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);
82 * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch)
83 * in blk_mq_run_hw_queue(). Its pair is the barrier in
84 * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART,
85 * meantime new request added to hctx->dispatch is missed to check in
86 * blk_mq_run_hw_queue().
90 blk_mq_run_hw_queue(hctx, true);
93 static int sched_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
95 struct request *rqa = container_of(a, struct request, queuelist);
96 struct request *rqb = container_of(b, struct request, queuelist);
98 return rqa->mq_hctx > rqb->mq_hctx;
101 static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list)
103 struct blk_mq_hw_ctx *hctx =
104 list_first_entry(rq_list, struct request, queuelist)->mq_hctx;
106 LIST_HEAD(hctx_list);
107 unsigned int count = 0;
109 list_for_each_entry(rq, rq_list, queuelist) {
110 if (rq->mq_hctx != hctx) {
111 list_cut_before(&hctx_list, rq_list, &rq->queuelist);
116 list_splice_tail_init(rq_list, &hctx_list);
119 return blk_mq_dispatch_rq_list(hctx, &hctx_list, count);
122 #define BLK_MQ_BUDGET_DELAY 3 /* ms units */
125 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
126 * its queue by itself in its completion handler, so we don't need to
127 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
129 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
130 * be run again. This is necessary to avoid starving flushes.
132 static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
134 struct request_queue *q = hctx->queue;
135 struct elevator_queue *e = q->elevator;
136 bool multi_hctxs = false, run_queue = false;
137 bool dispatched = false, busy = false;
138 unsigned int max_dispatch;
142 if (hctx->dispatch_busy)
145 max_dispatch = hctx->queue->nr_requests;
150 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
153 if (!list_empty_careful(&hctx->dispatch)) {
158 if (!blk_mq_get_dispatch_budget(q))
161 rq = e->type->ops.dispatch_request(hctx);
163 blk_mq_put_dispatch_budget(q);
165 * We're releasing without dispatching. Holding the
166 * budget could have blocked any "hctx"s with the
167 * same queue and if we didn't dispatch then there's
168 * no guarantee anyone will kick the queue. Kick it
176 * Now this rq owns the budget which has to be released
177 * if this rq won't be queued to driver via .queue_rq()
178 * in blk_mq_dispatch_rq_list().
180 list_add_tail(&rq->queuelist, &rq_list);
181 if (rq->mq_hctx != hctx)
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);
248 if (!list_empty_careful(&hctx->dispatch)) {
253 if (!sbitmap_any_bit_set(&hctx->ctx_map))
256 if (!blk_mq_get_dispatch_budget(q))
259 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
261 blk_mq_put_dispatch_budget(q);
263 * We're releasing without dispatching. Holding the
264 * budget could have blocked any "hctx"s with the
265 * same queue and if we didn't dispatch then there's
266 * no guarantee anyone will kick the queue. Kick it
269 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
274 * Now this rq owns the budget which has to be released
275 * if this rq won't be queued to driver via .queue_rq()
276 * in blk_mq_dispatch_rq_list().
278 list_add(&rq->queuelist, &rq_list);
280 /* round robin for fair dispatch */
281 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
283 } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1));
285 WRITE_ONCE(hctx->dispatch_from, ctx);
289 static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
291 struct request_queue *q = hctx->queue;
292 struct elevator_queue *e = q->elevator;
293 const bool has_sched_dispatch = e && e->type->ops.dispatch_request;
298 * If we have previous entries on our dispatch list, grab them first for
299 * more fair dispatch.
301 if (!list_empty_careful(&hctx->dispatch)) {
302 spin_lock(&hctx->lock);
303 if (!list_empty(&hctx->dispatch))
304 list_splice_init(&hctx->dispatch, &rq_list);
305 spin_unlock(&hctx->lock);
309 * Only ask the scheduler for requests, if we didn't have residual
310 * requests from the dispatch list. This is to avoid the case where
311 * we only ever dispatch a fraction of the requests available because
312 * of low device queue depth. Once we pull requests out of the IO
313 * scheduler, we can no longer merge or sort them. So it's best to
314 * leave them there for as long as we can. Mark the hw queue as
315 * needing a restart in that case.
317 * We want to dispatch from the scheduler if there was nothing
318 * on the dispatch list or we were able to dispatch from the
321 if (!list_empty(&rq_list)) {
322 blk_mq_sched_mark_restart_hctx(hctx);
323 if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) {
324 if (has_sched_dispatch)
325 ret = blk_mq_do_dispatch_sched(hctx);
327 ret = blk_mq_do_dispatch_ctx(hctx);
329 } else if (has_sched_dispatch) {
330 ret = blk_mq_do_dispatch_sched(hctx);
331 } else if (hctx->dispatch_busy) {
332 /* dequeue request one by one from sw queue if queue is busy */
333 ret = blk_mq_do_dispatch_ctx(hctx);
335 blk_mq_flush_busy_ctxs(hctx, &rq_list);
336 blk_mq_dispatch_rq_list(hctx, &rq_list, 0);
342 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
344 struct request_queue *q = hctx->queue;
346 /* RCU or SRCU read lock is needed before checking quiesced flag */
347 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
353 * A return of -EAGAIN is an indication that hctx->dispatch is not
354 * empty and we must run again in order to avoid starving flushes.
356 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) {
357 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN)
358 blk_mq_run_hw_queue(hctx, true);
362 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
363 unsigned int nr_segs, struct request **merged_request)
367 switch (elv_merge(q, &rq, bio)) {
368 case ELEVATOR_BACK_MERGE:
369 if (!blk_mq_sched_allow_merge(q, rq, bio))
371 if (bio_attempt_back_merge(rq, bio, nr_segs) != BIO_MERGE_OK)
373 *merged_request = attempt_back_merge(q, rq);
374 if (!*merged_request)
375 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
377 case ELEVATOR_FRONT_MERGE:
378 if (!blk_mq_sched_allow_merge(q, rq, bio))
380 if (bio_attempt_front_merge(rq, bio, nr_segs) != BIO_MERGE_OK)
382 *merged_request = attempt_front_merge(q, rq);
383 if (!*merged_request)
384 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
386 case ELEVATOR_DISCARD_MERGE:
387 return bio_attempt_discard_merge(q, rq, bio) == BIO_MERGE_OK;
392 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
394 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
395 unsigned int nr_segs)
397 struct elevator_queue *e = q->elevator;
398 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
399 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
403 if (e && e->type->ops.bio_merge)
404 return e->type->ops.bio_merge(hctx, bio, nr_segs);
407 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE) ||
408 list_empty_careful(&ctx->rq_lists[type]))
411 /* default per sw-queue merge */
412 spin_lock(&ctx->lock);
414 * Reverse check our software queue for entries that we could
415 * potentially merge with. Currently includes a hand-wavy stop
416 * count of 8, to not spend too much time checking for merges.
418 if (blk_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) {
423 spin_unlock(&ctx->lock);
428 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
430 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
432 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
434 void blk_mq_sched_request_inserted(struct request *rq)
436 trace_block_rq_insert(rq->q, rq);
438 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
440 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
445 * dispatch flush and passthrough rq directly
447 * passthrough request has to be added to hctx->dispatch directly.
448 * For some reason, device may be in one situation which can't
449 * handle FS request, so STS_RESOURCE is always returned and the
450 * FS request will be added to hctx->dispatch. However passthrough
451 * request may be required at that time for fixing the problem. If
452 * passthrough request is added to scheduler queue, there isn't any
453 * chance to dispatch it given we prioritize requests in hctx->dispatch.
455 if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq))
459 rq->rq_flags |= RQF_SORTED;
464 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
465 bool run_queue, bool async)
467 struct request_queue *q = rq->q;
468 struct elevator_queue *e = q->elevator;
469 struct blk_mq_ctx *ctx = rq->mq_ctx;
470 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
472 /* flush rq in flush machinery need to be dispatched directly */
473 if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) {
474 blk_insert_flush(rq);
478 WARN_ON(e && (rq->tag != BLK_MQ_NO_TAG));
480 if (blk_mq_sched_bypass_insert(hctx, !!e, rq)) {
482 * Firstly normal IO request is inserted to scheduler queue or
483 * sw queue, meantime we add flush request to dispatch queue(
484 * hctx->dispatch) directly and there is at most one in-flight
485 * flush request for each hw queue, so it doesn't matter to add
486 * flush request to tail or front of the dispatch queue.
488 * Secondly in case of NCQ, flush request belongs to non-NCQ
489 * command, and queueing it will fail when there is any
490 * in-flight normal IO request(NCQ command). When adding flush
491 * rq to the front of hctx->dispatch, it is easier to introduce
492 * extra time to flush rq's latency because of S_SCHED_RESTART
493 * compared with adding to the tail of dispatch queue, then
494 * chance of flush merge is increased, and less flush requests
495 * will be issued to controller. It is observed that ~10% time
496 * is saved in blktests block/004 on disk attached to AHCI/NCQ
497 * drive when adding flush rq to the front of hctx->dispatch.
499 * Simply queue flush rq to the front of hctx->dispatch so that
500 * intensive flush workloads can benefit in case of NCQ HW.
502 at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head;
503 blk_mq_request_bypass_insert(rq, at_head, false);
507 if (e && e->type->ops.insert_requests) {
510 list_add(&rq->queuelist, &list);
511 e->type->ops.insert_requests(hctx, &list, at_head);
513 spin_lock(&ctx->lock);
514 __blk_mq_insert_request(hctx, rq, at_head);
515 spin_unlock(&ctx->lock);
520 blk_mq_run_hw_queue(hctx, async);
523 void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
524 struct blk_mq_ctx *ctx,
525 struct list_head *list, bool run_queue_async)
527 struct elevator_queue *e;
528 struct request_queue *q = hctx->queue;
531 * blk_mq_sched_insert_requests() is called from flush plug
532 * context only, and hold one usage counter to prevent queue
533 * from being released.
535 percpu_ref_get(&q->q_usage_counter);
537 e = hctx->queue->elevator;
538 if (e && e->type->ops.insert_requests)
539 e->type->ops.insert_requests(hctx, list, false);
542 * try to issue requests directly if the hw queue isn't
543 * busy in case of 'none' scheduler, and this way may save
544 * us one extra enqueue & dequeue to sw queue.
546 if (!hctx->dispatch_busy && !e && !run_queue_async) {
547 blk_mq_try_issue_list_directly(hctx, list);
548 if (list_empty(list))
551 blk_mq_insert_requests(hctx, ctx, list);
554 blk_mq_run_hw_queue(hctx, run_queue_async);
556 percpu_ref_put(&q->q_usage_counter);
559 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
560 struct blk_mq_hw_ctx *hctx,
561 unsigned int hctx_idx)
563 unsigned int flags = set->flags & ~BLK_MQ_F_TAG_HCTX_SHARED;
565 if (hctx->sched_tags) {
566 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
567 blk_mq_free_rq_map(hctx->sched_tags, flags);
568 hctx->sched_tags = NULL;
572 static int blk_mq_sched_alloc_tags(struct request_queue *q,
573 struct blk_mq_hw_ctx *hctx,
574 unsigned int hctx_idx)
576 struct blk_mq_tag_set *set = q->tag_set;
577 /* Clear HCTX_SHARED so tags are init'ed */
578 unsigned int flags = set->flags & ~BLK_MQ_F_TAG_HCTX_SHARED;
581 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
582 set->reserved_tags, flags);
583 if (!hctx->sched_tags)
586 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
588 blk_mq_sched_free_tags(set, hctx, hctx_idx);
593 /* called in queue's release handler, tagset has gone away */
594 static void blk_mq_sched_tags_teardown(struct request_queue *q)
596 struct blk_mq_hw_ctx *hctx;
599 queue_for_each_hw_ctx(q, hctx, i) {
600 /* Clear HCTX_SHARED so tags are freed */
601 unsigned int flags = hctx->flags & ~BLK_MQ_F_TAG_HCTX_SHARED;
603 if (hctx->sched_tags) {
604 blk_mq_free_rq_map(hctx->sched_tags, flags);
605 hctx->sched_tags = NULL;
610 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
612 struct blk_mq_hw_ctx *hctx;
613 struct elevator_queue *eq;
619 q->nr_requests = q->tag_set->queue_depth;
624 * Default to double of smaller one between hw queue_depth and 128,
625 * since we don't split into sync/async like the old code did.
626 * Additionally, this is a per-hw queue depth.
628 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
631 queue_for_each_hw_ctx(q, hctx, i) {
632 ret = blk_mq_sched_alloc_tags(q, hctx, i);
637 ret = e->ops.init_sched(q, e);
641 blk_mq_debugfs_register_sched(q);
643 queue_for_each_hw_ctx(q, hctx, i) {
644 if (e->ops.init_hctx) {
645 ret = e->ops.init_hctx(hctx, i);
648 blk_mq_sched_free_requests(q);
649 blk_mq_exit_sched(q, eq);
650 kobject_put(&eq->kobj);
654 blk_mq_debugfs_register_sched_hctx(q, hctx);
660 blk_mq_sched_free_requests(q);
661 blk_mq_sched_tags_teardown(q);
667 * called in either blk_queue_cleanup or elevator_switch, tagset
668 * is required for freeing requests
670 void blk_mq_sched_free_requests(struct request_queue *q)
672 struct blk_mq_hw_ctx *hctx;
675 queue_for_each_hw_ctx(q, hctx, i) {
676 if (hctx->sched_tags)
677 blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i);
681 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
683 struct blk_mq_hw_ctx *hctx;
686 queue_for_each_hw_ctx(q, hctx, i) {
687 blk_mq_debugfs_unregister_sched_hctx(hctx);
688 if (e->type->ops.exit_hctx && hctx->sched_data) {
689 e->type->ops.exit_hctx(hctx, i);
690 hctx->sched_data = NULL;
693 blk_mq_debugfs_unregister_sched(q);
694 if (e->type->ops.exit_sched)
695 e->type->ops.exit_sched(e);
696 blk_mq_sched_tags_teardown(q);