Commit | Line | Data |
---|---|---|
3dcf60bc | 1 | // SPDX-License-Identifier: GPL-2.0 |
bd166ef1 JA |
2 | /* |
3 | * blk-mq scheduling framework | |
4 | * | |
5 | * Copyright (C) 2016 Jens Axboe | |
6 | */ | |
7 | #include <linux/kernel.h> | |
8 | #include <linux/module.h> | |
9 | #include <linux/blk-mq.h> | |
6e6fcbc2 | 10 | #include <linux/list_sort.h> |
bd166ef1 JA |
11 | |
12 | #include <trace/events/block.h> | |
13 | ||
14 | #include "blk.h" | |
15 | #include "blk-mq.h" | |
d332ce09 | 16 | #include "blk-mq-debugfs.h" |
bd166ef1 JA |
17 | #include "blk-mq-sched.h" |
18 | #include "blk-mq-tag.h" | |
19 | #include "blk-wbt.h" | |
20 | ||
e2b3fa5a | 21 | void blk_mq_sched_assign_ioc(struct request *rq) |
bd166ef1 | 22 | { |
44e8c2bf | 23 | struct request_queue *q = rq->q; |
0c62bff1 | 24 | struct io_context *ioc; |
bd166ef1 JA |
25 | struct io_cq *icq; |
26 | ||
0c62bff1 JA |
27 | /* |
28 | * May not have an IO context if it's a passthrough request | |
29 | */ | |
30 | ioc = current->io_context; | |
31 | if (!ioc) | |
32 | return; | |
33 | ||
0d945c1f | 34 | spin_lock_irq(&q->queue_lock); |
bd166ef1 | 35 | icq = ioc_lookup_icq(ioc, q); |
0d945c1f | 36 | spin_unlock_irq(&q->queue_lock); |
bd166ef1 JA |
37 | |
38 | if (!icq) { | |
39 | icq = ioc_create_icq(ioc, q, GFP_ATOMIC); | |
40 | if (!icq) | |
41 | return; | |
42 | } | |
ea511e3c | 43 | get_io_context(icq->ioc); |
44e8c2bf | 44 | rq->elv.icq = icq; |
bd166ef1 JA |
45 | } |
46 | ||
8e8320c9 JA |
47 | /* |
48 | * Mark a hardware queue as needing a restart. For shared queues, maintain | |
49 | * a count of how many hardware queues are marked for restart. | |
50 | */ | |
7211aef8 | 51 | void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx) |
8e8320c9 JA |
52 | { |
53 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
54 | return; | |
55 | ||
97889f9a | 56 | set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state); |
8e8320c9 | 57 | } |
7211aef8 | 58 | EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx); |
8e8320c9 | 59 | |
e9ea1596 | 60 | void __blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx) |
8e8320c9 | 61 | { |
97889f9a | 62 | clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state); |
8e8320c9 | 63 | |
d7d8535f ML |
64 | /* |
65 | * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch) | |
66 | * in blk_mq_run_hw_queue(). Its pair is the barrier in | |
67 | * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART, | |
68 | * meantime new request added to hctx->dispatch is missed to check in | |
69 | * blk_mq_run_hw_queue(). | |
70 | */ | |
71 | smp_mb(); | |
72 | ||
97889f9a | 73 | blk_mq_run_hw_queue(hctx, true); |
8e8320c9 JA |
74 | } |
75 | ||
4f0f586b ST |
76 | static int sched_rq_cmp(void *priv, const struct list_head *a, |
77 | const struct list_head *b) | |
6e6fcbc2 ML |
78 | { |
79 | struct request *rqa = container_of(a, struct request, queuelist); | |
80 | struct request *rqb = container_of(b, struct request, queuelist); | |
81 | ||
82 | return rqa->mq_hctx > rqb->mq_hctx; | |
83 | } | |
84 | ||
85 | static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list) | |
86 | { | |
87 | struct blk_mq_hw_ctx *hctx = | |
88 | list_first_entry(rq_list, struct request, queuelist)->mq_hctx; | |
89 | struct request *rq; | |
90 | LIST_HEAD(hctx_list); | |
91 | unsigned int count = 0; | |
6e6fcbc2 ML |
92 | |
93 | list_for_each_entry(rq, rq_list, queuelist) { | |
94 | if (rq->mq_hctx != hctx) { | |
95 | list_cut_before(&hctx_list, rq_list, &rq->queuelist); | |
96 | goto dispatch; | |
97 | } | |
98 | count++; | |
99 | } | |
100 | list_splice_tail_init(rq_list, &hctx_list); | |
101 | ||
102 | dispatch: | |
106e71c5 | 103 | return blk_mq_dispatch_rq_list(hctx, &hctx_list, count); |
6e6fcbc2 ML |
104 | } |
105 | ||
a0823421 DA |
106 | #define BLK_MQ_BUDGET_DELAY 3 /* ms units */ |
107 | ||
1f460b63 ML |
108 | /* |
109 | * Only SCSI implements .get_budget and .put_budget, and SCSI restarts | |
110 | * its queue by itself in its completion handler, so we don't need to | |
111 | * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE. | |
28d65729 SQ |
112 | * |
113 | * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to | |
114 | * be run again. This is necessary to avoid starving flushes. | |
1f460b63 | 115 | */ |
6e6fcbc2 | 116 | static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) |
caf8eb0d ML |
117 | { |
118 | struct request_queue *q = hctx->queue; | |
119 | struct elevator_queue *e = q->elevator; | |
6e6fcbc2 ML |
120 | bool multi_hctxs = false, run_queue = false; |
121 | bool dispatched = false, busy = false; | |
122 | unsigned int max_dispatch; | |
caf8eb0d | 123 | LIST_HEAD(rq_list); |
6e6fcbc2 ML |
124 | int count = 0; |
125 | ||
126 | if (hctx->dispatch_busy) | |
127 | max_dispatch = 1; | |
128 | else | |
129 | max_dispatch = hctx->queue->nr_requests; | |
caf8eb0d ML |
130 | |
131 | do { | |
6e6fcbc2 | 132 | struct request *rq; |
2a5a24aa | 133 | int budget_token; |
6e6fcbc2 | 134 | |
f9cd4bfe | 135 | if (e->type->ops.has_work && !e->type->ops.has_work(hctx)) |
caf8eb0d | 136 | break; |
de148297 | 137 | |
28d65729 | 138 | if (!list_empty_careful(&hctx->dispatch)) { |
6e6fcbc2 | 139 | busy = true; |
28d65729 SQ |
140 | break; |
141 | } | |
142 | ||
2a5a24aa ML |
143 | budget_token = blk_mq_get_dispatch_budget(q); |
144 | if (budget_token < 0) | |
1f460b63 | 145 | break; |
de148297 | 146 | |
f9cd4bfe | 147 | rq = e->type->ops.dispatch_request(hctx); |
de148297 | 148 | if (!rq) { |
2a5a24aa | 149 | blk_mq_put_dispatch_budget(q, budget_token); |
a0823421 DA |
150 | /* |
151 | * We're releasing without dispatching. Holding the | |
152 | * budget could have blocked any "hctx"s with the | |
153 | * same queue and if we didn't dispatch then there's | |
154 | * no guarantee anyone will kick the queue. Kick it | |
155 | * ourselves. | |
156 | */ | |
6e6fcbc2 | 157 | run_queue = true; |
de148297 | 158 | break; |
de148297 ML |
159 | } |
160 | ||
2a5a24aa ML |
161 | blk_mq_set_rq_budget_token(rq, budget_token); |
162 | ||
de148297 ML |
163 | /* |
164 | * Now this rq owns the budget which has to be released | |
165 | * if this rq won't be queued to driver via .queue_rq() | |
166 | * in blk_mq_dispatch_rq_list(). | |
167 | */ | |
6e6fcbc2 | 168 | list_add_tail(&rq->queuelist, &rq_list); |
61347154 | 169 | count++; |
6e6fcbc2 ML |
170 | if (rq->mq_hctx != hctx) |
171 | multi_hctxs = true; | |
61347154 JK |
172 | |
173 | /* | |
174 | * If we cannot get tag for the request, stop dequeueing | |
175 | * requests from the IO scheduler. We are unlikely to be able | |
176 | * to submit them anyway and it creates false impression for | |
177 | * scheduling heuristics that the device can take more IO. | |
178 | */ | |
179 | if (!blk_mq_get_driver_tag(rq)) | |
180 | break; | |
181 | } while (count < max_dispatch); | |
6e6fcbc2 ML |
182 | |
183 | if (!count) { | |
184 | if (run_queue) | |
185 | blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY); | |
186 | } else if (multi_hctxs) { | |
187 | /* | |
188 | * Requests from different hctx may be dequeued from some | |
189 | * schedulers, such as bfq and deadline. | |
190 | * | |
191 | * Sort the requests in the list according to their hctx, | |
192 | * dispatch batching requests from same hctx at a time. | |
193 | */ | |
194 | list_sort(NULL, &rq_list, sched_rq_cmp); | |
195 | do { | |
196 | dispatched |= blk_mq_dispatch_hctx_list(&rq_list); | |
197 | } while (!list_empty(&rq_list)); | |
198 | } else { | |
199 | dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, count); | |
200 | } | |
201 | ||
202 | if (busy) | |
203 | return -EAGAIN; | |
204 | return !!dispatched; | |
205 | } | |
206 | ||
207 | static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx) | |
208 | { | |
209 | int ret; | |
210 | ||
211 | do { | |
212 | ret = __blk_mq_do_dispatch_sched(hctx); | |
213 | } while (ret == 1); | |
28d65729 SQ |
214 | |
215 | return ret; | |
caf8eb0d ML |
216 | } |
217 | ||
b347689f ML |
218 | static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx, |
219 | struct blk_mq_ctx *ctx) | |
220 | { | |
f31967f0 | 221 | unsigned short idx = ctx->index_hw[hctx->type]; |
b347689f ML |
222 | |
223 | if (++idx == hctx->nr_ctx) | |
224 | idx = 0; | |
225 | ||
226 | return hctx->ctxs[idx]; | |
227 | } | |
228 | ||
1f460b63 ML |
229 | /* |
230 | * Only SCSI implements .get_budget and .put_budget, and SCSI restarts | |
231 | * its queue by itself in its completion handler, so we don't need to | |
232 | * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE. | |
28d65729 SQ |
233 | * |
234 | * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to | |
c4aecaa2 | 235 | * be run again. This is necessary to avoid starving flushes. |
1f460b63 | 236 | */ |
28d65729 | 237 | static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx) |
b347689f ML |
238 | { |
239 | struct request_queue *q = hctx->queue; | |
240 | LIST_HEAD(rq_list); | |
241 | struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from); | |
28d65729 | 242 | int ret = 0; |
445874e8 | 243 | struct request *rq; |
b347689f ML |
244 | |
245 | do { | |
2a5a24aa ML |
246 | int budget_token; |
247 | ||
28d65729 SQ |
248 | if (!list_empty_careful(&hctx->dispatch)) { |
249 | ret = -EAGAIN; | |
250 | break; | |
251 | } | |
252 | ||
b347689f ML |
253 | if (!sbitmap_any_bit_set(&hctx->ctx_map)) |
254 | break; | |
255 | ||
2a5a24aa ML |
256 | budget_token = blk_mq_get_dispatch_budget(q); |
257 | if (budget_token < 0) | |
1f460b63 | 258 | break; |
b347689f ML |
259 | |
260 | rq = blk_mq_dequeue_from_ctx(hctx, ctx); | |
261 | if (!rq) { | |
2a5a24aa | 262 | blk_mq_put_dispatch_budget(q, budget_token); |
a0823421 DA |
263 | /* |
264 | * We're releasing without dispatching. Holding the | |
265 | * budget could have blocked any "hctx"s with the | |
266 | * same queue and if we didn't dispatch then there's | |
267 | * no guarantee anyone will kick the queue. Kick it | |
268 | * ourselves. | |
269 | */ | |
270 | blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY); | |
b347689f | 271 | break; |
b347689f ML |
272 | } |
273 | ||
2a5a24aa ML |
274 | blk_mq_set_rq_budget_token(rq, budget_token); |
275 | ||
b347689f ML |
276 | /* |
277 | * Now this rq owns the budget which has to be released | |
278 | * if this rq won't be queued to driver via .queue_rq() | |
279 | * in blk_mq_dispatch_rq_list(). | |
280 | */ | |
281 | list_add(&rq->queuelist, &rq_list); | |
282 | ||
283 | /* round robin for fair dispatch */ | |
284 | ctx = blk_mq_next_ctx(hctx, rq->mq_ctx); | |
285 | ||
1fd40b5e | 286 | } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1)); |
b347689f ML |
287 | |
288 | WRITE_ONCE(hctx->dispatch_from, ctx); | |
28d65729 | 289 | return ret; |
b347689f ML |
290 | } |
291 | ||
e1b586f2 | 292 | static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx) |
bd166ef1 | 293 | { |
81380ca1 | 294 | struct request_queue *q = hctx->queue; |
e42cfb1d | 295 | const bool has_sched = q->elevator; |
28d65729 | 296 | int ret = 0; |
bd166ef1 JA |
297 | LIST_HEAD(rq_list); |
298 | ||
bd166ef1 JA |
299 | /* |
300 | * If we have previous entries on our dispatch list, grab them first for | |
301 | * more fair dispatch. | |
302 | */ | |
303 | if (!list_empty_careful(&hctx->dispatch)) { | |
304 | spin_lock(&hctx->lock); | |
305 | if (!list_empty(&hctx->dispatch)) | |
306 | list_splice_init(&hctx->dispatch, &rq_list); | |
307 | spin_unlock(&hctx->lock); | |
308 | } | |
309 | ||
310 | /* | |
311 | * Only ask the scheduler for requests, if we didn't have residual | |
312 | * requests from the dispatch list. This is to avoid the case where | |
313 | * we only ever dispatch a fraction of the requests available because | |
314 | * of low device queue depth. Once we pull requests out of the IO | |
315 | * scheduler, we can no longer merge or sort them. So it's best to | |
316 | * leave them there for as long as we can. Mark the hw queue as | |
317 | * needing a restart in that case. | |
caf8eb0d ML |
318 | * |
319 | * We want to dispatch from the scheduler if there was nothing | |
320 | * on the dispatch list or we were able to dispatch from the | |
321 | * dispatch list. | |
bd166ef1 | 322 | */ |
c13660a0 | 323 | if (!list_empty(&rq_list)) { |
d38d3515 | 324 | blk_mq_sched_mark_restart_hctx(hctx); |
1fd40b5e | 325 | if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) { |
e42cfb1d | 326 | if (has_sched) |
28d65729 | 327 | ret = blk_mq_do_dispatch_sched(hctx); |
b347689f | 328 | else |
28d65729 | 329 | ret = blk_mq_do_dispatch_ctx(hctx); |
b347689f | 330 | } |
e42cfb1d | 331 | } else if (has_sched) { |
28d65729 | 332 | ret = blk_mq_do_dispatch_sched(hctx); |
6e768717 ML |
333 | } else if (hctx->dispatch_busy) { |
334 | /* dequeue request one by one from sw queue if queue is busy */ | |
28d65729 | 335 | ret = blk_mq_do_dispatch_ctx(hctx); |
caf8eb0d | 336 | } else { |
c13660a0 | 337 | blk_mq_flush_busy_ctxs(hctx, &rq_list); |
1fd40b5e | 338 | blk_mq_dispatch_rq_list(hctx, &rq_list, 0); |
64765a75 | 339 | } |
28d65729 SQ |
340 | |
341 | return ret; | |
342 | } | |
343 | ||
344 | void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx) | |
345 | { | |
346 | struct request_queue *q = hctx->queue; | |
347 | ||
348 | /* RCU or SRCU read lock is needed before checking quiesced flag */ | |
349 | if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q))) | |
350 | return; | |
351 | ||
352 | hctx->run++; | |
353 | ||
354 | /* | |
355 | * A return of -EAGAIN is an indication that hctx->dispatch is not | |
356 | * empty and we must run again in order to avoid starving flushes. | |
357 | */ | |
358 | if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) { | |
359 | if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) | |
360 | blk_mq_run_hw_queue(hctx, true); | |
361 | } | |
bd166ef1 JA |
362 | } |
363 | ||
179ae84f | 364 | bool blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio, |
14ccb66b | 365 | unsigned int nr_segs) |
bd166ef1 JA |
366 | { |
367 | struct elevator_queue *e = q->elevator; | |
efed9a33 OS |
368 | struct blk_mq_ctx *ctx; |
369 | struct blk_mq_hw_ctx *hctx; | |
9bddeb2a | 370 | bool ret = false; |
c16d6b5a | 371 | enum hctx_type type; |
bd166ef1 | 372 | |
900e0807 JA |
373 | if (e && e->type->ops.bio_merge) { |
374 | ret = e->type->ops.bio_merge(q, bio, nr_segs); | |
375 | goto out_put; | |
376 | } | |
bd166ef1 | 377 | |
efed9a33 OS |
378 | ctx = blk_mq_get_ctx(q); |
379 | hctx = blk_mq_map_queue(q, bio->bi_opf, ctx); | |
c16d6b5a | 380 | type = hctx->type; |
cdfcef9e BW |
381 | if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE) || |
382 | list_empty_careful(&ctx->rq_lists[type])) | |
900e0807 | 383 | goto out_put; |
cdfcef9e BW |
384 | |
385 | /* default per sw-queue merge */ | |
386 | spin_lock(&ctx->lock); | |
387 | /* | |
388 | * Reverse check our software queue for entries that we could | |
389 | * potentially merge with. Currently includes a hand-wavy stop | |
390 | * count of 8, to not spend too much time checking for merges. | |
391 | */ | |
9a14d6ce | 392 | if (blk_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) |
cdfcef9e | 393 | ret = true; |
9bddeb2a | 394 | |
cdfcef9e | 395 | spin_unlock(&ctx->lock); |
900e0807 | 396 | out_put: |
9bddeb2a | 397 | return ret; |
bd166ef1 JA |
398 | } |
399 | ||
fd2ef39c JK |
400 | bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq, |
401 | struct list_head *free) | |
bd166ef1 | 402 | { |
fd2ef39c | 403 | return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq, free); |
bd166ef1 JA |
404 | } |
405 | EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge); | |
406 | ||
0cacba6c OS |
407 | static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx, |
408 | struct request *rq) | |
bd166ef1 | 409 | { |
01e99aec ML |
410 | /* |
411 | * dispatch flush and passthrough rq directly | |
412 | * | |
413 | * passthrough request has to be added to hctx->dispatch directly. | |
414 | * For some reason, device may be in one situation which can't | |
415 | * handle FS request, so STS_RESOURCE is always returned and the | |
416 | * FS request will be added to hctx->dispatch. However passthrough | |
417 | * request may be required at that time for fixing the problem. If | |
418 | * passthrough request is added to scheduler queue, there isn't any | |
419 | * chance to dispatch it given we prioritize requests in hctx->dispatch. | |
420 | */ | |
421 | if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq)) | |
a6a252e6 | 422 | return true; |
a6a252e6 | 423 | |
a6a252e6 | 424 | return false; |
bd166ef1 | 425 | } |
bd166ef1 | 426 | |
bd6737f1 | 427 | void blk_mq_sched_insert_request(struct request *rq, bool at_head, |
9e97d295 | 428 | bool run_queue, bool async) |
bd6737f1 JA |
429 | { |
430 | struct request_queue *q = rq->q; | |
431 | struct elevator_queue *e = q->elevator; | |
432 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 433 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
bd6737f1 | 434 | |
e44a6a23 | 435 | WARN_ON(e && (rq->tag != BLK_MQ_NO_TAG)); |
923218f6 | 436 | |
5218e12e | 437 | if (blk_mq_sched_bypass_insert(hctx, rq)) { |
cc3200ea ML |
438 | /* |
439 | * Firstly normal IO request is inserted to scheduler queue or | |
440 | * sw queue, meantime we add flush request to dispatch queue( | |
441 | * hctx->dispatch) directly and there is at most one in-flight | |
442 | * flush request for each hw queue, so it doesn't matter to add | |
443 | * flush request to tail or front of the dispatch queue. | |
444 | * | |
445 | * Secondly in case of NCQ, flush request belongs to non-NCQ | |
446 | * command, and queueing it will fail when there is any | |
447 | * in-flight normal IO request(NCQ command). When adding flush | |
448 | * rq to the front of hctx->dispatch, it is easier to introduce | |
449 | * extra time to flush rq's latency because of S_SCHED_RESTART | |
450 | * compared with adding to the tail of dispatch queue, then | |
451 | * chance of flush merge is increased, and less flush requests | |
452 | * will be issued to controller. It is observed that ~10% time | |
453 | * is saved in blktests block/004 on disk attached to AHCI/NCQ | |
454 | * drive when adding flush rq to the front of hctx->dispatch. | |
455 | * | |
456 | * Simply queue flush rq to the front of hctx->dispatch so that | |
457 | * intensive flush workloads can benefit in case of NCQ HW. | |
458 | */ | |
459 | at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head; | |
01e99aec | 460 | blk_mq_request_bypass_insert(rq, at_head, false); |
0cacba6c | 461 | goto run; |
01e99aec | 462 | } |
0cacba6c | 463 | |
e42cfb1d | 464 | if (e) { |
bd6737f1 JA |
465 | LIST_HEAD(list); |
466 | ||
467 | list_add(&rq->queuelist, &list); | |
f9cd4bfe | 468 | e->type->ops.insert_requests(hctx, &list, at_head); |
bd6737f1 JA |
469 | } else { |
470 | spin_lock(&ctx->lock); | |
471 | __blk_mq_insert_request(hctx, rq, at_head); | |
472 | spin_unlock(&ctx->lock); | |
473 | } | |
474 | ||
0cacba6c | 475 | run: |
bd6737f1 JA |
476 | if (run_queue) |
477 | blk_mq_run_hw_queue(hctx, async); | |
478 | } | |
479 | ||
67cae4c9 | 480 | void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx, |
bd6737f1 JA |
481 | struct blk_mq_ctx *ctx, |
482 | struct list_head *list, bool run_queue_async) | |
483 | { | |
f9afca4d | 484 | struct elevator_queue *e; |
e87eb301 ML |
485 | struct request_queue *q = hctx->queue; |
486 | ||
487 | /* | |
488 | * blk_mq_sched_insert_requests() is called from flush plug | |
489 | * context only, and hold one usage counter to prevent queue | |
490 | * from being released. | |
491 | */ | |
492 | percpu_ref_get(&q->q_usage_counter); | |
bd6737f1 | 493 | |
f9afca4d | 494 | e = hctx->queue->elevator; |
e42cfb1d | 495 | if (e) { |
f9cd4bfe | 496 | e->type->ops.insert_requests(hctx, list, false); |
e42cfb1d | 497 | } else { |
6ce3dd6e ML |
498 | /* |
499 | * try to issue requests directly if the hw queue isn't | |
500 | * busy in case of 'none' scheduler, and this way may save | |
501 | * us one extra enqueue & dequeue to sw queue. | |
502 | */ | |
ef1661ba | 503 | if (!hctx->dispatch_busy && !run_queue_async) { |
6ce3dd6e | 504 | blk_mq_try_issue_list_directly(hctx, list); |
fd9c40f6 | 505 | if (list_empty(list)) |
e87eb301 | 506 | goto out; |
fd9c40f6 BVA |
507 | } |
508 | blk_mq_insert_requests(hctx, ctx, list); | |
6ce3dd6e | 509 | } |
bd6737f1 JA |
510 | |
511 | blk_mq_run_hw_queue(hctx, run_queue_async); | |
e87eb301 ML |
512 | out: |
513 | percpu_ref_put(&q->q_usage_counter); | |
bd6737f1 JA |
514 | } |
515 | ||
d99a6bb3 JG |
516 | static int blk_mq_sched_alloc_map_and_rqs(struct request_queue *q, |
517 | struct blk_mq_hw_ctx *hctx, | |
518 | unsigned int hctx_idx) | |
6917ff0b | 519 | { |
079a2e3e JG |
520 | if (blk_mq_is_shared_tags(q->tag_set->flags)) { |
521 | hctx->sched_tags = q->sched_shared_tags; | |
e155b0c2 JG |
522 | return 0; |
523 | } | |
524 | ||
63064be1 JG |
525 | hctx->sched_tags = blk_mq_alloc_map_and_rqs(q->tag_set, hctx_idx, |
526 | q->nr_requests); | |
6917ff0b | 527 | |
6917ff0b OS |
528 | if (!hctx->sched_tags) |
529 | return -ENOMEM; | |
63064be1 | 530 | return 0; |
6917ff0b OS |
531 | } |
532 | ||
079a2e3e | 533 | static void blk_mq_exit_sched_shared_tags(struct request_queue *queue) |
e155b0c2 | 534 | { |
079a2e3e JG |
535 | blk_mq_free_rq_map(queue->sched_shared_tags); |
536 | queue->sched_shared_tags = NULL; | |
e155b0c2 JG |
537 | } |
538 | ||
c3e22192 | 539 | /* called in queue's release handler, tagset has gone away */ |
e155b0c2 | 540 | static void blk_mq_sched_tags_teardown(struct request_queue *q, unsigned int flags) |
bd166ef1 | 541 | { |
bd166ef1 | 542 | struct blk_mq_hw_ctx *hctx; |
6917ff0b OS |
543 | int i; |
544 | ||
c3e22192 ML |
545 | queue_for_each_hw_ctx(q, hctx, i) { |
546 | if (hctx->sched_tags) { | |
8bdf7b3f | 547 | if (!blk_mq_is_shared_tags(flags)) |
e155b0c2 | 548 | blk_mq_free_rq_map(hctx->sched_tags); |
c3e22192 ML |
549 | hctx->sched_tags = NULL; |
550 | } | |
551 | } | |
e155b0c2 | 552 | |
079a2e3e JG |
553 | if (blk_mq_is_shared_tags(flags)) |
554 | blk_mq_exit_sched_shared_tags(q); | |
6917ff0b OS |
555 | } |
556 | ||
079a2e3e | 557 | static int blk_mq_init_sched_shared_tags(struct request_queue *queue) |
d97e594c JG |
558 | { |
559 | struct blk_mq_tag_set *set = queue->tag_set; | |
d97e594c JG |
560 | |
561 | /* | |
562 | * Set initial depth at max so that we don't need to reallocate for | |
563 | * updating nr_requests. | |
564 | */ | |
079a2e3e | 565 | queue->sched_shared_tags = blk_mq_alloc_map_and_rqs(set, |
e155b0c2 JG |
566 | BLK_MQ_NO_HCTX_IDX, |
567 | MAX_SCHED_RQ); | |
079a2e3e | 568 | if (!queue->sched_shared_tags) |
e155b0c2 | 569 | return -ENOMEM; |
d97e594c | 570 | |
079a2e3e | 571 | blk_mq_tag_update_sched_shared_tags(queue); |
d97e594c JG |
572 | |
573 | return 0; | |
574 | } | |
575 | ||
6917ff0b OS |
576 | int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e) |
577 | { | |
e155b0c2 | 578 | unsigned int i, flags = q->tag_set->flags; |
6917ff0b | 579 | struct blk_mq_hw_ctx *hctx; |
ee056f98 | 580 | struct elevator_queue *eq; |
6917ff0b OS |
581 | int ret; |
582 | ||
583 | if (!e) { | |
584 | q->elevator = NULL; | |
32a50fab | 585 | q->nr_requests = q->tag_set->queue_depth; |
6917ff0b OS |
586 | return 0; |
587 | } | |
bd166ef1 JA |
588 | |
589 | /* | |
32825c45 ML |
590 | * Default to double of smaller one between hw queue_depth and 128, |
591 | * since we don't split into sync/async like the old code did. | |
592 | * Additionally, this is a per-hw queue depth. | |
bd166ef1 | 593 | */ |
32825c45 | 594 | q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth, |
d2a27964 | 595 | BLKDEV_DEFAULT_RQ); |
bd166ef1 | 596 | |
079a2e3e JG |
597 | if (blk_mq_is_shared_tags(flags)) { |
598 | ret = blk_mq_init_sched_shared_tags(q); | |
bd166ef1 | 599 | if (ret) |
e155b0c2 | 600 | return ret; |
d97e594c JG |
601 | } |
602 | ||
e155b0c2 JG |
603 | queue_for_each_hw_ctx(q, hctx, i) { |
604 | ret = blk_mq_sched_alloc_map_and_rqs(q, hctx, i); | |
d97e594c | 605 | if (ret) |
d99a6bb3 | 606 | goto err_free_map_and_rqs; |
bd166ef1 JA |
607 | } |
608 | ||
f9cd4bfe | 609 | ret = e->ops.init_sched(q, e); |
6917ff0b | 610 | if (ret) |
e155b0c2 | 611 | goto err_free_map_and_rqs; |
bd166ef1 | 612 | |
d332ce09 OS |
613 | blk_mq_debugfs_register_sched(q); |
614 | ||
615 | queue_for_each_hw_ctx(q, hctx, i) { | |
f9cd4bfe JA |
616 | if (e->ops.init_hctx) { |
617 | ret = e->ops.init_hctx(hctx, i); | |
ee056f98 OS |
618 | if (ret) { |
619 | eq = q->elevator; | |
1820f4f0 | 620 | blk_mq_sched_free_rqs(q); |
ee056f98 OS |
621 | blk_mq_exit_sched(q, eq); |
622 | kobject_put(&eq->kobj); | |
623 | return ret; | |
624 | } | |
625 | } | |
d332ce09 | 626 | blk_mq_debugfs_register_sched_hctx(q, hctx); |
ee056f98 OS |
627 | } |
628 | ||
bd166ef1 | 629 | return 0; |
bd166ef1 | 630 | |
d99a6bb3 | 631 | err_free_map_and_rqs: |
1820f4f0 | 632 | blk_mq_sched_free_rqs(q); |
e155b0c2 JG |
633 | blk_mq_sched_tags_teardown(q, flags); |
634 | ||
54d5329d | 635 | q->elevator = NULL; |
6917ff0b | 636 | return ret; |
bd166ef1 | 637 | } |
d3484991 | 638 | |
c3e22192 ML |
639 | /* |
640 | * called in either blk_queue_cleanup or elevator_switch, tagset | |
641 | * is required for freeing requests | |
642 | */ | |
1820f4f0 | 643 | void blk_mq_sched_free_rqs(struct request_queue *q) |
c3e22192 ML |
644 | { |
645 | struct blk_mq_hw_ctx *hctx; | |
646 | int i; | |
647 | ||
079a2e3e JG |
648 | if (blk_mq_is_shared_tags(q->tag_set->flags)) { |
649 | blk_mq_free_rqs(q->tag_set, q->sched_shared_tags, | |
e155b0c2 JG |
650 | BLK_MQ_NO_HCTX_IDX); |
651 | } else { | |
652 | queue_for_each_hw_ctx(q, hctx, i) { | |
653 | if (hctx->sched_tags) | |
654 | blk_mq_free_rqs(q->tag_set, | |
655 | hctx->sched_tags, i); | |
656 | } | |
c3e22192 ML |
657 | } |
658 | } | |
659 | ||
54d5329d OS |
660 | void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e) |
661 | { | |
ee056f98 OS |
662 | struct blk_mq_hw_ctx *hctx; |
663 | unsigned int i; | |
f0c1c4d2 | 664 | unsigned int flags = 0; |
ee056f98 | 665 | |
d332ce09 OS |
666 | queue_for_each_hw_ctx(q, hctx, i) { |
667 | blk_mq_debugfs_unregister_sched_hctx(hctx); | |
f9cd4bfe JA |
668 | if (e->type->ops.exit_hctx && hctx->sched_data) { |
669 | e->type->ops.exit_hctx(hctx, i); | |
d332ce09 | 670 | hctx->sched_data = NULL; |
ee056f98 | 671 | } |
f0c1c4d2 | 672 | flags = hctx->flags; |
ee056f98 | 673 | } |
d332ce09 | 674 | blk_mq_debugfs_unregister_sched(q); |
f9cd4bfe JA |
675 | if (e->type->ops.exit_sched) |
676 | e->type->ops.exit_sched(e); | |
e155b0c2 | 677 | blk_mq_sched_tags_teardown(q, flags); |
54d5329d OS |
678 | q->elevator = NULL; |
679 | } |