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"
22 * Mark a hardware queue as needing a restart. For shared queues, maintain
23 * a count of how many hardware queues are marked for restart.
25 void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
27 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
30 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
32 EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
34 void __blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
36 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
39 * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch)
40 * in blk_mq_run_hw_queue(). Its pair is the barrier in
41 * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART,
42 * meantime new request added to hctx->dispatch is missed to check in
43 * blk_mq_run_hw_queue().
47 blk_mq_run_hw_queue(hctx, true);
50 static int sched_rq_cmp(void *priv, const struct list_head *a,
51 const struct list_head *b)
53 struct request *rqa = container_of(a, struct request, queuelist);
54 struct request *rqb = container_of(b, struct request, queuelist);
56 return rqa->mq_hctx > rqb->mq_hctx;
59 static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list)
61 struct blk_mq_hw_ctx *hctx =
62 list_first_entry(rq_list, struct request, queuelist)->mq_hctx;
65 unsigned int count = 0;
67 list_for_each_entry(rq, rq_list, queuelist) {
68 if (rq->mq_hctx != hctx) {
69 list_cut_before(&hctx_list, rq_list, &rq->queuelist);
74 list_splice_tail_init(rq_list, &hctx_list);
77 return blk_mq_dispatch_rq_list(hctx, &hctx_list, count);
80 #define BLK_MQ_BUDGET_DELAY 3 /* ms units */
83 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
84 * its queue by itself in its completion handler, so we don't need to
85 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
87 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
88 * be run again. This is necessary to avoid starving flushes.
90 static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
92 struct request_queue *q = hctx->queue;
93 struct elevator_queue *e = q->elevator;
94 bool multi_hctxs = false, run_queue = false;
95 bool dispatched = false, busy = false;
96 unsigned int max_dispatch;
100 if (hctx->dispatch_busy)
103 max_dispatch = hctx->queue->nr_requests;
109 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
112 if (!list_empty_careful(&hctx->dispatch)) {
117 budget_token = blk_mq_get_dispatch_budget(q);
118 if (budget_token < 0)
121 rq = e->type->ops.dispatch_request(hctx);
123 blk_mq_put_dispatch_budget(q, budget_token);
125 * We're releasing without dispatching. Holding the
126 * budget could have blocked any "hctx"s with the
127 * same queue and if we didn't dispatch then there's
128 * no guarantee anyone will kick the queue. Kick it
135 blk_mq_set_rq_budget_token(rq, budget_token);
138 * Now this rq owns the budget which has to be released
139 * if this rq won't be queued to driver via .queue_rq()
140 * in blk_mq_dispatch_rq_list().
142 list_add_tail(&rq->queuelist, &rq_list);
144 if (rq->mq_hctx != hctx)
148 * If we cannot get tag for the request, stop dequeueing
149 * requests from the IO scheduler. We are unlikely to be able
150 * to submit them anyway and it creates false impression for
151 * scheduling heuristics that the device can take more IO.
153 if (!blk_mq_get_driver_tag(rq))
155 } while (count < max_dispatch);
159 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
160 } else if (multi_hctxs) {
162 * Requests from different hctx may be dequeued from some
163 * schedulers, such as bfq and deadline.
165 * Sort the requests in the list according to their hctx,
166 * dispatch batching requests from same hctx at a time.
168 list_sort(NULL, &rq_list, sched_rq_cmp);
170 dispatched |= blk_mq_dispatch_hctx_list(&rq_list);
171 } while (!list_empty(&rq_list));
173 dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, count);
181 static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
186 ret = __blk_mq_do_dispatch_sched(hctx);
192 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
193 struct blk_mq_ctx *ctx)
195 unsigned short idx = ctx->index_hw[hctx->type];
197 if (++idx == hctx->nr_ctx)
200 return hctx->ctxs[idx];
204 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
205 * its queue by itself in its completion handler, so we don't need to
206 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
208 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
209 * be run again. This is necessary to avoid starving flushes.
211 static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
213 struct request_queue *q = hctx->queue;
215 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
222 if (!list_empty_careful(&hctx->dispatch)) {
227 if (!sbitmap_any_bit_set(&hctx->ctx_map))
230 budget_token = blk_mq_get_dispatch_budget(q);
231 if (budget_token < 0)
234 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
236 blk_mq_put_dispatch_budget(q, budget_token);
238 * We're releasing without dispatching. Holding the
239 * budget could have blocked any "hctx"s with the
240 * same queue and if we didn't dispatch then there's
241 * no guarantee anyone will kick the queue. Kick it
244 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
248 blk_mq_set_rq_budget_token(rq, budget_token);
251 * Now this rq owns the budget which has to be released
252 * if this rq won't be queued to driver via .queue_rq()
253 * in blk_mq_dispatch_rq_list().
255 list_add(&rq->queuelist, &rq_list);
257 /* round robin for fair dispatch */
258 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
260 } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1));
262 WRITE_ONCE(hctx->dispatch_from, ctx);
266 static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
268 struct request_queue *q = hctx->queue;
269 const bool has_sched = q->elevator;
274 * If we have previous entries on our dispatch list, grab them first for
275 * more fair dispatch.
277 if (!list_empty_careful(&hctx->dispatch)) {
278 spin_lock(&hctx->lock);
279 if (!list_empty(&hctx->dispatch))
280 list_splice_init(&hctx->dispatch, &rq_list);
281 spin_unlock(&hctx->lock);
285 * Only ask the scheduler for requests, if we didn't have residual
286 * requests from the dispatch list. This is to avoid the case where
287 * we only ever dispatch a fraction of the requests available because
288 * of low device queue depth. Once we pull requests out of the IO
289 * scheduler, we can no longer merge or sort them. So it's best to
290 * leave them there for as long as we can. Mark the hw queue as
291 * needing a restart in that case.
293 * We want to dispatch from the scheduler if there was nothing
294 * on the dispatch list or we were able to dispatch from the
297 if (!list_empty(&rq_list)) {
298 blk_mq_sched_mark_restart_hctx(hctx);
299 if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) {
301 ret = blk_mq_do_dispatch_sched(hctx);
303 ret = blk_mq_do_dispatch_ctx(hctx);
305 } else if (has_sched) {
306 ret = blk_mq_do_dispatch_sched(hctx);
307 } else if (hctx->dispatch_busy) {
308 /* dequeue request one by one from sw queue if queue is busy */
309 ret = blk_mq_do_dispatch_ctx(hctx);
311 blk_mq_flush_busy_ctxs(hctx, &rq_list);
312 blk_mq_dispatch_rq_list(hctx, &rq_list, 0);
318 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
320 struct request_queue *q = hctx->queue;
322 /* RCU or SRCU read lock is needed before checking quiesced flag */
323 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
329 * A return of -EAGAIN is an indication that hctx->dispatch is not
330 * empty and we must run again in order to avoid starving flushes.
332 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) {
333 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN)
334 blk_mq_run_hw_queue(hctx, true);
338 bool blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
339 unsigned int nr_segs)
341 struct elevator_queue *e = q->elevator;
342 struct blk_mq_ctx *ctx;
343 struct blk_mq_hw_ctx *hctx;
347 if (e && e->type->ops.bio_merge) {
348 ret = e->type->ops.bio_merge(q, bio, nr_segs);
352 ctx = blk_mq_get_ctx(q);
353 hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
355 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE) ||
356 list_empty_careful(&ctx->rq_lists[type]))
359 /* default per sw-queue merge */
360 spin_lock(&ctx->lock);
362 * Reverse check our software queue for entries that we could
363 * potentially merge with. Currently includes a hand-wavy stop
364 * count of 8, to not spend too much time checking for merges.
366 if (blk_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs))
369 spin_unlock(&ctx->lock);
374 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq,
375 struct list_head *free)
377 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq, free);
379 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
381 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
385 * dispatch flush and passthrough rq directly
387 * passthrough request has to be added to hctx->dispatch directly.
388 * For some reason, device may be in one situation which can't
389 * handle FS request, so STS_RESOURCE is always returned and the
390 * FS request will be added to hctx->dispatch. However passthrough
391 * request may be required at that time for fixing the problem. If
392 * passthrough request is added to scheduler queue, there isn't any
393 * chance to dispatch it given we prioritize requests in hctx->dispatch.
395 if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq))
401 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
402 bool run_queue, bool async)
404 struct request_queue *q = rq->q;
405 struct elevator_queue *e = q->elevator;
406 struct blk_mq_ctx *ctx = rq->mq_ctx;
407 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
409 WARN_ON(e && (rq->tag != BLK_MQ_NO_TAG));
411 if (blk_mq_sched_bypass_insert(hctx, rq)) {
413 * Firstly normal IO request is inserted to scheduler queue or
414 * sw queue, meantime we add flush request to dispatch queue(
415 * hctx->dispatch) directly and there is at most one in-flight
416 * flush request for each hw queue, so it doesn't matter to add
417 * flush request to tail or front of the dispatch queue.
419 * Secondly in case of NCQ, flush request belongs to non-NCQ
420 * command, and queueing it will fail when there is any
421 * in-flight normal IO request(NCQ command). When adding flush
422 * rq to the front of hctx->dispatch, it is easier to introduce
423 * extra time to flush rq's latency because of S_SCHED_RESTART
424 * compared with adding to the tail of dispatch queue, then
425 * chance of flush merge is increased, and less flush requests
426 * will be issued to controller. It is observed that ~10% time
427 * is saved in blktests block/004 on disk attached to AHCI/NCQ
428 * drive when adding flush rq to the front of hctx->dispatch.
430 * Simply queue flush rq to the front of hctx->dispatch so that
431 * intensive flush workloads can benefit in case of NCQ HW.
433 at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head;
434 blk_mq_request_bypass_insert(rq, at_head, false);
441 list_add(&rq->queuelist, &list);
442 e->type->ops.insert_requests(hctx, &list, at_head);
444 spin_lock(&ctx->lock);
445 __blk_mq_insert_request(hctx, rq, at_head);
446 spin_unlock(&ctx->lock);
451 blk_mq_run_hw_queue(hctx, async);
454 void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
455 struct blk_mq_ctx *ctx,
456 struct list_head *list, bool run_queue_async)
458 struct elevator_queue *e;
459 struct request_queue *q = hctx->queue;
462 * blk_mq_sched_insert_requests() is called from flush plug
463 * context only, and hold one usage counter to prevent queue
464 * from being released.
466 percpu_ref_get(&q->q_usage_counter);
468 e = hctx->queue->elevator;
470 e->type->ops.insert_requests(hctx, list, false);
473 * try to issue requests directly if the hw queue isn't
474 * busy in case of 'none' scheduler, and this way may save
475 * us one extra enqueue & dequeue to sw queue.
477 if (!hctx->dispatch_busy && !run_queue_async) {
478 blk_mq_run_dispatch_ops(hctx->queue,
479 blk_mq_try_issue_list_directly(hctx, list));
480 if (list_empty(list))
483 blk_mq_insert_requests(hctx, ctx, list);
486 blk_mq_run_hw_queue(hctx, run_queue_async);
488 percpu_ref_put(&q->q_usage_counter);
491 static int blk_mq_sched_alloc_map_and_rqs(struct request_queue *q,
492 struct blk_mq_hw_ctx *hctx,
493 unsigned int hctx_idx)
495 if (blk_mq_is_shared_tags(q->tag_set->flags)) {
496 hctx->sched_tags = q->sched_shared_tags;
500 hctx->sched_tags = blk_mq_alloc_map_and_rqs(q->tag_set, hctx_idx,
503 if (!hctx->sched_tags)
508 static void blk_mq_exit_sched_shared_tags(struct request_queue *queue)
510 blk_mq_free_rq_map(queue->sched_shared_tags);
511 queue->sched_shared_tags = NULL;
514 /* called in queue's release handler, tagset has gone away */
515 static void blk_mq_sched_tags_teardown(struct request_queue *q, unsigned int flags)
517 struct blk_mq_hw_ctx *hctx;
520 queue_for_each_hw_ctx(q, hctx, i) {
521 if (hctx->sched_tags) {
522 if (!blk_mq_is_shared_tags(flags))
523 blk_mq_free_rq_map(hctx->sched_tags);
524 hctx->sched_tags = NULL;
528 if (blk_mq_is_shared_tags(flags))
529 blk_mq_exit_sched_shared_tags(q);
532 static int blk_mq_init_sched_shared_tags(struct request_queue *queue)
534 struct blk_mq_tag_set *set = queue->tag_set;
537 * Set initial depth at max so that we don't need to reallocate for
538 * updating nr_requests.
540 queue->sched_shared_tags = blk_mq_alloc_map_and_rqs(set,
543 if (!queue->sched_shared_tags)
546 blk_mq_tag_update_sched_shared_tags(queue);
551 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
553 unsigned int i, flags = q->tag_set->flags;
554 struct blk_mq_hw_ctx *hctx;
555 struct elevator_queue *eq;
560 q->nr_requests = q->tag_set->queue_depth;
565 * Default to double of smaller one between hw queue_depth and 128,
566 * since we don't split into sync/async like the old code did.
567 * Additionally, this is a per-hw queue depth.
569 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
572 if (blk_mq_is_shared_tags(flags)) {
573 ret = blk_mq_init_sched_shared_tags(q);
578 queue_for_each_hw_ctx(q, hctx, i) {
579 ret = blk_mq_sched_alloc_map_and_rqs(q, hctx, i);
581 goto err_free_map_and_rqs;
584 ret = e->ops.init_sched(q, e);
586 goto err_free_map_and_rqs;
588 blk_mq_debugfs_register_sched(q);
590 queue_for_each_hw_ctx(q, hctx, i) {
591 if (e->ops.init_hctx) {
592 ret = e->ops.init_hctx(hctx, i);
595 blk_mq_sched_free_rqs(q);
596 blk_mq_exit_sched(q, eq);
597 kobject_put(&eq->kobj);
601 blk_mq_debugfs_register_sched_hctx(q, hctx);
606 err_free_map_and_rqs:
607 blk_mq_sched_free_rqs(q);
608 blk_mq_sched_tags_teardown(q, flags);
615 * called in either blk_queue_cleanup or elevator_switch, tagset
616 * is required for freeing requests
618 void blk_mq_sched_free_rqs(struct request_queue *q)
620 struct blk_mq_hw_ctx *hctx;
623 if (blk_mq_is_shared_tags(q->tag_set->flags)) {
624 blk_mq_free_rqs(q->tag_set, q->sched_shared_tags,
627 queue_for_each_hw_ctx(q, hctx, i) {
628 if (hctx->sched_tags)
629 blk_mq_free_rqs(q->tag_set,
630 hctx->sched_tags, i);
635 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
637 struct blk_mq_hw_ctx *hctx;
639 unsigned int flags = 0;
641 queue_for_each_hw_ctx(q, hctx, i) {
642 blk_mq_debugfs_unregister_sched_hctx(hctx);
643 if (e->type->ops.exit_hctx && hctx->sched_data) {
644 e->type->ops.exit_hctx(hctx, i);
645 hctx->sched_data = NULL;
649 blk_mq_debugfs_unregister_sched(q);
650 if (e->type->ops.exit_sched)
651 e->type->ops.exit_sched(e);
652 blk_mq_sched_tags_teardown(q, flags);