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 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
64 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
67 * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch)
68 * in blk_mq_run_hw_queue(). Its pair is the barrier in
69 * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART,
70 * meantime new request added to hctx->dispatch is missed to check in
71 * blk_mq_run_hw_queue().
75 blk_mq_run_hw_queue(hctx, true);
78 static int sched_rq_cmp(void *priv, const struct list_head *a,
79 const struct list_head *b)
81 struct request *rqa = container_of(a, struct request, queuelist);
82 struct request *rqb = container_of(b, struct request, queuelist);
84 return rqa->mq_hctx > rqb->mq_hctx;
87 static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list)
89 struct blk_mq_hw_ctx *hctx =
90 list_first_entry(rq_list, struct request, queuelist)->mq_hctx;
93 unsigned int count = 0;
95 list_for_each_entry(rq, rq_list, queuelist) {
96 if (rq->mq_hctx != hctx) {
97 list_cut_before(&hctx_list, rq_list, &rq->queuelist);
102 list_splice_tail_init(rq_list, &hctx_list);
105 return blk_mq_dispatch_rq_list(hctx, &hctx_list, count);
108 #define BLK_MQ_BUDGET_DELAY 3 /* ms units */
111 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
112 * its queue by itself in its completion handler, so we don't need to
113 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
115 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
116 * be run again. This is necessary to avoid starving flushes.
118 static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
120 struct request_queue *q = hctx->queue;
121 struct elevator_queue *e = q->elevator;
122 bool multi_hctxs = false, run_queue = false;
123 bool dispatched = false, busy = false;
124 unsigned int max_dispatch;
128 if (hctx->dispatch_busy)
131 max_dispatch = hctx->queue->nr_requests;
137 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
140 if (!list_empty_careful(&hctx->dispatch)) {
145 budget_token = blk_mq_get_dispatch_budget(q);
146 if (budget_token < 0)
149 rq = e->type->ops.dispatch_request(hctx);
151 blk_mq_put_dispatch_budget(q, budget_token);
153 * We're releasing without dispatching. Holding the
154 * budget could have blocked any "hctx"s with the
155 * same queue and if we didn't dispatch then there's
156 * no guarantee anyone will kick the queue. Kick it
163 blk_mq_set_rq_budget_token(rq, budget_token);
166 * Now this rq owns the budget which has to be released
167 * if this rq won't be queued to driver via .queue_rq()
168 * in blk_mq_dispatch_rq_list().
170 list_add_tail(&rq->queuelist, &rq_list);
172 if (rq->mq_hctx != hctx)
176 * If we cannot get tag for the request, stop dequeueing
177 * requests from the IO scheduler. We are unlikely to be able
178 * to submit them anyway and it creates false impression for
179 * scheduling heuristics that the device can take more IO.
181 if (!blk_mq_get_driver_tag(rq))
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)
211 unsigned long end = jiffies + HZ;
215 ret = __blk_mq_do_dispatch_sched(hctx);
218 if (need_resched() || time_is_before_jiffies(end)) {
219 blk_mq_delay_run_hw_queue(hctx, 0);
227 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
228 struct blk_mq_ctx *ctx)
230 unsigned short idx = ctx->index_hw[hctx->type];
232 if (++idx == hctx->nr_ctx)
235 return hctx->ctxs[idx];
239 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
240 * its queue by itself in its completion handler, so we don't need to
241 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
243 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
244 * be run again. This is necessary to avoid starving flushes.
246 static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
248 struct request_queue *q = hctx->queue;
250 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
257 if (!list_empty_careful(&hctx->dispatch)) {
262 if (!sbitmap_any_bit_set(&hctx->ctx_map))
265 budget_token = blk_mq_get_dispatch_budget(q);
266 if (budget_token < 0)
269 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
271 blk_mq_put_dispatch_budget(q, budget_token);
273 * We're releasing without dispatching. Holding the
274 * budget could have blocked any "hctx"s with the
275 * same queue and if we didn't dispatch then there's
276 * no guarantee anyone will kick the queue. Kick it
279 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
283 blk_mq_set_rq_budget_token(rq, budget_token);
286 * Now this rq owns the budget which has to be released
287 * if this rq won't be queued to driver via .queue_rq()
288 * in blk_mq_dispatch_rq_list().
290 list_add(&rq->queuelist, &rq_list);
292 /* round robin for fair dispatch */
293 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
295 } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1));
297 WRITE_ONCE(hctx->dispatch_from, ctx);
301 static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
303 struct request_queue *q = hctx->queue;
304 const bool has_sched = q->elevator;
309 * If we have previous entries on our dispatch list, grab them first for
310 * more fair dispatch.
312 if (!list_empty_careful(&hctx->dispatch)) {
313 spin_lock(&hctx->lock);
314 if (!list_empty(&hctx->dispatch))
315 list_splice_init(&hctx->dispatch, &rq_list);
316 spin_unlock(&hctx->lock);
320 * Only ask the scheduler for requests, if we didn't have residual
321 * requests from the dispatch list. This is to avoid the case where
322 * we only ever dispatch a fraction of the requests available because
323 * of low device queue depth. Once we pull requests out of the IO
324 * scheduler, we can no longer merge or sort them. So it's best to
325 * leave them there for as long as we can. Mark the hw queue as
326 * needing a restart in that case.
328 * We want to dispatch from the scheduler if there was nothing
329 * on the dispatch list or we were able to dispatch from the
332 if (!list_empty(&rq_list)) {
333 blk_mq_sched_mark_restart_hctx(hctx);
334 if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) {
336 ret = blk_mq_do_dispatch_sched(hctx);
338 ret = blk_mq_do_dispatch_ctx(hctx);
340 } else if (has_sched) {
341 ret = blk_mq_do_dispatch_sched(hctx);
342 } else if (hctx->dispatch_busy) {
343 /* dequeue request one by one from sw queue if queue is busy */
344 ret = blk_mq_do_dispatch_ctx(hctx);
346 blk_mq_flush_busy_ctxs(hctx, &rq_list);
347 blk_mq_dispatch_rq_list(hctx, &rq_list, 0);
353 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
355 struct request_queue *q = hctx->queue;
357 /* RCU or SRCU read lock is needed before checking quiesced flag */
358 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
364 * A return of -EAGAIN is an indication that hctx->dispatch is not
365 * empty and we must run again in order to avoid starving flushes.
367 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) {
368 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN)
369 blk_mq_run_hw_queue(hctx, true);
373 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
374 unsigned int nr_segs)
376 struct elevator_queue *e = q->elevator;
377 struct blk_mq_ctx *ctx;
378 struct blk_mq_hw_ctx *hctx;
382 if (e && e->type->ops.bio_merge)
383 return e->type->ops.bio_merge(q, bio, nr_segs);
385 ctx = blk_mq_get_ctx(q);
386 hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
388 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE) ||
389 list_empty_careful(&ctx->rq_lists[type]))
392 /* default per sw-queue merge */
393 spin_lock(&ctx->lock);
395 * Reverse check our software queue for entries that we could
396 * potentially merge with. Currently includes a hand-wavy stop
397 * count of 8, to not spend too much time checking for merges.
399 if (blk_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) {
404 spin_unlock(&ctx->lock);
409 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq,
410 struct list_head *free)
412 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq, free);
414 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
416 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
420 * dispatch flush and passthrough rq directly
422 * passthrough request has to be added to hctx->dispatch directly.
423 * For some reason, device may be in one situation which can't
424 * handle FS request, so STS_RESOURCE is always returned and the
425 * FS request will be added to hctx->dispatch. However passthrough
426 * request may be required at that time for fixing the problem. If
427 * passthrough request is added to scheduler queue, there isn't any
428 * chance to dispatch it given we prioritize requests in hctx->dispatch.
430 if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq))
436 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
437 bool run_queue, bool async)
439 struct request_queue *q = rq->q;
440 struct elevator_queue *e = q->elevator;
441 struct blk_mq_ctx *ctx = rq->mq_ctx;
442 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
444 WARN_ON(e && (rq->tag != BLK_MQ_NO_TAG));
446 if (blk_mq_sched_bypass_insert(hctx, rq)) {
448 * Firstly normal IO request is inserted to scheduler queue or
449 * sw queue, meantime we add flush request to dispatch queue(
450 * hctx->dispatch) directly and there is at most one in-flight
451 * flush request for each hw queue, so it doesn't matter to add
452 * flush request to tail or front of the dispatch queue.
454 * Secondly in case of NCQ, flush request belongs to non-NCQ
455 * command, and queueing it will fail when there is any
456 * in-flight normal IO request(NCQ command). When adding flush
457 * rq to the front of hctx->dispatch, it is easier to introduce
458 * extra time to flush rq's latency because of S_SCHED_RESTART
459 * compared with adding to the tail of dispatch queue, then
460 * chance of flush merge is increased, and less flush requests
461 * will be issued to controller. It is observed that ~10% time
462 * is saved in blktests block/004 on disk attached to AHCI/NCQ
463 * drive when adding flush rq to the front of hctx->dispatch.
465 * Simply queue flush rq to the front of hctx->dispatch so that
466 * intensive flush workloads can benefit in case of NCQ HW.
468 at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head;
469 blk_mq_request_bypass_insert(rq, at_head, false);
476 list_add(&rq->queuelist, &list);
477 e->type->ops.insert_requests(hctx, &list, at_head);
479 spin_lock(&ctx->lock);
480 __blk_mq_insert_request(hctx, rq, at_head);
481 spin_unlock(&ctx->lock);
486 blk_mq_run_hw_queue(hctx, async);
489 void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
490 struct blk_mq_ctx *ctx,
491 struct list_head *list, bool run_queue_async)
493 struct elevator_queue *e;
494 struct request_queue *q = hctx->queue;
497 * blk_mq_sched_insert_requests() is called from flush plug
498 * context only, and hold one usage counter to prevent queue
499 * from being released.
501 percpu_ref_get(&q->q_usage_counter);
503 e = hctx->queue->elevator;
505 e->type->ops.insert_requests(hctx, list, false);
508 * try to issue requests directly if the hw queue isn't
509 * busy in case of 'none' scheduler, and this way may save
510 * us one extra enqueue & dequeue to sw queue.
512 if (!hctx->dispatch_busy && !e && !run_queue_async) {
513 blk_mq_try_issue_list_directly(hctx, list);
514 if (list_empty(list))
517 blk_mq_insert_requests(hctx, ctx, list);
520 blk_mq_run_hw_queue(hctx, run_queue_async);
522 percpu_ref_put(&q->q_usage_counter);
525 static int blk_mq_sched_alloc_tags(struct request_queue *q,
526 struct blk_mq_hw_ctx *hctx,
527 unsigned int hctx_idx)
529 struct blk_mq_tag_set *set = q->tag_set;
532 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
533 set->reserved_tags, set->flags);
534 if (!hctx->sched_tags)
537 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
539 blk_mq_free_rq_map(hctx->sched_tags, set->flags);
540 hctx->sched_tags = NULL;
546 /* called in queue's release handler, tagset has gone away */
547 static void blk_mq_sched_tags_teardown(struct request_queue *q)
549 struct blk_mq_hw_ctx *hctx;
552 queue_for_each_hw_ctx(q, hctx, i) {
553 if (hctx->sched_tags) {
554 blk_mq_free_rq_map(hctx->sched_tags, hctx->flags);
555 hctx->sched_tags = NULL;
560 static int blk_mq_init_sched_shared_sbitmap(struct request_queue *queue)
562 struct blk_mq_tag_set *set = queue->tag_set;
563 int alloc_policy = BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags);
564 struct blk_mq_hw_ctx *hctx;
568 * Set initial depth at max so that we don't need to reallocate for
569 * updating nr_requests.
571 ret = blk_mq_init_bitmaps(&queue->sched_bitmap_tags,
572 &queue->sched_breserved_tags,
573 MAX_SCHED_RQ, set->reserved_tags,
574 set->numa_node, alloc_policy);
578 queue_for_each_hw_ctx(queue, hctx, i) {
579 hctx->sched_tags->bitmap_tags =
580 &queue->sched_bitmap_tags;
581 hctx->sched_tags->breserved_tags =
582 &queue->sched_breserved_tags;
585 sbitmap_queue_resize(&queue->sched_bitmap_tags,
586 queue->nr_requests - set->reserved_tags);
591 static void blk_mq_exit_sched_shared_sbitmap(struct request_queue *queue)
593 sbitmap_queue_free(&queue->sched_bitmap_tags);
594 sbitmap_queue_free(&queue->sched_breserved_tags);
597 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
599 struct blk_mq_hw_ctx *hctx;
600 struct elevator_queue *eq;
606 q->nr_requests = q->tag_set->queue_depth;
611 * Default to double of smaller one between hw queue_depth and 128,
612 * since we don't split into sync/async like the old code did.
613 * Additionally, this is a per-hw queue depth.
615 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
618 queue_for_each_hw_ctx(q, hctx, i) {
619 ret = blk_mq_sched_alloc_tags(q, hctx, i);
624 if (blk_mq_is_sbitmap_shared(q->tag_set->flags)) {
625 ret = blk_mq_init_sched_shared_sbitmap(q);
630 ret = e->ops.init_sched(q, e);
632 goto err_free_sbitmap;
634 blk_mq_debugfs_register_sched(q);
636 queue_for_each_hw_ctx(q, hctx, i) {
637 if (e->ops.init_hctx) {
638 ret = e->ops.init_hctx(hctx, i);
641 blk_mq_sched_free_requests(q);
642 blk_mq_exit_sched(q, eq);
643 kobject_put(&eq->kobj);
647 blk_mq_debugfs_register_sched_hctx(q, hctx);
653 if (blk_mq_is_sbitmap_shared(q->tag_set->flags))
654 blk_mq_exit_sched_shared_sbitmap(q);
656 blk_mq_sched_free_requests(q);
657 blk_mq_sched_tags_teardown(q);
663 * called in either blk_queue_cleanup or elevator_switch, tagset
664 * is required for freeing requests
666 void blk_mq_sched_free_requests(struct request_queue *q)
668 struct blk_mq_hw_ctx *hctx;
671 queue_for_each_hw_ctx(q, hctx, i) {
672 if (hctx->sched_tags)
673 blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i);
677 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
679 struct blk_mq_hw_ctx *hctx;
681 unsigned int flags = 0;
683 queue_for_each_hw_ctx(q, hctx, i) {
684 blk_mq_debugfs_unregister_sched_hctx(hctx);
685 if (e->type->ops.exit_hctx && hctx->sched_data) {
686 e->type->ops.exit_hctx(hctx, i);
687 hctx->sched_data = NULL;
691 blk_mq_debugfs_unregister_sched(q);
692 if (e->type->ops.exit_sched)
693 e->type->ops.exit_sched(e);
694 blk_mq_sched_tags_teardown(q);
695 if (blk_mq_is_sbitmap_shared(flags))
696 blk_mq_exit_sched_shared_sbitmap(q);