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
;
137 if (e
->type
->ops
.has_work
&& !e
->type
->ops
.has_work(hctx
))
140 if (!list_empty_careful(&hctx
->dispatch
)) {
145 budget_token
= blk_mq_get_dispatch_budget(q
);
146 if (budget_token
< 0)
149 rq
= e
->type
->ops
.dispatch_request(hctx
);
151 blk_mq_put_dispatch_budget(q
, budget_token
);
153 * We're releasing without dispatching. Holding the
154 * budget could have blocked any "hctx"s with the
155 * same queue and if we didn't dispatch then there's
156 * no guarantee anyone will kick the queue. Kick it
163 blk_mq_set_rq_budget_token(rq
, budget_token
);
166 * Now this rq owns the budget which has to be released
167 * if this rq won't be queued to driver via .queue_rq()
168 * in blk_mq_dispatch_rq_list().
170 list_add_tail(&rq
->queuelist
, &rq_list
);
172 if (rq
->mq_hctx
!= hctx
)
176 * If we cannot get tag for the request, stop dequeueing
177 * requests from the IO scheduler. We are unlikely to be able
178 * to submit them anyway and it creates false impression for
179 * scheduling heuristics that the device can take more IO.
181 if (!blk_mq_get_driver_tag(rq
))
183 } while (count
< max_dispatch
);
187 blk_mq_delay_run_hw_queues(q
, BLK_MQ_BUDGET_DELAY
);
188 } else if (multi_hctxs
) {
190 * Requests from different hctx may be dequeued from some
191 * schedulers, such as bfq and deadline.
193 * Sort the requests in the list according to their hctx,
194 * dispatch batching requests from same hctx at a time.
196 list_sort(NULL
, &rq_list
, sched_rq_cmp
);
198 dispatched
|= blk_mq_dispatch_hctx_list(&rq_list
);
199 } while (!list_empty(&rq_list
));
201 dispatched
= blk_mq_dispatch_rq_list(hctx
, &rq_list
, count
);
209 static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx
*hctx
)
214 ret
= __blk_mq_do_dispatch_sched(hctx
);
220 static struct blk_mq_ctx
*blk_mq_next_ctx(struct blk_mq_hw_ctx
*hctx
,
221 struct blk_mq_ctx
*ctx
)
223 unsigned short idx
= ctx
->index_hw
[hctx
->type
];
225 if (++idx
== hctx
->nr_ctx
)
228 return hctx
->ctxs
[idx
];
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.
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.
239 static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx
*hctx
)
241 struct request_queue
*q
= hctx
->queue
;
243 struct blk_mq_ctx
*ctx
= READ_ONCE(hctx
->dispatch_from
);
250 if (!list_empty_careful(&hctx
->dispatch
)) {
255 if (!sbitmap_any_bit_set(&hctx
->ctx_map
))
258 budget_token
= blk_mq_get_dispatch_budget(q
);
259 if (budget_token
< 0)
262 rq
= blk_mq_dequeue_from_ctx(hctx
, ctx
);
264 blk_mq_put_dispatch_budget(q
, budget_token
);
266 * We're releasing without dispatching. Holding the
267 * budget could have blocked any "hctx"s with the
268 * same queue and if we didn't dispatch then there's
269 * no guarantee anyone will kick the queue. Kick it
272 blk_mq_delay_run_hw_queues(q
, BLK_MQ_BUDGET_DELAY
);
276 blk_mq_set_rq_budget_token(rq
, budget_token
);
279 * Now this rq owns the budget which has to be released
280 * if this rq won't be queued to driver via .queue_rq()
281 * in blk_mq_dispatch_rq_list().
283 list_add(&rq
->queuelist
, &rq_list
);
285 /* round robin for fair dispatch */
286 ctx
= blk_mq_next_ctx(hctx
, rq
->mq_ctx
);
288 } while (blk_mq_dispatch_rq_list(rq
->mq_hctx
, &rq_list
, 1));
290 WRITE_ONCE(hctx
->dispatch_from
, ctx
);
294 static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx
*hctx
)
296 struct request_queue
*q
= hctx
->queue
;
297 const bool has_sched
= q
->elevator
;
302 * If we have previous entries on our dispatch list, grab them first for
303 * more fair dispatch.
305 if (!list_empty_careful(&hctx
->dispatch
)) {
306 spin_lock(&hctx
->lock
);
307 if (!list_empty(&hctx
->dispatch
))
308 list_splice_init(&hctx
->dispatch
, &rq_list
);
309 spin_unlock(&hctx
->lock
);
313 * Only ask the scheduler for requests, if we didn't have residual
314 * requests from the dispatch list. This is to avoid the case where
315 * we only ever dispatch a fraction of the requests available because
316 * of low device queue depth. Once we pull requests out of the IO
317 * scheduler, we can no longer merge or sort them. So it's best to
318 * leave them there for as long as we can. Mark the hw queue as
319 * needing a restart in that case.
321 * We want to dispatch from the scheduler if there was nothing
322 * on the dispatch list or we were able to dispatch from the
325 if (!list_empty(&rq_list
)) {
326 blk_mq_sched_mark_restart_hctx(hctx
);
327 if (blk_mq_dispatch_rq_list(hctx
, &rq_list
, 0)) {
329 ret
= blk_mq_do_dispatch_sched(hctx
);
331 ret
= blk_mq_do_dispatch_ctx(hctx
);
333 } else if (has_sched
) {
334 ret
= blk_mq_do_dispatch_sched(hctx
);
335 } else if (hctx
->dispatch_busy
) {
336 /* dequeue request one by one from sw queue if queue is busy */
337 ret
= blk_mq_do_dispatch_ctx(hctx
);
339 blk_mq_flush_busy_ctxs(hctx
, &rq_list
);
340 blk_mq_dispatch_rq_list(hctx
, &rq_list
, 0);
346 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx
*hctx
)
348 struct request_queue
*q
= hctx
->queue
;
350 /* RCU or SRCU read lock is needed before checking quiesced flag */
351 if (unlikely(blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)))
357 * A return of -EAGAIN is an indication that hctx->dispatch is not
358 * empty and we must run again in order to avoid starving flushes.
360 if (__blk_mq_sched_dispatch_requests(hctx
) == -EAGAIN
) {
361 if (__blk_mq_sched_dispatch_requests(hctx
) == -EAGAIN
)
362 blk_mq_run_hw_queue(hctx
, true);
366 bool __blk_mq_sched_bio_merge(struct request_queue
*q
, struct bio
*bio
,
367 unsigned int nr_segs
)
369 struct elevator_queue
*e
= q
->elevator
;
370 struct blk_mq_ctx
*ctx
;
371 struct blk_mq_hw_ctx
*hctx
;
375 if (e
&& e
->type
->ops
.bio_merge
)
376 return e
->type
->ops
.bio_merge(q
, bio
, nr_segs
);
378 ctx
= blk_mq_get_ctx(q
);
379 hctx
= blk_mq_map_queue(q
, bio
->bi_opf
, ctx
);
381 if (!(hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) ||
382 list_empty_careful(&ctx
->rq_lists
[type
]))
385 /* default per sw-queue merge */
386 spin_lock(&ctx
->lock
);
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.
392 if (blk_bio_list_merge(q
, &ctx
->rq_lists
[type
], bio
, nr_segs
)) {
397 spin_unlock(&ctx
->lock
);
402 bool blk_mq_sched_try_insert_merge(struct request_queue
*q
, struct request
*rq
,
403 struct list_head
*free
)
405 return rq_mergeable(rq
) && elv_attempt_insert_merge(q
, rq
, free
);
407 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge
);
409 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx
*hctx
,
413 * dispatch flush and passthrough rq directly
415 * passthrough request has to be added to hctx->dispatch directly.
416 * For some reason, device may be in one situation which can't
417 * handle FS request, so STS_RESOURCE is always returned and the
418 * FS request will be added to hctx->dispatch. However passthrough
419 * request may be required at that time for fixing the problem. If
420 * passthrough request is added to scheduler queue, there isn't any
421 * chance to dispatch it given we prioritize requests in hctx->dispatch.
423 if ((rq
->rq_flags
& RQF_FLUSH_SEQ
) || blk_rq_is_passthrough(rq
))
429 void blk_mq_sched_insert_request(struct request
*rq
, bool at_head
,
430 bool run_queue
, bool async
)
432 struct request_queue
*q
= rq
->q
;
433 struct elevator_queue
*e
= q
->elevator
;
434 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
435 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
437 WARN_ON(e
&& (rq
->tag
!= BLK_MQ_NO_TAG
));
439 if (blk_mq_sched_bypass_insert(hctx
, rq
)) {
441 * Firstly normal IO request is inserted to scheduler queue or
442 * sw queue, meantime we add flush request to dispatch queue(
443 * hctx->dispatch) directly and there is at most one in-flight
444 * flush request for each hw queue, so it doesn't matter to add
445 * flush request to tail or front of the dispatch queue.
447 * Secondly in case of NCQ, flush request belongs to non-NCQ
448 * command, and queueing it will fail when there is any
449 * in-flight normal IO request(NCQ command). When adding flush
450 * rq to the front of hctx->dispatch, it is easier to introduce
451 * extra time to flush rq's latency because of S_SCHED_RESTART
452 * compared with adding to the tail of dispatch queue, then
453 * chance of flush merge is increased, and less flush requests
454 * will be issued to controller. It is observed that ~10% time
455 * is saved in blktests block/004 on disk attached to AHCI/NCQ
456 * drive when adding flush rq to the front of hctx->dispatch.
458 * Simply queue flush rq to the front of hctx->dispatch so that
459 * intensive flush workloads can benefit in case of NCQ HW.
461 at_head
= (rq
->rq_flags
& RQF_FLUSH_SEQ
) ? true : at_head
;
462 blk_mq_request_bypass_insert(rq
, at_head
, false);
469 list_add(&rq
->queuelist
, &list
);
470 e
->type
->ops
.insert_requests(hctx
, &list
, at_head
);
472 spin_lock(&ctx
->lock
);
473 __blk_mq_insert_request(hctx
, rq
, at_head
);
474 spin_unlock(&ctx
->lock
);
479 blk_mq_run_hw_queue(hctx
, async
);
482 void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx
*hctx
,
483 struct blk_mq_ctx
*ctx
,
484 struct list_head
*list
, bool run_queue_async
)
486 struct elevator_queue
*e
;
487 struct request_queue
*q
= hctx
->queue
;
490 * blk_mq_sched_insert_requests() is called from flush plug
491 * context only, and hold one usage counter to prevent queue
492 * from being released.
494 percpu_ref_get(&q
->q_usage_counter
);
496 e
= hctx
->queue
->elevator
;
498 e
->type
->ops
.insert_requests(hctx
, list
, false);
501 * try to issue requests directly if the hw queue isn't
502 * busy in case of 'none' scheduler, and this way may save
503 * us one extra enqueue & dequeue to sw queue.
505 if (!hctx
->dispatch_busy
&& !e
&& !run_queue_async
) {
506 blk_mq_try_issue_list_directly(hctx
, list
);
507 if (list_empty(list
))
510 blk_mq_insert_requests(hctx
, ctx
, list
);
513 blk_mq_run_hw_queue(hctx
, run_queue_async
);
515 percpu_ref_put(&q
->q_usage_counter
);
518 static int blk_mq_sched_alloc_tags(struct request_queue
*q
,
519 struct blk_mq_hw_ctx
*hctx
,
520 unsigned int hctx_idx
)
522 struct blk_mq_tag_set
*set
= q
->tag_set
;
525 hctx
->sched_tags
= blk_mq_alloc_rq_map(set
, hctx_idx
, q
->nr_requests
,
526 set
->reserved_tags
, set
->flags
);
527 if (!hctx
->sched_tags
)
530 ret
= blk_mq_alloc_rqs(set
, hctx
->sched_tags
, hctx_idx
, q
->nr_requests
);
532 blk_mq_free_rq_map(hctx
->sched_tags
, set
->flags
);
533 hctx
->sched_tags
= NULL
;
539 /* called in queue's release handler, tagset has gone away */
540 static void blk_mq_sched_tags_teardown(struct request_queue
*q
)
542 struct blk_mq_hw_ctx
*hctx
;
545 queue_for_each_hw_ctx(q
, hctx
, i
) {
546 if (hctx
->sched_tags
) {
547 blk_mq_free_rq_map(hctx
->sched_tags
, hctx
->flags
);
548 hctx
->sched_tags
= NULL
;
553 static int blk_mq_init_sched_shared_sbitmap(struct request_queue
*queue
)
555 struct blk_mq_tag_set
*set
= queue
->tag_set
;
556 int alloc_policy
= BLK_MQ_FLAG_TO_ALLOC_POLICY(set
->flags
);
557 struct blk_mq_hw_ctx
*hctx
;
561 * Set initial depth at max so that we don't need to reallocate for
562 * updating nr_requests.
564 ret
= blk_mq_init_bitmaps(&queue
->sched_bitmap_tags
,
565 &queue
->sched_breserved_tags
,
566 MAX_SCHED_RQ
, set
->reserved_tags
,
567 set
->numa_node
, alloc_policy
);
571 queue_for_each_hw_ctx(queue
, hctx
, i
) {
572 hctx
->sched_tags
->bitmap_tags
=
573 &queue
->sched_bitmap_tags
;
574 hctx
->sched_tags
->breserved_tags
=
575 &queue
->sched_breserved_tags
;
578 sbitmap_queue_resize(&queue
->sched_bitmap_tags
,
579 queue
->nr_requests
- set
->reserved_tags
);
584 static void blk_mq_exit_sched_shared_sbitmap(struct request_queue
*queue
)
586 sbitmap_queue_free(&queue
->sched_bitmap_tags
);
587 sbitmap_queue_free(&queue
->sched_breserved_tags
);
590 int blk_mq_init_sched(struct request_queue
*q
, struct elevator_type
*e
)
592 struct blk_mq_hw_ctx
*hctx
;
593 struct elevator_queue
*eq
;
599 q
->nr_requests
= q
->tag_set
->queue_depth
;
604 * Default to double of smaller one between hw queue_depth and 128,
605 * since we don't split into sync/async like the old code did.
606 * Additionally, this is a per-hw queue depth.
608 q
->nr_requests
= 2 * min_t(unsigned int, q
->tag_set
->queue_depth
,
611 queue_for_each_hw_ctx(q
, hctx
, i
) {
612 ret
= blk_mq_sched_alloc_tags(q
, hctx
, i
);
617 if (blk_mq_is_sbitmap_shared(q
->tag_set
->flags
)) {
618 ret
= blk_mq_init_sched_shared_sbitmap(q
);
623 ret
= e
->ops
.init_sched(q
, e
);
625 goto err_free_sbitmap
;
627 blk_mq_debugfs_register_sched(q
);
629 queue_for_each_hw_ctx(q
, hctx
, i
) {
630 if (e
->ops
.init_hctx
) {
631 ret
= e
->ops
.init_hctx(hctx
, i
);
634 blk_mq_sched_free_requests(q
);
635 blk_mq_exit_sched(q
, eq
);
636 kobject_put(&eq
->kobj
);
640 blk_mq_debugfs_register_sched_hctx(q
, hctx
);
646 if (blk_mq_is_sbitmap_shared(q
->tag_set
->flags
))
647 blk_mq_exit_sched_shared_sbitmap(q
);
649 blk_mq_sched_free_requests(q
);
650 blk_mq_sched_tags_teardown(q
);
656 * called in either blk_queue_cleanup or elevator_switch, tagset
657 * is required for freeing requests
659 void blk_mq_sched_free_requests(struct request_queue
*q
)
661 struct blk_mq_hw_ctx
*hctx
;
664 queue_for_each_hw_ctx(q
, hctx
, i
) {
665 if (hctx
->sched_tags
)
666 blk_mq_free_rqs(q
->tag_set
, hctx
->sched_tags
, i
);
670 void blk_mq_exit_sched(struct request_queue
*q
, struct elevator_queue
*e
)
672 struct blk_mq_hw_ctx
*hctx
;
674 unsigned int flags
= 0;
676 queue_for_each_hw_ctx(q
, hctx
, i
) {
677 blk_mq_debugfs_unregister_sched_hctx(hctx
);
678 if (e
->type
->ops
.exit_hctx
&& hctx
->sched_data
) {
679 e
->type
->ops
.exit_hctx(hctx
, i
);
680 hctx
->sched_data
= NULL
;
684 blk_mq_debugfs_unregister_sched(q
);
685 if (e
->type
->ops
.exit_sched
)
686 e
->type
->ops
.exit_sched(e
);
687 blk_mq_sched_tags_teardown(q
);
688 if (blk_mq_is_sbitmap_shared(flags
))
689 blk_mq_exit_sched_shared_sbitmap(q
);