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/list_sort.h>
11 #include <trace/events/block.h>
15 #include "blk-mq-debugfs.h"
16 #include "blk-mq-sched.h"
20 * Mark a hardware queue as needing a restart.
22 void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
24 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
27 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
29 EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
31 void __blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
33 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
36 * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch)
37 * in blk_mq_run_hw_queue(). Its pair is the barrier in
38 * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART,
39 * meantime new request added to hctx->dispatch is missed to check in
40 * blk_mq_run_hw_queue().
44 blk_mq_run_hw_queue(hctx, true);
47 static int sched_rq_cmp(void *priv, const struct list_head *a,
48 const struct list_head *b)
50 struct request *rqa = container_of(a, struct request, queuelist);
51 struct request *rqb = container_of(b, struct request, queuelist);
53 return rqa->mq_hctx > rqb->mq_hctx;
56 static bool blk_mq_dispatch_hctx_list(struct list_head *rq_list)
58 struct blk_mq_hw_ctx *hctx =
59 list_first_entry(rq_list, struct request, queuelist)->mq_hctx;
62 unsigned int count = 0;
64 list_for_each_entry(rq, rq_list, queuelist) {
65 if (rq->mq_hctx != hctx) {
66 list_cut_before(&hctx_list, rq_list, &rq->queuelist);
71 list_splice_tail_init(rq_list, &hctx_list);
74 return blk_mq_dispatch_rq_list(hctx, &hctx_list, count);
77 #define BLK_MQ_BUDGET_DELAY 3 /* ms units */
80 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
81 * its queue by itself in its completion handler, so we don't need to
82 * restart queue if .get_budget() fails to get the budget.
84 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
85 * be run again. This is necessary to avoid starving flushes.
87 static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
89 struct request_queue *q = hctx->queue;
90 struct elevator_queue *e = q->elevator;
91 bool multi_hctxs = false, run_queue = false;
92 bool dispatched = false, busy = false;
93 unsigned int max_dispatch;
97 if (hctx->dispatch_busy)
100 max_dispatch = hctx->queue->nr_requests;
106 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
109 if (!list_empty_careful(&hctx->dispatch)) {
114 budget_token = blk_mq_get_dispatch_budget(q);
115 if (budget_token < 0)
118 rq = e->type->ops.dispatch_request(hctx);
120 blk_mq_put_dispatch_budget(q, budget_token);
122 * We're releasing without dispatching. Holding the
123 * budget could have blocked any "hctx"s with the
124 * same queue and if we didn't dispatch then there's
125 * no guarantee anyone will kick the queue. Kick it
132 blk_mq_set_rq_budget_token(rq, budget_token);
135 * Now this rq owns the budget which has to be released
136 * if this rq won't be queued to driver via .queue_rq()
137 * in blk_mq_dispatch_rq_list().
139 list_add_tail(&rq->queuelist, &rq_list);
141 if (rq->mq_hctx != hctx)
145 * If we cannot get tag for the request, stop dequeueing
146 * requests from the IO scheduler. We are unlikely to be able
147 * to submit them anyway and it creates false impression for
148 * scheduling heuristics that the device can take more IO.
150 if (!blk_mq_get_driver_tag(rq))
152 } while (count < max_dispatch);
156 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
157 } else if (multi_hctxs) {
159 * Requests from different hctx may be dequeued from some
160 * schedulers, such as bfq and deadline.
162 * Sort the requests in the list according to their hctx,
163 * dispatch batching requests from same hctx at a time.
165 list_sort(NULL, &rq_list, sched_rq_cmp);
167 dispatched |= blk_mq_dispatch_hctx_list(&rq_list);
168 } while (!list_empty(&rq_list));
170 dispatched = blk_mq_dispatch_rq_list(hctx, &rq_list, count);
178 static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
180 unsigned long end = jiffies + HZ;
184 ret = __blk_mq_do_dispatch_sched(hctx);
187 if (need_resched() || time_is_before_jiffies(end)) {
188 blk_mq_delay_run_hw_queue(hctx, 0);
196 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
197 struct blk_mq_ctx *ctx)
199 unsigned short idx = ctx->index_hw[hctx->type];
201 if (++idx == hctx->nr_ctx)
204 return hctx->ctxs[idx];
208 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
209 * its queue by itself in its completion handler, so we don't need to
210 * restart queue if .get_budget() fails to get the budget.
212 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
213 * be run again. This is necessary to avoid starving flushes.
215 static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
217 struct request_queue *q = hctx->queue;
219 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
226 if (!list_empty_careful(&hctx->dispatch)) {
231 if (!sbitmap_any_bit_set(&hctx->ctx_map))
234 budget_token = blk_mq_get_dispatch_budget(q);
235 if (budget_token < 0)
238 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
240 blk_mq_put_dispatch_budget(q, budget_token);
242 * We're releasing without dispatching. Holding the
243 * budget could have blocked any "hctx"s with the
244 * same queue and if we didn't dispatch then there's
245 * no guarantee anyone will kick the queue. Kick it
248 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
252 blk_mq_set_rq_budget_token(rq, budget_token);
255 * Now this rq owns the budget which has to be released
256 * if this rq won't be queued to driver via .queue_rq()
257 * in blk_mq_dispatch_rq_list().
259 list_add(&rq->queuelist, &rq_list);
261 /* round robin for fair dispatch */
262 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
264 } while (blk_mq_dispatch_rq_list(rq->mq_hctx, &rq_list, 1));
266 WRITE_ONCE(hctx->dispatch_from, ctx);
270 static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
272 struct request_queue *q = hctx->queue;
273 const bool has_sched = q->elevator;
278 * If we have previous entries on our dispatch list, grab them first for
279 * more fair dispatch.
281 if (!list_empty_careful(&hctx->dispatch)) {
282 spin_lock(&hctx->lock);
283 if (!list_empty(&hctx->dispatch))
284 list_splice_init(&hctx->dispatch, &rq_list);
285 spin_unlock(&hctx->lock);
289 * Only ask the scheduler for requests, if we didn't have residual
290 * requests from the dispatch list. This is to avoid the case where
291 * we only ever dispatch a fraction of the requests available because
292 * of low device queue depth. Once we pull requests out of the IO
293 * scheduler, we can no longer merge or sort them. So it's best to
294 * leave them there for as long as we can. Mark the hw queue as
295 * needing a restart in that case.
297 * We want to dispatch from the scheduler if there was nothing
298 * on the dispatch list or we were able to dispatch from the
301 if (!list_empty(&rq_list)) {
302 blk_mq_sched_mark_restart_hctx(hctx);
303 if (blk_mq_dispatch_rq_list(hctx, &rq_list, 0)) {
305 ret = blk_mq_do_dispatch_sched(hctx);
307 ret = blk_mq_do_dispatch_ctx(hctx);
309 } else if (has_sched) {
310 ret = blk_mq_do_dispatch_sched(hctx);
311 } else if (hctx->dispatch_busy) {
312 /* dequeue request one by one from sw queue if queue is busy */
313 ret = blk_mq_do_dispatch_ctx(hctx);
315 blk_mq_flush_busy_ctxs(hctx, &rq_list);
316 blk_mq_dispatch_rq_list(hctx, &rq_list, 0);
322 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
324 struct request_queue *q = hctx->queue;
326 /* RCU or SRCU read lock is needed before checking quiesced flag */
327 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
333 * A return of -EAGAIN is an indication that hctx->dispatch is not
334 * empty and we must run again in order to avoid starving flushes.
336 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) {
337 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN)
338 blk_mq_run_hw_queue(hctx, true);
342 bool blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
343 unsigned int nr_segs)
345 struct elevator_queue *e = q->elevator;
346 struct blk_mq_ctx *ctx;
347 struct blk_mq_hw_ctx *hctx;
351 if (e && e->type->ops.bio_merge) {
352 ret = e->type->ops.bio_merge(q, bio, nr_segs);
356 ctx = blk_mq_get_ctx(q);
357 hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
359 if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE) ||
360 list_empty_careful(&ctx->rq_lists[type]))
363 /* default per sw-queue merge */
364 spin_lock(&ctx->lock);
366 * Reverse check our software queue for entries that we could
367 * potentially merge with. Currently includes a hand-wavy stop
368 * count of 8, to not spend too much time checking for merges.
370 if (blk_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs))
373 spin_unlock(&ctx->lock);
378 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq,
379 struct list_head *free)
381 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq, free);
383 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
385 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
389 * dispatch flush and passthrough rq directly
391 * passthrough request has to be added to hctx->dispatch directly.
392 * For some reason, device may be in one situation which can't
393 * handle FS request, so STS_RESOURCE is always returned and the
394 * FS request will be added to hctx->dispatch. However passthrough
395 * request may be required at that time for fixing the problem. If
396 * passthrough request is added to scheduler queue, there isn't any
397 * chance to dispatch it given we prioritize requests in hctx->dispatch.
399 if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq))
405 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
406 bool run_queue, bool async)
408 struct request_queue *q = rq->q;
409 struct elevator_queue *e = q->elevator;
410 struct blk_mq_ctx *ctx = rq->mq_ctx;
411 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
413 WARN_ON(e && (rq->tag != BLK_MQ_NO_TAG));
415 if (blk_mq_sched_bypass_insert(hctx, rq)) {
417 * Firstly normal IO request is inserted to scheduler queue or
418 * sw queue, meantime we add flush request to dispatch queue(
419 * hctx->dispatch) directly and there is at most one in-flight
420 * flush request for each hw queue, so it doesn't matter to add
421 * flush request to tail or front of the dispatch queue.
423 * Secondly in case of NCQ, flush request belongs to non-NCQ
424 * command, and queueing it will fail when there is any
425 * in-flight normal IO request(NCQ command). When adding flush
426 * rq to the front of hctx->dispatch, it is easier to introduce
427 * extra time to flush rq's latency because of S_SCHED_RESTART
428 * compared with adding to the tail of dispatch queue, then
429 * chance of flush merge is increased, and less flush requests
430 * will be issued to controller. It is observed that ~10% time
431 * is saved in blktests block/004 on disk attached to AHCI/NCQ
432 * drive when adding flush rq to the front of hctx->dispatch.
434 * Simply queue flush rq to the front of hctx->dispatch so that
435 * intensive flush workloads can benefit in case of NCQ HW.
437 at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head;
438 blk_mq_request_bypass_insert(rq, at_head, false);
445 list_add(&rq->queuelist, &list);
446 e->type->ops.insert_requests(hctx, &list, at_head);
448 spin_lock(&ctx->lock);
449 __blk_mq_insert_request(hctx, rq, at_head);
450 spin_unlock(&ctx->lock);
455 blk_mq_run_hw_queue(hctx, async);
458 static int blk_mq_sched_alloc_map_and_rqs(struct request_queue *q,
459 struct blk_mq_hw_ctx *hctx,
460 unsigned int hctx_idx)
462 if (blk_mq_is_shared_tags(q->tag_set->flags)) {
463 hctx->sched_tags = q->sched_shared_tags;
467 hctx->sched_tags = blk_mq_alloc_map_and_rqs(q->tag_set, hctx_idx,
470 if (!hctx->sched_tags)
475 static void blk_mq_exit_sched_shared_tags(struct request_queue *queue)
477 blk_mq_free_rq_map(queue->sched_shared_tags);
478 queue->sched_shared_tags = NULL;
481 /* called in queue's release handler, tagset has gone away */
482 static void blk_mq_sched_tags_teardown(struct request_queue *q, unsigned int flags)
484 struct blk_mq_hw_ctx *hctx;
487 queue_for_each_hw_ctx(q, hctx, i) {
488 if (hctx->sched_tags) {
489 if (!blk_mq_is_shared_tags(flags))
490 blk_mq_free_rq_map(hctx->sched_tags);
491 hctx->sched_tags = NULL;
495 if (blk_mq_is_shared_tags(flags))
496 blk_mq_exit_sched_shared_tags(q);
499 static int blk_mq_init_sched_shared_tags(struct request_queue *queue)
501 struct blk_mq_tag_set *set = queue->tag_set;
504 * Set initial depth at max so that we don't need to reallocate for
505 * updating nr_requests.
507 queue->sched_shared_tags = blk_mq_alloc_map_and_rqs(set,
510 if (!queue->sched_shared_tags)
513 blk_mq_tag_update_sched_shared_tags(queue);
518 /* caller must have a reference to @e, will grab another one if successful */
519 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
521 unsigned int flags = q->tag_set->flags;
522 struct blk_mq_hw_ctx *hctx;
523 struct elevator_queue *eq;
528 * Default to double of smaller one between hw queue_depth and 128,
529 * since we don't split into sync/async like the old code did.
530 * Additionally, this is a per-hw queue depth.
532 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
535 if (blk_mq_is_shared_tags(flags)) {
536 ret = blk_mq_init_sched_shared_tags(q);
541 queue_for_each_hw_ctx(q, hctx, i) {
542 ret = blk_mq_sched_alloc_map_and_rqs(q, hctx, i);
544 goto err_free_map_and_rqs;
547 ret = e->ops.init_sched(q, e);
549 goto err_free_map_and_rqs;
551 mutex_lock(&q->debugfs_mutex);
552 blk_mq_debugfs_register_sched(q);
553 mutex_unlock(&q->debugfs_mutex);
555 queue_for_each_hw_ctx(q, hctx, i) {
556 if (e->ops.init_hctx) {
557 ret = e->ops.init_hctx(hctx, i);
560 blk_mq_sched_free_rqs(q);
561 blk_mq_exit_sched(q, eq);
562 kobject_put(&eq->kobj);
566 mutex_lock(&q->debugfs_mutex);
567 blk_mq_debugfs_register_sched_hctx(q, hctx);
568 mutex_unlock(&q->debugfs_mutex);
573 err_free_map_and_rqs:
574 blk_mq_sched_free_rqs(q);
575 blk_mq_sched_tags_teardown(q, flags);
582 * called in either blk_queue_cleanup or elevator_switch, tagset
583 * is required for freeing requests
585 void blk_mq_sched_free_rqs(struct request_queue *q)
587 struct blk_mq_hw_ctx *hctx;
590 if (blk_mq_is_shared_tags(q->tag_set->flags)) {
591 blk_mq_free_rqs(q->tag_set, q->sched_shared_tags,
594 queue_for_each_hw_ctx(q, hctx, i) {
595 if (hctx->sched_tags)
596 blk_mq_free_rqs(q->tag_set,
597 hctx->sched_tags, i);
602 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
604 struct blk_mq_hw_ctx *hctx;
606 unsigned int flags = 0;
608 queue_for_each_hw_ctx(q, hctx, i) {
609 mutex_lock(&q->debugfs_mutex);
610 blk_mq_debugfs_unregister_sched_hctx(hctx);
611 mutex_unlock(&q->debugfs_mutex);
613 if (e->type->ops.exit_hctx && hctx->sched_data) {
614 e->type->ops.exit_hctx(hctx, i);
615 hctx->sched_data = NULL;
620 mutex_lock(&q->debugfs_mutex);
621 blk_mq_debugfs_unregister_sched(q);
622 mutex_unlock(&q->debugfs_mutex);
624 if (e->type->ops.exit_sched)
625 e->type->ops.exit_sched(e);
626 blk_mq_sched_tags_teardown(q, flags);