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>
11 #include <trace/events/block.h>
15 #include "blk-mq-debugfs.h"
16 #include "blk-mq-sched.h"
17 #include "blk-mq-tag.h"
20 void blk_mq_sched_free_hctx_data(struct request_queue *q,
21 void (*exit)(struct blk_mq_hw_ctx *))
23 struct blk_mq_hw_ctx *hctx;
26 queue_for_each_hw_ctx(q, hctx, i) {
27 if (exit && hctx->sched_data)
29 kfree(hctx->sched_data);
30 hctx->sched_data = NULL;
33 EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data);
35 void blk_mq_sched_assign_ioc(struct request *rq)
37 struct request_queue *q = rq->q;
38 struct io_context *ioc;
42 * May not have an IO context if it's a passthrough request
44 ioc = current->io_context;
48 spin_lock_irq(&q->queue_lock);
49 icq = ioc_lookup_icq(ioc, q);
50 spin_unlock_irq(&q->queue_lock);
53 icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
57 get_io_context(icq->ioc);
62 * Mark a hardware queue as needing a restart.
64 void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
66 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
69 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
71 EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
73 void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
75 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
77 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
80 * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch)
81 * in blk_mq_run_hw_queue(). Its pair is the barrier in
82 * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART,
83 * meantime new request added to hctx->dispatch is missed to check in
84 * blk_mq_run_hw_queue().
88 blk_mq_run_hw_queue(hctx, true);
92 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
93 * its queue by itself in its completion handler, so we don't need to
94 * restart queue if .get_budget() fails to get the budget.
96 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
97 * be run again. This is necessary to avoid starving flushes.
99 static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
101 struct request_queue *q = hctx->queue;
102 struct elevator_queue *e = q->elevator;
109 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
112 if (!list_empty_careful(&hctx->dispatch)) {
117 if (!blk_mq_get_dispatch_budget(hctx))
120 rq = e->type->ops.dispatch_request(hctx);
122 blk_mq_put_dispatch_budget(hctx);
127 * Now this rq owns the budget which has to be released
128 * if this rq won't be queued to driver via .queue_rq()
129 * in blk_mq_dispatch_rq_list().
131 list_add(&rq->queuelist, &rq_list);
132 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
137 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
138 struct blk_mq_ctx *ctx)
140 unsigned short idx = ctx->index_hw[hctx->type];
142 if (++idx == hctx->nr_ctx)
145 return hctx->ctxs[idx];
149 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
150 * its queue by itself in its completion handler, so we don't need to
151 * restart queue if .get_budget() fails to get the budget.
153 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
154 * to be run again. This is necessary to avoid starving flushes.
156 static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
158 struct request_queue *q = hctx->queue;
160 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
166 if (!list_empty_careful(&hctx->dispatch)) {
171 if (!sbitmap_any_bit_set(&hctx->ctx_map))
174 if (!blk_mq_get_dispatch_budget(hctx))
177 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
179 blk_mq_put_dispatch_budget(hctx);
184 * Now this rq owns the budget which has to be released
185 * if this rq won't be queued to driver via .queue_rq()
186 * in blk_mq_dispatch_rq_list().
188 list_add(&rq->queuelist, &rq_list);
190 /* round robin for fair dispatch */
191 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
193 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
195 WRITE_ONCE(hctx->dispatch_from, ctx);
199 int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
201 struct request_queue *q = hctx->queue;
202 struct elevator_queue *e = q->elevator;
203 const bool has_sched_dispatch = e && e->type->ops.dispatch_request;
208 * If we have previous entries on our dispatch list, grab them first for
209 * more fair dispatch.
211 if (!list_empty_careful(&hctx->dispatch)) {
212 spin_lock(&hctx->lock);
213 if (!list_empty(&hctx->dispatch))
214 list_splice_init(&hctx->dispatch, &rq_list);
215 spin_unlock(&hctx->lock);
219 * Only ask the scheduler for requests, if we didn't have residual
220 * requests from the dispatch list. This is to avoid the case where
221 * we only ever dispatch a fraction of the requests available because
222 * of low device queue depth. Once we pull requests out of the IO
223 * scheduler, we can no longer merge or sort them. So it's best to
224 * leave them there for as long as we can. Mark the hw queue as
225 * needing a restart in that case.
227 * We want to dispatch from the scheduler if there was nothing
228 * on the dispatch list or we were able to dispatch from the
231 if (!list_empty(&rq_list)) {
232 blk_mq_sched_mark_restart_hctx(hctx);
233 if (blk_mq_dispatch_rq_list(q, &rq_list, false)) {
234 if (has_sched_dispatch)
235 ret = blk_mq_do_dispatch_sched(hctx);
237 ret = blk_mq_do_dispatch_ctx(hctx);
239 } else if (has_sched_dispatch) {
240 ret = blk_mq_do_dispatch_sched(hctx);
241 } else if (hctx->dispatch_busy) {
242 /* dequeue request one by one from sw queue if queue is busy */
243 ret = blk_mq_do_dispatch_ctx(hctx);
245 blk_mq_flush_busy_ctxs(hctx, &rq_list);
246 blk_mq_dispatch_rq_list(q, &rq_list, false);
252 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
254 struct request_queue *q = hctx->queue;
256 /* RCU or SRCU read lock is needed before checking quiesced flag */
257 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
263 * A return of -EAGAIN is an indication that hctx->dispatch is not
264 * empty and we must run again in order to avoid starving flushes.
266 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) {
267 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN)
268 blk_mq_run_hw_queue(hctx, true);
272 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
273 unsigned int nr_segs, struct request **merged_request)
277 switch (elv_merge(q, &rq, bio)) {
278 case ELEVATOR_BACK_MERGE:
279 if (!blk_mq_sched_allow_merge(q, rq, bio))
281 if (!bio_attempt_back_merge(rq, bio, nr_segs))
283 *merged_request = attempt_back_merge(q, rq);
284 if (!*merged_request)
285 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
287 case ELEVATOR_FRONT_MERGE:
288 if (!blk_mq_sched_allow_merge(q, rq, bio))
290 if (!bio_attempt_front_merge(rq, bio, nr_segs))
292 *merged_request = attempt_front_merge(q, rq);
293 if (!*merged_request)
294 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
296 case ELEVATOR_DISCARD_MERGE:
297 return bio_attempt_discard_merge(q, rq, bio);
302 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
305 * Iterate list of requests and see if we can merge this bio with any
308 bool blk_mq_bio_list_merge(struct request_queue *q, struct list_head *list,
309 struct bio *bio, unsigned int nr_segs)
314 list_for_each_entry_reverse(rq, list, queuelist) {
320 if (!blk_rq_merge_ok(rq, bio))
323 switch (blk_try_merge(rq, bio)) {
324 case ELEVATOR_BACK_MERGE:
325 if (blk_mq_sched_allow_merge(q, rq, bio))
326 merged = bio_attempt_back_merge(rq, bio,
329 case ELEVATOR_FRONT_MERGE:
330 if (blk_mq_sched_allow_merge(q, rq, bio))
331 merged = bio_attempt_front_merge(rq, bio,
334 case ELEVATOR_DISCARD_MERGE:
335 merged = bio_attempt_discard_merge(q, rq, bio);
346 EXPORT_SYMBOL_GPL(blk_mq_bio_list_merge);
349 * Reverse check our software queue for entries that we could potentially
350 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
351 * too much time checking for merges.
353 static bool blk_mq_attempt_merge(struct request_queue *q,
354 struct blk_mq_hw_ctx *hctx,
355 struct blk_mq_ctx *ctx, struct bio *bio,
356 unsigned int nr_segs)
358 enum hctx_type type = hctx->type;
360 lockdep_assert_held(&ctx->lock);
362 if (blk_mq_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) {
370 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
371 unsigned int nr_segs)
373 struct elevator_queue *e = q->elevator;
374 struct blk_mq_ctx *ctx;
375 struct blk_mq_hw_ctx *hctx;
379 if (e && e->type->ops.bio_merge)
380 return e->type->ops.bio_merge(q, bio, nr_segs);
382 ctx = blk_mq_get_ctx(q);
383 hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
385 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
386 !list_empty_careful(&ctx->rq_lists[type])) {
387 /* default per sw-queue merge */
388 spin_lock(&ctx->lock);
389 ret = blk_mq_attempt_merge(q, hctx, ctx, bio, nr_segs);
390 spin_unlock(&ctx->lock);
396 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
398 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
400 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
402 void blk_mq_sched_request_inserted(struct request *rq)
404 trace_block_rq_insert(rq->q, rq);
406 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
408 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))
427 rq->rq_flags |= RQF_SORTED;
432 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
433 bool run_queue, bool async)
435 struct request_queue *q = rq->q;
436 struct elevator_queue *e = q->elevator;
437 struct blk_mq_ctx *ctx = rq->mq_ctx;
438 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
440 /* flush rq in flush machinery need to be dispatched directly */
441 if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) {
442 blk_insert_flush(rq);
446 WARN_ON(e && (rq->tag != -1));
448 if (blk_mq_sched_bypass_insert(hctx, !!e, rq)) {
450 * Firstly normal IO request is inserted to scheduler queue or
451 * sw queue, meantime we add flush request to dispatch queue(
452 * hctx->dispatch) directly and there is at most one in-flight
453 * flush request for each hw queue, so it doesn't matter to add
454 * flush request to tail or front of the dispatch queue.
456 * Secondly in case of NCQ, flush request belongs to non-NCQ
457 * command, and queueing it will fail when there is any
458 * in-flight normal IO request(NCQ command). When adding flush
459 * rq to the front of hctx->dispatch, it is easier to introduce
460 * extra time to flush rq's latency because of S_SCHED_RESTART
461 * compared with adding to the tail of dispatch queue, then
462 * chance of flush merge is increased, and less flush requests
463 * will be issued to controller. It is observed that ~10% time
464 * is saved in blktests block/004 on disk attached to AHCI/NCQ
465 * drive when adding flush rq to the front of hctx->dispatch.
467 * Simply queue flush rq to the front of hctx->dispatch so that
468 * intensive flush workloads can benefit in case of NCQ HW.
470 at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head;
471 blk_mq_request_bypass_insert(rq, at_head, false);
475 if (e && e->type->ops.insert_requests) {
478 list_add(&rq->queuelist, &list);
479 e->type->ops.insert_requests(hctx, &list, at_head);
481 spin_lock(&ctx->lock);
482 __blk_mq_insert_request(hctx, rq, at_head);
483 spin_unlock(&ctx->lock);
488 blk_mq_run_hw_queue(hctx, async);
491 void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
492 struct blk_mq_ctx *ctx,
493 struct list_head *list, bool run_queue_async)
495 struct elevator_queue *e;
496 struct request_queue *q = hctx->queue;
499 * blk_mq_sched_insert_requests() is called from flush plug
500 * context only, and hold one usage counter to prevent queue
501 * from being released.
503 percpu_ref_get(&q->q_usage_counter);
505 e = hctx->queue->elevator;
506 if (e && e->type->ops.insert_requests)
507 e->type->ops.insert_requests(hctx, list, false);
510 * try to issue requests directly if the hw queue isn't
511 * busy in case of 'none' scheduler, and this way may save
512 * us one extra enqueue & dequeue to sw queue.
514 if (!hctx->dispatch_busy && !e && !run_queue_async) {
515 blk_mq_try_issue_list_directly(hctx, list);
516 if (list_empty(list))
519 blk_mq_insert_requests(hctx, ctx, list);
522 blk_mq_run_hw_queue(hctx, run_queue_async);
524 percpu_ref_put(&q->q_usage_counter);
527 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
528 struct blk_mq_hw_ctx *hctx,
529 unsigned int hctx_idx)
531 if (hctx->sched_tags) {
532 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
533 blk_mq_free_rq_map(hctx->sched_tags);
534 hctx->sched_tags = NULL;
538 static int blk_mq_sched_alloc_tags(struct request_queue *q,
539 struct blk_mq_hw_ctx *hctx,
540 unsigned int hctx_idx)
542 struct blk_mq_tag_set *set = q->tag_set;
545 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
547 if (!hctx->sched_tags)
550 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
552 blk_mq_sched_free_tags(set, hctx, hctx_idx);
557 /* called in queue's release handler, tagset has gone away */
558 static void blk_mq_sched_tags_teardown(struct request_queue *q)
560 struct blk_mq_hw_ctx *hctx;
563 queue_for_each_hw_ctx(q, hctx, i) {
564 if (hctx->sched_tags) {
565 blk_mq_free_rq_map(hctx->sched_tags);
566 hctx->sched_tags = NULL;
571 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
573 struct blk_mq_hw_ctx *hctx;
574 struct elevator_queue *eq;
580 q->nr_requests = q->tag_set->queue_depth;
585 * Default to double of smaller one between hw queue_depth and 128,
586 * since we don't split into sync/async like the old code did.
587 * Additionally, this is a per-hw queue depth.
589 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
592 queue_for_each_hw_ctx(q, hctx, i) {
593 ret = blk_mq_sched_alloc_tags(q, hctx, i);
598 ret = e->ops.init_sched(q, e);
602 blk_mq_debugfs_register_sched(q);
604 queue_for_each_hw_ctx(q, hctx, i) {
605 if (e->ops.init_hctx) {
606 ret = e->ops.init_hctx(hctx, i);
609 blk_mq_sched_free_requests(q);
610 blk_mq_exit_sched(q, eq);
611 kobject_put(&eq->kobj);
615 blk_mq_debugfs_register_sched_hctx(q, hctx);
621 blk_mq_sched_free_requests(q);
622 blk_mq_sched_tags_teardown(q);
628 * called in either blk_queue_cleanup or elevator_switch, tagset
629 * is required for freeing requests
631 void blk_mq_sched_free_requests(struct request_queue *q)
633 struct blk_mq_hw_ctx *hctx;
636 queue_for_each_hw_ctx(q, hctx, i) {
637 if (hctx->sched_tags)
638 blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i);
642 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
644 struct blk_mq_hw_ctx *hctx;
647 queue_for_each_hw_ctx(q, hctx, i) {
648 blk_mq_debugfs_unregister_sched_hctx(hctx);
649 if (e->type->ops.exit_hctx && hctx->sched_data) {
650 e->type->ops.exit_hctx(hctx, i);
651 hctx->sched_data = NULL;
654 blk_mq_debugfs_unregister_sched(q);
655 if (e->type->ops.exit_sched)
656 e->type->ops.exit_sched(e);
657 blk_mq_sched_tags_teardown(q);