]>
Commit | Line | Data |
---|---|---|
75bb4625 | 1 | /* |
88459642 OS |
2 | * Tag allocation using scalable bitmaps. Uses active queue tracking to support |
3 | * fairer distribution of tags between multiple submitters when a shared tag map | |
4 | * is used. | |
75bb4625 JA |
5 | * |
6 | * Copyright (C) 2013-2014 Jens Axboe | |
7 | */ | |
320ae51f JA |
8 | #include <linux/kernel.h> |
9 | #include <linux/module.h> | |
320ae51f JA |
10 | |
11 | #include <linux/blk-mq.h> | |
12 | #include "blk.h" | |
13 | #include "blk-mq.h" | |
14 | #include "blk-mq-tag.h" | |
15 | ||
320ae51f JA |
16 | bool blk_mq_has_free_tags(struct blk_mq_tags *tags) |
17 | { | |
4bb659b1 JA |
18 | if (!tags) |
19 | return true; | |
20 | ||
88459642 | 21 | return sbitmap_any_bit_clear(&tags->bitmap_tags.sb); |
0d2602ca JA |
22 | } |
23 | ||
24 | /* | |
25 | * If a previously inactive queue goes active, bump the active user count. | |
d263ed99 JW |
26 | * We need to do this before try to allocate driver tag, then even if fail |
27 | * to get tag when first time, the other shared-tag users could reserve | |
28 | * budget for it. | |
0d2602ca JA |
29 | */ |
30 | bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx) | |
31 | { | |
32 | if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) && | |
33 | !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) | |
34 | atomic_inc(&hctx->tags->active_queues); | |
35 | ||
36 | return true; | |
37 | } | |
38 | ||
39 | /* | |
aed3ea94 | 40 | * Wakeup all potentially sleeping on tags |
0d2602ca | 41 | */ |
aed3ea94 | 42 | void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve) |
0d2602ca | 43 | { |
88459642 OS |
44 | sbitmap_queue_wake_all(&tags->bitmap_tags); |
45 | if (include_reserve) | |
46 | sbitmap_queue_wake_all(&tags->breserved_tags); | |
0d2602ca JA |
47 | } |
48 | ||
e3a2b3f9 JA |
49 | /* |
50 | * If a previously busy queue goes inactive, potential waiters could now | |
51 | * be allowed to queue. Wake them up and check. | |
52 | */ | |
53 | void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx) | |
54 | { | |
55 | struct blk_mq_tags *tags = hctx->tags; | |
56 | ||
57 | if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) | |
58 | return; | |
59 | ||
60 | atomic_dec(&tags->active_queues); | |
61 | ||
aed3ea94 | 62 | blk_mq_tag_wakeup_all(tags, false); |
e3a2b3f9 JA |
63 | } |
64 | ||
0d2602ca JA |
65 | /* |
66 | * For shared tag users, we track the number of currently active users | |
67 | * and attempt to provide a fair share of the tag depth for each of them. | |
68 | */ | |
69 | static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx, | |
88459642 | 70 | struct sbitmap_queue *bt) |
0d2602ca JA |
71 | { |
72 | unsigned int depth, users; | |
73 | ||
74 | if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED)) | |
75 | return true; | |
76 | if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) | |
77 | return true; | |
78 | ||
79 | /* | |
80 | * Don't try dividing an ant | |
81 | */ | |
88459642 | 82 | if (bt->sb.depth == 1) |
0d2602ca JA |
83 | return true; |
84 | ||
85 | users = atomic_read(&hctx->tags->active_queues); | |
86 | if (!users) | |
87 | return true; | |
88 | ||
89 | /* | |
90 | * Allow at least some tags | |
91 | */ | |
88459642 | 92 | depth = max((bt->sb.depth + users - 1) / users, 4U); |
0d2602ca JA |
93 | return atomic_read(&hctx->nr_active) < depth; |
94 | } | |
95 | ||
200e86b3 JA |
96 | static int __blk_mq_get_tag(struct blk_mq_alloc_data *data, |
97 | struct sbitmap_queue *bt) | |
4bb659b1 | 98 | { |
200e86b3 JA |
99 | if (!(data->flags & BLK_MQ_REQ_INTERNAL) && |
100 | !hctx_may_queue(data->hctx, bt)) | |
0d2602ca | 101 | return -1; |
229a9287 OS |
102 | if (data->shallow_depth) |
103 | return __sbitmap_queue_get_shallow(bt, data->shallow_depth); | |
104 | else | |
105 | return __sbitmap_queue_get(bt); | |
4bb659b1 JA |
106 | } |
107 | ||
4941115b | 108 | unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data) |
320ae51f | 109 | { |
4941115b JA |
110 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
111 | struct sbitmap_queue *bt; | |
88459642 | 112 | struct sbq_wait_state *ws; |
4bb659b1 | 113 | DEFINE_WAIT(wait); |
4941115b | 114 | unsigned int tag_offset; |
bd6737f1 | 115 | bool drop_ctx; |
320ae51f JA |
116 | int tag; |
117 | ||
4941115b JA |
118 | if (data->flags & BLK_MQ_REQ_RESERVED) { |
119 | if (unlikely(!tags->nr_reserved_tags)) { | |
120 | WARN_ON_ONCE(1); | |
121 | return BLK_MQ_TAG_FAIL; | |
122 | } | |
123 | bt = &tags->breserved_tags; | |
124 | tag_offset = 0; | |
125 | } else { | |
126 | bt = &tags->bitmap_tags; | |
127 | tag_offset = tags->nr_reserved_tags; | |
128 | } | |
129 | ||
200e86b3 | 130 | tag = __blk_mq_get_tag(data, bt); |
4bb659b1 | 131 | if (tag != -1) |
4941115b | 132 | goto found_tag; |
4bb659b1 | 133 | |
6f3b0e8b | 134 | if (data->flags & BLK_MQ_REQ_NOWAIT) |
4941115b | 135 | return BLK_MQ_TAG_FAIL; |
4bb659b1 | 136 | |
4941115b | 137 | ws = bt_wait_ptr(bt, data->hctx); |
bd6737f1 | 138 | drop_ctx = data->ctx == NULL; |
4bb659b1 | 139 | do { |
e6fc4649 ML |
140 | struct sbitmap_queue *bt_prev; |
141 | ||
b3223207 BVA |
142 | /* |
143 | * We're out of tags on this hardware queue, kick any | |
144 | * pending IO submits before going to sleep waiting for | |
8cecb07d | 145 | * some to complete. |
b3223207 | 146 | */ |
8cecb07d | 147 | blk_mq_run_hw_queue(data->hctx, false); |
b3223207 | 148 | |
080ff351 JA |
149 | /* |
150 | * Retry tag allocation after running the hardware queue, | |
151 | * as running the queue may also have found completions. | |
152 | */ | |
200e86b3 | 153 | tag = __blk_mq_get_tag(data, bt); |
080ff351 JA |
154 | if (tag != -1) |
155 | break; | |
156 | ||
4e5dff41 JA |
157 | prepare_to_wait_exclusive(&ws->wait, &wait, |
158 | TASK_UNINTERRUPTIBLE); | |
159 | ||
160 | tag = __blk_mq_get_tag(data, bt); | |
161 | if (tag != -1) | |
162 | break; | |
163 | ||
bd6737f1 JA |
164 | if (data->ctx) |
165 | blk_mq_put_ctx(data->ctx); | |
cb96a42c | 166 | |
e6fc4649 | 167 | bt_prev = bt; |
4bb659b1 | 168 | io_schedule(); |
cb96a42c ML |
169 | |
170 | data->ctx = blk_mq_get_ctx(data->q); | |
7d7e0f90 | 171 | data->hctx = blk_mq_map_queue(data->q, data->ctx->cpu); |
4941115b JA |
172 | tags = blk_mq_tags_from_data(data); |
173 | if (data->flags & BLK_MQ_REQ_RESERVED) | |
174 | bt = &tags->breserved_tags; | |
175 | else | |
176 | bt = &tags->bitmap_tags; | |
177 | ||
88459642 | 178 | finish_wait(&ws->wait, &wait); |
e6fc4649 ML |
179 | |
180 | /* | |
181 | * If destination hw queue is changed, fake wake up on | |
182 | * previous queue for compensating the wake up miss, so | |
183 | * other allocations on previous queue won't be starved. | |
184 | */ | |
185 | if (bt != bt_prev) | |
186 | sbitmap_queue_wake_up(bt_prev); | |
187 | ||
4941115b | 188 | ws = bt_wait_ptr(bt, data->hctx); |
4bb659b1 JA |
189 | } while (1); |
190 | ||
bd6737f1 JA |
191 | if (drop_ctx && data->ctx) |
192 | blk_mq_put_ctx(data->ctx); | |
193 | ||
88459642 | 194 | finish_wait(&ws->wait, &wait); |
320ae51f | 195 | |
4941115b JA |
196 | found_tag: |
197 | return tag + tag_offset; | |
320ae51f JA |
198 | } |
199 | ||
4941115b JA |
200 | void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, struct blk_mq_tags *tags, |
201 | struct blk_mq_ctx *ctx, unsigned int tag) | |
320ae51f | 202 | { |
415b806d | 203 | if (!blk_mq_tag_is_reserved(tags, tag)) { |
4bb659b1 JA |
204 | const int real_tag = tag - tags->nr_reserved_tags; |
205 | ||
70114c39 | 206 | BUG_ON(real_tag >= tags->nr_tags); |
f4a644db | 207 | sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu); |
70114c39 JA |
208 | } else { |
209 | BUG_ON(tag >= tags->nr_reserved_tags); | |
f4a644db | 210 | sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu); |
70114c39 | 211 | } |
320ae51f JA |
212 | } |
213 | ||
88459642 OS |
214 | struct bt_iter_data { |
215 | struct blk_mq_hw_ctx *hctx; | |
216 | busy_iter_fn *fn; | |
217 | void *data; | |
218 | bool reserved; | |
219 | }; | |
220 | ||
221 | static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data) | |
320ae51f | 222 | { |
88459642 OS |
223 | struct bt_iter_data *iter_data = data; |
224 | struct blk_mq_hw_ctx *hctx = iter_data->hctx; | |
225 | struct blk_mq_tags *tags = hctx->tags; | |
226 | bool reserved = iter_data->reserved; | |
81481eb4 | 227 | struct request *rq; |
4bb659b1 | 228 | |
88459642 OS |
229 | if (!reserved) |
230 | bitnr += tags->nr_reserved_tags; | |
231 | rq = tags->rqs[bitnr]; | |
4bb659b1 | 232 | |
7f5562d5 JA |
233 | /* |
234 | * We can hit rq == NULL here, because the tagging functions | |
235 | * test and set the bit before assining ->rqs[]. | |
236 | */ | |
237 | if (rq && rq->q == hctx->queue) | |
88459642 OS |
238 | iter_data->fn(hctx, rq, iter_data->data, reserved); |
239 | return true; | |
240 | } | |
4bb659b1 | 241 | |
88459642 OS |
242 | static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt, |
243 | busy_iter_fn *fn, void *data, bool reserved) | |
244 | { | |
245 | struct bt_iter_data iter_data = { | |
246 | .hctx = hctx, | |
247 | .fn = fn, | |
248 | .data = data, | |
249 | .reserved = reserved, | |
250 | }; | |
251 | ||
252 | sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data); | |
320ae51f JA |
253 | } |
254 | ||
88459642 OS |
255 | struct bt_tags_iter_data { |
256 | struct blk_mq_tags *tags; | |
257 | busy_tag_iter_fn *fn; | |
258 | void *data; | |
259 | bool reserved; | |
260 | }; | |
261 | ||
262 | static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data) | |
f26cdc85 | 263 | { |
88459642 OS |
264 | struct bt_tags_iter_data *iter_data = data; |
265 | struct blk_mq_tags *tags = iter_data->tags; | |
266 | bool reserved = iter_data->reserved; | |
f26cdc85 | 267 | struct request *rq; |
f26cdc85 | 268 | |
88459642 OS |
269 | if (!reserved) |
270 | bitnr += tags->nr_reserved_tags; | |
7f5562d5 JA |
271 | |
272 | /* | |
273 | * We can hit rq == NULL here, because the tagging functions | |
274 | * test and set the bit before assining ->rqs[]. | |
275 | */ | |
88459642 | 276 | rq = tags->rqs[bitnr]; |
2d5ba0e2 | 277 | if (rq && blk_mq_request_started(rq)) |
7f5562d5 | 278 | iter_data->fn(rq, iter_data->data, reserved); |
f26cdc85 | 279 | |
88459642 OS |
280 | return true; |
281 | } | |
282 | ||
283 | static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt, | |
284 | busy_tag_iter_fn *fn, void *data, bool reserved) | |
285 | { | |
286 | struct bt_tags_iter_data iter_data = { | |
287 | .tags = tags, | |
288 | .fn = fn, | |
289 | .data = data, | |
290 | .reserved = reserved, | |
291 | }; | |
292 | ||
293 | if (tags->rqs) | |
294 | sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data); | |
f26cdc85 KB |
295 | } |
296 | ||
e8f1e163 SG |
297 | static void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags, |
298 | busy_tag_iter_fn *fn, void *priv) | |
f26cdc85 KB |
299 | { |
300 | if (tags->nr_reserved_tags) | |
88459642 OS |
301 | bt_tags_for_each(tags, &tags->breserved_tags, fn, priv, true); |
302 | bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, false); | |
f26cdc85 | 303 | } |
f26cdc85 | 304 | |
e0489487 SG |
305 | void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset, |
306 | busy_tag_iter_fn *fn, void *priv) | |
307 | { | |
308 | int i; | |
309 | ||
310 | for (i = 0; i < tagset->nr_hw_queues; i++) { | |
311 | if (tagset->tags && tagset->tags[i]) | |
312 | blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv); | |
313 | } | |
314 | } | |
315 | EXPORT_SYMBOL(blk_mq_tagset_busy_iter); | |
316 | ||
0bf6cd5b | 317 | void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn, |
81481eb4 | 318 | void *priv) |
320ae51f | 319 | { |
0bf6cd5b CH |
320 | struct blk_mq_hw_ctx *hctx; |
321 | int i; | |
322 | ||
f5bbbbe4 JW |
323 | /* |
324 | * __blk_mq_update_nr_hw_queues will update the nr_hw_queues and | |
325 | * queue_hw_ctx after freeze the queue. So we could use q_usage_counter | |
326 | * to avoid race with it. __blk_mq_update_nr_hw_queues will users | |
327 | * synchronize_rcu to ensure all of the users go out of the critical | |
328 | * section below and see zeroed q_usage_counter. | |
329 | */ | |
330 | rcu_read_lock(); | |
331 | if (percpu_ref_is_zero(&q->q_usage_counter)) { | |
332 | rcu_read_unlock(); | |
333 | return; | |
334 | } | |
0bf6cd5b CH |
335 | |
336 | queue_for_each_hw_ctx(q, hctx, i) { | |
337 | struct blk_mq_tags *tags = hctx->tags; | |
338 | ||
339 | /* | |
340 | * If not software queues are currently mapped to this | |
341 | * hardware queue, there's nothing to check | |
342 | */ | |
343 | if (!blk_mq_hw_queue_mapped(hctx)) | |
344 | continue; | |
345 | ||
346 | if (tags->nr_reserved_tags) | |
88459642 OS |
347 | bt_for_each(hctx, &tags->breserved_tags, fn, priv, true); |
348 | bt_for_each(hctx, &tags->bitmap_tags, fn, priv, false); | |
4bb659b1 | 349 | } |
f5bbbbe4 | 350 | rcu_read_unlock(); |
4bb659b1 JA |
351 | } |
352 | ||
f4a644db OS |
353 | static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth, |
354 | bool round_robin, int node) | |
4bb659b1 | 355 | { |
f4a644db OS |
356 | return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL, |
357 | node); | |
4bb659b1 JA |
358 | } |
359 | ||
360 | static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags, | |
24391c0d | 361 | int node, int alloc_policy) |
4bb659b1 JA |
362 | { |
363 | unsigned int depth = tags->nr_tags - tags->nr_reserved_tags; | |
f4a644db | 364 | bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR; |
4bb659b1 | 365 | |
f4a644db | 366 | if (bt_alloc(&tags->bitmap_tags, depth, round_robin, node)) |
88459642 | 367 | goto free_tags; |
f4a644db OS |
368 | if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, round_robin, |
369 | node)) | |
88459642 | 370 | goto free_bitmap_tags; |
4bb659b1 JA |
371 | |
372 | return tags; | |
88459642 OS |
373 | free_bitmap_tags: |
374 | sbitmap_queue_free(&tags->bitmap_tags); | |
375 | free_tags: | |
4bb659b1 JA |
376 | kfree(tags); |
377 | return NULL; | |
378 | } | |
379 | ||
320ae51f | 380 | struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags, |
24391c0d SL |
381 | unsigned int reserved_tags, |
382 | int node, int alloc_policy) | |
320ae51f | 383 | { |
320ae51f | 384 | struct blk_mq_tags *tags; |
320ae51f JA |
385 | |
386 | if (total_tags > BLK_MQ_TAG_MAX) { | |
387 | pr_err("blk-mq: tag depth too large\n"); | |
388 | return NULL; | |
389 | } | |
390 | ||
391 | tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node); | |
392 | if (!tags) | |
393 | return NULL; | |
394 | ||
320ae51f JA |
395 | tags->nr_tags = total_tags; |
396 | tags->nr_reserved_tags = reserved_tags; | |
320ae51f | 397 | |
24391c0d | 398 | return blk_mq_init_bitmap_tags(tags, node, alloc_policy); |
320ae51f JA |
399 | } |
400 | ||
401 | void blk_mq_free_tags(struct blk_mq_tags *tags) | |
402 | { | |
88459642 OS |
403 | sbitmap_queue_free(&tags->bitmap_tags); |
404 | sbitmap_queue_free(&tags->breserved_tags); | |
320ae51f JA |
405 | kfree(tags); |
406 | } | |
407 | ||
70f36b60 JA |
408 | int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx, |
409 | struct blk_mq_tags **tagsptr, unsigned int tdepth, | |
410 | bool can_grow) | |
e3a2b3f9 | 411 | { |
70f36b60 JA |
412 | struct blk_mq_tags *tags = *tagsptr; |
413 | ||
414 | if (tdepth <= tags->nr_reserved_tags) | |
e3a2b3f9 JA |
415 | return -EINVAL; |
416 | ||
417 | /* | |
70f36b60 JA |
418 | * If we are allowed to grow beyond the original size, allocate |
419 | * a new set of tags before freeing the old one. | |
e3a2b3f9 | 420 | */ |
70f36b60 JA |
421 | if (tdepth > tags->nr_tags) { |
422 | struct blk_mq_tag_set *set = hctx->queue->tag_set; | |
423 | struct blk_mq_tags *new; | |
424 | bool ret; | |
425 | ||
426 | if (!can_grow) | |
427 | return -EINVAL; | |
428 | ||
429 | /* | |
430 | * We need some sort of upper limit, set it high enough that | |
431 | * no valid use cases should require more. | |
432 | */ | |
433 | if (tdepth > 16 * BLKDEV_MAX_RQ) | |
434 | return -EINVAL; | |
435 | ||
75d6e175 ML |
436 | new = blk_mq_alloc_rq_map(set, hctx->queue_num, tdepth, |
437 | tags->nr_reserved_tags); | |
70f36b60 JA |
438 | if (!new) |
439 | return -ENOMEM; | |
440 | ret = blk_mq_alloc_rqs(set, new, hctx->queue_num, tdepth); | |
441 | if (ret) { | |
442 | blk_mq_free_rq_map(new); | |
443 | return -ENOMEM; | |
444 | } | |
445 | ||
446 | blk_mq_free_rqs(set, *tagsptr, hctx->queue_num); | |
447 | blk_mq_free_rq_map(*tagsptr); | |
448 | *tagsptr = new; | |
449 | } else { | |
450 | /* | |
451 | * Don't need (or can't) update reserved tags here, they | |
452 | * remain static and should never need resizing. | |
453 | */ | |
75d6e175 ML |
454 | sbitmap_queue_resize(&tags->bitmap_tags, |
455 | tdepth - tags->nr_reserved_tags); | |
70f36b60 | 456 | } |
88459642 | 457 | |
e3a2b3f9 JA |
458 | return 0; |
459 | } | |
460 | ||
205fb5f5 BVA |
461 | /** |
462 | * blk_mq_unique_tag() - return a tag that is unique queue-wide | |
463 | * @rq: request for which to compute a unique tag | |
464 | * | |
465 | * The tag field in struct request is unique per hardware queue but not over | |
466 | * all hardware queues. Hence this function that returns a tag with the | |
467 | * hardware context index in the upper bits and the per hardware queue tag in | |
468 | * the lower bits. | |
469 | * | |
470 | * Note: When called for a request that is queued on a non-multiqueue request | |
471 | * queue, the hardware context index is set to zero. | |
472 | */ | |
473 | u32 blk_mq_unique_tag(struct request *rq) | |
474 | { | |
475 | struct request_queue *q = rq->q; | |
476 | struct blk_mq_hw_ctx *hctx; | |
477 | int hwq = 0; | |
478 | ||
479 | if (q->mq_ops) { | |
7d7e0f90 | 480 | hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu); |
205fb5f5 BVA |
481 | hwq = hctx->queue_num; |
482 | } | |
483 | ||
484 | return (hwq << BLK_MQ_UNIQUE_TAG_BITS) | | |
485 | (rq->tag & BLK_MQ_UNIQUE_TAG_MASK); | |
486 | } | |
487 | EXPORT_SYMBOL(blk_mq_unique_tag); |