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>
11 #include <trace/events/block.h>
15 #include "blk-mq-debugfs.h"
16 #include "blk-mq-sched.h"
17 #include "blk-mq-tag.h"
20 void blk_mq_sched_free_hctx_data(struct request_queue *q,
21 void (*exit)(struct blk_mq_hw_ctx *))
23 struct blk_mq_hw_ctx *hctx;
26 queue_for_each_hw_ctx(q, hctx, i) {
27 if (exit && hctx->sched_data)
29 kfree(hctx->sched_data);
30 hctx->sched_data = NULL;
33 EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data);
35 void blk_mq_sched_assign_ioc(struct request *rq)
37 struct request_queue *q = rq->q;
38 struct io_context *ioc;
42 * May not have an IO context if it's a passthrough request
44 ioc = current->io_context;
48 spin_lock_irq(&q->queue_lock);
49 icq = ioc_lookup_icq(ioc, q);
50 spin_unlock_irq(&q->queue_lock);
53 icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
57 get_io_context(icq->ioc);
62 * Mark a hardware queue as needing a restart. For shared queues, maintain
63 * a count of how many hardware queues are marked for restart.
65 void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
67 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
70 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
72 EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
74 void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
76 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
78 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
80 blk_mq_run_hw_queue(hctx, true);
83 #define BLK_MQ_BUDGET_DELAY 3 /* ms units */
86 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
87 * its queue by itself in its completion handler, so we don't need to
88 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
90 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
91 * be run again. This is necessary to avoid starving flushes.
93 static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
95 struct request_queue *q = hctx->queue;
96 struct elevator_queue *e = q->elevator;
103 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
106 if (!list_empty_careful(&hctx->dispatch)) {
111 if (!blk_mq_get_dispatch_budget(hctx))
114 rq = e->type->ops.dispatch_request(hctx);
116 blk_mq_put_dispatch_budget(hctx);
118 * We're releasing without dispatching. Holding the
119 * budget could have blocked any "hctx"s with the
120 * same queue and if we didn't dispatch then there's
121 * no guarantee anyone will kick the queue. Kick it
124 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
129 * Now this rq owns the budget which has to be released
130 * if this rq won't be queued to driver via .queue_rq()
131 * in blk_mq_dispatch_rq_list().
133 list_add(&rq->queuelist, &rq_list);
134 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
139 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
140 struct blk_mq_ctx *ctx)
142 unsigned short idx = ctx->index_hw[hctx->type];
144 if (++idx == hctx->nr_ctx)
147 return hctx->ctxs[idx];
151 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
152 * its queue by itself in its completion handler, so we don't need to
153 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
155 * Returns -EAGAIN if hctx->dispatch was found non-empty and run_work has to
156 * to be run again. This is necessary to avoid starving flushes.
158 static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
160 struct request_queue *q = hctx->queue;
162 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
168 if (!list_empty_careful(&hctx->dispatch)) {
173 if (!sbitmap_any_bit_set(&hctx->ctx_map))
176 if (!blk_mq_get_dispatch_budget(hctx))
179 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
181 blk_mq_put_dispatch_budget(hctx);
183 * We're releasing without dispatching. Holding the
184 * budget could have blocked any "hctx"s with the
185 * same queue and if we didn't dispatch then there's
186 * no guarantee anyone will kick the queue. Kick it
189 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
194 * Now this rq owns the budget which has to be released
195 * if this rq won't be queued to driver via .queue_rq()
196 * in blk_mq_dispatch_rq_list().
198 list_add(&rq->queuelist, &rq_list);
200 /* round robin for fair dispatch */
201 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
203 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
205 WRITE_ONCE(hctx->dispatch_from, ctx);
209 static int __blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
211 struct request_queue *q = hctx->queue;
212 struct elevator_queue *e = q->elevator;
213 const bool has_sched_dispatch = e && e->type->ops.dispatch_request;
218 * If we have previous entries on our dispatch list, grab them first for
219 * more fair dispatch.
221 if (!list_empty_careful(&hctx->dispatch)) {
222 spin_lock(&hctx->lock);
223 if (!list_empty(&hctx->dispatch))
224 list_splice_init(&hctx->dispatch, &rq_list);
225 spin_unlock(&hctx->lock);
229 * Only ask the scheduler for requests, if we didn't have residual
230 * requests from the dispatch list. This is to avoid the case where
231 * we only ever dispatch a fraction of the requests available because
232 * of low device queue depth. Once we pull requests out of the IO
233 * scheduler, we can no longer merge or sort them. So it's best to
234 * leave them there for as long as we can. Mark the hw queue as
235 * needing a restart in that case.
237 * We want to dispatch from the scheduler if there was nothing
238 * on the dispatch list or we were able to dispatch from the
241 if (!list_empty(&rq_list)) {
242 blk_mq_sched_mark_restart_hctx(hctx);
243 if (blk_mq_dispatch_rq_list(q, &rq_list, false)) {
244 if (has_sched_dispatch)
245 ret = blk_mq_do_dispatch_sched(hctx);
247 ret = blk_mq_do_dispatch_ctx(hctx);
249 } else if (has_sched_dispatch) {
250 ret = blk_mq_do_dispatch_sched(hctx);
251 } else if (hctx->dispatch_busy) {
252 /* dequeue request one by one from sw queue if queue is busy */
253 ret = blk_mq_do_dispatch_ctx(hctx);
255 blk_mq_flush_busy_ctxs(hctx, &rq_list);
256 blk_mq_dispatch_rq_list(q, &rq_list, false);
262 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
264 struct request_queue *q = hctx->queue;
266 /* RCU or SRCU read lock is needed before checking quiesced flag */
267 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
273 * A return of -EAGAIN is an indication that hctx->dispatch is not
274 * empty and we must run again in order to avoid starving flushes.
276 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN) {
277 if (__blk_mq_sched_dispatch_requests(hctx) == -EAGAIN)
278 blk_mq_run_hw_queue(hctx, true);
282 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
283 unsigned int nr_segs, struct request **merged_request)
287 switch (elv_merge(q, &rq, bio)) {
288 case ELEVATOR_BACK_MERGE:
289 if (!blk_mq_sched_allow_merge(q, rq, bio))
291 if (!bio_attempt_back_merge(rq, bio, nr_segs))
293 *merged_request = attempt_back_merge(q, rq);
294 if (!*merged_request)
295 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
297 case ELEVATOR_FRONT_MERGE:
298 if (!blk_mq_sched_allow_merge(q, rq, bio))
300 if (!bio_attempt_front_merge(rq, bio, nr_segs))
302 *merged_request = attempt_front_merge(q, rq);
303 if (!*merged_request)
304 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
306 case ELEVATOR_DISCARD_MERGE:
307 return bio_attempt_discard_merge(q, rq, bio);
312 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
315 * Iterate list of requests and see if we can merge this bio with any
318 bool blk_mq_bio_list_merge(struct request_queue *q, struct list_head *list,
319 struct bio *bio, unsigned int nr_segs)
324 list_for_each_entry_reverse(rq, list, queuelist) {
330 if (!blk_rq_merge_ok(rq, bio))
333 switch (blk_try_merge(rq, bio)) {
334 case ELEVATOR_BACK_MERGE:
335 if (blk_mq_sched_allow_merge(q, rq, bio))
336 merged = bio_attempt_back_merge(rq, bio,
339 case ELEVATOR_FRONT_MERGE:
340 if (blk_mq_sched_allow_merge(q, rq, bio))
341 merged = bio_attempt_front_merge(rq, bio,
344 case ELEVATOR_DISCARD_MERGE:
345 merged = bio_attempt_discard_merge(q, rq, bio);
356 EXPORT_SYMBOL_GPL(blk_mq_bio_list_merge);
359 * Reverse check our software queue for entries that we could potentially
360 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
361 * too much time checking for merges.
363 static bool blk_mq_attempt_merge(struct request_queue *q,
364 struct blk_mq_hw_ctx *hctx,
365 struct blk_mq_ctx *ctx, struct bio *bio,
366 unsigned int nr_segs)
368 enum hctx_type type = hctx->type;
370 lockdep_assert_held(&ctx->lock);
372 if (blk_mq_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) {
380 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
381 unsigned int nr_segs)
383 struct elevator_queue *e = q->elevator;
384 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
385 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
389 if (e && e->type->ops.bio_merge)
390 return e->type->ops.bio_merge(hctx, bio, nr_segs);
393 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
394 !list_empty_careful(&ctx->rq_lists[type])) {
395 /* default per sw-queue merge */
396 spin_lock(&ctx->lock);
397 ret = blk_mq_attempt_merge(q, hctx, ctx, bio, nr_segs);
398 spin_unlock(&ctx->lock);
404 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
406 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
408 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
410 void blk_mq_sched_request_inserted(struct request *rq)
412 trace_block_rq_insert(rq->q, rq);
414 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
416 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
421 * dispatch flush and passthrough rq directly
423 * passthrough request has to be added to hctx->dispatch directly.
424 * For some reason, device may be in one situation which can't
425 * handle FS request, so STS_RESOURCE is always returned and the
426 * FS request will be added to hctx->dispatch. However passthrough
427 * request may be required at that time for fixing the problem. If
428 * passthrough request is added to scheduler queue, there isn't any
429 * chance to dispatch it given we prioritize requests in hctx->dispatch.
431 if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq))
435 rq->rq_flags |= RQF_SORTED;
440 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
441 bool run_queue, bool async)
443 struct request_queue *q = rq->q;
444 struct elevator_queue *e = q->elevator;
445 struct blk_mq_ctx *ctx = rq->mq_ctx;
446 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
448 /* flush rq in flush machinery need to be dispatched directly */
449 if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) {
450 blk_insert_flush(rq);
454 WARN_ON(e && (rq->tag != -1));
456 if (blk_mq_sched_bypass_insert(hctx, !!e, rq)) {
458 * Firstly normal IO request is inserted to scheduler queue or
459 * sw queue, meantime we add flush request to dispatch queue(
460 * hctx->dispatch) directly and there is at most one in-flight
461 * flush request for each hw queue, so it doesn't matter to add
462 * flush request to tail or front of the dispatch queue.
464 * Secondly in case of NCQ, flush request belongs to non-NCQ
465 * command, and queueing it will fail when there is any
466 * in-flight normal IO request(NCQ command). When adding flush
467 * rq to the front of hctx->dispatch, it is easier to introduce
468 * extra time to flush rq's latency because of S_SCHED_RESTART
469 * compared with adding to the tail of dispatch queue, then
470 * chance of flush merge is increased, and less flush requests
471 * will be issued to controller. It is observed that ~10% time
472 * is saved in blktests block/004 on disk attached to AHCI/NCQ
473 * drive when adding flush rq to the front of hctx->dispatch.
475 * Simply queue flush rq to the front of hctx->dispatch so that
476 * intensive flush workloads can benefit in case of NCQ HW.
478 at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head;
479 blk_mq_request_bypass_insert(rq, at_head, false);
483 if (e && e->type->ops.insert_requests) {
486 list_add(&rq->queuelist, &list);
487 e->type->ops.insert_requests(hctx, &list, at_head);
489 spin_lock(&ctx->lock);
490 __blk_mq_insert_request(hctx, rq, at_head);
491 spin_unlock(&ctx->lock);
496 blk_mq_run_hw_queue(hctx, async);
499 void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
500 struct blk_mq_ctx *ctx,
501 struct list_head *list, bool run_queue_async)
503 struct elevator_queue *e;
504 struct request_queue *q = hctx->queue;
507 * blk_mq_sched_insert_requests() is called from flush plug
508 * context only, and hold one usage counter to prevent queue
509 * from being released.
511 percpu_ref_get(&q->q_usage_counter);
513 e = hctx->queue->elevator;
514 if (e && e->type->ops.insert_requests)
515 e->type->ops.insert_requests(hctx, list, false);
518 * try to issue requests directly if the hw queue isn't
519 * busy in case of 'none' scheduler, and this way may save
520 * us one extra enqueue & dequeue to sw queue.
522 if (!hctx->dispatch_busy && !e && !run_queue_async) {
523 blk_mq_try_issue_list_directly(hctx, list);
524 if (list_empty(list))
527 blk_mq_insert_requests(hctx, ctx, list);
530 blk_mq_run_hw_queue(hctx, run_queue_async);
532 percpu_ref_put(&q->q_usage_counter);
535 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
536 struct blk_mq_hw_ctx *hctx,
537 unsigned int hctx_idx)
539 if (hctx->sched_tags) {
540 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
541 blk_mq_free_rq_map(hctx->sched_tags);
542 hctx->sched_tags = NULL;
546 static int blk_mq_sched_alloc_tags(struct request_queue *q,
547 struct blk_mq_hw_ctx *hctx,
548 unsigned int hctx_idx)
550 struct blk_mq_tag_set *set = q->tag_set;
553 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
555 if (!hctx->sched_tags)
558 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
560 blk_mq_sched_free_tags(set, hctx, hctx_idx);
565 /* called in queue's release handler, tagset has gone away */
566 static void blk_mq_sched_tags_teardown(struct request_queue *q)
568 struct blk_mq_hw_ctx *hctx;
571 queue_for_each_hw_ctx(q, hctx, i) {
572 if (hctx->sched_tags) {
573 blk_mq_free_rq_map(hctx->sched_tags);
574 hctx->sched_tags = NULL;
579 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
581 struct blk_mq_hw_ctx *hctx;
582 struct elevator_queue *eq;
588 q->nr_requests = q->tag_set->queue_depth;
593 * Default to double of smaller one between hw queue_depth and 128,
594 * since we don't split into sync/async like the old code did.
595 * Additionally, this is a per-hw queue depth.
597 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
600 queue_for_each_hw_ctx(q, hctx, i) {
601 ret = blk_mq_sched_alloc_tags(q, hctx, i);
606 ret = e->ops.init_sched(q, e);
610 blk_mq_debugfs_register_sched(q);
612 queue_for_each_hw_ctx(q, hctx, i) {
613 if (e->ops.init_hctx) {
614 ret = e->ops.init_hctx(hctx, i);
617 blk_mq_sched_free_requests(q);
618 blk_mq_exit_sched(q, eq);
619 kobject_put(&eq->kobj);
623 blk_mq_debugfs_register_sched_hctx(q, hctx);
629 blk_mq_sched_free_requests(q);
630 blk_mq_sched_tags_teardown(q);
636 * called in either blk_queue_cleanup or elevator_switch, tagset
637 * is required for freeing requests
639 void blk_mq_sched_free_requests(struct request_queue *q)
641 struct blk_mq_hw_ctx *hctx;
644 queue_for_each_hw_ctx(q, hctx, i) {
645 if (hctx->sched_tags)
646 blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i);
650 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
652 struct blk_mq_hw_ctx *hctx;
655 queue_for_each_hw_ctx(q, hctx, i) {
656 blk_mq_debugfs_unregister_sched_hctx(hctx);
657 if (e->type->ops.exit_hctx && hctx->sched_data) {
658 e->type->ops.exit_hctx(hctx, i);
659 hctx->sched_data = NULL;
662 blk_mq_debugfs_unregister_sched(q);
663 if (e->type->ops.exit_sched)
664 e->type->ops.exit_sched(e);
665 blk_mq_sched_tags_teardown(q);