Commit a6b3f7614ca690e49e934c291f707b0c19312194
Committed by
Jens Axboe
1 parent
a5faeaf910
Exists in
master
and in
20 other branches
block: Reserve only one queue tag for sync IO if only 3 tags are available
In case a device has three tags available we still reserve two of them for sync IO. That leaves only a single tag for async IO such as writeback from flusher thread which results in poor performance. Allow async IO to consume two tags in case queue has three tag availabe to get a decent async write performance. This patch improves streaming write performance on a machine with such disk from ~21 MB/s to ~52 MB/s. Also postmark throughput in presence of streaming writer improves from 8 to 12 transactions per second so sync IO doesn't seem to be harmed in presence of heavy async writer. Signed-off-by: Jan Kara <jack@suse.cz> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Showing 1 changed file with 9 additions and 2 deletions Inline Diff
block/blk-tag.c
1 | /* | 1 | /* |
2 | * Functions related to tagged command queuing | 2 | * Functions related to tagged command queuing |
3 | */ | 3 | */ |
4 | #include <linux/kernel.h> | 4 | #include <linux/kernel.h> |
5 | #include <linux/module.h> | 5 | #include <linux/module.h> |
6 | #include <linux/bio.h> | 6 | #include <linux/bio.h> |
7 | #include <linux/blkdev.h> | 7 | #include <linux/blkdev.h> |
8 | #include <linux/slab.h> | 8 | #include <linux/slab.h> |
9 | 9 | ||
10 | #include "blk.h" | 10 | #include "blk.h" |
11 | 11 | ||
12 | /** | 12 | /** |
13 | * blk_queue_find_tag - find a request by its tag and queue | 13 | * blk_queue_find_tag - find a request by its tag and queue |
14 | * @q: The request queue for the device | 14 | * @q: The request queue for the device |
15 | * @tag: The tag of the request | 15 | * @tag: The tag of the request |
16 | * | 16 | * |
17 | * Notes: | 17 | * Notes: |
18 | * Should be used when a device returns a tag and you want to match | 18 | * Should be used when a device returns a tag and you want to match |
19 | * it with a request. | 19 | * it with a request. |
20 | * | 20 | * |
21 | * no locks need be held. | 21 | * no locks need be held. |
22 | **/ | 22 | **/ |
23 | struct request *blk_queue_find_tag(struct request_queue *q, int tag) | 23 | struct request *blk_queue_find_tag(struct request_queue *q, int tag) |
24 | { | 24 | { |
25 | return blk_map_queue_find_tag(q->queue_tags, tag); | 25 | return blk_map_queue_find_tag(q->queue_tags, tag); |
26 | } | 26 | } |
27 | EXPORT_SYMBOL(blk_queue_find_tag); | 27 | EXPORT_SYMBOL(blk_queue_find_tag); |
28 | 28 | ||
29 | /** | 29 | /** |
30 | * __blk_free_tags - release a given set of tag maintenance info | 30 | * __blk_free_tags - release a given set of tag maintenance info |
31 | * @bqt: the tag map to free | 31 | * @bqt: the tag map to free |
32 | * | 32 | * |
33 | * Tries to free the specified @bqt. Returns true if it was | 33 | * Tries to free the specified @bqt. Returns true if it was |
34 | * actually freed and false if there are still references using it | 34 | * actually freed and false if there are still references using it |
35 | */ | 35 | */ |
36 | static int __blk_free_tags(struct blk_queue_tag *bqt) | 36 | static int __blk_free_tags(struct blk_queue_tag *bqt) |
37 | { | 37 | { |
38 | int retval; | 38 | int retval; |
39 | 39 | ||
40 | retval = atomic_dec_and_test(&bqt->refcnt); | 40 | retval = atomic_dec_and_test(&bqt->refcnt); |
41 | if (retval) { | 41 | if (retval) { |
42 | BUG_ON(find_first_bit(bqt->tag_map, bqt->max_depth) < | 42 | BUG_ON(find_first_bit(bqt->tag_map, bqt->max_depth) < |
43 | bqt->max_depth); | 43 | bqt->max_depth); |
44 | 44 | ||
45 | kfree(bqt->tag_index); | 45 | kfree(bqt->tag_index); |
46 | bqt->tag_index = NULL; | 46 | bqt->tag_index = NULL; |
47 | 47 | ||
48 | kfree(bqt->tag_map); | 48 | kfree(bqt->tag_map); |
49 | bqt->tag_map = NULL; | 49 | bqt->tag_map = NULL; |
50 | 50 | ||
51 | kfree(bqt); | 51 | kfree(bqt); |
52 | } | 52 | } |
53 | 53 | ||
54 | return retval; | 54 | return retval; |
55 | } | 55 | } |
56 | 56 | ||
57 | /** | 57 | /** |
58 | * __blk_queue_free_tags - release tag maintenance info | 58 | * __blk_queue_free_tags - release tag maintenance info |
59 | * @q: the request queue for the device | 59 | * @q: the request queue for the device |
60 | * | 60 | * |
61 | * Notes: | 61 | * Notes: |
62 | * blk_cleanup_queue() will take care of calling this function, if tagging | 62 | * blk_cleanup_queue() will take care of calling this function, if tagging |
63 | * has been used. So there's no need to call this directly. | 63 | * has been used. So there's no need to call this directly. |
64 | **/ | 64 | **/ |
65 | void __blk_queue_free_tags(struct request_queue *q) | 65 | void __blk_queue_free_tags(struct request_queue *q) |
66 | { | 66 | { |
67 | struct blk_queue_tag *bqt = q->queue_tags; | 67 | struct blk_queue_tag *bqt = q->queue_tags; |
68 | 68 | ||
69 | if (!bqt) | 69 | if (!bqt) |
70 | return; | 70 | return; |
71 | 71 | ||
72 | __blk_free_tags(bqt); | 72 | __blk_free_tags(bqt); |
73 | 73 | ||
74 | q->queue_tags = NULL; | 74 | q->queue_tags = NULL; |
75 | queue_flag_clear_unlocked(QUEUE_FLAG_QUEUED, q); | 75 | queue_flag_clear_unlocked(QUEUE_FLAG_QUEUED, q); |
76 | } | 76 | } |
77 | 77 | ||
78 | /** | 78 | /** |
79 | * blk_free_tags - release a given set of tag maintenance info | 79 | * blk_free_tags - release a given set of tag maintenance info |
80 | * @bqt: the tag map to free | 80 | * @bqt: the tag map to free |
81 | * | 81 | * |
82 | * For externally managed @bqt frees the map. Callers of this | 82 | * For externally managed @bqt frees the map. Callers of this |
83 | * function must guarantee to have released all the queues that | 83 | * function must guarantee to have released all the queues that |
84 | * might have been using this tag map. | 84 | * might have been using this tag map. |
85 | */ | 85 | */ |
86 | void blk_free_tags(struct blk_queue_tag *bqt) | 86 | void blk_free_tags(struct blk_queue_tag *bqt) |
87 | { | 87 | { |
88 | if (unlikely(!__blk_free_tags(bqt))) | 88 | if (unlikely(!__blk_free_tags(bqt))) |
89 | BUG(); | 89 | BUG(); |
90 | } | 90 | } |
91 | EXPORT_SYMBOL(blk_free_tags); | 91 | EXPORT_SYMBOL(blk_free_tags); |
92 | 92 | ||
93 | /** | 93 | /** |
94 | * blk_queue_free_tags - release tag maintenance info | 94 | * blk_queue_free_tags - release tag maintenance info |
95 | * @q: the request queue for the device | 95 | * @q: the request queue for the device |
96 | * | 96 | * |
97 | * Notes: | 97 | * Notes: |
98 | * This is used to disable tagged queuing to a device, yet leave | 98 | * This is used to disable tagged queuing to a device, yet leave |
99 | * queue in function. | 99 | * queue in function. |
100 | **/ | 100 | **/ |
101 | void blk_queue_free_tags(struct request_queue *q) | 101 | void blk_queue_free_tags(struct request_queue *q) |
102 | { | 102 | { |
103 | queue_flag_clear_unlocked(QUEUE_FLAG_QUEUED, q); | 103 | queue_flag_clear_unlocked(QUEUE_FLAG_QUEUED, q); |
104 | } | 104 | } |
105 | EXPORT_SYMBOL(blk_queue_free_tags); | 105 | EXPORT_SYMBOL(blk_queue_free_tags); |
106 | 106 | ||
107 | static int | 107 | static int |
108 | init_tag_map(struct request_queue *q, struct blk_queue_tag *tags, int depth) | 108 | init_tag_map(struct request_queue *q, struct blk_queue_tag *tags, int depth) |
109 | { | 109 | { |
110 | struct request **tag_index; | 110 | struct request **tag_index; |
111 | unsigned long *tag_map; | 111 | unsigned long *tag_map; |
112 | int nr_ulongs; | 112 | int nr_ulongs; |
113 | 113 | ||
114 | if (q && depth > q->nr_requests * 2) { | 114 | if (q && depth > q->nr_requests * 2) { |
115 | depth = q->nr_requests * 2; | 115 | depth = q->nr_requests * 2; |
116 | printk(KERN_ERR "%s: adjusted depth to %d\n", | 116 | printk(KERN_ERR "%s: adjusted depth to %d\n", |
117 | __func__, depth); | 117 | __func__, depth); |
118 | } | 118 | } |
119 | 119 | ||
120 | tag_index = kzalloc(depth * sizeof(struct request *), GFP_ATOMIC); | 120 | tag_index = kzalloc(depth * sizeof(struct request *), GFP_ATOMIC); |
121 | if (!tag_index) | 121 | if (!tag_index) |
122 | goto fail; | 122 | goto fail; |
123 | 123 | ||
124 | nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG; | 124 | nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG; |
125 | tag_map = kzalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC); | 125 | tag_map = kzalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC); |
126 | if (!tag_map) | 126 | if (!tag_map) |
127 | goto fail; | 127 | goto fail; |
128 | 128 | ||
129 | tags->real_max_depth = depth; | 129 | tags->real_max_depth = depth; |
130 | tags->max_depth = depth; | 130 | tags->max_depth = depth; |
131 | tags->tag_index = tag_index; | 131 | tags->tag_index = tag_index; |
132 | tags->tag_map = tag_map; | 132 | tags->tag_map = tag_map; |
133 | 133 | ||
134 | return 0; | 134 | return 0; |
135 | fail: | 135 | fail: |
136 | kfree(tag_index); | 136 | kfree(tag_index); |
137 | return -ENOMEM; | 137 | return -ENOMEM; |
138 | } | 138 | } |
139 | 139 | ||
140 | static struct blk_queue_tag *__blk_queue_init_tags(struct request_queue *q, | 140 | static struct blk_queue_tag *__blk_queue_init_tags(struct request_queue *q, |
141 | int depth) | 141 | int depth) |
142 | { | 142 | { |
143 | struct blk_queue_tag *tags; | 143 | struct blk_queue_tag *tags; |
144 | 144 | ||
145 | tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC); | 145 | tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC); |
146 | if (!tags) | 146 | if (!tags) |
147 | goto fail; | 147 | goto fail; |
148 | 148 | ||
149 | if (init_tag_map(q, tags, depth)) | 149 | if (init_tag_map(q, tags, depth)) |
150 | goto fail; | 150 | goto fail; |
151 | 151 | ||
152 | atomic_set(&tags->refcnt, 1); | 152 | atomic_set(&tags->refcnt, 1); |
153 | return tags; | 153 | return tags; |
154 | fail: | 154 | fail: |
155 | kfree(tags); | 155 | kfree(tags); |
156 | return NULL; | 156 | return NULL; |
157 | } | 157 | } |
158 | 158 | ||
159 | /** | 159 | /** |
160 | * blk_init_tags - initialize the tag info for an external tag map | 160 | * blk_init_tags - initialize the tag info for an external tag map |
161 | * @depth: the maximum queue depth supported | 161 | * @depth: the maximum queue depth supported |
162 | **/ | 162 | **/ |
163 | struct blk_queue_tag *blk_init_tags(int depth) | 163 | struct blk_queue_tag *blk_init_tags(int depth) |
164 | { | 164 | { |
165 | return __blk_queue_init_tags(NULL, depth); | 165 | return __blk_queue_init_tags(NULL, depth); |
166 | } | 166 | } |
167 | EXPORT_SYMBOL(blk_init_tags); | 167 | EXPORT_SYMBOL(blk_init_tags); |
168 | 168 | ||
169 | /** | 169 | /** |
170 | * blk_queue_init_tags - initialize the queue tag info | 170 | * blk_queue_init_tags - initialize the queue tag info |
171 | * @q: the request queue for the device | 171 | * @q: the request queue for the device |
172 | * @depth: the maximum queue depth supported | 172 | * @depth: the maximum queue depth supported |
173 | * @tags: the tag to use | 173 | * @tags: the tag to use |
174 | * | 174 | * |
175 | * Queue lock must be held here if the function is called to resize an | 175 | * Queue lock must be held here if the function is called to resize an |
176 | * existing map. | 176 | * existing map. |
177 | **/ | 177 | **/ |
178 | int blk_queue_init_tags(struct request_queue *q, int depth, | 178 | int blk_queue_init_tags(struct request_queue *q, int depth, |
179 | struct blk_queue_tag *tags) | 179 | struct blk_queue_tag *tags) |
180 | { | 180 | { |
181 | int rc; | 181 | int rc; |
182 | 182 | ||
183 | BUG_ON(tags && q->queue_tags && tags != q->queue_tags); | 183 | BUG_ON(tags && q->queue_tags && tags != q->queue_tags); |
184 | 184 | ||
185 | if (!tags && !q->queue_tags) { | 185 | if (!tags && !q->queue_tags) { |
186 | tags = __blk_queue_init_tags(q, depth); | 186 | tags = __blk_queue_init_tags(q, depth); |
187 | 187 | ||
188 | if (!tags) | 188 | if (!tags) |
189 | return -ENOMEM; | 189 | return -ENOMEM; |
190 | 190 | ||
191 | } else if (q->queue_tags) { | 191 | } else if (q->queue_tags) { |
192 | rc = blk_queue_resize_tags(q, depth); | 192 | rc = blk_queue_resize_tags(q, depth); |
193 | if (rc) | 193 | if (rc) |
194 | return rc; | 194 | return rc; |
195 | queue_flag_set(QUEUE_FLAG_QUEUED, q); | 195 | queue_flag_set(QUEUE_FLAG_QUEUED, q); |
196 | return 0; | 196 | return 0; |
197 | } else | 197 | } else |
198 | atomic_inc(&tags->refcnt); | 198 | atomic_inc(&tags->refcnt); |
199 | 199 | ||
200 | /* | 200 | /* |
201 | * assign it, all done | 201 | * assign it, all done |
202 | */ | 202 | */ |
203 | q->queue_tags = tags; | 203 | q->queue_tags = tags; |
204 | queue_flag_set_unlocked(QUEUE_FLAG_QUEUED, q); | 204 | queue_flag_set_unlocked(QUEUE_FLAG_QUEUED, q); |
205 | INIT_LIST_HEAD(&q->tag_busy_list); | 205 | INIT_LIST_HEAD(&q->tag_busy_list); |
206 | return 0; | 206 | return 0; |
207 | } | 207 | } |
208 | EXPORT_SYMBOL(blk_queue_init_tags); | 208 | EXPORT_SYMBOL(blk_queue_init_tags); |
209 | 209 | ||
210 | /** | 210 | /** |
211 | * blk_queue_resize_tags - change the queueing depth | 211 | * blk_queue_resize_tags - change the queueing depth |
212 | * @q: the request queue for the device | 212 | * @q: the request queue for the device |
213 | * @new_depth: the new max command queueing depth | 213 | * @new_depth: the new max command queueing depth |
214 | * | 214 | * |
215 | * Notes: | 215 | * Notes: |
216 | * Must be called with the queue lock held. | 216 | * Must be called with the queue lock held. |
217 | **/ | 217 | **/ |
218 | int blk_queue_resize_tags(struct request_queue *q, int new_depth) | 218 | int blk_queue_resize_tags(struct request_queue *q, int new_depth) |
219 | { | 219 | { |
220 | struct blk_queue_tag *bqt = q->queue_tags; | 220 | struct blk_queue_tag *bqt = q->queue_tags; |
221 | struct request **tag_index; | 221 | struct request **tag_index; |
222 | unsigned long *tag_map; | 222 | unsigned long *tag_map; |
223 | int max_depth, nr_ulongs; | 223 | int max_depth, nr_ulongs; |
224 | 224 | ||
225 | if (!bqt) | 225 | if (!bqt) |
226 | return -ENXIO; | 226 | return -ENXIO; |
227 | 227 | ||
228 | /* | 228 | /* |
229 | * if we already have large enough real_max_depth. just | 229 | * if we already have large enough real_max_depth. just |
230 | * adjust max_depth. *NOTE* as requests with tag value | 230 | * adjust max_depth. *NOTE* as requests with tag value |
231 | * between new_depth and real_max_depth can be in-flight, tag | 231 | * between new_depth and real_max_depth can be in-flight, tag |
232 | * map can not be shrunk blindly here. | 232 | * map can not be shrunk blindly here. |
233 | */ | 233 | */ |
234 | if (new_depth <= bqt->real_max_depth) { | 234 | if (new_depth <= bqt->real_max_depth) { |
235 | bqt->max_depth = new_depth; | 235 | bqt->max_depth = new_depth; |
236 | return 0; | 236 | return 0; |
237 | } | 237 | } |
238 | 238 | ||
239 | /* | 239 | /* |
240 | * Currently cannot replace a shared tag map with a new | 240 | * Currently cannot replace a shared tag map with a new |
241 | * one, so error out if this is the case | 241 | * one, so error out if this is the case |
242 | */ | 242 | */ |
243 | if (atomic_read(&bqt->refcnt) != 1) | 243 | if (atomic_read(&bqt->refcnt) != 1) |
244 | return -EBUSY; | 244 | return -EBUSY; |
245 | 245 | ||
246 | /* | 246 | /* |
247 | * save the old state info, so we can copy it back | 247 | * save the old state info, so we can copy it back |
248 | */ | 248 | */ |
249 | tag_index = bqt->tag_index; | 249 | tag_index = bqt->tag_index; |
250 | tag_map = bqt->tag_map; | 250 | tag_map = bqt->tag_map; |
251 | max_depth = bqt->real_max_depth; | 251 | max_depth = bqt->real_max_depth; |
252 | 252 | ||
253 | if (init_tag_map(q, bqt, new_depth)) | 253 | if (init_tag_map(q, bqt, new_depth)) |
254 | return -ENOMEM; | 254 | return -ENOMEM; |
255 | 255 | ||
256 | memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *)); | 256 | memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *)); |
257 | nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG; | 257 | nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG; |
258 | memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long)); | 258 | memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long)); |
259 | 259 | ||
260 | kfree(tag_index); | 260 | kfree(tag_index); |
261 | kfree(tag_map); | 261 | kfree(tag_map); |
262 | return 0; | 262 | return 0; |
263 | } | 263 | } |
264 | EXPORT_SYMBOL(blk_queue_resize_tags); | 264 | EXPORT_SYMBOL(blk_queue_resize_tags); |
265 | 265 | ||
266 | /** | 266 | /** |
267 | * blk_queue_end_tag - end tag operations for a request | 267 | * blk_queue_end_tag - end tag operations for a request |
268 | * @q: the request queue for the device | 268 | * @q: the request queue for the device |
269 | * @rq: the request that has completed | 269 | * @rq: the request that has completed |
270 | * | 270 | * |
271 | * Description: | 271 | * Description: |
272 | * Typically called when end_that_request_first() returns %0, meaning | 272 | * Typically called when end_that_request_first() returns %0, meaning |
273 | * all transfers have been done for a request. It's important to call | 273 | * all transfers have been done for a request. It's important to call |
274 | * this function before end_that_request_last(), as that will put the | 274 | * this function before end_that_request_last(), as that will put the |
275 | * request back on the free list thus corrupting the internal tag list. | 275 | * request back on the free list thus corrupting the internal tag list. |
276 | * | 276 | * |
277 | * Notes: | 277 | * Notes: |
278 | * queue lock must be held. | 278 | * queue lock must be held. |
279 | **/ | 279 | **/ |
280 | void blk_queue_end_tag(struct request_queue *q, struct request *rq) | 280 | void blk_queue_end_tag(struct request_queue *q, struct request *rq) |
281 | { | 281 | { |
282 | struct blk_queue_tag *bqt = q->queue_tags; | 282 | struct blk_queue_tag *bqt = q->queue_tags; |
283 | unsigned tag = rq->tag; /* negative tags invalid */ | 283 | unsigned tag = rq->tag; /* negative tags invalid */ |
284 | 284 | ||
285 | BUG_ON(tag >= bqt->real_max_depth); | 285 | BUG_ON(tag >= bqt->real_max_depth); |
286 | 286 | ||
287 | list_del_init(&rq->queuelist); | 287 | list_del_init(&rq->queuelist); |
288 | rq->cmd_flags &= ~REQ_QUEUED; | 288 | rq->cmd_flags &= ~REQ_QUEUED; |
289 | rq->tag = -1; | 289 | rq->tag = -1; |
290 | 290 | ||
291 | if (unlikely(bqt->tag_index[tag] == NULL)) | 291 | if (unlikely(bqt->tag_index[tag] == NULL)) |
292 | printk(KERN_ERR "%s: tag %d is missing\n", | 292 | printk(KERN_ERR "%s: tag %d is missing\n", |
293 | __func__, tag); | 293 | __func__, tag); |
294 | 294 | ||
295 | bqt->tag_index[tag] = NULL; | 295 | bqt->tag_index[tag] = NULL; |
296 | 296 | ||
297 | if (unlikely(!test_bit(tag, bqt->tag_map))) { | 297 | if (unlikely(!test_bit(tag, bqt->tag_map))) { |
298 | printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n", | 298 | printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n", |
299 | __func__, tag); | 299 | __func__, tag); |
300 | return; | 300 | return; |
301 | } | 301 | } |
302 | /* | 302 | /* |
303 | * The tag_map bit acts as a lock for tag_index[bit], so we need | 303 | * The tag_map bit acts as a lock for tag_index[bit], so we need |
304 | * unlock memory barrier semantics. | 304 | * unlock memory barrier semantics. |
305 | */ | 305 | */ |
306 | clear_bit_unlock(tag, bqt->tag_map); | 306 | clear_bit_unlock(tag, bqt->tag_map); |
307 | } | 307 | } |
308 | EXPORT_SYMBOL(blk_queue_end_tag); | 308 | EXPORT_SYMBOL(blk_queue_end_tag); |
309 | 309 | ||
310 | /** | 310 | /** |
311 | * blk_queue_start_tag - find a free tag and assign it | 311 | * blk_queue_start_tag - find a free tag and assign it |
312 | * @q: the request queue for the device | 312 | * @q: the request queue for the device |
313 | * @rq: the block request that needs tagging | 313 | * @rq: the block request that needs tagging |
314 | * | 314 | * |
315 | * Description: | 315 | * Description: |
316 | * This can either be used as a stand-alone helper, or possibly be | 316 | * This can either be used as a stand-alone helper, or possibly be |
317 | * assigned as the queue &prep_rq_fn (in which case &struct request | 317 | * assigned as the queue &prep_rq_fn (in which case &struct request |
318 | * automagically gets a tag assigned). Note that this function | 318 | * automagically gets a tag assigned). Note that this function |
319 | * assumes that any type of request can be queued! if this is not | 319 | * assumes that any type of request can be queued! if this is not |
320 | * true for your device, you must check the request type before | 320 | * true for your device, you must check the request type before |
321 | * calling this function. The request will also be removed from | 321 | * calling this function. The request will also be removed from |
322 | * the request queue, so it's the drivers responsibility to readd | 322 | * the request queue, so it's the drivers responsibility to readd |
323 | * it if it should need to be restarted for some reason. | 323 | * it if it should need to be restarted for some reason. |
324 | * | 324 | * |
325 | * Notes: | 325 | * Notes: |
326 | * queue lock must be held. | 326 | * queue lock must be held. |
327 | **/ | 327 | **/ |
328 | int blk_queue_start_tag(struct request_queue *q, struct request *rq) | 328 | int blk_queue_start_tag(struct request_queue *q, struct request *rq) |
329 | { | 329 | { |
330 | struct blk_queue_tag *bqt = q->queue_tags; | 330 | struct blk_queue_tag *bqt = q->queue_tags; |
331 | unsigned max_depth; | 331 | unsigned max_depth; |
332 | int tag; | 332 | int tag; |
333 | 333 | ||
334 | if (unlikely((rq->cmd_flags & REQ_QUEUED))) { | 334 | if (unlikely((rq->cmd_flags & REQ_QUEUED))) { |
335 | printk(KERN_ERR | 335 | printk(KERN_ERR |
336 | "%s: request %p for device [%s] already tagged %d", | 336 | "%s: request %p for device [%s] already tagged %d", |
337 | __func__, rq, | 337 | __func__, rq, |
338 | rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag); | 338 | rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag); |
339 | BUG(); | 339 | BUG(); |
340 | } | 340 | } |
341 | 341 | ||
342 | /* | 342 | /* |
343 | * Protect against shared tag maps, as we may not have exclusive | 343 | * Protect against shared tag maps, as we may not have exclusive |
344 | * access to the tag map. | 344 | * access to the tag map. |
345 | * | 345 | * |
346 | * We reserve a few tags just for sync IO, since we don't want | 346 | * We reserve a few tags just for sync IO, since we don't want |
347 | * to starve sync IO on behalf of flooding async IO. | 347 | * to starve sync IO on behalf of flooding async IO. |
348 | */ | 348 | */ |
349 | max_depth = bqt->max_depth; | 349 | max_depth = bqt->max_depth; |
350 | if (!rq_is_sync(rq) && max_depth > 1) { | 350 | if (!rq_is_sync(rq) && max_depth > 1) { |
351 | max_depth -= 2; | 351 | switch (max_depth) { |
352 | if (!max_depth) | 352 | case 2: |
353 | max_depth = 1; | 353 | max_depth = 1; |
354 | break; | ||
355 | case 3: | ||
356 | max_depth = 2; | ||
357 | break; | ||
358 | default: | ||
359 | max_depth -= 2; | ||
360 | } | ||
354 | if (q->in_flight[BLK_RW_ASYNC] > max_depth) | 361 | if (q->in_flight[BLK_RW_ASYNC] > max_depth) |
355 | return 1; | 362 | return 1; |
356 | } | 363 | } |
357 | 364 | ||
358 | do { | 365 | do { |
359 | tag = find_first_zero_bit(bqt->tag_map, max_depth); | 366 | tag = find_first_zero_bit(bqt->tag_map, max_depth); |
360 | if (tag >= max_depth) | 367 | if (tag >= max_depth) |
361 | return 1; | 368 | return 1; |
362 | 369 | ||
363 | } while (test_and_set_bit_lock(tag, bqt->tag_map)); | 370 | } while (test_and_set_bit_lock(tag, bqt->tag_map)); |
364 | /* | 371 | /* |
365 | * We need lock ordering semantics given by test_and_set_bit_lock. | 372 | * We need lock ordering semantics given by test_and_set_bit_lock. |
366 | * See blk_queue_end_tag for details. | 373 | * See blk_queue_end_tag for details. |
367 | */ | 374 | */ |
368 | 375 | ||
369 | rq->cmd_flags |= REQ_QUEUED; | 376 | rq->cmd_flags |= REQ_QUEUED; |
370 | rq->tag = tag; | 377 | rq->tag = tag; |
371 | bqt->tag_index[tag] = rq; | 378 | bqt->tag_index[tag] = rq; |
372 | blk_start_request(rq); | 379 | blk_start_request(rq); |
373 | list_add(&rq->queuelist, &q->tag_busy_list); | 380 | list_add(&rq->queuelist, &q->tag_busy_list); |
374 | return 0; | 381 | return 0; |
375 | } | 382 | } |
376 | EXPORT_SYMBOL(blk_queue_start_tag); | 383 | EXPORT_SYMBOL(blk_queue_start_tag); |
377 | 384 | ||
378 | /** | 385 | /** |
379 | * blk_queue_invalidate_tags - invalidate all pending tags | 386 | * blk_queue_invalidate_tags - invalidate all pending tags |
380 | * @q: the request queue for the device | 387 | * @q: the request queue for the device |
381 | * | 388 | * |
382 | * Description: | 389 | * Description: |
383 | * Hardware conditions may dictate a need to stop all pending requests. | 390 | * Hardware conditions may dictate a need to stop all pending requests. |
384 | * In this case, we will safely clear the block side of the tag queue and | 391 | * In this case, we will safely clear the block side of the tag queue and |
385 | * readd all requests to the request queue in the right order. | 392 | * readd all requests to the request queue in the right order. |
386 | * | 393 | * |
387 | * Notes: | 394 | * Notes: |
388 | * queue lock must be held. | 395 | * queue lock must be held. |
389 | **/ | 396 | **/ |
390 | void blk_queue_invalidate_tags(struct request_queue *q) | 397 | void blk_queue_invalidate_tags(struct request_queue *q) |
391 | { | 398 | { |
392 | struct list_head *tmp, *n; | 399 | struct list_head *tmp, *n; |
393 | 400 | ||
394 | list_for_each_safe(tmp, n, &q->tag_busy_list) | 401 | list_for_each_safe(tmp, n, &q->tag_busy_list) |
395 | blk_requeue_request(q, list_entry_rq(tmp)); | 402 | blk_requeue_request(q, list_entry_rq(tmp)); |
396 | } | 403 | } |
397 | EXPORT_SYMBOL(blk_queue_invalidate_tags); | 404 | EXPORT_SYMBOL(blk_queue_invalidate_tags); |
398 | 405 |