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 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
65 * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch)
66 * in blk_mq_run_hw_queue(). Its pair is the barrier in
67 * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART,
68 * meantime new request added to hctx->dispatch is missed to check in
69 * blk_mq_run_hw_queue().
73 blk_mq_run_hw_queue(hctx, true);
76 static int sched_rq_cmp(void *priv, const struct list_head *a,
77 const struct list_head *b)
79 struct request *rqa = container_of(a, struct request, queuelist);
80 struct request *rqb = container_of(b, struct request, queuelist);
82 return rqa->mq_hctx > rqb->mq_hctx;
85 static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list)
87 struct blk_mq_hw_ctx *hctx =
88 list_first_entry(rq_list, struct request, queuelist)->mq_hctx;
91 unsigned int count = 0;
93 list_for_each_entry(rq, rq_list, queuelist) {
94 if (rq->mq_hctx != hctx) {
95 list_cut_before(&hctx_list, rq_list, &rq->queuelist);
100 list_splice_tail_init(rq_list, &hctx_list);
103 return blk_mq_dispatch_rq_list(hctx, &hctx_list, count);
106 #define BLK_MQ_BUDGET_DELAY 3 /* ms units */
109 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
110 * its queue by itself in its completion handler, so we don't need to
111 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
113 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
114 * be run again. This is necessary to avoid starving flushes.
116 static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
118 struct request_queue *q = hctx->queue;
119 struct elevator_queue *e = q->elevator;
120 bool multi_hctxs = false, run_queue = false;
121 bool dispatched = false, busy = false;
122 unsigned int max_dispatch;
126 if (hctx->dispatch_busy)
129 max_dispatch = hctx->queue->nr_requests;
135 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
138 if (!list_empty_careful(&hctx->dispatch)) {
143 budget_token = blk_mq_get_dispatch_budget(q);
144 if (budget_token < 0)
147 rq = e->type->ops.dispatch_request(hctx);
149 blk_mq_put_dispatch_budget(q, budget_token);
151 * We're releasing without dispatching. Holding the
152 * budget could have blocked any "hctx"s with the
153 * same queue and if we didn't dispatch then there's
154 * no guarantee anyone will kick the queue. Kick it
161 blk_mq_set_rq_budget_token(rq, budget_token);
164 * Now this rq owns the budget which has to be released
165 * if this rq won't be queued to driver via .queue_rq()
166 * in blk_mq_dispatch_rq_list().
168 list_add_tail(&rq->queuelist, &rq_list);
170 if (rq->mq_hctx != hctx)
174 * If we cannot get tag for the request, stop dequeueing
175 * requests from the IO scheduler. We are unlikely to be able
176 * to submit them anyway and it creates false impression for
177 * scheduling heuristics that the device can take more IO.
179 if (!blk_mq_get_driver_tag(rq))
181 } while (count < max_dispatch);
185 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
186 } else if (multi_hctxs) {
188 * Requests from different hctx may be dequeued from some
189 * schedulers, such as bfq and deadline.
191 * Sort the requests in the list according to their hctx,
192 * dispatch batching requests from same hctx at a time.
194 list_sort(NULL, &rq_list, sched_rq_cmp);
196 dispatched |= blk_mq_dispatch_hctx_list(&rq_list);
197 } while (!list_empty(&rq_list));
199 dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, count);
207 static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
212 ret = __blk_mq_do_dispatch_sched(hctx);
218 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
219 struct blk_mq_ctx *ctx)
221 unsigned short idx = ctx->index_hw[hctx->type];
223 if (++idx == hctx->nr_ctx)
226 return hctx->ctxs[idx];
230 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
231 * its queue by itself in its completion handler, so we don't need to
232 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
234 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
235 * be run again. This is necessary to avoid starving flushes.
237 static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
239 struct request_queue *q = hctx->queue;
241 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 budget_token = blk_mq_get_dispatch_budget(q);
257 if (budget_token < 0)
260 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
262 blk_mq_put_dispatch_budget(q, budget_token);
264 * We're releasing without dispatching. Holding the
265 * budget could have blocked any "hctx"s with the
266 * same queue and if we didn't dispatch then there's
267 * no guarantee anyone will kick the queue. Kick it
270 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
274 blk_mq_set_rq_budget_token(rq, budget_token);
277 * Now this rq owns the budget which has to be released
278 * if this rq won't be queued to driver via .queue_rq()
279 * in blk_mq_dispatch_rq_list().
281 list_add(&rq->queuelist, &rq_list);
283 /* round robin for fair dispatch */
284 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
286 } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1));
288 WRITE_ONCE(hctx->dispatch_from, ctx);
292 static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
294 struct request_queue *q = hctx->queue;
295 const bool has_sched = q->elevator;
300 * If we have previous entries on our dispatch list, grab them first for
301 * more fair dispatch.
303 if (!list_empty_careful(&hctx->dispatch)) {
304 spin_lock(&hctx->lock);
305 if (!list_empty(&hctx->dispatch))
306 list_splice_init(&hctx->dispatch, &rq_list);
307 spin_unlock(&hctx->lock);
311 * Only ask the scheduler for requests, if we didn't have residual
312 * requests from the dispatch list. This is to avoid the case where
313 * we only ever dispatch a fraction of the requests available because
314 * of low device queue depth. Once we pull requests out of the IO
315 * scheduler, we can no longer merge or sort them. So it's best to
316 * leave them there for as long as we can. Mark the hw queue as
317 * needing a restart in that case.
319 * We want to dispatch from the scheduler if there was nothing
320 * on the dispatch list or we were able to dispatch from the
323 if (!list_empty(&rq_list)) {
324 blk_mq_sched_mark_restart_hctx(hctx);
325 if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) {
327 ret = blk_mq_do_dispatch_sched(hctx);
329 ret = blk_mq_do_dispatch_ctx(hctx);
331 } else if (has_sched) {
332 ret = blk_mq_do_dispatch_sched(hctx);
333 } else if (hctx->dispatch_busy) {
334 /* dequeue request one by one from sw queue if queue is busy */
335 ret = blk_mq_do_dispatch_ctx(hctx);
337 blk_mq_flush_busy_ctxs(hctx, &rq_list);
338 blk_mq_dispatch_rq_list(hctx, &rq_list, 0);
344 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
346 struct request_queue *q = hctx->queue;
348 /* RCU or SRCU read lock is needed before checking quiesced flag */
349 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
355 * A return of -EAGAIN is an indication that hctx->dispatch is not
356 * empty and we must run again in order to avoid starving flushes.
358 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) {
359 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN)
360 blk_mq_run_hw_queue(hctx, true);
364 bool blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
365 unsigned int nr_segs)
367 struct elevator_queue *e = q->elevator;
368 struct blk_mq_ctx *ctx;
369 struct blk_mq_hw_ctx *hctx;
373 if (e && e->type->ops.bio_merge) {
374 ret = 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))
395 spin_unlock(&ctx->lock);
400 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq,
401 struct list_head *free)
403 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq, free);
405 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
407 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
411 * dispatch flush and passthrough rq directly
413 * passthrough request has to be added to hctx->dispatch directly.
414 * For some reason, device may be in one situation which can't
415 * handle FS request, so STS_RESOURCE is always returned and the
416 * FS request will be added to hctx->dispatch. However passthrough
417 * request may be required at that time for fixing the problem. If
418 * passthrough request is added to scheduler queue, there isn't any
419 * chance to dispatch it given we prioritize requests in hctx->dispatch.
421 if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq))
427 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
428 bool run_queue, bool async)
430 struct request_queue *q = rq->q;
431 struct elevator_queue *e = q->elevator;
432 struct blk_mq_ctx *ctx = rq->mq_ctx;
433 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
435 WARN_ON(e && (rq->tag != BLK_MQ_NO_TAG));
437 if (blk_mq_sched_bypass_insert(hctx, rq)) {
439 * Firstly normal IO request is inserted to scheduler queue or
440 * sw queue, meantime we add flush request to dispatch queue(
441 * hctx->dispatch) directly and there is at most one in-flight
442 * flush request for each hw queue, so it doesn't matter to add
443 * flush request to tail or front of the dispatch queue.
445 * Secondly in case of NCQ, flush request belongs to non-NCQ
446 * command, and queueing it will fail when there is any
447 * in-flight normal IO request(NCQ command). When adding flush
448 * rq to the front of hctx->dispatch, it is easier to introduce
449 * extra time to flush rq's latency because of S_SCHED_RESTART
450 * compared with adding to the tail of dispatch queue, then
451 * chance of flush merge is increased, and less flush requests
452 * will be issued to controller. It is observed that ~10% time
453 * is saved in blktests block/004 on disk attached to AHCI/NCQ
454 * drive when adding flush rq to the front of hctx->dispatch.
456 * Simply queue flush rq to the front of hctx->dispatch so that
457 * intensive flush workloads can benefit in case of NCQ HW.
459 at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head;
460 blk_mq_request_bypass_insert(rq, at_head, false);
467 list_add(&rq->queuelist, &list);
468 e->type->ops.insert_requests(hctx, &list, at_head);
470 spin_lock(&ctx->lock);
471 __blk_mq_insert_request(hctx, rq, at_head);
472 spin_unlock(&ctx->lock);
477 blk_mq_run_hw_queue(hctx, async);
480 void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
481 struct blk_mq_ctx *ctx,
482 struct list_head *list, bool run_queue_async)
484 struct elevator_queue *e;
485 struct request_queue *q = hctx->queue;
488 * blk_mq_sched_insert_requests() is called from flush plug
489 * context only, and hold one usage counter to prevent queue
490 * from being released.
492 percpu_ref_get(&q->q_usage_counter);
494 e = hctx->queue->elevator;
496 e->type->ops.insert_requests(hctx, list, false);
499 * try to issue requests directly if the hw queue isn't
500 * busy in case of 'none' scheduler, and this way may save
501 * us one extra enqueue & dequeue to sw queue.
503 if (!hctx->dispatch_busy && !run_queue_async) {
504 blk_mq_try_issue_list_directly(hctx, list);
505 if (list_empty(list))
508 blk_mq_insert_requests(hctx, ctx, list);
511 blk_mq_run_hw_queue(hctx, run_queue_async);
513 percpu_ref_put(&q->q_usage_counter);
516 static int blk_mq_sched_alloc_map_and_rqs(struct request_queue *q,
517 struct blk_mq_hw_ctx *hctx,
518 unsigned int hctx_idx)
520 if (blk_mq_is_shared_tags(q->tag_set->flags)) {
521 hctx->sched_tags = q->sched_shared_tags;
525 hctx->sched_tags = blk_mq_alloc_map_and_rqs(q->tag_set, hctx_idx,
528 if (!hctx->sched_tags)
533 static void blk_mq_exit_sched_shared_tags(struct request_queue *queue)
535 blk_mq_free_rq_map(queue->sched_shared_tags);
536 queue->sched_shared_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, unsigned int flags)
542 struct blk_mq_hw_ctx *hctx;
545 queue_for_each_hw_ctx(q, hctx, i) {
546 if (hctx->sched_tags) {
547 if (!blk_mq_is_shared_tags(flags))
548 blk_mq_free_rq_map(hctx->sched_tags);
549 hctx->sched_tags = NULL;
553 if (blk_mq_is_shared_tags(flags))
554 blk_mq_exit_sched_shared_tags(q);
557 static int blk_mq_init_sched_shared_tags(struct request_queue *queue)
559 struct blk_mq_tag_set *set = queue->tag_set;
562 * Set initial depth at max so that we don't need to reallocate for
563 * updating nr_requests.
565 queue->sched_shared_tags = blk_mq_alloc_map_and_rqs(set,
568 if (!queue->sched_shared_tags)
571 blk_mq_tag_update_sched_shared_tags(queue);
576 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
578 unsigned int i, flags = q->tag_set->flags;
579 struct blk_mq_hw_ctx *hctx;
580 struct elevator_queue *eq;
585 q->nr_requests = q->tag_set->queue_depth;
590 * Default to double of smaller one between hw queue_depth and 128,
591 * since we don't split into sync/async like the old code did.
592 * Additionally, this is a per-hw queue depth.
594 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
597 if (blk_mq_is_shared_tags(flags)) {
598 ret = blk_mq_init_sched_shared_tags(q);
603 queue_for_each_hw_ctx(q, hctx, i) {
604 ret = blk_mq_sched_alloc_map_and_rqs(q, hctx, i);
606 goto err_free_map_and_rqs;
609 ret = e->ops.init_sched(q, e);
611 goto err_free_map_and_rqs;
613 blk_mq_debugfs_register_sched(q);
615 queue_for_each_hw_ctx(q, hctx, i) {
616 if (e->ops.init_hctx) {
617 ret = e->ops.init_hctx(hctx, i);
620 blk_mq_sched_free_rqs(q);
621 blk_mq_exit_sched(q, eq);
622 kobject_put(&eq->kobj);
626 blk_mq_debugfs_register_sched_hctx(q, hctx);
631 err_free_map_and_rqs:
632 blk_mq_sched_free_rqs(q);
633 blk_mq_sched_tags_teardown(q, flags);
640 * called in either blk_queue_cleanup or elevator_switch, tagset
641 * is required for freeing requests
643 void blk_mq_sched_free_rqs(struct request_queue *q)
645 struct blk_mq_hw_ctx *hctx;
648 if (blk_mq_is_shared_tags(q->tag_set->flags)) {
649 blk_mq_free_rqs(q->tag_set, q->sched_shared_tags,
652 queue_for_each_hw_ctx(q, hctx, i) {
653 if (hctx->sched_tags)
654 blk_mq_free_rqs(q->tag_set,
655 hctx->sched_tags, i);
660 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
662 struct blk_mq_hw_ctx *hctx;
664 unsigned int flags = 0;
666 queue_for_each_hw_ctx(q, hctx, i) {
667 blk_mq_debugfs_unregister_sched_hctx(hctx);
668 if (e->type->ops.exit_hctx && hctx->sched_data) {
669 e->type->ops.exit_hctx(hctx, i);
670 hctx->sched_data = NULL;
674 blk_mq_debugfs_unregister_sched(q);
675 if (e->type->ops.exit_sched)
676 e->type->ops.exit_sched(e);
677 blk_mq_sched_tags_teardown(q, flags);