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
);
171 if (rq
->mq_hctx
!= hctx
)
173 } while (++count
< max_dispatch
);
177 blk_mq_delay_run_hw_queues(q
, BLK_MQ_BUDGET_DELAY
);
178 } else if (multi_hctxs
) {
180 * Requests from different hctx may be dequeued from some
181 * schedulers, such as bfq and deadline.
183 * Sort the requests in the list according to their hctx,
184 * dispatch batching requests from same hctx at a time.
186 list_sort(NULL
, &rq_list
, sched_rq_cmp
);
188 dispatched
|= blk_mq_dispatch_hctx_list(&rq_list
);
189 } while (!list_empty(&rq_list
));
191 dispatched
= blk_mq_dispatch_rq_list(hctx
, &rq_list
, count
);
199 static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx
*hctx
)
204 ret
= __blk_mq_do_dispatch_sched(hctx
);
210 static struct blk_mq_ctx
*blk_mq_next_ctx(struct blk_mq_hw_ctx
*hctx
,
211 struct blk_mq_ctx
*ctx
)
213 unsigned short idx
= ctx
->index_hw
[hctx
->type
];
215 if (++idx
== hctx
->nr_ctx
)
218 return hctx
->ctxs
[idx
];
222 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
223 * its queue by itself in its completion handler, so we don't need to
224 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
226 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
227 * be run again. This is necessary to avoid starving flushes.
229 static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx
*hctx
)
231 struct request_queue
*q
= hctx
->queue
;
233 struct blk_mq_ctx
*ctx
= READ_ONCE(hctx
->dispatch_from
);
240 if (!list_empty_careful(&hctx
->dispatch
)) {
245 if (!sbitmap_any_bit_set(&hctx
->ctx_map
))
248 budget_token
= blk_mq_get_dispatch_budget(q
);
249 if (budget_token
< 0)
252 rq
= blk_mq_dequeue_from_ctx(hctx
, ctx
);
254 blk_mq_put_dispatch_budget(q
, budget_token
);
256 * We're releasing without dispatching. Holding the
257 * budget could have blocked any "hctx"s with the
258 * same queue and if we didn't dispatch then there's
259 * no guarantee anyone will kick the queue. Kick it
262 blk_mq_delay_run_hw_queues(q
, BLK_MQ_BUDGET_DELAY
);
266 blk_mq_set_rq_budget_token(rq
, budget_token
);
269 * Now this rq owns the budget which has to be released
270 * if this rq won't be queued to driver via .queue_rq()
271 * in blk_mq_dispatch_rq_list().
273 list_add(&rq
->queuelist
, &rq_list
);
275 /* round robin for fair dispatch */
276 ctx
= blk_mq_next_ctx(hctx
, rq
->mq_ctx
);
278 } while (blk_mq_dispatch_rq_list(rq
->mq_hctx
, &rq_list
, 1));
280 WRITE_ONCE(hctx
->dispatch_from
, ctx
);
284 static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx
*hctx
)
286 struct request_queue
*q
= hctx
->queue
;
287 struct elevator_queue
*e
= q
->elevator
;
288 const bool has_sched_dispatch
= e
&& e
->type
->ops
.dispatch_request
;
293 * If we have previous entries on our dispatch list, grab them first for
294 * more fair dispatch.
296 if (!list_empty_careful(&hctx
->dispatch
)) {
297 spin_lock(&hctx
->lock
);
298 if (!list_empty(&hctx
->dispatch
))
299 list_splice_init(&hctx
->dispatch
, &rq_list
);
300 spin_unlock(&hctx
->lock
);
304 * Only ask the scheduler for requests, if we didn't have residual
305 * requests from the dispatch list. This is to avoid the case where
306 * we only ever dispatch a fraction of the requests available because
307 * of low device queue depth. Once we pull requests out of the IO
308 * scheduler, we can no longer merge or sort them. So it's best to
309 * leave them there for as long as we can. Mark the hw queue as
310 * needing a restart in that case.
312 * We want to dispatch from the scheduler if there was nothing
313 * on the dispatch list or we were able to dispatch from the
316 if (!list_empty(&rq_list
)) {
317 blk_mq_sched_mark_restart_hctx(hctx
);
318 if (blk_mq_dispatch_rq_list(hctx
, &rq_list
, 0)) {
319 if (has_sched_dispatch
)
320 ret
= blk_mq_do_dispatch_sched(hctx
);
322 ret
= blk_mq_do_dispatch_ctx(hctx
);
324 } else if (has_sched_dispatch
) {
325 ret
= blk_mq_do_dispatch_sched(hctx
);
326 } else if (hctx
->dispatch_busy
) {
327 /* dequeue request one by one from sw queue if queue is busy */
328 ret
= blk_mq_do_dispatch_ctx(hctx
);
330 blk_mq_flush_busy_ctxs(hctx
, &rq_list
);
331 blk_mq_dispatch_rq_list(hctx
, &rq_list
, 0);
337 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx
*hctx
)
339 struct request_queue
*q
= hctx
->queue
;
341 /* RCU or SRCU read lock is needed before checking quiesced flag */
342 if (unlikely(blk_mq_hctx_stopped(hctx
) || blk_queue_quiesced(q
)))
348 * A return of -EAGAIN is an indication that hctx->dispatch is not
349 * empty and we must run again in order to avoid starving flushes.
351 if (__blk_mq_sched_dispatch_requests(hctx
) == -EAGAIN
) {
352 if (__blk_mq_sched_dispatch_requests(hctx
) == -EAGAIN
)
353 blk_mq_run_hw_queue(hctx
, true);
357 bool __blk_mq_sched_bio_merge(struct request_queue
*q
, struct bio
*bio
,
358 unsigned int nr_segs
)
360 struct elevator_queue
*e
= q
->elevator
;
361 struct blk_mq_ctx
*ctx
;
362 struct blk_mq_hw_ctx
*hctx
;
366 if (e
&& e
->type
->ops
.bio_merge
)
367 return e
->type
->ops
.bio_merge(q
, bio
, nr_segs
);
369 ctx
= blk_mq_get_ctx(q
);
370 hctx
= blk_mq_map_queue(q
, bio
->bi_opf
, ctx
);
372 if (!(hctx
->flags
& BLK_MQ_F_SHOULD_MERGE
) ||
373 list_empty_careful(&ctx
->rq_lists
[type
]))
376 /* default per sw-queue merge */
377 spin_lock(&ctx
->lock
);
379 * Reverse check our software queue for entries that we could
380 * potentially merge with. Currently includes a hand-wavy stop
381 * count of 8, to not spend too much time checking for merges.
383 if (blk_bio_list_merge(q
, &ctx
->rq_lists
[type
], bio
, nr_segs
)) {
388 spin_unlock(&ctx
->lock
);
393 bool blk_mq_sched_try_insert_merge(struct request_queue
*q
, struct request
*rq
)
395 return rq_mergeable(rq
) && elv_attempt_insert_merge(q
, rq
);
397 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge
);
399 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx
*hctx
,
403 * dispatch flush and passthrough rq directly
405 * passthrough request has to be added to hctx->dispatch directly.
406 * For some reason, device may be in one situation which can't
407 * handle FS request, so STS_RESOURCE is always returned and the
408 * FS request will be added to hctx->dispatch. However passthrough
409 * request may be required at that time for fixing the problem. If
410 * passthrough request is added to scheduler queue, there isn't any
411 * chance to dispatch it given we prioritize requests in hctx->dispatch.
413 if ((rq
->rq_flags
& RQF_FLUSH_SEQ
) || blk_rq_is_passthrough(rq
))
419 void blk_mq_sched_insert_request(struct request
*rq
, bool at_head
,
420 bool run_queue
, bool async
)
422 struct request_queue
*q
= rq
->q
;
423 struct elevator_queue
*e
= q
->elevator
;
424 struct blk_mq_ctx
*ctx
= rq
->mq_ctx
;
425 struct blk_mq_hw_ctx
*hctx
= rq
->mq_hctx
;
427 WARN_ON(e
&& (rq
->tag
!= BLK_MQ_NO_TAG
));
429 if (blk_mq_sched_bypass_insert(hctx
, rq
)) {
431 * Firstly normal IO request is inserted to scheduler queue or
432 * sw queue, meantime we add flush request to dispatch queue(
433 * hctx->dispatch) directly and there is at most one in-flight
434 * flush request for each hw queue, so it doesn't matter to add
435 * flush request to tail or front of the dispatch queue.
437 * Secondly in case of NCQ, flush request belongs to non-NCQ
438 * command, and queueing it will fail when there is any
439 * in-flight normal IO request(NCQ command). When adding flush
440 * rq to the front of hctx->dispatch, it is easier to introduce
441 * extra time to flush rq's latency because of S_SCHED_RESTART
442 * compared with adding to the tail of dispatch queue, then
443 * chance of flush merge is increased, and less flush requests
444 * will be issued to controller. It is observed that ~10% time
445 * is saved in blktests block/004 on disk attached to AHCI/NCQ
446 * drive when adding flush rq to the front of hctx->dispatch.
448 * Simply queue flush rq to the front of hctx->dispatch so that
449 * intensive flush workloads can benefit in case of NCQ HW.
451 at_head
= (rq
->rq_flags
& RQF_FLUSH_SEQ
) ? true : at_head
;
452 blk_mq_request_bypass_insert(rq
, at_head
, false);
456 if (e
&& e
->type
->ops
.insert_requests
) {
459 list_add(&rq
->queuelist
, &list
);
460 e
->type
->ops
.insert_requests(hctx
, &list
, at_head
);
462 spin_lock(&ctx
->lock
);
463 __blk_mq_insert_request(hctx
, rq
, at_head
);
464 spin_unlock(&ctx
->lock
);
469 blk_mq_run_hw_queue(hctx
, async
);
472 void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx
*hctx
,
473 struct blk_mq_ctx
*ctx
,
474 struct list_head
*list
, bool run_queue_async
)
476 struct elevator_queue
*e
;
477 struct request_queue
*q
= hctx
->queue
;
480 * blk_mq_sched_insert_requests() is called from flush plug
481 * context only, and hold one usage counter to prevent queue
482 * from being released.
484 percpu_ref_get(&q
->q_usage_counter
);
486 e
= hctx
->queue
->elevator
;
487 if (e
&& e
->type
->ops
.insert_requests
)
488 e
->type
->ops
.insert_requests(hctx
, list
, false);
491 * try to issue requests directly if the hw queue isn't
492 * busy in case of 'none' scheduler, and this way may save
493 * us one extra enqueue & dequeue to sw queue.
495 if (!hctx
->dispatch_busy
&& !e
&& !run_queue_async
) {
496 blk_mq_try_issue_list_directly(hctx
, list
);
497 if (list_empty(list
))
500 blk_mq_insert_requests(hctx
, ctx
, list
);
503 blk_mq_run_hw_queue(hctx
, run_queue_async
);
505 percpu_ref_put(&q
->q_usage_counter
);
508 static void blk_mq_sched_free_tags(struct blk_mq_tag_set
*set
,
509 struct blk_mq_hw_ctx
*hctx
,
510 unsigned int hctx_idx
)
512 unsigned int flags
= set
->flags
& ~BLK_MQ_F_TAG_HCTX_SHARED
;
514 if (hctx
->sched_tags
) {
515 blk_mq_free_rqs(set
, hctx
->sched_tags
, hctx_idx
);
516 blk_mq_free_rq_map(hctx
->sched_tags
, flags
);
517 hctx
->sched_tags
= NULL
;
521 static int blk_mq_sched_alloc_tags(struct request_queue
*q
,
522 struct blk_mq_hw_ctx
*hctx
,
523 unsigned int hctx_idx
)
525 struct blk_mq_tag_set
*set
= q
->tag_set
;
526 /* Clear HCTX_SHARED so tags are init'ed */
527 unsigned int flags
= set
->flags
& ~BLK_MQ_F_TAG_HCTX_SHARED
;
530 hctx
->sched_tags
= blk_mq_alloc_rq_map(set
, hctx_idx
, q
->nr_requests
,
531 set
->reserved_tags
, flags
);
532 if (!hctx
->sched_tags
)
535 ret
= blk_mq_alloc_rqs(set
, hctx
->sched_tags
, hctx_idx
, q
->nr_requests
);
537 blk_mq_sched_free_tags(set
, hctx
, hctx_idx
);
542 /* called in queue's release handler, tagset has gone away */
543 static void blk_mq_sched_tags_teardown(struct request_queue
*q
)
545 struct blk_mq_hw_ctx
*hctx
;
548 queue_for_each_hw_ctx(q
, hctx
, i
) {
549 /* Clear HCTX_SHARED so tags are freed */
550 unsigned int flags
= hctx
->flags
& ~BLK_MQ_F_TAG_HCTX_SHARED
;
552 if (hctx
->sched_tags
) {
553 blk_mq_free_rq_map(hctx
->sched_tags
, flags
);
554 hctx
->sched_tags
= NULL
;
559 int blk_mq_init_sched(struct request_queue
*q
, struct elevator_type
*e
)
561 struct blk_mq_hw_ctx
*hctx
;
562 struct elevator_queue
*eq
;
568 q
->nr_requests
= q
->tag_set
->queue_depth
;
573 * Default to double of smaller one between hw queue_depth and 128,
574 * since we don't split into sync/async like the old code did.
575 * Additionally, this is a per-hw queue depth.
577 q
->nr_requests
= 2 * min_t(unsigned int, q
->tag_set
->queue_depth
,
580 queue_for_each_hw_ctx(q
, hctx
, i
) {
581 ret
= blk_mq_sched_alloc_tags(q
, hctx
, i
);
586 ret
= e
->ops
.init_sched(q
, e
);
590 blk_mq_debugfs_register_sched(q
);
592 queue_for_each_hw_ctx(q
, hctx
, i
) {
593 if (e
->ops
.init_hctx
) {
594 ret
= e
->ops
.init_hctx(hctx
, i
);
597 blk_mq_sched_free_requests(q
);
598 blk_mq_exit_sched(q
, eq
);
599 kobject_put(&eq
->kobj
);
603 blk_mq_debugfs_register_sched_hctx(q
, hctx
);
609 blk_mq_sched_free_requests(q
);
610 blk_mq_sched_tags_teardown(q
);
616 * called in either blk_queue_cleanup or elevator_switch, tagset
617 * is required for freeing requests
619 void blk_mq_sched_free_requests(struct request_queue
*q
)
621 struct blk_mq_hw_ctx
*hctx
;
624 queue_for_each_hw_ctx(q
, hctx
, i
) {
625 if (hctx
->sched_tags
)
626 blk_mq_free_rqs(q
->tag_set
, hctx
->sched_tags
, i
);
630 void blk_mq_exit_sched(struct request_queue
*q
, struct elevator_queue
*e
)
632 struct blk_mq_hw_ctx
*hctx
;
635 queue_for_each_hw_ctx(q
, hctx
, i
) {
636 blk_mq_debugfs_unregister_sched_hctx(hctx
);
637 if (e
->type
->ops
.exit_hctx
&& hctx
->sched_data
) {
638 e
->type
->ops
.exit_hctx(hctx
, i
);
639 hctx
->sched_data
= NULL
;
642 blk_mq_debugfs_unregister_sched(q
);
643 if (e
->type
->ops
.exit_sched
)
644 e
->type
->ops
.exit_sched(e
);
645 blk_mq_sched_tags_teardown(q
);