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"
21 void blk_mq_sched_assign_ioc(struct request
*rq
)
23 struct request_queue
*q
= rq
->q
;
24 struct io_context
*ioc
;
28 * May not have an IO context if it's a passthrough request
30 ioc
= current
->io_context
;
34 spin_lock_irq(&q
->queue_lock
);
35 icq
= ioc_lookup_icq(ioc
, q
);
36 spin_unlock_irq(&q
->queue_lock
);
39 icq
= ioc_create_icq(ioc
, q
, GFP_ATOMIC
);
43 get_io_context(icq
->ioc
);
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.
51 void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx
*hctx
)
53 if (test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
56 set_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
58 EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx
);
60 void blk_mq_sched_restart(struct blk_mq_hw_ctx
*hctx
)
62 if (!test_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
))
64 clear_bit(BLK_MQ_S_SCHED_RESTART
, &hctx
->state
);
67 * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch)
68 * in blk_mq_run_hw_queue(). Its pair is the barrier in
69 * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART,
70 * meantime new request added to hctx->dispatch is missed to check in
71 * blk_mq_run_hw_queue().
75 blk_mq_run_hw_queue(hctx
, true);
78 static int sched_rq_cmp(void *priv
, const struct list_head
*a
,
79 const struct list_head
*b
)
81 struct request
*rqa
= container_of(a
, struct request
, queuelist
);
82 struct request
*rqb
= container_of(b
, struct request
, queuelist
);
84 return rqa
->mq_hctx
> rqb
->mq_hctx
;
87 static bool blk_mq_dispatch_hctx_list(struct list_head
*rq_list
)
89 struct blk_mq_hw_ctx
*hctx
=
90 list_first_entry(rq_list
, struct request
, queuelist
)->mq_hctx
;
93 unsigned int count
= 0;
95 list_for_each_entry(rq
, rq_list
, queuelist
) {
96 if (rq
->mq_hctx
!= hctx
) {
97 list_cut_before(&hctx_list
, rq_list
, &rq
->queuelist
);
102 list_splice_tail_init(rq_list
, &hctx_list
);
105 return blk_mq_dispatch_rq_list(hctx
, &hctx_list
, count
);
108 #define BLK_MQ_BUDGET_DELAY 3 /* ms units */
111 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
112 * its queue by itself in its completion handler, so we don't need to
113 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
115 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
116 * be run again. This is necessary to avoid starving flushes.
118 static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx
*hctx
)
120 struct request_queue
*q
= hctx
->queue
;
121 struct elevator_queue
*e
= q
->elevator
;
122 bool multi_hctxs
= false, run_queue
= false;
123 bool dispatched
= false, busy
= false;
124 unsigned int max_dispatch
;
128 if (hctx
->dispatch_busy
)
131 max_dispatch
= hctx
->queue
->nr_requests
;
136 if (e
->type
->ops
.has_work
&& !e
->type
->ops
.has_work(hctx
))
139 if (!list_empty_careful(&hctx
->dispatch
)) {
144 if (!blk_mq_get_dispatch_budget(q
))
147 rq
= e
->type
->ops
.dispatch_request(hctx
);
149 blk_mq_put_dispatch_budget(q
);
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
162 * Now this rq owns the budget which has to be released
163 * if this rq won't be queued to driver via .queue_rq()
164 * in blk_mq_dispatch_rq_list().
166 list_add_tail(&rq
->queuelist
, &rq_list
);
167 if (rq
->mq_hctx
!= hctx
)
169 } while (++count
< max_dispatch
);
173 blk_mq_delay_run_hw_queues(q
, BLK_MQ_BUDGET_DELAY
);
174 } else if (multi_hctxs
) {
176 * Requests from different hctx may be dequeued from some
177 * schedulers, such as bfq and deadline.
179 * Sort the requests in the list according to their hctx,
180 * dispatch batching requests from same hctx at a time.
182 list_sort(NULL
, &rq_list
, sched_rq_cmp
);
184 dispatched
|= blk_mq_dispatch_hctx_list(&rq_list
);
185 } while (!list_empty(&rq_list
));
187 dispatched
= blk_mq_dispatch_rq_list(hctx
, &rq_list
, count
);
195 static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx
*hctx
)
200 ret
= __blk_mq_do_dispatch_sched(hctx
);
206 static struct blk_mq_ctx
*blk_mq_next_ctx(struct blk_mq_hw_ctx
*hctx
,
207 struct blk_mq_ctx
*ctx
)
209 unsigned short idx
= ctx
->index_hw
[hctx
->type
];
211 if (++idx
== hctx
->nr_ctx
)
214 return hctx
->ctxs
[idx
];
218 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
219 * its queue by itself in its completion handler, so we don't need to
220 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
222 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
223 * be run again. This is necessary to avoid starving flushes.
225 static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx
*hctx
)
227 struct request_queue
*q
= hctx
->queue
;
229 struct blk_mq_ctx
*ctx
= READ_ONCE(hctx
->dispatch_from
);
234 if (!list_empty_careful(&hctx
->dispatch
)) {
239 if (!sbitmap_any_bit_set(&hctx
->ctx_map
))
242 if (!blk_mq_get_dispatch_budget(q
))
245 rq
= blk_mq_dequeue_from_ctx(hctx
, ctx
);
247 blk_mq_put_dispatch_budget(q
);
249 * We're releasing without dispatching. Holding the
250 * budget could have blocked any "hctx"s with the
251 * same queue and if we didn't dispatch then there's
252 * no guarantee anyone will kick the queue. Kick it
255 blk_mq_delay_run_hw_queues(q
, BLK_MQ_BUDGET_DELAY
);
260 * Now this rq owns the budget which has to be released
261 * if this rq won't be queued to driver via .queue_rq()
262 * in blk_mq_dispatch_rq_list().
264 list_add(&rq
->queuelist
, &rq_list
);
266 /* round robin for fair dispatch */
267 ctx
= blk_mq_next_ctx(hctx
, rq
->mq_ctx
);
269 } while (blk_mq_dispatch_rq_list(rq
->mq_hctx
, &rq_list
, 1));
271 WRITE_ONCE(hctx
->dispatch_from
, ctx
);
275 static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx
*hctx
)
277 struct request_queue
*q
= hctx
->queue
;
278 struct elevator_queue
*e
= q
->elevator
;
279 const bool has_sched_dispatch
= e
&& e
->type
->ops
.dispatch_request
;
284 * If we have previous entries on our dispatch list, grab them first for
285 * more fair dispatch.
287 if (!list_empty_careful(&hctx
->dispatch
)) {
288 spin_lock(&hctx
->lock
);
289 if (!list_empty(&hctx
->dispatch
))
290 list_splice_init(&hctx
->dispatch
, &rq_list
);
291 spin_unlock(&hctx
->lock
);
295 * Only ask the scheduler for requests, if we didn't have residual
296 * requests from the dispatch list. This is to avoid the case where
297 * we only ever dispatch a fraction of the requests available because
298 * of low device queue depth. Once we pull requests out of the IO
299 * scheduler, we can no longer merge or sort them. So it's best to
300 * leave them there for as long as we can. Mark the hw queue as
301 * needing a restart in that case.
303 * We want to dispatch from the scheduler if there was nothing
304 * on the dispatch list or we were able to dispatch from the
307 if (!list_empty(&rq_list
)) {
308 blk_mq_sched_mark_restart_hctx(hctx
);
309 if (blk_mq_dispatch_rq_list(hctx
, &rq_list
, 0)) {
310 if (has_sched_dispatch
)
311 ret
= blk_mq_do_dispatch_sched(hctx
);
313 ret
= blk_mq_do_dispatch_ctx(hctx
);
315 } else if (has_sched_dispatch
) {
316 ret
= blk_mq_do_dispatch_sched(hctx
);
317 } else if (hctx
->dispatch_busy
) {
318 /* dequeue request one by one from sw queue if queue is busy */
319 ret
= blk_mq_do_dispatch_ctx(hctx
);
321 blk_mq_flush_busy_ctxs(hctx
, &rq_list
);
322 blk_mq_dispatch_rq_list(hctx
, &rq_list
, 0);
328 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx
*hctx
)
330 struct request_queue
*q
= hctx
->queue
;
332 /* RCU or SRCU read lock is needed before checking quiesced flag */
333 if (unlikely(blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)))
339 * A return of -EAGAIN is an indication that hctx->dispatch is not
340 * empty and we must run again in order to avoid starving flushes.
342 if (__blk_mq_sched_dispatch_requests(hctx
) == -EAGAIN
) {
343 if (__blk_mq_sched_dispatch_requests(hctx
) == -EAGAIN
)
344 blk_mq_run_hw_queue(hctx
, true);
348 bool __blk_mq_sched_bio_merge(struct request_queue
*q
, struct bio
*bio
,
349 unsigned int nr_segs
)
351 struct elevator_queue
*e
= q
->elevator
;
352 struct blk_mq_ctx
*ctx
;
353 struct blk_mq_hw_ctx
*hctx
;
357 if (e
&& e
->type
->ops
.bio_merge
)
358 return e
->type
->ops
.bio_merge(q
, bio
, nr_segs
);
360 ctx
= blk_mq_get_ctx(q
);
361 hctx
= blk_mq_map_queue(q
, bio
->bi_opf
, ctx
);
363 if (!(hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) ||
364 list_empty_careful(&ctx
->rq_lists
[type
]))
367 /* default per sw-queue merge */
368 spin_lock(&ctx
->lock
);
370 * Reverse check our software queue for entries that we could
371 * potentially merge with. Currently includes a hand-wavy stop
372 * count of 8, to not spend too much time checking for merges.
374 if (blk_bio_list_merge(q
, &ctx
->rq_lists
[type
], bio
, nr_segs
)) {
379 spin_unlock(&ctx
->lock
);
384 bool blk_mq_sched_try_insert_merge(struct request_queue
*q
, struct request
*rq
)
386 return rq_mergeable(rq
) && elv_attempt_insert_merge(q
, rq
);
388 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge
);
390 void blk_mq_sched_request_inserted(struct request
*rq
)
392 trace_block_rq_insert(rq
);
394 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted
);
396 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx
*hctx
,
401 * dispatch flush and passthrough rq directly
403 * passthrough request has to be added to hctx->dispatch directly.
404 * For some reason, device may be in one situation which can't
405 * handle FS request, so STS_RESOURCE is always returned and the
406 * FS request will be added to hctx->dispatch. However passthrough
407 * request may be required at that time for fixing the problem. If
408 * passthrough request is added to scheduler queue, there isn't any
409 * chance to dispatch it given we prioritize requests in hctx->dispatch.
411 if ((rq
->rq_flags
& RQF_FLUSH_SEQ
) || blk_rq_is_passthrough(rq
))
415 rq
->rq_flags
|= RQF_SORTED
;
420 void blk_mq_sched_insert_request(struct request
*rq
, bool at_head
,
421 bool run_queue
, bool async
)
423 struct request_queue
*q
= rq
->q
;
424 struct elevator_queue
*e
= q
->elevator
;
425 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
426 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
428 WARN_ON(e
&& (rq
->tag
!= BLK_MQ_NO_TAG
));
430 if (blk_mq_sched_bypass_insert(hctx
, !!e
, rq
)) {
432 * Firstly normal IO request is inserted to scheduler queue or
433 * sw queue, meantime we add flush request to dispatch queue(
434 * hctx->dispatch) directly and there is at most one in-flight
435 * flush request for each hw queue, so it doesn't matter to add
436 * flush request to tail or front of the dispatch queue.
438 * Secondly in case of NCQ, flush request belongs to non-NCQ
439 * command, and queueing it will fail when there is any
440 * in-flight normal IO request(NCQ command). When adding flush
441 * rq to the front of hctx->dispatch, it is easier to introduce
442 * extra time to flush rq's latency because of S_SCHED_RESTART
443 * compared with adding to the tail of dispatch queue, then
444 * chance of flush merge is increased, and less flush requests
445 * will be issued to controller. It is observed that ~10% time
446 * is saved in blktests block/004 on disk attached to AHCI/NCQ
447 * drive when adding flush rq to the front of hctx->dispatch.
449 * Simply queue flush rq to the front of hctx->dispatch so that
450 * intensive flush workloads can benefit in case of NCQ HW.
452 at_head
= (rq
->rq_flags
& RQF_FLUSH_SEQ
) ? true : at_head
;
453 blk_mq_request_bypass_insert(rq
, at_head
, false);
457 if (e
&& e
->type
->ops
.insert_requests
) {
460 list_add(&rq
->queuelist
, &list
);
461 e
->type
->ops
.insert_requests(hctx
, &list
, at_head
);
463 spin_lock(&ctx
->lock
);
464 __blk_mq_insert_request(hctx
, rq
, at_head
);
465 spin_unlock(&ctx
->lock
);
470 blk_mq_run_hw_queue(hctx
, async
);
473 void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx
*hctx
,
474 struct blk_mq_ctx
*ctx
,
475 struct list_head
*list
, bool run_queue_async
)
477 struct elevator_queue
*e
;
478 struct request_queue
*q
= hctx
->queue
;
481 * blk_mq_sched_insert_requests() is called from flush plug
482 * context only, and hold one usage counter to prevent queue
483 * from being released.
485 percpu_ref_get(&q
->q_usage_counter
);
487 e
= hctx
->queue
->elevator
;
488 if (e
&& e
->type
->ops
.insert_requests
)
489 e
->type
->ops
.insert_requests(hctx
, list
, false);
492 * try to issue requests directly if the hw queue isn't
493 * busy in case of 'none' scheduler, and this way may save
494 * us one extra enqueue & dequeue to sw queue.
496 if (!hctx
->dispatch_busy
&& !e
&& !run_queue_async
) {
497 blk_mq_try_issue_list_directly(hctx
, list
);
498 if (list_empty(list
))
501 blk_mq_insert_requests(hctx
, ctx
, list
);
504 blk_mq_run_hw_queue(hctx
, run_queue_async
);
506 percpu_ref_put(&q
->q_usage_counter
);
509 static void blk_mq_sched_free_tags(struct blk_mq_tag_set
*set
,
510 struct blk_mq_hw_ctx
*hctx
,
511 unsigned int hctx_idx
)
513 unsigned int flags
= set
->flags
& ~BLK_MQ_F_TAG_HCTX_SHARED
;
515 if (hctx
->sched_tags
) {
516 blk_mq_free_rqs(set
, hctx
->sched_tags
, hctx_idx
);
517 blk_mq_free_rq_map(hctx
->sched_tags
, flags
);
518 hctx
->sched_tags
= NULL
;
522 static int blk_mq_sched_alloc_tags(struct request_queue
*q
,
523 struct blk_mq_hw_ctx
*hctx
,
524 unsigned int hctx_idx
)
526 struct blk_mq_tag_set
*set
= q
->tag_set
;
527 /* Clear HCTX_SHARED so tags are init'ed */
528 unsigned int flags
= set
->flags
& ~BLK_MQ_F_TAG_HCTX_SHARED
;
531 hctx
->sched_tags
= blk_mq_alloc_rq_map(set
, hctx_idx
, q
->nr_requests
,
532 set
->reserved_tags
, flags
);
533 if (!hctx
->sched_tags
)
536 ret
= blk_mq_alloc_rqs(set
, hctx
->sched_tags
, hctx_idx
, q
->nr_requests
);
538 blk_mq_sched_free_tags(set
, hctx
, hctx_idx
);
543 /* called in queue's release handler, tagset has gone away */
544 static void blk_mq_sched_tags_teardown(struct request_queue
*q
)
546 struct blk_mq_hw_ctx
*hctx
;
549 queue_for_each_hw_ctx(q
, hctx
, i
) {
550 /* Clear HCTX_SHARED so tags are freed */
551 unsigned int flags
= hctx
->flags
& ~BLK_MQ_F_TAG_HCTX_SHARED
;
553 if (hctx
->sched_tags
) {
554 blk_mq_free_rq_map(hctx
->sched_tags
, flags
);
555 hctx
->sched_tags
= NULL
;
560 int blk_mq_init_sched(struct request_queue
*q
, struct elevator_type
*e
)
562 struct blk_mq_hw_ctx
*hctx
;
563 struct elevator_queue
*eq
;
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 queue_for_each_hw_ctx(q
, hctx
, i
) {
582 ret
= blk_mq_sched_alloc_tags(q
, hctx
, i
);
587 ret
= e
->ops
.init_sched(q
, e
);
591 blk_mq_debugfs_register_sched(q
);
593 queue_for_each_hw_ctx(q
, hctx
, i
) {
594 if (e
->ops
.init_hctx
) {
595 ret
= e
->ops
.init_hctx(hctx
, i
);
598 blk_mq_sched_free_requests(q
);
599 blk_mq_exit_sched(q
, eq
);
600 kobject_put(&eq
->kobj
);
604 blk_mq_debugfs_register_sched_hctx(q
, hctx
);
610 blk_mq_sched_free_requests(q
);
611 blk_mq_sched_tags_teardown(q
);
617 * called in either blk_queue_cleanup or elevator_switch, tagset
618 * is required for freeing requests
620 void blk_mq_sched_free_requests(struct request_queue
*q
)
622 struct blk_mq_hw_ctx
*hctx
;
625 queue_for_each_hw_ctx(q
, hctx
, i
) {
626 if (hctx
->sched_tags
)
627 blk_mq_free_rqs(q
->tag_set
, hctx
->sched_tags
, i
);
631 void blk_mq_exit_sched(struct request_queue
*q
, struct elevator_queue
*e
)
633 struct blk_mq_hw_ctx
*hctx
;
636 queue_for_each_hw_ctx(q
, hctx
, i
) {
637 blk_mq_debugfs_unregister_sched_hctx(hctx
);
638 if (e
->type
->ops
.exit_hctx
&& hctx
->sched_data
) {
639 e
->type
->ops
.exit_hctx(hctx
, i
);
640 hctx
->sched_data
= NULL
;
643 blk_mq_debugfs_unregister_sched(q
);
644 if (e
->type
->ops
.exit_sched
)
645 e
->type
->ops
.exit_sched(e
);
646 blk_mq_sched_tags_teardown(q
);