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.
50 void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
52 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
55 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
57 EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
59 void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
61 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
63 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
66 * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch)
67 * in blk_mq_run_hw_queue(). Its pair is the barrier in
68 * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART,
69 * meantime new request added to hctx->dispatch is missed to check in
70 * blk_mq_run_hw_queue().
74 blk_mq_run_hw_queue(hctx, true);
77 static int sched_rq_cmp(void *priv, const struct list_head *a,
78 const struct list_head *b)
80 struct request *rqa = container_of(a, struct request, queuelist);
81 struct request *rqb = container_of(b, struct request, queuelist);
83 return rqa->mq_hctx > rqb->mq_hctx;
86 static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list)
88 struct blk_mq_hw_ctx *hctx =
89 list_first_entry(rq_list, struct request, queuelist)->mq_hctx;
92 unsigned int count = 0;
94 list_for_each_entry(rq, rq_list, queuelist) {
95 if (rq->mq_hctx != hctx) {
96 list_cut_before(&hctx_list, rq_list, &rq->queuelist);
101 list_splice_tail_init(rq_list, &hctx_list);
104 return blk_mq_dispatch_rq_list(hctx, &hctx_list, count);
107 #define BLK_MQ_BUDGET_DELAY 3 /* ms units */
110 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
111 * its queue by itself in its completion handler, so we don't need to
112 * restart queue if .get_budget() fails to get the budget.
114 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
115 * be run again. This is necessary to avoid starving flushes.
117 static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
119 struct request_queue *q = hctx->queue;
120 struct elevator_queue *e = q->elevator;
121 bool multi_hctxs = false, run_queue = false;
122 bool dispatched = false, busy = false;
123 unsigned int max_dispatch;
127 if (hctx->dispatch_busy)
130 max_dispatch = hctx->queue->nr_requests;
136 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
139 if (!list_empty_careful(&hctx->dispatch)) {
144 budget_token = blk_mq_get_dispatch_budget(q);
145 if (budget_token < 0)
148 rq = e->type->ops.dispatch_request(hctx);
150 blk_mq_put_dispatch_budget(q, budget_token);
152 * We're releasing without dispatching. Holding the
153 * budget could have blocked any "hctx"s with the
154 * same queue and if we didn't dispatch then there's
155 * no guarantee anyone will kick the queue. Kick it
162 blk_mq_set_rq_budget_token(rq, budget_token);
165 * Now this rq owns the budget which has to be released
166 * if this rq won't be queued to driver via .queue_rq()
167 * in blk_mq_dispatch_rq_list().
169 list_add_tail(&rq->queuelist, &rq_list);
171 if (rq->mq_hctx != hctx)
175 * If we cannot get tag for the request, stop dequeueing
176 * requests from the IO scheduler. We are unlikely to be able
177 * to submit them anyway and it creates false impression for
178 * scheduling heuristics that the device can take more IO.
180 if (!blk_mq_get_driver_tag(rq))
182 } while (count < max_dispatch);
186 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
187 } else if (multi_hctxs) {
189 * Requests from different hctx may be dequeued from some
190 * schedulers, such as bfq and deadline.
192 * Sort the requests in the list according to their hctx,
193 * dispatch batching requests from same hctx at a time.
195 list_sort(NULL, &rq_list, sched_rq_cmp);
197 dispatched |= blk_mq_dispatch_hctx_list(&rq_list);
198 } while (!list_empty(&rq_list));
200 dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, count);
208 static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
210 unsigned long end = jiffies + HZ;
214 ret = __blk_mq_do_dispatch_sched(hctx);
217 if (need_resched() || time_is_before_jiffies(end)) {
218 blk_mq_delay_run_hw_queue(hctx, 0);
226 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
227 struct blk_mq_ctx *ctx)
229 unsigned short idx = ctx->index_hw[hctx->type];
231 if (++idx == hctx->nr_ctx)
234 return hctx->ctxs[idx];
238 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
239 * its queue by itself in its completion handler, so we don't need to
240 * restart queue if .get_budget() fails to get the budget.
242 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
243 * be run again. This is necessary to avoid starving flushes.
245 static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
247 struct request_queue *q = hctx->queue;
249 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
256 if (!list_empty_careful(&hctx->dispatch)) {
261 if (!sbitmap_any_bit_set(&hctx->ctx_map))
264 budget_token = blk_mq_get_dispatch_budget(q);
265 if (budget_token < 0)
268 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
270 blk_mq_put_dispatch_budget(q, budget_token);
272 * We're releasing without dispatching. Holding the
273 * budget could have blocked any "hctx"s with the
274 * same queue and if we didn't dispatch then there's
275 * no guarantee anyone will kick the queue. Kick it
278 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
282 blk_mq_set_rq_budget_token(rq, budget_token);
285 * Now this rq owns the budget which has to be released
286 * if this rq won't be queued to driver via .queue_rq()
287 * in blk_mq_dispatch_rq_list().
289 list_add(&rq->queuelist, &rq_list);
291 /* round robin for fair dispatch */
292 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
294 } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1));
296 WRITE_ONCE(hctx->dispatch_from, ctx);
300 static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
302 struct request_queue *q = hctx->queue;
303 const bool has_sched = q->elevator;
308 * If we have previous entries on our dispatch list, grab them first for
309 * more fair dispatch.
311 if (!list_empty_careful(&hctx->dispatch)) {
312 spin_lock(&hctx->lock);
313 if (!list_empty(&hctx->dispatch))
314 list_splice_init(&hctx->dispatch, &rq_list);
315 spin_unlock(&hctx->lock);
319 * Only ask the scheduler for requests, if we didn't have residual
320 * requests from the dispatch list. This is to avoid the case where
321 * we only ever dispatch a fraction of the requests available because
322 * of low device queue depth. Once we pull requests out of the IO
323 * scheduler, we can no longer merge or sort them. So it's best to
324 * leave them there for as long as we can. Mark the hw queue as
325 * needing a restart in that case.
327 * We want to dispatch from the scheduler if there was nothing
328 * on the dispatch list or we were able to dispatch from the
331 if (!list_empty(&rq_list)) {
332 blk_mq_sched_mark_restart_hctx(hctx);
333 if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) {
335 ret = blk_mq_do_dispatch_sched(hctx);
337 ret = blk_mq_do_dispatch_ctx(hctx);
339 } else if (has_sched) {
340 ret = blk_mq_do_dispatch_sched(hctx);
341 } else if (hctx->dispatch_busy) {
342 /* dequeue request one by one from sw queue if queue is busy */
343 ret = blk_mq_do_dispatch_ctx(hctx);
345 blk_mq_flush_busy_ctxs(hctx, &rq_list);
346 blk_mq_dispatch_rq_list(hctx, &rq_list, 0);
352 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
354 struct request_queue *q = hctx->queue;
356 /* RCU or SRCU read lock is needed before checking quiesced flag */
357 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
363 * A return of -EAGAIN is an indication that hctx->dispatch is not
364 * empty and we must run again in order to avoid starving flushes.
366 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) {
367 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN)
368 blk_mq_run_hw_queue(hctx, true);
372 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
373 unsigned int nr_segs)
375 struct elevator_queue *e = q->elevator;
376 struct blk_mq_ctx *ctx;
377 struct blk_mq_hw_ctx *hctx;
381 if (e && e->type->ops.bio_merge)
382 return e->type->ops.bio_merge(q, bio, nr_segs);
384 ctx = blk_mq_get_ctx(q);
385 hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
387 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE) ||
388 list_empty_careful(&ctx->rq_lists[type]))
391 /* default per sw-queue merge */
392 spin_lock(&ctx->lock);
394 * Reverse check our software queue for entries that we could
395 * potentially merge with. Currently includes a hand-wavy stop
396 * count of 8, to not spend too much time checking for merges.
398 if (blk_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) {
403 spin_unlock(&ctx->lock);
408 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq,
409 struct list_head *free)
411 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq, free);
413 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
415 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
419 * dispatch flush and passthrough rq directly
421 * passthrough request has to be added to hctx->dispatch directly.
422 * For some reason, device may be in one situation which can't
423 * handle FS request, so STS_RESOURCE is always returned and the
424 * FS request will be added to hctx->dispatch. However passthrough
425 * request may be required at that time for fixing the problem. If
426 * passthrough request is added to scheduler queue, there isn't any
427 * chance to dispatch it given we prioritize requests in hctx->dispatch.
429 if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq))
435 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
436 bool run_queue, bool async)
438 struct request_queue *q = rq->q;
439 struct elevator_queue *e = q->elevator;
440 struct blk_mq_ctx *ctx = rq->mq_ctx;
441 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
443 WARN_ON(e && (rq->tag != BLK_MQ_NO_TAG));
445 if (blk_mq_sched_bypass_insert(hctx, rq)) {
447 * Firstly normal IO request is inserted to scheduler queue or
448 * sw queue, meantime we add flush request to dispatch queue(
449 * hctx->dispatch) directly and there is at most one in-flight
450 * flush request for each hw queue, so it doesn't matter to add
451 * flush request to tail or front of the dispatch queue.
453 * Secondly in case of NCQ, flush request belongs to non-NCQ
454 * command, and queueing it will fail when there is any
455 * in-flight normal IO request(NCQ command). When adding flush
456 * rq to the front of hctx->dispatch, it is easier to introduce
457 * extra time to flush rq's latency because of S_SCHED_RESTART
458 * compared with adding to the tail of dispatch queue, then
459 * chance of flush merge is increased, and less flush requests
460 * will be issued to controller. It is observed that ~10% time
461 * is saved in blktests block/004 on disk attached to AHCI/NCQ
462 * drive when adding flush rq to the front of hctx->dispatch.
464 * Simply queue flush rq to the front of hctx->dispatch so that
465 * intensive flush workloads can benefit in case of NCQ HW.
467 at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head;
468 blk_mq_request_bypass_insert(rq, at_head, false);
475 list_add(&rq->queuelist, &list);
476 e->type->ops.insert_requests(hctx, &list, at_head);
478 spin_lock(&ctx->lock);
479 __blk_mq_insert_request(hctx, rq, at_head);
480 spin_unlock(&ctx->lock);
485 blk_mq_run_hw_queue(hctx, async);
488 void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
489 struct blk_mq_ctx *ctx,
490 struct list_head *list, bool run_queue_async)
492 struct elevator_queue *e;
493 struct request_queue *q = hctx->queue;
496 * blk_mq_sched_insert_requests() is called from flush plug
497 * context only, and hold one usage counter to prevent queue
498 * from being released.
500 percpu_ref_get(&q->q_usage_counter);
502 e = hctx->queue->elevator;
504 e->type->ops.insert_requests(hctx, list, false);
507 * try to issue requests directly if the hw queue isn't
508 * busy in case of 'none' scheduler, and this way may save
509 * us one extra enqueue & dequeue to sw queue.
511 if (!hctx->dispatch_busy && !e && !run_queue_async) {
512 blk_mq_try_issue_list_directly(hctx, list);
513 if (list_empty(list))
516 blk_mq_insert_requests(hctx, ctx, list);
519 blk_mq_run_hw_queue(hctx, run_queue_async);
521 percpu_ref_put(&q->q_usage_counter);
524 static int blk_mq_sched_alloc_tags(struct request_queue *q,
525 struct blk_mq_hw_ctx *hctx,
526 unsigned int hctx_idx)
528 struct blk_mq_tag_set *set = q->tag_set;
531 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
532 set->reserved_tags, set->flags);
533 if (!hctx->sched_tags)
536 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
538 blk_mq_free_rq_map(hctx->sched_tags, set->flags);
539 hctx->sched_tags = NULL;
545 /* called in queue's release handler, tagset has gone away */
546 static void blk_mq_sched_tags_teardown(struct request_queue *q)
548 struct blk_mq_hw_ctx *hctx;
551 queue_for_each_hw_ctx(q, hctx, i) {
552 if (hctx->sched_tags) {
553 blk_mq_free_rq_map(hctx->sched_tags, hctx->flags);
554 hctx->sched_tags = NULL;
559 static int blk_mq_init_sched_shared_sbitmap(struct request_queue *queue)
561 struct blk_mq_tag_set *set = queue->tag_set;
562 int alloc_policy = BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags);
563 struct blk_mq_hw_ctx *hctx;
567 * Set initial depth at max so that we don't need to reallocate for
568 * updating nr_requests.
570 ret = blk_mq_init_bitmaps(&queue->sched_bitmap_tags,
571 &queue->sched_breserved_tags,
572 MAX_SCHED_RQ, set->reserved_tags,
573 set->numa_node, alloc_policy);
577 queue_for_each_hw_ctx(queue, hctx, i) {
578 hctx->sched_tags->bitmap_tags =
579 &queue->sched_bitmap_tags;
580 hctx->sched_tags->breserved_tags =
581 &queue->sched_breserved_tags;
584 sbitmap_queue_resize(&queue->sched_bitmap_tags,
585 queue->nr_requests - set->reserved_tags);
590 static void blk_mq_exit_sched_shared_sbitmap(struct request_queue *queue)
592 sbitmap_queue_free(&queue->sched_bitmap_tags);
593 sbitmap_queue_free(&queue->sched_breserved_tags);
596 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
598 struct blk_mq_hw_ctx *hctx;
599 struct elevator_queue *eq;
605 q->nr_requests = q->tag_set->queue_depth;
610 * Default to double of smaller one between hw queue_depth and 128,
611 * since we don't split into sync/async like the old code did.
612 * Additionally, this is a per-hw queue depth.
614 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
617 queue_for_each_hw_ctx(q, hctx, i) {
618 ret = blk_mq_sched_alloc_tags(q, hctx, i);
623 if (blk_mq_is_sbitmap_shared(q->tag_set->flags)) {
624 ret = blk_mq_init_sched_shared_sbitmap(q);
629 ret = e->ops.init_sched(q, e);
631 goto err_free_sbitmap;
633 blk_mq_debugfs_register_sched(q);
635 queue_for_each_hw_ctx(q, hctx, i) {
636 if (e->ops.init_hctx) {
637 ret = e->ops.init_hctx(hctx, i);
640 blk_mq_sched_free_requests(q);
641 blk_mq_exit_sched(q, eq);
642 kobject_put(&eq->kobj);
646 blk_mq_debugfs_register_sched_hctx(q, hctx);
652 if (blk_mq_is_sbitmap_shared(q->tag_set->flags))
653 blk_mq_exit_sched_shared_sbitmap(q);
655 blk_mq_sched_free_requests(q);
656 blk_mq_sched_tags_teardown(q);
662 * called in either blk_queue_cleanup or elevator_switch, tagset
663 * is required for freeing requests
665 void blk_mq_sched_free_requests(struct request_queue *q)
667 struct blk_mq_hw_ctx *hctx;
670 queue_for_each_hw_ctx(q, hctx, i) {
671 if (hctx->sched_tags)
672 blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i);
676 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
678 struct blk_mq_hw_ctx *hctx;
680 unsigned int flags = 0;
682 queue_for_each_hw_ctx(q, hctx, i) {
683 blk_mq_debugfs_unregister_sched_hctx(hctx);
684 if (e->type->ops.exit_hctx && hctx->sched_data) {
685 e->type->ops.exit_hctx(hctx, i);
686 hctx->sched_data = NULL;
690 blk_mq_debugfs_unregister_sched(q);
691 if (e->type->ops.exit_sched)
692 e->type->ops.exit_sched(e);
693 blk_mq_sched_tags_teardown(q);
694 if (blk_mq_is_sbitmap_shared(flags))
695 blk_mq_exit_sched_shared_sbitmap(q);