1 // SPDX-License-Identifier: GPL-2.0
3 * blk-mq scheduling framework
5 * Copyright (C) 2016 Jens Axboe
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/blk-mq.h>
10 #include <linux/list_sort.h>
12 #include <trace/events/block.h>
16 #include "blk-mq-debugfs.h"
17 #include "blk-mq-sched.h"
18 #include "blk-mq-tag.h"
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.
25 void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx
*hctx
)
27 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
30 set_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
32 EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx
);
34 void __blk_mq_sched_restart(struct blk_mq_hw_ctx
*hctx
)
36 clear_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
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().
47 blk_mq_run_hw_queue(hctx
, true);
50 static int sched_rq_cmp(void *priv
, const struct list_head
*a
,
51 const struct list_head
*b
)
53 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
54 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
56 return rqa
->mq_hctx
> rqb
->mq_hctx
;
59 static bool blk_mq_dispatch_hctx_list(struct list_head
*rq_list
)
61 struct blk_mq_hw_ctx
*hctx
=
62 list_first_entry(rq_list
, struct request
, queuelist
)->mq_hctx
;
65 unsigned int count
= 0;
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
);
74 list_splice_tail_init(rq_list
, &hctx_list
);
77 return blk_mq_dispatch_rq_list(hctx
, &hctx_list
, count
);
80 #define BLK_MQ_BUDGET_DELAY 3 /* ms units */
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.
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.
90 static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx
*hctx
)
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
;
100 if (hctx
->dispatch_busy
)
103 max_dispatch
= hctx
->queue
->nr_requests
;
109 if (e
->type
->ops
.has_work
&& !e
->type
->ops
.has_work(hctx
))
112 if (!list_empty_careful(&hctx
->dispatch
)) {
117 budget_token
= blk_mq_get_dispatch_budget(q
);
118 if (budget_token
< 0)
121 rq
= e
->type
->ops
.dispatch_request(hctx
);
123 blk_mq_put_dispatch_budget(q
, budget_token
);
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
135 blk_mq_set_rq_budget_token(rq
, budget_token
);
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().
142 list_add_tail(&rq
->queuelist
, &rq_list
);
144 if (rq
->mq_hctx
!= hctx
)
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.
153 if (!blk_mq_get_driver_tag(rq
))
155 } while (count
< max_dispatch
);
159 blk_mq_delay_run_hw_queues(q
, BLK_MQ_BUDGET_DELAY
);
160 } else if (multi_hctxs
) {
162 * Requests from different hctx may be dequeued from some
163 * schedulers, such as bfq and deadline.
165 * Sort the requests in the list according to their hctx,
166 * dispatch batching requests from same hctx at a time.
168 list_sort(NULL
, &rq_list
, sched_rq_cmp
);
170 dispatched
|= blk_mq_dispatch_hctx_list(&rq_list
);
171 } while (!list_empty(&rq_list
));
173 dispatched
= blk_mq_dispatch_rq_list(hctx
, &rq_list
, count
);
181 static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx
*hctx
)
183 unsigned long end
= jiffies
+ HZ
;
187 ret
= __blk_mq_do_dispatch_sched(hctx
);
190 if (need_resched() || time_is_before_jiffies(end
)) {
191 blk_mq_delay_run_hw_queue(hctx
, 0);
199 static struct blk_mq_ctx
*blk_mq_next_ctx(struct blk_mq_hw_ctx
*hctx
,
200 struct blk_mq_ctx
*ctx
)
202 unsigned short idx
= ctx
->index_hw
[hctx
->type
];
204 if (++idx
== hctx
->nr_ctx
)
207 return hctx
->ctxs
[idx
];
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.
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.
218 static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx
*hctx
)
220 struct request_queue
*q
= hctx
->queue
;
222 struct blk_mq_ctx
*ctx
= READ_ONCE(hctx
->dispatch_from
);
229 if (!list_empty_careful(&hctx
->dispatch
)) {
234 if (!sbitmap_any_bit_set(&hctx
->ctx_map
))
237 budget_token
= blk_mq_get_dispatch_budget(q
);
238 if (budget_token
< 0)
241 rq
= blk_mq_dequeue_from_ctx(hctx
, ctx
);
243 blk_mq_put_dispatch_budget(q
, budget_token
);
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
251 blk_mq_delay_run_hw_queues(q
, BLK_MQ_BUDGET_DELAY
);
255 blk_mq_set_rq_budget_token(rq
, budget_token
);
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().
262 list_add(&rq
->queuelist
, &rq_list
);
264 /* round robin for fair dispatch */
265 ctx
= blk_mq_next_ctx(hctx
, rq
->mq_ctx
);
267 } while (blk_mq_dispatch_rq_list(rq
->mq_hctx
, &rq_list
, 1));
269 WRITE_ONCE(hctx
->dispatch_from
, ctx
);
273 static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx
*hctx
)
275 struct request_queue
*q
= hctx
->queue
;
276 const bool has_sched
= q
->elevator
;
281 * If we have previous entries on our dispatch list, grab them first for
282 * more fair dispatch.
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
);
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.
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
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)) {
308 ret
= blk_mq_do_dispatch_sched(hctx
);
310 ret
= blk_mq_do_dispatch_ctx(hctx
);
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
);
318 blk_mq_flush_busy_ctxs(hctx
, &rq_list
);
319 blk_mq_dispatch_rq_list(hctx
, &rq_list
, 0);
325 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx
*hctx
)
327 struct request_queue
*q
= hctx
->queue
;
329 /* RCU or SRCU read lock is needed before checking quiesced flag */
330 if (unlikely(blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)))
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.
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);
345 bool blk_mq_sched_bio_merge(struct request_queue
*q
, struct bio
*bio
,
346 unsigned int nr_segs
)
348 struct elevator_queue
*e
= q
->elevator
;
349 struct blk_mq_ctx
*ctx
;
350 struct blk_mq_hw_ctx
*hctx
;
354 if (e
&& e
->type
->ops
.bio_merge
) {
355 ret
= e
->type
->ops
.bio_merge(q
, bio
, nr_segs
);
359 ctx
= blk_mq_get_ctx(q
);
360 hctx
= blk_mq_map_queue(q
, bio
->bi_opf
, ctx
);
362 if (!(hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) ||
363 list_empty_careful(&ctx
->rq_lists
[type
]))
366 /* default per sw-queue merge */
367 spin_lock(&ctx
->lock
);
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.
373 if (blk_bio_list_merge(q
, &ctx
->rq_lists
[type
], bio
, nr_segs
))
376 spin_unlock(&ctx
->lock
);
381 bool blk_mq_sched_try_insert_merge(struct request_queue
*q
, struct request
*rq
,
382 struct list_head
*free
)
384 return rq_mergeable(rq
) && elv_attempt_insert_merge(q
, rq
, free
);
386 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge
);
388 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx
*hctx
,
392 * dispatch flush and passthrough rq directly
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.
402 if ((rq
->rq_flags
& RQF_FLUSH_SEQ
) || blk_rq_is_passthrough(rq
))
408 void blk_mq_sched_insert_request(struct request
*rq
, bool at_head
,
409 bool run_queue
, bool async
)
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
;
416 WARN_ON(e
&& (rq
->tag
!= BLK_MQ_NO_TAG
));
418 if (blk_mq_sched_bypass_insert(hctx
, rq
)) {
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.
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.
437 * Simply queue flush rq to the front of hctx->dispatch so that
438 * intensive flush workloads can benefit in case of NCQ HW.
440 at_head
= (rq
->rq_flags
& RQF_FLUSH_SEQ
) ? true : at_head
;
441 blk_mq_request_bypass_insert(rq
, at_head
, false);
448 list_add(&rq
->queuelist
, &list
);
449 e
->type
->ops
.insert_requests(hctx
, &list
, at_head
);
451 spin_lock(&ctx
->lock
);
452 __blk_mq_insert_request(hctx
, rq
, at_head
);
453 spin_unlock(&ctx
->lock
);
458 blk_mq_run_hw_queue(hctx
, async
);
461 void 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
)
465 struct elevator_queue
*e
;
466 struct request_queue
*q
= hctx
->queue
;
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.
473 percpu_ref_get(&q
->q_usage_counter
);
475 e
= hctx
->queue
->elevator
;
477 e
->type
->ops
.insert_requests(hctx
, list
, false);
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.
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
))
490 blk_mq_insert_requests(hctx
, ctx
, list
);
493 blk_mq_run_hw_queue(hctx
, run_queue_async
);
495 percpu_ref_put(&q
->q_usage_counter
);
498 static int blk_mq_sched_alloc_map_and_rqs(struct request_queue
*q
,
499 struct blk_mq_hw_ctx
*hctx
,
500 unsigned int hctx_idx
)
502 if (blk_mq_is_shared_tags(q
->tag_set
->flags
)) {
503 hctx
->sched_tags
= q
->sched_shared_tags
;
507 hctx
->sched_tags
= blk_mq_alloc_map_and_rqs(q
->tag_set
, hctx_idx
,
510 if (!hctx
->sched_tags
)
515 static void blk_mq_exit_sched_shared_tags(struct request_queue
*queue
)
517 blk_mq_free_rq_map(queue
->sched_shared_tags
);
518 queue
->sched_shared_tags
= NULL
;
521 /* called in queue's release handler, tagset has gone away */
522 static void blk_mq_sched_tags_teardown(struct request_queue
*q
, unsigned int flags
)
524 struct blk_mq_hw_ctx
*hctx
;
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
;
535 if (blk_mq_is_shared_tags(flags
))
536 blk_mq_exit_sched_shared_tags(q
);
539 static int blk_mq_init_sched_shared_tags(struct request_queue
*queue
)
541 struct blk_mq_tag_set
*set
= queue
->tag_set
;
544 * Set initial depth at max so that we don't need to reallocate for
545 * updating nr_requests.
547 queue
->sched_shared_tags
= blk_mq_alloc_map_and_rqs(set
,
550 if (!queue
->sched_shared_tags
)
553 blk_mq_tag_update_sched_shared_tags(queue
);
558 int blk_mq_init_sched(struct request_queue
*q
, struct elevator_type
*e
)
560 unsigned int flags
= q
->tag_set
->flags
;
561 struct blk_mq_hw_ctx
*hctx
;
562 struct elevator_queue
*eq
;
567 blk_queue_flag_clear(QUEUE_FLAG_SQ_SCHED
, q
);
569 q
->nr_requests
= q
->tag_set
->queue_depth
;
574 * Default to double of smaller one between hw queue_depth and 128,
575 * since we don't split into sync/async like the old code did.
576 * Additionally, this is a per-hw queue depth.
578 q
->nr_requests
= 2 * min_t(unsigned int, q
->tag_set
->queue_depth
,
581 if (blk_mq_is_shared_tags(flags
)) {
582 ret
= blk_mq_init_sched_shared_tags(q
);
587 queue_for_each_hw_ctx(q
, hctx
, i
) {
588 ret
= blk_mq_sched_alloc_map_and_rqs(q
, hctx
, i
);
590 goto err_free_map_and_rqs
;
593 ret
= e
->ops
.init_sched(q
, e
);
595 goto err_free_map_and_rqs
;
597 mutex_lock(&q
->debugfs_mutex
);
598 blk_mq_debugfs_register_sched(q
);
599 mutex_unlock(&q
->debugfs_mutex
);
601 queue_for_each_hw_ctx(q
, hctx
, i
) {
602 if (e
->ops
.init_hctx
) {
603 ret
= e
->ops
.init_hctx(hctx
, i
);
606 blk_mq_sched_free_rqs(q
);
607 blk_mq_exit_sched(q
, eq
);
608 kobject_put(&eq
->kobj
);
612 mutex_lock(&q
->debugfs_mutex
);
613 blk_mq_debugfs_register_sched_hctx(q
, hctx
);
614 mutex_unlock(&q
->debugfs_mutex
);
619 err_free_map_and_rqs
:
620 blk_mq_sched_free_rqs(q
);
621 blk_mq_sched_tags_teardown(q
, flags
);
628 * called in either blk_queue_cleanup or elevator_switch, tagset
629 * is required for freeing requests
631 void blk_mq_sched_free_rqs(struct request_queue
*q
)
633 struct blk_mq_hw_ctx
*hctx
;
636 if (blk_mq_is_shared_tags(q
->tag_set
->flags
)) {
637 blk_mq_free_rqs(q
->tag_set
, q
->sched_shared_tags
,
640 queue_for_each_hw_ctx(q
, hctx
, i
) {
641 if (hctx
->sched_tags
)
642 blk_mq_free_rqs(q
->tag_set
,
643 hctx
->sched_tags
, i
);
648 void blk_mq_exit_sched(struct request_queue
*q
, struct elevator_queue
*e
)
650 struct blk_mq_hw_ctx
*hctx
;
652 unsigned int flags
= 0;
654 queue_for_each_hw_ctx(q
, hctx
, i
) {
655 mutex_lock(&q
->debugfs_mutex
);
656 blk_mq_debugfs_unregister_sched_hctx(hctx
);
657 mutex_unlock(&q
->debugfs_mutex
);
659 if (e
->type
->ops
.exit_hctx
&& hctx
->sched_data
) {
660 e
->type
->ops
.exit_hctx(hctx
, i
);
661 hctx
->sched_data
= NULL
;
666 mutex_lock(&q
->debugfs_mutex
);
667 blk_mq_debugfs_unregister_sched(q
);
668 mutex_unlock(&q
->debugfs_mutex
);
670 if (e
->type
->ops
.exit_sched
)
671 e
->type
->ops
.exit_sched(e
);
672 blk_mq_sched_tags_teardown(q
, flags
);