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_free_hctx_data(struct request_queue *q,
22 void (*exit)(struct blk_mq_hw_ctx *))
24 struct blk_mq_hw_ctx *hctx;
27 queue_for_each_hw_ctx(q, hctx, i) {
28 if (exit && hctx->sched_data)
30 kfree(hctx->sched_data);
31 hctx->sched_data = NULL;
34 EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data);
36 void blk_mq_sched_assign_ioc(struct request *rq)
38 struct request_queue *q = rq->q;
39 struct io_context *ioc;
43 * May not have an IO context if it's a passthrough request
45 ioc = current->io_context;
49 spin_lock_irq(&q->queue_lock);
50 icq = ioc_lookup_icq(ioc, q);
51 spin_unlock_irq(&q->queue_lock);
54 icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
58 get_io_context(icq->ioc);
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.
66 void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
68 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
71 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
73 EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
75 void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
77 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
79 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
81 blk_mq_run_hw_queue(hctx, true);
84 static int sched_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
86 struct request *rqa = container_of(a, struct request, queuelist);
87 struct request *rqb = container_of(b, struct request, queuelist);
89 return rqa->mq_hctx > rqb->mq_hctx;
92 static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list)
94 struct blk_mq_hw_ctx *hctx =
95 list_first_entry(rq_list, struct request, queuelist)->mq_hctx;
98 unsigned int count = 0;
100 list_for_each_entry(rq, rq_list, queuelist) {
101 if (rq->mq_hctx != hctx) {
102 list_cut_before(&hctx_list, rq_list, &rq->queuelist);
107 list_splice_tail_init(rq_list, &hctx_list);
110 return blk_mq_dispatch_rq_list(hctx, &hctx_list, count);
113 #define BLK_MQ_BUDGET_DELAY 3 /* ms units */
116 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
117 * its queue by itself in its completion handler, so we don't need to
118 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
120 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
121 * be run again. This is necessary to avoid starving flushes.
123 static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
125 struct request_queue *q = hctx->queue;
126 struct elevator_queue *e = q->elevator;
127 bool multi_hctxs = false, run_queue = false;
128 bool dispatched = false, busy = false;
129 unsigned int max_dispatch;
133 if (hctx->dispatch_busy)
136 max_dispatch = hctx->queue->nr_requests;
141 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
144 if (!list_empty_careful(&hctx->dispatch)) {
149 if (!blk_mq_get_dispatch_budget(q))
152 rq = e->type->ops.dispatch_request(hctx);
154 blk_mq_put_dispatch_budget(q);
156 * We're releasing without dispatching. Holding the
157 * budget could have blocked any "hctx"s with the
158 * same queue and if we didn't dispatch then there's
159 * no guarantee anyone will kick the queue. Kick it
167 * Now this rq owns the budget which has to be released
168 * if this rq won't be queued to driver via .queue_rq()
169 * in blk_mq_dispatch_rq_list().
171 list_add_tail(&rq->queuelist, &rq_list);
172 if (rq->mq_hctx != hctx)
174 } while (++count < max_dispatch);
178 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
179 } else if (multi_hctxs) {
181 * Requests from different hctx may be dequeued from some
182 * schedulers, such as bfq and deadline.
184 * Sort the requests in the list according to their hctx,
185 * dispatch batching requests from same hctx at a time.
187 list_sort(NULL, &rq_list, sched_rq_cmp);
189 dispatched |= blk_mq_dispatch_hctx_list(&rq_list);
190 } while (!list_empty(&rq_list));
192 dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, count);
200 static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
205 ret = __blk_mq_do_dispatch_sched(hctx);
211 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
212 struct blk_mq_ctx *ctx)
214 unsigned short idx = ctx->index_hw[hctx->type];
216 if (++idx == hctx->nr_ctx)
219 return hctx->ctxs[idx];
223 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
224 * its queue by itself in its completion handler, so we don't need to
225 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
227 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
228 * be run again. This is necessary to avoid starving flushes.
230 static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
232 struct request_queue *q = hctx->queue;
234 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
239 if (!list_empty_careful(&hctx->dispatch)) {
244 if (!sbitmap_any_bit_set(&hctx->ctx_map))
247 if (!blk_mq_get_dispatch_budget(q))
250 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
252 blk_mq_put_dispatch_budget(q);
254 * We're releasing without dispatching. Holding the
255 * budget could have blocked any "hctx"s with the
256 * same queue and if we didn't dispatch then there's
257 * no guarantee anyone will kick the queue. Kick it
260 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
265 * Now this rq owns the budget which has to be released
266 * if this rq won't be queued to driver via .queue_rq()
267 * in blk_mq_dispatch_rq_list().
269 list_add(&rq->queuelist, &rq_list);
271 /* round robin for fair dispatch */
272 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
274 } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1));
276 WRITE_ONCE(hctx->dispatch_from, ctx);
280 static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
282 struct request_queue *q = hctx->queue;
283 struct elevator_queue *e = q->elevator;
284 const bool has_sched_dispatch = e && e->type->ops.dispatch_request;
289 * If we have previous entries on our dispatch list, grab them first for
290 * more fair dispatch.
292 if (!list_empty_careful(&hctx->dispatch)) {
293 spin_lock(&hctx->lock);
294 if (!list_empty(&hctx->dispatch))
295 list_splice_init(&hctx->dispatch, &rq_list);
296 spin_unlock(&hctx->lock);
300 * Only ask the scheduler for requests, if we didn't have residual
301 * requests from the dispatch list. This is to avoid the case where
302 * we only ever dispatch a fraction of the requests available because
303 * of low device queue depth. Once we pull requests out of the IO
304 * scheduler, we can no longer merge or sort them. So it's best to
305 * leave them there for as long as we can. Mark the hw queue as
306 * needing a restart in that case.
308 * We want to dispatch from the scheduler if there was nothing
309 * on the dispatch list or we were able to dispatch from the
312 if (!list_empty(&rq_list)) {
313 blk_mq_sched_mark_restart_hctx(hctx);
314 if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) {
315 if (has_sched_dispatch)
316 ret = blk_mq_do_dispatch_sched(hctx);
318 ret = blk_mq_do_dispatch_ctx(hctx);
320 } else if (has_sched_dispatch) {
321 ret = blk_mq_do_dispatch_sched(hctx);
322 } else if (hctx->dispatch_busy) {
323 /* dequeue request one by one from sw queue if queue is busy */
324 ret = blk_mq_do_dispatch_ctx(hctx);
326 blk_mq_flush_busy_ctxs(hctx, &rq_list);
327 blk_mq_dispatch_rq_list(hctx, &rq_list, 0);
333 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
335 struct request_queue *q = hctx->queue;
337 /* RCU or SRCU read lock is needed before checking quiesced flag */
338 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
344 * A return of -EAGAIN is an indication that hctx->dispatch is not
345 * empty and we must run again in order to avoid starving flushes.
347 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) {
348 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN)
349 blk_mq_run_hw_queue(hctx, true);
353 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
354 unsigned int nr_segs, struct request **merged_request)
358 switch (elv_merge(q, &rq, bio)) {
359 case ELEVATOR_BACK_MERGE:
360 if (!blk_mq_sched_allow_merge(q, rq, bio))
362 if (!bio_attempt_back_merge(rq, bio, nr_segs))
364 *merged_request = attempt_back_merge(q, rq);
365 if (!*merged_request)
366 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
368 case ELEVATOR_FRONT_MERGE:
369 if (!blk_mq_sched_allow_merge(q, rq, bio))
371 if (!bio_attempt_front_merge(rq, bio, nr_segs))
373 *merged_request = attempt_front_merge(q, rq);
374 if (!*merged_request)
375 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
377 case ELEVATOR_DISCARD_MERGE:
378 return bio_attempt_discard_merge(q, rq, bio);
383 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
386 * Iterate list of requests and see if we can merge this bio with any
389 bool blk_mq_bio_list_merge(struct request_queue *q, struct list_head *list,
390 struct bio *bio, unsigned int nr_segs)
395 list_for_each_entry_reverse(rq, list, queuelist) {
401 if (!blk_rq_merge_ok(rq, bio))
404 switch (blk_try_merge(rq, bio)) {
405 case ELEVATOR_BACK_MERGE:
406 if (blk_mq_sched_allow_merge(q, rq, bio))
407 merged = bio_attempt_back_merge(rq, bio,
410 case ELEVATOR_FRONT_MERGE:
411 if (blk_mq_sched_allow_merge(q, rq, bio))
412 merged = bio_attempt_front_merge(rq, bio,
415 case ELEVATOR_DISCARD_MERGE:
416 merged = bio_attempt_discard_merge(q, rq, bio);
427 EXPORT_SYMBOL_GPL(blk_mq_bio_list_merge);
430 * Reverse check our software queue for entries that we could potentially
431 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
432 * too much time checking for merges.
434 static bool blk_mq_attempt_merge(struct request_queue *q,
435 struct blk_mq_hw_ctx *hctx,
436 struct blk_mq_ctx *ctx, struct bio *bio,
437 unsigned int nr_segs)
439 enum hctx_type type = hctx->type;
441 lockdep_assert_held(&ctx->lock);
443 if (blk_mq_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) {
451 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
452 unsigned int nr_segs)
454 struct elevator_queue *e = q->elevator;
455 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
456 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
460 if (e && e->type->ops.bio_merge)
461 return e->type->ops.bio_merge(hctx, bio, nr_segs);
464 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
465 !list_empty_careful(&ctx->rq_lists[type])) {
466 /* default per sw-queue merge */
467 spin_lock(&ctx->lock);
468 ret = blk_mq_attempt_merge(q, hctx, ctx, bio, nr_segs);
469 spin_unlock(&ctx->lock);
475 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
477 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
479 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
481 void blk_mq_sched_request_inserted(struct request *rq)
483 trace_block_rq_insert(rq->q, rq);
485 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
487 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
492 * dispatch flush and passthrough rq directly
494 * passthrough request has to be added to hctx->dispatch directly.
495 * For some reason, device may be in one situation which can't
496 * handle FS request, so STS_RESOURCE is always returned and the
497 * FS request will be added to hctx->dispatch. However passthrough
498 * request may be required at that time for fixing the problem. If
499 * passthrough request is added to scheduler queue, there isn't any
500 * chance to dispatch it given we prioritize requests in hctx->dispatch.
502 if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq))
506 rq->rq_flags |= RQF_SORTED;
511 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
512 bool run_queue, bool async)
514 struct request_queue *q = rq->q;
515 struct elevator_queue *e = q->elevator;
516 struct blk_mq_ctx *ctx = rq->mq_ctx;
517 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
519 /* flush rq in flush machinery need to be dispatched directly */
520 if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) {
521 blk_insert_flush(rq);
525 WARN_ON(e && (rq->tag != -1));
527 if (blk_mq_sched_bypass_insert(hctx, !!e, rq)) {
529 * Firstly normal IO request is inserted to scheduler queue or
530 * sw queue, meantime we add flush request to dispatch queue(
531 * hctx->dispatch) directly and there is at most one in-flight
532 * flush request for each hw queue, so it doesn't matter to add
533 * flush request to tail or front of the dispatch queue.
535 * Secondly in case of NCQ, flush request belongs to non-NCQ
536 * command, and queueing it will fail when there is any
537 * in-flight normal IO request(NCQ command). When adding flush
538 * rq to the front of hctx->dispatch, it is easier to introduce
539 * extra time to flush rq's latency because of S_SCHED_RESTART
540 * compared with adding to the tail of dispatch queue, then
541 * chance of flush merge is increased, and less flush requests
542 * will be issued to controller. It is observed that ~10% time
543 * is saved in blktests block/004 on disk attached to AHCI/NCQ
544 * drive when adding flush rq to the front of hctx->dispatch.
546 * Simply queue flush rq to the front of hctx->dispatch so that
547 * intensive flush workloads can benefit in case of NCQ HW.
549 at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head;
550 blk_mq_request_bypass_insert(rq, at_head, false);
554 if (e && e->type->ops.insert_requests) {
557 list_add(&rq->queuelist, &list);
558 e->type->ops.insert_requests(hctx, &list, at_head);
560 spin_lock(&ctx->lock);
561 __blk_mq_insert_request(hctx, rq, at_head);
562 spin_unlock(&ctx->lock);
567 blk_mq_run_hw_queue(hctx, async);
570 void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
571 struct blk_mq_ctx *ctx,
572 struct list_head *list, bool run_queue_async)
574 struct elevator_queue *e;
575 struct request_queue *q = hctx->queue;
578 * blk_mq_sched_insert_requests() is called from flush plug
579 * context only, and hold one usage counter to prevent queue
580 * from being released.
582 percpu_ref_get(&q->q_usage_counter);
584 e = hctx->queue->elevator;
585 if (e && e->type->ops.insert_requests)
586 e->type->ops.insert_requests(hctx, list, false);
589 * try to issue requests directly if the hw queue isn't
590 * busy in case of 'none' scheduler, and this way may save
591 * us one extra enqueue & dequeue to sw queue.
593 if (!hctx->dispatch_busy && !e && !run_queue_async) {
594 blk_mq_try_issue_list_directly(hctx, list);
595 if (list_empty(list))
598 blk_mq_insert_requests(hctx, ctx, list);
601 blk_mq_run_hw_queue(hctx, run_queue_async);
603 percpu_ref_put(&q->q_usage_counter);
606 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
607 struct blk_mq_hw_ctx *hctx,
608 unsigned int hctx_idx)
610 if (hctx->sched_tags) {
611 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
612 blk_mq_free_rq_map(hctx->sched_tags);
613 hctx->sched_tags = NULL;
617 static int blk_mq_sched_alloc_tags(struct request_queue *q,
618 struct blk_mq_hw_ctx *hctx,
619 unsigned int hctx_idx)
621 struct blk_mq_tag_set *set = q->tag_set;
624 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
626 if (!hctx->sched_tags)
629 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
631 blk_mq_sched_free_tags(set, hctx, hctx_idx);
636 /* called in queue's release handler, tagset has gone away */
637 static void blk_mq_sched_tags_teardown(struct request_queue *q)
639 struct blk_mq_hw_ctx *hctx;
642 queue_for_each_hw_ctx(q, hctx, i) {
643 if (hctx->sched_tags) {
644 blk_mq_free_rq_map(hctx->sched_tags);
645 hctx->sched_tags = NULL;
650 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
652 struct blk_mq_hw_ctx *hctx;
653 struct elevator_queue *eq;
659 q->nr_requests = q->tag_set->queue_depth;
664 * Default to double of smaller one between hw queue_depth and 128,
665 * since we don't split into sync/async like the old code did.
666 * Additionally, this is a per-hw queue depth.
668 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
671 queue_for_each_hw_ctx(q, hctx, i) {
672 ret = blk_mq_sched_alloc_tags(q, hctx, i);
677 ret = e->ops.init_sched(q, e);
681 blk_mq_debugfs_register_sched(q);
683 queue_for_each_hw_ctx(q, hctx, i) {
684 if (e->ops.init_hctx) {
685 ret = e->ops.init_hctx(hctx, i);
688 blk_mq_sched_free_requests(q);
689 blk_mq_exit_sched(q, eq);
690 kobject_put(&eq->kobj);
694 blk_mq_debugfs_register_sched_hctx(q, hctx);
700 blk_mq_sched_free_requests(q);
701 blk_mq_sched_tags_teardown(q);
707 * called in either blk_queue_cleanup or elevator_switch, tagset
708 * is required for freeing requests
710 void blk_mq_sched_free_requests(struct request_queue *q)
712 struct blk_mq_hw_ctx *hctx;
715 queue_for_each_hw_ctx(q, hctx, i) {
716 if (hctx->sched_tags)
717 blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i);
721 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
723 struct blk_mq_hw_ctx *hctx;
726 queue_for_each_hw_ctx(q, hctx, i) {
727 blk_mq_debugfs_unregister_sched_hctx(hctx);
728 if (e->type->ops.exit_hctx && hctx->sched_data) {
729 e->type->ops.exit_hctx(hctx, i);
730 hctx->sched_data = NULL;
733 blk_mq_debugfs_unregister_sched(q);
734 if (e->type->ops.exit_sched)
735 e->type->ops.exit_sched(e);
736 blk_mq_sched_tags_teardown(q);