Loading...
1// SPDX-License-Identifier: GPL-2.0
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>
10#include <linux/list_sort.h>
11
12#include <trace/events/block.h>
13
14#include "blk.h"
15#include "blk-mq.h"
16#include "blk-mq-debugfs.h"
17#include "blk-mq-sched.h"
18#include "blk-mq-tag.h"
19#include "blk-wbt.h"
20
21/*
22 * Mark a hardware queue as needing a restart. For shared queues, maintain
23 * a count of how many hardware queues are marked for restart.
24 */
25void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
26{
27 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
28 return;
29
30 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
31}
32EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
33
34void __blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
35{
36 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
37
38 /*
39 * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch)
40 * in blk_mq_run_hw_queue(). Its pair is the barrier in
41 * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART,
42 * meantime new request added to hctx->dispatch is missed to check in
43 * blk_mq_run_hw_queue().
44 */
45 smp_mb();
46
47 blk_mq_run_hw_queue(hctx, true);
48}
49
50static int sched_rq_cmp(void *priv, const struct list_head *a,
51 const struct list_head *b)
52{
53 struct request *rqa = container_of(a, struct request, queuelist);
54 struct request *rqb = container_of(b, struct request, queuelist);
55
56 return rqa->mq_hctx > rqb->mq_hctx;
57}
58
59static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list)
60{
61 struct blk_mq_hw_ctx *hctx =
62 list_first_entry(rq_list, struct request, queuelist)->mq_hctx;
63 struct request *rq;
64 LIST_HEAD(hctx_list);
65 unsigned int count = 0;
66
67 list_for_each_entry(rq, rq_list, queuelist) {
68 if (rq->mq_hctx != hctx) {
69 list_cut_before(&hctx_list, rq_list, &rq->queuelist);
70 goto dispatch;
71 }
72 count++;
73 }
74 list_splice_tail_init(rq_list, &hctx_list);
75
76dispatch:
77 return blk_mq_dispatch_rq_list(hctx, &hctx_list, count);
78}
79
80#define BLK_MQ_BUDGET_DELAY 3 /* ms units */
81
82/*
83 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
84 * its queue by itself in its completion handler, so we don't need to
85 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
86 *
87 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
88 * be run again. This is necessary to avoid starving flushes.
89 */
90static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
91{
92 struct request_queue *q = hctx->queue;
93 struct elevator_queue *e = q->elevator;
94 bool multi_hctxs = false, run_queue = false;
95 bool dispatched = false, busy = false;
96 unsigned int max_dispatch;
97 LIST_HEAD(rq_list);
98 int count = 0;
99
100 if (hctx->dispatch_busy)
101 max_dispatch = 1;
102 else
103 max_dispatch = hctx->queue->nr_requests;
104
105 do {
106 struct request *rq;
107 int budget_token;
108
109 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
110 break;
111
112 if (!list_empty_careful(&hctx->dispatch)) {
113 busy = true;
114 break;
115 }
116
117 budget_token = blk_mq_get_dispatch_budget(q);
118 if (budget_token < 0)
119 break;
120
121 rq = e->type->ops.dispatch_request(hctx);
122 if (!rq) {
123 blk_mq_put_dispatch_budget(q, budget_token);
124 /*
125 * We're releasing without dispatching. Holding the
126 * budget could have blocked any "hctx"s with the
127 * same queue and if we didn't dispatch then there's
128 * no guarantee anyone will kick the queue. Kick it
129 * ourselves.
130 */
131 run_queue = true;
132 break;
133 }
134
135 blk_mq_set_rq_budget_token(rq, budget_token);
136
137 /*
138 * Now this rq owns the budget which has to be released
139 * if this rq won't be queued to driver via .queue_rq()
140 * in blk_mq_dispatch_rq_list().
141 */
142 list_add_tail(&rq->queuelist, &rq_list);
143 count++;
144 if (rq->mq_hctx != hctx)
145 multi_hctxs = true;
146
147 /*
148 * If we cannot get tag for the request, stop dequeueing
149 * requests from the IO scheduler. We are unlikely to be able
150 * to submit them anyway and it creates false impression for
151 * scheduling heuristics that the device can take more IO.
152 */
153 if (!blk_mq_get_driver_tag(rq))
154 break;
155 } while (count < max_dispatch);
156
157 if (!count) {
158 if (run_queue)
159 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
160 } else if (multi_hctxs) {
161 /*
162 * Requests from different hctx may be dequeued from some
163 * schedulers, such as bfq and deadline.
164 *
165 * Sort the requests in the list according to their hctx,
166 * dispatch batching requests from same hctx at a time.
167 */
168 list_sort(NULL, &rq_list, sched_rq_cmp);
169 do {
170 dispatched |= blk_mq_dispatch_hctx_list(&rq_list);
171 } while (!list_empty(&rq_list));
172 } else {
173 dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, count);
174 }
175
176 if (busy)
177 return -EAGAIN;
178 return !!dispatched;
179}
180
181static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
182{
183 unsigned long end = jiffies + HZ;
184 int ret;
185
186 do {
187 ret = __blk_mq_do_dispatch_sched(hctx);
188 if (ret != 1)
189 break;
190 if (need_resched() || time_is_before_jiffies(end)) {
191 blk_mq_delay_run_hw_queue(hctx, 0);
192 break;
193 }
194 } while (1);
195
196 return ret;
197}
198
199static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
200 struct blk_mq_ctx *ctx)
201{
202 unsigned short idx = ctx->index_hw[hctx->type];
203
204 if (++idx == hctx->nr_ctx)
205 idx = 0;
206
207 return hctx->ctxs[idx];
208}
209
210/*
211 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
212 * its queue by itself in its completion handler, so we don't need to
213 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
214 *
215 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
216 * be run again. This is necessary to avoid starving flushes.
217 */
218static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
219{
220 struct request_queue *q = hctx->queue;
221 LIST_HEAD(rq_list);
222 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
223 int ret = 0;
224 struct request *rq;
225
226 do {
227 int budget_token;
228
229 if (!list_empty_careful(&hctx->dispatch)) {
230 ret = -EAGAIN;
231 break;
232 }
233
234 if (!sbitmap_any_bit_set(&hctx->ctx_map))
235 break;
236
237 budget_token = blk_mq_get_dispatch_budget(q);
238 if (budget_token < 0)
239 break;
240
241 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
242 if (!rq) {
243 blk_mq_put_dispatch_budget(q, budget_token);
244 /*
245 * We're releasing without dispatching. Holding the
246 * budget could have blocked any "hctx"s with the
247 * same queue and if we didn't dispatch then there's
248 * no guarantee anyone will kick the queue. Kick it
249 * ourselves.
250 */
251 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
252 break;
253 }
254
255 blk_mq_set_rq_budget_token(rq, budget_token);
256
257 /*
258 * Now this rq owns the budget which has to be released
259 * if this rq won't be queued to driver via .queue_rq()
260 * in blk_mq_dispatch_rq_list().
261 */
262 list_add(&rq->queuelist, &rq_list);
263
264 /* round robin for fair dispatch */
265 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
266
267 } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1));
268
269 WRITE_ONCE(hctx->dispatch_from, ctx);
270 return ret;
271}
272
273static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
274{
275 struct request_queue *q = hctx->queue;
276 const bool has_sched = q->elevator;
277 int ret = 0;
278 LIST_HEAD(rq_list);
279
280 /*
281 * If we have previous entries on our dispatch list, grab them first for
282 * more fair dispatch.
283 */
284 if (!list_empty_careful(&hctx->dispatch)) {
285 spin_lock(&hctx->lock);
286 if (!list_empty(&hctx->dispatch))
287 list_splice_init(&hctx->dispatch, &rq_list);
288 spin_unlock(&hctx->lock);
289 }
290
291 /*
292 * Only ask the scheduler for requests, if we didn't have residual
293 * requests from the dispatch list. This is to avoid the case where
294 * we only ever dispatch a fraction of the requests available because
295 * of low device queue depth. Once we pull requests out of the IO
296 * scheduler, we can no longer merge or sort them. So it's best to
297 * leave them there for as long as we can. Mark the hw queue as
298 * needing a restart in that case.
299 *
300 * We want to dispatch from the scheduler if there was nothing
301 * on the dispatch list or we were able to dispatch from the
302 * dispatch list.
303 */
304 if (!list_empty(&rq_list)) {
305 blk_mq_sched_mark_restart_hctx(hctx);
306 if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) {
307 if (has_sched)
308 ret = blk_mq_do_dispatch_sched(hctx);
309 else
310 ret = blk_mq_do_dispatch_ctx(hctx);
311 }
312 } else if (has_sched) {
313 ret = blk_mq_do_dispatch_sched(hctx);
314 } else if (hctx->dispatch_busy) {
315 /* dequeue request one by one from sw queue if queue is busy */
316 ret = blk_mq_do_dispatch_ctx(hctx);
317 } else {
318 blk_mq_flush_busy_ctxs(hctx, &rq_list);
319 blk_mq_dispatch_rq_list(hctx, &rq_list, 0);
320 }
321
322 return ret;
323}
324
325void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
326{
327 struct request_queue *q = hctx->queue;
328
329 /* RCU or SRCU read lock is needed before checking quiesced flag */
330 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
331 return;
332
333 hctx->run++;
334
335 /*
336 * A return of -EAGAIN is an indication that hctx->dispatch is not
337 * empty and we must run again in order to avoid starving flushes.
338 */
339 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) {
340 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN)
341 blk_mq_run_hw_queue(hctx, true);
342 }
343}
344
345bool blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
346 unsigned int nr_segs)
347{
348 struct elevator_queue *e = q->elevator;
349 struct blk_mq_ctx *ctx;
350 struct blk_mq_hw_ctx *hctx;
351 bool ret = false;
352 enum hctx_type type;
353
354 if (e && e->type->ops.bio_merge) {
355 ret = e->type->ops.bio_merge(q, bio, nr_segs);
356 goto out_put;
357 }
358
359 ctx = blk_mq_get_ctx(q);
360 hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
361 type = hctx->type;
362 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE) ||
363 list_empty_careful(&ctx->rq_lists[type]))
364 goto out_put;
365
366 /* default per sw-queue merge */
367 spin_lock(&ctx->lock);
368 /*
369 * Reverse check our software queue for entries that we could
370 * potentially merge with. Currently includes a hand-wavy stop
371 * count of 8, to not spend too much time checking for merges.
372 */
373 if (blk_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs))
374 ret = true;
375
376 spin_unlock(&ctx->lock);
377out_put:
378 return ret;
379}
380
381bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq,
382 struct list_head *free)
383{
384 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq, free);
385}
386EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
387
388static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
389 struct request *rq)
390{
391 /*
392 * dispatch flush and passthrough rq directly
393 *
394 * passthrough request has to be added to hctx->dispatch directly.
395 * For some reason, device may be in one situation which can't
396 * handle FS request, so STS_RESOURCE is always returned and the
397 * FS request will be added to hctx->dispatch. However passthrough
398 * request may be required at that time for fixing the problem. If
399 * passthrough request is added to scheduler queue, there isn't any
400 * chance to dispatch it given we prioritize requests in hctx->dispatch.
401 */
402 if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq))
403 return true;
404
405 return false;
406}
407
408void blk_mq_sched_insert_request(struct request *rq, bool at_head,
409 bool run_queue, bool async)
410{
411 struct request_queue *q = rq->q;
412 struct elevator_queue *e = q->elevator;
413 struct blk_mq_ctx *ctx = rq->mq_ctx;
414 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
415
416 WARN_ON(e && (rq->tag != BLK_MQ_NO_TAG));
417
418 if (blk_mq_sched_bypass_insert(hctx, rq)) {
419 /*
420 * Firstly normal IO request is inserted to scheduler queue or
421 * sw queue, meantime we add flush request to dispatch queue(
422 * hctx->dispatch) directly and there is at most one in-flight
423 * flush request for each hw queue, so it doesn't matter to add
424 * flush request to tail or front of the dispatch queue.
425 *
426 * Secondly in case of NCQ, flush request belongs to non-NCQ
427 * command, and queueing it will fail when there is any
428 * in-flight normal IO request(NCQ command). When adding flush
429 * rq to the front of hctx->dispatch, it is easier to introduce
430 * extra time to flush rq's latency because of S_SCHED_RESTART
431 * compared with adding to the tail of dispatch queue, then
432 * chance of flush merge is increased, and less flush requests
433 * will be issued to controller. It is observed that ~10% time
434 * is saved in blktests block/004 on disk attached to AHCI/NCQ
435 * drive when adding flush rq to the front of hctx->dispatch.
436 *
437 * Simply queue flush rq to the front of hctx->dispatch so that
438 * intensive flush workloads can benefit in case of NCQ HW.
439 */
440 at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head;
441 blk_mq_request_bypass_insert(rq, at_head, false);
442 goto run;
443 }
444
445 if (e) {
446 LIST_HEAD(list);
447
448 list_add(&rq->queuelist, &list);
449 e->type->ops.insert_requests(hctx, &list, at_head);
450 } else {
451 spin_lock(&ctx->lock);
452 __blk_mq_insert_request(hctx, rq, at_head);
453 spin_unlock(&ctx->lock);
454 }
455
456run:
457 if (run_queue)
458 blk_mq_run_hw_queue(hctx, async);
459}
460
461void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
462 struct blk_mq_ctx *ctx,
463 struct list_head *list, bool run_queue_async)
464{
465 struct elevator_queue *e;
466 struct request_queue *q = hctx->queue;
467
468 /*
469 * blk_mq_sched_insert_requests() is called from flush plug
470 * context only, and hold one usage counter to prevent queue
471 * from being released.
472 */
473 percpu_ref_get(&q->q_usage_counter);
474
475 e = hctx->queue->elevator;
476 if (e) {
477 e->type->ops.insert_requests(hctx, list, false);
478 } else {
479 /*
480 * try to issue requests directly if the hw queue isn't
481 * busy in case of 'none' scheduler, and this way may save
482 * us one extra enqueue & dequeue to sw queue.
483 */
484 if (!hctx->dispatch_busy && !run_queue_async) {
485 blk_mq_run_dispatch_ops(hctx->queue,
486 blk_mq_try_issue_list_directly(hctx, list));
487 if (list_empty(list))
488 goto out;
489 }
490 blk_mq_insert_requests(hctx, ctx, list);
491 }
492
493 blk_mq_run_hw_queue(hctx, run_queue_async);
494 out:
495 percpu_ref_put(&q->q_usage_counter);
496}
497
498static int blk_mq_sched_alloc_map_and_rqs(struct request_queue *q,
499 struct blk_mq_hw_ctx *hctx,
500 unsigned int hctx_idx)
501{
502 if (blk_mq_is_shared_tags(q->tag_set->flags)) {
503 hctx->sched_tags = q->sched_shared_tags;
504 return 0;
505 }
506
507 hctx->sched_tags = blk_mq_alloc_map_and_rqs(q->tag_set, hctx_idx,
508 q->nr_requests);
509
510 if (!hctx->sched_tags)
511 return -ENOMEM;
512 return 0;
513}
514
515static void blk_mq_exit_sched_shared_tags(struct request_queue *queue)
516{
517 blk_mq_free_rq_map(queue->sched_shared_tags);
518 queue->sched_shared_tags = NULL;
519}
520
521/* called in queue's release handler, tagset has gone away */
522static void blk_mq_sched_tags_teardown(struct request_queue *q, unsigned int flags)
523{
524 struct blk_mq_hw_ctx *hctx;
525 unsigned long i;
526
527 queue_for_each_hw_ctx(q, hctx, i) {
528 if (hctx->sched_tags) {
529 if (!blk_mq_is_shared_tags(flags))
530 blk_mq_free_rq_map(hctx->sched_tags);
531 hctx->sched_tags = NULL;
532 }
533 }
534
535 if (blk_mq_is_shared_tags(flags))
536 blk_mq_exit_sched_shared_tags(q);
537}
538
539static int blk_mq_init_sched_shared_tags(struct request_queue *queue)
540{
541 struct blk_mq_tag_set *set = queue->tag_set;
542
543 /*
544 * Set initial depth at max so that we don't need to reallocate for
545 * updating nr_requests.
546 */
547 queue->sched_shared_tags = blk_mq_alloc_map_and_rqs(set,
548 BLK_MQ_NO_HCTX_IDX,
549 MAX_SCHED_RQ);
550 if (!queue->sched_shared_tags)
551 return -ENOMEM;
552
553 blk_mq_tag_update_sched_shared_tags(queue);
554
555 return 0;
556}
557
558/* caller must have a reference to @e, will grab another one if successful */
559int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
560{
561 unsigned int flags = q->tag_set->flags;
562 struct blk_mq_hw_ctx *hctx;
563 struct elevator_queue *eq;
564 unsigned long i;
565 int ret;
566
567 /*
568 * Default to double of smaller one between hw queue_depth and 128,
569 * since we don't split into sync/async like the old code did.
570 * Additionally, this is a per-hw queue depth.
571 */
572 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
573 BLKDEV_DEFAULT_RQ);
574
575 if (blk_mq_is_shared_tags(flags)) {
576 ret = blk_mq_init_sched_shared_tags(q);
577 if (ret)
578 return ret;
579 }
580
581 queue_for_each_hw_ctx(q, hctx, i) {
582 ret = blk_mq_sched_alloc_map_and_rqs(q, hctx, i);
583 if (ret)
584 goto err_free_map_and_rqs;
585 }
586
587 ret = e->ops.init_sched(q, e);
588 if (ret)
589 goto err_free_map_and_rqs;
590
591 mutex_lock(&q->debugfs_mutex);
592 blk_mq_debugfs_register_sched(q);
593 mutex_unlock(&q->debugfs_mutex);
594
595 queue_for_each_hw_ctx(q, hctx, i) {
596 if (e->ops.init_hctx) {
597 ret = e->ops.init_hctx(hctx, i);
598 if (ret) {
599 eq = q->elevator;
600 blk_mq_sched_free_rqs(q);
601 blk_mq_exit_sched(q, eq);
602 kobject_put(&eq->kobj);
603 return ret;
604 }
605 }
606 mutex_lock(&q->debugfs_mutex);
607 blk_mq_debugfs_register_sched_hctx(q, hctx);
608 mutex_unlock(&q->debugfs_mutex);
609 }
610
611 return 0;
612
613err_free_map_and_rqs:
614 blk_mq_sched_free_rqs(q);
615 blk_mq_sched_tags_teardown(q, flags);
616
617 q->elevator = NULL;
618 return ret;
619}
620
621/*
622 * called in either blk_queue_cleanup or elevator_switch, tagset
623 * is required for freeing requests
624 */
625void blk_mq_sched_free_rqs(struct request_queue *q)
626{
627 struct blk_mq_hw_ctx *hctx;
628 unsigned long i;
629
630 if (blk_mq_is_shared_tags(q->tag_set->flags)) {
631 blk_mq_free_rqs(q->tag_set, q->sched_shared_tags,
632 BLK_MQ_NO_HCTX_IDX);
633 } else {
634 queue_for_each_hw_ctx(q, hctx, i) {
635 if (hctx->sched_tags)
636 blk_mq_free_rqs(q->tag_set,
637 hctx->sched_tags, i);
638 }
639 }
640}
641
642void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
643{
644 struct blk_mq_hw_ctx *hctx;
645 unsigned long i;
646 unsigned int flags = 0;
647
648 queue_for_each_hw_ctx(q, hctx, i) {
649 mutex_lock(&q->debugfs_mutex);
650 blk_mq_debugfs_unregister_sched_hctx(hctx);
651 mutex_unlock(&q->debugfs_mutex);
652
653 if (e->type->ops.exit_hctx && hctx->sched_data) {
654 e->type->ops.exit_hctx(hctx, i);
655 hctx->sched_data = NULL;
656 }
657 flags = hctx->flags;
658 }
659
660 mutex_lock(&q->debugfs_mutex);
661 blk_mq_debugfs_unregister_sched(q);
662 mutex_unlock(&q->debugfs_mutex);
663
664 if (e->type->ops.exit_sched)
665 e->type->ops.exit_sched(e);
666 blk_mq_sched_tags_teardown(q, flags);
667 q->elevator = NULL;
668}
1// SPDX-License-Identifier: GPL-2.0
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>
10#include <linux/list_sort.h>
11
12#include <trace/events/block.h>
13
14#include "blk.h"
15#include "blk-mq.h"
16#include "blk-mq-debugfs.h"
17#include "blk-mq-sched.h"
18#include "blk-mq-tag.h"
19#include "blk-wbt.h"
20
21void blk_mq_sched_free_hctx_data(struct request_queue *q,
22 void (*exit)(struct blk_mq_hw_ctx *))
23{
24 struct blk_mq_hw_ctx *hctx;
25 int i;
26
27 queue_for_each_hw_ctx(q, hctx, i) {
28 if (exit && hctx->sched_data)
29 exit(hctx);
30 kfree(hctx->sched_data);
31 hctx->sched_data = NULL;
32 }
33}
34EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data);
35
36void blk_mq_sched_assign_ioc(struct request *rq)
37{
38 struct request_queue *q = rq->q;
39 struct io_context *ioc;
40 struct io_cq *icq;
41
42 /*
43 * May not have an IO context if it's a passthrough request
44 */
45 ioc = current->io_context;
46 if (!ioc)
47 return;
48
49 spin_lock_irq(&q->queue_lock);
50 icq = ioc_lookup_icq(ioc, q);
51 spin_unlock_irq(&q->queue_lock);
52
53 if (!icq) {
54 icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
55 if (!icq)
56 return;
57 }
58 get_io_context(icq->ioc);
59 rq->elv.icq = icq;
60}
61
62/*
63 * Mark a hardware queue as needing a restart. For shared queues, maintain
64 * a count of how many hardware queues are marked for restart.
65 */
66void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
67{
68 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
69 return;
70
71 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
72}
73EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
74
75void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
76{
77 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
78 return;
79 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
80
81 /*
82 * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch)
83 * in blk_mq_run_hw_queue(). Its pair is the barrier in
84 * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART,
85 * meantime new request added to hctx->dispatch is missed to check in
86 * blk_mq_run_hw_queue().
87 */
88 smp_mb();
89
90 blk_mq_run_hw_queue(hctx, true);
91}
92
93static int sched_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
94{
95 struct request *rqa = container_of(a, struct request, queuelist);
96 struct request *rqb = container_of(b, struct request, queuelist);
97
98 return rqa->mq_hctx > rqb->mq_hctx;
99}
100
101static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list)
102{
103 struct blk_mq_hw_ctx *hctx =
104 list_first_entry(rq_list, struct request, queuelist)->mq_hctx;
105 struct request *rq;
106 LIST_HEAD(hctx_list);
107 unsigned int count = 0;
108
109 list_for_each_entry(rq, rq_list, queuelist) {
110 if (rq->mq_hctx != hctx) {
111 list_cut_before(&hctx_list, rq_list, &rq->queuelist);
112 goto dispatch;
113 }
114 count++;
115 }
116 list_splice_tail_init(rq_list, &hctx_list);
117
118dispatch:
119 return blk_mq_dispatch_rq_list(hctx, &hctx_list, count);
120}
121
122#define BLK_MQ_BUDGET_DELAY 3 /* ms units */
123
124/*
125 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
126 * its queue by itself in its completion handler, so we don't need to
127 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
128 *
129 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
130 * be run again. This is necessary to avoid starving flushes.
131 */
132static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
133{
134 struct request_queue *q = hctx->queue;
135 struct elevator_queue *e = q->elevator;
136 bool multi_hctxs = false, run_queue = false;
137 bool dispatched = false, busy = false;
138 unsigned int max_dispatch;
139 LIST_HEAD(rq_list);
140 int count = 0;
141
142 if (hctx->dispatch_busy)
143 max_dispatch = 1;
144 else
145 max_dispatch = hctx->queue->nr_requests;
146
147 do {
148 struct request *rq;
149
150 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
151 break;
152
153 if (!list_empty_careful(&hctx->dispatch)) {
154 busy = true;
155 break;
156 }
157
158 if (!blk_mq_get_dispatch_budget(q))
159 break;
160
161 rq = e->type->ops.dispatch_request(hctx);
162 if (!rq) {
163 blk_mq_put_dispatch_budget(q);
164 /*
165 * We're releasing without dispatching. Holding the
166 * budget could have blocked any "hctx"s with the
167 * same queue and if we didn't dispatch then there's
168 * no guarantee anyone will kick the queue. Kick it
169 * ourselves.
170 */
171 run_queue = true;
172 break;
173 }
174
175 /*
176 * Now this rq owns the budget which has to be released
177 * if this rq won't be queued to driver via .queue_rq()
178 * in blk_mq_dispatch_rq_list().
179 */
180 list_add_tail(&rq->queuelist, &rq_list);
181 if (rq->mq_hctx != hctx)
182 multi_hctxs = true;
183 } while (++count < max_dispatch);
184
185 if (!count) {
186 if (run_queue)
187 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
188 } else if (multi_hctxs) {
189 /*
190 * Requests from different hctx may be dequeued from some
191 * schedulers, such as bfq and deadline.
192 *
193 * Sort the requests in the list according to their hctx,
194 * dispatch batching requests from same hctx at a time.
195 */
196 list_sort(NULL, &rq_list, sched_rq_cmp);
197 do {
198 dispatched |= blk_mq_dispatch_hctx_list(&rq_list);
199 } while (!list_empty(&rq_list));
200 } else {
201 dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, count);
202 }
203
204 if (busy)
205 return -EAGAIN;
206 return !!dispatched;
207}
208
209static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
210{
211 int ret;
212
213 do {
214 ret = __blk_mq_do_dispatch_sched(hctx);
215 } while (ret == 1);
216
217 return ret;
218}
219
220static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
221 struct blk_mq_ctx *ctx)
222{
223 unsigned short idx = ctx->index_hw[hctx->type];
224
225 if (++idx == hctx->nr_ctx)
226 idx = 0;
227
228 return hctx->ctxs[idx];
229}
230
231/*
232 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
233 * its queue by itself in its completion handler, so we don't need to
234 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
235 *
236 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
237 * be run again. This is necessary to avoid starving flushes.
238 */
239static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
240{
241 struct request_queue *q = hctx->queue;
242 LIST_HEAD(rq_list);
243 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
244 int ret = 0;
245 struct request *rq;
246
247 do {
248 if (!list_empty_careful(&hctx->dispatch)) {
249 ret = -EAGAIN;
250 break;
251 }
252
253 if (!sbitmap_any_bit_set(&hctx->ctx_map))
254 break;
255
256 if (!blk_mq_get_dispatch_budget(q))
257 break;
258
259 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
260 if (!rq) {
261 blk_mq_put_dispatch_budget(q);
262 /*
263 * We're releasing without dispatching. Holding the
264 * budget could have blocked any "hctx"s with the
265 * same queue and if we didn't dispatch then there's
266 * no guarantee anyone will kick the queue. Kick it
267 * ourselves.
268 */
269 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
270 break;
271 }
272
273 /*
274 * Now this rq owns the budget which has to be released
275 * if this rq won't be queued to driver via .queue_rq()
276 * in blk_mq_dispatch_rq_list().
277 */
278 list_add(&rq->queuelist, &rq_list);
279
280 /* round robin for fair dispatch */
281 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
282
283 } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1));
284
285 WRITE_ONCE(hctx->dispatch_from, ctx);
286 return ret;
287}
288
289static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
290{
291 struct request_queue *q = hctx->queue;
292 struct elevator_queue *e = q->elevator;
293 const bool has_sched_dispatch = e && e->type->ops.dispatch_request;
294 int ret = 0;
295 LIST_HEAD(rq_list);
296
297 /*
298 * If we have previous entries on our dispatch list, grab them first for
299 * more fair dispatch.
300 */
301 if (!list_empty_careful(&hctx->dispatch)) {
302 spin_lock(&hctx->lock);
303 if (!list_empty(&hctx->dispatch))
304 list_splice_init(&hctx->dispatch, &rq_list);
305 spin_unlock(&hctx->lock);
306 }
307
308 /*
309 * Only ask the scheduler for requests, if we didn't have residual
310 * requests from the dispatch list. This is to avoid the case where
311 * we only ever dispatch a fraction of the requests available because
312 * of low device queue depth. Once we pull requests out of the IO
313 * scheduler, we can no longer merge or sort them. So it's best to
314 * leave them there for as long as we can. Mark the hw queue as
315 * needing a restart in that case.
316 *
317 * We want to dispatch from the scheduler if there was nothing
318 * on the dispatch list or we were able to dispatch from the
319 * dispatch list.
320 */
321 if (!list_empty(&rq_list)) {
322 blk_mq_sched_mark_restart_hctx(hctx);
323 if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) {
324 if (has_sched_dispatch)
325 ret = blk_mq_do_dispatch_sched(hctx);
326 else
327 ret = blk_mq_do_dispatch_ctx(hctx);
328 }
329 } else if (has_sched_dispatch) {
330 ret = blk_mq_do_dispatch_sched(hctx);
331 } else if (hctx->dispatch_busy) {
332 /* dequeue request one by one from sw queue if queue is busy */
333 ret = blk_mq_do_dispatch_ctx(hctx);
334 } else {
335 blk_mq_flush_busy_ctxs(hctx, &rq_list);
336 blk_mq_dispatch_rq_list(hctx, &rq_list, 0);
337 }
338
339 return ret;
340}
341
342void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
343{
344 struct request_queue *q = hctx->queue;
345
346 /* RCU or SRCU read lock is needed before checking quiesced flag */
347 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
348 return;
349
350 hctx->run++;
351
352 /*
353 * A return of -EAGAIN is an indication that hctx->dispatch is not
354 * empty and we must run again in order to avoid starving flushes.
355 */
356 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) {
357 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN)
358 blk_mq_run_hw_queue(hctx, true);
359 }
360}
361
362bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
363 unsigned int nr_segs, struct request **merged_request)
364{
365 struct request *rq;
366
367 switch (elv_merge(q, &rq, bio)) {
368 case ELEVATOR_BACK_MERGE:
369 if (!blk_mq_sched_allow_merge(q, rq, bio))
370 return false;
371 if (!bio_attempt_back_merge(rq, bio, nr_segs))
372 return false;
373 *merged_request = attempt_back_merge(q, rq);
374 if (!*merged_request)
375 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
376 return true;
377 case ELEVATOR_FRONT_MERGE:
378 if (!blk_mq_sched_allow_merge(q, rq, bio))
379 return false;
380 if (!bio_attempt_front_merge(rq, bio, nr_segs))
381 return false;
382 *merged_request = attempt_front_merge(q, rq);
383 if (!*merged_request)
384 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
385 return true;
386 case ELEVATOR_DISCARD_MERGE:
387 return bio_attempt_discard_merge(q, rq, bio);
388 default:
389 return false;
390 }
391}
392EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
393
394/*
395 * Iterate list of requests and see if we can merge this bio with any
396 * of them.
397 */
398bool blk_mq_bio_list_merge(struct request_queue *q, struct list_head *list,
399 struct bio *bio, unsigned int nr_segs)
400{
401 struct request *rq;
402 int checked = 8;
403
404 list_for_each_entry_reverse(rq, list, queuelist) {
405 bool merged = false;
406
407 if (!checked--)
408 break;
409
410 if (!blk_rq_merge_ok(rq, bio))
411 continue;
412
413 switch (blk_try_merge(rq, bio)) {
414 case ELEVATOR_BACK_MERGE:
415 if (blk_mq_sched_allow_merge(q, rq, bio))
416 merged = bio_attempt_back_merge(rq, bio,
417 nr_segs);
418 break;
419 case ELEVATOR_FRONT_MERGE:
420 if (blk_mq_sched_allow_merge(q, rq, bio))
421 merged = bio_attempt_front_merge(rq, bio,
422 nr_segs);
423 break;
424 case ELEVATOR_DISCARD_MERGE:
425 merged = bio_attempt_discard_merge(q, rq, bio);
426 break;
427 default:
428 continue;
429 }
430
431 return merged;
432 }
433
434 return false;
435}
436EXPORT_SYMBOL_GPL(blk_mq_bio_list_merge);
437
438/*
439 * Reverse check our software queue for entries that we could potentially
440 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
441 * too much time checking for merges.
442 */
443static bool blk_mq_attempt_merge(struct request_queue *q,
444 struct blk_mq_hw_ctx *hctx,
445 struct blk_mq_ctx *ctx, struct bio *bio,
446 unsigned int nr_segs)
447{
448 enum hctx_type type = hctx->type;
449
450 lockdep_assert_held(&ctx->lock);
451
452 if (blk_mq_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) {
453 ctx->rq_merged++;
454 return true;
455 }
456
457 return false;
458}
459
460bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
461 unsigned int nr_segs)
462{
463 struct elevator_queue *e = q->elevator;
464 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
465 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
466 bool ret = false;
467 enum hctx_type type;
468
469 if (e && e->type->ops.bio_merge)
470 return e->type->ops.bio_merge(hctx, bio, nr_segs);
471
472 type = hctx->type;
473 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
474 !list_empty_careful(&ctx->rq_lists[type])) {
475 /* default per sw-queue merge */
476 spin_lock(&ctx->lock);
477 ret = blk_mq_attempt_merge(q, hctx, ctx, bio, nr_segs);
478 spin_unlock(&ctx->lock);
479 }
480
481 return ret;
482}
483
484bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
485{
486 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
487}
488EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
489
490void blk_mq_sched_request_inserted(struct request *rq)
491{
492 trace_block_rq_insert(rq->q, rq);
493}
494EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
495
496static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
497 bool has_sched,
498 struct request *rq)
499{
500 /*
501 * dispatch flush and passthrough rq directly
502 *
503 * passthrough request has to be added to hctx->dispatch directly.
504 * For some reason, device may be in one situation which can't
505 * handle FS request, so STS_RESOURCE is always returned and the
506 * FS request will be added to hctx->dispatch. However passthrough
507 * request may be required at that time for fixing the problem. If
508 * passthrough request is added to scheduler queue, there isn't any
509 * chance to dispatch it given we prioritize requests in hctx->dispatch.
510 */
511 if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq))
512 return true;
513
514 if (has_sched)
515 rq->rq_flags |= RQF_SORTED;
516
517 return false;
518}
519
520void blk_mq_sched_insert_request(struct request *rq, bool at_head,
521 bool run_queue, bool async)
522{
523 struct request_queue *q = rq->q;
524 struct elevator_queue *e = q->elevator;
525 struct blk_mq_ctx *ctx = rq->mq_ctx;
526 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
527
528 /* flush rq in flush machinery need to be dispatched directly */
529 if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) {
530 blk_insert_flush(rq);
531 goto run;
532 }
533
534 WARN_ON(e && (rq->tag != -1));
535
536 if (blk_mq_sched_bypass_insert(hctx, !!e, rq)) {
537 /*
538 * Firstly normal IO request is inserted to scheduler queue or
539 * sw queue, meantime we add flush request to dispatch queue(
540 * hctx->dispatch) directly and there is at most one in-flight
541 * flush request for each hw queue, so it doesn't matter to add
542 * flush request to tail or front of the dispatch queue.
543 *
544 * Secondly in case of NCQ, flush request belongs to non-NCQ
545 * command, and queueing it will fail when there is any
546 * in-flight normal IO request(NCQ command). When adding flush
547 * rq to the front of hctx->dispatch, it is easier to introduce
548 * extra time to flush rq's latency because of S_SCHED_RESTART
549 * compared with adding to the tail of dispatch queue, then
550 * chance of flush merge is increased, and less flush requests
551 * will be issued to controller. It is observed that ~10% time
552 * is saved in blktests block/004 on disk attached to AHCI/NCQ
553 * drive when adding flush rq to the front of hctx->dispatch.
554 *
555 * Simply queue flush rq to the front of hctx->dispatch so that
556 * intensive flush workloads can benefit in case of NCQ HW.
557 */
558 at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head;
559 blk_mq_request_bypass_insert(rq, at_head, false);
560 goto run;
561 }
562
563 if (e && e->type->ops.insert_requests) {
564 LIST_HEAD(list);
565
566 list_add(&rq->queuelist, &list);
567 e->type->ops.insert_requests(hctx, &list, at_head);
568 } else {
569 spin_lock(&ctx->lock);
570 __blk_mq_insert_request(hctx, rq, at_head);
571 spin_unlock(&ctx->lock);
572 }
573
574run:
575 if (run_queue)
576 blk_mq_run_hw_queue(hctx, async);
577}
578
579void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
580 struct blk_mq_ctx *ctx,
581 struct list_head *list, bool run_queue_async)
582{
583 struct elevator_queue *e;
584 struct request_queue *q = hctx->queue;
585
586 /*
587 * blk_mq_sched_insert_requests() is called from flush plug
588 * context only, and hold one usage counter to prevent queue
589 * from being released.
590 */
591 percpu_ref_get(&q->q_usage_counter);
592
593 e = hctx->queue->elevator;
594 if (e && e->type->ops.insert_requests)
595 e->type->ops.insert_requests(hctx, list, false);
596 else {
597 /*
598 * try to issue requests directly if the hw queue isn't
599 * busy in case of 'none' scheduler, and this way may save
600 * us one extra enqueue & dequeue to sw queue.
601 */
602 if (!hctx->dispatch_busy && !e && !run_queue_async) {
603 blk_mq_try_issue_list_directly(hctx, list);
604 if (list_empty(list))
605 goto out;
606 }
607 blk_mq_insert_requests(hctx, ctx, list);
608 }
609
610 blk_mq_run_hw_queue(hctx, run_queue_async);
611 out:
612 percpu_ref_put(&q->q_usage_counter);
613}
614
615static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
616 struct blk_mq_hw_ctx *hctx,
617 unsigned int hctx_idx)
618{
619 if (hctx->sched_tags) {
620 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
621 blk_mq_free_rq_map(hctx->sched_tags);
622 hctx->sched_tags = NULL;
623 }
624}
625
626static int blk_mq_sched_alloc_tags(struct request_queue *q,
627 struct blk_mq_hw_ctx *hctx,
628 unsigned int hctx_idx)
629{
630 struct blk_mq_tag_set *set = q->tag_set;
631 int ret;
632
633 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
634 set->reserved_tags);
635 if (!hctx->sched_tags)
636 return -ENOMEM;
637
638 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
639 if (ret)
640 blk_mq_sched_free_tags(set, hctx, hctx_idx);
641
642 return ret;
643}
644
645/* called in queue's release handler, tagset has gone away */
646static void blk_mq_sched_tags_teardown(struct request_queue *q)
647{
648 struct blk_mq_hw_ctx *hctx;
649 int i;
650
651 queue_for_each_hw_ctx(q, hctx, i) {
652 if (hctx->sched_tags) {
653 blk_mq_free_rq_map(hctx->sched_tags);
654 hctx->sched_tags = NULL;
655 }
656 }
657}
658
659int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
660{
661 struct blk_mq_hw_ctx *hctx;
662 struct elevator_queue *eq;
663 unsigned int i;
664 int ret;
665
666 if (!e) {
667 q->elevator = NULL;
668 q->nr_requests = q->tag_set->queue_depth;
669 return 0;
670 }
671
672 /*
673 * Default to double of smaller one between hw queue_depth and 128,
674 * since we don't split into sync/async like the old code did.
675 * Additionally, this is a per-hw queue depth.
676 */
677 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
678 BLKDEV_MAX_RQ);
679
680 queue_for_each_hw_ctx(q, hctx, i) {
681 ret = blk_mq_sched_alloc_tags(q, hctx, i);
682 if (ret)
683 goto err;
684 }
685
686 ret = e->ops.init_sched(q, e);
687 if (ret)
688 goto err;
689
690 blk_mq_debugfs_register_sched(q);
691
692 queue_for_each_hw_ctx(q, hctx, i) {
693 if (e->ops.init_hctx) {
694 ret = e->ops.init_hctx(hctx, i);
695 if (ret) {
696 eq = q->elevator;
697 blk_mq_sched_free_requests(q);
698 blk_mq_exit_sched(q, eq);
699 kobject_put(&eq->kobj);
700 return ret;
701 }
702 }
703 blk_mq_debugfs_register_sched_hctx(q, hctx);
704 }
705
706 return 0;
707
708err:
709 blk_mq_sched_free_requests(q);
710 blk_mq_sched_tags_teardown(q);
711 q->elevator = NULL;
712 return ret;
713}
714
715/*
716 * called in either blk_queue_cleanup or elevator_switch, tagset
717 * is required for freeing requests
718 */
719void blk_mq_sched_free_requests(struct request_queue *q)
720{
721 struct blk_mq_hw_ctx *hctx;
722 int i;
723
724 queue_for_each_hw_ctx(q, hctx, i) {
725 if (hctx->sched_tags)
726 blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i);
727 }
728}
729
730void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
731{
732 struct blk_mq_hw_ctx *hctx;
733 unsigned int i;
734
735 queue_for_each_hw_ctx(q, hctx, i) {
736 blk_mq_debugfs_unregister_sched_hctx(hctx);
737 if (e->type->ops.exit_hctx && hctx->sched_data) {
738 e->type->ops.exit_hctx(hctx, i);
739 hctx->sched_data = NULL;
740 }
741 }
742 blk_mq_debugfs_unregister_sched(q);
743 if (e->type->ops.exit_sched)
744 e->type->ops.exit_sched(e);
745 blk_mq_sched_tags_teardown(q);
746 q->elevator = NULL;
747}