Eric Lee / smarc-ti-linux-kernel

18
Download as
Browse Code ยป

Commit 0f397f2c90ce68821ee864c2c53baafe78de765d

Authored by Kirill Tkhai
Committed by Ingo Molnar
1 parent b14ed2c273
Exists in ti-lsk-linux-4.1.y and in 12 other branches dlt-processor-sdk-linux-03.00.00.04, processor-sdk-linux-02.00.01, smarc-ti-linux-3.15.y, smarc-ti-lsk-linux-4.1.y, smarct3x-processor-sdk-04.01.00.06, smarct3x-processor-sdk-linux-02.00.01, smarct3x-processor-sdk-linux-03.00.00.04, smarct4x-800-processor-sdk-linux-02.00.01, smarct4x-processor-sdk-04.01.00.06, smarct4x-processor-sdk-linux-02.00.01, smarct4x-processor-sdk-linux-03.00.00.04, ti-linux-3.15.y

sched/dl: Fix race in dl_task_timer() 

Throttled task is still on rq, and it may be moved to other cpu
if user is playing with sched_setaffinity(). Therefore, unlocked
task_rq() access makes the race.

Juri Lelli reports he got this race when dl_bandwidth_enabled()
was not set.

Other thing, pointed by Peter Zijlstra:

   "Now I suppose the problem can still actually happen when
    you change the root domain and trigger a effective affinity
    change that way".

To fix that we do the same as made in __task_rq_lock(). We do not
use __task_rq_lock() itself, because it has a useful lockdep check,
which is not correct in case of dl_task_timer(). We do not need
pi_lock locked here. This case is an exception (PeterZ):

   "The only reason we don't strictly need ->pi_lock now is because
    we're guaranteed to have p->state == TASK_RUNNING here and are
    thus free of ttwu races".

Signed-off-by: Kirill Tkhai <tkhai@yandex.ru>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: <stable@vger.kernel.org> # v3.14+
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/3056991400578422@web14g.yandex.ru
Signed-off-by: Ingo Molnar <mingo@kernel.org>

Showing 1 changed file with 9 additions and 1 deletions Inline Diff

kernel/sched/deadline.c
Diff comments   View file @ 0f397f2
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 extern unsigned long to_ratio(u64 period, u64 runtime); 60 extern unsigned long to_ratio(u64 period, u64 runtime);
61 61
62 void init_dl_bw(struct dl_bw *dl_b) 62 void init_dl_bw(struct dl_bw *dl_b)
63 { 63 {
64 raw_spin_lock_init(&dl_b->lock); 64 raw_spin_lock_init(&dl_b->lock);
65 raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock); 65 raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock);
66 if (global_rt_runtime() == RUNTIME_INF) 66 if (global_rt_runtime() == RUNTIME_INF)
67 dl_b->bw = -1; 67 dl_b->bw = -1;
68 else 68 else
69 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime()); 69 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
70 raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock); 70 raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock);
71 dl_b->total_bw = 0; 71 dl_b->total_bw = 0;
72 } 72 }
73 73
74 void init_dl_rq(struct dl_rq *dl_rq, struct rq *rq) 74 void init_dl_rq(struct dl_rq *dl_rq, struct rq *rq)
75 { 75 {
76 dl_rq->rb_root = RB_ROOT; 76 dl_rq->rb_root = RB_ROOT;
77 77
78 #ifdef CONFIG_SMP 78 #ifdef CONFIG_SMP
79 /* zero means no -deadline tasks */ 79 /* zero means no -deadline tasks */
80 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0; 80 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
81 81
82 dl_rq->dl_nr_migratory = 0; 82 dl_rq->dl_nr_migratory = 0;
83 dl_rq->overloaded = 0; 83 dl_rq->overloaded = 0;
84 dl_rq->pushable_dl_tasks_root = RB_ROOT; 84 dl_rq->pushable_dl_tasks_root = RB_ROOT;
85 #else 85 #else
86 init_dl_bw(&dl_rq->dl_bw); 86 init_dl_bw(&dl_rq->dl_bw);
87 #endif 87 #endif
88 } 88 }
89 89
90 #ifdef CONFIG_SMP 90 #ifdef CONFIG_SMP
91 91
92 static inline int dl_overloaded(struct rq *rq) 92 static inline int dl_overloaded(struct rq *rq)
93 { 93 {
94 return atomic_read(&rq->rd->dlo_count); 94 return atomic_read(&rq->rd->dlo_count);
95 } 95 }
96 96
97 static inline void dl_set_overload(struct rq *rq) 97 static inline void dl_set_overload(struct rq *rq)
98 { 98 {
99 if (!rq->online) 99 if (!rq->online)
100 return; 100 return;
101 101
102 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask); 102 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
103 /* 103 /*
104 * Must be visible before the overload count is 104 * Must be visible before the overload count is
105 * set (as in sched_rt.c). 105 * set (as in sched_rt.c).
106 * 106 *
107 * Matched by the barrier in pull_dl_task(). 107 * Matched by the barrier in pull_dl_task().
108 */ 108 */
109 smp_wmb(); 109 smp_wmb();
110 atomic_inc(&rq->rd->dlo_count); 110 atomic_inc(&rq->rd->dlo_count);
111 } 111 }
112 112
113 static inline void dl_clear_overload(struct rq *rq) 113 static inline void dl_clear_overload(struct rq *rq)
114 { 114 {
115 if (!rq->online) 115 if (!rq->online)
116 return; 116 return;
117 117
118 atomic_dec(&rq->rd->dlo_count); 118 atomic_dec(&rq->rd->dlo_count);
119 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask); 119 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
120 } 120 }
121 121
122 static void update_dl_migration(struct dl_rq *dl_rq) 122 static void update_dl_migration(struct dl_rq *dl_rq)
123 { 123 {
124 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) { 124 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
125 if (!dl_rq->overloaded) { 125 if (!dl_rq->overloaded) {
126 dl_set_overload(rq_of_dl_rq(dl_rq)); 126 dl_set_overload(rq_of_dl_rq(dl_rq));
127 dl_rq->overloaded = 1; 127 dl_rq->overloaded = 1;
128 } 128 }
129 } else if (dl_rq->overloaded) { 129 } else if (dl_rq->overloaded) {
130 dl_clear_overload(rq_of_dl_rq(dl_rq)); 130 dl_clear_overload(rq_of_dl_rq(dl_rq));
131 dl_rq->overloaded = 0; 131 dl_rq->overloaded = 0;
132 } 132 }
133 } 133 }
134 134
135 static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) 135 static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
136 { 136 {
137 struct task_struct *p = dl_task_of(dl_se); 137 struct task_struct *p = dl_task_of(dl_se);
138 138
139 if (p->nr_cpus_allowed > 1) 139 if (p->nr_cpus_allowed > 1)
140 dl_rq->dl_nr_migratory++; 140 dl_rq->dl_nr_migratory++;
141 141
142 update_dl_migration(dl_rq); 142 update_dl_migration(dl_rq);
143 } 143 }
144 144
145 static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) 145 static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
146 { 146 {
147 struct task_struct *p = dl_task_of(dl_se); 147 struct task_struct *p = dl_task_of(dl_se);
148 148
149 if (p->nr_cpus_allowed > 1) 149 if (p->nr_cpus_allowed > 1)
150 dl_rq->dl_nr_migratory--; 150 dl_rq->dl_nr_migratory--;
151 151
152 update_dl_migration(dl_rq); 152 update_dl_migration(dl_rq);
153 } 153 }
154 154
155 /* 155 /*
156 * The list of pushable -deadline task is not a plist, like in 156 * The list of pushable -deadline task is not a plist, like in
157 * sched_rt.c, it is an rb-tree with tasks ordered by deadline. 157 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
158 */ 158 */
159 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) 159 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
160 { 160 {
161 struct dl_rq *dl_rq = &rq->dl; 161 struct dl_rq *dl_rq = &rq->dl;
162 struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_node; 162 struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_node;
163 struct rb_node *parent = NULL; 163 struct rb_node *parent = NULL;
164 struct task_struct *entry; 164 struct task_struct *entry;
165 int leftmost = 1; 165 int leftmost = 1;
166 166
167 BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks)); 167 BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
168 168
169 while (*link) { 169 while (*link) {
170 parent = *link; 170 parent = *link;
171 entry = rb_entry(parent, struct task_struct, 171 entry = rb_entry(parent, struct task_struct,
172 pushable_dl_tasks); 172 pushable_dl_tasks);
173 if (dl_entity_preempt(&p->dl, &entry->dl)) 173 if (dl_entity_preempt(&p->dl, &entry->dl))
174 link = &parent->rb_left; 174 link = &parent->rb_left;
175 else { 175 else {
176 link = &parent->rb_right; 176 link = &parent->rb_right;
177 leftmost = 0; 177 leftmost = 0;
178 } 178 }
179 } 179 }
180 180
181 if (leftmost) 181 if (leftmost)
182 dl_rq->pushable_dl_tasks_leftmost = &p->pushable_dl_tasks; 182 dl_rq->pushable_dl_tasks_leftmost = &p->pushable_dl_tasks;
183 183
184 rb_link_node(&p->pushable_dl_tasks, parent, link); 184 rb_link_node(&p->pushable_dl_tasks, parent, link);
185 rb_insert_color(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root); 185 rb_insert_color(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
186 } 186 }
187 187
188 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) 188 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
189 { 189 {
190 struct dl_rq *dl_rq = &rq->dl; 190 struct dl_rq *dl_rq = &rq->dl;
191 191
192 if (RB_EMPTY_NODE(&p->pushable_dl_tasks)) 192 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
193 return; 193 return;
194 194
195 if (dl_rq->pushable_dl_tasks_leftmost == &p->pushable_dl_tasks) { 195 if (dl_rq->pushable_dl_tasks_leftmost == &p->pushable_dl_tasks) {
196 struct rb_node *next_node; 196 struct rb_node *next_node;
197 197
198 next_node = rb_next(&p->pushable_dl_tasks); 198 next_node = rb_next(&p->pushable_dl_tasks);
199 dl_rq->pushable_dl_tasks_leftmost = next_node; 199 dl_rq->pushable_dl_tasks_leftmost = next_node;
200 } 200 }
201 201
202 rb_erase(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root); 202 rb_erase(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
203 RB_CLEAR_NODE(&p->pushable_dl_tasks); 203 RB_CLEAR_NODE(&p->pushable_dl_tasks);
204 } 204 }
205 205
206 static inline int has_pushable_dl_tasks(struct rq *rq) 206 static inline int has_pushable_dl_tasks(struct rq *rq)
207 { 207 {
208 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root); 208 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root);
209 } 209 }
210 210
211 static int push_dl_task(struct rq *rq); 211 static int push_dl_task(struct rq *rq);
212 212
213 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev) 213 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
214 { 214 {
215 return dl_task(prev); 215 return dl_task(prev);
216 } 216 }
217 217
218 static inline void set_post_schedule(struct rq *rq) 218 static inline void set_post_schedule(struct rq *rq)
219 { 219 {
220 rq->post_schedule = has_pushable_dl_tasks(rq); 220 rq->post_schedule = has_pushable_dl_tasks(rq);
221 } 221 }
222 222
223 #else 223 #else
224 224
225 static inline 225 static inline
226 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p) 226 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
227 { 227 {
228 } 228 }
229 229
230 static inline 230 static inline
231 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p) 231 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
232 { 232 {
233 } 233 }
234 234
235 static inline 235 static inline
236 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) 236 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
237 { 237 {
238 } 238 }
239 239
240 static inline 240 static inline
241 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) 241 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
242 { 242 {
243 } 243 }
244 244
245 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev) 245 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
246 { 246 {
247 return false; 247 return false;
248 } 248 }
249 249
250 static inline int pull_dl_task(struct rq *rq) 250 static inline int pull_dl_task(struct rq *rq)
251 { 251 {
252 return 0; 252 return 0;
253 } 253 }
254 254
255 static inline void set_post_schedule(struct rq *rq) 255 static inline void set_post_schedule(struct rq *rq)
256 { 256 {
257 } 257 }
258 #endif /* CONFIG_SMP */ 258 #endif /* CONFIG_SMP */
259 259
260 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags); 260 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
261 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags); 261 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
262 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, 262 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
263 int flags); 263 int flags);
264 264
265 /* 265 /*
266 * We are being explicitly informed that a new instance is starting, 266 * We are being explicitly informed that a new instance is starting,
267 * and this means that: 267 * and this means that:
268 * - the absolute deadline of the entity has to be placed at 268 * - the absolute deadline of the entity has to be placed at
269 * current time + relative deadline; 269 * current time + relative deadline;
270 * - the runtime of the entity has to be set to the maximum value. 270 * - the runtime of the entity has to be set to the maximum value.
271 * 271 *
272 * The capability of specifying such event is useful whenever a -deadline 272 * The capability of specifying such event is useful whenever a -deadline
273 * entity wants to (try to!) synchronize its behaviour with the scheduler's 273 * entity wants to (try to!) synchronize its behaviour with the scheduler's
274 * one, and to (try to!) reconcile itself with its own scheduling 274 * one, and to (try to!) reconcile itself with its own scheduling
275 * parameters. 275 * parameters.
276 */ 276 */
277 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se, 277 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se,
278 struct sched_dl_entity *pi_se) 278 struct sched_dl_entity *pi_se)
279 { 279 {
280 struct dl_rq *dl_rq = dl_rq_of_se(dl_se); 280 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
281 struct rq *rq = rq_of_dl_rq(dl_rq); 281 struct rq *rq = rq_of_dl_rq(dl_rq);
282 282
283 WARN_ON(!dl_se->dl_new || dl_se->dl_throttled); 283 WARN_ON(!dl_se->dl_new || dl_se->dl_throttled);
284 284
285 /* 285 /*
286 * We use the regular wall clock time to set deadlines in the 286 * We use the regular wall clock time to set deadlines in the
287 * future; in fact, we must consider execution overheads (time 287 * future; in fact, we must consider execution overheads (time
288 * spent on hardirq context, etc.). 288 * spent on hardirq context, etc.).
289 */ 289 */
290 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; 290 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
291 dl_se->runtime = pi_se->dl_runtime; 291 dl_se->runtime = pi_se->dl_runtime;
292 dl_se->dl_new = 0; 292 dl_se->dl_new = 0;
293 } 293 }
294 294
295 /* 295 /*
296 * Pure Earliest Deadline First (EDF) scheduling does not deal with the 296 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
297 * possibility of a entity lasting more than what it declared, and thus 297 * possibility of a entity lasting more than what it declared, and thus
298 * exhausting its runtime. 298 * exhausting its runtime.
299 * 299 *
300 * Here we are interested in making runtime overrun possible, but we do 300 * Here we are interested in making runtime overrun possible, but we do
301 * not want a entity which is misbehaving to affect the scheduling of all 301 * not want a entity which is misbehaving to affect the scheduling of all
302 * other entities. 302 * other entities.
303 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS) 303 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
304 * is used, in order to confine each entity within its own bandwidth. 304 * is used, in order to confine each entity within its own bandwidth.
305 * 305 *
306 * This function deals exactly with that, and ensures that when the runtime 306 * This function deals exactly with that, and ensures that when the runtime
307 * of a entity is replenished, its deadline is also postponed. That ensures 307 * of a entity is replenished, its deadline is also postponed. That ensures
308 * the overrunning entity can't interfere with other entity in the system and 308 * the overrunning entity can't interfere with other entity in the system and
309 * can't make them miss their deadlines. Reasons why this kind of overruns 309 * can't make them miss their deadlines. Reasons why this kind of overruns
310 * could happen are, typically, a entity voluntarily trying to overcome its 310 * could happen are, typically, a entity voluntarily trying to overcome its
311 * runtime, or it just underestimated it during sched_setscheduler_ex(). 311 * runtime, or it just underestimated it during sched_setscheduler_ex().
312 */ 312 */
313 static void replenish_dl_entity(struct sched_dl_entity *dl_se, 313 static void replenish_dl_entity(struct sched_dl_entity *dl_se,
314 struct sched_dl_entity *pi_se) 314 struct sched_dl_entity *pi_se)
315 { 315 {
316 struct dl_rq *dl_rq = dl_rq_of_se(dl_se); 316 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
317 struct rq *rq = rq_of_dl_rq(dl_rq); 317 struct rq *rq = rq_of_dl_rq(dl_rq);
318 318
319 BUG_ON(pi_se->dl_runtime <= 0); 319 BUG_ON(pi_se->dl_runtime <= 0);
320 320
321 /* 321 /*
322 * This could be the case for a !-dl task that is boosted. 322 * This could be the case for a !-dl task that is boosted.
323 * Just go with full inherited parameters. 323 * Just go with full inherited parameters.
324 */ 324 */
325 if (dl_se->dl_deadline == 0) { 325 if (dl_se->dl_deadline == 0) {
326 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; 326 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
327 dl_se->runtime = pi_se->dl_runtime; 327 dl_se->runtime = pi_se->dl_runtime;
328 } 328 }
329 329
330 /* 330 /*
331 * We keep moving the deadline away until we get some 331 * We keep moving the deadline away until we get some
332 * available runtime for the entity. This ensures correct 332 * available runtime for the entity. This ensures correct
333 * handling of situations where the runtime overrun is 333 * handling of situations where the runtime overrun is
334 * arbitrary large. 334 * arbitrary large.
335 */ 335 */
336 while (dl_se->runtime <= 0) { 336 while (dl_se->runtime <= 0) {
337 dl_se->deadline += pi_se->dl_period; 337 dl_se->deadline += pi_se->dl_period;
338 dl_se->runtime += pi_se->dl_runtime; 338 dl_se->runtime += pi_se->dl_runtime;
339 } 339 }
340 340
341 /* 341 /*
342 * At this point, the deadline really should be "in 342 * At this point, the deadline really should be "in
343 * the future" with respect to rq->clock. If it's 343 * the future" with respect to rq->clock. If it's
344 * not, we are, for some reason, lagging too much! 344 * not, we are, for some reason, lagging too much!
345 * Anyway, after having warn userspace abut that, 345 * Anyway, after having warn userspace abut that,
346 * we still try to keep the things running by 346 * we still try to keep the things running by
347 * resetting the deadline and the budget of the 347 * resetting the deadline and the budget of the
348 * entity. 348 * entity.
349 */ 349 */
350 if (dl_time_before(dl_se->deadline, rq_clock(rq))) { 350 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
351 static bool lag_once = false; 351 static bool lag_once = false;
352 352
353 if (!lag_once) { 353 if (!lag_once) {
354 lag_once = true; 354 lag_once = true;
355 printk_sched("sched: DL replenish lagged to much\n"); 355 printk_sched("sched: DL replenish lagged to much\n");
356 } 356 }
357 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; 357 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
358 dl_se->runtime = pi_se->dl_runtime; 358 dl_se->runtime = pi_se->dl_runtime;
359 } 359 }
360 } 360 }
361 361
362 /* 362 /*
363 * Here we check if --at time t-- an entity (which is probably being 363 * Here we check if --at time t-- an entity (which is probably being
364 * [re]activated or, in general, enqueued) can use its remaining runtime 364 * [re]activated or, in general, enqueued) can use its remaining runtime
365 * and its current deadline _without_ exceeding the bandwidth it is 365 * and its current deadline _without_ exceeding the bandwidth it is
366 * assigned (function returns true if it can't). We are in fact applying 366 * assigned (function returns true if it can't). We are in fact applying
367 * one of the CBS rules: when a task wakes up, if the residual runtime 367 * one of the CBS rules: when a task wakes up, if the residual runtime
368 * over residual deadline fits within the allocated bandwidth, then we 368 * over residual deadline fits within the allocated bandwidth, then we
369 * can keep the current (absolute) deadline and residual budget without 369 * can keep the current (absolute) deadline and residual budget without
370 * disrupting the schedulability of the system. Otherwise, we should 370 * disrupting the schedulability of the system. Otherwise, we should
371 * refill the runtime and set the deadline a period in the future, 371 * refill the runtime and set the deadline a period in the future,
372 * because keeping the current (absolute) deadline of the task would 372 * because keeping the current (absolute) deadline of the task would
373 * result in breaking guarantees promised to other tasks (refer to 373 * result in breaking guarantees promised to other tasks (refer to
374 * Documentation/scheduler/sched-deadline.txt for more informations). 374 * Documentation/scheduler/sched-deadline.txt for more informations).
375 * 375 *
376 * This function returns true if: 376 * This function returns true if:
377 * 377 *
378 * runtime / (deadline - t) > dl_runtime / dl_period , 378 * runtime / (deadline - t) > dl_runtime / dl_period ,
379 * 379 *
380 * IOW we can't recycle current parameters. 380 * IOW we can't recycle current parameters.
381 * 381 *
382 * Notice that the bandwidth check is done against the period. For 382 * Notice that the bandwidth check is done against the period. For
383 * task with deadline equal to period this is the same of using 383 * task with deadline equal to period this is the same of using
384 * dl_deadline instead of dl_period in the equation above. 384 * dl_deadline instead of dl_period in the equation above.
385 */ 385 */
386 static bool dl_entity_overflow(struct sched_dl_entity *dl_se, 386 static bool dl_entity_overflow(struct sched_dl_entity *dl_se,
387 struct sched_dl_entity *pi_se, u64 t) 387 struct sched_dl_entity *pi_se, u64 t)
388 { 388 {
389 u64 left, right; 389 u64 left, right;
390 390
391 /* 391 /*
392 * left and right are the two sides of the equation above, 392 * left and right are the two sides of the equation above,
393 * after a bit of shuffling to use multiplications instead 393 * after a bit of shuffling to use multiplications instead
394 * of divisions. 394 * of divisions.
395 * 395 *
396 * Note that none of the time values involved in the two 396 * Note that none of the time values involved in the two
397 * multiplications are absolute: dl_deadline and dl_runtime 397 * multiplications are absolute: dl_deadline and dl_runtime
398 * are the relative deadline and the maximum runtime of each 398 * are the relative deadline and the maximum runtime of each
399 * instance, runtime is the runtime left for the last instance 399 * instance, runtime is the runtime left for the last instance
400 * and (deadline - t), since t is rq->clock, is the time left 400 * and (deadline - t), since t is rq->clock, is the time left
401 * to the (absolute) deadline. Even if overflowing the u64 type 401 * to the (absolute) deadline. Even if overflowing the u64 type
402 * is very unlikely to occur in both cases, here we scale down 402 * is very unlikely to occur in both cases, here we scale down
403 * as we want to avoid that risk at all. Scaling down by 10 403 * as we want to avoid that risk at all. Scaling down by 10
404 * means that we reduce granularity to 1us. We are fine with it, 404 * means that we reduce granularity to 1us. We are fine with it,
405 * since this is only a true/false check and, anyway, thinking 405 * since this is only a true/false check and, anyway, thinking
406 * of anything below microseconds resolution is actually fiction 406 * of anything below microseconds resolution is actually fiction
407 * (but still we want to give the user that illusion >;). 407 * (but still we want to give the user that illusion >;).
408 */ 408 */
409 left = (pi_se->dl_period >> DL_SCALE) * (dl_se->runtime >> DL_SCALE); 409 left = (pi_se->dl_period >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
410 right = ((dl_se->deadline - t) >> DL_SCALE) * 410 right = ((dl_se->deadline - t) >> DL_SCALE) *
411 (pi_se->dl_runtime >> DL_SCALE); 411 (pi_se->dl_runtime >> DL_SCALE);
412 412
413 return dl_time_before(right, left); 413 return dl_time_before(right, left);
414 } 414 }
415 415
416 /* 416 /*
417 * When a -deadline entity is queued back on the runqueue, its runtime and 417 * When a -deadline entity is queued back on the runqueue, its runtime and
418 * deadline might need updating. 418 * deadline might need updating.
419 * 419 *
420 * The policy here is that we update the deadline of the entity only if: 420 * The policy here is that we update the deadline of the entity only if:
421 * - the current deadline is in the past, 421 * - the current deadline is in the past,
422 * - using the remaining runtime with the current deadline would make 422 * - using the remaining runtime with the current deadline would make
423 * the entity exceed its bandwidth. 423 * the entity exceed its bandwidth.
424 */ 424 */
425 static void update_dl_entity(struct sched_dl_entity *dl_se, 425 static void update_dl_entity(struct sched_dl_entity *dl_se,
426 struct sched_dl_entity *pi_se) 426 struct sched_dl_entity *pi_se)
427 { 427 {
428 struct dl_rq *dl_rq = dl_rq_of_se(dl_se); 428 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
429 struct rq *rq = rq_of_dl_rq(dl_rq); 429 struct rq *rq = rq_of_dl_rq(dl_rq);
430 430
431 /* 431 /*
432 * The arrival of a new instance needs special treatment, i.e., 432 * The arrival of a new instance needs special treatment, i.e.,
433 * the actual scheduling parameters have to be "renewed". 433 * the actual scheduling parameters have to be "renewed".
434 */ 434 */
435 if (dl_se->dl_new) { 435 if (dl_se->dl_new) {
436 setup_new_dl_entity(dl_se, pi_se); 436 setup_new_dl_entity(dl_se, pi_se);
437 return; 437 return;
438 } 438 }
439 439
440 if (dl_time_before(dl_se->deadline, rq_clock(rq)) || 440 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
441 dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) { 441 dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) {
442 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline; 442 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
443 dl_se->runtime = pi_se->dl_runtime; 443 dl_se->runtime = pi_se->dl_runtime;
444 } 444 }
445 } 445 }
446 446
447 /* 447 /*
448 * If the entity depleted all its runtime, and if we want it to sleep 448 * If the entity depleted all its runtime, and if we want it to sleep
449 * while waiting for some new execution time to become available, we 449 * while waiting for some new execution time to become available, we
450 * set the bandwidth enforcement timer to the replenishment instant 450 * set the bandwidth enforcement timer to the replenishment instant
451 * and try to activate it. 451 * and try to activate it.
452 * 452 *
453 * Notice that it is important for the caller to know if the timer 453 * Notice that it is important for the caller to know if the timer
454 * actually started or not (i.e., the replenishment instant is in 454 * actually started or not (i.e., the replenishment instant is in
455 * the future or in the past). 455 * the future or in the past).
456 */ 456 */
457 static int start_dl_timer(struct sched_dl_entity *dl_se, bool boosted) 457 static int start_dl_timer(struct sched_dl_entity *dl_se, bool boosted)
458 { 458 {
459 struct dl_rq *dl_rq = dl_rq_of_se(dl_se); 459 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
460 struct rq *rq = rq_of_dl_rq(dl_rq); 460 struct rq *rq = rq_of_dl_rq(dl_rq);
461 ktime_t now, act; 461 ktime_t now, act;
462 ktime_t soft, hard; 462 ktime_t soft, hard;
463 unsigned long range; 463 unsigned long range;
464 s64 delta; 464 s64 delta;
465 465
466 if (boosted) 466 if (boosted)
467 return 0; 467 return 0;
468 /* 468 /*
469 * We want the timer to fire at the deadline, but considering 469 * We want the timer to fire at the deadline, but considering
470 * that it is actually coming from rq->clock and not from 470 * that it is actually coming from rq->clock and not from
471 * hrtimer's time base reading. 471 * hrtimer's time base reading.
472 */ 472 */
473 act = ns_to_ktime(dl_se->deadline); 473 act = ns_to_ktime(dl_se->deadline);
474 now = hrtimer_cb_get_time(&dl_se->dl_timer); 474 now = hrtimer_cb_get_time(&dl_se->dl_timer);
475 delta = ktime_to_ns(now) - rq_clock(rq); 475 delta = ktime_to_ns(now) - rq_clock(rq);
476 act = ktime_add_ns(act, delta); 476 act = ktime_add_ns(act, delta);
477 477
478 /* 478 /*
479 * If the expiry time already passed, e.g., because the value 479 * If the expiry time already passed, e.g., because the value
480 * chosen as the deadline is too small, don't even try to 480 * chosen as the deadline is too small, don't even try to
481 * start the timer in the past! 481 * start the timer in the past!
482 */ 482 */
483 if (ktime_us_delta(act, now) < 0) 483 if (ktime_us_delta(act, now) < 0)
484 return 0; 484 return 0;
485 485
486 hrtimer_set_expires(&dl_se->dl_timer, act); 486 hrtimer_set_expires(&dl_se->dl_timer, act);
487 487
488 soft = hrtimer_get_softexpires(&dl_se->dl_timer); 488 soft = hrtimer_get_softexpires(&dl_se->dl_timer);
489 hard = hrtimer_get_expires(&dl_se->dl_timer); 489 hard = hrtimer_get_expires(&dl_se->dl_timer);
490 range = ktime_to_ns(ktime_sub(hard, soft)); 490 range = ktime_to_ns(ktime_sub(hard, soft));
491 __hrtimer_start_range_ns(&dl_se->dl_timer, soft, 491 __hrtimer_start_range_ns(&dl_se->dl_timer, soft,
492 range, HRTIMER_MODE_ABS, 0); 492 range, HRTIMER_MODE_ABS, 0);
493 493
494 return hrtimer_active(&dl_se->dl_timer); 494 return hrtimer_active(&dl_se->dl_timer);
495 } 495 }
496 496
497 /* 497 /*
498 * This is the bandwidth enforcement timer callback. If here, we know 498 * This is the bandwidth enforcement timer callback. If here, we know
499 * a task is not on its dl_rq, since the fact that the timer was running 499 * a task is not on its dl_rq, since the fact that the timer was running
500 * means the task is throttled and needs a runtime replenishment. 500 * means the task is throttled and needs a runtime replenishment.
501 * 501 *
502 * However, what we actually do depends on the fact the task is active, 502 * However, what we actually do depends on the fact the task is active,
503 * (it is on its rq) or has been removed from there by a call to 503 * (it is on its rq) or has been removed from there by a call to
504 * dequeue_task_dl(). In the former case we must issue the runtime 504 * dequeue_task_dl(). In the former case we must issue the runtime
505 * replenishment and add the task back to the dl_rq; in the latter, we just 505 * replenishment and add the task back to the dl_rq; in the latter, we just
506 * do nothing but clearing dl_throttled, so that runtime and deadline 506 * do nothing but clearing dl_throttled, so that runtime and deadline
507 * updating (and the queueing back to dl_rq) will be done by the 507 * updating (and the queueing back to dl_rq) will be done by the
508 * next call to enqueue_task_dl(). 508 * next call to enqueue_task_dl().
509 */ 509 */
510 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer) 510 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
511 { 511 {
512 struct sched_dl_entity *dl_se = container_of(timer, 512 struct sched_dl_entity *dl_se = container_of(timer,
513 struct sched_dl_entity, 513 struct sched_dl_entity,
514 dl_timer); 514 dl_timer);
515 struct task_struct *p = dl_task_of(dl_se); 515 struct task_struct *p = dl_task_of(dl_se);
516 struct rq *rq = task_rq(p); 516 struct rq *rq;
517 again:
518 rq = task_rq(p);
517 raw_spin_lock(&rq->lock); 519 raw_spin_lock(&rq->lock);
520
521 if (rq != task_rq(p)) {
522 /* Task was moved, retrying. */
523 raw_spin_unlock(&rq->lock);
524 goto again;
525 }
518 526
519 /* 527 /*
520 * We need to take care of a possible races here. In fact, the 528 * We need to take care of a possible races here. In fact, the
521 * task might have changed its scheduling policy to something 529 * task might have changed its scheduling policy to something
522 * different from SCHED_DEADLINE or changed its reservation 530 * different from SCHED_DEADLINE or changed its reservation
523 * parameters (through sched_setscheduler()). 531 * parameters (through sched_setscheduler()).
524 */ 532 */
525 if (!dl_task(p) || dl_se->dl_new) 533 if (!dl_task(p) || dl_se->dl_new)
526 goto unlock; 534 goto unlock;
527 535
528 sched_clock_tick(); 536 sched_clock_tick();
529 update_rq_clock(rq); 537 update_rq_clock(rq);
530 dl_se->dl_throttled = 0; 538 dl_se->dl_throttled = 0;
531 dl_se->dl_yielded = 0; 539 dl_se->dl_yielded = 0;
532 if (p->on_rq) { 540 if (p->on_rq) {
533 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH); 541 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
534 if (task_has_dl_policy(rq->curr)) 542 if (task_has_dl_policy(rq->curr))
535 check_preempt_curr_dl(rq, p, 0); 543 check_preempt_curr_dl(rq, p, 0);
536 else 544 else
537 resched_task(rq->curr); 545 resched_task(rq->curr);
538 #ifdef CONFIG_SMP 546 #ifdef CONFIG_SMP
539 /* 547 /*
540 * Queueing this task back might have overloaded rq, 548 * Queueing this task back might have overloaded rq,
541 * check if we need to kick someone away. 549 * check if we need to kick someone away.
542 */ 550 */
543 if (has_pushable_dl_tasks(rq)) 551 if (has_pushable_dl_tasks(rq))
544 push_dl_task(rq); 552 push_dl_task(rq);
545 #endif 553 #endif
546 } 554 }
547 unlock: 555 unlock:
548 raw_spin_unlock(&rq->lock); 556 raw_spin_unlock(&rq->lock);
549 557
550 return HRTIMER_NORESTART; 558 return HRTIMER_NORESTART;
551 } 559 }
552 560
553 void init_dl_task_timer(struct sched_dl_entity *dl_se) 561 void init_dl_task_timer(struct sched_dl_entity *dl_se)
554 { 562 {
555 struct hrtimer *timer = &dl_se->dl_timer; 563 struct hrtimer *timer = &dl_se->dl_timer;
556 564
557 if (hrtimer_active(timer)) { 565 if (hrtimer_active(timer)) {
558 hrtimer_try_to_cancel(timer); 566 hrtimer_try_to_cancel(timer);
559 return; 567 return;
560 } 568 }
561 569
562 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 570 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
563 timer->function = dl_task_timer; 571 timer->function = dl_task_timer;
564 } 572 }
565 573
566 static 574 static
567 int dl_runtime_exceeded(struct rq *rq, struct sched_dl_entity *dl_se) 575 int dl_runtime_exceeded(struct rq *rq, struct sched_dl_entity *dl_se)
568 { 576 {
569 int dmiss = dl_time_before(dl_se->deadline, rq_clock(rq)); 577 int dmiss = dl_time_before(dl_se->deadline, rq_clock(rq));
570 int rorun = dl_se->runtime <= 0; 578 int rorun = dl_se->runtime <= 0;
571 579
572 if (!rorun && !dmiss) 580 if (!rorun && !dmiss)
573 return 0; 581 return 0;
574 582
575 /* 583 /*
576 * If we are beyond our current deadline and we are still 584 * If we are beyond our current deadline and we are still
577 * executing, then we have already used some of the runtime of 585 * executing, then we have already used some of the runtime of
578 * the next instance. Thus, if we do not account that, we are 586 * the next instance. Thus, if we do not account that, we are
579 * stealing bandwidth from the system at each deadline miss! 587 * stealing bandwidth from the system at each deadline miss!
580 */ 588 */
581 if (dmiss) { 589 if (dmiss) {
582 dl_se->runtime = rorun ? dl_se->runtime : 0; 590 dl_se->runtime = rorun ? dl_se->runtime : 0;
583 dl_se->runtime -= rq_clock(rq) - dl_se->deadline; 591 dl_se->runtime -= rq_clock(rq) - dl_se->deadline;
584 } 592 }
585 593
586 return 1; 594 return 1;
587 } 595 }
588 596
589 extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq); 597 extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
590 598
591 /* 599 /*
592 * Update the current task's runtime statistics (provided it is still 600 * Update the current task's runtime statistics (provided it is still
593 * a -deadline task and has not been removed from the dl_rq). 601 * a -deadline task and has not been removed from the dl_rq).
594 */ 602 */
595 static void update_curr_dl(struct rq *rq) 603 static void update_curr_dl(struct rq *rq)
596 { 604 {
597 struct task_struct *curr = rq->curr; 605 struct task_struct *curr = rq->curr;
598 struct sched_dl_entity *dl_se = &curr->dl; 606 struct sched_dl_entity *dl_se = &curr->dl;
599 u64 delta_exec; 607 u64 delta_exec;
600 608
601 if (!dl_task(curr) || !on_dl_rq(dl_se)) 609 if (!dl_task(curr) || !on_dl_rq(dl_se))
602 return; 610 return;
603 611
604 /* 612 /*
605 * Consumed budget is computed considering the time as 613 * Consumed budget is computed considering the time as
606 * observed by schedulable tasks (excluding time spent 614 * observed by schedulable tasks (excluding time spent
607 * in hardirq context, etc.). Deadlines are instead 615 * in hardirq context, etc.). Deadlines are instead
608 * computed using hard walltime. This seems to be the more 616 * computed using hard walltime. This seems to be the more
609 * natural solution, but the full ramifications of this 617 * natural solution, but the full ramifications of this
610 * approach need further study. 618 * approach need further study.
611 */ 619 */
612 delta_exec = rq_clock_task(rq) - curr->se.exec_start; 620 delta_exec = rq_clock_task(rq) - curr->se.exec_start;
613 if (unlikely((s64)delta_exec <= 0)) 621 if (unlikely((s64)delta_exec <= 0))
614 return; 622 return;
615 623
616 schedstat_set(curr->se.statistics.exec_max, 624 schedstat_set(curr->se.statistics.exec_max,
617 max(curr->se.statistics.exec_max, delta_exec)); 625 max(curr->se.statistics.exec_max, delta_exec));
618 626
619 curr->se.sum_exec_runtime += delta_exec; 627 curr->se.sum_exec_runtime += delta_exec;
620 account_group_exec_runtime(curr, delta_exec); 628 account_group_exec_runtime(curr, delta_exec);
621 629
622 curr->se.exec_start = rq_clock_task(rq); 630 curr->se.exec_start = rq_clock_task(rq);
623 cpuacct_charge(curr, delta_exec); 631 cpuacct_charge(curr, delta_exec);
624 632
625 sched_rt_avg_update(rq, delta_exec); 633 sched_rt_avg_update(rq, delta_exec);
626 634
627 dl_se->runtime -= delta_exec; 635 dl_se->runtime -= delta_exec;
628 if (dl_runtime_exceeded(rq, dl_se)) { 636 if (dl_runtime_exceeded(rq, dl_se)) {
629 __dequeue_task_dl(rq, curr, 0); 637 __dequeue_task_dl(rq, curr, 0);
630 if (likely(start_dl_timer(dl_se, curr->dl.dl_boosted))) 638 if (likely(start_dl_timer(dl_se, curr->dl.dl_boosted)))
631 dl_se->dl_throttled = 1; 639 dl_se->dl_throttled = 1;
632 else 640 else
633 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH); 641 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
634 642
635 if (!is_leftmost(curr, &rq->dl)) 643 if (!is_leftmost(curr, &rq->dl))
636 resched_task(curr); 644 resched_task(curr);
637 } 645 }
638 646
639 /* 647 /*
640 * Because -- for now -- we share the rt bandwidth, we need to 648 * Because -- for now -- we share the rt bandwidth, we need to
641 * account our runtime there too, otherwise actual rt tasks 649 * account our runtime there too, otherwise actual rt tasks
642 * would be able to exceed the shared quota. 650 * would be able to exceed the shared quota.
643 * 651 *
644 * Account to the root rt group for now. 652 * Account to the root rt group for now.
645 * 653 *
646 * The solution we're working towards is having the RT groups scheduled 654 * The solution we're working towards is having the RT groups scheduled
647 * using deadline servers -- however there's a few nasties to figure 655 * using deadline servers -- however there's a few nasties to figure
648 * out before that can happen. 656 * out before that can happen.
649 */ 657 */
650 if (rt_bandwidth_enabled()) { 658 if (rt_bandwidth_enabled()) {
651 struct rt_rq *rt_rq = &rq->rt; 659 struct rt_rq *rt_rq = &rq->rt;
652 660
653 raw_spin_lock(&rt_rq->rt_runtime_lock); 661 raw_spin_lock(&rt_rq->rt_runtime_lock);
654 /* 662 /*
655 * We'll let actual RT tasks worry about the overflow here, we 663 * We'll let actual RT tasks worry about the overflow here, we
656 * have our own CBS to keep us inline; only account when RT 664 * have our own CBS to keep us inline; only account when RT
657 * bandwidth is relevant. 665 * bandwidth is relevant.
658 */ 666 */
659 if (sched_rt_bandwidth_account(rt_rq)) 667 if (sched_rt_bandwidth_account(rt_rq))
660 rt_rq->rt_time += delta_exec; 668 rt_rq->rt_time += delta_exec;
661 raw_spin_unlock(&rt_rq->rt_runtime_lock); 669 raw_spin_unlock(&rt_rq->rt_runtime_lock);
662 } 670 }
663 } 671 }
664 672
665 #ifdef CONFIG_SMP 673 #ifdef CONFIG_SMP
666 674
667 static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu); 675 static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu);
668 676
669 static inline u64 next_deadline(struct rq *rq) 677 static inline u64 next_deadline(struct rq *rq)
670 { 678 {
671 struct task_struct *next = pick_next_earliest_dl_task(rq, rq->cpu); 679 struct task_struct *next = pick_next_earliest_dl_task(rq, rq->cpu);
672 680
673 if (next && dl_prio(next->prio)) 681 if (next && dl_prio(next->prio))
674 return next->dl.deadline; 682 return next->dl.deadline;
675 else 683 else
676 return 0; 684 return 0;
677 } 685 }
678 686
679 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) 687 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
680 { 688 {
681 struct rq *rq = rq_of_dl_rq(dl_rq); 689 struct rq *rq = rq_of_dl_rq(dl_rq);
682 690
683 if (dl_rq->earliest_dl.curr == 0 || 691 if (dl_rq->earliest_dl.curr == 0 ||
684 dl_time_before(deadline, dl_rq->earliest_dl.curr)) { 692 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
685 /* 693 /*
686 * If the dl_rq had no -deadline tasks, or if the new task 694 * If the dl_rq had no -deadline tasks, or if the new task
687 * has shorter deadline than the current one on dl_rq, we 695 * has shorter deadline than the current one on dl_rq, we
688 * know that the previous earliest becomes our next earliest, 696 * know that the previous earliest becomes our next earliest,
689 * as the new task becomes the earliest itself. 697 * as the new task becomes the earliest itself.
690 */ 698 */
691 dl_rq->earliest_dl.next = dl_rq->earliest_dl.curr; 699 dl_rq->earliest_dl.next = dl_rq->earliest_dl.curr;
692 dl_rq->earliest_dl.curr = deadline; 700 dl_rq->earliest_dl.curr = deadline;
693 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline, 1); 701 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline, 1);
694 } else if (dl_rq->earliest_dl.next == 0 || 702 } else if (dl_rq->earliest_dl.next == 0 ||
695 dl_time_before(deadline, dl_rq->earliest_dl.next)) { 703 dl_time_before(deadline, dl_rq->earliest_dl.next)) {
696 /* 704 /*
697 * On the other hand, if the new -deadline task has a 705 * On the other hand, if the new -deadline task has a
698 * a later deadline than the earliest one on dl_rq, but 706 * a later deadline than the earliest one on dl_rq, but
699 * it is earlier than the next (if any), we must 707 * it is earlier than the next (if any), we must
700 * recompute the next-earliest. 708 * recompute the next-earliest.
701 */ 709 */
702 dl_rq->earliest_dl.next = next_deadline(rq); 710 dl_rq->earliest_dl.next = next_deadline(rq);
703 } 711 }
704 } 712 }
705 713
706 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) 714 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
707 { 715 {
708 struct rq *rq = rq_of_dl_rq(dl_rq); 716 struct rq *rq = rq_of_dl_rq(dl_rq);
709 717
710 /* 718 /*
711 * Since we may have removed our earliest (and/or next earliest) 719 * Since we may have removed our earliest (and/or next earliest)
712 * task we must recompute them. 720 * task we must recompute them.
713 */ 721 */
714 if (!dl_rq->dl_nr_running) { 722 if (!dl_rq->dl_nr_running) {
715 dl_rq->earliest_dl.curr = 0; 723 dl_rq->earliest_dl.curr = 0;
716 dl_rq->earliest_dl.next = 0; 724 dl_rq->earliest_dl.next = 0;
717 cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0); 725 cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0);
718 } else { 726 } else {
719 struct rb_node *leftmost = dl_rq->rb_leftmost; 727 struct rb_node *leftmost = dl_rq->rb_leftmost;
720 struct sched_dl_entity *entry; 728 struct sched_dl_entity *entry;
721 729
722 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node); 730 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
723 dl_rq->earliest_dl.curr = entry->deadline; 731 dl_rq->earliest_dl.curr = entry->deadline;
724 dl_rq->earliest_dl.next = next_deadline(rq); 732 dl_rq->earliest_dl.next = next_deadline(rq);
725 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline, 1); 733 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline, 1);
726 } 734 }
727 } 735 }
728 736
729 #else 737 #else
730 738
731 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} 739 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
732 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {} 740 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
733 741
734 #endif /* CONFIG_SMP */ 742 #endif /* CONFIG_SMP */
735 743
736 static inline 744 static inline
737 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) 745 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
738 { 746 {
739 int prio = dl_task_of(dl_se)->prio; 747 int prio = dl_task_of(dl_se)->prio;
740 u64 deadline = dl_se->deadline; 748 u64 deadline = dl_se->deadline;
741 749
742 WARN_ON(!dl_prio(prio)); 750 WARN_ON(!dl_prio(prio));
743 dl_rq->dl_nr_running++; 751 dl_rq->dl_nr_running++;
744 inc_nr_running(rq_of_dl_rq(dl_rq)); 752 inc_nr_running(rq_of_dl_rq(dl_rq));
745 753
746 inc_dl_deadline(dl_rq, deadline); 754 inc_dl_deadline(dl_rq, deadline);
747 inc_dl_migration(dl_se, dl_rq); 755 inc_dl_migration(dl_se, dl_rq);
748 } 756 }
749 757
750 static inline 758 static inline
751 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq) 759 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
752 { 760 {
753 int prio = dl_task_of(dl_se)->prio; 761 int prio = dl_task_of(dl_se)->prio;
754 762
755 WARN_ON(!dl_prio(prio)); 763 WARN_ON(!dl_prio(prio));
756 WARN_ON(!dl_rq->dl_nr_running); 764 WARN_ON(!dl_rq->dl_nr_running);
757 dl_rq->dl_nr_running--; 765 dl_rq->dl_nr_running--;
758 dec_nr_running(rq_of_dl_rq(dl_rq)); 766 dec_nr_running(rq_of_dl_rq(dl_rq));
759 767
760 dec_dl_deadline(dl_rq, dl_se->deadline); 768 dec_dl_deadline(dl_rq, dl_se->deadline);
761 dec_dl_migration(dl_se, dl_rq); 769 dec_dl_migration(dl_se, dl_rq);
762 } 770 }
763 771
764 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se) 772 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
765 { 773 {
766 struct dl_rq *dl_rq = dl_rq_of_se(dl_se); 774 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
767 struct rb_node **link = &dl_rq->rb_root.rb_node; 775 struct rb_node **link = &dl_rq->rb_root.rb_node;
768 struct rb_node *parent = NULL; 776 struct rb_node *parent = NULL;
769 struct sched_dl_entity *entry; 777 struct sched_dl_entity *entry;
770 int leftmost = 1; 778 int leftmost = 1;
771 779
772 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node)); 780 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
773 781
774 while (*link) { 782 while (*link) {
775 parent = *link; 783 parent = *link;
776 entry = rb_entry(parent, struct sched_dl_entity, rb_node); 784 entry = rb_entry(parent, struct sched_dl_entity, rb_node);
777 if (dl_time_before(dl_se->deadline, entry->deadline)) 785 if (dl_time_before(dl_se->deadline, entry->deadline))
778 link = &parent->rb_left; 786 link = &parent->rb_left;
779 else { 787 else {
780 link = &parent->rb_right; 788 link = &parent->rb_right;
781 leftmost = 0; 789 leftmost = 0;
782 } 790 }
783 } 791 }
784 792
785 if (leftmost) 793 if (leftmost)
786 dl_rq->rb_leftmost = &dl_se->rb_node; 794 dl_rq->rb_leftmost = &dl_se->rb_node;
787 795
788 rb_link_node(&dl_se->rb_node, parent, link); 796 rb_link_node(&dl_se->rb_node, parent, link);
789 rb_insert_color(&dl_se->rb_node, &dl_rq->rb_root); 797 rb_insert_color(&dl_se->rb_node, &dl_rq->rb_root);
790 798
791 inc_dl_tasks(dl_se, dl_rq); 799 inc_dl_tasks(dl_se, dl_rq);
792 } 800 }
793 801
794 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se) 802 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
795 { 803 {
796 struct dl_rq *dl_rq = dl_rq_of_se(dl_se); 804 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
797 805
798 if (RB_EMPTY_NODE(&dl_se->rb_node)) 806 if (RB_EMPTY_NODE(&dl_se->rb_node))
799 return; 807 return;
800 808
801 if (dl_rq->rb_leftmost == &dl_se->rb_node) { 809 if (dl_rq->rb_leftmost == &dl_se->rb_node) {
802 struct rb_node *next_node; 810 struct rb_node *next_node;
803 811
804 next_node = rb_next(&dl_se->rb_node); 812 next_node = rb_next(&dl_se->rb_node);
805 dl_rq->rb_leftmost = next_node; 813 dl_rq->rb_leftmost = next_node;
806 } 814 }
807 815
808 rb_erase(&dl_se->rb_node, &dl_rq->rb_root); 816 rb_erase(&dl_se->rb_node, &dl_rq->rb_root);
809 RB_CLEAR_NODE(&dl_se->rb_node); 817 RB_CLEAR_NODE(&dl_se->rb_node);
810 818
811 dec_dl_tasks(dl_se, dl_rq); 819 dec_dl_tasks(dl_se, dl_rq);
812 } 820 }
813 821
814 static void 822 static void
815 enqueue_dl_entity(struct sched_dl_entity *dl_se, 823 enqueue_dl_entity(struct sched_dl_entity *dl_se,
816 struct sched_dl_entity *pi_se, int flags) 824 struct sched_dl_entity *pi_se, int flags)
817 { 825 {
818 BUG_ON(on_dl_rq(dl_se)); 826 BUG_ON(on_dl_rq(dl_se));
819 827
820 /* 828 /*
821 * If this is a wakeup or a new instance, the scheduling 829 * If this is a wakeup or a new instance, the scheduling
822 * parameters of the task might need updating. Otherwise, 830 * parameters of the task might need updating. Otherwise,
823 * we want a replenishment of its runtime. 831 * we want a replenishment of its runtime.
824 */ 832 */
825 if (!dl_se->dl_new && flags & ENQUEUE_REPLENISH) 833 if (!dl_se->dl_new && flags & ENQUEUE_REPLENISH)
826 replenish_dl_entity(dl_se, pi_se); 834 replenish_dl_entity(dl_se, pi_se);
827 else 835 else
828 update_dl_entity(dl_se, pi_se); 836 update_dl_entity(dl_se, pi_se);
829 837
830 __enqueue_dl_entity(dl_se); 838 __enqueue_dl_entity(dl_se);
831 } 839 }
832 840
833 static void dequeue_dl_entity(struct sched_dl_entity *dl_se) 841 static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
834 { 842 {
835 __dequeue_dl_entity(dl_se); 843 __dequeue_dl_entity(dl_se);
836 } 844 }
837 845
838 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags) 846 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
839 { 847 {
840 struct task_struct *pi_task = rt_mutex_get_top_task(p); 848 struct task_struct *pi_task = rt_mutex_get_top_task(p);
841 struct sched_dl_entity *pi_se = &p->dl; 849 struct sched_dl_entity *pi_se = &p->dl;
842 850
843 /* 851 /*
844 * Use the scheduling parameters of the top pi-waiter 852 * Use the scheduling parameters of the top pi-waiter
845 * task if we have one and its (relative) deadline is 853 * task if we have one and its (relative) deadline is
846 * smaller than our one... OTW we keep our runtime and 854 * smaller than our one... OTW we keep our runtime and
847 * deadline. 855 * deadline.
848 */ 856 */
849 if (pi_task && p->dl.dl_boosted && dl_prio(pi_task->normal_prio)) 857 if (pi_task && p->dl.dl_boosted && dl_prio(pi_task->normal_prio))
850 pi_se = &pi_task->dl; 858 pi_se = &pi_task->dl;
851 859
852 /* 860 /*
853 * If p is throttled, we do nothing. In fact, if it exhausted 861 * If p is throttled, we do nothing. In fact, if it exhausted
854 * its budget it needs a replenishment and, since it now is on 862 * its budget it needs a replenishment and, since it now is on
855 * its rq, the bandwidth timer callback (which clearly has not 863 * its rq, the bandwidth timer callback (which clearly has not
856 * run yet) will take care of this. 864 * run yet) will take care of this.
857 */ 865 */
858 if (p->dl.dl_throttled) 866 if (p->dl.dl_throttled)
859 return; 867 return;
860 868
861 enqueue_dl_entity(&p->dl, pi_se, flags); 869 enqueue_dl_entity(&p->dl, pi_se, flags);
862 870
863 if (!task_current(rq, p) && p->nr_cpus_allowed > 1) 871 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
864 enqueue_pushable_dl_task(rq, p); 872 enqueue_pushable_dl_task(rq, p);
865 } 873 }
866 874
867 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) 875 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
868 { 876 {
869 dequeue_dl_entity(&p->dl); 877 dequeue_dl_entity(&p->dl);
870 dequeue_pushable_dl_task(rq, p); 878 dequeue_pushable_dl_task(rq, p);
871 } 879 }
872 880
873 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags) 881 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
874 { 882 {
875 update_curr_dl(rq); 883 update_curr_dl(rq);
876 __dequeue_task_dl(rq, p, flags); 884 __dequeue_task_dl(rq, p, flags);
877 } 885 }
878 886
879 /* 887 /*
880 * Yield task semantic for -deadline tasks is: 888 * Yield task semantic for -deadline tasks is:
881 * 889 *
882 * get off from the CPU until our next instance, with 890 * get off from the CPU until our next instance, with
883 * a new runtime. This is of little use now, since we 891 * a new runtime. This is of little use now, since we
884 * don't have a bandwidth reclaiming mechanism. Anyway, 892 * don't have a bandwidth reclaiming mechanism. Anyway,
885 * bandwidth reclaiming is planned for the future, and 893 * bandwidth reclaiming is planned for the future, and
886 * yield_task_dl will indicate that some spare budget 894 * yield_task_dl will indicate that some spare budget
887 * is available for other task instances to use it. 895 * is available for other task instances to use it.
888 */ 896 */
889 static void yield_task_dl(struct rq *rq) 897 static void yield_task_dl(struct rq *rq)
890 { 898 {
891 struct task_struct *p = rq->curr; 899 struct task_struct *p = rq->curr;
892 900
893 /* 901 /*
894 * We make the task go to sleep until its current deadline by 902 * We make the task go to sleep until its current deadline by
895 * forcing its runtime to zero. This way, update_curr_dl() stops 903 * forcing its runtime to zero. This way, update_curr_dl() stops
896 * it and the bandwidth timer will wake it up and will give it 904 * it and the bandwidth timer will wake it up and will give it
897 * new scheduling parameters (thanks to dl_yielded=1). 905 * new scheduling parameters (thanks to dl_yielded=1).
898 */ 906 */
899 if (p->dl.runtime > 0) { 907 if (p->dl.runtime > 0) {
900 rq->curr->dl.dl_yielded = 1; 908 rq->curr->dl.dl_yielded = 1;
901 p->dl.runtime = 0; 909 p->dl.runtime = 0;
902 } 910 }
903 update_curr_dl(rq); 911 update_curr_dl(rq);
904 } 912 }
905 913
906 #ifdef CONFIG_SMP 914 #ifdef CONFIG_SMP
907 915
908 static int find_later_rq(struct task_struct *task); 916 static int find_later_rq(struct task_struct *task);
909 917
910 static int 918 static int
911 select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags) 919 select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags)
912 { 920 {
913 struct task_struct *curr; 921 struct task_struct *curr;
914 struct rq *rq; 922 struct rq *rq;
915 923
916 if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK) 924 if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
917 goto out; 925 goto out;
918 926
919 rq = cpu_rq(cpu); 927 rq = cpu_rq(cpu);
920 928
921 rcu_read_lock(); 929 rcu_read_lock();
922 curr = ACCESS_ONCE(rq->curr); /* unlocked access */ 930 curr = ACCESS_ONCE(rq->curr); /* unlocked access */
923 931
924 /* 932 /*
925 * If we are dealing with a -deadline task, we must 933 * If we are dealing with a -deadline task, we must
926 * decide where to wake it up. 934 * decide where to wake it up.
927 * If it has a later deadline and the current task 935 * If it has a later deadline and the current task
928 * on this rq can't move (provided the waking task 936 * on this rq can't move (provided the waking task
929 * can!) we prefer to send it somewhere else. On the 937 * can!) we prefer to send it somewhere else. On the
930 * other hand, if it has a shorter deadline, we 938 * other hand, if it has a shorter deadline, we
931 * try to make it stay here, it might be important. 939 * try to make it stay here, it might be important.
932 */ 940 */
933 if (unlikely(dl_task(curr)) && 941 if (unlikely(dl_task(curr)) &&
934 (curr->nr_cpus_allowed < 2 || 942 (curr->nr_cpus_allowed < 2 ||
935 !dl_entity_preempt(&p->dl, &curr->dl)) && 943 !dl_entity_preempt(&p->dl, &curr->dl)) &&
936 (p->nr_cpus_allowed > 1)) { 944 (p->nr_cpus_allowed > 1)) {
937 int target = find_later_rq(p); 945 int target = find_later_rq(p);
938 946
939 if (target != -1) 947 if (target != -1)
940 cpu = target; 948 cpu = target;
941 } 949 }
942 rcu_read_unlock(); 950 rcu_read_unlock();
943 951
944 out: 952 out:
945 return cpu; 953 return cpu;
946 } 954 }
947 955
948 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p) 956 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
949 { 957 {
950 /* 958 /*
951 * Current can't be migrated, useless to reschedule, 959 * Current can't be migrated, useless to reschedule,
952 * let's hope p can move out. 960 * let's hope p can move out.
953 */ 961 */
954 if (rq->curr->nr_cpus_allowed == 1 || 962 if (rq->curr->nr_cpus_allowed == 1 ||
955 cpudl_find(&rq->rd->cpudl, rq->curr, NULL) == -1) 963 cpudl_find(&rq->rd->cpudl, rq->curr, NULL) == -1)
956 return; 964 return;
957 965
958 /* 966 /*
959 * p is migratable, so let's not schedule it and 967 * p is migratable, so let's not schedule it and
960 * see if it is pushed or pulled somewhere else. 968 * see if it is pushed or pulled somewhere else.
961 */ 969 */
962 if (p->nr_cpus_allowed != 1 && 970 if (p->nr_cpus_allowed != 1 &&
963 cpudl_find(&rq->rd->cpudl, p, NULL) != -1) 971 cpudl_find(&rq->rd->cpudl, p, NULL) != -1)
964 return; 972 return;
965 973
966 resched_task(rq->curr); 974 resched_task(rq->curr);
967 } 975 }
968 976
969 static int pull_dl_task(struct rq *this_rq); 977 static int pull_dl_task(struct rq *this_rq);
970 978
971 #endif /* CONFIG_SMP */ 979 #endif /* CONFIG_SMP */
972 980
973 /* 981 /*
974 * Only called when both the current and waking task are -deadline 982 * Only called when both the current and waking task are -deadline
975 * tasks. 983 * tasks.
976 */ 984 */
977 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, 985 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
978 int flags) 986 int flags)
979 { 987 {
980 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) { 988 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
981 resched_task(rq->curr); 989 resched_task(rq->curr);
982 return; 990 return;
983 } 991 }
984 992
985 #ifdef CONFIG_SMP 993 #ifdef CONFIG_SMP
986 /* 994 /*
987 * In the unlikely case current and p have the same deadline 995 * In the unlikely case current and p have the same deadline
988 * let us try to decide what's the best thing to do... 996 * let us try to decide what's the best thing to do...
989 */ 997 */
990 if ((p->dl.deadline == rq->curr->dl.deadline) && 998 if ((p->dl.deadline == rq->curr->dl.deadline) &&
991 !test_tsk_need_resched(rq->curr)) 999 !test_tsk_need_resched(rq->curr))
992 check_preempt_equal_dl(rq, p); 1000 check_preempt_equal_dl(rq, p);
993 #endif /* CONFIG_SMP */ 1001 #endif /* CONFIG_SMP */
994 } 1002 }
995 1003
996 #ifdef CONFIG_SCHED_HRTICK 1004 #ifdef CONFIG_SCHED_HRTICK
997 static void start_hrtick_dl(struct rq *rq, struct task_struct *p) 1005 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
998 { 1006 {
999 s64 delta = p->dl.dl_runtime - p->dl.runtime; 1007 s64 delta = p->dl.dl_runtime - p->dl.runtime;
1000 1008
1001 if (delta > 10000) 1009 if (delta > 10000)
1002 hrtick_start(rq, p->dl.runtime); 1010 hrtick_start(rq, p->dl.runtime);
1003 } 1011 }
1004 #endif 1012 #endif
1005 1013
1006 static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq, 1014 static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq,
1007 struct dl_rq *dl_rq) 1015 struct dl_rq *dl_rq)
1008 { 1016 {
1009 struct rb_node *left = dl_rq->rb_leftmost; 1017 struct rb_node *left = dl_rq->rb_leftmost;
1010 1018
1011 if (!left) 1019 if (!left)
1012 return NULL; 1020 return NULL;
1013 1021
1014 return rb_entry(left, struct sched_dl_entity, rb_node); 1022 return rb_entry(left, struct sched_dl_entity, rb_node);
1015 } 1023 }
1016 1024
1017 struct task_struct *pick_next_task_dl(struct rq *rq, struct task_struct *prev) 1025 struct task_struct *pick_next_task_dl(struct rq *rq, struct task_struct *prev)
1018 { 1026 {
1019 struct sched_dl_entity *dl_se; 1027 struct sched_dl_entity *dl_se;
1020 struct task_struct *p; 1028 struct task_struct *p;
1021 struct dl_rq *dl_rq; 1029 struct dl_rq *dl_rq;
1022 1030
1023 dl_rq = &rq->dl; 1031 dl_rq = &rq->dl;
1024 1032
1025 if (need_pull_dl_task(rq, prev)) { 1033 if (need_pull_dl_task(rq, prev)) {
1026 pull_dl_task(rq); 1034 pull_dl_task(rq);
1027 /* 1035 /*
1028 * pull_rt_task() can drop (and re-acquire) rq->lock; this 1036 * pull_rt_task() can drop (and re-acquire) rq->lock; this
1029 * means a stop task can slip in, in which case we need to 1037 * means a stop task can slip in, in which case we need to
1030 * re-start task selection. 1038 * re-start task selection.
1031 */ 1039 */
1032 if (rq->stop && rq->stop->on_rq) 1040 if (rq->stop && rq->stop->on_rq)
1033 return RETRY_TASK; 1041 return RETRY_TASK;
1034 } 1042 }
1035 1043
1036 /* 1044 /*
1037 * When prev is DL, we may throttle it in put_prev_task(). 1045 * When prev is DL, we may throttle it in put_prev_task().
1038 * So, we update time before we check for dl_nr_running. 1046 * So, we update time before we check for dl_nr_running.
1039 */ 1047 */
1040 if (prev->sched_class == &dl_sched_class) 1048 if (prev->sched_class == &dl_sched_class)
1041 update_curr_dl(rq); 1049 update_curr_dl(rq);
1042 1050
1043 if (unlikely(!dl_rq->dl_nr_running)) 1051 if (unlikely(!dl_rq->dl_nr_running))
1044 return NULL; 1052 return NULL;
1045 1053
1046 put_prev_task(rq, prev); 1054 put_prev_task(rq, prev);
1047 1055
1048 dl_se = pick_next_dl_entity(rq, dl_rq); 1056 dl_se = pick_next_dl_entity(rq, dl_rq);
1049 BUG_ON(!dl_se); 1057 BUG_ON(!dl_se);
1050 1058
1051 p = dl_task_of(dl_se); 1059 p = dl_task_of(dl_se);
1052 p->se.exec_start = rq_clock_task(rq); 1060 p->se.exec_start = rq_clock_task(rq);
1053 1061
1054 /* Running task will never be pushed. */ 1062 /* Running task will never be pushed. */
1055 dequeue_pushable_dl_task(rq, p); 1063 dequeue_pushable_dl_task(rq, p);
1056 1064
1057 #ifdef CONFIG_SCHED_HRTICK 1065 #ifdef CONFIG_SCHED_HRTICK
1058 if (hrtick_enabled(rq)) 1066 if (hrtick_enabled(rq))
1059 start_hrtick_dl(rq, p); 1067 start_hrtick_dl(rq, p);
1060 #endif 1068 #endif
1061 1069
1062 set_post_schedule(rq); 1070 set_post_schedule(rq);
1063 1071
1064 return p; 1072 return p;
1065 } 1073 }
1066 1074
1067 static void put_prev_task_dl(struct rq *rq, struct task_struct *p) 1075 static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
1068 { 1076 {
1069 update_curr_dl(rq); 1077 update_curr_dl(rq);
1070 1078
1071 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1) 1079 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
1072 enqueue_pushable_dl_task(rq, p); 1080 enqueue_pushable_dl_task(rq, p);
1073 } 1081 }
1074 1082
1075 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued) 1083 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
1076 { 1084 {
1077 update_curr_dl(rq); 1085 update_curr_dl(rq);
1078 1086
1079 #ifdef CONFIG_SCHED_HRTICK 1087 #ifdef CONFIG_SCHED_HRTICK
1080 if (hrtick_enabled(rq) && queued && p->dl.runtime > 0) 1088 if (hrtick_enabled(rq) && queued && p->dl.runtime > 0)
1081 start_hrtick_dl(rq, p); 1089 start_hrtick_dl(rq, p);
1082 #endif 1090 #endif
1083 } 1091 }
1084 1092
1085 static void task_fork_dl(struct task_struct *p) 1093 static void task_fork_dl(struct task_struct *p)
1086 { 1094 {
1087 /* 1095 /*
1088 * SCHED_DEADLINE tasks cannot fork and this is achieved through 1096 * SCHED_DEADLINE tasks cannot fork and this is achieved through
1089 * sched_fork() 1097 * sched_fork()
1090 */ 1098 */
1091 } 1099 }
1092 1100
1093 static void task_dead_dl(struct task_struct *p) 1101 static void task_dead_dl(struct task_struct *p)
1094 { 1102 {
1095 struct hrtimer *timer = &p->dl.dl_timer; 1103 struct hrtimer *timer = &p->dl.dl_timer;
1096 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 1104 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1097 1105
1098 /* 1106 /*
1099 * Since we are TASK_DEAD we won't slip out of the domain! 1107 * Since we are TASK_DEAD we won't slip out of the domain!
1100 */ 1108 */
1101 raw_spin_lock_irq(&dl_b->lock); 1109 raw_spin_lock_irq(&dl_b->lock);
1102 dl_b->total_bw -= p->dl.dl_bw; 1110 dl_b->total_bw -= p->dl.dl_bw;
1103 raw_spin_unlock_irq(&dl_b->lock); 1111 raw_spin_unlock_irq(&dl_b->lock);
1104 1112
1105 hrtimer_cancel(timer); 1113 hrtimer_cancel(timer);
1106 } 1114 }
1107 1115
1108 static void set_curr_task_dl(struct rq *rq) 1116 static void set_curr_task_dl(struct rq *rq)
1109 { 1117 {
1110 struct task_struct *p = rq->curr; 1118 struct task_struct *p = rq->curr;
1111 1119
1112 p->se.exec_start = rq_clock_task(rq); 1120 p->se.exec_start = rq_clock_task(rq);
1113 1121
1114 /* You can't push away the running task */ 1122 /* You can't push away the running task */
1115 dequeue_pushable_dl_task(rq, p); 1123 dequeue_pushable_dl_task(rq, p);
1116 } 1124 }
1117 1125
1118 #ifdef CONFIG_SMP 1126 #ifdef CONFIG_SMP
1119 1127
1120 /* Only try algorithms three times */ 1128 /* Only try algorithms three times */
1121 #define DL_MAX_TRIES 3 1129 #define DL_MAX_TRIES 3
1122 1130
1123 static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu) 1131 static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
1124 { 1132 {
1125 if (!task_running(rq, p) && 1133 if (!task_running(rq, p) &&
1126 (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) && 1134 (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
1127 (p->nr_cpus_allowed > 1)) 1135 (p->nr_cpus_allowed > 1))
1128 return 1; 1136 return 1;
1129 1137
1130 return 0; 1138 return 0;
1131 } 1139 }
1132 1140
1133 /* Returns the second earliest -deadline task, NULL otherwise */ 1141 /* Returns the second earliest -deadline task, NULL otherwise */
1134 static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu) 1142 static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu)
1135 { 1143 {
1136 struct rb_node *next_node = rq->dl.rb_leftmost; 1144 struct rb_node *next_node = rq->dl.rb_leftmost;
1137 struct sched_dl_entity *dl_se; 1145 struct sched_dl_entity *dl_se;
1138 struct task_struct *p = NULL; 1146 struct task_struct *p = NULL;
1139 1147
1140 next_node: 1148 next_node:
1141 next_node = rb_next(next_node); 1149 next_node = rb_next(next_node);
1142 if (next_node) { 1150 if (next_node) {
1143 dl_se = rb_entry(next_node, struct sched_dl_entity, rb_node); 1151 dl_se = rb_entry(next_node, struct sched_dl_entity, rb_node);
1144 p = dl_task_of(dl_se); 1152 p = dl_task_of(dl_se);
1145 1153
1146 if (pick_dl_task(rq, p, cpu)) 1154 if (pick_dl_task(rq, p, cpu))
1147 return p; 1155 return p;
1148 1156
1149 goto next_node; 1157 goto next_node;
1150 } 1158 }
1151 1159
1152 return NULL; 1160 return NULL;
1153 } 1161 }
1154 1162
1155 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl); 1163 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
1156 1164
1157 static int find_later_rq(struct task_struct *task) 1165 static int find_later_rq(struct task_struct *task)
1158 { 1166 {
1159 struct sched_domain *sd; 1167 struct sched_domain *sd;
1160 struct cpumask *later_mask = __get_cpu_var(local_cpu_mask_dl); 1168 struct cpumask *later_mask = __get_cpu_var(local_cpu_mask_dl);
1161 int this_cpu = smp_processor_id(); 1169 int this_cpu = smp_processor_id();
1162 int best_cpu, cpu = task_cpu(task); 1170 int best_cpu, cpu = task_cpu(task);
1163 1171
1164 /* Make sure the mask is initialized first */ 1172 /* Make sure the mask is initialized first */
1165 if (unlikely(!later_mask)) 1173 if (unlikely(!later_mask))
1166 return -1; 1174 return -1;
1167 1175
1168 if (task->nr_cpus_allowed == 1) 1176 if (task->nr_cpus_allowed == 1)
1169 return -1; 1177 return -1;
1170 1178
1171 best_cpu = cpudl_find(&task_rq(task)->rd->cpudl, 1179 best_cpu = cpudl_find(&task_rq(task)->rd->cpudl,
1172 task, later_mask); 1180 task, later_mask);
1173 if (best_cpu == -1) 1181 if (best_cpu == -1)
1174 return -1; 1182 return -1;
1175 1183
1176 /* 1184 /*
1177 * If we are here, some target has been found, 1185 * If we are here, some target has been found,
1178 * the most suitable of which is cached in best_cpu. 1186 * the most suitable of which is cached in best_cpu.
1179 * This is, among the runqueues where the current tasks 1187 * This is, among the runqueues where the current tasks
1180 * have later deadlines than the task's one, the rq 1188 * have later deadlines than the task's one, the rq
1181 * with the latest possible one. 1189 * with the latest possible one.
1182 * 1190 *
1183 * Now we check how well this matches with task's 1191 * Now we check how well this matches with task's
1184 * affinity and system topology. 1192 * affinity and system topology.
1185 * 1193 *
1186 * The last cpu where the task run is our first 1194 * The last cpu where the task run is our first
1187 * guess, since it is most likely cache-hot there. 1195 * guess, since it is most likely cache-hot there.
1188 */ 1196 */
1189 if (cpumask_test_cpu(cpu, later_mask)) 1197 if (cpumask_test_cpu(cpu, later_mask))
1190 return cpu; 1198 return cpu;
1191 /* 1199 /*
1192 * Check if this_cpu is to be skipped (i.e., it is 1200 * Check if this_cpu is to be skipped (i.e., it is
1193 * not in the mask) or not. 1201 * not in the mask) or not.
1194 */ 1202 */
1195 if (!cpumask_test_cpu(this_cpu, later_mask)) 1203 if (!cpumask_test_cpu(this_cpu, later_mask))
1196 this_cpu = -1; 1204 this_cpu = -1;
1197 1205
1198 rcu_read_lock(); 1206 rcu_read_lock();
1199 for_each_domain(cpu, sd) { 1207 for_each_domain(cpu, sd) {
1200 if (sd->flags & SD_WAKE_AFFINE) { 1208 if (sd->flags & SD_WAKE_AFFINE) {
1201 1209
1202 /* 1210 /*
1203 * If possible, preempting this_cpu is 1211 * If possible, preempting this_cpu is
1204 * cheaper than migrating. 1212 * cheaper than migrating.
1205 */ 1213 */
1206 if (this_cpu != -1 && 1214 if (this_cpu != -1 &&
1207 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) { 1215 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
1208 rcu_read_unlock(); 1216 rcu_read_unlock();
1209 return this_cpu; 1217 return this_cpu;
1210 } 1218 }
1211 1219
1212 /* 1220 /*
1213 * Last chance: if best_cpu is valid and is 1221 * Last chance: if best_cpu is valid and is
1214 * in the mask, that becomes our choice. 1222 * in the mask, that becomes our choice.
1215 */ 1223 */
1216 if (best_cpu < nr_cpu_ids && 1224 if (best_cpu < nr_cpu_ids &&
1217 cpumask_test_cpu(best_cpu, sched_domain_span(sd))) { 1225 cpumask_test_cpu(best_cpu, sched_domain_span(sd))) {
1218 rcu_read_unlock(); 1226 rcu_read_unlock();
1219 return best_cpu; 1227 return best_cpu;
1220 } 1228 }
1221 } 1229 }
1222 } 1230 }
1223 rcu_read_unlock(); 1231 rcu_read_unlock();
1224 1232
1225 /* 1233 /*
1226 * At this point, all our guesses failed, we just return 1234 * At this point, all our guesses failed, we just return
1227 * 'something', and let the caller sort the things out. 1235 * 'something', and let the caller sort the things out.
1228 */ 1236 */
1229 if (this_cpu != -1) 1237 if (this_cpu != -1)
1230 return this_cpu; 1238 return this_cpu;
1231 1239
1232 cpu = cpumask_any(later_mask); 1240 cpu = cpumask_any(later_mask);
1233 if (cpu < nr_cpu_ids) 1241 if (cpu < nr_cpu_ids)
1234 return cpu; 1242 return cpu;
1235 1243
1236 return -1; 1244 return -1;
1237 } 1245 }
1238 1246
1239 /* Locks the rq it finds */ 1247 /* Locks the rq it finds */
1240 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq) 1248 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
1241 { 1249 {
1242 struct rq *later_rq = NULL; 1250 struct rq *later_rq = NULL;
1243 int tries; 1251 int tries;
1244 int cpu; 1252 int cpu;
1245 1253
1246 for (tries = 0; tries < DL_MAX_TRIES; tries++) { 1254 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
1247 cpu = find_later_rq(task); 1255 cpu = find_later_rq(task);
1248 1256
1249 if ((cpu == -1) || (cpu == rq->cpu)) 1257 if ((cpu == -1) || (cpu == rq->cpu))
1250 break; 1258 break;
1251 1259
1252 later_rq = cpu_rq(cpu); 1260 later_rq = cpu_rq(cpu);
1253 1261
1254 /* Retry if something changed. */ 1262 /* Retry if something changed. */
1255 if (double_lock_balance(rq, later_rq)) { 1263 if (double_lock_balance(rq, later_rq)) {
1256 if (unlikely(task_rq(task) != rq || 1264 if (unlikely(task_rq(task) != rq ||
1257 !cpumask_test_cpu(later_rq->cpu, 1265 !cpumask_test_cpu(later_rq->cpu,
1258 &task->cpus_allowed) || 1266 &task->cpus_allowed) ||
1259 task_running(rq, task) || !task->on_rq)) { 1267 task_running(rq, task) || !task->on_rq)) {
1260 double_unlock_balance(rq, later_rq); 1268 double_unlock_balance(rq, later_rq);
1261 later_rq = NULL; 1269 later_rq = NULL;
1262 break; 1270 break;
1263 } 1271 }
1264 } 1272 }
1265 1273
1266 /* 1274 /*
1267 * If the rq we found has no -deadline task, or 1275 * If the rq we found has no -deadline task, or
1268 * its earliest one has a later deadline than our 1276 * its earliest one has a later deadline than our
1269 * task, the rq is a good one. 1277 * task, the rq is a good one.
1270 */ 1278 */
1271 if (!later_rq->dl.dl_nr_running || 1279 if (!later_rq->dl.dl_nr_running ||
1272 dl_time_before(task->dl.deadline, 1280 dl_time_before(task->dl.deadline,
1273 later_rq->dl.earliest_dl.curr)) 1281 later_rq->dl.earliest_dl.curr))
1274 break; 1282 break;
1275 1283
1276 /* Otherwise we try again. */ 1284 /* Otherwise we try again. */
1277 double_unlock_balance(rq, later_rq); 1285 double_unlock_balance(rq, later_rq);
1278 later_rq = NULL; 1286 later_rq = NULL;
1279 } 1287 }
1280 1288
1281 return later_rq; 1289 return later_rq;
1282 } 1290 }
1283 1291
1284 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq) 1292 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
1285 { 1293 {
1286 struct task_struct *p; 1294 struct task_struct *p;
1287 1295
1288 if (!has_pushable_dl_tasks(rq)) 1296 if (!has_pushable_dl_tasks(rq))
1289 return NULL; 1297 return NULL;
1290 1298
1291 p = rb_entry(rq->dl.pushable_dl_tasks_leftmost, 1299 p = rb_entry(rq->dl.pushable_dl_tasks_leftmost,
1292 struct task_struct, pushable_dl_tasks); 1300 struct task_struct, pushable_dl_tasks);
1293 1301
1294 BUG_ON(rq->cpu != task_cpu(p)); 1302 BUG_ON(rq->cpu != task_cpu(p));
1295 BUG_ON(task_current(rq, p)); 1303 BUG_ON(task_current(rq, p));
1296 BUG_ON(p->nr_cpus_allowed <= 1); 1304 BUG_ON(p->nr_cpus_allowed <= 1);
1297 1305
1298 BUG_ON(!p->on_rq); 1306 BUG_ON(!p->on_rq);
1299 BUG_ON(!dl_task(p)); 1307 BUG_ON(!dl_task(p));
1300 1308
1301 return p; 1309 return p;
1302 } 1310 }
1303 1311
1304 /* 1312 /*
1305 * See if the non running -deadline tasks on this rq 1313 * See if the non running -deadline tasks on this rq
1306 * can be sent to some other CPU where they can preempt 1314 * can be sent to some other CPU where they can preempt
1307 * and start executing. 1315 * and start executing.
1308 */ 1316 */
1309 static int push_dl_task(struct rq *rq) 1317 static int push_dl_task(struct rq *rq)
1310 { 1318 {
1311 struct task_struct *next_task; 1319 struct task_struct *next_task;
1312 struct rq *later_rq; 1320 struct rq *later_rq;
1313 1321
1314 if (!rq->dl.overloaded) 1322 if (!rq->dl.overloaded)
1315 return 0; 1323 return 0;
1316 1324
1317 next_task = pick_next_pushable_dl_task(rq); 1325 next_task = pick_next_pushable_dl_task(rq);
1318 if (!next_task) 1326 if (!next_task)
1319 return 0; 1327 return 0;
1320 1328
1321 retry: 1329 retry:
1322 if (unlikely(next_task == rq->curr)) { 1330 if (unlikely(next_task == rq->curr)) {
1323 WARN_ON(1); 1331 WARN_ON(1);
1324 return 0; 1332 return 0;
1325 } 1333 }
1326 1334
1327 /* 1335 /*
1328 * If next_task preempts rq->curr, and rq->curr 1336 * If next_task preempts rq->curr, and rq->curr
1329 * can move away, it makes sense to just reschedule 1337 * can move away, it makes sense to just reschedule
1330 * without going further in pushing next_task. 1338 * without going further in pushing next_task.
1331 */ 1339 */
1332 if (dl_task(rq->curr) && 1340 if (dl_task(rq->curr) &&
1333 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) && 1341 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
1334 rq->curr->nr_cpus_allowed > 1) { 1342 rq->curr->nr_cpus_allowed > 1) {
1335 resched_task(rq->curr); 1343 resched_task(rq->curr);
1336 return 0; 1344 return 0;
1337 } 1345 }
1338 1346
1339 /* We might release rq lock */ 1347 /* We might release rq lock */
1340 get_task_struct(next_task); 1348 get_task_struct(next_task);
1341 1349
1342 /* Will lock the rq it'll find */ 1350 /* Will lock the rq it'll find */
1343 later_rq = find_lock_later_rq(next_task, rq); 1351 later_rq = find_lock_later_rq(next_task, rq);
1344 if (!later_rq) { 1352 if (!later_rq) {
1345 struct task_struct *task; 1353 struct task_struct *task;
1346 1354
1347 /* 1355 /*
1348 * We must check all this again, since 1356 * We must check all this again, since
1349 * find_lock_later_rq releases rq->lock and it is 1357 * find_lock_later_rq releases rq->lock and it is
1350 * then possible that next_task has migrated. 1358 * then possible that next_task has migrated.
1351 */ 1359 */
1352 task = pick_next_pushable_dl_task(rq); 1360 task = pick_next_pushable_dl_task(rq);
1353 if (task_cpu(next_task) == rq->cpu && task == next_task) { 1361 if (task_cpu(next_task) == rq->cpu && task == next_task) {
1354 /* 1362 /*
1355 * The task is still there. We don't try 1363 * The task is still there. We don't try
1356 * again, some other cpu will pull it when ready. 1364 * again, some other cpu will pull it when ready.
1357 */ 1365 */
1358 dequeue_pushable_dl_task(rq, next_task); 1366 dequeue_pushable_dl_task(rq, next_task);
1359 goto out; 1367 goto out;
1360 } 1368 }
1361 1369
1362 if (!task) 1370 if (!task)
1363 /* No more tasks */ 1371 /* No more tasks */
1364 goto out; 1372 goto out;
1365 1373
1366 put_task_struct(next_task); 1374 put_task_struct(next_task);
1367 next_task = task; 1375 next_task = task;
1368 goto retry; 1376 goto retry;
1369 } 1377 }
1370 1378
1371 deactivate_task(rq, next_task, 0); 1379 deactivate_task(rq, next_task, 0);
1372 set_task_cpu(next_task, later_rq->cpu); 1380 set_task_cpu(next_task, later_rq->cpu);
1373 activate_task(later_rq, next_task, 0); 1381 activate_task(later_rq, next_task, 0);
1374 1382
1375 resched_task(later_rq->curr); 1383 resched_task(later_rq->curr);
1376 1384
1377 double_unlock_balance(rq, later_rq); 1385 double_unlock_balance(rq, later_rq);
1378 1386
1379 out: 1387 out:
1380 put_task_struct(next_task); 1388 put_task_struct(next_task);
1381 1389
1382 return 1; 1390 return 1;
1383 } 1391 }
1384 1392
1385 static void push_dl_tasks(struct rq *rq) 1393 static void push_dl_tasks(struct rq *rq)
1386 { 1394 {
1387 /* Terminates as it moves a -deadline task */ 1395 /* Terminates as it moves a -deadline task */
1388 while (push_dl_task(rq)) 1396 while (push_dl_task(rq))
1389 ; 1397 ;
1390 } 1398 }
1391 1399
1392 static int pull_dl_task(struct rq *this_rq) 1400 static int pull_dl_task(struct rq *this_rq)
1393 { 1401 {
1394 int this_cpu = this_rq->cpu, ret = 0, cpu; 1402 int this_cpu = this_rq->cpu, ret = 0, cpu;
1395 struct task_struct *p; 1403 struct task_struct *p;
1396 struct rq *src_rq; 1404 struct rq *src_rq;
1397 u64 dmin = LONG_MAX; 1405 u64 dmin = LONG_MAX;
1398 1406
1399 if (likely(!dl_overloaded(this_rq))) 1407 if (likely(!dl_overloaded(this_rq)))
1400 return 0; 1408 return 0;
1401 1409
1402 /* 1410 /*
1403 * Match the barrier from dl_set_overloaded; this guarantees that if we 1411 * Match the barrier from dl_set_overloaded; this guarantees that if we
1404 * see overloaded we must also see the dlo_mask bit. 1412 * see overloaded we must also see the dlo_mask bit.
1405 */ 1413 */
1406 smp_rmb(); 1414 smp_rmb();
1407 1415
1408 for_each_cpu(cpu, this_rq->rd->dlo_mask) { 1416 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
1409 if (this_cpu == cpu) 1417 if (this_cpu == cpu)
1410 continue; 1418 continue;
1411 1419
1412 src_rq = cpu_rq(cpu); 1420 src_rq = cpu_rq(cpu);
1413 1421
1414 /* 1422 /*
1415 * It looks racy, abd it is! However, as in sched_rt.c, 1423 * It looks racy, abd it is! However, as in sched_rt.c,
1416 * we are fine with this. 1424 * we are fine with this.
1417 */ 1425 */
1418 if (this_rq->dl.dl_nr_running && 1426 if (this_rq->dl.dl_nr_running &&
1419 dl_time_before(this_rq->dl.earliest_dl.curr, 1427 dl_time_before(this_rq->dl.earliest_dl.curr,
1420 src_rq->dl.earliest_dl.next)) 1428 src_rq->dl.earliest_dl.next))
1421 continue; 1429 continue;
1422 1430
1423 /* Might drop this_rq->lock */ 1431 /* Might drop this_rq->lock */
1424 double_lock_balance(this_rq, src_rq); 1432 double_lock_balance(this_rq, src_rq);
1425 1433
1426 /* 1434 /*
1427 * If there are no more pullable tasks on the 1435 * If there are no more pullable tasks on the
1428 * rq, we're done with it. 1436 * rq, we're done with it.
1429 */ 1437 */
1430 if (src_rq->dl.dl_nr_running <= 1) 1438 if (src_rq->dl.dl_nr_running <= 1)
1431 goto skip; 1439 goto skip;
1432 1440
1433 p = pick_next_earliest_dl_task(src_rq, this_cpu); 1441 p = pick_next_earliest_dl_task(src_rq, this_cpu);
1434 1442
1435 /* 1443 /*
1436 * We found a task to be pulled if: 1444 * We found a task to be pulled if:
1437 * - it preempts our current (if there's one), 1445 * - it preempts our current (if there's one),
1438 * - it will preempt the last one we pulled (if any). 1446 * - it will preempt the last one we pulled (if any).
1439 */ 1447 */
1440 if (p && dl_time_before(p->dl.deadline, dmin) && 1448 if (p && dl_time_before(p->dl.deadline, dmin) &&
1441 (!this_rq->dl.dl_nr_running || 1449 (!this_rq->dl.dl_nr_running ||
1442 dl_time_before(p->dl.deadline, 1450 dl_time_before(p->dl.deadline,
1443 this_rq->dl.earliest_dl.curr))) { 1451 this_rq->dl.earliest_dl.curr))) {
1444 WARN_ON(p == src_rq->curr); 1452 WARN_ON(p == src_rq->curr);
1445 WARN_ON(!p->on_rq); 1453 WARN_ON(!p->on_rq);
1446 1454
1447 /* 1455 /*
1448 * Then we pull iff p has actually an earlier 1456 * Then we pull iff p has actually an earlier
1449 * deadline than the current task of its runqueue. 1457 * deadline than the current task of its runqueue.
1450 */ 1458 */
1451 if (dl_time_before(p->dl.deadline, 1459 if (dl_time_before(p->dl.deadline,
1452 src_rq->curr->dl.deadline)) 1460 src_rq->curr->dl.deadline))
1453 goto skip; 1461 goto skip;
1454 1462
1455 ret = 1; 1463 ret = 1;
1456 1464
1457 deactivate_task(src_rq, p, 0); 1465 deactivate_task(src_rq, p, 0);
1458 set_task_cpu(p, this_cpu); 1466 set_task_cpu(p, this_cpu);
1459 activate_task(this_rq, p, 0); 1467 activate_task(this_rq, p, 0);
1460 dmin = p->dl.deadline; 1468 dmin = p->dl.deadline;
1461 1469
1462 /* Is there any other task even earlier? */ 1470 /* Is there any other task even earlier? */
1463 } 1471 }
1464 skip: 1472 skip:
1465 double_unlock_balance(this_rq, src_rq); 1473 double_unlock_balance(this_rq, src_rq);
1466 } 1474 }
1467 1475
1468 return ret; 1476 return ret;
1469 } 1477 }
1470 1478
1471 static void post_schedule_dl(struct rq *rq) 1479 static void post_schedule_dl(struct rq *rq)
1472 { 1480 {
1473 push_dl_tasks(rq); 1481 push_dl_tasks(rq);
1474 } 1482 }
1475 1483
1476 /* 1484 /*
1477 * Since the task is not running and a reschedule is not going to happen 1485 * Since the task is not running and a reschedule is not going to happen
1478 * anytime soon on its runqueue, we try pushing it away now. 1486 * anytime soon on its runqueue, we try pushing it away now.
1479 */ 1487 */
1480 static void task_woken_dl(struct rq *rq, struct task_struct *p) 1488 static void task_woken_dl(struct rq *rq, struct task_struct *p)
1481 { 1489 {
1482 if (!task_running(rq, p) && 1490 if (!task_running(rq, p) &&
1483 !test_tsk_need_resched(rq->curr) && 1491 !test_tsk_need_resched(rq->curr) &&
1484 has_pushable_dl_tasks(rq) && 1492 has_pushable_dl_tasks(rq) &&
1485 p->nr_cpus_allowed > 1 && 1493 p->nr_cpus_allowed > 1 &&
1486 dl_task(rq->curr) && 1494 dl_task(rq->curr) &&
1487 (rq->curr->nr_cpus_allowed < 2 || 1495 (rq->curr->nr_cpus_allowed < 2 ||
1488 dl_entity_preempt(&rq->curr->dl, &p->dl))) { 1496 dl_entity_preempt(&rq->curr->dl, &p->dl))) {
1489 push_dl_tasks(rq); 1497 push_dl_tasks(rq);
1490 } 1498 }
1491 } 1499 }
1492 1500
1493 static void set_cpus_allowed_dl(struct task_struct *p, 1501 static void set_cpus_allowed_dl(struct task_struct *p,
1494 const struct cpumask *new_mask) 1502 const struct cpumask *new_mask)
1495 { 1503 {
1496 struct rq *rq; 1504 struct rq *rq;
1497 int weight; 1505 int weight;
1498 1506
1499 BUG_ON(!dl_task(p)); 1507 BUG_ON(!dl_task(p));
1500 1508
1501 /* 1509 /*
1502 * Update only if the task is actually running (i.e., 1510 * Update only if the task is actually running (i.e.,
1503 * it is on the rq AND it is not throttled). 1511 * it is on the rq AND it is not throttled).
1504 */ 1512 */
1505 if (!on_dl_rq(&p->dl)) 1513 if (!on_dl_rq(&p->dl))
1506 return; 1514 return;
1507 1515
1508 weight = cpumask_weight(new_mask); 1516 weight = cpumask_weight(new_mask);
1509 1517
1510 /* 1518 /*
1511 * Only update if the process changes its state from whether it 1519 * Only update if the process changes its state from whether it
1512 * can migrate or not. 1520 * can migrate or not.
1513 */ 1521 */
1514 if ((p->nr_cpus_allowed > 1) == (weight > 1)) 1522 if ((p->nr_cpus_allowed > 1) == (weight > 1))
1515 return; 1523 return;
1516 1524
1517 rq = task_rq(p); 1525 rq = task_rq(p);
1518 1526
1519 /* 1527 /*
1520 * The process used to be able to migrate OR it can now migrate 1528 * The process used to be able to migrate OR it can now migrate
1521 */ 1529 */
1522 if (weight <= 1) { 1530 if (weight <= 1) {
1523 if (!task_current(rq, p)) 1531 if (!task_current(rq, p))
1524 dequeue_pushable_dl_task(rq, p); 1532 dequeue_pushable_dl_task(rq, p);
1525 BUG_ON(!rq->dl.dl_nr_migratory); 1533 BUG_ON(!rq->dl.dl_nr_migratory);
1526 rq->dl.dl_nr_migratory--; 1534 rq->dl.dl_nr_migratory--;
1527 } else { 1535 } else {
1528 if (!task_current(rq, p)) 1536 if (!task_current(rq, p))
1529 enqueue_pushable_dl_task(rq, p); 1537 enqueue_pushable_dl_task(rq, p);
1530 rq->dl.dl_nr_migratory++; 1538 rq->dl.dl_nr_migratory++;
1531 } 1539 }
1532 1540
1533 update_dl_migration(&rq->dl); 1541 update_dl_migration(&rq->dl);
1534 } 1542 }
1535 1543
1536 /* Assumes rq->lock is held */ 1544 /* Assumes rq->lock is held */
1537 static void rq_online_dl(struct rq *rq) 1545 static void rq_online_dl(struct rq *rq)
1538 { 1546 {
1539 if (rq->dl.overloaded) 1547 if (rq->dl.overloaded)
1540 dl_set_overload(rq); 1548 dl_set_overload(rq);
1541 1549
1542 if (rq->dl.dl_nr_running > 0) 1550 if (rq->dl.dl_nr_running > 0)
1543 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr, 1); 1551 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr, 1);
1544 } 1552 }
1545 1553
1546 /* Assumes rq->lock is held */ 1554 /* Assumes rq->lock is held */
1547 static void rq_offline_dl(struct rq *rq) 1555 static void rq_offline_dl(struct rq *rq)
1548 { 1556 {
1549 if (rq->dl.overloaded) 1557 if (rq->dl.overloaded)
1550 dl_clear_overload(rq); 1558 dl_clear_overload(rq);
1551 1559
1552 cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0); 1560 cpudl_set(&rq->rd->cpudl, rq->cpu, 0, 0);
1553 } 1561 }
1554 1562
1555 void init_sched_dl_class(void) 1563 void init_sched_dl_class(void)
1556 { 1564 {
1557 unsigned int i; 1565 unsigned int i;
1558 1566
1559 for_each_possible_cpu(i) 1567 for_each_possible_cpu(i)
1560 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i), 1568 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
1561 GFP_KERNEL, cpu_to_node(i)); 1569 GFP_KERNEL, cpu_to_node(i));
1562 } 1570 }
1563 1571
1564 #endif /* CONFIG_SMP */ 1572 #endif /* CONFIG_SMP */
1565 1573
1566 static void switched_from_dl(struct rq *rq, struct task_struct *p) 1574 static void switched_from_dl(struct rq *rq, struct task_struct *p)
1567 { 1575 {
1568 if (hrtimer_active(&p->dl.dl_timer) && !dl_policy(p->policy)) 1576 if (hrtimer_active(&p->dl.dl_timer) && !dl_policy(p->policy))
1569 hrtimer_try_to_cancel(&p->dl.dl_timer); 1577 hrtimer_try_to_cancel(&p->dl.dl_timer);
1570 1578
1571 #ifdef CONFIG_SMP 1579 #ifdef CONFIG_SMP
1572 /* 1580 /*
1573 * Since this might be the only -deadline task on the rq, 1581 * Since this might be the only -deadline task on the rq,
1574 * this is the right place to try to pull some other one 1582 * this is the right place to try to pull some other one
1575 * from an overloaded cpu, if any. 1583 * from an overloaded cpu, if any.
1576 */ 1584 */
1577 if (!rq->dl.dl_nr_running) 1585 if (!rq->dl.dl_nr_running)
1578 pull_dl_task(rq); 1586 pull_dl_task(rq);
1579 #endif 1587 #endif
1580 } 1588 }
1581 1589
1582 /* 1590 /*
1583 * When switching to -deadline, we may overload the rq, then 1591 * When switching to -deadline, we may overload the rq, then
1584 * we try to push someone off, if possible. 1592 * we try to push someone off, if possible.
1585 */ 1593 */
1586 static void switched_to_dl(struct rq *rq, struct task_struct *p) 1594 static void switched_to_dl(struct rq *rq, struct task_struct *p)
1587 { 1595 {
1588 int check_resched = 1; 1596 int check_resched = 1;
1589 1597
1590 /* 1598 /*
1591 * If p is throttled, don't consider the possibility 1599 * If p is throttled, don't consider the possibility
1592 * of preempting rq->curr, the check will be done right 1600 * of preempting rq->curr, the check will be done right
1593 * after its runtime will get replenished. 1601 * after its runtime will get replenished.
1594 */ 1602 */
1595 if (unlikely(p->dl.dl_throttled)) 1603 if (unlikely(p->dl.dl_throttled))
1596 return; 1604 return;
1597 1605
1598 if (p->on_rq && rq->curr != p) { 1606 if (p->on_rq && rq->curr != p) {
1599 #ifdef CONFIG_SMP 1607 #ifdef CONFIG_SMP
1600 if (rq->dl.overloaded && push_dl_task(rq) && rq != task_rq(p)) 1608 if (rq->dl.overloaded && push_dl_task(rq) && rq != task_rq(p))
1601 /* Only reschedule if pushing failed */ 1609 /* Only reschedule if pushing failed */
1602 check_resched = 0; 1610 check_resched = 0;
1603 #endif /* CONFIG_SMP */ 1611 #endif /* CONFIG_SMP */
1604 if (check_resched && task_has_dl_policy(rq->curr)) 1612 if (check_resched && task_has_dl_policy(rq->curr))
1605 check_preempt_curr_dl(rq, p, 0); 1613 check_preempt_curr_dl(rq, p, 0);
1606 } 1614 }
1607 } 1615 }
1608 1616
1609 /* 1617 /*
1610 * If the scheduling parameters of a -deadline task changed, 1618 * If the scheduling parameters of a -deadline task changed,
1611 * a push or pull operation might be needed. 1619 * a push or pull operation might be needed.
1612 */ 1620 */
1613 static void prio_changed_dl(struct rq *rq, struct task_struct *p, 1621 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
1614 int oldprio) 1622 int oldprio)
1615 { 1623 {
1616 if (p->on_rq || rq->curr == p) { 1624 if (p->on_rq || rq->curr == p) {
1617 #ifdef CONFIG_SMP 1625 #ifdef CONFIG_SMP
1618 /* 1626 /*
1619 * This might be too much, but unfortunately 1627 * This might be too much, but unfortunately
1620 * we don't have the old deadline value, and 1628 * we don't have the old deadline value, and
1621 * we can't argue if the task is increasing 1629 * we can't argue if the task is increasing
1622 * or lowering its prio, so... 1630 * or lowering its prio, so...
1623 */ 1631 */
1624 if (!rq->dl.overloaded) 1632 if (!rq->dl.overloaded)
1625 pull_dl_task(rq); 1633 pull_dl_task(rq);
1626 1634
1627 /* 1635 /*
1628 * If we now have a earlier deadline task than p, 1636 * If we now have a earlier deadline task than p,
1629 * then reschedule, provided p is still on this 1637 * then reschedule, provided p is still on this
1630 * runqueue. 1638 * runqueue.
1631 */ 1639 */
1632 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline) && 1640 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline) &&
1633 rq->curr == p) 1641 rq->curr == p)
1634 resched_task(p); 1642 resched_task(p);
1635 #else 1643 #else
1636 /* 1644 /*
1637 * Again, we don't know if p has a earlier 1645 * Again, we don't know if p has a earlier
1638 * or later deadline, so let's blindly set a 1646 * or later deadline, so let's blindly set a
1639 * (maybe not needed) rescheduling point. 1647 * (maybe not needed) rescheduling point.
1640 */ 1648 */
1641 resched_task(p); 1649 resched_task(p);
1642 #endif /* CONFIG_SMP */ 1650 #endif /* CONFIG_SMP */
1643 } else 1651 } else
1644 switched_to_dl(rq, p); 1652 switched_to_dl(rq, p);
1645 } 1653 }
1646 1654
1647 const struct sched_class dl_sched_class = { 1655 const struct sched_class dl_sched_class = {
1648 .next = &rt_sched_class, 1656 .next = &rt_sched_class,
1649 .enqueue_task = enqueue_task_dl, 1657 .enqueue_task = enqueue_task_dl,
1650 .dequeue_task = dequeue_task_dl, 1658 .dequeue_task = dequeue_task_dl,
1651 .yield_task = yield_task_dl, 1659 .yield_task = yield_task_dl,
1652 1660
1653 .check_preempt_curr = check_preempt_curr_dl, 1661 .check_preempt_curr = check_preempt_curr_dl,
1654 1662
1655 .pick_next_task = pick_next_task_dl, 1663 .pick_next_task = pick_next_task_dl,
1656 .put_prev_task = put_prev_task_dl, 1664 .put_prev_task = put_prev_task_dl,
1657 1665
1658 #ifdef CONFIG_SMP 1666 #ifdef CONFIG_SMP
1659 .select_task_rq = select_task_rq_dl, 1667 .select_task_rq = select_task_rq_dl,
1660 .set_cpus_allowed = set_cpus_allowed_dl, 1668 .set_cpus_allowed = set_cpus_allowed_dl,
1661 .rq_online = rq_online_dl, 1669 .rq_online = rq_online_dl,
1662 .rq_offline = rq_offline_dl, 1670 .rq_offline = rq_offline_dl,
1663 .post_schedule = post_schedule_dl, 1671 .post_schedule = post_schedule_dl,
1664 .task_woken = task_woken_dl, 1672 .task_woken = task_woken_dl,
1665 #endif 1673 #endif
1666 1674
1667 .set_curr_task = set_curr_task_dl, 1675 .set_curr_task = set_curr_task_dl,
1668 .task_tick = task_tick_dl, 1676 .task_tick = task_tick_dl,
1669 .task_fork = task_fork_dl, 1677 .task_fork = task_fork_dl,
1670 .task_dead = task_dead_dl, 1678 .task_dead = task_dead_dl,
1671 1679
1672 .prio_changed = prio_changed_dl, 1680 .prio_changed = prio_changed_dl,
1673 .switched_from = switched_from_dl, 1681 .switched_from = switched_from_dl,
1674 .switched_to = switched_to_dl, 1682 .switched_to = switched_to_dl,
1675 }; 1683 };
1676 1684