned long total = weighted_cpuload(cpu); |
4286 | 4290 | ||
4287 | if (type == 0 || !sched_feat(LB_BIAS)) | 4291 | if (type == 0 || !sched_feat(LB_BIAS)) |
4288 | return total; | 4292 | return total; |
4289 | 4293 | ||
4290 | return max(rq->cpu_load[type-1], total); | 4294 | return max(rq->cpu_load[type-1], total); |
4291 | } | 4295 | } |
4292 | 4296 | ||
4293 | static unsigned long capacity_of(int cpu) | 4297 | static unsigned long capacity_of(int cpu) |
4294 | { | 4298 | { |
4295 | return cpu_rq(cpu)->cpu_capacity; | 4299 | return cpu_rq(cpu)->cpu_capacity; |
4296 | } | 4300 | } |
4297 | 4301 | ||
4298 | static unsigned long cpu_avg_load_per_task(int cpu) | 4302 | static unsigned long cpu_avg_load_per_task(int cpu) |
4299 | { | 4303 | { |
4300 | struct rq *rq = cpu_rq(cpu); | 4304 | struct rq *rq = cpu_rq(cpu); |
4301 | unsigned long nr_running = ACCESS_ONCE(rq->cfs.h_nr_running); | 4305 | unsigned long nr_running = ACCESS_ONCE(rq->cfs.h_nr_running); |
4302 | unsigned long load_avg = rq->cfs.runnable_load_avg; | 4306 | unsigned long load_avg = rq->cfs.runnable_load_avg; |
4303 | 4307 | ||
4304 | if (nr_running) | 4308 | if (nr_running) |
4305 | return load_avg / nr_running; | 4309 | return load_avg / nr_running; |
4306 | 4310 | ||
4307 | return 0; | 4311 | return 0; |
4308 | } | 4312 | } |
4309 | 4313 | ||
4310 | static void record_wakee(struct task_struct *p) | 4314 | static void record_wakee(struct task_struct *p) |
4311 | { | 4315 | { |
4312 | /* | 4316 | /* |
4313 | * Rough decay (wiping) for cost saving, don't worry | 4317 | * Rough decay (wiping) for cost saving, don't worry |
4314 | * about the boundary, really active task won't care | 4318 | * about the boundary, really active task won't care |
4315 | * about the loss. | 4319 | * about the loss. |
4316 | */ | 4320 | */ |
4317 | if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { | 4321 | if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { |
4318 | current->wakee_flips >>= 1; | 4322 | current->wakee_flips >>= 1; |
4319 | current->wakee_flip_decay_ts = jiffies; | 4323 | current->wakee_flip_decay_ts = jiffies; |
4320 | } | 4324 | } |
4321 | 4325 | ||
4322 | if (current->last_wakee != p) { | 4326 | if (current->last_wakee != p) { |
4323 | current->last_wakee = p; | 4327 | current->last_wakee = p; |
4324 | current->wakee_flips++; | 4328 | current->wakee_flips++; |
4325 | } | 4329 | } |
4326 | } | 4330 | } |
4327 | 4331 | ||
4328 | static void task_waking_fair(struct task_struct *p) | 4332 | static void task_waking_fair(struct task_struct *p) |
4329 | { | 4333 | { |
4330 | struct sched_entity *se = &p->se; | 4334 | struct sched_entity *se = &p->se; |
4331 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | 4335 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
4332 | u64 min_vruntime; | 4336 | u64 min_vruntime; |
4333 | 4337 | ||
4334 | #ifndef CONFIG_64BIT | 4338 | #ifndef CONFIG_64BIT |
4335 | u64 min_vruntime_copy; | 4339 | u64 min_vruntime_copy; |
4336 | 4340 | ||
4337 | do { | 4341 | do { |
4338 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | 4342 | min_vruntime_copy = cfs_rq->min_vruntime_copy; |
4339 | smp_rmb(); | 4343 | smp_rmb(); |
4340 | min_vruntime = cfs_rq->min_vruntime; | 4344 | min_vruntime = cfs_rq->min_vruntime; |
4341 | } while (min_vruntime != min_vruntime_copy); | 4345 | } while (min_vruntime != min_vruntime_copy); |
4342 | #else | 4346 | #else |
4343 | min_vruntime = cfs_rq->min_vruntime; | 4347 | min_vruntime = cfs_rq->min_vruntime; |
4344 | #endif | 4348 | #endif |
4345 | 4349 | ||
4346 | se->vruntime -= min_vruntime; | 4350 | se->vruntime -= min_vruntime; |
4347 | record_wakee(p); | 4351 | record_wakee(p); |
4348 | } | 4352 | } |
4349 | 4353 | ||
4350 | #ifdef CONFIG_FAIR_GROUP_SCHED | 4354 | #ifdef CONFIG_FAIR_GROUP_SCHED |
4351 | /* | 4355 | /* |
4352 | * effective_load() calculates the load change as seen from the root_task_group | 4356 | * effective_load() calculates the load change as seen from the root_task_group |
4353 | * | 4357 | * |
4354 | * Adding load to a group doesn't make a group heavier, but can cause movement | 4358 | * Adding load to a group doesn't make a group heavier, but can cause movement |
4355 | * of group shares between cpus. Assuming the shares were perfectly aligned one | 4359 | * of group shares between cpus. Assuming the shares were perfectly aligned one |
4356 | * can calculate the shift in shares. | 4360 | * can calculate the shift in shares. |
4357 | * | 4361 | * |
4358 | * Calculate the effective load difference if @wl is added (subtracted) to @tg | 4362 | * Calculate the effective load difference if @wl is added (subtracted) to @tg |
4359 | * on this @cpu and results in a total addition (subtraction) of @wg to the | 4363 | * on this @cpu and results in a total addition (subtraction) of @wg to the |
4360 | * total group weight. | 4364 | * total group weight. |
4361 | * | 4365 | * |
4362 | * Given a runqueue weight distribution (rw_i) we can compute a shares | 4366 | * Given a runqueue weight distribution (rw_i) we can compute a shares |
4363 | * distribution (s_i) using: | 4367 | * distribution (s_i) using: |
4364 | * | 4368 | * |
4365 | * s_i = rw_i / \Sum rw_j (1) | 4369 | * s_i = rw_i / \Sum rw_j (1) |
4366 | * | 4370 | * |
4367 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and | 4371 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and |
4368 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting | 4372 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting |
4369 | * shares distribution (s_i): | 4373 | * shares distribution (s_i): |
4370 | * | 4374 | * |
4371 | * rw_i = { 2, 4, 1, 0 } | 4375 | * rw_i = { 2, 4, 1, 0 } |
4372 | * s_i = { 2/7, 4/7, 1/7, 0 } | 4376 | * s_i = { 2/7, 4/7, 1/7, 0 } |
4373 | * | 4377 | * |
4374 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the | 4378 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the |
4375 | * task used to run on and the CPU the waker is running on), we need to | 4379 | * task used to run on and the CPU the waker is running on), we need to |
4376 | * compute the effect of waking a task on either CPU and, in case of a sync | 4380 | * compute the effect of waking a task on either CPU and, in case of a sync |
4377 | * wakeup, compute the effect of the current task going to sleep. | 4381 | * wakeup, compute the effect of the current task going to sleep. |
4378 | * | 4382 | * |
4379 | * So for a change of @wl to the local @cpu with an overall group weight change | 4383 | * So for a change of @wl to the local @cpu with an overall group weight change |
4380 | * of @wl we can compute the new shares distribution (s'_i) using: | 4384 | * of @wl we can compute the new shares distribution (s'_i) using: |
4381 | * | 4385 | * |
4382 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) | 4386 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) |
4383 | * | 4387 | * |
4384 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load | 4388 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load |
4385 | * differences in waking a task to CPU 0. The additional task changes the | 4389 | * differences in waking a task to CPU 0. The additional task changes the |
4386 | * weight and shares distributions like: | 4390 | * weight and shares distributions like: |
4387 | * | 4391 | * |
4388 | * rw'_i = { 3, 4, 1, 0 } | 4392 | * rw'_i = { 3, 4, 1, 0 } |
4389 | * s'_i = { 3/8, 4/8, 1/8, 0 } | 4393 | * s'_i = { 3/8, 4/8, 1/8, 0 } |
4390 | * | 4394 | * |
4391 | * We can then compute the difference in effective weight by using: | 4395 | * We can then compute the difference in effective weight by using: |
4392 | * | 4396 | * |
4393 | * dw_i = S * (s'_i - s_i) (3) | 4397 | * dw_i = S * (s'_i - s_i) (3) |
4394 | * | 4398 | * |
4395 | * Where 'S' is the group weight as seen by its parent. | 4399 | * Where 'S' is the group weight as seen by its parent. |
4396 | * | 4400 | * |
4397 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) | 4401 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) |
4398 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - | 4402 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - |
4399 | * 4/7) times the weight of the group. | 4403 | * 4/7) times the weight of the group. |
4400 | */ | 4404 | */ |
4401 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) | 4405 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
4402 | { | 4406 | { |
4403 | struct sched_entity *se = tg->se[cpu]; | 4407 | struct sched_entity *se = tg->se[cpu]; |
4404 | 4408 | ||
4405 | if (!tg->parent) /* the trivial, non-cgroup case */ | 4409 | if (!tg->parent) /* the trivial, non-cgroup case */ |
4406 | return wl; | 4410 | return wl; |
4407 | 4411 | ||
4408 | for_each_sched_entity(se) { | 4412 | for_each_sched_entity(se) { |
4409 | long w, W; | 4413 | long w, W; |
4410 | 4414 | ||
4411 | tg = se->my_q->tg; | 4415 | tg = se->my_q->tg; |
4412 | 4416 | ||
4413 | /* | 4417 | /* |
4414 | * W = @wg + \Sum rw_j | 4418 | * W = @wg + \Sum rw_j |
4415 | */ | 4419 | */ |
4416 | W = wg + calc_tg_weight(tg, se->my_q); | 4420 | W = wg + calc_tg_weight(tg, se->my_q); |
4417 | 4421 | ||
4418 | /* | 4422 | /* |
4419 | * w = rw_i + @wl | 4423 | * w = rw_i + @wl |
4420 | */ | 4424 | */ |
4421 | w = se->my_q->load.weight + wl; | 4425 | w = se->my_q->load.weight + wl; |
4422 | 4426 | ||
4423 | /* | 4427 | /* |
4424 | * wl = S * s'_i; see (2) | 4428 | * wl = S * s'_i; see (2) |
4425 | */ | 4429 | */ |
4426 | if (W > 0 && w < W) | 4430 | if (W > 0 && w < W) |
4427 | wl = (w * tg->shares) / W; | 4431 | wl = (w * (long)tg->shares) / W; |
4428 | else | 4432 | else |
4429 | wl = tg->shares; | 4433 | wl = tg->shares; |
4430 | 4434 | ||
4431 | /* | 4435 | /* |
4432 | * Per the above, wl is the new se->load.weight value; since | 4436 | * Per the above, wl is the new se->load.weight value; since |
4433 | * those are clipped to [MIN_SHARES, ...) do so now. See | 4437 | * those are clipped to [MIN_SHARES, ...) do so now. See |
4434 | * calc_cfs_shares(). | 4438 | * calc_cfs_shares(). |
4435 | */ | 4439 | */ |
4436 | if (wl < MIN_SHARES) | 4440 | if (wl < MIN_SHARES) |
4437 | wl = MIN_SHARES; | 4441 | wl = MIN_SHARES; |
4438 | 4442 | ||
4439 | /* | 4443 | /* |
4440 | * wl = dw_i = S * (s'_i - s_i); see (3) | 4444 | * wl = dw_i = S * (s'_i - s_i); see (3) |
4441 | */ | 4445 | */ |
4442 | wl -= se->load.weight; | 4446 | wl -= se->load.weight; |
4443 | 4447 | ||
4444 | /* | 4448 | /* |
4445 | * Recursively apply this logic to all parent groups to compute | 4449 | * Recursively apply this logic to all parent groups to compute |
4446 | * the final effective load change on the root group. Since | 4450 | * the final effective load change on the root group. Since |
4447 | * only the @tg group gets extra weight, all parent groups can | 4451 | * only the @tg group gets extra weight, all parent groups can |
4448 | * only redistribute existing shares. @wl is the shift in shares | 4452 | * only redistribute existing shares. @wl is the shift in shares |
4449 | * resulting from this level per the above. | 4453 | * resulting from this level per the above. |
4450 | */ | 4454 | */ |
4451 | wg = 0; | 4455 | wg = 0; |
4452 | } | 4456 | } |
4453 | 4457 | ||
4454 | return wl; | 4458 | return wl; |
4455 | } | 4459 | } |
4456 | #else | 4460 | #else |
4457 | 4461 | ||
4458 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) | 4462 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
4459 | { | 4463 | { |
4460 | return wl; | 4464 | return wl; |
4461 | } | 4465 | } |
4462 | 4466 | ||
4463 | #endif | 4467 | #endif |
4464 | 4468 | ||
4465 | static int wake_wide(struct task_struct *p) | 4469 | static int wake_wide(struct task_struct *p) |
4466 | { | 4470 | { |
4467 | int factor = this_cpu_read(sd_llc_size); | 4471 | int factor = this_cpu_read(sd_llc_size); |
4468 | 4472 | ||
4469 | /* | 4473 | /* |
4470 | * Yeah, it's the switching-frequency, could means many wakee or | 4474 | * Yeah, it's the switching-frequency, could means many wakee or |
4471 | * rapidly switch, use factor here will just help to automatically | 4475 | * rapidly switch, use factor here will just help to automatically |
4472 | * adjust the loose-degree, so bigger node will lead to more pull. | 4476 | * adjust the loose-degree, so bigger node will lead to more pull. |
4473 | */ | 4477 | */ |
4474 | if (p->wakee_flips > factor) { | 4478 | if (p->wakee_flips > factor) { |
4475 | /* | 4479 | /* |
4476 | * wakee is somewhat hot, it needs certain amount of cpu | 4480 | * wakee is somewhat hot, it needs certain amount of cpu |
4477 | * resource, so if waker is far more hot, prefer to leave | 4481 | * resource, so if waker is far more hot, prefer to leave |
4478 | * it alone. | 4482 | * it alone. |
4479 | */ | 4483 | */ |
4480 | if (current->wakee_flips > (factor * p->wakee_flips)) | 4484 | if (current->wakee_flips > (factor * p->wakee_flips)) |
4481 | return 1; | 4485 | return 1; |
4482 | } | 4486 | } |
4483 | 4487 | ||
4484 | return 0; | 4488 | return 0; |
4485 | } | 4489 | } |
4486 | 4490 | ||
4487 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) | 4491 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) |
4488 | { | 4492 | { |
4489 | s64 this_load, load; | 4493 | s64 this_load, load; |
4490 | s64 this_eff_load, prev_eff_load; | 4494 | s64 this_eff_load, prev_eff_load; |
4491 | int idx, this_cpu, prev_cpu; | 4495 | int idx, this_cpu, prev_cpu; |
4492 | struct task_group *tg; | 4496 | struct task_group *tg; |
4493 | unsigned long weight; | 4497 | unsigned long weight; |
4494 | int balanced; | 4498 | int balanced; |
4495 | 4499 | ||
4496 | /* | 4500 | /* |
4497 | * If we wake multiple tasks be careful to not bounce | 4501 | * If we wake multiple tasks be careful to not bounce |
4498 | * ourselves around too much. | 4502 | * ourselves around too much. |
4499 | */ | 4503 | */ |
4500 | if (wake_wide(p)) | 4504 | if (wake_wide(p)) |
4501 | return 0; | 4505 | return 0; |
4502 | 4506 | ||
4503 | idx = sd->wake_idx; | 4507 | idx = sd->wake_idx; |
4504 | this_cpu = smp_processor_id(); | 4508 | this_cpu = smp_processor_id(); |
4505 | prev_cpu = task_cpu(p); | 4509 | prev_cpu = task_cpu(p); |
4506 | load = source_load(prev_cpu, idx); | 4510 | load = source_load(prev_cpu, idx); |
4507 | this_load = target_load(this_cpu, idx); | 4511 | this_load = target_load(this_cpu, idx); |
4508 | 4512 | ||
4509 | /* | 4513 | /* |
4510 | * If sync wakeup then subtract the (maximum possible) | 4514 | * If sync wakeup then subtract the (maximum possible) |
4511 | * effect of the currently running task from the load | 4515 | * effect of the currently running task from the load |
4512 | * of the current CPU: | 4516 | * of the current CPU: |
4513 | */ | 4517 | */ |
4514 | if (sync) { | 4518 | if (sync) { |
4515 | tg = task_group(current); | 4519 | tg = task_group(current); |
4516 | weight = current->se.load.weight; | 4520 | weight = current->se.load.weight; |
4517 | 4521 | ||
4518 | this_load += effective_load(tg, this_cpu, -weight, -weight); | 4522 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
4519 | load += effective_load(tg, prev_cpu, 0, -weight); | 4523 | load += effective_load(tg, prev_cpu, 0, -weight); |
4520 | } | 4524 | } |
4521 | 4525 | ||
4522 | tg = task_group(p); | 4526 | tg = task_group(p); |
4523 | weight = p->se.load.weight; | 4527 | weight = p->se.load.weight; |
4524 | 4528 | ||
4525 | /* | 4529 | /* |
4526 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | 4530 | * In low-load situations, where prev_cpu is idle and this_cpu is idle |
4527 | * due to the sync cause above having dropped this_load to 0, we'll | 4531 | * due to the sync cause above having dropped this_load to 0, we'll |
4528 | * always have an imbalance, but there's really nothing you can do | 4532 | * always have an imbalance, but there's really nothing you can do |
4529 | * about that, so that's good too. | 4533 | * about that, so that's good too. |
4530 | * | 4534 | * |
4531 | * Otherwise check if either cpus are near enough in load to allow this | 4535 | * Otherwise check if either cpus are near enough in load to allow this |
4532 | * task to be woken on this_cpu. | 4536 | * task to be woken on this_cpu. |
4533 | */ | 4537 | */ |
4534 | this_eff_load = 100; | 4538 | this_eff_load = 100; |
4535 | this_eff_load *= capacity_of(prev_cpu); | 4539 | this_eff_load *= capacity_of(prev_cpu); |
4536 | 4540 | ||
4537 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; | 4541 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; |
4538 | prev_eff_load *= capacity_of(this_cpu); | 4542 | prev_eff_load *= capacity_of(this_cpu); |
4539 | 4543 | ||
4540 | if (this_load > 0) { | 4544 | if (this_load > 0) { |
4541 | this_eff_load *= this_load + | 4545 | this_eff_load *= this_load + |
4542 | effective_load(tg, this_cpu, weight, weight); | 4546 | effective_load(tg, this_cpu, weight, weight); |
4543 | 4547 | ||
4544 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); | 4548 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); |
4545 | } | 4549 | } |
4546 | 4550 | ||
4547 | balanced = this_eff_load <= prev_eff_load; | 4551 | balanced = this_eff_load <= prev_eff_load; |
4548 | 4552 | ||
4549 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); | 4553 | schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); |
4550 | 4554 | ||
4551 | if (!balanced) | 4555 | if (!balanced) |
4552 | return 0; | 4556 | return 0; |
4553 | 4557 | ||
4554 | schedstat_inc(sd, ttwu_move_affine); | 4558 | schedstat_inc(sd, ttwu_move_affine); |
4555 | schedstat_inc(p, se.statistics.nr_wakeups_affine); | 4559 | schedstat_inc(p, se.statistics.nr_wakeups_affine); |
4556 | 4560 | ||
4557 | return 1; | 4561 | return 1; |
4558 | } | 4562 | } |
4559 | 4563 | ||
4560 | /* | 4564 | /* |
4561 | * find_idlest_group finds and returns the least busy CPU group within the | 4565 | * find_idlest_group finds and returns the least busy CPU group within the |
4562 | * domain. | 4566 | * domain. |
4563 | */ | 4567 | */ |
4564 | static struct sched_group * | 4568 | static struct sched_group * |
4565 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, | 4569 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
4566 | int this_cpu, int sd_flag) | 4570 | int this_cpu, int sd_flag) |
4567 | { | 4571 | { |
4568 | struct sched_group *idlest = NULL, *group = sd->groups; | 4572 | struct sched_group *idlest = NULL, *group = sd->groups; |
4569 | unsigned long min_load = ULONG_MAX, this_load = 0; | 4573 | unsigned long min_load = ULONG_MAX, this_load = 0; |
4570 | int load_idx = sd->forkexec_idx; | 4574 | int load_idx = sd->forkexec_idx; |
4571 | int imbalance = 100 + (sd->imbalance_pct-100)/2; | 4575 | int imbalance = 100 + (sd->imbalance_pct-100)/2; |
4572 | 4576 | ||
4573 | if (sd_flag & SD_BALANCE_WAKE) | 4577 | if (sd_flag & SD_BALANCE_WAKE) |
4574 | load_idx = sd->wake_idx; | 4578 | load_idx = sd->wake_idx; |
4575 | 4579 | ||
4576 | do { | 4580 | do { |
4577 | unsigned long load, avg_load; | 4581 | unsigned long load, avg_load; |
4578 | int local_group; | 4582 | int local_group; |
4579 | int i; | 4583 | int i; |
4580 | 4584 | ||
4581 | /* Skip over this group if it has no CPUs allowed */ | 4585 | /* Skip over this group if it has no CPUs allowed */ |
4582 | if (!cpumask_intersects(sched_group_cpus(group), | 4586 | if (!cpumask_intersects(sched_group_cpus(group), |
4583 | tsk_cpus_allowed(p))) | 4587 | tsk_cpus_allowed(p))) |
4584 | continue; | 4588 | continue; |
4585 | 4589 | ||
4586 | local_group = cpumask_test_cpu(this_cpu, | 4590 | local_group = cpumask_test_cpu(this_cpu, |
4587 | sched_group_cpus(group)); | 4591 | sched_group_cpus(group)); |
4588 | 4592 | ||
4589 | /* Tally up the load of all CPUs in the group */ | 4593 | /* Tally up the load of all CPUs in the group */ |
4590 | avg_load = 0; | 4594 | avg_load = 0; |
4591 | 4595 | ||
4592 | for_each_cpu(i, sched_group_cpus(group)) { | 4596 | for_each_cpu(i, sched_group_cpus(group)) { |
4593 | /* Bias balancing toward cpus of our domain */ | 4597 | /* Bias balancing toward cpus of our domain */ |
4594 | if (local_group) | 4598 | if (local_group) |
4595 | load = source_load(i, load_idx); | 4599 | load = source_load(i, load_idx); |
4596 | else | 4600 | else |
4597 | load = target_load(i, load_idx); | 4601 | load = target_load(i, load_idx); |
4598 | 4602 | ||
4599 | avg_load += load; | 4603 | avg_load += load; |
4600 | } | 4604 | } |
4601 | 4605 | ||
4602 | /* Adjust by relative CPU capacity of the group */ | 4606 | /* Adjust by relative CPU capacity of the group */ |
4603 | avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity; | 4607 | avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity; |
4604 | 4608 | ||
4605 | if (local_group) { | 4609 | if (local_group) { |
4606 | this_load = avg_load; | 4610 | this_load = avg_load; |
4607 | } else if (avg_load < min_load) { | 4611 | } else if (avg_load < min_load) { |
4608 | min_load = avg_load; | 4612 | min_load = avg_load; |
4609 | idlest = group; | 4613 | idlest = group; |
4610 | } | 4614 | } |
4611 | } while (group = group->next, group != sd->groups); | 4615 | } while (group = group->next, group != sd->groups); |
4612 | 4616 | ||
4613 | if (!idlest || 100*this_load < imbalance*min_load) | 4617 | if (!idlest || 100*this_load < imbalance*min_load) |
4614 | return NULL; | 4618 | return NULL; |
4615 | return idlest; | 4619 | return idlest; |
4616 | } | 4620 | } |
4617 | 4621 | ||
4618 | /* | 4622 | /* |
4619 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | 4623 | * find_idlest_cpu - find the idlest cpu among the cpus in group. |
4620 | */ | 4624 | */ |
4621 | static int | 4625 | static int |
4622 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | 4626 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) |
4623 | { | 4627 | { |
4624 | unsigned long load, min_load = ULONG_MAX; | 4628 | unsigned long load, min_load = ULONG_MAX; |
4625 | unsigned int min_exit_latency = UINT_MAX; | 4629 | unsigned int min_exit_latency = UINT_MAX; |
4626 | u64 latest_idle_timestamp = 0; | 4630 | u64 latest_idle_timestamp = 0; |
4627 | int least_loaded_cpu = this_cpu; | 4631 | int least_loaded_cpu = this_cpu; |
4628 | int shallowest_idle_cpu = -1; | 4632 | int shallowest_idle_cpu = -1; |
4629 | int i; | 4633 | int i; |
4630 | 4634 | ||
4631 | /* Traverse only the allowed CPUs */ | 4635 | /* Traverse only the allowed CPUs */ |
4632 | for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { | 4636 | for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { |
4633 | if (idle_cpu(i)) { | 4637 | if (idle_cpu(i)) { |
4634 | struct rq *rq = cpu_rq(i); | 4638 | struct rq *rq = cpu_rq(i); |
4635 | struct cpuidle_state *idle = idle_get_state(rq); | 4639 | struct cpuidle_state *idle = idle_get_state(rq); |
4636 | if (idle && idle->exit_latency < min_exit_latency) { | 4640 | if (idle && idle->exit_latency < min_exit_latency) { |
4637 | /* | 4641 | /* |
4638 | * We give priority to a CPU whose idle state | 4642 | * We give priority to a CPU whose idle state |
4639 | * has the smallest exit latency irrespective | 4643 | * has the smallest exit latency irrespective |
4640 | * of any idle timestamp. | 4644 | * of any idle timestamp. |
4641 | */ | 4645 | */ |
4642 | min_exit_latency = idle->exit_latency; | 4646 | min_exit_latency = idle->exit_latency; |
4643 | latest_idle_timestamp = rq->idle_stamp; | 4647 | latest_idle_timestamp = rq->idle_stamp; |
4644 | shallowest_idle_cpu = i; | 4648 | shallowest_idle_cpu = i; |
4645 | } else if ((!idle || idle->exit_latency == min_exit_latency) && | 4649 | } else if ((!idle || idle->exit_latency == min_exit_latency) && |
4646 | rq->idle_stamp > latest_idle_timestamp) { | 4650 | rq->idle_stamp > latest_idle_timestamp) { |
4647 | /* | 4651 | /* |
4648 | * If equal or no active idle state, then | 4652 | * If equal or no active idle state, then |
4649 | * the most recently idled CPU might have | 4653 | * the most recently idled CPU might have |
4650 | * a warmer cache. | 4654 | * a warmer cache. |
4651 | */ | 4655 | */ |
4652 | latest_idle_timestamp = rq->idle_stamp; | 4656 | latest_idle_timestamp = rq->idle_stamp; |
4653 | shallowest_idle_cpu = i; | 4657 | shallowest_idle_cpu = i; |
4654 | } | 4658 | } |
4655 | } else if (shallowest_idle_cpu == -1) { | 4659 | } else if (shallowest_idle_cpu == -1) { |
4656 | load = weighted_cpuload(i); | 4660 | load = weighted_cpuload(i); |
4657 | if (load < min_load || (load == min_load && i == this_cpu)) { | 4661 | if (load < min_load || (load == min_load && i == this_cpu)) { |
4658 | min_load = load; | 4662 | min_load = load; |
4659 | least_loaded_cpu = i; | 4663 | least_loaded_cpu = i; |
4660 | } | 4664 | } |
4661 | } | 4665 | } |
4662 | } | 4666 | } |
4663 | 4667 | ||
4664 | return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; | 4668 | return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; |
4665 | } | 4669 | } |
4666 | 4670 | ||
4667 | /* | 4671 | /* |
4668 | * Try and locate an idle CPU in the sched_domain. | 4672 | * Try and locate an idle CPU in the sched_domain. |
4669 | */ | 4673 | */ |
4670 | static int select_idle_sibling(struct task_struct *p, int target) | 4674 | static int select_idle_sibling(struct task_struct *p, int target) |
4671 | { | 4675 | { |
4672 | struct sched_domain *sd; | 4676 | struct sched_domain *sd; |
4673 | struct sched_group *sg; | 4677 | struct sched_group *sg; |
4674 | int i = task_cpu(p); | 4678 | int i = task_cpu(p); |
4675 | 4679 | ||
4676 | if (idle_cpu(target)) | 4680 | if (idle_cpu(target)) |
4677 | return target; | 4681 | return target; |
4678 | 4682 | ||
4679 | /* | 4683 | /* |
4680 | * If the prevous cpu is cache affine and idle, don't be stupid. | 4684 | * If the prevous cpu is cache affine and idle, don't be stupid. |
4681 | */ | 4685 | */ |
4682 | if (i != target && cpus_share_cache(i, target) && idle_cpu(i)) | 4686 | if (i != target && cpus_share_cache(i, target) && idle_cpu(i)) |
4683 | return i; | 4687 | return i; |
4684 | 4688 | ||
4685 | /* | 4689 | /* |
4686 | * Otherwise, iterate the domains and find an elegible idle cpu. | 4690 | * Otherwise, iterate the domains and find an elegible idle cpu. |
4687 | */ | 4691 | */ |
4688 | sd = rcu_dereference(per_cpu(sd_llc, target)); | 4692 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
4689 | for_each_lower_domain(sd) { | 4693 | for_each_lower_domain(sd) { |
4690 | sg = sd->groups; | 4694 | sg = sd->groups; |
4691 | do { | 4695 | do { |
4692 | if (!cpumask_intersects(sched_group_cpus(sg), | 4696 | if (!cpumask_intersects(sched_group_cpus(sg), |
4693 | tsk_cpus_allowed(p))) | 4697 | tsk_cpus_allowed(p))) |
4694 | goto next; | 4698 | goto next; |
4695 | 4699 | ||
4696 | for_each_cpu(i, sched_group_cpus(sg)) { | 4700 | for_each_cpu(i, sched_group_cpus(sg)) { |
4697 | if (i == target || !idle_cpu(i)) | 4701 | if (i == target || !idle_cpu(i)) |
4698 | goto next; | 4702 | goto next; |
4699 | } | 4703 | } |
4700 | 4704 | ||
4701 | target = cpumask_first_and(sched_group_cpus(sg), | 4705 | target = cpumask_first_and(sched_group_cpus(sg), |
4702 | tsk_cpus_allowed(p)); | 4706 | tsk_cpus_allowed(p)); |
4703 | goto done; | 4707 | goto done; |
4704 | next: | 4708 | next: |
4705 | sg = sg->next; | 4709 | sg = sg->next; |
4706 | } while (sg != sd->groups); | 4710 | } while (sg != sd->groups); |
4707 | } | 4711 | } |
4708 | done: | 4712 | done: |
4709 | return target; | 4713 | return target; |
4710 | } | 4714 | } |
4711 | 4715 | ||
4712 | /* | 4716 | /* |
4713 | * select_task_rq_fair: Select target runqueue for the waking task in domains | 4717 | * select_task_rq_fair: Select target runqueue for the waking task in domains |
4714 | * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE, | 4718 | * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE, |
4715 | * SD_BALANCE_FORK, or SD_BALANCE_EXEC. | 4719 | * SD_BALANCE_FORK, or SD_BALANCE_EXEC. |
4716 | * | 4720 | * |
4717 | * Balances load by selecting the idlest cpu in the idlest group, or under | 4721 | * Balances load by selecting the idlest cpu in the idlest group, or under |
4718 | * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set. | 4722 | * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set. |
4719 | * | 4723 | * |
4720 | * Returns the target cpu number. | 4724 | * Returns the target cpu number. |
4721 | * | 4725 | * |
4722 | * preempt must be disabled. | 4726 | * preempt must be disabled. |
4723 | */ | 4727 | */ |
4724 | static int | 4728 | static int |
4725 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) | 4729 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) |
4726 | { | 4730 | { |
4727 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; | 4731 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
4728 | int cpu = smp_processor_id(); | 4732 | int cpu = smp_processor_id(); |
4729 | int new_cpu = cpu; | 4733 | int new_cpu = cpu; |
4730 | int want_affine = 0; | 4734 | int want_affine = 0; |
4731 | int sync = wake_flags & WF_SYNC; | 4735 | int sync = wake_flags & WF_SYNC; |
4732 | 4736 | ||
4733 | if (sd_flag & SD_BALANCE_WAKE) | 4737 | if (sd_flag & SD_BALANCE_WAKE) |
4734 | want_affine = cpumask_test_cpu(cpu, tsk_cpus_allowed(p)); | 4738 | want_affine = cpumask_test_cpu(cpu, tsk_cpus_allowed(p)); |
4735 | 4739 | ||
4736 | rcu_read_lock(); | 4740 | rcu_read_lock(); |
4737 | for_each_domain(cpu, tmp) { | 4741 | for_each_domain(cpu, tmp) { |
4738 | if (!(tmp->flags & SD_LOAD_BALANCE)) | 4742 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
4739 | continue; | 4743 | continue; |
4740 | 4744 | ||
4741 | /* | 4745 | /* |
4742 | * If both cpu and prev_cpu are part of this domain, | 4746 | * If both cpu and prev_cpu are part of this domain, |
4743 | * cpu is a valid SD_WAKE_AFFINE target. | 4747 | * cpu is a valid SD_WAKE_AFFINE target. |
4744 | */ | 4748 | */ |
4745 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && | 4749 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
4746 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | 4750 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { |
4747 | affine_sd = tmp; | 4751 | affine_sd = tmp; |
4748 | break; | 4752 | break; |
4749 | } | 4753 | } |
4750 | 4754 | ||
4751 | if (tmp->flags & sd_flag) | 4755 | if (tmp->flags & sd_flag) |
4752 | sd = tmp; | 4756 | sd = tmp; |
4753 | } | 4757 | } |
4754 | 4758 | ||
4755 | if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync)) | 4759 | if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync)) |
4756 | prev_cpu = cpu; | 4760 | prev_cpu = cpu; |
4757 | 4761 | ||
4758 | if (sd_flag & SD_BALANCE_WAKE) { | 4762 | if (sd_flag & SD_BALANCE_WAKE) { |
4759 | new_cpu = select_idle_sibling(p, prev_cpu); | 4763 | new_cpu = select_idle_sibling(p, prev_cpu); |
4760 | goto unlock; | 4764 | goto unlock; |
4761 | } | 4765 | } |
4762 | 4766 | ||
4763 | while (sd) { | 4767 | while (sd) { |
4764 | struct sched_group *group; | 4768 | struct sched_group *group; |
4765 | int weight; | 4769 | int weight; |
4766 | 4770 | ||
4767 | if (!(sd->flags & sd_flag)) { | 4771 | if (!(sd->flags & sd_flag)) { |
4768 | sd = sd->child; | 4772 | sd = sd->child; |
4769 | continue; | 4773 | continue; |
4770 | } | 4774 | } |
4771 | 4775 | ||
4772 | group = find_idlest_group(sd, p, cpu, sd_flag); | 4776 | group = find_idlest_group(sd, p, cpu, sd_flag); |
4773 | if (!group) { | 4777 | if (!group) { |
4774 | sd = sd->child; | 4778 | sd = sd->child; |
4775 | continue; | 4779 | continue; |
4776 | } | 4780 | } |
4777 | 4781 | ||
4778 | new_cpu = find_idlest_cpu(group, p, cpu); | 4782 | new_cpu = find_idlest_cpu(group, p, cpu); |
4779 | if (new_cpu == -1 || new_cpu == cpu) { | 4783 | if (new_cpu == -1 || new_cpu == cpu) { |
4780 | /* Now try balancing at a lower domain level of cpu */ | 4784 | /* Now try balancing at a lower domain level of cpu */ |
4781 | sd = sd->child; | 4785 | sd = sd->child; |
4782 | continue; | 4786 | continue; |
4783 | } | 4787 | } |
4784 | 4788 | ||
4785 | /* Now try balancing at a lower domain level of new_cpu */ | 4789 | /* Now try balancing at a lower domain level of new_cpu */ |
4786 | cpu = new_cpu; | 4790 | cpu = new_cpu; |
4787 | weight = sd->span_weight; | 4791 | weight = sd->span_weight; |
4788 | sd = NULL; | 4792 | sd = NULL; |
4789 | for_each_domain(cpu, tmp) { | 4793 | for_each_domain(cpu, tmp) { |
4790 | if (weight <= tmp->span_weight) | 4794 | if (weight <= tmp->span_weight) |
4791 | break; | 4795 | break; |
4792 | if (tmp->flags & sd_flag) | 4796 | if (tmp->flags & sd_flag) |
4793 | sd = tmp; | 4797 | sd = tmp; |
4794 | } | 4798 | } |
4795 | /* while loop will break here if sd == NULL */ | 4799 | /* while loop will break here if sd == NULL */ |
4796 | } | 4800 | } |
4797 | unlock: | 4801 | unlock: |
4798 | rcu_read_unlock(); | 4802 | rcu_read_unlock(); |
4799 | 4803 | ||
4800 | return new_cpu; | 4804 | return new_cpu; |
4801 | } | 4805 | } |
4802 | 4806 | ||
4803 | /* | 4807 | /* |
4804 | * Called immediately before a task is migrated to a new cpu; task_cpu(p) and | 4808 | * Called immediately before a task is migrated to a new cpu; task_cpu(p) and |
4805 | * cfs_rq_of(p) references at time of call are still valid and identify the | 4809 | * cfs_rq_of(p) references at time of call are still valid and identify the |
4806 | * previous cpu. However, the caller only guarantees p->pi_lock is held; no | 4810 | * previous cpu. However, the caller only guarantees p->pi_lock is held; no |
4807 | * other assumptions, including the state of rq->lock, should be made. | 4811 | * other assumptions, including the state of rq->lock, should be made. |
4808 | */ | 4812 | */ |
4809 | static void | 4813 | static void |
4810 | migrate_task_rq_fair(struct task_struct *p, int next_cpu) | 4814 | migrate_task_rq_fair(struct task_struct *p, int next_cpu) |
4811 | { | 4815 | { |
4812 | struct sched_entity *se = &p->se; | 4816 | struct sched_entity *se = &p->se; |
4813 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | 4817 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
4814 | 4818 | ||
4815 | /* | 4819 | /* |
4816 | * Load tracking: accumulate removed load so that it can be processed | 4820 | * Load tracking: accumulate removed load so that it can be processed |
4817 | * when we next update owning cfs_rq under rq->lock. Tasks contribute | 4821 | * when we next update owning cfs_rq under rq->lock. Tasks contribute |
4818 | * to blocked load iff they have a positive decay-count. It can never | 4822 | * to blocked load iff they have a positive decay-count. It can never |
4819 | * be negative here since on-rq tasks have decay-count == 0. | 4823 | * be negative here since on-rq tasks have decay-count == 0. |
4820 | */ | 4824 | */ |
4821 | if (se->avg.decay_count) { | 4825 | if (se->avg.decay_count) { |
4822 | se->avg.decay_count = -__synchronize_entity_decay(se); | 4826 | se->avg.decay_count = -__synchronize_entity_decay(se); |
4823 | atomic_long_add(se->avg.load_avg_contrib, | 4827 | atomic_long_add(se->avg.load_avg_contrib, |
4824 | &cfs_rq->removed_load); | 4828 | &cfs_rq->removed_load); |
4825 | } | 4829 | } |
4826 | 4830 | ||
4827 | /* We have migrated, no longer consider this task hot */ | 4831 | /* We have migrated, no longer consider this task hot */ |
4828 | se->exec_start = 0; | 4832 | se->exec_start = 0; |
4829 | } | 4833 | } |
4830 | #endif /* CONFIG_SMP */ | 4834 | #endif /* CONFIG_SMP */ |
4831 | 4835 | ||
4832 | static unsigned long | 4836 | static unsigned long |
4833 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | 4837 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) |
4834 | { | 4838 | { |
4835 | unsigned long gran = sysctl_sched_wakeup_granularity; | 4839 | unsigned long gran = sysctl_sched_wakeup_granularity; |
4836 | 4840 | ||
4837 | /* | 4841 | /* |
4838 | * Since its curr running now, convert the gran from real-time | 4842 | * Since its curr running now, convert the gran from real-time |
4839 | * to virtual-time in his units. | 4843 | * to virtual-time in his units. |
4840 | * | 4844 | * |
4841 | * By using 'se' instead of 'curr' we penalize light tasks, so | 4845 | * By using 'se' instead of 'curr' we penalize light tasks, so |
4842 | * they get preempted easier. That is, if 'se' < 'curr' then | 4846 | * they get preempted easier. That is, if 'se' < 'curr' then |
4843 | * the resulting gran will be larger, therefore penalizing the | 4847 | * the resulting gran will be larger, therefore penalizing the |
4844 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | 4848 | * lighter, if otoh 'se' > 'curr' then the resulting gran will |
4845 | * be smaller, again penalizing the lighter task. | 4849 | * be smaller, again penalizing the lighter task. |
4846 | * | 4850 | * |
4847 | * This is especially important for buddies when the leftmost | 4851 | * This is especially important for buddies when the leftmost |
4848 | * task is higher priority than the buddy. | 4852 | * task is higher priority than the buddy. |
4849 | */ | 4853 | */ |
4850 | return calc_delta_fair(gran, se); | 4854 | return calc_delta_fair(gran, se); |
4851 | } | 4855 | } |
4852 | 4856 | ||
4853 | /* | 4857 | /* |
4854 | * Should 'se' preempt 'curr'. | 4858 | * Should 'se' preempt 'curr'. |
4855 | * | 4859 | * |
4856 | * |s1 | 4860 | * |s1 |
4857 | * |s2 | 4861 | * |s2 |
4858 | * |s3 | 4862 | * |s3 |
4859 | * g | 4863 | * g |
4860 | * |<--->|c | 4864 | * |<--->|c |
4861 | * | 4865 | * |
4862 | * w(c, s1) = -1 | 4866 | * w(c, s1) = -1 |
4863 | * w(c, s2) = 0 | 4867 | * w(c, s2) = 0 |
4864 | * w(c, s3) = 1 | 4868 | * w(c, s3) = 1 |
4865 | * | 4869 | * |
4866 | */ | 4870 | */ |
4867 | static int | 4871 | static int |
4868 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | 4872 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) |
4869 | { | 4873 | { |
4870 | s64 gran, vdiff = curr->vruntime - se->vruntime; | 4874 | s64 gran, vdiff = curr->vruntime - se->vruntime; |
4871 | 4875 | ||
4872 | if (vdiff <= 0) | 4876 | if (vdiff <= 0) |
4873 | return -1; | 4877 | return -1; |
4874 | 4878 | ||
4875 | gran = wakeup_gran(curr, se); | 4879 | gran = wakeup_gran(curr, se); |
4876 | if (vdiff > gran) | 4880 | if (vdiff > gran) |
4877 | return 1; | 4881 | return 1; |
4878 | 4882 | ||
4879 | return 0; | 4883 | return 0; |
4880 | } | 4884 | } |
4881 | 4885 | ||
4882 | static void set_last_buddy(struct sched_entity *se) | 4886 | static void set_last_buddy(struct sched_entity *se) |
4883 | { | 4887 | { |
4884 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) | 4888 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
4885 | return; | 4889 | return; |
4886 | 4890 | ||
4887 | for_each_sched_entity(se) | 4891 | for_each_sched_entity(se) |
4888 | cfs_rq_of(se)->last = se; | 4892 | cfs_rq_of(se)->last = se; |
4889 | } | 4893 | } |
4890 | 4894 | ||
4891 | static void set_next_buddy(struct sched_entity *se) | 4895 | static void set_next_buddy(struct sched_entity *se) |
4892 | { | 4896 | { |
4893 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) | 4897 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
4894 | return; | 4898 | return; |
4895 | 4899 | ||
4896 | for_each_sched_entity(se) | 4900 | for_each_sched_entity(se) |
4897 | cfs_rq_of(se)->next = se; | 4901 | cfs_rq_of(se)->next = se; |
4898 | } | 4902 | } |
4899 | 4903 | ||
4900 | static void set_skip_buddy(struct sched_entity *se) | 4904 | static void set_skip_buddy(struct sched_entity *se) |
4901 | { | 4905 | { |
4902 | for_each_sched_entity(se) | 4906 | for_each_sched_entity(se) |
4903 | cfs_rq_of(se)->skip = se; | 4907 | cfs_rq_of(se)->skip = se; |
4904 | } | 4908 | } |
4905 | 4909 | ||
4906 | /* | 4910 | /* |
4907 | * Preempt the current task with a newly woken task if needed: | 4911 | * Preempt the current task with a newly woken task if needed: |
4908 | */ | 4912 | */ |
4909 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) | 4913 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
4910 | { | 4914 | { |
4911 | struct task_struct *curr = rq->curr; | 4915 | struct task_struct *curr = rq->curr; |
4912 | struct sched_entity *se = &curr->se, *pse = &p->se; | 4916 | struct sched_entity *se = &curr->se, *pse = &p->se; |
4913 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | 4917 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
4914 | int scale = cfs_rq->nr_running >= sched_nr_latency; | 4918 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
4915 | int next_buddy_marked = 0; | 4919 | int next_buddy_marked = 0; |
4916 | 4920 | ||
4917 | if (unlikely(se == pse)) | 4921 | if (unlikely(se == pse)) |
4918 | return; | 4922 | return; |
4919 | 4923 | ||
4920 | /* | 4924 | /* |
4921 | * This is possible from callers such as attach_tasks(), in which we | 4925 | * This is possible from callers such as attach_tasks(), in which we |
4922 | * unconditionally check_prempt_curr() after an enqueue (which may have | 4926 | * unconditionally check_prempt_curr() after an enqueue (which may have |
4923 | * lead to a throttle). This both saves work and prevents false | 4927 | * lead to a throttle). This both saves work and prevents false |
4924 | * next-buddy nomination below. | 4928 | * next-buddy nomination below. |
4925 | */ | 4929 | */ |
4926 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | 4930 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) |
4927 | return; | 4931 | return; |
4928 | 4932 | ||
4929 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { | 4933 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
4930 | set_next_buddy(pse); | 4934 | set_next_buddy(pse); |
4931 | next_buddy_marked = 1; | 4935 | next_buddy_marked = 1; |
4932 | } | 4936 | } |
4933 | 4937 | ||
4934 | /* | 4938 | /* |
4935 | * We can come here with TIF_NEED_RESCHED already set from new task | 4939 | * We can come here with TIF_NEED_RESCHED already set from new task |
4936 | * wake up path. | 4940 | * wake up path. |
4937 | * | 4941 | * |
4938 | * Note: this also catches the edge-case of curr being in a throttled | 4942 | * Note: this also catches the edge-case of curr being in a throttled |
4939 | * group (e.g. via set_curr_task), since update_curr() (in the | 4943 | * group (e.g. via set_curr_task), since update_curr() (in the |
4940 | * enqueue of curr) will have resulted in resched being set. This | 4944 | * enqueue of curr) will have resulted in resched being set. This |
4941 | * prevents us from potentially nominating it as a false LAST_BUDDY | 4945 | * prevents us from potentially nominating it as a false LAST_BUDDY |
4942 | * below. | 4946 | * below. |
4943 | */ | 4947 | */ |
4944 | if (test_tsk_need_resched(curr)) | 4948 | if (test_tsk_need_resched(curr)) |
4945 | return; | 4949 | return; |
4946 | 4950 | ||
4947 | /* Idle tasks are by definition preempted by non-idle tasks. */ | 4951 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
4948 | if (unlikely(curr->policy == SCHED_IDLE) && | 4952 | if (unlikely(curr->policy == SCHED_IDLE) && |
4949 | likely(p->policy != SCHED_IDLE)) | 4953 | likely(p->policy != SCHED_IDLE)) |
4950 | goto preempt; | 4954 | goto preempt; |
4951 | 4955 | ||
4952 | /* | 4956 | /* |
4953 | * Batch and idle tasks do not preempt non-idle tasks (their preemption | 4957 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
4954 | * is driven by the tick): | 4958 | * is driven by the tick): |
4955 | */ | 4959 | */ |
4956 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) | 4960 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
4957 | return; | 4961 | return; |
4958 | 4962 | ||
4959 | find_matching_se(&se, &pse); | 4963 | find_matching_se(&se, &pse); |
4960 | update_curr(cfs_rq_of(se)); | 4964 | update_curr(cfs_rq_of(se)); |
4961 | BUG_ON(!pse); | 4965 | BUG_ON(!pse); |
4962 | if (wakeup_preempt_entity(se, pse) == 1) { | 4966 | if (wakeup_preempt_entity(se, pse) == 1) { |
4963 | /* | 4967 | /* |
4964 | * Bias pick_next to pick the sched entity that is | 4968 | * Bias pick_next to pick the sched entity that is |
4965 | * triggering this preemption. | 4969 | * triggering this preemption. |
4966 | */ | 4970 | */ |
4967 | if (!next_buddy_marked) | 4971 | if (!next_buddy_marked) |
4968 | set_next_buddy(pse); | 4972 | set_next_buddy(pse); |
4969 | goto preempt; | 4973 | goto preempt; |
4970 | } | 4974 | } |
4971 | 4975 | ||
4972 | return; | 4976 | return; |
4973 | 4977 | ||
4974 | preempt: | 4978 | preempt: |
4975 | resched_curr(rq); | 4979 | resched_curr(rq); |
4976 | /* | 4980 | /* |
4977 | * Only set the backward buddy when the current task is still | 4981 | * Only set the backward buddy when the current task is still |
4978 | * on the rq. This can happen when a wakeup gets interleaved | 4982 | * on the rq. This can happen when a wakeup gets interleaved |
4979 | * with schedule on the ->pre_schedule() or idle_balance() | 4983 | * with schedule on the ->pre_schedule() or idle_balance() |
4980 | * point, either of which can * drop the rq lock. | 4984 | * point, either of which can * drop the rq lock. |
4981 | * | 4985 | * |
4982 | * Also, during early boot the idle thread is in the fair class, | 4986 | * Also, during early boot the idle thread is in the fair class, |
4983 | * for obvious reasons its a bad idea to schedule back to it. | 4987 | * for obvious reasons its a bad idea to schedule back to it. |
4984 | */ | 4988 | */ |
4985 | if (unlikely(!se->on_rq || curr == rq->idle)) | 4989 | if (unlikely(!se->on_rq || curr == rq->idle)) |
4986 | return; | 4990 | return; |
4987 | 4991 | ||
4988 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | 4992 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) |
4989 | set_last_buddy(se); | 4993 | set_last_buddy(se); |
4990 | } | 4994 | } |
4991 | 4995 | ||
4992 | static struct task_struct * | 4996 | static struct task_struct * |
4993 | pick_next_task_fair(struct rq *rq, struct task_struct *prev) | 4997 | pick_next_task_fair(struct rq *rq, struct task_struct *prev) |
4994 | { | 4998 | { |
4995 | struct cfs_rq *cfs_rq = &rq->cfs; | 4999 | struct cfs_rq *cfs_rq = &rq->cfs; |
4996 | struct sched_entity *se; | 5000 | struct sched_entity *se; |
4997 | struct task_struct *p; | 5001 | struct task_struct *p; |
4998 | int new_tasks; | 5002 | int new_tasks; |
4999 | 5003 | ||
5000 | again: | 5004 | again: |
5001 | #ifdef CONFIG_FAIR_GROUP_SCHED | 5005 | #ifdef CONFIG_FAIR_GROUP_SCHED |
5002 | if (!cfs_rq->nr_running) | 5006 | if (!cfs_rq->nr_running) |
5003 | goto idle; | 5007 | goto idle; |
5004 | 5008 | ||
5005 | if (prev->sched_class != &fair_sched_class) | 5009 | if (prev->sched_class != &fair_sched_class) |
5006 | goto simple; | 5010 | goto simple; |
5007 | 5011 | ||
5008 | /* | 5012 | /* |
5009 | * Because of the set_next_buddy() in dequeue_task_fair() it is rather | 5013 | * Because of the set_next_buddy() in dequeue_task_fair() it is rather |
5010 | * likely that a next task is from the same cgroup as the current. | 5014 | * likely that a next task is from the same cgroup as the current. |
5011 | * | 5015 | * |
5012 | * Therefore attempt to avoid putting and setting the entire cgroup | 5016 | * Therefore attempt to avoid putting and setting the entire cgroup |
5013 | * hierarchy, only change the part that actually changes. | 5017 | * hierarchy, only change the part that actually changes. |
5014 | */ | 5018 | */ |
5015 | 5019 | ||
5016 | do { | 5020 | do { |
5017 | struct sched_entity *curr = cfs_rq->curr; | 5021 | struct sched_entity *curr = cfs_rq->curr; |
5018 | 5022 | ||
5019 | /* | 5023 | /* |
5020 | * Since we got here without doing put_prev_entity() we also | 5024 | * Since we got here without doing put_prev_entity() we also |
5021 | * have to consider cfs_rq->curr. If it is still a runnable | 5025 | * have to consider cfs_rq->curr. If it is still a runnable |
5022 | * entity, update_curr() will update its vruntime, otherwise | 5026 | * entity, update_curr() will update its vruntime, otherwise |
5023 | * forget we've ever seen it. | 5027 | * forget we've ever seen it. |
5024 | */ | 5028 | */ |
5025 | if (curr && curr->on_rq) | 5029 | if (curr && curr->on_rq) |
5026 | update_curr(cfs_rq); | 5030 | update_curr(cfs_rq); |
5027 | else | 5031 | else |
5028 | curr = NULL; | 5032 | curr = NULL; |
5029 | 5033 | ||
5030 | /* | 5034 | /* |
5031 | * This call to check_cfs_rq_runtime() will do the throttle and | 5035 | * This call to check_cfs_rq_runtime() will do the throttle and |
5032 | * dequeue its entity in the parent(s). Therefore the 'simple' | 5036 | * dequeue its entity in the parent(s). Therefore the 'simple' |
5033 | * nr_running test will indeed be correct. | 5037 | * nr_running test will indeed be correct. |
5034 | */ | 5038 | */ |
5035 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) | 5039 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) |
5036 | goto simple; | 5040 | goto simple; |
5037 | 5041 | ||
5038 | se = pick_next_entity(cfs_rq, curr); | 5042 | se = pick_next_entity(cfs_rq, curr); |
5039 | cfs_rq = group_cfs_rq(se); | 5043 | cfs_rq = group_cfs_rq(se); |
5040 | } while (cfs_rq); | 5044 | } while (cfs_rq); |
5041 | 5045 | ||
5042 | p = task_of(se); | 5046 | p = task_of(se); |
5043 | 5047 | ||
5044 | /* | 5048 | /* |
5045 | * Since we haven't yet done put_prev_entity and if the selected task | 5049 | * Since we haven't yet done put_prev_entity and if the selected task |
5046 | * is a different task than we started out with, try and touch the | 5050 | * is a different task than we started out with, try and touch the |
5047 | * least amount of cfs_rqs. | 5051 | * least amount of cfs_rqs. |
5048 | */ | 5052 | */ |
5049 | if (prev != p) { | 5053 | if (prev != p) { |
5050 | struct sched_entity *pse = &prev->se; | 5054 | struct sched_entity *pse = &prev->se; |
5051 | 5055 | ||
5052 | while (!(cfs_rq = is_same_group(se, pse))) { | 5056 | while (!(cfs_rq = is_same_group(se, pse))) { |
5053 | int se_depth = se->depth; | 5057 | int se_depth = se->depth; |
5054 | int pse_depth = pse->depth; | 5058 | int pse_depth = pse->depth; |
5055 | 5059 | ||
5056 | if (se_depth <= pse_depth) { | 5060 | if (se_depth <= pse_depth) { |
5057 | put_prev_entity(cfs_rq_of(pse), pse); | 5061 | put_prev_entity(cfs_rq_of(pse), pse); |
5058 | pse = parent_entity(pse); | 5062 | pse = parent_entity(pse); |
5059 | } | 5063 | } |
5060 | if (se_depth >= pse_depth) { | 5064 | if (se_depth >= pse_depth) { |
5061 | set_next_entity(cfs_rq_of(se), se); | 5065 | set_next_entity(cfs_rq_of(se), se); |
5062 | se = parent_entity(se); | 5066 | se = parent_entity(se); |
5063 | } | 5067 | } |
5064 | } | 5068 | } |
5065 | 5069 | ||
5066 | put_prev_entity(cfs_rq, pse); | 5070 | put_prev_entity(cfs_rq, pse); |
5067 | set_next_entity(cfs_rq, se); | 5071 | set_next_entity(cfs_rq, se); |
5068 | } | 5072 | } |
5069 | 5073 | ||
5070 | if (hrtick_enabled(rq)) | 5074 | if (hrtick_enabled(rq)) |
5071 | hrtick_start_fair(rq, p); | 5075 | hrtick_start_fair(rq, p); |
5072 | 5076 | ||
5073 | return p; | 5077 | return p; |
5074 | simple: | 5078 | simple: |
5075 | cfs_rq = &rq->cfs; | 5079 | cfs_rq = &rq->cfs; |
5076 | #endif | 5080 | #endif |
5077 | 5081 | ||
5078 | if (!cfs_rq->nr_running) | 5082 | if (!cfs_rq->nr_running) |
5079 | goto idle; | 5083 | goto idle; |
5080 | 5084 | ||
5081 | put_prev_task(rq, prev); | 5085 | put_prev_task(rq, prev); |
5082 | 5086 | ||
5083 | do { | 5087 | do { |
5084 | se = pick_next_entity(cfs_rq, NULL); | 5088 | se = pick_next_entity(cfs_rq, NULL); |
5085 | set_next_entity(cfs_rq, se); | 5089 | set_next_entity(cfs_rq, se); |
5086 | cfs_rq = group_cfs_rq(se); | 5090 | cfs_rq = group_cfs_rq(se); |
5087 | } while (cfs_rq); | 5091 | } while (cfs_rq); |
5088 | 5092 | ||
5089 | p = task_of(se); | 5093 | p = task_of(se); |
5090 | 5094 | ||
5091 | if (hrtick_enabled(rq)) | 5095 | if (hrtick_enabled(rq)) |
5092 | hrtick_start_fair(rq, p); | 5096 | hrtick_start_fair(rq, p); |
5093 | 5097 | ||
5094 | return p; | 5098 | return p; |
5095 | 5099 | ||
5096 | idle: | 5100 | idle: |
5097 | new_tasks = idle_balance(rq); | 5101 | new_tasks = idle_balance(rq); |
5098 | /* | 5102 | /* |
5099 | * Because idle_balance() releases (and re-acquires) rq->lock, it is | 5103 | * Because idle_balance() releases (and re-acquires) rq->lock, it is |
5100 | * possible for any higher priority task to appear. In that case we | 5104 | * possible for any higher priority task to appear. In that case we |
5101 | * must re-start the pick_next_entity() loop. | 5105 | * must re-start the pick_next_entity() loop. |
5102 | */ | 5106 | */ |
5103 | if (new_tasks < 0) | 5107 | if (new_tasks < 0) |
5104 | return RETRY_TASK; | 5108 | return RETRY_TASK; |
5105 | 5109 | ||
5106 | if (new_tasks > 0) | 5110 | if (new_tasks > 0) |
5107 | goto again; | 5111 | goto again; |
5108 | 5112 | ||
5109 | return NULL; | 5113 | return NULL; |
5110 | } | 5114 | } |
5111 | 5115 | ||
5112 | /* | 5116 | /* |
5113 | * Account for a descheduled task: | 5117 | * Account for a descheduled task: |
5114 | */ | 5118 | */ |
5115 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) | 5119 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
5116 | { | 5120 | { |
5117 | struct sched_entity *se = &prev->se; | 5121 | struct sched_entity *se = &prev->se; |
5118 | struct cfs_rq *cfs_rq; | 5122 | struct cfs_rq *cfs_rq; |
5119 | 5123 | ||
5120 | for_each_sched_entity(se) { | 5124 | for_each_sched_entity(se) { |
5121 | cfs_rq = cfs_rq_of(se); | 5125 | cfs_rq = cfs_rq_of(se); |
5122 | put_prev_entity(cfs_rq, se); | 5126 | put_prev_entity(cfs_rq, se); |
5123 | } | 5127 | } |
5124 | } | 5128 | } |
5125 | 5129 | ||
5126 | /* | 5130 | /* |
5127 | * sched_yield() is very simple | 5131 | * sched_yield() is very simple |
5128 | * | 5132 | * |
5129 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | 5133 | * The magic of dealing with the ->skip buddy is in pick_next_entity. |
5130 | */ | 5134 | */ |
5131 | static void yield_task_fair(struct rq *rq) | 5135 | static void yield_task_fair(struct rq *rq) |
5132 | { | 5136 | { |
5133 | struct task_struct *curr = rq->curr; | 5137 | struct task_struct *curr = rq->curr; |
5134 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | 5138 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
5135 | struct sched_entity *se = &curr->se; | 5139 | struct sched_entity *se = &curr->se; |
5136 | 5140 | ||
5137 | /* | 5141 | /* |
5138 | * Are we the only task in the tree? | 5142 | * Are we the only task in the tree? |
5139 | */ | 5143 | */ |
5140 | if (unlikely(rq->nr_running == 1)) | 5144 | if (unlikely(rq->nr_running == 1)) |
5141 | return; | 5145 | return; |
5142 | 5146 | ||
5143 | clear_buddies(cfs_rq, se); | 5147 | clear_buddies(cfs_rq, se); |
5144 | 5148 | ||
5145 | if (curr->policy != SCHED_BATCH) { | 5149 | if (curr->policy != SCHED_BATCH) { |
5146 | update_rq_clock(rq); | 5150 | update_rq_clock(rq); |
5147 | /* | 5151 | /* |
5148 | * Update run-time statistics of the 'current'. | 5152 | * Update run-time statistics of the 'current'. |
5149 | */ | 5153 | */ |
5150 | update_curr(cfs_rq); | 5154 | update_curr(cfs_rq); |
5151 | /* | 5155 | /* |
5152 | * Tell update_rq_clock() that we've just updated, | 5156 | * Tell update_rq_clock() that we've just updated, |
5153 | * so we don't do microscopic update in schedule() | 5157 | * so we don't do microscopic update in schedule() |
5154 | * and double the fastpath cost. | 5158 | * and double the fastpath cost. |
5155 | */ | 5159 | */ |
5156 | rq->skip_clock_update = 1; | 5160 | rq->skip_clock_update = 1; |
5157 | } | 5161 | } |
5158 | 5162 | ||
5159 | set_skip_buddy(se); | 5163 | set_skip_buddy(se); |
5160 | } | 5164 | } |
5161 | 5165 | ||
5162 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) | 5166 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
5163 | { | 5167 | { |
5164 | struct sched_entity *se = &p->se; | 5168 | struct sched_entity *se = &p->se; |
5165 | 5169 | ||
5166 | /* throttled hierarchies are not runnable */ | 5170 | /* throttled hierarchies are not runnable */ |
5167 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | 5171 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) |
5168 | return false; | 5172 | return false; |
5169 | 5173 | ||
5170 | /* Tell the scheduler that we'd really like pse to run next. */ | 5174 | /* Tell the scheduler that we'd really like pse to run next. */ |
5171 | set_next_buddy(se); | 5175 | set_next_buddy(se); |
5172 | 5176 | ||
5173 | yield_task_fair(rq); | 5177 | yield_task_fair(rq); |
5174 | 5178 | ||
5175 | return true; | 5179 | return true; |
5176 | } | 5180 | } |
5177 | 5181 | ||
5178 | #ifdef CONFIG_SMP | 5182 | #ifdef CONFIG_SMP |
5179 | /************************************************** | 5183 | /************************************************** |
5180 | * Fair scheduling class load-balancing methods. | 5184 | * Fair scheduling class load-balancing methods. |
5181 | * | 5185 | * |
5182 | * BASICS | 5186 | * BASICS |
5183 | * | 5187 | * |
5184 | * The purpose of load-balancing is to achieve the same basic fairness the | 5188 | * The purpose of load-balancing is to achieve the same basic fairness the |
5185 | * per-cpu scheduler provides, namely provide a proportional amount of compute | 5189 | * per-cpu scheduler provides, namely provide a proportional amount of compute |
5186 | * time to each task. This is expressed in the following equation: | 5190 | * time to each task. This is expressed in the following equation: |
5187 | * | 5191 | * |
5188 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | 5192 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) |
5189 | * | 5193 | * |
5190 | * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight | 5194 | * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight |
5191 | * W_i,0 is defined as: | 5195 | * W_i,0 is defined as: |
5192 | * | 5196 | * |
5193 | * W_i,0 = \Sum_j w_i,j (2) | 5197 | * W_i,0 = \Sum_j w_i,j (2) |
5194 | * | 5198 | * |
5195 | * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight | 5199 | * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight |
5196 | * is derived from the nice value as per prio_to_weight[]. | 5200 | * is derived from the nice value as per prio_to_weight[]. |
5197 | * | 5201 | * |
5198 | * The weight average is an exponential decay average of the instantaneous | 5202 | * The weight average is an exponential decay average of the instantaneous |
5199 | * weight: | 5203 | * weight: |
5200 | * | 5204 | * |
5201 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | 5205 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) |
5202 | * | 5206 | * |
5203 | * C_i is the compute capacity of cpu i, typically it is the | 5207 | * C_i is the compute capacity of cpu i, typically it is the |
5204 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it | 5208 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it |
5205 | * can also include other factors [XXX]. | 5209 | * can also include other factors [XXX]. |
5206 | * | 5210 | * |
5207 | * To achieve this balance we define a measure of imbalance which follows | 5211 | * To achieve this balance we define a measure of imbalance which follows |
5208 | * directly from (1): | 5212 | * directly from (1): |
5209 | * | 5213 | * |
5210 | * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) | 5214 | * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) |
5211 | * | 5215 | * |
5212 | * We them move tasks around to minimize the imbalance. In the continuous | 5216 | * We them move tasks around to minimize the imbalance. In the continuous |
5213 | * function space it is obvious this converges, in the discrete case we get | 5217 | * function space it is obvious this converges, in the discrete case we get |
5214 | * a few fun cases generally called infeasible weight scenarios. | 5218 | * a few fun cases generally called infeasible weight scenarios. |
5215 | * | 5219 | * |
5216 | * [XXX expand on: | 5220 | * [XXX expand on: |
5217 | * - infeasible weights; | 5221 | * - infeasible weights; |
5218 | * - local vs global optima in the discrete case. ] | 5222 | * - local vs global optima in the discrete case. ] |
5219 | * | 5223 | * |
5220 | * | 5224 | * |
5221 | * SCHED DOMAINS | 5225 | * SCHED DOMAINS |
5222 | * | 5226 | * |
5223 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | 5227 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) |
5224 | * for all i,j solution, we create a tree of cpus that follows the hardware | 5228 | * for all i,j solution, we create a tree of cpus that follows the hardware |
5225 | * topology where each level pairs two lower groups (or better). This results | 5229 | * topology where each level pairs two lower groups (or better). This results |
5226 | * in O(log n) layers. Furthermore we reduce the number of cpus going up the | 5230 | * in O(log n) layers. Furthermore we reduce the number of cpus going up the |
5227 | * tree to only the first of the previous level and we decrease the frequency | 5231 | * tree to only the first of the previous level and we decrease the frequency |
5228 | * of load-balance at each level inv. proportional to the number of cpus in | 5232 | * of load-balance at each level inv. proportional to the number of cpus in |
5229 | * the groups. | 5233 | * the groups. |
5230 | * | 5234 | * |
5231 | * This yields: | 5235 | * This yields: |
5232 | * | 5236 | * |
5233 | * log_2 n 1 n | 5237 | * log_2 n 1 n |
5234 | * \Sum { --- * --- * 2^i } = O(n) (5) | 5238 | * \Sum { --- * --- * 2^i } = O(n) (5) |
5235 | * i = 0 2^i 2^i | 5239 | * i = 0 2^i 2^i |
5236 | * `- size of each group | 5240 | * `- size of each group |
5237 | * | | `- number of cpus doing load-balance | 5241 | * | | `- number of cpus doing load-balance |
5238 | * | `- freq | 5242 | * | `- freq |
5239 | * `- sum over all levels | 5243 | * `- sum over all levels |
5240 | * | 5244 | * |
5241 | * Coupled with a limit on how many tasks we can migrate every balance pass, | 5245 | * Coupled with a limit on how many tasks we can migrate every balance pass, |
5242 | * this makes (5) the runtime complexity of the balancer. | 5246 | * this makes (5) the runtime complexity of the balancer. |
5243 | * | 5247 | * |
5244 | * An important property here is that each CPU is still (indirectly) connected | 5248 | * An important property here is that each CPU is still (indirectly) connected |
5245 | * to every other cpu in at most O(log n) steps: | 5249 | * to every other cpu in at most O(log n) steps: |
5246 | * | 5250 | * |
5247 | * The adjacency matrix of the resulting graph is given by: | 5251 | * The adjacency matrix of the resulting graph is given by: |
5248 | * | 5252 | * |
5249 | * log_2 n | 5253 | * log_2 n |
5250 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) | 5254 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) |
5251 | * k = 0 | 5255 | * k = 0 |
5252 | * | 5256 | * |
5253 | * And you'll find that: | 5257 | * And you'll find that: |
5254 | * | 5258 | * |
5255 | * A^(log_2 n)_i,j != 0 for all i,j (7) | 5259 | * A^(log_2 n)_i,j != 0 for all i,j (7) |
5256 | * | 5260 | * |
5257 | * Showing there's indeed a path between every cpu in at most O(log n) steps. | 5261 | * Showing there's indeed a path between every cpu in at most O(log n) steps. |
5258 | * The task movement gives a factor of O(m), giving a convergence complexity | 5262 | * The task movement gives a factor of O(m), giving a convergence complexity |
5259 | * of: | 5263 | * of: |
5260 | * | 5264 | * |
5261 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | 5265 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) |
5262 | * | 5266 | * |
5263 | * | 5267 | * |
5264 | * WORK CONSERVING | 5268 | * WORK CONSERVING |
5265 | * | 5269 | * |
5266 | * In order to avoid CPUs going idle while there's still work to do, new idle | 5270 | * In order to avoid CPUs going idle while there's still work to do, new idle |
5267 | * balancing is more aggressive and has the newly idle cpu iterate up the domain | 5271 | * balancing is more aggressive and has the newly idle cpu iterate up the domain |
5268 | * tree itself instead of relying on other CPUs to bring it work. | 5272 | * tree itself instead of relying on other CPUs to bring it work. |
5269 | * | 5273 | * |
5270 | * This adds some complexity to both (5) and (8) but it reduces the total idle | 5274 | * This adds some complexity to both (5) and (8) but it reduces the total idle |
5271 | * time. | 5275 | * time. |
5272 | * | 5276 | * |
5273 | * [XXX more?] | 5277 | * [XXX more?] |
5274 | * | 5278 | * |
5275 | * | 5279 | * |
5276 | * CGROUPS | 5280 | * CGROUPS |
5277 | * | 5281 | * |
5278 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | 5282 | * Cgroups make a horror show out of (2), instead of a simple sum we get: |
5279 | * | 5283 | * |
5280 | * s_k,i | 5284 | * s_k,i |
5281 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | 5285 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) |
5282 | * S_k | 5286 | * S_k |
5283 | * | 5287 | * |
5284 | * Where | 5288 | * Where |
5285 | * | 5289 | * |
5286 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | 5290 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) |
5287 | * | 5291 | * |
5288 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. | 5292 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. |
5289 | * | 5293 | * |
5290 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | 5294 | * The big problem is S_k, its a global sum needed to compute a local (W_i) |
5291 | * property. | 5295 | * property. |
5292 | * | 5296 | * |
5293 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | 5297 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that |
5294 | * rewrite all of this once again.] | 5298 | * rewrite all of this once again.] |
5295 | */ | 5299 | */ |
5296 | 5300 | ||
5297 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; | 5301 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
5298 | 5302 | ||
5299 | enum fbq_type { regular, remote, all }; | 5303 | enum fbq_type { regular, remote, all }; |
5300 | 5304 | ||
5301 | #define LBF_ALL_PINNED 0x01 | 5305 | #define LBF_ALL_PINNED 0x01 |
5302 | #define LBF_NEED_BREAK 0x02 | 5306 | #define LBF_NEED_BREAK 0x02 |
5303 | #define LBF_DST_PINNED 0x04 | 5307 | #define LBF_DST_PINNED 0x04 |
5304 | #define LBF_SOME_PINNED 0x08 | 5308 | #define LBF_SOME_PINNED 0x08 |
5305 | 5309 | ||
5306 | struct lb_env { | 5310 | struct lb_env { |
5307 | struct sched_domain *sd; | 5311 | struct sched_domain *sd; |
5308 | 5312 | ||
5309 | struct rq *src_rq; | 5313 | struct rq *src_rq; |
5310 | int src_cpu; | 5314 | int src_cpu; |
5311 | 5315 | ||
5312 | int dst_cpu; | 5316 | int dst_cpu; |
5313 | struct rq *dst_rq; | 5317 | struct rq *dst_rq; |
5314 | 5318 | ||
5315 | struct cpumask *dst_grpmask; | 5319 | struct cpumask *dst_grpmask; |
5316 | int new_dst_cpu; | 5320 | int new_dst_cpu; |
5317 | enum cpu_idle_type idle; | 5321 | enum cpu_idle_type idle; |
5318 | long imbalance; | 5322 | long imbalance; |
5319 | /* The set of CPUs under consideration for load-balancing */ | 5323 | /* The set of CPUs under consideration for load-balancing */ |
5320 | struct cpumask *cpus; | 5324 | struct cpumask *cpus; |
5321 | 5325 | ||
5322 | unsigned int flags; | 5326 | unsigned int flags; |
5323 | 5327 | ||
5324 | unsigned int loop; | 5328 | unsigned int loop; |
5325 | unsigned int loop_break; | 5329 | unsigned int loop_break; |
5326 | unsigned int loop_max; | 5330 | unsigned int loop_max; |
5327 | 5331 | ||
5328 | enum fbq_type fbq_type; | 5332 | enum fbq_type fbq_type; |
5329 | struct list_head tasks; | 5333 | struct list_head tasks; |
5330 | }; | 5334 | }; |
5331 | 5335 | ||
5332 | /* | 5336 | /* |
5333 | * Is this task likely cache-hot: | 5337 | * Is this task likely cache-hot: |
5334 | */ | 5338 | */ |
5335 | static int task_hot(struct task_struct *p, struct lb_env *env) | 5339 | static int task_hot(struct task_struct *p, struct lb_env *env) |
5336 | { | 5340 | { |
5337 | s64 delta; | 5341 | s64 delta; |
5338 | 5342 | ||
5339 | lockdep_assert_held(&env->src_rq->lock); | 5343 | lockdep_assert_held(&env->src_rq->lock); |
5340 | 5344 | ||
5341 | if (p->sched_class != &fair_sched_class) | 5345 | if (p->sched_class != &fair_sched_class) |
5342 | return 0; | 5346 | return 0; |
5343 | 5347 | ||
5344 | if (unlikely(p->policy == SCHED_IDLE)) | 5348 | if (unlikely(p->policy == SCHED_IDLE)) |
5345 | return 0; | 5349 | return 0; |
5346 | 5350 | ||
5347 | /* | 5351 | /* |
5348 | * Buddy candidates are cache hot: | 5352 | * Buddy candidates are cache hot: |
5349 | */ | 5353 | */ |
5350 | if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && | 5354 | if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && |
5351 | (&p->se == cfs_rq_of(&p->se)->next || | 5355 | (&p->se == cfs_rq_of(&p->se)->next || |
5352 | &p->se == cfs_rq_of(&p->se)->last)) | 5356 | &p->se == cfs_rq_of(&p->se)->last)) |
5353 | return 1; | 5357 | return 1; |
5354 | 5358 | ||
5355 | if (sysctl_sched_migration_cost == -1) | 5359 | if (sysctl_sched_migration_cost == -1) |
5356 | return 1; | 5360 | return 1; |
5357 | if (sysctl_sched_migration_cost == 0) | 5361 | if (sysctl_sched_migration_cost == 0) |
5358 | return 0; | 5362 | return 0; |
5359 | 5363 | ||
5360 | delta = rq_clock_task(env->src_rq) - p->se.exec_start; | 5364 | delta = rq_clock_task(env->src_rq) - p->se.exec_start; |
5361 | 5365 | ||
5362 | return delta < (s64)sysctl_sched_migration_cost; | 5366 | return delta < (s64)sysctl_sched_migration_cost; |
5363 | } | 5367 | } |
5364 | 5368 | ||
5365 | #ifdef CONFIG_NUMA_BALANCING | 5369 | #ifdef CONFIG_NUMA_BALANCING |
5366 | /* Returns true if the destination node has incurred more faults */ | 5370 | /* Returns true if the destination node has incurred more faults */ |
5367 | static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env) | 5371 | static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env) |
5368 | { | 5372 | { |
5369 | struct numa_group *numa_group = rcu_dereference(p->numa_group); | 5373 | struct numa_group *numa_group = rcu_dereference(p->numa_group); |
5370 | int src_nid, dst_nid; | 5374 | int src_nid, dst_nid; |
5371 | 5375 | ||
5372 | if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults || | 5376 | if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults || |
5373 | !(env->sd->flags & SD_NUMA)) { | 5377 | !(env->sd->flags & SD_NUMA)) { |
5374 | return false; | 5378 | return false; |
5375 | } | 5379 | } |
5376 | 5380 | ||
5377 | src_nid = cpu_to_node(env->src_cpu); | 5381 | src_nid = cpu_to_node(env->src_cpu); |
5378 | dst_nid = cpu_to_node(env->dst_cpu); | 5382 | dst_nid = cpu_to_node(env->dst_cpu); |
5379 | 5383 | ||
5380 | if (src_nid == dst_nid) | 5384 | if (src_nid == dst_nid) |
5381 | return false; | 5385 | return false; |
5382 | 5386 | ||
5383 | if (numa_group) { | 5387 | if (numa_group) { |
5384 | /* Task is already in the group's interleave set. */ | 5388 | /* Task is already in the group's interleave set. */ |
5385 | if (node_isset(src_nid, numa_group->active_nodes)) | 5389 | if (node_isset(src_nid, numa_group->active_nodes)) |
5386 | return false; | 5390 | return false; |
5387 | 5391 | ||
5388 | /* Task is moving into the group's interleave set. */ | 5392 | /* Task is moving into the group's interleave set. */ |
5389 | if (node_isset(dst_nid, numa_group->active_nodes)) | 5393 | if (node_isset(dst_nid, numa_group->active_nodes)) |
5390 | return true; | 5394 | return true; |
5391 | 5395 | ||
5392 | return group_faults(p, dst_nid) > group_faults(p, src_nid); | 5396 | return group_faults(p, dst_nid) > group_faults(p, src_nid); |
5393 | } | 5397 | } |
5394 | 5398 | ||
5395 | /* Encourage migration to the preferred node. */ | 5399 | /* Encourage migration to the preferred node. */ |
5396 | if (dst_nid == p->numa_preferred_nid) | 5400 | if (dst_nid == p->numa_preferred_nid) |
5397 | return true; | 5401 | return true; |
5398 | 5402 | ||
5399 | return task_faults(p, dst_nid) > task_faults(p, src_nid); | 5403 | return task_faults(p, dst_nid) > task_faults(p, src_nid); |
5400 | } | 5404 | } |
5401 | 5405 | ||
5402 | 5406 | ||
5403 | static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env) | 5407 | static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env) |
5404 | { | 5408 | { |
5405 | struct numa_group *numa_group = rcu_dereference(p->numa_group); | 5409 | struct numa_group *numa_group = rcu_dereference(p->numa_group); |
5406 | int src_nid, dst_nid; | 5410 | int src_nid, dst_nid; |
5407 | 5411 | ||
5408 | if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER)) | 5412 | if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER)) |
5409 | return false; | 5413 | return false; |
5410 | 5414 | ||
5411 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) | 5415 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) |
5412 | return false; | 5416 | return false; |
5413 | 5417 | ||
5414 | src_nid = cpu_to_node(env->src_cpu); | 5418 | src_nid = cpu_to_node(env->src_cpu); |
5415 | dst_nid = cpu_to_node(env->dst_cpu); | 5419 | dst_nid = cpu_to_node(env->dst_cpu); |
5416 | 5420 | ||
5417 | if (src_nid == dst_nid) | 5421 | if (src_nid == dst_nid) |
5418 | return false; | 5422 | return false; |
5419 | 5423 | ||
5420 | if (numa_group) { | 5424 | if (numa_group) { |
5421 | /* Task is moving within/into the group's interleave set. */ | 5425 | /* Task is moving within/into the group's interleave set. */ |
5422 | if (node_isset(dst_nid, numa_group->active_nodes)) | 5426 | if (node_isset(dst_nid, numa_group->active_nodes)) |
5423 | return false; | 5427 | return false; |
5424 | 5428 | ||
5425 | /* Task is moving out of the group's interleave set. */ | 5429 | /* Task is moving out of the group's interleave set. */ |
5426 | if (node_isset(src_nid, numa_group->active_nodes)) | 5430 | if (node_isset(src_nid, numa_group->active_nodes)) |
5427 | return true; | 5431 | return true; |
5428 | 5432 | ||
5429 | return group_faults(p, dst_nid) < group_faults(p, src_nid); | 5433 | return group_faults(p, dst_nid) < group_faults(p, src_nid); |
5430 | } | 5434 | } |
5431 | 5435 | ||
5432 | /* Migrating away from the preferred node is always bad. */ | 5436 | /* Migrating away from the preferred node is always bad. */ |
5433 | if (src_nid == p->numa_preferred_nid) | 5437 | if (src_nid == p->numa_preferred_nid) |
5434 | return true; | 5438 | return true; |
5435 | 5439 | ||
5436 | return task_faults(p, dst_nid) < task_faults(p, src_nid); | 5440 | return task_faults(p, dst_nid) < task_faults(p, src_nid); |
5437 | } | 5441 | } |
5438 | 5442 | ||
5439 | #else | 5443 | #else |
5440 | static inline bool migrate_improves_locality(struct task_struct *p, | 5444 | static inline bool migrate_improves_locality(struct task_struct *p, |
5441 | struct lb_env *env) | 5445 | struct lb_env *env) |
5442 | { | 5446 | { |
5443 | return false; | 5447 | return false; |
5444 | } | 5448 | } |
5445 | 5449 | ||
5446 | static inline bool migrate_degrades_locality(struct task_struct *p, | 5450 | static inline bool migrate_degrades_locality(struct task_struct *p, |
5447 | struct lb_env *env) | 5451 | struct lb_env *env) |
5448 | { | 5452 | { |
5449 | return false; | 5453 | return false; |
5450 | } | 5454 | } |
5451 | #endif | 5455 | #endif |
5452 | 5456 | ||
5453 | /* | 5457 | /* |
5454 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | 5458 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? |
5455 | */ | 5459 | */ |
5456 | static | 5460 | static |
5457 | int can_migrate_task(struct task_struct *p, struct lb_env *env) | 5461 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
5458 | { | 5462 | { |
5459 | int tsk_cache_hot = 0; | 5463 | int tsk_cache_hot = 0; |
5460 | 5464 | ||
5461 | lockdep_assert_held(&env->src_rq->lock); | 5465 | lockdep_assert_held(&env->src_rq->lock); |
5462 | 5466 | ||
5463 | /* | 5467 | /* |
5464 | * We do not migrate tasks that are: | 5468 | * We do not migrate tasks that are: |
5465 | * 1) throttled_lb_pair, or | 5469 | * 1) throttled_lb_pair, or |
5466 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | 5470 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
5467 | * 3) running (obviously), or | 5471 | * 3) running (obviously), or |
5468 | * 4) are cache-hot on their current CPU. | 5472 | * 4) are cache-hot on their current CPU. |
5469 | */ | 5473 | */ |
5470 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) | 5474 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
5471 | return 0; | 5475 | return 0; |
5472 | 5476 | ||
5473 | if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { | 5477 | if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { |
5474 | int cpu; | 5478 | int cpu; |
5475 | 5479 | ||
5476 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); | 5480 | schedstat_inc(p, se.statistics.nr_failed_migrations_affine); |
5477 | 5481 | ||
5478 | env->flags |= LBF_SOME_PINNED; | 5482 | env->flags |= LBF_SOME_PINNED; |
5479 | 5483 | ||
5480 | /* | 5484 | /* |
5481 | * Remember if this task can be migrated to any other cpu in | 5485 | * Remember if this task can be migrated to any other cpu in |
5482 | * our sched_group. We may want to revisit it if we couldn't | 5486 | * our sched_group. We may want to revisit it if we couldn't |
5483 | * meet load balance goals by pulling other tasks on src_cpu. | 5487 | * meet load balance goals by pulling other tasks on src_cpu. |
5484 | * | 5488 | * |
5485 | * Also avoid computing new_dst_cpu if we have already computed | 5489 | * Also avoid computing new_dst_cpu if we have already computed |
5486 | * one in current iteration. | 5490 | * one in current iteration. |
5487 | */ | 5491 | */ |
5488 | if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED)) | 5492 | if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED)) |
5489 | return 0; | 5493 | return 0; |
5490 | 5494 | ||
5491 | /* Prevent to re-select dst_cpu via env's cpus */ | 5495 | /* Prevent to re-select dst_cpu via env's cpus */ |
5492 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { | 5496 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { |
5493 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) { | 5497 | if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) { |
5494 | env->flags |= LBF_DST_PINNED; | 5498 | env->flags |= LBF_DST_PINNED; |
5495 | env->new_dst_cpu = cpu; | 5499 | env->new_dst_cpu = cpu; |
5496 | break; | 5500 | break; |
5497 | } | 5501 | } |
5498 | } | 5502 | } |
5499 | 5503 | ||
5500 | return 0; | 5504 | return 0; |
5501 | } | 5505 | } |
5502 | 5506 | ||
5503 | /* Record that we found atleast one task that could run on dst_cpu */ | 5507 | /* Record that we found atleast one task that could run on dst_cpu */ |
5504 | env->flags &= ~LBF_ALL_PINNED; | 5508 | env->flags &= ~LBF_ALL_PINNED; |
5505 | 5509 | ||
5506 | if (task_running(env->src_rq, p)) { | 5510 | if (task_running(env->src_rq, p)) { |
5507 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); | 5511 | schedstat_inc(p, se.statistics.nr_failed_migrations_running); |
5508 | return 0; | 5512 | return 0; |
5509 | } | 5513 | } |
5510 | 5514 | ||
5511 | /* | 5515 | /* |
5512 | * Aggressive migration if: | 5516 | * Aggressive migration if: |
5513 | * 1) destination numa is preferred | 5517 | * 1) destination numa is preferred |
5514 | * 2) task is cache cold, or | 5518 | * 2) task is cache cold, or |
5515 | * 3) too many balance attempts have failed. | 5519 | * 3) too many balance attempts have failed. |
5516 | */ | 5520 | */ |
5517 | tsk_cache_hot = task_hot(p, env); | 5521 | tsk_cache_hot = task_hot(p, env); |
5518 | if (!tsk_cache_hot) | 5522 | if (!tsk_cache_hot) |
5519 | tsk_cache_hot = migrate_degrades_locality(p, env); | 5523 | tsk_cache_hot = migrate_degrades_locality(p, env); |
5520 | 5524 | ||
5521 | if (migrate_improves_locality(p, env) || !tsk_cache_hot || | 5525 | if (migrate_improves_locality(p, env) || !tsk_cache_hot || |
5522 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { | 5526 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
5523 | if (tsk_cache_hot) { | 5527 | if (tsk_cache_hot) { |
5524 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); | 5528 | schedstat_inc(env->sd, lb_hot_gained[env->idle]); |
5525 | schedstat_inc(p, se.statistics.nr_forced_migrations); | 5529 | schedstat_inc(p, se.statistics.nr_forced_migrations); |
5526 | } | 5530 | } |
5527 | return 1; | 5531 | return 1; |
5528 | } | 5532 | } |
5529 | 5533 | ||
5530 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); | 5534 | schedstat_inc(p, se.statistics.nr_failed_migrations_hot); |
5531 | return 0; | 5535 | return 0; |
5532 | } | 5536 | } |
5533 | 5537 | ||
5534 | /* | 5538 | /* |
5535 | * detach_task() -- detach the task for the migration specified in env | 5539 | * detach_task() -- detach the task for the migration specified in env |
5536 | */ | 5540 | */ |
5537 | static void detach_task(struct task_struct *p, struct lb_env *env) | 5541 | static void detach_task(struct task_struct *p, struct lb_env *env) |
5538 | { | 5542 | { |
5539 | lockdep_assert_held(&env->src_rq->lock); | 5543 | lockdep_assert_held(&env->src_rq->lock); |
5540 | 5544 | ||
5541 | deactivate_task(env->src_rq, p, 0); | 5545 | deactivate_task(env->src_rq, p, 0); |
5542 | p->on_rq = TASK_ON_RQ_MIGRATING; | 5546 | p->on_rq = TASK_ON_RQ_MIGRATING; |
5543 | set_task_cpu(p, env->dst_cpu); | 5547 | set_task_cpu(p, env->dst_cpu); |
5544 | } | 5548 | } |
5545 | 5549 | ||
5546 | /* | 5550 | /* |
5547 | * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as | 5551 | * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as |
5548 | * part of active balancing operations within "domain". | 5552 | * part of active balancing operations within "domain". |
5549 | * | 5553 | * |
5550 | * Returns a task if successful and NULL otherwise. | 5554 | * Returns a task if successful and NULL otherwise. |
5551 | */ | 5555 | */ |
5552 | static struct task_struct *detach_one_task(struct lb_env *env) | 5556 | static struct task_struct *detach_one_task(struct lb_env *env) |
5553 | { | 5557 | { |
5554 | struct task_struct *p, *n; | 5558 | struct task_struct *p, *n; |
5555 | 5559 | ||
5556 | lockdep_assert_held(&env->src_rq->lock); | 5560 | lockdep_assert_held(&env->src_rq->lock); |
5557 | 5561 | ||
5558 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { | 5562 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { |
5559 | if (!can_migrate_task(p, env)) | 5563 | if (!can_migrate_task(p, env)) |
5560 | continue; | 5564 | continue; |
5561 | 5565 | ||
5562 | detach_task(p, env); | 5566 | detach_task(p, env); |
5563 | 5567 | ||
5564 | /* | 5568 | /* |
5565 | * Right now, this is only the second place where | 5569 | * Right now, this is only the second place where |
5566 | * lb_gained[env->idle] is updated (other is detach_tasks) | 5570 | * lb_gained[env->idle] is updated (other is detach_tasks) |
5567 | * so we can safely collect stats here rather than | 5571 | * so we can safely collect stats here rather than |
5568 | * inside detach_tasks(). | 5572 | * inside detach_tasks(). |
5569 | */ | 5573 | */ |
5570 | schedstat_inc(env->sd, lb_gained[env->idle]); | 5574 | schedstat_inc(env->sd, lb_gained[env->idle]); |
5571 | return p; | 5575 | return p; |
5572 | } | 5576 | } |
5573 | return NULL; | 5577 | return NULL; |
5574 | } | 5578 | } |
5575 | 5579 | ||
5576 | static const unsigned int sched_nr_migrate_break = 32; | 5580 | static const unsigned int sched_nr_migrate_break = 32; |
5577 | 5581 | ||
5578 | /* | 5582 | /* |
5579 | * detach_tasks() -- tries to detach up to imbalance weighted load from | 5583 | * detach_tasks() -- tries to detach up to imbalance weighted load from |
5580 | * busiest_rq, as part of a balancing operation within domain "sd". | 5584 | * busiest_rq, as part of a balancing operation within domain "sd". |
5581 | * | 5585 | * |
5582 | * Returns number of detached tasks if successful and 0 otherwise. | 5586 | * Returns number of detached tasks if successful and 0 otherwise. |
5583 | */ | 5587 | */ |
5584 | static int detach_tasks(struct lb_env *env) | 5588 | static int detach_tasks(struct lb_env *env) |
5585 | { | 5589 | { |
5586 | struct list_head *tasks = &env->src_rq->cfs_tasks; | 5590 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
5587 | struct task_struct *p; | 5591 | struct task_struct *p; |
5588 | unsigned long load; | 5592 | unsigned long load; |
5589 | int detached = 0; | 5593 | int detached = 0; |
5590 | 5594 | ||
5591 | lockdep_assert_held(&env->src_rq->lock); | 5595 | lockdep_assert_held(&env->src_rq->lock); |
5592 | 5596 | ||
5593 | if (env->imbalance <= 0) | 5597 | if (env->imbalance <= 0) |
5594 | return 0; | 5598 | return 0; |
5595 | 5599 | ||
5596 | while (!list_empty(tasks)) { | 5600 | while (!list_empty(tasks)) { |
5597 | p = list_first_entry(tasks, struct task_struct, se.group_node); | 5601 | p = list_first_entry(tasks, struct task_struct, se.group_node); |
5598 | 5602 | ||
5599 | env->loop++; | 5603 | env->loop++; |
5600 | /* We've more or less seen every task there is, call it quits */ | 5604 | /* We've more or less seen every task there is, call it quits */ |
5601 | if (env->loop > env->loop_max) | 5605 | if (env->loop > env->loop_max) |
5602 | break; | 5606 | break; |
5603 | 5607 | ||
5604 | /* take a breather every nr_migrate tasks */ | 5608 | /* take a breather every nr_migrate tasks */ |
5605 | if (env->loop > env->loop_break) { | 5609 | if (env->loop > env->loop_break) { |
5606 | env->loop_break += sched_nr_migrate_break; | 5610 | env->loop_break += sched_nr_migrate_break; |
5607 | env->flags |= LBF_NEED_BREAK; | 5611 | env->flags |= LBF_NEED_BREAK; |
5608 | break; | 5612 | break; |
5609 | } | 5613 | } |
5610 | 5614 | ||
5611 | if (!can_migrate_task(p, env)) | 5615 | if (!can_migrate_task(p, env)) |
5612 | goto next; | 5616 | goto next; |
5613 | 5617 | ||
5614 | load = task_h_load(p); | 5618 | load = task_h_load(p); |
5615 | 5619 | ||
5616 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) | 5620 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
5617 | goto next; | 5621 | goto next; |
5618 | 5622 | ||
5619 | if ((load / 2) > env->imbalance) | 5623 | if ((load / 2) > env->imbalance) |
5620 | goto next; | 5624 | goto next; |
5621 | 5625 | ||
5622 | detach_task(p, env); | 5626 | detach_task(p, env); |
5623 | list_add(&p->se.group_node, &env->tasks); | 5627 | list_add(&p->se.group_node, &env->tasks); |
5624 | 5628 | ||
5625 | detached++; | 5629 | detached++; |
5626 | env->imbalance -= load; | 5630 | env->imbalance -= load; |
5627 | 5631 | ||
5628 | #ifdef CONFIG_PREEMPT | 5632 | #ifdef CONFIG_PREEMPT |
5629 | /* | 5633 | /* |
5630 | * NEWIDLE balancing is a source of latency, so preemptible | 5634 | * NEWIDLE balancing is a source of latency, so preemptible |
5631 | * kernels will stop after the first task is detached to minimize | 5635 | * kernels will stop after the first task is detached to minimize |
5632 | * the critical section. | 5636 | * the critical section. |
5633 | */ | 5637 | */ |
5634 | if (env->idle == CPU_NEWLY_IDLE) | 5638 | if (env->idle == CPU_NEWLY_IDLE) |
5635 | break; | 5639 | break; |
5636 | #endif | 5640 | #endif |
5637 | 5641 | ||
5638 | /* | 5642 | /* |
5639 | * We only want to steal up to the prescribed amount of | 5643 | * We only want to steal up to the prescribed amount of |
5640 | * weighted load. | 5644 | * weighted load. |
5641 | */ | 5645 | */ |
5642 | if (env->imbalance <= 0) | 5646 | if (env->imbalance <= 0) |
5643 | break; | 5647 | break; |
5644 | 5648 | ||
5645 | continue; | 5649 | continue; |
5646 | next: | 5650 | next: |
5647 | list_move_tail(&p->se.group_node, tasks); | 5651 | list_move_tail(&p->se.group_node, tasks); |
5648 | } | 5652 | } |
5649 | 5653 | ||
5650 | /* | 5654 | /* |
5651 | * Right now, this is one of only two places we collect this stat | 5655 | * Right now, this is one of only two places we collect this stat |
5652 | * so we can safely collect detach_one_task() stats here rather | 5656 | * so we can safely collect detach_one_task() stats here rather |
5653 | * than inside detach_one_task(). | 5657 | * than inside detach_one_task(). |
5654 | */ | 5658 | */ |
5655 | schedstat_add(env->sd, lb_gained[env->idle], detached); | 5659 | schedstat_add(env->sd, lb_gained[env->idle], detached); |
5656 | 5660 | ||
5657 | return detached; | 5661 | return detached; |
5658 | } | 5662 | } |
5659 | 5663 | ||
5660 | /* | 5664 | /* |
5661 | * attach_task() -- attach the task detached by detach_task() to its new rq. | 5665 | * attach_task() -- attach the task detached by detach_task() to its new rq. |
5662 | */ | 5666 | */ |
5663 | static void attach_task(struct rq *rq, struct task_struct *p) | 5667 | static void attach_task(struct rq *rq, struct task_struct *p) |
5664 | { | 5668 | { |
5665 | lockdep_assert_held(&rq->lock); | 5669 | lockdep_assert_held(&rq->lock); |
5666 | 5670 | ||
5667 | BUG_ON(task_rq(p) != rq); | 5671 | BUG_ON(task_rq(p) != rq); |
5668 | p->on_rq = TASK_ON_RQ_QUEUED; | 5672 | p->on_rq = TASK_ON_RQ_QUEUED; |
5669 | activate_task(rq, p, 0); | 5673 | activate_task(rq, p, 0); |
5670 | check_preempt_curr(rq, p, 0); | 5674 | check_preempt_curr(rq, p, 0); |
5671 | } | 5675 | } |
5672 | 5676 | ||
5673 | /* | 5677 | /* |
5674 | * attach_one_task() -- attaches the task returned from detach_one_task() to | 5678 | * attach_one_task() -- attaches the task returned from detach_one_task() to |
5675 | * its new rq. | 5679 | * its new rq. |
5676 | */ | 5680 | */ |
5677 | static void attach_one_task(struct rq *rq, struct task_struct *p) | 5681 | static void attach_one_task(struct rq *rq, struct task_struct *p) |
5678 | { | 5682 | { |
5679 | raw_spin_lock(&rq->lock); | 5683 | raw_spin_lock(&rq->lock); |
5680 | attach_task(rq, p); | 5684 | attach_task(rq, p); |
5681 | raw_spin_unlock(&rq->lock); | 5685 | raw_spin_unlock(&rq->lock); |
5682 | } | 5686 | } |
5683 | 5687 | ||
5684 | /* | 5688 | /* |
5685 | * attach_tasks() -- attaches all tasks detached by detach_tasks() to their | 5689 | * attach_tasks() -- attaches all tasks detached by detach_tasks() to their |
5686 | * new rq. | 5690 | * new rq. |
5687 | */ | 5691 | */ |
5688 | static void attach_tasks(struct lb_env *env) | 5692 | static void attach_tasks(struct lb_env *env) |
5689 | { | 5693 | { |
5690 | struct list_head *tasks = &env->tasks; | 5694 | struct list_head *tasks = &env->tasks; |
5691 | struct task_struct *p; | 5695 | struct task_struct *p; |
5692 | 5696 | ||
5693 | raw_spin_lock(&env->dst_rq->lock); | 5697 | raw_spin_lock(&env->dst_rq->lock); |
5694 | 5698 | ||
5695 | while (!list_empty(tasks)) { | 5699 | while (!list_empty(tasks)) { |
5696 | p = list_first_entry(tasks, struct task_struct, se.group_node); | 5700 | p = list_first_entry(tasks, struct task_struct, se.group_node); |
5697 | list_del_init(&p->se.group_node); | 5701 | list_del_init(&p->se.group_node); |
5698 | 5702 | ||
5699 | attach_task(env->dst_rq, p); | 5703 | attach_task(env->dst_rq, p); |
5700 | } | 5704 | } |
5701 | 5705 | ||
5702 | raw_spin_unlock(&env->dst_rq->lock); | 5706 | raw_spin_unlock(&env->dst_rq->lock); |
5703 | } | 5707 | } |
5704 | 5708 | ||
5705 | #ifdef CONFIG_FAIR_GROUP_SCHED | 5709 | #ifdef CONFIG_FAIR_GROUP_SCHED |
5706 | /* | 5710 | /* |
5707 | * update tg->load_weight by folding this cpu's load_avg | 5711 | * update tg->load_weight by folding this cpu's load_avg |
5708 | */ | 5712 | */ |
5709 | static void __update_blocked_averages_cpu(struct task_group *tg, int cpu) | 5713 | static void __update_blocked_averages_cpu(struct task_group *tg, int cpu) |
5710 | { | 5714 | { |
5711 | struct sched_entity *se = tg->se[cpu]; | 5715 | struct sched_entity *se = tg->se[cpu]; |
5712 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu]; | 5716 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu]; |
5713 | 5717 | ||
5714 | /* throttled entities do not contribute to load */ | 5718 | /* throttled entities do not contribute to load */ |
5715 | if (throttled_hierarchy(cfs_rq)) | 5719 | if (throttled_hierarchy(cfs_rq)) |
5716 | return; | 5720 | return; |
5717 | 5721 | ||
5718 | update_cfs_rq_blocked_load(cfs_rq, 1); | 5722 | update_cfs_rq_blocked_load(cfs_rq, 1); |
5719 | 5723 | ||
5720 | if (se) { | 5724 | if (se) { |
5721 | update_entity_load_avg(se, 1); | 5725 | update_entity_load_avg(se, 1); |
5722 | /* | 5726 | /* |
5723 | * We pivot on our runnable average having decayed to zero for | 5727 | * We pivot on our runnable average having decayed to zero for |
5724 | * list removal. This generally implies that all our children | 5728 | * list removal. This generally implies that all our children |
5725 | * have also been removed (modulo rounding error or bandwidth | 5729 | * have also been removed (modulo rounding error or bandwidth |
5726 | * control); however, such cases are rare and we can fix these | 5730 | * control); however, such cases are rare and we can fix these |
5727 | * at enqueue. | 5731 | * at enqueue. |
5728 | * | 5732 | * |
5729 | * TODO: fix up out-of-order children on enqueue. | 5733 | * TODO: fix up out-of-order children on enqueue. |
5730 | */ | 5734 | */ |
5731 | if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running) | 5735 | if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running) |
5732 | list_del_leaf_cfs_rq(cfs_rq); | 5736 | list_del_leaf_cfs_rq(cfs_rq); |
5733 | } else { | 5737 | } else { |
5734 | struct rq *rq = rq_of(cfs_rq); | 5738 | struct rq *rq = rq_of(cfs_rq); |
5735 | update_rq_runnable_avg(rq, rq->nr_running); | 5739 | update_rq_runnable_avg(rq, rq->nr_running); |
5736 | } | 5740 | } |
5737 | } | 5741 | } |
5738 | 5742 | ||
5739 | static void update_blocked_averages(int cpu) | 5743 | static void update_blocked_averages(int cpu) |
5740 | { | 5744 | { |
5741 | struct rq *rq = cpu_rq(cpu); | 5745 | struct rq *rq = cpu_rq(cpu); |
5742 | struct cfs_rq *cfs_rq; | 5746 | struct cfs_rq *cfs_rq; |
5743 | unsigned long flags; | 5747 | unsigned long flags; |
5744 | 5748 | ||
5745 | raw_spin_lock_irqsave(&rq->lock, flags); | 5749 | raw_spin_lock_irqsave(&rq->lock, flags); |
5746 | update_rq_clock(rq); | 5750 | update_rq_clock(rq); |
5747 | /* | 5751 | /* |
5748 | * Iterates the task_group tree in a bottom up fashion, see | 5752 | * Iterates the task_group tree in a bottom up fashion, see |
5749 | * list_add_leaf_cfs_rq() for details. | 5753 | * list_add_leaf_cfs_rq() for details. |
5750 | */ | 5754 | */ |
5751 | for_each_leaf_cfs_rq(rq, cfs_rq) { | 5755 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
5752 | /* | 5756 | /* |
5753 | * Note: We may want to consider periodically releasing | 5757 | * Note: We may want to consider periodically releasing |
5754 | * rq->lock about these updates so that creating many task | 5758 | * rq->lock about these updates so that creating many task |
5755 | * groups does not result in continually extending hold time. | 5759 | * groups does not result in continually extending hold time. |
5756 | */ | 5760 | */ |
5757 | __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu); | 5761 | __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu); |
5758 | } | 5762 | } |
5759 | 5763 | ||
5760 | raw_spin_unlock_irqrestore(&rq->lock, flags); | 5764 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
5761 | } | 5765 | } |
5762 | 5766 | ||
5763 | /* | 5767 | /* |
5764 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. | 5768 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
5765 | * This needs to be done in a top-down fashion because the load of a child | 5769 | * This needs to be done in a top-down fashion because the load of a child |
5766 | * group is a fraction of its parents load. | 5770 | * group is a fraction of its parents load. |
5767 | */ | 5771 | */ |
5768 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) | 5772 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
5769 | { | 5773 | { |
5770 | struct rq *rq = rq_of(cfs_rq); | 5774 | struct rq *rq = rq_of(cfs_rq); |
5771 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | 5775 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; |
5772 | unsigned long now = jiffies; | 5776 | unsigned long now = jiffies; |
5773 | unsigned long load; | 5777 | unsigned long load; |
5774 | 5778 | ||
5775 | if (cfs_rq->last_h_load_update == now) | 5779 | if (cfs_rq->last_h_load_update == now) |
5776 | return; | 5780 | return; |
5777 | 5781 | ||
5778 | cfs_rq->h_load_next = NULL; | 5782 | cfs_rq->h_load_next = NULL; |
5779 | for_each_sched_entity(se) { | 5783 | for_each_sched_entity(se) { |
5780 | cfs_rq = cfs_rq_of(se); | 5784 | cfs_rq = cfs_rq_of(se); |
5781 | cfs_rq->h_load_next = se; | 5785 | cfs_rq->h_load_next = se; |
5782 | if (cfs_rq->last_h_load_update == now) | 5786 | if (cfs_rq->last_h_load_update == now) |
5783 | break; | 5787 | break; |
5784 | } | 5788 | } |
5785 | 5789 | ||
5786 | if (!se) { | 5790 | if (!se) { |
5787 | cfs_rq->h_load = cfs_rq->runnable_load_avg; | 5791 | cfs_rq->h_load = cfs_rq->runnable_load_avg; |
5788 | cfs_rq->last_h_load_update = now; | 5792 | cfs_rq->last_h_load_update = now; |
5789 | } | 5793 | } |
5790 | 5794 | ||
5791 | while ((se = cfs_rq->h_load_next) != NULL) { | 5795 | while ((se = cfs_rq->h_load_next) != NULL) { |
5792 | load = cfs_rq->h_load; | 5796 | load = cfs_rq->h_load; |
5793 | load = div64_ul(load * se->avg.load_avg_contrib, | 5797 | load = div64_ul(load * se->avg.load_avg_contrib, |
5794 | cfs_rq->runnable_load_avg + 1); | 5798 | cfs_rq->runnable_load_avg + 1); |
5795 | cfs_rq = group_cfs_rq(se); | 5799 | cfs_rq = group_cfs_rq(se); |
5796 | cfs_rq->h_load = load; | 5800 | cfs_rq->h_load = load; |
5797 | cfs_rq->last_h_load_update = now; | 5801 | cfs_rq->last_h_load_update = now; |
5798 | } | 5802 | } |
5799 | } | 5803 | } |
5800 | 5804 | ||
5801 | static unsigned long task_h_load(struct task_struct *p) | 5805 | static unsigned long task_h_load(struct task_struct *p) |
5802 | { | 5806 | { |
5803 | struct cfs_rq *cfs_rq = task_cfs_rq(p); | 5807 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
5804 | 5808 | ||
5805 | update_cfs_rq_h_load(cfs_rq); | 5809 | update_cfs_rq_h_load(cfs_rq); |
5806 | return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load, | 5810 | return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load, |
5807 | cfs_rq->runnable_load_avg + 1); | 5811 | cfs_rq->runnable_load_avg + 1); |
5808 | } | 5812 | } |
5809 | #else | 5813 | #else |
5810 | static inline void update_blocked_averages(int cpu) | 5814 | static inline void update_blocked_averages(int cpu) |
5811 | { | 5815 | { |
5812 | } | 5816 | } |
5813 | 5817 | ||
5814 | static unsigned long task_h_load(struct task_struct *p) | 5818 | static unsigned long task_h_load(struct task_struct *p) |
5815 | { | 5819 | { |
5816 | return p->se.avg.load_avg_contrib; | 5820 | return p->se.avg.load_avg_contrib; |
5817 | } | 5821 | } |
5818 | #endif | 5822 | #endif |
5819 | 5823 | ||
5820 | /********** Helpers for find_busiest_group ************************/ | 5824 | /********** Helpers for find_busiest_group ************************/ |
5821 | 5825 | ||
5822 | enum group_type { | 5826 | enum group_type { |
5823 | group_other = 0, | 5827 | group_other = 0, |
5824 | group_imbalanced, | 5828 | group_imbalanced, |
5825 | group_overloaded, | 5829 | group_overloaded, |
5826 | }; | 5830 | }; |
5827 | 5831 | ||
5828 | /* | 5832 | /* |
5829 | * sg_lb_stats - stats of a sched_group required for load_balancing | 5833 | * sg_lb_stats - stats of a sched_group required for load_balancing |
5830 | */ | 5834 | */ |
5831 | struct sg_lb_stats { | 5835 | struct sg_lb_stats { |
5832 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | 5836 | unsigned long avg_load; /*Avg load across the CPUs of the group */ |
5833 | unsigned long group_load; /* Total load over the CPUs of the group */ | 5837 | unsigned long group_load; /* Total load over the CPUs of the group */ |
5834 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ | 5838 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ |
5835 | unsigned long load_per_task; | 5839 | unsigned long load_per_task; |
5836 | unsigned long group_capacity; | 5840 | unsigned long group_capacity; |
5837 | unsigned int sum_nr_running; /* Nr tasks running in the group */ | 5841 | unsigned int sum_nr_running; /* Nr tasks running in the group */ |
5838 | unsigned int group_capacity_factor; | 5842 | unsigned int group_capacity_factor; |
5839 | unsigned int idle_cpus; | 5843 | unsigned int idle_cpus; |
5840 | unsigned int group_weight; | 5844 | unsigned int group_weight; |
5841 | enum group_type group_type; | 5845 | enum group_type group_type; |
5842 | int group_has_free_capacity; | 5846 | int group_has_free_capacity; |
5843 | #ifdef CONFIG_NUMA_BALANCING | 5847 | #ifdef CONFIG_NUMA_BALANCING |
5844 | unsigned int nr_numa_running; | 5848 | unsigned int nr_numa_running; |
5845 | unsigned int nr_preferred_running; | 5849 | unsigned int nr_preferred_running; |
5846 | #endif | 5850 | #endif |
5847 | }; | 5851 | }; |
5848 | 5852 | ||
5849 | /* | 5853 | /* |
5850 | * sd_lb_stats - Structure to store the statistics of a sched_domain | 5854 | * sd_lb_stats - Structure to store the statistics of a sched_domain |
5851 | * during load balancing. | 5855 | * during load balancing. |
5852 | */ | 5856 | */ |
5853 | struct sd_lb_stats { | 5857 | struct sd_lb_stats { |
5854 | struct sched_group *busiest; /* Busiest group in this sd */ | 5858 | struct sched_group *busiest; /* Busiest group in this sd */ |
5855 | struct sched_group *local; /* Local group in this sd */ | 5859 | struct sched_group *local; /* Local group in this sd */ |
5856 | unsigned long total_load; /* Total load of all groups in sd */ | 5860 | unsigned long total_load; /* Total load of all groups in sd */ |
5857 | unsigned long total_capacity; /* Total capacity of all groups in sd */ | 5861 | unsigned long total_capacity; /* Total capacity of all groups in sd */ |
5858 | unsigned long avg_load; /* Average load across all groups in sd */ | 5862 | unsigned long avg_load; /* Average load across all groups in sd */ |
5859 | 5863 | ||
5860 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ | 5864 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
5861 | struct sg_lb_stats local_stat; /* Statistics of the local group */ | 5865 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
5862 | }; | 5866 | }; |
5863 | 5867 | ||
5864 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) | 5868 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
5865 | { | 5869 | { |
5866 | /* | 5870 | /* |
5867 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | 5871 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing |
5868 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | 5872 | * local_stat because update_sg_lb_stats() does a full clear/assignment. |
5869 | * We must however clear busiest_stat::avg_load because | 5873 | * We must however clear busiest_stat::avg_load because |
5870 | * update_sd_pick_busiest() reads this before assignment. | 5874 | * update_sd_pick_busiest() reads this before assignment. |
5871 | */ | 5875 | */ |
5872 | *sds = (struct sd_lb_stats){ | 5876 | *sds = (struct sd_lb_stats){ |
5873 | .busiest = NULL, | 5877 | .busiest = NULL, |
5874 | .local = NULL, | 5878 | .local = NULL, |
5875 | .total_load = 0UL, | 5879 | .total_load = 0UL, |
5876 | .total_capacity = 0UL, | 5880 | .total_capacity = 0UL, |
5877 | .busiest_stat = { | 5881 | .busiest_stat = { |
5878 | .avg_load = 0UL, | 5882 | .avg_load = 0UL, |
5879 | .sum_nr_running = 0, | 5883 | .sum_nr_running = 0, |
5880 | .group_type = group_other, | 5884 | .group_type = group_other, |
5881 | }, | 5885 | }, |
5882 | }; | 5886 | }; |
5883 | } | 5887 | } |
5884 | 5888 | ||
5885 | /** | 5889 | /** |
5886 | * get_sd_load_idx - Obtain the load index for a given sched domain. | 5890 | * get_sd_load_idx - Obtain the load index for a given sched domain. |
5887 | * @sd: The sched_domain whose load_idx is to be obtained. | 5891 | * @sd: The sched_domain whose load_idx is to be obtained. |
5888 | * @idle: The idle status of the CPU for whose sd load_idx is obtained. | 5892 | * @idle: The idle status of the CPU for whose sd load_idx is obtained. |
5889 | * | 5893 | * |
5890 | * Return: The load index. | 5894 | * Return: The load index. |
5891 | */ | 5895 | */ |
5892 | static inline int get_sd_load_idx(struct sched_domain *sd, | 5896 | static inline int get_sd_load_idx(struct sched_domain *sd, |
5893 | enum cpu_idle_type idle) | 5897 | enum cpu_idle_type idle) |
5894 | { | 5898 | { |
5895 | int load_idx; | 5899 | int load_idx; |
5896 | 5900 | ||
5897 | switch (idle) { | 5901 | switch (idle) { |
5898 | case CPU_NOT_IDLE: | 5902 | case CPU_NOT_IDLE: |
5899 | load_idx = sd->busy_idx; | 5903 | load_idx = sd->busy_idx; |
5900 | break; | 5904 | break; |
5901 | 5905 | ||
5902 | case CPU_NEWLY_IDLE: | 5906 | case CPU_NEWLY_IDLE: |
5903 | load_idx = sd->newidle_idx; | 5907 | load_idx = sd->newidle_idx; |
5904 | break; | 5908 | break; |
5905 | default: | 5909 | default: |
5906 | load_idx = sd->idle_idx; | 5910 | load_idx = sd->idle_idx; |
5907 | break; | 5911 | break; |
5908 | } | 5912 | } |
5909 | 5913 | ||
5910 | return load_idx; | 5914 | return load_idx; |
5911 | } | 5915 | } |
5912 | 5916 | ||
5913 | static unsigned long default_scale_capacity(struct sched_domain *sd, int cpu) | 5917 | static unsigned long default_scale_capacity(struct sched_domain *sd, int cpu) |
5914 | { | 5918 | { |
5915 | return SCHED_CAPACITY_SCALE; | 5919 | return SCHED_CAPACITY_SCALE; |
5916 | } | 5920 | } |
5917 | 5921 | ||
5918 | unsigned long __weak arch_scale_freq_capacity(struct sched_domain *sd, int cpu) | 5922 | unsigned long __weak arch_scale_freq_capacity(struct sched_domain *sd, int cpu) |
5919 | { | 5923 | { |
5920 | return default_scale_capacity(sd, cpu); | 5924 | return default_scale_capacity(sd, cpu); |
5921 | } | 5925 | } |
5922 | 5926 | ||
5923 | static unsigned long default_scale_cpu_capacity(struct sched_domain *sd, int cpu) | 5927 | static unsigned long default_scale_cpu_capacity(struct sched_domain *sd, int cpu) |
5924 | { | 5928 | { |
5925 | if ((sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1)) | 5929 | if ((sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1)) |
5926 | return sd->smt_gain / sd->span_weight; | 5930 | return sd->smt_gain / sd->span_weight; |
5927 | 5931 | ||
5928 | return SCHED_CAPACITY_SCALE; | 5932 | return SCHED_CAPACITY_SCALE; |
5929 | } | 5933 | } |
5930 | 5934 | ||
5931 | unsigned long __weak arch_scale_cpu_capacity(struct sched_domain *sd, int cpu) | 5935 | unsigned long __weak arch_scale_cpu_capacity(struct sched_domain *sd, int cpu) |
5932 | { | 5936 | { |
5933 | return default_scale_cpu_capacity(sd, cpu); | 5937 | return default_scale_cpu_capacity(sd, cpu); |
5934 | } | 5938 | } |
5935 | 5939 | ||
5936 | static unsigned long scale_rt_capacity(int cpu) | 5940 | static unsigned long scale_rt_capacity(int cpu) |
5937 | { | 5941 | { |
5938 | struct rq *rq = cpu_rq(cpu); | 5942 | struct rq *rq = cpu_rq(cpu); |
5939 | u64 total, available, age_stamp, avg; | 5943 | u64 total, available, age_stamp, avg; |
5940 | s64 delta; | 5944 | s64 delta; |
5941 | 5945 | ||
5942 | /* | 5946 | /* |
5943 | * Since we're reading these variables without serialization make sure | 5947 | * Since we're reading these variables without serialization make sure |
5944 | * we read them once before doing sanity checks on them. | 5948 | * we read them once before doing sanity checks on them. |
5945 | */ | 5949 | */ |
5946 | age_stamp = ACCESS_ONCE(rq->age_stamp); | 5950 | age_stamp = ACCESS_ONCE(rq->age_stamp); |
5947 | avg = ACCESS_ONCE(rq->rt_avg); | 5951 | avg = ACCESS_ONCE(rq->rt_avg); |
5948 | 5952 | ||
5949 | delta = rq_clock(rq) - age_stamp; | 5953 | delta = rq_clock(rq) - age_stamp; |
5950 | if (unlikely(delta < 0)) | 5954 | if (unlikely(delta < 0)) |
5951 | delta = 0; | 5955 | delta = 0; |
5952 | 5956 | ||
5953 | total = sched_avg_period() + delta; | 5957 | total = sched_avg_period() + delta; |
5954 | 5958 | ||
5955 | if (unlikely(total < avg)) { | 5959 | if (unlikely(total < avg)) { |
5956 | /* Ensures that capacity won't end up being negative */ | 5960 | /* Ensures that capacity won't end up being negative */ |
5957 | available = 0; | 5961 | available = 0; |
5958 | } else { | 5962 | } else { |
5959 | available = total - avg; | 5963 | available = total - avg; |
5960 | } | 5964 | } |
5961 | 5965 | ||
5962 | if (unlikely((s64)total < SCHED_CAPACITY_SCALE)) | 5966 | if (unlikely((s64)total < SCHED_CAPACITY_SCALE)) |
5963 | total = SCHED_CAPACITY_SCALE; | 5967 | total = SCHED_CAPACITY_SCALE; |
5964 | 5968 | ||
5965 | total >>= SCHED_CAPACITY_SHIFT; | 5969 | total >>= SCHED_CAPACITY_SHIFT; |
5966 | 5970 | ||
5967 | return div_u64(available, total); | 5971 | return div_u64(available, total); |
5968 | } | 5972 | } |
5969 | 5973 | ||
5970 | static void update_cpu_capacity(struct sched_domain *sd, int cpu) | 5974 | static void update_cpu_capacity(struct sched_domain *sd, int cpu) |
5971 | { | 5975 | { |
5972 | unsigned long capacity = SCHED_CAPACITY_SCALE; | 5976 | unsigned long capacity = SCHED_CAPACITY_SCALE; |
5973 | struct sched_group *sdg = sd->groups; | 5977 | struct sched_group *sdg = sd->groups; |
5974 | 5978 | ||
5975 | if (sched_feat(ARCH_CAPACITY)) | 5979 | if (sched_feat(ARCH_CAPACITY)) |
5976 | capacity *= arch_scale_cpu_capacity(sd, cpu); | 5980 | capacity *= arch_scale_cpu_capacity(sd, cpu); |
5977 | else | 5981 | else |
5978 | capacity *= default_scale_cpu_capacity(sd, cpu); | 5982 | capacity *= default_scale_cpu_capacity(sd, cpu); |
5979 | 5983 | ||
5980 | capacity >>= SCHED_CAPACITY_SHIFT; | 5984 | capacity >>= SCHED_CAPACITY_SHIFT; |
5981 | 5985 | ||
5982 | sdg->sgc->capacity_orig = capacity; | 5986 | sdg->sgc->capacity_orig = capacity; |
5983 | 5987 | ||
5984 | if (sched_feat(ARCH_CAPACITY)) | 5988 | if (sched_feat(ARCH_CAPACITY)) |
5985 | capacity *= arch_scale_freq_capacity(sd, cpu); | 5989 | capacity *= arch_scale_freq_capacity(sd, cpu); |
5986 | else | 5990 | else |
5987 | capacity *= default_scale_capacity(sd, cpu); | 5991 | capacity *= default_scale_capacity(sd, cpu); |
5988 | 5992 | ||
5989 | capacity >>= SCHED_CAPACITY_SHIFT; | 5993 | capacity >>= SCHED_CAPACITY_SHIFT; |
5990 | 5994 | ||
5991 | capacity *= scale_rt_capacity(cpu); | 5995 | capacity *= scale_rt_capacity(cpu); |
5992 | capacity >>= SCHED_CAPACITY_SHIFT; | 5996 | capacity >>= SCHED_CAPACITY_SHIFT; |
5993 | 5997 | ||
5994 | if (!capacity) | 5998 | if (!capacity) |
5995 | capacity = 1; | 5999 | capacity = 1; |
5996 | 6000 | ||
5997 | cpu_rq(cpu)->cpu_capacity = capacity; | 6001 | cpu_rq(cpu)->cpu_capacity = capacity; |
5998 | sdg->sgc->capacity = capacity; | 6002 | sdg->sgc->capacity = capacity; |
5999 | } | 6003 | } |
6000 | 6004 | ||
6001 | void update_group_capacity(struct sched_domain *sd, int cpu) | 6005 | void update_group_capacity(struct sched_domain *sd, int cpu) |
6002 | { | 6006 | { |
6003 | struct sched_domain *child = sd->child; | 6007 | struct sched_domain *child = sd->child; |
6004 | struct sched_group *group, *sdg = sd->groups; | 6008 | struct sched_group *group, *sdg = sd->groups; |
6005 | unsigned long capacity, capacity_orig; | 6009 | unsigned long capacity, capacity_orig; |
6006 | unsigned long interval; | 6010 | unsigned long interval; |
6007 | 6011 | ||
6008 | interval = msecs_to_jiffies(sd->balance_interval); | 6012 | interval = msecs_to_jiffies(sd->balance_interval); |
6009 | interval = clamp(interval, 1UL, max_load_balance_interval); | 6013 | interval = clamp(interval, 1UL, max_load_balance_interval); |
6010 | sdg->sgc->next_update = jiffies + interval; | 6014 | sdg->sgc->next_update = jiffies + interval; |
6011 | 6015 | ||
6012 | if (!child) { | 6016 | if (!child) { |
6013 | update_cpu_capacity(sd, cpu); | 6017 | update_cpu_capacity(sd, cpu); |
6014 | return; | 6018 | return; |
6015 | } | 6019 | } |
6016 | 6020 | ||
6017 | capacity_orig = capacity = 0; | 6021 | capacity_orig = capacity = 0; |
6018 | 6022 | ||
6019 | if (child->flags & SD_OVERLAP) { | 6023 | if (child->flags & SD_OVERLAP) { |
6020 | /* | 6024 | /* |
6021 | * SD_OVERLAP domains cannot assume that child groups | 6025 | * SD_OVERLAP domains cannot assume that child groups |
6022 | * span the current group. | 6026 | * span the current group. |
6023 | */ | 6027 | */ |
6024 | 6028 | ||
6025 | for_each_cpu(cpu, sched_group_cpus(sdg)) { | 6029 | for_each_cpu(cpu, sched_group_cpus(sdg)) { |
6026 | struct sched_group_capacity *sgc; | 6030 | struct sched_group_capacity *sgc; |
6027 | struct rq *rq = cpu_rq(cpu); | 6031 | struct rq *rq = cpu_rq(cpu); |
6028 | 6032 | ||
6029 | /* | 6033 | /* |
6030 | * build_sched_domains() -> init_sched_groups_capacity() | 6034 | * build_sched_domains() -> init_sched_groups_capacity() |
6031 | * gets here before we've attached the domains to the | 6035 | * gets here before we've attached the domains to the |
6032 | * runqueues. | 6036 | * runqueues. |
6033 | * | 6037 | * |
6034 | * Use capacity_of(), which is set irrespective of domains | 6038 | * Use capacity_of(), which is set irrespective of domains |
6035 | * in update_cpu_capacity(). | 6039 | * in update_cpu_capacity(). |
6036 | * | 6040 | * |
6037 | * This avoids capacity/capacity_orig from being 0 and | 6041 | * This avoids capacity/capacity_orig from being 0 and |
6038 | * causing divide-by-zero issues on boot. | 6042 | * causing divide-by-zero issues on boot. |
6039 | * | 6043 | * |
6040 | * Runtime updates will correct capacity_orig. | 6044 | * Runtime updates will correct capacity_orig. |
6041 | */ | 6045 | */ |
6042 | if (unlikely(!rq->sd)) { | 6046 | if (unlikely(!rq->sd)) { |
6043 | capacity_orig += capacity_of(cpu); | 6047 | capacity_orig += capacity_of(cpu); |
6044 | capacity += capacity_of(cpu); | 6048 | capacity += capacity_of(cpu); |
6045 | continue; | 6049 | continue; |
6046 | } | 6050 | } |
6047 | 6051 | ||
6048 | sgc = rq->sd->groups->sgc; | 6052 | sgc = rq->sd->groups->sgc; |
6049 | capacity_orig += sgc->capacity_orig; | 6053 | capacity_orig += sgc->capacity_orig; |
6050 | capacity += sgc->capacity; | 6054 | capacity += sgc->capacity; |
6051 | } | 6055 | } |
6052 | } else { | 6056 | } else { |
6053 | /* | 6057 | /* |
6054 | * !SD_OVERLAP domains can assume that child groups | 6058 | * !SD_OVERLAP domains can assume that child groups |
6055 | * span the current group. | 6059 | * span the current group. |
6056 | */ | 6060 | */ |
6057 | 6061 | ||
6058 | group = child->groups; | 6062 | group = child->groups; |
6059 | do { | 6063 | do { |
6060 | capacity_orig += group->sgc->capacity_orig; | 6064 | capacity_orig += group->sgc->capacity_orig; |
6061 | capacity += group->sgc->capacity; | 6065 | capacity += group->sgc->capacity; |
6062 | group = group->next; | 6066 | group = group->next; |
6063 | } while (group != child->groups); | 6067 | } while (group != child->groups); |
6064 | } | 6068 | } |
6065 | 6069 | ||
6066 | sdg->sgc->capacity_orig = capacity_orig; | 6070 | sdg->sgc->capacity_orig = capacity_orig; |
6067 | sdg->sgc->capacity = capacity; | 6071 | sdg->sgc->capacity = capacity; |
6068 | } | 6072 | } |
6069 | 6073 | ||
6070 | /* | 6074 | /* |
6071 | * Try and fix up capacity for tiny siblings, this is needed when | 6075 | * Try and fix up capacity for tiny siblings, this is needed when |
6072 | * things like SD_ASYM_PACKING need f_b_g to select another sibling | 6076 | * things like SD_ASYM_PACKING need f_b_g to select another sibling |
6073 | * which on its own isn't powerful enough. | 6077 | * which on its own isn't powerful enough. |
6074 | * | 6078 | * |
6075 | * See update_sd_pick_busiest() and check_asym_packing(). | 6079 | * See update_sd_pick_busiest() and check_asym_packing(). |
6076 | */ | 6080 | */ |
6077 | static inline int | 6081 | static inline int |
6078 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) | 6082 | fix_small_capacity(struct sched_domain *sd, struct sched_group *group) |
6079 | { | 6083 | { |
6080 | /* | 6084 | /* |
6081 | * Only siblings can have significantly less than SCHED_CAPACITY_SCALE | 6085 | * Only siblings can have significantly less than SCHED_CAPACITY_SCALE |
6082 | */ | 6086 | */ |
6083 | if (!(sd->flags & SD_SHARE_CPUCAPACITY)) | 6087 | if (!(sd->flags & SD_SHARE_CPUCAPACITY)) |
6084 | return 0; | 6088 | return 0; |
6085 | 6089 | ||
6086 | /* | 6090 | /* |
6087 | * If ~90% of the cpu_capacity is still there, we're good. | 6091 | * If ~90% of the cpu_capacity is still there, we're good. |
6088 | */ | 6092 | */ |
6089 | if (group->sgc->capacity * 32 > group->sgc->capacity_orig * 29) | 6093 | if (group->sgc->capacity * 32 > group->sgc->capacity_orig * 29) |
6090 | return 1; | 6094 | return 1; |
6091 | 6095 | ||
6092 | return 0; | 6096 | return 0; |
6093 | } | 6097 | } |
6094 | 6098 | ||
6095 | /* | 6099 | /* |
6096 | * Group imbalance indicates (and tries to solve) the problem where balancing | 6100 | * Group imbalance indicates (and tries to solve) the problem where balancing |
6097 | * groups is inadequate due to tsk_cpus_allowed() constraints. | 6101 | * groups is inadequate due to tsk_cpus_allowed() constraints. |
6098 | * | 6102 | * |
6099 | * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a | 6103 | * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a |
6100 | * cpumask covering 1 cpu of the first group and 3 cpus of the second group. | 6104 | * cpumask covering 1 cpu of the first group and 3 cpus of the second group. |
6101 | * Something like: | 6105 | * Something like: |
6102 | * | 6106 | * |
6103 | * { 0 1 2 3 } { 4 5 6 7 } | 6107 | * { 0 1 2 3 } { 4 5 6 7 } |
6104 | * * * * * | 6108 | * * * * * |
6105 | * | 6109 | * |
6106 | * If we were to balance group-wise we'd place two tasks in the first group and | 6110 | * If we were to balance group-wise we'd place two tasks in the first group and |
6107 | * two tasks in the second group. Clearly this is undesired as it will overload | 6111 | * two tasks in the second group. Clearly this is undesired as it will overload |
6108 | * cpu 3 and leave one of the cpus in the second group unused. | 6112 | * cpu 3 and leave one of the cpus in the second group unused. |
6109 | * | 6113 | * |
6110 | * The current solution to this issue is detecting the skew in the first group | 6114 | * The current solution to this issue is detecting the skew in the first group |
6111 | * by noticing the lower domain failed to reach balance and had difficulty | 6115 | * by noticing the lower domain failed to reach balance and had difficulty |
6112 | * moving tasks due to affinity constraints. | 6116 | * moving tasks due to affinity constraints. |
6113 | * | 6117 | * |
6114 | * When this is so detected; this group becomes a candidate for busiest; see | 6118 | * When this is so detected; this group becomes a candidate for busiest; see |
6115 | * update_sd_pick_busiest(). And calculate_imbalance() and | 6119 | * update_sd_pick_busiest(). And calculate_imbalance() and |
6116 | * find_busiest_group() avoid some of the usual balance conditions to allow it | 6120 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
6117 | * to create an effective group imbalance. | 6121 | * to create an effective group imbalance. |
6118 | * | 6122 | * |
6119 | * This is a somewhat tricky proposition since the next run might not find the | 6123 | * This is a somewhat tricky proposition since the next run might not find the |
6120 | * group imbalance and decide the groups need to be balanced again. A most | 6124 | * group imbalance and decide the groups need to be balanced again. A most |
6121 | * subtle and fragile situation. | 6125 | * subtle and fragile situation. |
6122 | */ | 6126 | */ |
6123 | 6127 | ||
6124 | static inline int sg_imbalanced(struct sched_group *group) | 6128 | static inline int sg_imbalanced(struct sched_group *group) |
6125 | { | 6129 | { |
6126 | return group->sgc->imbalance; | 6130 | return group->sgc->imbalance; |
6127 | } | 6131 | } |
6128 | 6132 | ||
6129 | /* | 6133 | /* |
6130 | * Compute the group capacity factor. | 6134 | * Compute the group capacity factor. |
6131 | * | 6135 | * |
6132 | * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by | 6136 | * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by |
6133 | * first dividing out the smt factor and computing the actual number of cores | 6137 | * first dividing out the smt factor and computing the actual number of cores |
6134 | * and limit unit capacity with that. | 6138 | * and limit unit capacity with that. |
6135 | */ | 6139 | */ |
6136 | static inline int sg_capacity_factor(struct lb_env *env, struct sched_group *group) | 6140 | static inline int sg_capacity_factor(struct lb_env *env, struct sched_group *group) |
6137 | { | 6141 | { |
6138 | unsigned int capacity_factor, smt, cpus; | 6142 | unsigned int capacity_factor, smt, cpus; |
6139 | unsigned int capacity, capacity_orig; | 6143 | unsigned int capacity, capacity_orig; |
6140 | 6144 | ||
6141 | capacity = group->sgc->capacity; | 6145 | capacity = group->sgc->capacity; |
6142 | capacity_orig = group->sgc->capacity_orig; | 6146 | capacity_orig = group->sgc->capacity_orig; |
6143 | cpus = group->group_weight; | 6147 | cpus = group->group_weight; |
6144 | 6148 | ||
6145 | /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */ | 6149 | /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */ |
6146 | smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, capacity_orig); | 6150 | smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, capacity_orig); |
6147 | capacity_factor = cpus / smt; /* cores */ | 6151 | capacity_factor = cpus / smt; /* cores */ |
6148 | 6152 | ||
6149 | capacity_factor = min_t(unsigned, | 6153 | capacity_factor = min_t(unsigned, |
6150 | capacity_factor, DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE)); | 6154 | capacity_factor, DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE)); |
6151 | if (!capacity_factor) | 6155 | if (!capacity_factor) |
6152 | capacity_factor = fix_small_capacity(env->sd, group); | 6156 | capacity_factor = fix_small_capacity(env->sd, group); |
6153 | 6157 | ||
6154 | return capacity_factor; | 6158 | return capacity_factor; |
6155 | } | 6159 | } |
6156 | 6160 | ||
6157 | static enum group_type | 6161 | static enum group_type |
6158 | group_classify(struct sched_group *group, struct sg_lb_stats *sgs) | 6162 | group_classify(struct sched_group *group, struct sg_lb_stats *sgs) |
6159 | { | 6163 | { |
6160 | if (sgs->sum_nr_running > sgs->group_capacity_factor) | 6164 | if (sgs->sum_nr_running > sgs->group_capacity_factor) |
6161 | return group_overloaded; | 6165 | return group_overloaded; |
6162 | 6166 | ||
6163 | if (sg_imbalanced(group)) | 6167 | if (sg_imbalanced(group)) |
6164 | return group_imbalanced; | 6168 | return group_imbalanced; |
6165 | 6169 | ||
6166 | return group_other; | 6170 | return group_other; |
6167 | } | 6171 | } |
6168 | 6172 | ||
6169 | /** | 6173 | /** |
6170 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | 6174 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. |
6171 | * @env: The load balancing environment. | 6175 | * @env: The load balancing environment. |
6172 | * @group: sched_group whose statistics are to be updated. | 6176 | * @group: sched_group whose statistics are to be updated. |
6173 | * @load_idx: Load index of sched_domain of this_cpu for load calc. | 6177 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
6174 | * @local_group: Does group contain this_cpu. | 6178 | * @local_group: Does group contain this_cpu. |
6175 | * @sgs: variable to hold the statistics for this group. | 6179 | * @sgs: variable to hold the statistics for this group. |
6176 | * @overload: Indicate more than one runnable task for any CPU. | 6180 | * @overload: Indicate more than one runnable task for any CPU. |
6177 | */ | 6181 | */ |
6178 | static inline void update_sg_lb_stats(struct lb_env *env, | 6182 | static inline void update_sg_lb_stats(struct lb_env *env, |
6179 | struct sched_group *group, int load_idx, | 6183 | struct sched_group *group, int load_idx, |
6180 | int local_group, struct sg_lb_stats *sgs, | 6184 | int local_group, struct sg_lb_stats *sgs, |
6181 | bool *overload) | 6185 | bool *overload) |
6182 | { | 6186 | { |
6183 | unsigned long load; | 6187 | unsigned long load; |
6184 | int i; | 6188 | int i; |
6185 | 6189 | ||
6186 | memset(sgs, 0, sizeof(*sgs)); | 6190 | memset(sgs, 0, sizeof(*sgs)); |
6187 | 6191 | ||
6188 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { | 6192 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
6189 | struct rq *rq = cpu_rq(i); | 6193 | struct rq *rq = cpu_rq(i); |
6190 | 6194 | ||
6191 | /* Bias balancing toward cpus of our domain */ | 6195 | /* Bias balancing toward cpus of our domain */ |
6192 | if (local_group) | 6196 | if (local_group) |
6193 | load = target_load(i, load_idx); | 6197 | load = target_load(i, load_idx); |
6194 | else | 6198 | else |
6195 | load = source_load(i, load_idx); | 6199 | load = source_load(i, load_idx); |
6196 | 6200 | ||
6197 | sgs->group_load += load; | 6201 | sgs->group_load += load; |
6198 | sgs->sum_nr_running += rq->cfs.h_nr_running; | 6202 | sgs->sum_nr_running += rq->cfs.h_nr_running; |
6199 | 6203 | ||
6200 | if (rq->nr_running > 1) | 6204 | if (rq->nr_running > 1) |
6201 | *overload = true; | 6205 | *overload = true; |
6202 | 6206 | ||
6203 | #ifdef CONFIG_NUMA_BALANCING | 6207 | #ifdef CONFIG_NUMA_BALANCING |
6204 | sgs->nr_numa_running += rq->nr_numa_running; | 6208 | sgs->nr_numa_running += rq->nr_numa_running; |
6205 | sgs->nr_preferred_running += rq->nr_preferred_running; | 6209 | sgs->nr_preferred_running += rq->nr_preferred_running; |
6206 | #endif | 6210 | #endif |
6207 | sgs->sum_weighted_load += weighted_cpuload(i); | 6211 | sgs->sum_weighted_load += weighted_cpuload(i); |
6208 | if (idle_cpu(i)) | 6212 | if (idle_cpu(i)) |
6209 | sgs->idle_cpus++; | 6213 | sgs->idle_cpus++; |
6210 | } | 6214 | } |
6211 | 6215 | ||
6212 | /* Adjust by relative CPU capacity of the group */ | 6216 | /* Adjust by relative CPU capacity of the group */ |
6213 | sgs->group_capacity = group->sgc->capacity; | 6217 | sgs->group_capacity = group->sgc->capacity; |
6214 | sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity; | 6218 | sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity; |
6215 | 6219 | ||
6216 | if (sgs->sum_nr_running) | 6220 | if (sgs->sum_nr_running) |
6217 | sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; | 6221 | sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; |
6218 | 6222 | ||
6219 | sgs->group_weight = group->group_weight; | 6223 | sgs->group_weight = group->group_weight; |
6220 | sgs->group_capacity_factor = sg_capacity_factor(env, group); | 6224 | sgs->group_capacity_factor = sg_capacity_factor(env, group); |
6221 | sgs->group_type = group_classify(group, sgs); | 6225 | sgs->group_type = group_classify(group, sgs); |
6222 | 6226 | ||
6223 | if (sgs->group_capacity_factor > sgs->sum_nr_running) | 6227 | if (sgs->group_capacity_factor > sgs->sum_nr_running) |
6224 | sgs->group_has_free_capacity = 1; | 6228 | sgs->group_has_free_capacity = 1; |
6225 | } | 6229 | } |
6226 | 6230 | ||
6227 | /** | 6231 | /** |
6228 | * update_sd_pick_busiest - return 1 on busiest group | 6232 | * update_sd_pick_busiest - return 1 on busiest group |
6229 | * @env: The load balancing environment. | 6233 | * @env: The load balancing environment. |
6230 | * @sds: sched_domain statistics | 6234 | * @sds: sched_domain statistics |
6231 | * @sg: sched_group candidate to be checked for being the busiest | 6235 | * @sg: sched_group candidate to be checked for being the busiest |
6232 | * @sgs: sched_group statistics | 6236 | * @sgs: sched_group statistics |
6233 | * | 6237 | * |
6234 | * Determine if @sg is a busier group than the previously selected | 6238 | * Determine if @sg is a busier group than the previously selected |
6235 | * busiest group. | 6239 | * busiest group. |
6236 | * | 6240 | * |
6237 | * Return: %true if @sg is a busier group than the previously selected | 6241 | * Return: %true if @sg is a busier group than the previously selected |
6238 | * busiest group. %false otherwise. | 6242 | * busiest group. %false otherwise. |
6239 | */ | 6243 | */ |
6240 | static bool update_sd_pick_busiest(struct lb_env *env, | 6244 | static bool update_sd_pick_busiest(struct lb_env *env, |
6241 | struct sd_lb_stats *sds, | 6245 | struct sd_lb_stats *sds, |
6242 | struct sched_group *sg, | 6246 | struct sched_group *sg, |
6243 | struct sg_lb_stats *sgs) | 6247 | struct sg_lb_stats *sgs) |
6244 | { | 6248 | { |
6245 | struct sg_lb_stats *busiest = &sds->busiest_stat; | 6249 | struct sg_lb_stats *busiest = &sds->busiest_stat; |
6246 | 6250 | ||
6247 | if (sgs->group_type > busiest->group_type) | 6251 | if (sgs->group_type > busiest->group_type) |
6248 | return true; | 6252 | return true; |
6249 | 6253 | ||
6250 | if (sgs->group_type < busiest->group_type) | 6254 | if (sgs->group_type < busiest->group_type) |
6251 | return false; | 6255 | return false; |
6252 | 6256 | ||
6253 | if (sgs->avg_load <= busiest->avg_load) | 6257 | if (sgs->avg_load <= busiest->avg_load) |
6254 | return false; | 6258 | return false; |
6255 | 6259 | ||
6256 | /* This is the busiest node in its class. */ | 6260 | /* This is the busiest node in its class. */ |
6257 | if (!(env->sd->flags & SD_ASYM_PACKING)) | 6261 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
6258 | return true; | 6262 | return true; |
6259 | 6263 | ||
6260 | /* | 6264 | /* |
6261 | * ASYM_PACKING needs to move all the work to the lowest | 6265 | * ASYM_PACKING needs to move all the work to the lowest |
6262 | * numbered CPUs in the group, therefore mark all groups | 6266 | * numbered CPUs in the group, therefore mark all groups |
6263 | * higher than ourself as busy. | 6267 | * higher than ourself as busy. |
6264 | */ | 6268 | */ |
6265 | if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) { | 6269 | if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) { |
6266 | if (!sds->busiest) | 6270 | if (!sds->busiest) |
6267 | return true; | 6271 | return true; |
6268 | 6272 | ||
6269 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) | 6273 | if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) |
6270 | return true; | 6274 | return true; |
6271 | } | 6275 | } |
6272 | 6276 | ||
6273 | return false; | 6277 | return false; |
6274 | } | 6278 | } |
6275 | 6279 | ||
6276 | #ifdef CONFIG_NUMA_BALANCING | 6280 | #ifdef CONFIG_NUMA_BALANCING |
6277 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | 6281 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) |
6278 | { | 6282 | { |
6279 | if (sgs->sum_nr_running > sgs->nr_numa_running) | 6283 | if (sgs->sum_nr_running > sgs->nr_numa_running) |
6280 | return regular; | 6284 | return regular; |
6281 | if (sgs->sum_nr_running > sgs->nr_preferred_running) | 6285 | if (sgs->sum_nr_running > sgs->nr_preferred_running) |
6282 | return remote; | 6286 | return remote; |
6283 | return all; | 6287 | return all; |
6284 | } | 6288 | } |
6285 | 6289 | ||
6286 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | 6290 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) |
6287 | { | 6291 | { |
6288 | if (rq->nr_running > rq->nr_numa_running) | 6292 | if (rq->nr_running > rq->nr_numa_running) |
6289 | return regular; | 6293 | return regular; |
6290 | if (rq->nr_running > rq->nr_preferred_running) | 6294 | if (rq->nr_running > rq->nr_preferred_running) |
6291 | return remote; | 6295 | return remote; |
6292 | return all; | 6296 | return all; |
6293 | } | 6297 | } |
6294 | #else | 6298 | #else |
6295 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | 6299 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) |
6296 | { | 6300 | { |
6297 | return all; | 6301 | return all; |
6298 | } | 6302 | } |
6299 | 6303 | ||
6300 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | 6304 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) |
6301 | { | 6305 | { |
6302 | return regular; | 6306 | return regular; |
6303 | } | 6307 | } |
6304 | #endif /* CONFIG_NUMA_BALANCING */ | 6308 | #endif /* CONFIG_NUMA_BALANCING */ |
6305 | 6309 | ||
6306 | /** | 6310 | /** |
6307 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. | 6311 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
6308 | * @env: The load balancing environment. | 6312 | * @env: The load balancing environment. |
6309 | * @sds: variable to hold the statistics for this sched_domain. | 6313 | * @sds: variable to hold the statistics for this sched_domain. |
6310 | */ | 6314 | */ |
6311 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) | 6315 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) |
6312 | { | 6316 | { |
6313 | struct sched_domain *child = env->sd->child; | 6317 | struct sched_domain *child = env->sd->child; |
6314 | struct sched_group *sg = env->sd->groups; | 6318 | struct sched_group *sg = env->sd->groups; |
6315 | struct sg_lb_stats tmp_sgs; | 6319 | struct sg_lb_stats tmp_sgs; |
6316 | int load_idx, prefer_sibling = 0; | 6320 | int load_idx, prefer_sibling = 0; |
6317 | bool overload = false; | 6321 | bool overload = false; |
6318 | 6322 | ||
6319 | if (child && child->flags & SD_PREFER_SIBLING) | 6323 | if (child && child->flags & SD_PREFER_SIBLING) |
6320 | prefer_sibling = 1; | 6324 | prefer_sibling = 1; |
6321 | 6325 | ||
6322 | load_idx = get_sd_load_idx(env->sd, env->idle); | 6326 | load_idx = get_sd_load_idx(env->sd, env->idle); |
6323 | 6327 | ||
6324 | do { | 6328 | do { |
6325 | struct sg_lb_stats *sgs = &tmp_sgs; | 6329 | struct sg_lb_stats *sgs = &tmp_sgs; |
6326 | int local_group; | 6330 | int local_group; |
6327 | 6331 | ||
6328 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); | 6332 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); |
6329 | if (local_group) { | 6333 | if (local_group) { |
6330 | sds->local = sg; | 6334 | sds->local = sg; |
6331 | sgs = &sds->local_stat; | 6335 | sgs = &sds->local_stat; |
6332 | 6336 | ||
6333 | if (env->idle != CPU_NEWLY_IDLE || | 6337 | if (env->idle != CPU_NEWLY_IDLE || |
6334 | time_after_eq(jiffies, sg->sgc->next_update)) | 6338 | time_after_eq(jiffies, sg->sgc->next_update)) |
6335 | update_group_capacity(env->sd, env->dst_cpu); | 6339 | update_group_capacity(env->sd, env->dst_cpu); |
6336 | } | 6340 | } |
6337 | 6341 | ||
6338 | update_sg_lb_stats(env, sg, load_idx, local_group, sgs, | 6342 | update_sg_lb_stats(env, sg, load_idx, local_group, sgs, |
6339 | &overload); | 6343 | &overload); |
6340 | 6344 | ||
6341 | if (local_group) | 6345 | if (local_group) |
6342 | goto next_group; | 6346 | goto next_group; |
6343 | 6347 | ||
6344 | /* | 6348 | /* |
6345 | * In case the child domain prefers tasks go to siblings | 6349 | * In case the child domain prefers tasks go to siblings |
6346 | * first, lower the sg capacity factor to one so that we'll try | 6350 | * first, lower the sg capacity factor to one so that we'll try |
6347 | * and move all the excess tasks away. We lower the capacity | 6351 | * and move all the excess tasks away. We lower the capacity |
6348 | * of a group only if the local group has the capacity to fit | 6352 | * of a group only if the local group has the capacity to fit |
6349 | * these excess tasks, i.e. nr_running < group_capacity_factor. The | 6353 | * these excess tasks, i.e. nr_running < group_capacity_factor. The |
6350 | * extra check prevents the case where you always pull from the | 6354 | * extra check prevents the case where you always pull from the |
6351 | * heaviest group when it is already under-utilized (possible | 6355 | * heaviest group when it is already under-utilized (possible |
6352 | * with a large weight task outweighs the tasks on the system). | 6356 | * with a large weight task outweighs the tasks on the system). |
6353 | */ | 6357 | */ |
6354 | if (prefer_sibling && sds->local && | 6358 | if (prefer_sibling && sds->local && |
6355 | sds->local_stat.group_has_free_capacity) { | 6359 | sds->local_stat.group_has_free_capacity) { |
6356 | sgs->group_capacity_factor = min(sgs->group_capacity_factor, 1U); | 6360 | sgs->group_capacity_factor = min(sgs->group_capacity_factor, 1U); |
6357 | sgs->group_type = group_classify(sg, sgs); | 6361 | sgs->group_type = group_classify(sg, sgs); |
6358 | } | 6362 | } |
6359 | 6363 | ||
6360 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { | 6364 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
6361 | sds->busiest = sg; | 6365 | sds->busiest = sg; |
6362 | sds->busiest_stat = *sgs; | 6366 | sds->busiest_stat = *sgs; |
6363 | } | 6367 | } |
6364 | 6368 | ||
6365 | next_group: | 6369 | next_group: |
6366 | /* Now, start updating sd_lb_stats */ | 6370 | /* Now, start updating sd_lb_stats */ |
6367 | sds->total_load += sgs->group_load; | 6371 | sds->total_load += sgs->group_load; |
6368 | sds->total_capacity += sgs->group_capacity; | 6372 | sds->total_capacity += sgs->group_capacity; |
6369 | 6373 | ||
6370 | sg = sg->next; | 6374 | sg = sg->next; |
6371 | } while (sg != env->sd->groups); | 6375 | } while (sg != env->sd->groups); |
6372 | 6376 | ||
6373 | if (env->sd->flags & SD_NUMA) | 6377 | if (env->sd->flags & SD_NUMA) |
6374 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); | 6378 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); |
6375 | 6379 | ||
6376 | if (!env->sd->parent) { | 6380 | if (!env->sd->parent) { |
6377 | /* update overload indicator if we are at root domain */ | 6381 | /* update overload indicator if we are at root domain */ |
6378 | if (env->dst_rq->rd->overload != overload) | 6382 | if (env->dst_rq->rd->overload != overload) |
6379 | env->dst_rq->rd->overload = overload; | 6383 | env->dst_rq->rd->overload = overload; |
6380 | } | 6384 | } |
6381 | 6385 | ||
6382 | } | 6386 | } |
6383 | 6387 | ||
6384 | /** | 6388 | /** |
6385 | * check_asym_packing - Check to see if the group is packed into the | 6389 | * check_asym_packing - Check to see if the group is packed into the |
6386 | * sched doman. | 6390 | * sched doman. |
6387 | * | 6391 | * |
6388 | * This is primarily intended to used at the sibling level. Some | 6392 | * This is primarily intended to used at the sibling level. Some |
6389 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | 6393 | * cores like POWER7 prefer to use lower numbered SMT threads. In the |
6390 | * case of POWER7, it can move to lower SMT modes only when higher | 6394 | * case of POWER7, it can move to lower SMT modes only when higher |
6391 | * threads are idle. When in lower SMT modes, the threads will | 6395 | * threads are idle. When in lower SMT modes, the threads will |
6392 | * perform better since they share less core resources. Hence when we | 6396 | * perform better since they share less core resources. Hence when we |
6393 | * have idle threads, we want them to be the higher ones. | 6397 | * have idle threads, we want them to be the higher ones. |
6394 | * | 6398 | * |
6395 | * This packing function is run on idle threads. It checks to see if | 6399 | * This packing function is run on idle threads. It checks to see if |
6396 | * the busiest CPU in this domain (core in the P7 case) has a higher | 6400 | * the busiest CPU in this domain (core in the P7 case) has a higher |
6397 | * CPU number than the packing function is being run on. Here we are | 6401 | * CPU number than the packing function is being run on. Here we are |
6398 | * assuming lower CPU number will be equivalent to lower a SMT thread | 6402 | * assuming lower CPU number will be equivalent to lower a SMT thread |
6399 | * number. | 6403 | * number. |
6400 | * | 6404 | * |
6401 | * Return: 1 when packing is required and a task should be moved to | 6405 | * Return: 1 when packing is required and a task should be moved to |
6402 | * this CPU. The amount of the imbalance is returned in *imbalance. | 6406 | * this CPU. The amount of the imbalance is returned in *imbalance. |
6403 | * | 6407 | * |
6404 | * @env: The load balancing environment. | 6408 | * @env: The load balancing environment. |
6405 | * @sds: Statistics of the sched_domain which is to be packed | 6409 | * @sds: Statistics of the sched_domain which is to be packed |
6406 | */ | 6410 | */ |
6407 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) | 6411 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
6408 | { | 6412 | { |
6409 | int busiest_cpu; | 6413 | int busiest_cpu; |
6410 | 6414 | ||
6411 | if (!(env->sd->flags & SD_ASYM_PACKING)) | 6415 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
6412 | return 0; | 6416 | return 0; |
6413 | 6417 | ||
6414 | if (!sds->busiest) | 6418 | if (!sds->busiest) |
6415 | return 0; | 6419 | return 0; |
6416 | 6420 | ||
6417 | busiest_cpu = group_first_cpu(sds->busiest); | 6421 | busiest_cpu = group_first_cpu(sds->busiest); |
6418 | if (env->dst_cpu > busiest_cpu) | 6422 | if (env->dst_cpu > busiest_cpu) |
6419 | return 0; | 6423 | return 0; |
6420 | 6424 | ||
6421 | env->imbalance = DIV_ROUND_CLOSEST( | 6425 | env->imbalance = DIV_ROUND_CLOSEST( |
6422 | sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity, | 6426 | sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity, |
6423 | SCHED_CAPACITY_SCALE); | 6427 | SCHED_CAPACITY_SCALE); |
6424 | 6428 | ||
6425 | return 1; | 6429 | return 1; |
6426 | } | 6430 | } |
6427 | 6431 | ||
6428 | /** | 6432 | /** |
6429 | * fix_small_imbalance - Calculate the minor imbalance that exists | 6433 | * fix_small_imbalance - Calculate the minor imbalance that exists |
6430 | * amongst the groups of a sched_domain, during | 6434 | * amongst the groups of a sched_domain, during |
6431 | * load balancing. | 6435 | * load balancing. |
6432 | * @env: The load balancing environment. | 6436 | * @env: The load balancing environment. |
6433 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. | 6437 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
6434 | */ | 6438 | */ |
6435 | static inline | 6439 | static inline |
6436 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | 6440 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
6437 | { | 6441 | { |
6438 | unsigned long tmp, capa_now = 0, capa_move = 0; | 6442 | unsigned long tmp, capa_now = 0, capa_move = 0; |
6439 | unsigned int imbn = 2; | 6443 | unsigned int imbn = 2; |
6440 | unsigned long scaled_busy_load_per_task; | 6444 | unsigned long scaled_busy_load_per_task; |
6441 | struct sg_lb_stats *local, *busiest; | 6445 | struct sg_lb_stats *local, *busiest; |
6442 | 6446 | ||
6443 | local = &sds->local_stat; | 6447 | local = &sds->local_stat; |
6444 | busiest = &sds->busiest_stat; | 6448 | busiest = &sds->busiest_stat; |
6445 | 6449 | ||
6446 | if (!local->sum_nr_running) | 6450 | if (!local->sum_nr_running) |
6447 | local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); | 6451 | local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); |
6448 | else if (busiest->load_per_task > local->load_per_task) | 6452 | else if (busiest->load_per_task > local->load_per_task) |
6449 | imbn = 1; | 6453 | imbn = 1; |
6450 | 6454 | ||
6451 | scaled_busy_load_per_task = | 6455 | scaled_busy_load_per_task = |
6452 | (busiest->load_per_task * SCHED_CAPACITY_SCALE) / | 6456 | (busiest->load_per_task * SCHED_CAPACITY_SCALE) / |
6453 | busiest->group_capacity; | 6457 | busiest->group_capacity; |
6454 | 6458 | ||
6455 | if (busiest->avg_load + scaled_busy_load_per_task >= | 6459 | if (busiest->avg_load + scaled_busy_load_per_task >= |
6456 | local->avg_load + (scaled_busy_load_per_task * imbn)) { | 6460 | local->avg_load + (scaled_busy_load_per_task * imbn)) { |
6457 | env->imbalance = busiest->load_per_task; | 6461 | env->imbalance = busiest->load_per_task; |
6458 | return; | 6462 | return; |
6459 | } | 6463 | } |
6460 | 6464 | ||
6461 | /* | 6465 | /* |
6462 | * OK, we don't have enough imbalance to justify moving tasks, | 6466 | * OK, we don't have enough imbalance to justify moving tasks, |
6463 | * however we may be able to increase total CPU capacity used by | 6467 | * however we may be able to increase total CPU capacity used by |
6464 | * moving them. | 6468 | * moving them. |
6465 | */ | 6469 | */ |
6466 | 6470 | ||
6467 | capa_now += busiest->group_capacity * | 6471 | capa_now += busiest->group_capacity * |
6468 | min(busiest->load_per_task, busiest->avg_load); | 6472 | min(busiest->load_per_task, busiest->avg_load); |
6469 | capa_now += local->group_capacity * | 6473 | capa_now += local->group_capacity * |
6470 | min(local->load_per_task, local->avg_load); | 6474 | min(local->load_per_task, local->avg_load); |
6471 | capa_now /= SCHED_CAPACITY_SCALE; | 6475 | capa_now /= SCHED_CAPACITY_SCALE; |
6472 | 6476 | ||
6473 | /* Amount of load we'd subtract */ | 6477 | /* Amount of load we'd subtract */ |
6474 | if (busiest->avg_load > scaled_busy_load_per_task) { | 6478 | if (busiest->avg_load > scaled_busy_load_per_task) { |
6475 | capa_move += busiest->group_capacity * | 6479 | capa_move += busiest->group_capacity * |
6476 | min(busiest->load_per_task, | 6480 | min(busiest->load_per_task, |
6477 | busiest->avg_load - scaled_busy_load_per_task); | 6481 | busiest->avg_load - scaled_busy_load_per_task); |
6478 | } | 6482 | } |
6479 | 6483 | ||
6480 | /* Amount of load we'd add */ | 6484 | /* Amount of load we'd add */ |
6481 | if (busiest->avg_load * busiest->group_capacity < | 6485 | if (busiest->avg_load * busiest->group_capacity < |
6482 | busiest->load_per_task * SCHED_CAPACITY_SCALE) { | 6486 | busiest->load_per_task * SCHED_CAPACITY_SCALE) { |
6483 | tmp = (busiest->avg_load * busiest->group_capacity) / | 6487 | tmp = (busiest->avg_load * busiest->group_capacity) / |
6484 | local->group_capacity; | 6488 | local->group_capacity; |
6485 | } else { | 6489 | } else { |
6486 | tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) / | 6490 | tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) / |
6487 | local->group_capacity; | 6491 | local->group_capacity; |
6488 | } | 6492 | } |
6489 | capa_move += local->group_capacity * | 6493 | capa_move += local->group_capacity * |
6490 | min(local->load_per_task, local->avg_load + tmp); | 6494 | min(local->load_per_task, local->avg_load + tmp); |
6491 | capa_move /= SCHED_CAPACITY_SCALE; | 6495 | capa_move /= SCHED_CAPACITY_SCALE; |
6492 | 6496 | ||
6493 | /* Move if we gain throughput */ | 6497 | /* Move if we gain throughput */ |
6494 | if (capa_move > capa_now) | 6498 | if (capa_move > capa_now) |
6495 | env->imbalance = busiest->load_per_task; | 6499 | env->imbalance = busiest->load_per_task; |
6496 | } | 6500 | } |
6497 | 6501 | ||
6498 | /** | 6502 | /** |
6499 | * calculate_imbalance - Calculate the amount of imbalance present within the | 6503 | * calculate_imbalance - Calculate the amount of imbalance present within the |
6500 | * groups of a given sched_domain during load balance. | 6504 | * groups of a given sched_domain during load balance. |
6501 | * @env: load balance environment | 6505 | * @env: load balance environment |
6502 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. | 6506 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
6503 | */ | 6507 | */ |
6504 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | 6508 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
6505 | { | 6509 | { |
6506 | unsigned long max_pull, load_above_capacity = ~0UL; | 6510 | unsigned long max_pull, load_above_capacity = ~0UL; |
6507 | struct sg_lb_stats *local, *busiest; | 6511 | struct sg_lb_stats *local, *busiest; |
6508 | 6512 | ||
6509 | local = &sds->local_stat; | 6513 | local = &sds->local_stat; |
6510 | busiest = &sds->busiest_stat; | 6514 | busiest = &sds->busiest_stat; |
6511 | 6515 | ||
6512 | if (busiest->group_type == group_imbalanced) { | 6516 | if (busiest->group_type == group_imbalanced) { |
6513 | /* | 6517 | /* |
6514 | * In the group_imb case we cannot rely on group-wide averages | 6518 | * In the group_imb case we cannot rely on group-wide averages |
6515 | * to ensure cpu-load equilibrium, look at wider averages. XXX | 6519 | * to ensure cpu-load equilibrium, look at wider averages. XXX |
6516 | */ | 6520 | */ |
6517 | busiest->load_per_task = | 6521 | busiest->load_per_task = |
6518 | min(busiest->load_per_task, sds->avg_load); | 6522 | min(busiest->load_per_task, sds->avg_load); |
6519 | } | 6523 | } |
6520 | 6524 | ||
6521 | /* | 6525 | /* |
6522 | * In the presence of smp nice balancing, certain scenarios can have | 6526 | * In the presence of smp nice balancing, certain scenarios can have |
6523 | * max load less than avg load(as we skip the groups at or below | 6527 | * max load less than avg load(as we skip the groups at or below |
6524 | * its cpu_capacity, while calculating max_load..) | 6528 | * its cpu_capacity, while calculating max_load..) |
6525 | */ | 6529 | */ |
6526 | if (busiest->avg_load <= sds->avg_load || | 6530 | if (busiest->avg_load <= sds->avg_load || |
6527 | local->avg_load >= sds->avg_load) { | 6531 | local->avg_load >= sds->avg_load) { |
6528 | env->imbalance = 0; | 6532 | env->imbalance = 0; |
6529 | return fix_small_imbalance(env, sds); | 6533 | return fix_small_imbalance(env, sds); |
6530 | } | 6534 | } |
6531 | 6535 | ||
6532 | /* | 6536 | /* |
6533 | * If there aren't any idle cpus, avoid creating some. | 6537 | * If there aren't any idle cpus, avoid creating some. |
6534 | */ | 6538 | */ |
6535 | if (busiest->group_type == group_overloaded && | 6539 | if (busiest->group_type == group_overloaded && |
6536 | local->group_type == group_overloaded) { | 6540 | local->group_type == group_overloaded) { |
6537 | load_above_capacity = | 6541 | load_above_capacity = |
6538 | (busiest->sum_nr_running - busiest->group_capacity_factor); | 6542 | (busiest->sum_nr_running - busiest->group_capacity_factor); |
6539 | 6543 | ||
6540 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_CAPACITY_SCALE); | 6544 | load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_CAPACITY_SCALE); |
6541 | load_above_capacity /= busiest->group_capacity; | 6545 | load_above_capacity /= busiest->group_capacity; |
6542 | } | 6546 | } |
6543 | 6547 | ||
6544 | /* | 6548 | /* |
6545 | * We're trying to get all the cpus to the average_load, so we don't | 6549 | * We're trying to get all the cpus to the average_load, so we don't |
6546 | * want to push ourselves above the average load, nor do we wish to | 6550 | * want to push ourselves above the average load, nor do we wish to |
6547 | * reduce the max loaded cpu below the average load. At the same time, | 6551 | * reduce the max loaded cpu below the average load. At the same time, |
6548 | * we also don't want to reduce the group load below the group capacity | 6552 | * we also don't want to reduce the group load below the group capacity |
6549 | * (so that we can implement power-savings policies etc). Thus we look | 6553 | * (so that we can implement power-savings policies etc). Thus we look |
6550 | * for the minimum possible imbalance. | 6554 | * for the minimum possible imbalance. |
6551 | */ | 6555 | */ |
6552 | max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); | 6556 | max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); |
6553 | 6557 | ||
6554 | /* How much load to actually move to equalise the imbalance */ | 6558 | /* How much load to actually move to equalise the imbalance */ |
6555 | env->imbalance = min( | 6559 | env->imbalance = min( |
6556 | max_pull * busiest->group_capacity, | 6560 | max_pull * busiest->group_capacity, |
6557 | (sds->avg_load - local->avg_load) * local->group_capacity | 6561 | (sds->avg_load - local->avg_load) * local->group_capacity |
6558 | ) / SCHED_CAPACITY_SCALE; | 6562 | ) / SCHED_CAPACITY_SCALE; |
6559 | 6563 | ||
6560 | /* | 6564 | /* |
6561 | * if *imbalance is less than the average load per runnable task | 6565 | * if *imbalance is less than the average load per runnable task |
6562 | * there is no guarantee that any tasks will be moved so we'll have | 6566 | * there is no guarantee that any tasks will be moved so we'll have |
6563 | * a think about bumping its value to force at least one task to be | 6567 | * a think about bumping its value to force at least one task to be |
6564 | * moved | 6568 | * moved |
6565 | */ | 6569 | */ |
6566 | if (env->imbalance < busiest->load_per_task) | 6570 | if (env->imbalance < busiest->load_per_task) |
6567 | return fix_small_imbalance(env, sds); | 6571 | return fix_small_imbalance(env, sds); |
6568 | } | 6572 | } |
6569 | 6573 | ||
6570 | /******* find_busiest_group() helpers end here *********************/ | 6574 | /******* find_busiest_group() helpers end here *********************/ |
6571 | 6575 | ||
6572 | /** | 6576 | /** |
6573 | * find_busiest_group - Returns the busiest group within the sched_domain | 6577 | * find_busiest_group - Returns the busiest group within the sched_domain |
6574 | * if there is an imbalance. If there isn't an imbalance, and | 6578 | * if there is an imbalance. If there isn't an imbalance, and |
6575 | * the user has opted for power-savings, it returns a group whose | 6579 | * the user has opted for power-savings, it returns a group whose |
6576 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if | 6580 | * CPUs can be put to idle by rebalancing those tasks elsewhere, if |
6577 | * such a group exists. | 6581 | * such a group exists. |
6578 | * | 6582 | * |
6579 | * Also calculates the amount of weighted load which should be moved | 6583 | * Also calculates the amount of weighted load which should be moved |
6580 | * to restore balance. | 6584 | * to restore balance. |
6581 | * | 6585 | * |
6582 | * @env: The load balancing environment. | 6586 | * @env: The load balancing environment. |
6583 | * | 6587 | * |
6584 | * Return: - The busiest group if imbalance exists. | 6588 | * Return: - The busiest group if imbalance exists. |
6585 | * - If no imbalance and user has opted for power-savings balance, | 6589 | * - If no imbalance and user has opted for power-savings balance, |
6586 | * return the least loaded group whose CPUs can be | 6590 | * return the least loaded group whose CPUs can be |
6587 | * put to idle by rebalancing its tasks onto our group. | 6591 | * put to idle by rebalancing its tasks onto our group. |
6588 | */ | 6592 | */ |
6589 | static struct sched_group *find_busiest_group(struct lb_env *env) | 6593 | static struct sched_group *find_busiest_group(struct lb_env *env) |
6590 | { | 6594 | { |
6591 | struct sg_lb_stats *local, *busiest; | 6595 | struct sg_lb_stats *local, *busiest; |
6592 | struct sd_lb_stats sds; | 6596 | struct sd_lb_stats sds; |
6593 | 6597 | ||
6594 | init_sd_lb_stats(&sds); | 6598 | init_sd_lb_stats(&sds); |
6595 | 6599 | ||
6596 | /* | 6600 | /* |
6597 | * Compute the various statistics relavent for load balancing at | 6601 | * Compute the various statistics relavent for load balancing at |
6598 | * this level. | 6602 | * this level. |
6599 | */ | 6603 | */ |
6600 | update_sd_lb_stats(env, &sds); | 6604 | update_sd_lb_stats(env, &sds); |
6601 | local = &sds.local_stat; | 6605 | local = &sds.local_stat; |
6602 | busiest = &sds.busiest_stat; | 6606 | busiest = &sds.busiest_stat; |
6603 | 6607 | ||
6604 | if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && | 6608 | if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && |
6605 | check_asym_packing(env, &sds)) | 6609 | check_asym_packing(env, &sds)) |
6606 | return sds.busiest; | 6610 | return sds.busiest; |
6607 | 6611 | ||
6608 | /* There is no busy sibling group to pull tasks from */ | 6612 | /* There is no busy sibling group to pull tasks from */ |
6609 | if (!sds.busiest || busiest->sum_nr_running == 0) | 6613 | if (!sds.busiest || busiest->sum_nr_running == 0) |
6610 | goto out_balanced; | 6614 | goto out_balanced; |
6611 | 6615 | ||
6612 | sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load) | 6616 | sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load) |
6613 | / sds.total_capacity; | 6617 | / sds.total_capacity; |
6614 | 6618 | ||
6615 | /* | 6619 | /* |
6616 | * If the busiest group is imbalanced the below checks don't | 6620 | * If the busiest group is imbalanced the below checks don't |
6617 | * work because they assume all things are equal, which typically | 6621 | * work because they assume all things are equal, which typically |
6618 | * isn't true due to cpus_allowed constraints and the like. | 6622 | * isn't true due to cpus_allowed constraints and the like. |
6619 | */ | 6623 | */ |
6620 | if (busiest->group_type == group_imbalanced) | 6624 | if (busiest->group_type == group_imbalanced) |
6621 | goto force_balance; | 6625 | goto force_balance; |
6622 | 6626 | ||
6623 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ | 6627 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
6624 | if (env->idle == CPU_NEWLY_IDLE && local->group_has_free_capacity && | 6628 | if (env->idle == CPU_NEWLY_IDLE && local->group_has_free_capacity && |
6625 | !busiest->group_has_free_capacity) | 6629 | !busiest->group_has_free_capacity) |
6626 | goto force_balance; | 6630 | goto force_balance; |
6627 | 6631 | ||
6628 | /* | 6632 | /* |
6629 | * If the local group is busier than the selected busiest group | 6633 | * If the local group is busier than the selected busiest group |
6630 | * don't try and pull any tasks. | 6634 | * don't try and pull any tasks. |
6631 | */ | 6635 | */ |
6632 | if (local->avg_load >= busiest->avg_load) | 6636 | if (local->avg_load >= busiest->avg_load) |
6633 | goto out_balanced; | 6637 | goto out_balanced; |
6634 | 6638 | ||
6635 | /* | 6639 | /* |
6636 | * Don't pull any tasks if this group is already above the domain | 6640 | * Don't pull any tasks if this group is already above the domain |
6637 | * average load. | 6641 | * average load. |
6638 | */ | 6642 | */ |
6639 | if (local->avg_load >= sds.avg_load) | 6643 | if (local->avg_load >= sds.avg_load) |
6640 | goto out_balanced; | 6644 | goto out_balanced; |
6641 | 6645 | ||
6642 | if (env->idle == CPU_IDLE) { | 6646 | if (env->idle == CPU_IDLE) { |
6643 | /* | 6647 | /* |
6644 | * This cpu is idle. If the busiest group is not overloaded | 6648 | * This cpu is idle. If the busiest group is not overloaded |
6645 | * and there is no imbalance between this and busiest group | 6649 | * and there is no imbalance between this and busiest group |
6646 | * wrt idle cpus, it is balanced. The imbalance becomes | 6650 | * wrt idle cpus, it is balanced. The imbalance becomes |
6647 | * significant if the diff is greater than 1 otherwise we | 6651 | * significant if the diff is greater than 1 otherwise we |
6648 | * might end up to just move the imbalance on another group | 6652 | * might end up to just move the imbalance on another group |
6649 | */ | 6653 | */ |
6650 | if ((busiest->group_type != group_overloaded) && | 6654 | if ((busiest->group_type != group_overloaded) && |
6651 | (local->idle_cpus <= (busiest->idle_cpus + 1))) | 6655 | (local->idle_cpus <= (busiest->idle_cpus + 1))) |
6652 | goto out_balanced; | 6656 | goto out_balanced; |
6653 | } else { | 6657 | } else { |
6654 | /* | 6658 | /* |
6655 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | 6659 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use |
6656 | * imbalance_pct to be conservative. | 6660 | * imbalance_pct to be conservative. |
6657 | */ | 6661 | */ |
6658 | if (100 * busiest->avg_load <= | 6662 | if (100 * busiest->avg_load <= |
6659 | env->sd->imbalance_pct * local->avg_load) | 6663 | env->sd->imbalance_pct * local->avg_load) |
6660 | goto out_balanced; | 6664 | goto out_balanced; |
6661 | } | 6665 | } |
6662 | 6666 | ||
6663 | force_balance: | 6667 | force_balance: |
6664 | /* Looks like there is an imbalance. Compute it */ | 6668 | /* Looks like there is an imbalance. Compute it */ |
6665 | calculate_imbalance(env, &sds); | 6669 | calculate_imbalance(env, &sds); |
6666 | return sds.busiest; | 6670 | return sds.busiest; |
6667 | 6671 | ||
6668 | out_balanced: | 6672 | out_balanced: |
6669 | env->imbalance = 0; | 6673 | env->imbalance = 0; |
6670 | return NULL; | 6674 | return NULL; |
6671 | } | 6675 | } |
6672 | 6676 | ||
6673 | /* | 6677 | /* |
6674 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | 6678 | * find_busiest_queue - find the busiest runqueue among the cpus in group. |
6675 | */ | 6679 | */ |
6676 | static struct rq *find_busiest_queue(struct lb_env *env, | 6680 | static struct rq *find_busiest_queue(struct lb_env *env, |
6677 | struct sched_group *group) | 6681 | struct sched_group *group) |
6678 | { | 6682 | { |
6679 | struct rq *busiest = NULL, *rq; | 6683 | struct rq *busiest = NULL, *rq; |
6680 | unsigned long busiest_load = 0, busiest_capacity = 1; | 6684 | unsigned long busiest_load = 0, busiest_capacity = 1; |
6681 | int i; | 6685 | int i; |
6682 | 6686 | ||
6683 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { | 6687 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
6684 | unsigned long capacity, capacity_factor, wl; | 6688 | unsigned long capacity, capacity_factor, wl; |
6685 | enum fbq_type rt; | 6689 | enum fbq_type rt; |
6686 | 6690 | ||
6687 | rq = cpu_rq(i); | 6691 | rq = cpu_rq(i); |
6688 | rt = fbq_classify_rq(rq); | 6692 | rt = fbq_classify_rq(rq); |
6689 | 6693 | ||
6690 | /* | 6694 | /* |
6691 | * We classify groups/runqueues into three groups: | 6695 | * We classify groups/runqueues into three groups: |
6692 | * - regular: there are !numa tasks | 6696 | * - regular: there are !numa tasks |
6693 | * - remote: there are numa tasks that run on the 'wrong' node | 6697 | * - remote: there are numa tasks that run on the 'wrong' node |
6694 | * - all: there is no distinction | 6698 | * - all: there is no distinction |
6695 | * | 6699 | * |
6696 | * In order to avoid migrating ideally placed numa tasks, | 6700 | * In order to avoid migrating ideally placed numa tasks, |
6697 | * ignore those when there's better options. | 6701 | * ignore those when there's better options. |
6698 | * | 6702 | * |
6699 | * If we ignore the actual busiest queue to migrate another | 6703 | * If we ignore the actual busiest queue to migrate another |
6700 | * task, the next balance pass can still reduce the busiest | 6704 | * task, the next balance pass can still reduce the busiest |
6701 | * queue by moving tasks around inside the node. | 6705 | * queue by moving tasks around inside the node. |
6702 | * | 6706 | * |
6703 | * If we cannot move enough load due to this classification | 6707 | * If we cannot move enough load due to this classification |
6704 | * the next pass will adjust the group classification and | 6708 | * the next pass will adjust the group classification and |
6705 | * allow migration of more tasks. | 6709 | * allow migration of more tasks. |
6706 | * | 6710 | * |
6707 | * Both cases only affect the total convergence complexity. | 6711 | * Both cases only affect the total convergence complexity. |
6708 | */ | 6712 | */ |
6709 | if (rt > env->fbq_type) | 6713 | if (rt > env->fbq_type) |
6710 | continue; | 6714 | continue; |
6711 | 6715 | ||
6712 | capacity = capacity_of(i); | 6716 | capacity = capacity_of(i); |
6713 | capacity_factor = DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE); | 6717 | capacity_factor = DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE); |
6714 | if (!capacity_factor) | 6718 | if (!capacity_factor) |
6715 | capacity_factor = fix_small_capacity(env->sd, group); | 6719 | capacity_factor = fix_small_capacity(env->sd, group); |
6716 | 6720 | ||
6717 | wl = weighted_cpuload(i); | 6721 | wl = weighted_cpuload(i); |
6718 | 6722 | ||
6719 | /* | 6723 | /* |
6720 | * When comparing with imbalance, use weighted_cpuload() | 6724 | * When comparing with imbalance, use weighted_cpuload() |
6721 | * which is not scaled with the cpu capacity. | 6725 | * which is not scaled with the cpu capacity. |
6722 | */ | 6726 | */ |
6723 | if (capacity_factor && rq->nr_running == 1 && wl > env->imbalance) | 6727 | if (capacity_factor && rq->nr_running == 1 && wl > env->imbalance) |
6724 | continue; | 6728 | continue; |
6725 | 6729 | ||
6726 | /* | 6730 | /* |
6727 | * For the load comparisons with the other cpu's, consider | 6731 | * For the load comparisons with the other cpu's, consider |
6728 | * the weighted_cpuload() scaled with the cpu capacity, so | 6732 | * the weighted_cpuload() scaled with the cpu capacity, so |
6729 | * that the load can be moved away from the cpu that is | 6733 | * that the load can be moved away from the cpu that is |
6730 | * potentially running at a lower capacity. | 6734 | * potentially running at a lower capacity. |
6731 | * | 6735 | * |
6732 | * Thus we're looking for max(wl_i / capacity_i), crosswise | 6736 | * Thus we're looking for max(wl_i / capacity_i), crosswise |
6733 | * multiplication to rid ourselves of the division works out | 6737 | * multiplication to rid ourselves of the division works out |
6734 | * to: wl_i * capacity_j > wl_j * capacity_i; where j is | 6738 | * to: wl_i * capacity_j > wl_j * capacity_i; where j is |
6735 | * our previous maximum. | 6739 | * our previous maximum. |
6736 | */ | 6740 | */ |
6737 | if (wl * busiest_capacity > busiest_load * capacity) { | 6741 | if (wl * busiest_capacity > busiest_load * capacity) { |
6738 | busiest_load = wl; | 6742 | busiest_load = wl; |
6739 | busiest_capacity = capacity; | 6743 | busiest_capacity = capacity; |
6740 | busiest = rq; | 6744 | busiest = rq; |
6741 | } | 6745 | } |
6742 | } | 6746 | } |
6743 | 6747 | ||
6744 | return busiest; | 6748 | return busiest; |
6745 | } | 6749 | } |
6746 | 6750 | ||
6747 | /* | 6751 | /* |
6748 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | 6752 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but |
6749 | * so long as it is large enough. | 6753 | * so long as it is large enough. |
6750 | */ | 6754 | */ |
6751 | #define MAX_PINNED_INTERVAL 512 | 6755 | #define MAX_PINNED_INTERVAL 512 |
6752 | 6756 | ||
6753 | /* Working cpumask for load_balance and load_balance_newidle. */ | 6757 | /* Working cpumask for load_balance and load_balance_newidle. */ |
6754 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); | 6758 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); |
6755 | 6759 | ||
6756 | static int need_active_balance(struct lb_env *env) | 6760 | static int need_active_balance(struct lb_env *env) |
6757 | { | 6761 | { |
6758 | struct sched_domain *sd = env->sd; | 6762 | struct sched_domain *sd = env->sd; |
6759 | 6763 | ||
6760 | if (env->idle == CPU_NEWLY_IDLE) { | 6764 | if (env->idle == CPU_NEWLY_IDLE) { |
6761 | 6765 | ||
6762 | /* | 6766 | /* |
6763 | * ASYM_PACKING needs to force migrate tasks from busy but | 6767 | * ASYM_PACKING needs to force migrate tasks from busy but |
6764 | * higher numbered CPUs in order to pack all tasks in the | 6768 | * higher numbered CPUs in order to pack all tasks in the |
6765 | * lowest numbered CPUs. | 6769 | * lowest numbered CPUs. |
6766 | */ | 6770 | */ |
6767 | if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) | 6771 | if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) |
6768 | return 1; | 6772 | return 1; |
6769 | } | 6773 | } |
6770 | 6774 | ||
6771 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); | 6775 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); |
6772 | } | 6776 | } |
6773 | 6777 | ||
6774 | static int active_load_balance_cpu_stop(void *data); | 6778 | static int active_load_balance_cpu_stop(void *data); |
6775 | 6779 | ||
6776 | static int should_we_balance(struct lb_env *env) | 6780 | static int should_we_balance(struct lb_env *env) |
6777 | { | 6781 | { |
6778 | struct sched_group *sg = env->sd->groups; | 6782 | struct sched_group *sg = env->sd->groups; |
6779 | struct cpumask *sg_cpus, *sg_mask; | 6783 | struct cpumask *sg_cpus, *sg_mask; |
6780 | int cpu, balance_cpu = -1; | 6784 | int cpu, balance_cpu = -1; |
6781 | 6785 | ||
6782 | /* | 6786 | /* |
6783 | * In the newly idle case, we will allow all the cpu's | 6787 | * In the newly idle case, we will allow all the cpu's |
6784 | * to do the newly idle load balance. | 6788 | * to do the newly idle load balance. |
6785 | */ | 6789 | */ |
6786 | if (env->idle == CPU_NEWLY_IDLE) | 6790 | if (env->idle == CPU_NEWLY_IDLE) |
6787 | return 1; | 6791 | return 1; |
6788 | 6792 | ||
6789 | sg_cpus = sched_group_cpus(sg); | 6793 | sg_cpus = sched_group_cpus(sg); |
6790 | sg_mask = sched_group_mask(sg); | 6794 | sg_mask = sched_group_mask(sg); |
6791 | /* Try to find first idle cpu */ | 6795 | /* Try to find first idle cpu */ |
6792 | for_each_cpu_and(cpu, sg_cpus, env->cpus) { | 6796 | for_each_cpu_and(cpu, sg_cpus, env->cpus) { |
6793 | if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu)) | 6797 | if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu)) |
6794 | continue; | 6798 | continue; |
6795 | 6799 | ||
6796 | balance_cpu = cpu; | 6800 | balance_cpu = cpu; |
6797 | break; | 6801 | break; |
6798 | } | 6802 | } |
6799 | 6803 | ||
6800 | if (balance_cpu == -1) | 6804 | if (balance_cpu == -1) |
6801 | balance_cpu = group_balance_cpu(sg); | 6805 | balance_cpu = group_balance_cpu(sg); |
6802 | 6806 | ||
6803 | /* | 6807 | /* |
6804 | * First idle cpu or the first cpu(busiest) in this sched group | 6808 | * First idle cpu or the first cpu(busiest) in this sched group |
6805 | * is eligible for doing load balancing at this and above domains. | 6809 | * is eligible for doing load balancing at this and above domains. |
6806 | */ | 6810 | */ |
6807 | return balance_cpu == env->dst_cpu; | 6811 | return balance_cpu == env->dst_cpu; |
6808 | } | 6812 | } |
6809 | 6813 | ||
6810 | /* | 6814 | /* |
6811 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | 6815 | * Check this_cpu to ensure it is balanced within domain. Attempt to move |
6812 | * tasks if there is an imbalance. | 6816 | * tasks if there is an imbalance. |
6813 | */ | 6817 | */ |
6814 | static int load_balance(int this_cpu, struct rq *this_rq, | 6818 | static int load_balance(int this_cpu, struct rq *this_rq, |
6815 | struct sched_domain *sd, enum cpu_idle_type idle, | 6819 | struct sched_domain *sd, enum cpu_idle_type idle, |
6816 | int *continue_balancing) | 6820 | int *continue_balancing) |
6817 | { | 6821 | { |
6818 | int ld_moved, cur_ld_moved, active_balance = 0; | 6822 | int ld_moved, cur_ld_moved, active_balance = 0; |
6819 | struct sched_domain *sd_parent = sd->parent; | 6823 | struct sched_domain *sd_parent = sd->parent; |
6820 | struct sched_group *group; | 6824 | struct sched_group *group; |
6821 | struct rq *busiest; | 6825 | struct rq *busiest; |
6822 | unsigned long flags; | 6826 | unsigned long flags; |
6823 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); | 6827 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); |
6824 | 6828 | ||
6825 | struct lb_env env = { | 6829 | struct lb_env env = { |
6826 | .sd = sd, | 6830 | .sd = sd, |
6827 | .dst_cpu = this_cpu, | 6831 | .dst_cpu = this_cpu, |
6828 | .dst_rq = this_rq, | 6832 | .dst_rq = this_rq, |
6829 | .dst_grpmask = sched_group_cpus(sd->groups), | 6833 | .dst_grpmask = sched_group_cpus(sd->groups), |
6830 | .idle = idle, | 6834 | .idle = idle, |
6831 | .loop_break = sched_nr_migrate_break, | 6835 | .loop_break = sched_nr_migrate_break, |
6832 | .cpus = cpus, | 6836 | .cpus = cpus, |
6833 | .fbq_type = all, | 6837 | .fbq_type = all, |
6834 | .tasks = LIST_HEAD_INIT(env.tasks), | 6838 | .tasks = LIST_HEAD_INIT(env.tasks), |
6835 | }; | 6839 | }; |
6836 | 6840 | ||
6837 | /* | 6841 | /* |
6838 | * For NEWLY_IDLE load_balancing, we don't need to consider | 6842 | * For NEWLY_IDLE load_balancing, we don't need to consider |
6839 | * other cpus in our group | 6843 | * other cpus in our group |
6840 | */ | 6844 | */ |
6841 | if (idle == CPU_NEWLY_IDLE) | 6845 | if (idle == CPU_NEWLY_IDLE) |
6842 | env.dst_grpmask = NULL; | 6846 | env.dst_grpmask = NULL; |
6843 | 6847 | ||
6844 | cpumask_copy(cpus, cpu_active_mask); | 6848 | cpumask_copy(cpus, cpu_active_mask); |
6845 | 6849 | ||
6846 | schedstat_inc(sd, lb_count[idle]); | 6850 | schedstat_inc(sd, lb_count[idle]); |
6847 | 6851 | ||
6848 | redo: | 6852 | redo: |
6849 | if (!should_we_balance(&env)) { | 6853 | if (!should_we_balance(&env)) { |
6850 | *continue_balancing = 0; | 6854 | *continue_balancing = 0; |
6851 | goto out_balanced; | 6855 | goto out_balanced; |
6852 | } | 6856 | } |
6853 | 6857 | ||
6854 | group = find_busiest_group(&env); | 6858 | group = find_busiest_group(&env); |
6855 | if (!group) { | 6859 | if (!group) { |
6856 | schedstat_inc(sd, lb_nobusyg[idle]); | 6860 | schedstat_inc(sd, lb_nobusyg[idle]); |
6857 | goto out_balanced; | 6861 | goto out_balanced; |
6858 | } | 6862 | } |
6859 | 6863 | ||
6860 | busiest = find_busiest_queue(&env, group); | 6864 | busiest = find_busiest_queue(&env, group); |
6861 | if (!busiest) { | 6865 | if (!busiest) { |
6862 | schedstat_inc(sd, lb_nobusyq[idle]); | 6866 | schedstat_inc(sd, lb_nobusyq[idle]); |
6863 | goto out_balanced; | 6867 | goto out_balanced; |
6864 | } | 6868 | } |
6865 | 6869 | ||
6866 | BUG_ON(busiest == env.dst_rq); | 6870 | BUG_ON(busiest == env.dst_rq); |
6867 | 6871 | ||
6868 | schedstat_add(sd, lb_imbalance[idle], env.imbalance); | 6872 | schedstat_add(sd, lb_imbalance[idle], env.imbalance); |
6869 | 6873 | ||
6870 | ld_moved = 0; | 6874 | ld_moved = 0; |
6871 | if (busiest->nr_running > 1) { | 6875 | if (busiest->nr_running > 1) { |
6872 | /* | 6876 | /* |
6873 | * Attempt to move tasks. If find_busiest_group has found | 6877 | * Attempt to move tasks. If find_busiest_group has found |
6874 | * an imbalance but busiest->nr_running <= 1, the group is | 6878 | * an imbalance but busiest->nr_running <= 1, the group is |
6875 | * still unbalanced. ld_moved simply stays zero, so it is | 6879 | * still unbalanced. ld_moved simply stays zero, so it is |
6876 | * correctly treated as an imbalance. | 6880 | * correctly treated as an imbalance. |
6877 | */ | 6881 | */ |
6878 | env.flags |= LBF_ALL_PINNED; | 6882 | env.flags |= LBF_ALL_PINNED; |
6879 | env.src_cpu = busiest->cpu; | 6883 | env.src_cpu = busiest->cpu; |
6880 | env.src_rq = busiest; | 6884 | env.src_rq = busiest; |
6881 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); | 6885 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); |
6882 | 6886 | ||
6883 | more_balance: | 6887 | more_balance: |
6884 | raw_spin_lock_irqsave(&busiest->lock, flags); | 6888 | raw_spin_lock_irqsave(&busiest->lock, flags); |
6885 | 6889 | ||
6886 | /* | 6890 | /* |
6887 | * cur_ld_moved - load moved in current iteration | 6891 | * cur_ld_moved - load moved in current iteration |
6888 | * ld_moved - cumulative load moved across iterations | 6892 | * ld_moved - cumulative load moved across iterations |
6889 | */ | 6893 | */ |
6890 | cur_ld_moved = detach_tasks(&env); | 6894 | cur_ld_moved = detach_tasks(&env); |
6891 | 6895 | ||
6892 | /* | 6896 | /* |
6893 | * We've detached some tasks from busiest_rq. Every | 6897 | * We've detached some tasks from busiest_rq. Every |
6894 | * task is masked "TASK_ON_RQ_MIGRATING", so we can safely | 6898 | * task is masked "TASK_ON_RQ_MIGRATING", so we can safely |
6895 | * unlock busiest->lock, and we are able to be sure | 6899 | * unlock busiest->lock, and we are able to be sure |
6896 | * that nobody can manipulate the tasks in parallel. | 6900 | * that nobody can manipulate the tasks in parallel. |
6897 | * See task_rq_lock() family for the details. | 6901 | * See task_rq_lock() family for the details. |
6898 | */ | 6902 | */ |
6899 | 6903 | ||
6900 | raw_spin_unlock(&busiest->lock); | 6904 | raw_spin_unlock(&busiest->lock); |
6901 | 6905 | ||
6902 | if (cur_ld_moved) { | 6906 | if (cur_ld_moved) { |
6903 | attach_tasks(&env); | 6907 | attach_tasks(&env); |
6904 | ld_moved += cur_ld_moved; | 6908 | ld_moved += cur_ld_moved; |
6905 | } | 6909 | } |
6906 | 6910 | ||
6907 | local_irq_restore(flags); | 6911 | local_irq_restore(flags); |
6908 | 6912 | ||
6909 | if (env.flags & LBF_NEED_BREAK) { | 6913 | if (env.flags & LBF_NEED_BREAK) { |
6910 | env.flags &= ~LBF_NEED_BREAK; | 6914 | env.flags &= ~LBF_NEED_BREAK; |
6911 | goto more_balance; | 6915 | goto more_balance; |
6912 | } | 6916 | } |
6913 | 6917 | ||
6914 | /* | 6918 | /* |
6915 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | 6919 | * Revisit (affine) tasks on src_cpu that couldn't be moved to |
6916 | * us and move them to an alternate dst_cpu in our sched_group | 6920 | * us and move them to an alternate dst_cpu in our sched_group |
6917 | * where they can run. The upper limit on how many times we | 6921 | * where they can run. The upper limit on how many times we |
6918 | * iterate on same src_cpu is dependent on number of cpus in our | 6922 | * iterate on same src_cpu is dependent on number of cpus in our |
6919 | * sched_group. | 6923 | * sched_group. |
6920 | * | 6924 | * |
6921 | * This changes load balance semantics a bit on who can move | 6925 | * This changes load balance semantics a bit on who can move |
6922 | * load to a given_cpu. In addition to the given_cpu itself | 6926 | * load to a given_cpu. In addition to the given_cpu itself |
6923 | * (or a ilb_cpu acting on its behalf where given_cpu is | 6927 | * (or a ilb_cpu acting on its behalf where given_cpu is |
6924 | * nohz-idle), we now have balance_cpu in a position to move | 6928 | * nohz-idle), we now have balance_cpu in a position to move |
6925 | * load to given_cpu. In rare situations, this may cause | 6929 | * load to given_cpu. In rare situations, this may cause |
6926 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | 6930 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding |
6927 | * _independently_ and at _same_ time to move some load to | 6931 | * _independently_ and at _same_ time to move some load to |
6928 | * given_cpu) causing exceess load to be moved to given_cpu. | 6932 | * given_cpu) causing exceess load to be moved to given_cpu. |
6929 | * This however should not happen so much in practice and | 6933 | * This however should not happen so much in practice and |
6930 | * moreover subsequent load balance cycles should correct the | 6934 | * moreover subsequent load balance cycles should correct the |
6931 | * excess load moved. | 6935 | * excess load moved. |
6932 | */ | 6936 | */ |
6933 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { | 6937 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
6934 | 6938 | ||
6935 | /* Prevent to re-select dst_cpu via env's cpus */ | 6939 | /* Prevent to re-select dst_cpu via env's cpus */ |
6936 | cpumask_clear_cpu(env.dst_cpu, env.cpus); | 6940 | cpumask_clear_cpu(env.dst_cpu, env.cpus); |
6937 | 6941 | ||
6938 | env.dst_rq = cpu_rq(env.new_dst_cpu); | 6942 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
6939 | env.dst_cpu = env.new_dst_cpu; | 6943 | env.dst_cpu = env.new_dst_cpu; |
6940 | env.flags &= ~LBF_DST_PINNED; | 6944 | env.flags &= ~LBF_DST_PINNED; |
6941 | env.loop = 0; | 6945 | env.loop = 0; |
6942 | env.loop_break = sched_nr_migrate_break; | 6946 | env.loop_break = sched_nr_migrate_break; |
6943 | 6947 | ||
6944 | /* | 6948 | /* |
6945 | * Go back to "more_balance" rather than "redo" since we | 6949 | * Go back to "more_balance" rather than "redo" since we |
6946 | * need to continue with same src_cpu. | 6950 | * need to continue with same src_cpu. |
6947 | */ | 6951 | */ |
6948 | goto more_balance; | 6952 | goto more_balance; |
6949 | } | 6953 | } |
6950 | 6954 | ||
6951 | /* | 6955 | /* |
6952 | * We failed to reach balance because of affinity. | 6956 | * We failed to reach balance because of affinity. |
6953 | */ | 6957 | */ |
6954 | if (sd_parent) { | 6958 | if (sd_parent) { |
6955 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; | 6959 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
6956 | 6960 | ||
6957 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) | 6961 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) |
6958 | *group_imbalance = 1; | 6962 | *group_imbalance = 1; |
6959 | } | 6963 | } |
6960 | 6964 | ||
6961 | /* All tasks on this runqueue were pinned by CPU affinity */ | 6965 | /* All tasks on this runqueue were pinned by CPU affinity */ |
6962 | if (unlikely(env.flags & LBF_ALL_PINNED)) { | 6966 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
6963 | cpumask_clear_cpu(cpu_of(busiest), cpus); | 6967 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
6964 | if (!cpumask_empty(cpus)) { | 6968 | if (!cpumask_empty(cpus)) { |
6965 | env.loop = 0; | 6969 | env.loop = 0; |
6966 | env.loop_break = sched_nr_migrate_break; | 6970 | env.loop_break = sched_nr_migrate_break; |
6967 | goto redo; | 6971 | goto redo; |
6968 | } | 6972 | } |
6969 | goto out_all_pinned; | 6973 | goto out_all_pinned; |
6970 | } | 6974 | } |
6971 | } | 6975 | } |
6972 | 6976 | ||
6973 | if (!ld_moved) { | 6977 | if (!ld_moved) { |
6974 | schedstat_inc(sd, lb_failed[idle]); | 6978 | schedstat_inc(sd, lb_failed[idle]); |
6975 | /* | 6979 | /* |
6976 | * Increment the failure counter only on periodic balance. | 6980 | * Increment the failure counter only on periodic balance. |
6977 | * We do not want newidle balance, which can be very | 6981 | * We do not want newidle balance, which can be very |
6978 | * frequent, pollute the failure counter causing | 6982 | * frequent, pollute the failure counter causing |
6979 | * excessive cache_hot migrations and active balances. | 6983 | * excessive cache_hot migrations and active balances. |
6980 | */ | 6984 | */ |
6981 | if (idle != CPU_NEWLY_IDLE) | 6985 | if (idle != CPU_NEWLY_IDLE) |
6982 | sd->nr_balance_failed++; | 6986 | sd->nr_balance_failed++; |
6983 | 6987 | ||
6984 | if (need_active_balance(&env)) { | 6988 | if (need_active_balance(&env)) { |
6985 | raw_spin_lock_irqsave(&busiest->lock, flags); | 6989 | raw_spin_lock_irqsave(&busiest->lock, flags); |
6986 | 6990 | ||
6987 | /* don't kick the active_load_balance_cpu_stop, | 6991 | /* don't kick the active_load_balance_cpu_stop, |
6988 | * if the curr task on busiest cpu can't be | 6992 | * if the curr task on busiest cpu can't be |
6989 | * moved to this_cpu | 6993 | * moved to this_cpu |
6990 | */ | 6994 | */ |
6991 | if (!cpumask_test_cpu(this_cpu, | 6995 | if (!cpumask_test_cpu(this_cpu, |
6992 | tsk_cpus_allowed(busiest->curr))) { | 6996 | tsk_cpus_allowed(busiest->curr))) { |
6993 | raw_spin_unlock_irqrestore(&busiest->lock, | 6997 | raw_spin_unlock_irqrestore(&busiest->lock, |
6994 | flags); | 6998 | flags); |
6995 | env.flags |= LBF_ALL_PINNED; | 6999 | env.flags |= LBF_ALL_PINNED; |
6996 | goto out_one_pinned; | 7000 | goto out_one_pinned; |
6997 | } | 7001 | } |
6998 | 7002 | ||
6999 | /* | 7003 | /* |
7000 | * ->active_balance synchronizes accesses to | 7004 | * ->active_balance synchronizes accesses to |
7001 | * ->active_balance_work. Once set, it's cleared | 7005 | * ->active_balance_work. Once set, it's cleared |
7002 | * only after active load balance is finished. | 7006 | * only after active load balance is finished. |
7003 | */ | 7007 | */ |
7004 | if (!busiest->active_balance) { | 7008 | if (!busiest->active_balance) { |
7005 | busiest->active_balance = 1; | 7009 | busiest->active_balance = 1; |
7006 | busiest->push_cpu = this_cpu; | 7010 | busiest->push_cpu = this_cpu; |
7007 | active_balance = 1; | 7011 | active_balance = 1; |
7008 | } | 7012 | } |
7009 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | 7013 | raw_spin_unlock_irqrestore(&busiest->lock, flags); |
7010 | 7014 | ||
7011 | if (active_balance) { | 7015 | if (active_balance) { |
7012 | stop_one_cpu_nowait(cpu_of(busiest), | 7016 | stop_one_cpu_nowait(cpu_of(busiest), |
7013 | active_load_balance_cpu_stop, busiest, | 7017 | active_load_balance_cpu_stop, busiest, |
7014 | &busiest->active_balance_work); | 7018 | &busiest->active_balance_work); |
7015 | } | 7019 | } |
7016 | 7020 | ||
7017 | /* | 7021 | /* |
7018 | * We've kicked active balancing, reset the failure | 7022 | * We've kicked active balancing, reset the failure |
7019 | * counter. | 7023 | * counter. |
7020 | */ | 7024 | */ |
7021 | sd->nr_balance_failed = sd->cache_nice_tries+1; | 7025 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
7022 | } | 7026 | } |
7023 | } else | 7027 | } else |
7024 | sd->nr_balance_failed = 0; | 7028 | sd->nr_balance_failed = 0; |
7025 | 7029 | ||
7026 | if (likely(!active_balance)) { | 7030 | if (likely(!active_balance)) { |
7027 | /* We were unbalanced, so reset the balancing interval */ | 7031 | /* We were unbalanced, so reset the balancing interval */ |
7028 | sd->balance_interval = sd->min_interval; | 7032 | sd->balance_interval = sd->min_interval; |
7029 | } else { | 7033 | } else { |
7030 | /* | 7034 | /* |
7031 | * If we've begun active balancing, start to back off. This | 7035 | * If we've begun active balancing, start to back off. This |
7032 | * case may not be covered by the all_pinned logic if there | 7036 | * case may not be covered by the all_pinned logic if there |
7033 | * is only 1 task on the busy runqueue (because we don't call | 7037 | * is only 1 task on the busy runqueue (because we don't call |
7034 | * detach_tasks). | 7038 | * detach_tasks). |
7035 | */ | 7039 | */ |
7036 | if (sd->balance_interval < sd->max_interval) | 7040 | if (sd->balance_interval < sd->max_interval) |
7037 | sd->balance_interval *= 2; | 7041 | sd->balance_interval *= 2; |
7038 | } | 7042 | } |
7039 | 7043 | ||
7040 | goto out; | 7044 | goto out; |
7041 | 7045 | ||
7042 | out_balanced: | 7046 | out_balanced: |
7043 | /* | 7047 | /* |
7044 | * We reach balance although we may have faced some affinity | 7048 | * We reach balance although we may have faced some affinity |
7045 | * constraints. Clear the imbalance flag if it was set. | 7049 | * constraints. Clear the imbalance flag if it was set. |
7046 | */ | 7050 | */ |
7047 | if (sd_parent) { | 7051 | if (sd_parent) { |
7048 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; | 7052 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
7049 | 7053 | ||
7050 | if (*group_imbalance) | 7054 | if (*group_imbalance) |
7051 | *group_imbalance = 0; | 7055 | *group_imbalance = 0; |
7052 | } | 7056 | } |
7053 | 7057 | ||
7054 | out_all_pinned: | 7058 | out_all_pinned: |
7055 | /* | 7059 | /* |
7056 | * We reach balance because all tasks are pinned at this level so | 7060 | * We reach balance because all tasks are pinned at this level so |
7057 | * we can't migrate them. Let the imbalance flag set so parent level | 7061 | * we can't migrate them. Let the imbalance flag set so parent level |
7058 | * can try to migrate them. | 7062 | * can try to migrate them. |
7059 | */ | 7063 | */ |
7060 | schedstat_inc(sd, lb_balanced[idle]); | 7064 | schedstat_inc(sd, lb_balanced[idle]); |
7061 | 7065 | ||
7062 | sd->nr_balance_failed = 0; | 7066 | sd->nr_balance_failed = 0; |
7063 | 7067 | ||
7064 | out_one_pinned: | 7068 | out_one_pinned: |
7065 | /* tune up the balancing interval */ | 7069 | /* tune up the balancing interval */ |
7066 | if (((env.flags & LBF_ALL_PINNED) && | 7070 | if (((env.flags & LBF_ALL_PINNED) && |
7067 | sd->balance_interval < MAX_PINNED_INTERVAL) || | 7071 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
7068 | (sd->balance_interval < sd->max_interval)) | 7072 | (sd->balance_interval < sd->max_interval)) |
7069 | sd->balance_interval *= 2; | 7073 | sd->balance_interval *= 2; |
7070 | 7074 | ||
7071 | ld_moved = 0; | 7075 | ld_moved = 0; |
7072 | out: | 7076 | out: |
7073 | return ld_moved; | 7077 | return ld_moved; |
7074 | } | 7078 | } |
7075 | 7079 | ||
7076 | static inline unsigned long | 7080 | static inline unsigned long |
7077 | get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) | 7081 | get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) |
7078 | { | 7082 | { |
7079 | unsigned long interval = sd->balance_interval; | 7083 | unsigned long interval = sd->balance_interval; |
7080 | 7084 | ||
7081 | if (cpu_busy) | 7085 | if (cpu_busy) |
7082 | interval *= sd->busy_factor; | 7086 | interval *= sd->busy_factor; |
7083 | 7087 | ||
7084 | /* scale ms to jiffies */ | 7088 | /* scale ms to jiffies */ |
7085 | interval = msecs_to_jiffies(interval); | 7089 | interval = msecs_to_jiffies(interval); |
7086 | interval = clamp(interval, 1UL, max_load_balance_interval); | 7090 | interval = clamp(interval, 1UL, max_load_balance_interval); |
7087 | 7091 | ||
7088 | return interval; | 7092 | return interval; |
7089 | } | 7093 | } |
7090 | 7094 | ||
7091 | static inline void | 7095 | static inline void |
7092 | update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance) | 7096 | update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance) |
7093 | { | 7097 | { |
7094 | unsigned long interval, next; | 7098 | unsigned long interval, next; |
7095 | 7099 | ||
7096 | interval = get_sd_balance_interval(sd, cpu_busy); | 7100 | interval = get_sd_balance_interval(sd, cpu_busy); |
7097 | next = sd->last_balance + interval; | 7101 | next = sd->last_balance + interval; |
7098 | 7102 | ||
7099 | if (time_after(*next_balance, next)) | 7103 | if (time_after(*next_balance, next)) |
7100 | *next_balance = next; | 7104 | *next_balance = next; |
7101 | } | 7105 | } |
7102 | 7106 | ||
7103 | /* | 7107 | /* |
7104 | * idle_balance is called by schedule() if this_cpu is about to become | 7108 | * idle_balance is called by schedule() if this_cpu is about to become |
7105 | * idle. Attempts to pull tasks from other CPUs. | 7109 | * idle. Attempts to pull tasks from other CPUs. |
7106 | */ | 7110 | */ |
7107 | static int idle_balance(struct rq *this_rq) | 7111 | static int idle_balance(struct rq *this_rq) |
7108 | { | 7112 | { |
7109 | unsigned long next_balance = jiffies + HZ; | 7113 | unsigned long next_balance = jiffies + HZ; |
7110 | int this_cpu = this_rq->cpu; | 7114 | int this_cpu = this_rq->cpu; |
7111 | struct sched_domain *sd; | 7115 | struct sched_domain *sd; |
7112 | int pulled_task = 0; | 7116 | int pulled_task = 0; |
7113 | u64 curr_cost = 0; | 7117 | u64 curr_cost = 0; |
7114 | 7118 | ||
7115 | idle_enter_fair(this_rq); | 7119 | idle_enter_fair(this_rq); |
7116 | 7120 | ||
7117 | /* | 7121 | /* |
7118 | * We must set idle_stamp _before_ calling idle_balance(), such that we | 7122 | * We must set idle_stamp _before_ calling idle_balance(), such that we |
7119 | * measure the duration of idle_balance() as idle time. | 7123 | * measure the duration of idle_balance() as idle time. |
7120 | */ | 7124 | */ |
7121 | this_rq->idle_stamp = rq_clock(this_rq); | 7125 | this_rq->idle_stamp = rq_clock(this_rq); |
7122 | 7126 | ||
7123 | if (this_rq->avg_idle < sysctl_sched_migration_cost || | 7127 | if (this_rq->avg_idle < sysctl_sched_migration_cost || |
7124 | !this_rq->rd->overload) { | 7128 | !this_rq->rd->overload) { |
7125 | rcu_read_lock(); | 7129 | rcu_read_lock(); |
7126 | sd = rcu_dereference_check_sched_domain(this_rq->sd); | 7130 | sd = rcu_dereference_check_sched_domain(this_rq->sd); |
7127 | if (sd) | 7131 | if (sd) |
7128 | update_next_balance(sd, 0, &next_balance); | 7132 | update_next_balance(sd, 0, &next_balance); |
7129 | rcu_read_unlock(); | 7133 | rcu_read_unlock(); |
7130 | 7134 | ||
7131 | goto out; | 7135 | goto out; |
7132 | } | 7136 | } |
7133 | 7137 | ||
7134 | /* | 7138 | /* |
7135 | * Drop the rq->lock, but keep IRQ/preempt disabled. | 7139 | * Drop the rq->lock, but keep IRQ/preempt disabled. |
7136 | */ | 7140 | */ |
7137 | raw_spin_unlock(&this_rq->lock); | 7141 | raw_spin_unlock(&this_rq->lock); |
7138 | 7142 | ||
7139 | update_blocked_averages(this_cpu); | 7143 | update_blocked_averages(this_cpu); |
7140 | rcu_read_lock(); | 7144 | rcu_read_lock(); |
7141 | for_each_domain(this_cpu, sd) { | 7145 | for_each_domain(this_cpu, sd) { |
7142 | int continue_balancing = 1; | 7146 | int continue_balancing = 1; |
7143 | u64 t0, domain_cost; | 7147 | u64 t0, domain_cost; |
7144 | 7148 | ||
7145 | if (!(sd->flags & SD_LOAD_BALANCE)) | 7149 | if (!(sd->flags & SD_LOAD_BALANCE)) |
7146 | continue; | 7150 | continue; |
7147 | 7151 | ||
7148 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) { | 7152 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) { |
7149 | update_next_balance(sd, 0, &next_balance); | 7153 | update_next_balance(sd, 0, &next_balance); |
7150 | break; | 7154 | break; |
7151 | } | 7155 | } |
7152 | 7156 | ||
7153 | if (sd->flags & SD_BALANCE_NEWIDLE) { | 7157 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
7154 | t0 = sched_clock_cpu(this_cpu); | 7158 | t0 = sched_clock_cpu(this_cpu); |
7155 | 7159 | ||
7156 | pulled_task = load_balance(this_cpu, this_rq, | 7160 | pulled_task = load_balance(this_cpu, this_rq, |
7157 | sd, CPU_NEWLY_IDLE, | 7161 | sd, CPU_NEWLY_IDLE, |
7158 | &continue_balancing); | 7162 | &continue_balancing); |
7159 | 7163 | ||
7160 | domain_cost = sched_clock_cpu(this_cpu) - t0; | 7164 | domain_cost = sched_clock_cpu(this_cpu) - t0; |
7161 | if (domain_cost > sd->max_newidle_lb_cost) | 7165 | if (domain_cost > sd->max_newidle_lb_cost) |
7162 | sd->max_newidle_lb_cost = domain_cost; | 7166 | sd->max_newidle_lb_cost = domain_cost; |
7163 | 7167 | ||
7164 | curr_cost += domain_cost; | 7168 | curr_cost += domain_cost; |
7165 | } | 7169 | } |
7166 | 7170 | ||
7167 | update_next_balance(sd, 0, &next_balance); | 7171 | update_next_balance(sd, 0, &next_balance); |
7168 | 7172 | ||
7169 | /* | 7173 | /* |
7170 | * Stop searching for tasks to pull if there are | 7174 | * Stop searching for tasks to pull if there are |
7171 | * now runnable tasks on this rq. | 7175 | * now runnable tasks on this rq. |
7172 | */ | 7176 | */ |
7173 | if (pulled_task || this_rq->nr_running > 0) | 7177 | if (pulled_task || this_rq->nr_running > 0) |
7174 | break; | 7178 | break; |
7175 | } | 7179 | } |
7176 | rcu_read_unlock(); | 7180 | rcu_read_unlock(); |
7177 | 7181 | ||
7178 | raw_spin_lock(&this_rq->lock); | 7182 | raw_spin_lock(&this_rq->lock); |
7179 | 7183 | ||
7180 | if (curr_cost > this_rq->max_idle_balance_cost) | 7184 | if (curr_cost > this_rq->max_idle_balance_cost) |
7181 | this_rq->max_idle_balance_cost = curr_cost; | 7185 | this_rq->max_idle_balance_cost = curr_cost; |
7182 | 7186 | ||
7183 | /* | 7187 | /* |
7184 | * While browsing the domains, we released the rq lock, a task could | 7188 | * While browsing the domains, we released the rq lock, a task could |
7185 | * have been enqueued in the meantime. Since we're not going idle, | 7189 | * have been enqueued in the meantime. Since we're not going idle, |
7186 | * pretend we pulled a task. | 7190 | * pretend we pulled a task. |
7187 | */ | 7191 | */ |
7188 | if (this_rq->cfs.h_nr_running && !pulled_task) | 7192 | if (this_rq->cfs.h_nr_running && !pulled_task) |
7189 | pulled_task = 1; | 7193 | pulled_task = 1; |
7190 | 7194 | ||
7191 | out: | 7195 | out: |
7192 | /* Move the next balance forward */ | 7196 | /* Move the next balance forward */ |
7193 | if (time_after(this_rq->next_balance, next_balance)) | 7197 | if (time_after(this_rq->next_balance, next_balance)) |
7194 | this_rq->next_balance = next_balance; | 7198 | this_rq->next_balance = next_balance; |
7195 | 7199 | ||
7196 | /* Is there a task of a high priority class? */ | 7200 | /* Is there a task of a high priority class? */ |
7197 | if (this_rq->nr_running != this_rq->cfs.h_nr_running) | 7201 | if (this_rq->nr_running != this_rq->cfs.h_nr_running) |
7198 | pulled_task = -1; | 7202 | pulled_task = -1; |
7199 | 7203 | ||
7200 | if (pulled_task) { | 7204 | if (pulled_task) { |
7201 | idle_exit_fair(this_rq); | 7205 | idle_exit_fair(this_rq); |
7202 | this_rq->idle_stamp = 0; | 7206 | this_rq->idle_stamp = 0; |
7203 | } | 7207 | } |
7204 | 7208 | ||
7205 | return pulled_task; | 7209 | return pulled_task; |
7206 | } | 7210 | } |
7207 | 7211 | ||
7208 | /* | 7212 | /* |
7209 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes | 7213 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
7210 | * running tasks off the busiest CPU onto idle CPUs. It requires at | 7214 | * running tasks off the busiest CPU onto idle CPUs. It requires at |
7211 | * least 1 task to be running on each physical CPU where possible, and | 7215 | * least 1 task to be running on each physical CPU where possible, and |
7212 | * avoids physical / logical imbalances. | 7216 | * avoids physical / logical imbalances. |
7213 | */ | 7217 | */ |
7214 | static int active_load_balance_cpu_stop(void *data) | 7218 | static int active_load_balance_cpu_stop(void *data) |
7215 | { | 7219 | { |
7216 | struct rq *busiest_rq = data; | 7220 | struct rq *busiest_rq = data; |
7217 | int busiest_cpu = cpu_of(busiest_rq); | 7221 | int busiest_cpu = cpu_of(busiest_rq); |
7218 | int target_cpu = busiest_rq->push_cpu; | 7222 | int target_cpu = busiest_rq->push_cpu; |
7219 | struct rq *target_rq = cpu_rq(target_cpu); | 7223 | struct rq *target_rq = cpu_rq(target_cpu); |
7220 | struct sched_domain *sd; | 7224 | struct sched_domain *sd; |
7221 | struct task_struct *p = NULL; | 7225 | struct task_struct *p = NULL; |
7222 | 7226 | ||
7223 | raw_spin_lock_irq(&busiest_rq->lock); | 7227 | raw_spin_lock_irq(&busiest_rq->lock); |
7224 | 7228 | ||
7225 | /* make sure the requested cpu hasn't gone down in the meantime */ | 7229 | /* make sure the requested cpu hasn't gone down in the meantime */ |
7226 | if (unlikely(busiest_cpu != smp_processor_id() || | 7230 | if (unlikely(busiest_cpu != smp_processor_id() || |
7227 | !busiest_rq->active_balance)) | 7231 | !busiest_rq->active_balance)) |
7228 | goto out_unlock; | 7232 | goto out_unlock; |
7229 | 7233 | ||
7230 | /* Is there any task to move? */ | 7234 | /* Is there any task to move? */ |
7231 | if (busiest_rq->nr_running <= 1) | 7235 | if (busiest_rq->nr_running <= 1) |
7232 | goto out_unlock; | 7236 | goto out_unlock; |
7233 | 7237 | ||
7234 | /* | 7238 | /* |
7235 | * This condition is "impossible", if it occurs | 7239 | * This condition is "impossible", if it occurs |
7236 | * we need to fix it. Originally reported by | 7240 | * we need to fix it. Originally reported by |
7237 | * Bjorn Helgaas on a 128-cpu setup. | 7241 | * Bjorn Helgaas on a 128-cpu setup. |
7238 | */ | 7242 | */ |
7239 | BUG_ON(busiest_rq == target_rq); | 7243 | BUG_ON(busiest_rq == target_rq); |
7240 | 7244 | ||
7241 | /* Search for an sd spanning us and the target CPU. */ | 7245 | /* Search for an sd spanning us and the target CPU. */ |
7242 | rcu_read_lock(); | 7246 | rcu_read_lock(); |
7243 | for_each_domain(target_cpu, sd) { | 7247 | for_each_domain(target_cpu, sd) { |
7244 | if ((sd->flags & SD_LOAD_BALANCE) && | 7248 | if ((sd->flags & SD_LOAD_BALANCE) && |
7245 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | 7249 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) |
7246 | break; | 7250 | break; |
7247 | } | 7251 | } |
7248 | 7252 | ||
7249 | if (likely(sd)) { | 7253 | if (likely(sd)) { |
7250 | struct lb_env env = { | 7254 | struct lb_env env = { |
7251 | .sd = sd, | 7255 | .sd = sd, |
7252 | .dst_cpu = target_cpu, | 7256 | .dst_cpu = target_cpu, |
7253 | .dst_rq = target_rq, | 7257 | .dst_rq = target_rq, |
7254 | .src_cpu = busiest_rq->cpu, | 7258 | .src_cpu = busiest_rq->cpu, |
7255 | .src_rq = busiest_rq, | 7259 | .src_rq = busiest_rq, |
7256 | .idle = CPU_IDLE, | 7260 | .idle = CPU_IDLE, |
7257 | }; | 7261 | }; |
7258 | 7262 | ||
7259 | schedstat_inc(sd, alb_count); | 7263 | schedstat_inc(sd, alb_count); |
7260 | 7264 | ||
7261 | p = detach_one_task(&env); | 7265 | p = detach_one_task(&env); |
7262 | if (p) | 7266 | if (p) |
7263 | schedstat_inc(sd, alb_pushed); | 7267 | schedstat_inc(sd, alb_pushed); |
7264 | else | 7268 | else |
7265 | schedstat_inc(sd, alb_failed); | 7269 | schedstat_inc(sd, alb_failed); |
7266 | } | 7270 | } |
7267 | rcu_read_unlock(); | 7271 | rcu_read_unlock(); |
7268 | out_unlock: | 7272 | out_unlock: |
7269 | busiest_rq->active_balance = 0; | 7273 | busiest_rq->active_balance = 0; |
7270 | raw_spin_unlock(&busiest_rq->lock); | 7274 | raw_spin_unlock(&busiest_rq->lock); |
7271 | 7275 | ||
7272 | if (p) | 7276 | if (p) |
7273 | attach_one_task(target_rq, p); | 7277 | attach_one_task(target_rq, p); |
7274 | 7278 | ||
7275 | local_irq_enable(); | 7279 | local_irq_enable(); |
7276 | 7280 | ||
7277 | return 0; | 7281 | return 0; |
7278 | } | 7282 | } |
7279 | 7283 | ||
7280 | static inline int on_null_domain(struct rq *rq) | 7284 | static inline int on_null_domain(struct rq *rq) |
7281 | { | 7285 | { |
7282 | return unlikely(!rcu_dereference_sched(rq->sd)); | 7286 | return unlikely(!rcu_dereference_sched(rq->sd)); |
7283 | } | 7287 | } |
7284 | 7288 | ||
7285 | #ifdef CONFIG_NO_HZ_COMMON | 7289 | #ifdef CONFIG_NO_HZ_COMMON |
7286 | /* | 7290 | /* |
7287 | * idle load balancing details | 7291 | * idle load balancing details |
7288 | * - When one of the busy CPUs notice that there may be an idle rebalancing | 7292 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
7289 | * needed, they will kick the idle load balancer, which then does idle | 7293 | * needed, they will kick the idle load balancer, which then does idle |
7290 | * load balancing for all the idle CPUs. | 7294 | * load balancing for all the idle CPUs. |
7291 | */ | 7295 | */ |
7292 | static struct { | 7296 | static struct { |
7293 | cpumask_var_t idle_cpus_mask; | 7297 | cpumask_var_t idle_cpus_mask; |
7294 | atomic_t nr_cpus; | 7298 | atomic_t nr_cpus; |
7295 | unsigned long next_balance; /* in jiffy units */ | 7299 | unsigned long next_balance; /* in jiffy units */ |
7296 | } nohz ____cacheline_aligned; | 7300 | } nohz ____cacheline_aligned; |
7297 | 7301 | ||
7298 | static inline int find_new_ilb(void) | 7302 | static inline int find_new_ilb(void) |
7299 | { | 7303 | { |
7300 | int ilb = cpumask_first(nohz.idle_cpus_mask); | 7304 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
7301 | 7305 | ||
7302 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) | 7306 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
7303 | return ilb; | 7307 | return ilb; |
7304 | 7308 | ||
7305 | return nr_cpu_ids; | 7309 | return nr_cpu_ids; |
7306 | } | 7310 | } |
7307 | 7311 | ||
7308 | /* | 7312 | /* |
7309 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | 7313 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the |
7310 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | 7314 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle |
7311 | * CPU (if there is one). | 7315 | * CPU (if there is one). |
7312 | */ | 7316 | */ |
7313 | static void nohz_balancer_kick(void) | 7317 | static void nohz_balancer_kick(void) |
7314 | { | 7318 | { |
7315 | int ilb_cpu; | 7319 | int ilb_cpu; |
7316 | 7320 | ||
7317 | nohz.next_balance++; | 7321 | nohz.next_balance++; |
7318 | 7322 | ||
7319 | ilb_cpu = find_new_ilb(); | 7323 | ilb_cpu = find_new_ilb(); |
7320 | 7324 | ||
7321 | if (ilb_cpu >= nr_cpu_ids) | 7325 | if (ilb_cpu >= nr_cpu_ids) |
7322 | return; | 7326 | return; |
7323 | 7327 | ||
7324 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) | 7328 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
7325 | return; | 7329 | return; |
7326 | /* | 7330 | /* |
7327 | * Use smp_send_reschedule() instead of resched_cpu(). | 7331 | * Use smp_send_reschedule() instead of resched_cpu(). |
7328 | * This way we generate a sched IPI on the target cpu which | 7332 | * This way we generate a sched IPI on the target cpu which |
7329 | * is idle. And the softirq performing nohz idle load balance | 7333 | * is idle. And the softirq performing nohz idle load balance |
7330 | * will be run before returning from the IPI. | 7334 | * will be run before returning from the IPI. |
7331 | */ | 7335 | */ |
7332 | smp_send_reschedule(ilb_cpu); | 7336 | smp_send_reschedule(ilb_cpu); |
7333 | return; | 7337 | return; |
7334 | } | 7338 | } |
7335 | 7339 | ||
7336 | static inline void nohz_balance_exit_idle(int cpu) | 7340 | static inline void nohz_balance_exit_idle(int cpu) |
7337 | { | 7341 | { |
7338 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | 7342 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { |
7339 | /* | 7343 | /* |
7340 | * Completely isolated CPUs don't ever set, so we must test. | 7344 | * Completely isolated CPUs don't ever set, so we must test. |
7341 | */ | 7345 | */ |
7342 | if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) { | 7346 | if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) { |
7343 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | 7347 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); |
7344 | atomic_dec(&nohz.nr_cpus); | 7348 | atomic_dec(&nohz.nr_cpus); |
7345 | } | 7349 | } |
7346 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | 7350 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); |
7347 | } | 7351 | } |
7348 | } | 7352 | } |
7349 | 7353 | ||
7350 | static inline void set_cpu_sd_state_busy(void) | 7354 | static inline void set_cpu_sd_state_busy(void) |
7351 | { | 7355 | { |
7352 | struct sched_domain *sd; | 7356 | struct sched_domain *sd; |
7353 | int cpu = smp_processor_id(); | 7357 | int cpu = smp_processor_id(); |
7354 | 7358 | ||
7355 | rcu_read_lock(); | 7359 | rcu_read_lock(); |
7356 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); | 7360 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
7357 | 7361 | ||
7358 | if (!sd || !sd->nohz_idle) | 7362 | if (!sd || !sd->nohz_idle) |
7359 | goto unlock; | 7363 | goto unlock; |
7360 | sd->nohz_idle = 0; | 7364 | sd->nohz_idle = 0; |
7361 | 7365 | ||
7362 | atomic_inc(&sd->groups->sgc->nr_busy_cpus); | 7366 | atomic_inc(&sd->groups->sgc->nr_busy_cpus); |
7363 | unlock: | 7367 | unlock: |
7364 | rcu_read_unlock(); | 7368 | rcu_read_unlock(); |
7365 | } | 7369 | } |
7366 | 7370 | ||
7367 | void set_cpu_sd_state_idle(void) | 7371 | void set_cpu_sd_state_idle(void) |
7368 | { | 7372 | { |
7369 | struct sched_domain *sd; | 7373 | struct sched_domain *sd; |
7370 | int cpu = smp_processor_id(); | 7374 | int cpu = smp_processor_id(); |
7371 | 7375 | ||
7372 | rcu_read_lock(); | 7376 | rcu_read_lock(); |
7373 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); | 7377 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
7374 | 7378 | ||
7375 | if (!sd || sd->nohz_idle) | 7379 | if (!sd || sd->nohz_idle) |
7376 | goto unlock; | 7380 | goto unlock; |
7377 | sd->nohz_idle = 1; | 7381 | sd->nohz_idle = 1; |
7378 | 7382 | ||
7379 | atomic_dec(&sd->groups->sgc->nr_busy_cpus); | 7383 | atomic_dec(&sd->groups->sgc->nr_busy_cpus); |
7380 | unlock: | 7384 | unlock: |
7381 | rcu_read_unlock(); | 7385 | rcu_read_unlock(); |
7382 | } | 7386 | } |
7383 | 7387 | ||
7384 | /* | 7388 | /* |
7385 | * This routine will record that the cpu is going idle with tick stopped. | 7389 | * This routine will record that the cpu is going idle with tick stopped. |
7386 | * This info will be used in performing idle load balancing in the future. | 7390 | * This info will be used in performing idle load balancing in the future. |
7387 | */ | 7391 | */ |
7388 | void nohz_balance_enter_idle(int cpu) | 7392 | void nohz_balance_enter_idle(int cpu) |
7389 | { | 7393 | { |
7390 | /* | 7394 | /* |
7391 | * If this cpu is going down, then nothing needs to be done. | 7395 | * If this cpu is going down, then nothing needs to be done. |
7392 | */ | 7396 | */ |
7393 | if (!cpu_active(cpu)) | 7397 | if (!cpu_active(cpu)) |
7394 | return; | 7398 | return; |
7395 | 7399 | ||
7396 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) | 7400 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
7397 | return; | 7401 | return; |
7398 | 7402 | ||
7399 | /* | 7403 | /* |
7400 | * If we're a completely isolated CPU, we don't play. | 7404 | * If we're a completely isolated CPU, we don't play. |
7401 | */ | 7405 | */ |
7402 | if (on_null_domain(cpu_rq(cpu))) | 7406 | if (on_null_domain(cpu_rq(cpu))) |
7403 | return; | 7407 | return; |
7404 | 7408 | ||
7405 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); | 7409 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
7406 | atomic_inc(&nohz.nr_cpus); | 7410 | atomic_inc(&nohz.nr_cpus); |
7407 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | 7411 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); |
7408 | } | 7412 | } |
7409 | 7413 | ||
7410 | static int sched_ilb_notifier(struct notifier_block *nfb, | 7414 | static int sched_ilb_notifier(struct notifier_block *nfb, |
7411 | unsigned long action, void *hcpu) | 7415 | unsigned long action, void *hcpu) |
7412 | { | 7416 | { |
7413 | switch (action & ~CPU_TASKS_FROZEN) { | 7417 | switch (action & ~CPU_TASKS_FROZEN) { |
7414 | case CPU_DYING: | 7418 | case CPU_DYING: |
7415 | nohz_balance_exit_idle(smp_processor_id()); | 7419 | nohz_balance_exit_idle(smp_processor_id()); |
7416 | return NOTIFY_OK; | 7420 | return NOTIFY_OK; |
7417 | default: | 7421 | default: |
7418 | return NOTIFY_DONE; | 7422 | return NOTIFY_DONE; |
7419 | } | 7423 | } |
7420 | } | 7424 | } |
7421 | #endif | 7425 | #endif |
7422 | 7426 | ||
7423 | static DEFINE_SPINLOCK(balancing); | 7427 | static DEFINE_SPINLOCK(balancing); |
7424 | 7428 | ||
7425 | /* | 7429 | /* |
7426 | * Scale the max load_balance interval with the number of CPUs in the system. | 7430 | * Scale the max load_balance interval with the number of CPUs in the system. |
7427 | * This trades load-balance latency on larger machines for less cross talk. | 7431 | * This trades load-balance latency on larger machines for less cross talk. |
7428 | */ | 7432 | */ |
7429 | void update_max_interval(void) | 7433 | void update_max_interval(void) |
7430 | { | 7434 | { |
7431 | max_load_balance_interval = HZ*num_online_cpus()/10; | 7435 | max_load_balance_interval = HZ*num_online_cpus()/10; |
7432 | } | 7436 | } |
7433 | 7437 | ||
7434 | /* | 7438 | /* |
7435 | * It checks each scheduling domain to see if it is due to be balanced, | 7439 | * It checks each scheduling domain to see if it is due to be balanced, |
7436 | * and initiates a balancing operation if so. | 7440 | * and initiates a balancing operation if so. |
7437 | * | 7441 | * |
7438 | * Balancing parameters are set up in init_sched_domains. | 7442 | * Balancing parameters are set up in init_sched_domains. |
7439 | */ | 7443 | */ |
7440 | static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) | 7444 | static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) |
7441 | { | 7445 | { |
7442 | int continue_balancing = 1; | 7446 | int continue_balancing = 1; |
7443 | int cpu = rq->cpu; | 7447 | int cpu = rq->cpu; |
7444 | unsigned long interval; | 7448 | unsigned long interval; |
7445 | struct sched_domain *sd; | 7449 | struct sched_domain *sd; |
7446 | /* Earliest time when we have to do rebalance again */ | 7450 | /* Earliest time when we have to do rebalance again */ |
7447 | unsigned long next_balance = jiffies + 60*HZ; | 7451 | unsigned long next_balance = jiffies + 60*HZ; |
7448 | int update_next_balance = 0; | 7452 | int update_next_balance = 0; |
7449 | int need_serialize, need_decay = 0; | 7453 | int need_serialize, need_decay = 0; |
7450 | u64 max_cost = 0; | 7454 | u64 max_cost = 0; |
7451 | 7455 | ||
7452 | update_blocked_averages(cpu); | 7456 | update_blocked_averages(cpu); |
7453 | 7457 | ||
7454 | rcu_read_lock(); | 7458 | rcu_read_lock(); |
7455 | for_each_domain(cpu, sd) { | 7459 | for_each_domain(cpu, sd) { |
7456 | /* | 7460 | /* |
7457 | * Decay the newidle max times here because this is a regular | 7461 | * Decay the newidle max times here because this is a regular |
7458 | * visit to all the domains. Decay ~1% per second. | 7462 | * visit to all the domains. Decay ~1% per second. |
7459 | */ | 7463 | */ |
7460 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | 7464 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { |
7461 | sd->max_newidle_lb_cost = | 7465 | sd->max_newidle_lb_cost = |
7462 | (sd->max_newidle_lb_cost * 253) / 256; | 7466 | (sd->max_newidle_lb_cost * 253) / 256; |
7463 | sd->next_decay_max_lb_cost = jiffies + HZ; | 7467 | sd->next_decay_max_lb_cost = jiffies + HZ; |
7464 | need_decay = 1; | 7468 | need_decay = 1; |
7465 | } | 7469 | } |
7466 | max_cost += sd->max_newidle_lb_cost; | 7470 | max_cost += sd->max_newidle_lb_cost; |
7467 | 7471 | ||
7468 | if (!(sd->flags & SD_LOAD_BALANCE)) | 7472 | if (!(sd->flags & SD_LOAD_BALANCE)) |
7469 | continue; | 7473 | continue; |
7470 | 7474 | ||
7471 | /* | 7475 | /* |
7472 | * Stop the load balance at this level. There is another | 7476 | * Stop the load balance at this level. There is another |
7473 | * CPU in our sched group which is doing load balancing more | 7477 | * CPU in our sched group which is doing load balancing more |
7474 | * actively. | 7478 | * actively. |
7475 | */ | 7479 | */ |
7476 | if (!continue_balancing) { | 7480 | if (!continue_balancing) { |
7477 | if (need_decay) | 7481 | if (need_decay) |
7478 | continue; | 7482 | continue; |
7479 | break; | 7483 | break; |
7480 | } | 7484 | } |
7481 | 7485 | ||
7482 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); | 7486 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); |
7483 | 7487 | ||
7484 | need_serialize = sd->flags & SD_SERIALIZE; | 7488 | need_serialize = sd->flags & SD_SERIALIZE; |
7485 | if (need_serialize) { | 7489 | if (need_serialize) { |
7486 | if (!spin_trylock(&balancing)) | 7490 | if (!spin_trylock(&balancing)) |
7487 | goto out; | 7491 | goto out; |
7488 | } | 7492 | } |
7489 | 7493 | ||
7490 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | 7494 | if (time_after_eq(jiffies, sd->last_balance + interval)) { |
7491 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { | 7495 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { |
7492 | /* | 7496 | /* |
7493 | * The LBF_DST_PINNED logic could have changed | 7497 | * The LBF_DST_PINNED logic could have changed |
7494 | * env->dst_cpu, so we can't know our idle | 7498 | * env->dst_cpu, so we can't know our idle |
7495 | * state even if we migrated tasks. Update it. | 7499 | * state even if we migrated tasks. Update it. |
7496 | */ | 7500 | */ |
7497 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; | 7501 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; |
7498 | } | 7502 | } |
7499 | sd->last_balance = jiffies; | 7503 | sd->last_balance = jiffies; |
7500 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); | 7504 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); |
7501 | } | 7505 | } |
7502 | if (need_serialize) | 7506 | if (need_serialize) |
7503 | spin_unlock(&balancing); | 7507 | spin_unlock(&balancing); |
7504 | out: | 7508 | out: |
7505 | if (time_after(next_balance, sd->last_balance + interval)) { | 7509 | if (time_after(next_balance, sd->last_balance + interval)) { |
7506 | next_balance = sd->last_balance + interval; | 7510 | next_balance = sd->last_balance + interval; |
7507 | update_next_balance = 1; | 7511 | update_next_balance = 1; |
7508 | } | 7512 | } |
7509 | } | 7513 | } |
7510 | if (need_decay) { | 7514 | if (need_decay) { |
7511 | /* | 7515 | /* |
7512 | * Ensure the rq-wide value also decays but keep it at a | 7516 | * Ensure the rq-wide value also decays but keep it at a |
7513 | * reasonable floor to avoid funnies with rq->avg_idle. | 7517 | * reasonable floor to avoid funnies with rq->avg_idle. |
7514 | */ | 7518 | */ |
7515 | rq->max_idle_balance_cost = | 7519 | rq->max_idle_balance_cost = |
7516 | max((u64)sysctl_sched_migration_cost, max_cost); | 7520 | max((u64)sysctl_sched_migration_cost, max_cost); |
7517 | } | 7521 | } |
7518 | rcu_read_unlock(); | 7522 | rcu_read_unlock(); |
7519 | 7523 | ||
7520 | /* | 7524 | /* |
7521 | * next_balance will be updated only when there is a need. | 7525 | * next_balance will be updated only when there is a need. |
7522 | * When the cpu is attached to null domain for ex, it will not be | 7526 | * When the cpu is attached to null domain for ex, it will not be |
7523 | * updated. | 7527 | * updated. |
7524 | */ | 7528 | */ |
7525 | if (likely(update_next_balance)) | 7529 | if (likely(update_next_balance)) |
7526 | rq->next_balance = next_balance; | 7530 | rq->next_balance = next_balance; |
7527 | } | 7531 | } |
7528 | 7532 | ||
7529 | #ifdef CONFIG_NO_HZ_COMMON | 7533 | #ifdef CONFIG_NO_HZ_COMMON |
7530 | /* | 7534 | /* |
7531 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the | 7535 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the |
7532 | * rebalancing for all the cpus for whom scheduler ticks are stopped. | 7536 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
7533 | */ | 7537 | */ |
7534 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) | 7538 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) |
7535 | { | 7539 | { |
7536 | int this_cpu = this_rq->cpu; | 7540 | int this_cpu = this_rq->cpu; |
7537 | struct rq *rq; | 7541 | struct rq *rq; |
7538 | int balance_cpu; | 7542 | int balance_cpu; |
7539 | 7543 | ||
7540 | if (idle != CPU_IDLE || | 7544 | if (idle != CPU_IDLE || |
7541 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | 7545 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) |
7542 | goto end; | 7546 | goto end; |
7543 | 7547 | ||
7544 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | 7548 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { |
7545 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) | 7549 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
7546 | continue; | 7550 | continue; |
7547 | 7551 | ||
7548 | /* | 7552 | /* |
7549 | * If this cpu gets work to do, stop the load balancing | 7553 | * If this cpu gets work to do, stop the load balancing |
7550 | * work being done for other cpus. Next load | 7554 | * work being done for other cpus. Next load |
7551 | * balancing owner will pick it up. | 7555 | * balancing owner will pick it up. |
7552 | */ | 7556 | */ |
7553 | if (need_resched()) | 7557 | if (need_resched()) |
7554 | break; | 7558 | break; |
7555 | 7559 | ||
7556 | rq = cpu_rq(balance_cpu); | 7560 | rq = cpu_rq(balance_cpu); |
7557 | 7561 | ||
7558 | /* | 7562 | /* |
7559 | * If time for next balance is due, | 7563 | * If time for next balance is due, |
7560 | * do the balance. | 7564 | * do the balance. |
7561 | */ | 7565 | */ |
7562 | if (time_after_eq(jiffies, rq->next_balance)) { | 7566 | if (time_after_eq(jiffies, rq->next_balance)) { |
7563 | raw_spin_lock_irq(&rq->lock); | 7567 | raw_spin_lock_irq(&rq->lock); |
7564 | update_rq_clock(rq); | 7568 | update_rq_clock(rq); |
7565 | update_idle_cpu_load(rq); | 7569 | update_idle_cpu_load(rq); |
7566 | raw_spin_unlock_irq(&rq->lock); | 7570 | raw_spin_unlock_irq(&rq->lock); |
7567 | rebalance_domains(rq, CPU_IDLE); | 7571 | rebalance_domains(rq, CPU_IDLE); |
7568 | } | 7572 | } |
7569 | 7573 | ||
7570 | if (time_after(this_rq->next_balance, rq->next_balance)) | 7574 | if (time_after(this_rq->next_balance, rq->next_balance)) |
7571 | this_rq->next_balance = rq->next_balance; | 7575 | this_rq->next_balance = rq->next_balance; |
7572 | } | 7576 | } |
7573 | nohz.next_balance = this_rq->next_balance; | 7577 | nohz.next_balance = this_rq->next_balance; |
7574 | end: | 7578 | end: |
7575 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | 7579 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); |
7576 | } | 7580 | } |
7577 | 7581 | ||
7578 | /* | 7582 | /* |
7579 | * Current heuristic for kicking the idle load balancer in the presence | 7583 | * Current heuristic for kicking the idle load balancer in the presence |
7580 | * of an idle cpu is the system. | 7584 | * of an idle cpu is the system. |
7581 | * - This rq has more than one task. | 7585 | * - This rq has more than one task. |
7582 | * - At any scheduler domain level, this cpu's scheduler group has multiple | 7586 | * - At any scheduler domain level, this cpu's scheduler group has multiple |
7583 | * busy cpu's exceeding the group's capacity. | 7587 | * busy cpu's exceeding the group's capacity. |
7584 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler | 7588 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler |
7585 | * domain span are idle. | 7589 | * domain span are idle. |
7586 | */ | 7590 | */ |
7587 | static inline int nohz_kick_needed(struct rq *rq) | 7591 | static inline int nohz_kick_needed(struct rq *rq) |
7588 | { | 7592 | { |
7589 | unsigned long now = jiffies; | 7593 | unsigned long now = jiffies; |
7590 | struct sched_domain *sd; | 7594 | struct sched_domain *sd; |
7591 | struct sched_group_capacity *sgc; | 7595 | struct sched_group_capacity *sgc; |
7592 | int nr_busy, cpu = rq->cpu; | 7596 | int nr_busy, cpu = rq->cpu; |
7593 | 7597 | ||
7594 | if (unlikely(rq->idle_balance)) | 7598 | if (unlikely(rq->idle_balance)) |
7595 | return 0; | 7599 | return 0; |
7596 | 7600 | ||
7597 | /* | 7601 | /* |
7598 | * We may be recently in ticked or tickless idle mode. At the first | 7602 | * We may be recently in ticked or tickless idle mode. At the first |
7599 | * busy tick after returning from idle, we will update the busy stats. | 7603 | * busy tick after returning from idle, we will update the busy stats. |
7600 | */ | 7604 | */ |
7601 | set_cpu_sd_state_busy(); | 7605 | set_cpu_sd_state_busy(); |
7602 | nohz_balance_exit_idle(cpu); | 7606 | nohz_balance_exit_idle(cpu); |
7603 | 7607 | ||
7604 | /* | 7608 | /* |
7605 | * None are in tickless mode and hence no need for NOHZ idle load | 7609 | * None are in tickless mode and hence no need for NOHZ idle load |
7606 | * balancing. | 7610 | * balancing. |
7607 | */ | 7611 | */ |
7608 | if (likely(!atomic_read(&nohz.nr_cpus))) | 7612 | if (likely(!atomic_read(&nohz.nr_cpus))) |
7609 | return 0; | 7613 | return 0; |
7610 | 7614 | ||
7611 | if (time_before(now, nohz.next_balance)) | 7615 | if (time_before(now, nohz.next_balance)) |
7612 | return 0; | 7616 | return 0; |
7613 | 7617 | ||
7614 | if (rq->nr_running >= 2) | 7618 | if (rq->nr_running >= 2) |
7615 | goto need_kick; | 7619 | goto need_kick; |
7616 | 7620 | ||
7617 | rcu_read_lock(); | 7621 | rcu_read_lock(); |
7618 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); | 7622 | sd = rcu_dereference(per_cpu(sd_busy, cpu)); |
7619 | 7623 | ||
7620 | if (sd) { | 7624 | if (sd) { |
7621 | sgc = sd->groups->sgc; | 7625 | sgc = sd->groups->sgc; |
7622 | nr_busy = atomic_read(&sgc->nr_busy_cpus); | 7626 | nr_busy = atomic_read(&sgc->nr_busy_cpus); |
7623 | 7627 | ||
7624 | if (nr_busy > 1) | 7628 | if (nr_busy > 1) |
7625 | goto need_kick_unlock; | 7629 | goto need_kick_unlock; |
7626 | } | 7630 | } |
7627 | 7631 | ||
7628 | sd = rcu_dereference(per_cpu(sd_asym, cpu)); | 7632 | sd = rcu_dereference(per_cpu(sd_asym, cpu)); |
7629 | 7633 | ||
7630 | if (sd && (cpumask_first_and(nohz.idle_cpus_mask, | 7634 | if (sd && (cpumask_first_and(nohz.idle_cpus_mask, |
7631 | sched_domain_span(sd)) < cpu)) | 7635 | sched_domain_span(sd)) < cpu)) |
7632 | goto need_kick_unlock; | 7636 | goto need_kick_unlock; |
7633 | 7637 | ||
7634 | rcu_read_unlock(); | 7638 | rcu_read_unlock(); |
7635 | return 0; | 7639 | return 0; |
7636 | 7640 | ||
7637 | need_kick_unlock: | 7641 | need_kick_unlock: |
7638 | rcu_read_unlock(); | 7642 | rcu_read_unlock(); |
7639 | need_kick: | 7643 | need_kick: |
7640 | return 1; | 7644 | return 1; |
7641 | } | 7645 | } |
7642 | #else | 7646 | #else |
7643 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { } | 7647 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { } |
7644 | #endif | 7648 | #endif |
7645 | 7649 | ||
7646 | /* | 7650 | /* |
7647 | * run_rebalance_domains is triggered when needed from the scheduler tick. | 7651 | * run_rebalance_domains is triggered when needed from the scheduler tick. |
7648 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | 7652 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). |
7649 | */ | 7653 | */ |
7650 | static void run_rebalance_domains(struct softirq_action *h) | 7654 | static void run_rebalance_domains(struct softirq_action *h) |
7651 | { | 7655 | { |
7652 | struct rq *this_rq = this_rq(); | 7656 | struct rq *this_rq = this_rq(); |
7653 | enum cpu_idle_type idle = this_rq->idle_balance ? | 7657 | enum cpu_idle_type idle = this_rq->idle_balance ? |
7654 | CPU_IDLE : CPU_NOT_IDLE; | 7658 | CPU_IDLE : CPU_NOT_IDLE; |
7655 | 7659 | ||
7656 | rebalance_domains(this_rq, idle); | 7660 | rebalance_domains(this_rq, idle); |
7657 | 7661 | ||
7658 | /* | 7662 | /* |
7659 | * If this cpu has a pending nohz_balance_kick, then do the | 7663 | * If this cpu has a pending nohz_balance_kick, then do the |
7660 | * balancing on behalf of the other idle cpus whose ticks are | 7664 | * balancing on behalf of the other idle cpus whose ticks are |
7661 | * stopped. | 7665 | * stopped. |
7662 | */ | 7666 | */ |
7663 | nohz_idle_balance(this_rq, idle); | 7667 | nohz_idle_balance(this_rq, idle); |
7664 | } | 7668 | } |
7665 | 7669 | ||
7666 | /* | 7670 | /* |
7667 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | 7671 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. |
7668 | */ | 7672 | */ |
7669 | void trigger_load_balance(struct rq *rq) | 7673 | void trigger_load_balance(struct rq *rq) |
7670 | { | 7674 | { |
7671 | /* Don't need to rebalance while attached to NULL domain */ | 7675 | /* Don't need to rebalance while attached to NULL domain */ |
7672 | if (unlikely(on_null_domain(rq))) | 7676 | if (unlikely(on_null_domain(rq))) |
7673 | return; | 7677 | return; |
7674 | 7678 | ||
7675 | if (time_after_eq(jiffies, rq->next_balance)) | 7679 | if (time_after_eq(jiffies, rq->next_balance)) |
7676 | raise_softirq(SCHED_SOFTIRQ); | 7680 | raise_softirq(SCHED_SOFTIRQ); |
7677 | #ifdef CONFIG_NO_HZ_COMMON | 7681 | #ifdef CONFIG_NO_HZ_COMMON |
7678 | if (nohz_kick_needed(rq)) | 7682 | if (nohz_kick_needed(rq)) |
7679 | nohz_balancer_kick(); | 7683 | nohz_balancer_kick(); |
7680 | #endif | 7684 | #endif |
7681 | } | 7685 | } |
7682 | 7686 | ||
7683 | static void rq_online_fair(struct rq *rq) | 7687 | static void rq_online_fair(struct rq *rq) |
7684 | { | 7688 | { |
7685 | update_sysctl(); | 7689 | update_sysctl(); |
7686 | 7690 | ||
7687 | update_runtime_enabled(rq); | 7691 | update_runtime_enabled(rq); |
7688 | } | 7692 | } |
7689 | 7693 | ||
7690 | static void rq_offline_fair(struct rq *rq) | 7694 | static void rq_offline_fair(struct rq *rq) |
7691 | { | 7695 | { |
7692 | update_sysctl(); | 7696 | update_sysctl(); |
7693 | 7697 | ||
7694 | /* Ensure any throttled groups are reachable by pick_next_task */ | 7698 | /* Ensure any throttled groups are reachable by pick_next_task */ |
7695 | unthrottle_offline_cfs_rqs(rq); | 7699 | unthrottle_offline_cfs_rqs(rq); |
7696 | } | 7700 | } |
7697 | 7701 | ||
7698 | #endif /* CONFIG_SMP */ | 7702 | #endif /* CONFIG_SMP */ |
7699 | 7703 | ||
7700 | /* | 7704 | /* |
7701 | * scheduler tick hitting a task of our scheduling class: | 7705 | * scheduler tick hitting a task of our scheduling class: |
7702 | */ | 7706 | */ |
7703 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) | 7707 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
7704 | { | 7708 | { |
7705 | struct cfs_rq *cfs_rq; | 7709 | struct cfs_rq *cfs_rq; |
7706 | struct sched_entity *se = &curr->se; | 7710 | struct sched_entity *se = &curr->se; |
7707 | 7711 | ||
7708 | for_each_sched_entity(se) { | 7712 | for_each_sched_entity(se) { |
7709 | cfs_rq = cfs_rq_of(se); | 7713 | cfs_rq = cfs_rq_of(se); |
7710 | entity_tick(cfs_rq, se, queued); | 7714 | entity_tick(cfs_rq, se, queued); |
7711 | } | 7715 | } |
7712 | 7716 | ||
7713 | if (numabalancing_enabled) | 7717 | if (numabalancing_enabled) |
7714 | task_tick_numa(rq, curr); | 7718 | task_tick_numa(rq, curr); |
7715 | 7719 | ||
7716 | update_rq_runnable_avg(rq, 1); | 7720 | update_rq_runnable_avg(rq, 1); |
7717 | } | 7721 | } |
7718 | 7722 | ||
7719 | /* | 7723 | /* |
7720 | * called on fork with the child task as argument from the parent's context | 7724 | * called on fork with the child task as argument from the parent's context |
7721 | * - child not yet on the tasklist | 7725 | * - child not yet on the tasklist |
7722 | * - preemption disabled | 7726 | * - preemption disabled |
7723 | */ | 7727 | */ |
7724 | static void task_fork_fair(struct task_struct *p) | 7728 | static void task_fork_fair(struct task_struct *p) |
7725 | { | 7729 | { |
7726 | struct cfs_rq *cfs_rq; | 7730 | struct cfs_rq *cfs_rq; |
7727 | struct sched_entity *se = &p->se, *curr; | 7731 | struct sched_entity *se = &p->se, *curr; |
7728 | int this_cpu = smp_processor_id(); | 7732 | int this_cpu = smp_processor_id(); |
7729 | struct rq *rq = this_rq(); | 7733 | struct rq *rq = this_rq(); |
7730 | unsigned long flags; | 7734 | unsigned long flags; |
7731 | 7735 | ||
7732 | raw_spin_lock_irqsave(&rq->lock, flags); | 7736 | raw_spin_lock_irqsave(&rq->lock, flags); |
7733 | 7737 | ||
7734 | update_rq_clock(rq); | 7738 | update_rq_clock(rq); |
7735 | 7739 | ||
7736 | cfs_rq = task_cfs_rq(current); | 7740 | cfs_rq = task_cfs_rq(current); |
7737 | curr = cfs_rq->curr; | 7741 | curr = cfs_rq->curr; |
7738 | 7742 | ||
7739 | /* | 7743 | /* |
7740 | * Not only the cpu but also the task_group of the parent might have | 7744 | * Not only the cpu but also the task_group of the parent might have |
7741 | * been changed after parent->se.parent,cfs_rq were copied to | 7745 | * been changed after parent->se.parent,cfs_rq were copied to |
7742 | * child->se.parent,cfs_rq. So call __set_task_cpu() to make those | 7746 | * child->se.parent,cfs_rq. So call __set_task_cpu() to make those |
7743 | * of child point to valid ones. | 7747 | * of child point to valid ones. |
7744 | */ | 7748 | */ |
7745 | rcu_read_lock(); | 7749 | rcu_read_lock(); |
7746 | __set_task_cpu(p, this_cpu); | 7750 | __set_task_cpu(p, this_cpu); |
7747 | rcu_read_unlock(); | 7751 | rcu_read_unlock(); |
7748 | 7752 | ||
7749 | update_curr(cfs_rq); | 7753 | update_curr(cfs_rq); |
7750 | 7754 | ||
7751 | if (curr) | 7755 | if (curr) |
7752 | se->vruntime = curr->vruntime; | 7756 | se->vruntime = curr->vruntime; |
7753 | place_entity(cfs_rq, se, 1); | 7757 | place_entity(cfs_rq, se, 1); |
7754 | 7758 | ||
7755 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { | 7759 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
7756 | /* | 7760 | /* |
7757 | * Upon rescheduling, sched_class::put_prev_task() will place | 7761 | * Upon rescheduling, sched_class::put_prev_task() will place |
7758 | * 'current' within the tree based on its new key value. | 7762 | * 'current' within the tree based on its new key value. |
7759 | */ | 7763 | */ |
7760 | swap(curr->vruntime, se->vruntime); | 7764 | swap(curr->vruntime, se->vruntime); |
7761 | resched_curr(rq); | 7765 | resched_curr(rq); |
7762 | } | 7766 | } |
7763 | 7767 | ||
7764 | se->vruntime -= cfs_rq->min_vruntime; | 7768 | se->vruntime -= cfs_rq->min_vruntime; |
7765 | 7769 | ||
7766 | raw_spin_unlock_irqrestore(&rq->lock, flags); | 7770 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
7767 | } | 7771 | } |
7768 | 7772 | ||
7769 | /* | 7773 | /* |
7770 | * Priority of the task has changed. Check to see if we preempt | 7774 | * Priority of the task has changed. Check to see if we preempt |
7771 | * the current task. | 7775 | * the current task. |
7772 | */ | 7776 | */ |
7773 | static void | 7777 | static void |
7774 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | 7778 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) |
7775 | { | 7779 | { |
7776 | if (!task_on_rq_queued(p)) | 7780 | if (!task_on_rq_queued(p)) |
7777 | return; | 7781 | return; |
7778 | 7782 | ||
7779 | /* | 7783 | /* |
7780 | * Reschedule if we are currently running on this runqueue and | 7784 | * Reschedule if we are currently running on this runqueue and |
7781 | * our priority decreased, or if we are not currently running on | 7785 | * our priority decreased, or if we are not currently running on |
7782 | * this runqueue and our priority is higher than the current's | 7786 | * this runqueue and our priority is higher than the current's |
7783 | */ | 7787 | */ |
7784 | if (rq->curr == p) { | 7788 | if (rq->curr == p) { |
7785 | if (p->prio > oldprio) | 7789 | if (p->prio > oldprio) |
7786 | resched_curr(rq); | 7790 | resched_curr(rq); |
7787 | } else | 7791 | } else |
7788 | check_preempt_curr(rq, p, 0); | 7792 | check_preempt_curr(rq, p, 0); |
7789 | } | 7793 | } |
7790 | 7794 | ||
7791 | static void switched_from_fair(struct rq *rq, struct task_struct *p) | 7795 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
7792 | { | 7796 | { |
7793 | struct sched_entity *se = &p->se; | 7797 | struct sched_entity *se = &p->se; |
7794 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | 7798 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
7795 | 7799 | ||
7796 | /* | 7800 | /* |
7797 | * Ensure the task's vruntime is normalized, so that when it's | 7801 | * Ensure the task's vruntime is normalized, so that when it's |
7798 | * switched back to the fair class the enqueue_entity(.flags=0) will | 7802 | * switched back to the fair class the enqueue_entity(.flags=0) will |
7799 | * do the right thing. | 7803 | * do the right thing. |
7800 | * | 7804 | * |
7801 | * If it's queued, then the dequeue_entity(.flags=0) will already | 7805 | * If it's queued, then the dequeue_entity(.flags=0) will already |
7802 | * have normalized the vruntime, if it's !queued, then only when | 7806 | * have normalized the vruntime, if it's !queued, then only when |
7803 | * the task is sleeping will it still have non-normalized vruntime. | 7807 | * the task is sleeping will it still have non-normalized vruntime. |
7804 | */ | 7808 | */ |
7805 | if (!task_on_rq_queued(p) && p->state != TASK_RUNNING) { | 7809 | if (!task_on_rq_queued(p) && p->state != TASK_RUNNING) { |
7806 | /* | 7810 | /* |
7807 | * Fix up our vruntime so that the current sleep doesn't | 7811 | * Fix up our vruntime so that the current sleep doesn't |
7808 | * cause 'unlimited' sleep bonus. | 7812 | * cause 'unlimited' sleep bonus. |
7809 | */ | 7813 | */ |
7810 | place_entity(cfs_rq, se, 0); | 7814 | place_entity(cfs_rq, se, 0); |
7811 | se->vruntime -= cfs_rq->min_vruntime; | 7815 | se->vruntime -= cfs_rq->min_vruntime; |
7812 | } | 7816 | } |
7813 | 7817 | ||
7814 | #ifdef CONFIG_SMP | 7818 | #ifdef CONFIG_SMP |
7815 | /* | 7819 | /* |
7816 | * Remove our load from contribution when we leave sched_fair | 7820 | * Remove our load from contribution when we leave sched_fair |
7817 | * and ensure we don't carry in an old decay_count if we | 7821 | * and ensure we don't carry in an old decay_count if we |
7818 | * switch back. | 7822 | * switch back. |
7819 | */ | 7823 | */ |
7820 | if (se->avg.decay_count) { | 7824 | if (se->avg.decay_count) { |
7821 | __synchronize_entity_decay(se); | 7825 | __synchronize_entity_decay(se); |
7822 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); | 7826 | subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); |
7823 | } | 7827 | } |
7824 | #endif | 7828 | #endif |
7825 | } | 7829 | } |
7826 | 7830 | ||
7827 | /* | 7831 | /* |
7828 | * We switched to the sched_fair class. | 7832 | * We switched to the sched_fair class. |
7829 | */ | 7833 | */ |
7830 | static void switched_to_fair(struct rq *rq, struct task_struct *p) | 7834 | static void switched_to_fair(struct rq *rq, struct task_struct *p) |
7831 | { | 7835 | { |
7832 | #ifdef CONFIG_FAIR_GROUP_SCHED | 7836 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7833 | struct sched_entity *se = &p->se; | 7837 | struct sched_entity *se = &p->se; |
7834 | /* | 7838 | /* |
7835 | * Since the real-depth could have been changed (only FAIR | 7839 | * Since the real-depth could have been changed (only FAIR |
7836 | * class maintain depth value), reset depth properly. | 7840 | * class maintain depth value), reset depth properly. |
7837 | */ | 7841 | */ |
7838 | se->depth = se->parent ? se->parent->depth + 1 : 0; | 7842 | se->depth = se->parent ? se->parent->depth + 1 : 0; |
7839 | #endif | 7843 | #endif |
7840 | if (!task_on_rq_queued(p)) | 7844 | if (!task_on_rq_queued(p)) |
7841 | return; | 7845 | return; |
7842 | 7846 | ||
7843 | /* | 7847 | /* |
7844 | * We were most likely switched from sched_rt, so | 7848 | * We were most likely switched from sched_rt, so |
7845 | * kick off the schedule if running, otherwise just see | 7849 | * kick off the schedule if running, otherwise just see |
7846 | * if we can still preempt the current task. | 7850 | * if we can still preempt the current task. |
7847 | */ | 7851 | */ |
7848 | if (rq->curr == p) | 7852 | if (rq->curr == p) |
7849 | resched_curr(rq); | 7853 | resched_curr(rq); |
7850 | else | 7854 | else |
7851 | check_preempt_curr(rq, p, 0); | 7855 | check_preempt_curr(rq, p, 0); |
7852 | } | 7856 | } |
7853 | 7857 | ||
7854 | /* Account for a task changing its policy or group. | 7858 | /* Account for a task changing its policy or group. |
7855 | * | 7859 | * |
7856 | * This routine is mostly called to set cfs_rq->curr field when a task | 7860 | * This routine is mostly called to set cfs_rq->curr field when a task |
7857 | * migrates between groups/classes. | 7861 | * migrates between groups/classes. |
7858 | */ | 7862 | */ |
7859 | static void set_curr_task_fair(struct rq *rq) | 7863 | static void set_curr_task_fair(struct rq *rq) |
7860 | { | 7864 | { |
7861 | struct sched_entity *se = &rq->curr->se; | 7865 | struct sched_entity *se = &rq->curr->se; |
7862 | 7866 | ||
7863 | for_each_sched_entity(se) { | 7867 | for_each_sched_entity(se) { |
7864 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | 7868 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
7865 | 7869 | ||
7866 | set_next_entity(cfs_rq, se); | 7870 | set_next_entity(cfs_rq, se); |
7867 | /* ensure bandwidth has been allocated on our new cfs_rq */ | 7871 | /* ensure bandwidth has been allocated on our new cfs_rq */ |
7868 | account_cfs_rq_runtime(cfs_rq, 0); | 7872 | account_cfs_rq_runtime(cfs_rq, 0); |
7869 | } | 7873 | } |
7870 | } | 7874 | } |
7871 | 7875 | ||
7872 | void init_cfs_rq(struct cfs_rq *cfs_rq) | 7876 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
7873 | { | 7877 | { |
7874 | cfs_rq->tasks_timeline = RB_ROOT; | 7878 | cfs_rq->tasks_timeline = RB_ROOT; |
7875 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); | 7879 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
7876 | #ifndef CONFIG_64BIT | 7880 | #ifndef CONFIG_64BIT |
7877 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | 7881 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; |
7878 | #endif | 7882 | #endif |
7879 | #ifdef CONFIG_SMP | 7883 | #ifdef CONFIG_SMP |
7880 | atomic64_set(&cfs_rq->decay_counter, 1); | 7884 | atomic64_set(&cfs_rq->decay_counter, 1); |
7881 | atomic_long_set(&cfs_rq->removed_load, 0); | 7885 | atomic_long_set(&cfs_rq->removed_load, 0); |
7882 | #endif | 7886 | #endif |
7883 | } | 7887 | } |
7884 | 7888 | ||
7885 | #ifdef CONFIG_FAIR_GROUP_SCHED | 7889 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7886 | static void task_move_group_fair(struct task_struct *p, int queued) | 7890 | static void task_move_group_fair(struct task_struct *p, int queued) |
7887 | { | 7891 | { |
7888 | struct sched_entity *se = &p->se; | 7892 | struct sched_entity *se = &p->se; |
7889 | struct cfs_rq *cfs_rq; | 7893 | struct cfs_rq *cfs_rq; |
7890 | 7894 | ||
7891 | /* | 7895 | /* |
7892 | * If the task was not on the rq at the time of this cgroup movement | 7896 | * If the task was not on the rq at the time of this cgroup movement |
7893 | * it must have been asleep, sleeping tasks keep their ->vruntime | 7897 | * it must have been asleep, sleeping tasks keep their ->vruntime |
7894 | * absolute on their old rq until wakeup (needed for the fair sleeper | 7898 | * absolute on their old rq until wakeup (needed for the fair sleeper |
7895 | * bonus in place_entity()). | 7899 | * bonus in place_entity()). |
7896 | * | 7900 | * |
7897 | * If it was on the rq, we've just 'preempted' it, which does convert | 7901 | * If it was on the rq, we've just 'preempted' it, which does convert |
7898 | * ->vruntime to a relative base. | 7902 | * ->vruntime to a relative base. |
7899 | * | 7903 | * |
7900 | * Make sure both cases convert their relative position when migrating | 7904 | * Make sure both cases convert their relative position when migrating |
7901 | * to another cgroup's rq. This does somewhat interfere with the | 7905 | * to another cgroup's rq. This does somewhat interfere with the |
7902 | * fair sleeper stuff for the first placement, but who cares. | 7906 | * fair sleeper stuff for the first placement, but who cares. |
7903 | */ | 7907 | */ |
7904 | /* | 7908 | /* |
7905 | * When !queued, vruntime of the task has usually NOT been normalized. | 7909 | * When !queued, vruntime of the task has usually NOT been normalized. |
7906 | * But there are some cases where it has already been normalized: | 7910 | * But there are some cases where it has already been normalized: |
7907 | * | 7911 | * |
7908 | * - Moving a forked child which is waiting for being woken up by | 7912 | * - Moving a forked child which is waiting for being woken up by |
7909 | * wake_up_new_task(). | 7913 | * wake_up_new_task(). |
7910 | * - Moving a task which has been woken up by try_to_wake_up() and | 7914 | * - Moving a task which has been woken up by try_to_wake_up() and |
7911 | * waiting for actually being woken up by sched_ttwu_pending(). | 7915 | * waiting for actually being woken up by sched_ttwu_pending(). |
7912 | * | 7916 | * |
7913 | * To prevent boost or penalty in the new cfs_rq caused by delta | 7917 | * To prevent boost or penalty in the new cfs_rq caused by delta |
7914 | * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. | 7918 | * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. |
7915 | */ | 7919 | */ |
7916 | if (!queued && (!se->sum_exec_runtime || p->state == TASK_WAKING)) | 7920 | if (!queued && (!se->sum_exec_runtime || p->state == TASK_WAKING)) |
7917 | queued = 1; | 7921 | queued = 1; |
7918 | 7922 | ||
7919 | if (!queued) | 7923 | if (!queued) |
7920 | se->vruntime -= cfs_rq_of(se)->min_vruntime; | 7924 | se->vruntime -= cfs_rq_of(se)->min_vruntime; |
7921 | set_task_rq(p, task_cpu(p)); | 7925 | set_task_rq(p, task_cpu(p)); |
7922 | se->depth = se->parent ? se->parent->depth + 1 : 0; | 7926 | se->depth = se->parent ? se->parent->depth + 1 : 0; |
7923 | if (!queued) { | 7927 | if (!queued) { |
7924 | cfs_rq = cfs_rq_of(se); | 7928 | cfs_rq = cfs_rq_of(se); |
7925 | se->vruntime += cfs_rq->min_vruntime; | 7929 | se->vruntime += cfs_rq->min_vruntime; |
7926 | #ifdef CONFIG_SMP | 7930 | #ifdef CONFIG_SMP |
7927 | /* | 7931 | /* |
7928 | * migrate_task_rq_fair() will have removed our previous | 7932 | * migrate_task_rq_fair() will have removed our previous |
7929 | * contribution, but we must synchronize for ongoing future | 7933 | * contribution, but we must synchronize for ongoing future |
7930 | * decay. | 7934 | * decay. |
7931 | */ | 7935 | */ |
7932 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); | 7936 | se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); |
7933 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; | 7937 | cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; |
7934 | #endif | 7938 | #endif |
7935 | } | 7939 | } |
7936 | } | 7940 | } |
7937 | 7941 | ||
7938 | void free_fair_sched_group(struct task_group *tg) | 7942 | void free_fair_sched_group(struct task_group *tg) |
7939 | { | 7943 | { |
7940 | int i; | 7944 | int i; |
7941 | 7945 | ||
7942 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | 7946 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); |
7943 | 7947 | ||
7944 | for_each_possible_cpu(i) { | 7948 | for_each_possible_cpu(i) { |
7945 | if (tg->cfs_rq) | 7949 | if (tg->cfs_rq) |
7946 | kfree(tg->cfs_rq[i]); | 7950 | kfree(tg->cfs_rq[i]); |
7947 | if (tg->se) | 7951 | if (tg->se) |
7948 | kfree(tg->se[i]); | 7952 | kfree(tg->se[i]); |
7949 | } | 7953 | } |
7950 | 7954 | ||
7951 | kfree(tg->cfs_rq); | 7955 | kfree(tg->cfs_rq); |
7952 | kfree(tg->se); | 7956 | kfree(tg->se); |
7953 | } | 7957 | } |
7954 | 7958 | ||
7955 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | 7959 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) |
7956 | { | 7960 | { |
7957 | struct cfs_rq *cfs_rq; | 7961 | struct cfs_rq *cfs_rq; |
7958 | struct sched_entity *se; | 7962 | struct sched_entity *se; |
7959 | int i; | 7963 | int i; |
7960 | 7964 | ||
7961 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | 7965 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); |
7962 | if (!tg->cfs_rq) | 7966 | if (!tg->cfs_rq) |
7963 | goto err; | 7967 | goto err; |
7964 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | 7968 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); |
7965 | if (!tg->se) | 7969 | if (!tg->se) |
7966 | goto err; | 7970 | goto err; |
7967 | 7971 | ||
7968 | tg->shares = NICE_0_LOAD; | 7972 | tg->shares = NICE_0_LOAD; |
7969 | 7973 | ||
7970 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | 7974 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); |
7971 | 7975 | ||
7972 | for_each_possible_cpu(i) { | 7976 | for_each_possible_cpu(i) { |
7973 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | 7977 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), |
7974 | GFP_KERNEL, cpu_to_node(i)); | 7978 | GFP_KERNEL, cpu_to_node(i)); |
7975 | if (!cfs_rq) | 7979 | if (!cfs_rq) |
7976 | goto err; | 7980 | goto err; |
7977 | 7981 | ||
7978 | se = kzalloc_node(sizeof(struct sched_entity), | 7982 | se = kzalloc_node(sizeof(struct sched_entity), |
7979 | GFP_KERNEL, cpu_to_node(i)); | 7983 | GFP_KERNEL, cpu_to_node(i)); |
7980 | if (!se) | 7984 | if (!se) |
7981 | goto err_free_rq; | 7985 | goto err_free_rq; |
7982 | 7986 | ||
7983 | init_cfs_rq(cfs_rq); | 7987 | init_cfs_rq(cfs_rq); |
7984 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | 7988 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); |
7985 | } | 7989 | } |
7986 | 7990 | ||
7987 | return 1; | 7991 | return 1; |
7988 | 7992 | ||
7989 | err_free_rq: | 7993 | err_free_rq: |
7990 | kfree(cfs_rq); | 7994 | kfree(cfs_rq); |
7991 | err: | 7995 | err: |
7992 | return 0; | 7996 | return 0; |
7993 | } | 7997 | } |
7994 | 7998 | ||
7995 | void unregister_fair_sched_group(struct task_group *tg, int cpu) | 7999 | void unregister_fair_sched_group(struct task_group *tg, int cpu) |
7996 | { | 8000 | { |
7997 | struct rq *rq = cpu_rq(cpu); | 8001 | struct rq *rq = cpu_rq(cpu); |
7998 | unsigned long flags; | 8002 | unsigned long flags; |
7999 | 8003 | ||
8000 | /* | 8004 | /* |
8001 | * Only empty task groups can be destroyed; so we can speculatively | 8005 | * Only empty task groups can be destroyed; so we can speculatively |
8002 | * check on_list without danger of it being re-added. | 8006 | * check on_list without danger of it being re-added. |
8003 | */ | 8007 | */ |
8004 | if (!tg->cfs_rq[cpu]->on_list) | 8008 | if (!tg->cfs_rq[cpu]->on_list) |
8005 | return; | 8009 | return; |
8006 | 8010 | ||
8007 | raw_spin_lock_irqsave(&rq->lock, flags); | 8011 | raw_spin_lock_irqsave(&rq->lock, flags); |
8008 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | 8012 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); |
8009 | raw_spin_unlock_irqrestore(&rq->lock, flags); | 8013 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
8010 | } | 8014 | } |
8011 | 8015 | ||
8012 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | 8016 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, |
8013 | struct sched_entity *se, int cpu, | 8017 | struct sched_entity *se, int cpu, |
8014 | struct sched_entity *parent) | 8018 | struct sched_entity *parent) |
8015 | { | 8019 | { |
8016 | struct rq *rq = cpu_rq(cpu); | 8020 | struct rq *rq = cpu_rq(cpu); |
8017 | 8021 | ||
8018 | cfs_rq->tg = tg; | 8022 | cfs_rq->tg = tg; |
8019 | cfs_rq->rq = rq; | 8023 | cfs_rq->rq = rq; |
8020 | init_cfs_rq_runtime(cfs_rq); | 8024 | init_cfs_rq_runtime(cfs_rq); |
8021 | 8025 | ||
8022 | tg->cfs_rq[cpu] = cfs_rq; | 8026 | tg->cfs_rq[cpu] = cfs_rq; |
8023 | tg->se[cpu] = se; | 8027 | tg->se[cpu] = se; |
8024 | 8028 | ||
8025 | /* se could be NULL for root_task_group */ | 8029 | /* se could be NULL for root_task_group */ |
8026 | if (!se) | 8030 | if (!se) |
8027 | return; | 8031 | return; |
8028 | 8032 | ||
8029 | if (!parent) { | 8033 | if (!parent) { |
8030 | se->cfs_rq = &rq->cfs; | 8034 | se->cfs_rq = &rq->cfs; |
8031 | se->depth = 0; | 8035 | se->depth = 0; |
8032 | } else { | 8036 | } else { |
8033 | se->cfs_rq = parent->my_q; | 8037 | se->cfs_rq = parent->my_q; |
8034 | se->depth = parent->depth + 1; | 8038 | se->depth = parent->depth + 1; |
8035 | } | 8039 | } |
8036 | 8040 | ||
8037 | se->my_q = cfs_rq; | 8041 | se->my_q = cfs_rq; |
8038 | /* guarantee group entities always have weight */ | 8042 | /* guarantee group entities always have weight */ |
8039 | update_load_set(&se->load, NICE_0_LOAD); | 8043 | update_load_set(&se->load, NICE_0_LOAD); |
8040 | se->parent = parent; | 8044 | se->parent = parent; |
8041 | } | 8045 | } |
8042 | 8046 | ||
8043 | static DEFINE_MUTEX(shares_mutex); | 8047 | static DEFINE_MUTEX(shares_mutex); |
8044 | 8048 | ||
8045 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | 8049 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) |
8046 | { | 8050 | { |
8047 | int i; | 8051 | int i; |
8048 | unsigned long flags; | 8052 | unsigned long flags; |
8049 | 8053 | ||
8050 | /* | 8054 | /* |
8051 | * We can't change the weight of the root cgroup. | 8055 | * We can't change the weight of the root cgroup. |
8052 | */ | 8056 | */ |
8053 | if (!tg->se[0]) | 8057 | if (!tg->se[0]) |
8054 | return -EINVAL; | 8058 | return -EINVAL; |
8055 | 8059 | ||
8056 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | 8060 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); |
8057 | 8061 | ||
8058 | mutex_lock(&shares_mutex); | 8062 | mutex_lock(&shares_mutex); |
8059 | if (tg->shares == shares) | 8063 | if (tg->shares == shares) |
8060 | goto done; | 8064 | goto done; |
8061 | 8065 | ||
8062 | tg->shares = shares; | 8066 | tg->shares = shares; |
8063 | for_each_possible_cpu(i) { | 8067 | for_each_possible_cpu(i) { |
8064 | struct rq *rq = cpu_rq(i); | 8068 | struct rq *rq = cpu_rq(i); |
8065 | struct sched_entity *se; | 8069 | struct sched_entity *se; |
8066 | 8070 | ||
8067 | se = tg->se[i]; | 8071 | se = tg->se[i]; |
8068 | /* Propagate contribution to hierarchy */ | 8072 | /* Propagate contribution to hierarchy */ |
8069 | raw_spin_lock_irqsave(&rq->lock, flags); | 8073 | raw_spin_lock_irqsave(&rq->lock, flags); |
8070 | 8074 | ||
8071 | /* Possible calls to update_curr() need rq clock */ | 8075 | /* Possible calls to update_curr() need rq clock */ |
8072 | update_rq_clock(rq); | 8076 | update_rq_clock(rq); |
8073 | for_each_sched_entity(se) | 8077 | for_each_sched_entity(se) |
8074 | update_cfs_shares(group_cfs_rq(se)); | 8078 | update_cfs_shares(group_cfs_rq(se)); |
8075 | raw_spin_unlock_irqrestore(&rq->lock, flags); | 8079 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
8076 | } | 8080 | } |
8077 | 8081 | ||
8078 | done: | 8082 | done: |
8079 | mutex_unlock(&shares_mutex); | 8083 | mutex_unlock(&shares_mutex); |
8080 | return 0; | 8084 | return 0; |
8081 | } | 8085 | } |
8082 | #else /* CONFIG_FAIR_GROUP_SCHED */ | 8086 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
8083 | 8087 | ||
8084 | void free_fair_sched_group(struct task_group *tg) { } | 8088 | void free_fair_sched_group(struct task_group *tg) { } |
8085 | 8089 | ||
8086 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | 8090 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) |
8087 | { | 8091 | { |
8088 | return 1; | 8092 | return 1; |
8089 | } | 8093 | } |
8090 | 8094 | ||
8091 | void unregister_fair_sched_group(struct task_group *tg, int cpu) { } | 8095 | void unregister_fair_sched_group(struct task_group *tg, int cpu) { } |
8092 | 8096 | ||
8093 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | 8097 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
8094 | 8098 | ||
8095 | 8099 | ||
8096 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) | 8100 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
8097 | { | 8101 | { |
8098 | struct sched_entity *se = &task->se; | 8102 | struct sched_entity *se = &task->se; |
8099 | unsigned int rr_interval = 0; | 8103 | unsigned int rr_interval = 0; |
8100 | 8104 | ||
8101 | /* | 8105 | /* |
8102 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | 8106 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise |
8103 | * idle runqueue: | 8107 | * idle runqueue: |
8104 | */ | 8108 | */ |
8105 | if (rq->cfs.load.weight) | 8109 | if (rq->cfs.load.weight) |
8106 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); | 8110 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
8107 | 8111 | ||
8108 | return rr_interval; | 8112 | return rr_interval; |
8109 | } | 8113 | } |
8110 | 8114 | ||
8111 | /* | 8115 | /* |
8112 | * All the scheduling class methods: | 8116 | * All the scheduling class methods: |
8113 | */ | 8117 | */ |
8114 | const struct sched_class fair_sched_class = { | 8118 | const struct sched_class fair_sched_class = { |
8115 | .next = &idle_sched_class, | 8119 | .next = &idle_sched_class, |
8116 | .enqueue_task = enqueue_task_fair, | 8120 | .enqueue_task = enqueue_task_fair, |
8117 | .dequeue_task = dequeue_task_fair, | 8121 | .dequeue_task = dequeue_task_fair, |
8118 | .yield_task = yield_task_fair, | 8122 | .yield_task = yield_task_fair, |
8119 | .yield_to_task = yield_to_task_fair, | 8123 | .yield_to_task = yield_to_task_fair, |
8120 | 8124 | ||
8121 | .check_preempt_curr = check_preempt_wakeup, | 8125 | .check_preempt_curr = check_preempt_wakeup, |
8122 | 8126 | ||
8123 | .pick_next_task = pick_next_task_fair, | 8127 | .pick_next_task = pick_next_task_fair, |
8124 | .put_prev_task = put_prev_task_fair, | 8128 | .put_prev_task = put_prev_task_fair, |
8125 | 8129 | ||
8126 | #ifdef CONFIG_SMP | 8130 | #ifdef CONFIG_SMP |
8127 | .select_task_rq = select_task_rq_fair, | 8131 | .select_task_rq = select_task_rq_fair, |
8128 | .migrate_task_rq = migrate_task_rq_fair, | 8132 | .migrate_task_rq = migrate_task_rq_fair, |
8129 | 8133 | ||
8130 | .rq_online = rq_online_fair, | 8134 | .rq_online = rq_online_fair, |
8131 | .rq_offline = rq_offline_fair, | 8135 | .rq_offline = rq_offline_fair, |
8132 | 8136 | ||
8133 | .task_waking = task_waking_fair, | 8137 | .task_waking = task_waking_fair, |
8134 | #endif | 8138 | #endif |
8135 | 8139 | ||
8136 | .set_curr_task = set_curr_task_fair, | 8140 | .set_curr_task = set_curr_task_fair, |
8137 | .task_tick = task_tick_fair, | 8141 | .task_tick = task_tick_fair, |
8138 | .task_fork = task_fork_fair, | 8142 | .task_fork = task_fork_fair, |
8139 | 8143 | ||
8140 | .prio_changed = prio_changed_fair, | 8144 | .prio_changed = prio_changed_fair, |
8141 | .switched_from = switched_from_fair, | 8145 | .switched_from = switched_from_fair, |
8142 | .switched_to = switched_to_fair, | 8146 | .switched_to = switched_to_fair, |
8143 | 8147 | ||
8144 | .get_rr_interval = get_rr_interval_fair, | 8148 | .get_rr_interval = get_rr_interval_fair, |
8145 | 8149 | ||
8146 | .update_curr = update_curr_fair, | 8150 | .update_curr = update_curr_fair, |
8147 | 8151 | ||
8148 | #ifdef CONFIG_FAIR_GROUP_SCHED | 8152 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8149 | .task_move_group = task_move_group_fair, | 8153 | .task_move_group = task_move_group_fair, |
8150 | #endif | 8154 | #endif |
8151 | }; | 8155 | }; |
8152 | 8156 | ||
8153 | #ifdef CONFIG_SCHED_DEBUG | 8157 | #ifdef CONFIG_SCHED_DEBUG |
8154 | void print_cfs_stats(struct seq_file *m, int cpu) | 8158 | void print_cfs_stats(struct seq_file *m, int cpu) |
8155 | { | 8159 | { |
8156 | struct cfs_rq *cfs_rq; | 8160 | struct cfs_rq *cfs_rq; |
8157 | 8161 | ||
8158 | rcu_read_lock(); | 8162 | rcu_read_lock(); |
8159 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) | 8163 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
8160 | print_cfs_rq(m, cpu, cfs_rq); | 8164 | print_cfs_rq(m, cpu, cfs_rq); |
8161 | rcu_read_unlock(); | 8165 | rcu_read_unlock(); |
8162 | } | 8166 | } |
8163 | #endif | 8167 | #endif |
8164 | 8168 | ||
8165 | __init void init_sched_fair_class(void) | 8169 | __init void init_sched_fair_class(void) |
8166 | { | 8170 | { |
8167 | #ifdef CONFIG_SMP | 8171 | #ifdef CONFIG_SMP |
8168 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | 8172 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); |
8169 | 8173 | ||
8170 | #ifdef CONFIG_NO_HZ_COMMON | 8174 | #ifdef CONFIG_NO_HZ_COMMON |
8171 | nohz.next_balance = jiffies; | 8175 | nohz.next_balance = jiffies; |
8172 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); | 8176 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
8173 | cpu_notifier(sched_ilb_notifier, 0); | 8177 | cpu_notifier(sched_ilb_notifier, 0); |
8174 | #endif | 8178 | #endif |
8175 | #endif /* SMP */ | 8179 | #endif /* SMP */ |
8176 | 8180 | ||
8177 | } | 8181 | } |
8178 | 8182 |