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);
81 * Order clearing SCHED_RESTART and list_empty_careful(&hctx->dispatch)
82 * in blk_mq_run_hw_queue(). Its pair is the barrier in
83 * blk_mq_dispatch_rq_list(). So dispatch code won't see SCHED_RESTART,
84 * meantime new request added to hctx->dispatch is missed to check in
85 * blk_mq_run_hw_queue().
89 blk_mq_run_hw_queue(hctx, true);
93 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
94 * its queue by itself in its completion handler, so we don't need to
95 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
97 static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
99 struct request_queue *q = hctx->queue;
100 struct elevator_queue *e = q->elevator;
106 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
109 if (!blk_mq_get_dispatch_budget(hctx))
112 rq = e->type->ops.dispatch_request(hctx);
114 blk_mq_put_dispatch_budget(hctx);
119 * Now this rq owns the budget which has to be released
120 * if this rq won't be queued to driver via .queue_rq()
121 * in blk_mq_dispatch_rq_list().
123 list_add(&rq->queuelist, &rq_list);
124 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
127 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
128 struct blk_mq_ctx *ctx)
130 unsigned short idx = ctx->index_hw[hctx->type];
132 if (++idx == hctx->nr_ctx)
135 return hctx->ctxs[idx];
139 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
140 * its queue by itself in its completion handler, so we don't need to
141 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
143 static void blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
145 struct request_queue *q = hctx->queue;
147 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
152 if (!sbitmap_any_bit_set(&hctx->ctx_map))
155 if (!blk_mq_get_dispatch_budget(hctx))
158 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
160 blk_mq_put_dispatch_budget(hctx);
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(&rq->queuelist, &rq_list);
171 /* round robin for fair dispatch */
172 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
174 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
176 WRITE_ONCE(hctx->dispatch_from, ctx);
179 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
181 struct request_queue *q = hctx->queue;
182 struct elevator_queue *e = q->elevator;
183 const bool has_sched_dispatch = e && e->type->ops.dispatch_request;
186 /* RCU or SRCU read lock is needed before checking quiesced flag */
187 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
193 * If we have previous entries on our dispatch list, grab them first for
194 * more fair dispatch.
196 if (!list_empty_careful(&hctx->dispatch)) {
197 spin_lock(&hctx->lock);
198 if (!list_empty(&hctx->dispatch))
199 list_splice_init(&hctx->dispatch, &rq_list);
200 spin_unlock(&hctx->lock);
204 * Only ask the scheduler for requests, if we didn't have residual
205 * requests from the dispatch list. This is to avoid the case where
206 * we only ever dispatch a fraction of the requests available because
207 * of low device queue depth. Once we pull requests out of the IO
208 * scheduler, we can no longer merge or sort them. So it's best to
209 * leave them there for as long as we can. Mark the hw queue as
210 * needing a restart in that case.
212 * We want to dispatch from the scheduler if there was nothing
213 * on the dispatch list or we were able to dispatch from the
216 if (!list_empty(&rq_list)) {
217 blk_mq_sched_mark_restart_hctx(hctx);
218 if (blk_mq_dispatch_rq_list(q, &rq_list, false)) {
219 if (has_sched_dispatch)
220 blk_mq_do_dispatch_sched(hctx);
222 blk_mq_do_dispatch_ctx(hctx);
224 } else if (has_sched_dispatch) {
225 blk_mq_do_dispatch_sched(hctx);
226 } else if (hctx->dispatch_busy) {
227 /* dequeue request one by one from sw queue if queue is busy */
228 blk_mq_do_dispatch_ctx(hctx);
230 blk_mq_flush_busy_ctxs(hctx, &rq_list);
231 blk_mq_dispatch_rq_list(q, &rq_list, false);
235 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
236 unsigned int nr_segs, struct request **merged_request)
240 switch (elv_merge(q, &rq, bio)) {
241 case ELEVATOR_BACK_MERGE:
242 if (!blk_mq_sched_allow_merge(q, rq, bio))
244 if (!bio_attempt_back_merge(rq, bio, nr_segs))
246 *merged_request = attempt_back_merge(q, rq);
247 if (!*merged_request)
248 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
250 case ELEVATOR_FRONT_MERGE:
251 if (!blk_mq_sched_allow_merge(q, rq, bio))
253 if (!bio_attempt_front_merge(rq, bio, nr_segs))
255 *merged_request = attempt_front_merge(q, rq);
256 if (!*merged_request)
257 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
259 case ELEVATOR_DISCARD_MERGE:
260 return bio_attempt_discard_merge(q, rq, bio);
265 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
268 * Iterate list of requests and see if we can merge this bio with any
271 bool blk_mq_bio_list_merge(struct request_queue *q, struct list_head *list,
272 struct bio *bio, unsigned int nr_segs)
277 list_for_each_entry_reverse(rq, list, queuelist) {
283 if (!blk_rq_merge_ok(rq, bio))
286 switch (blk_try_merge(rq, bio)) {
287 case ELEVATOR_BACK_MERGE:
288 if (blk_mq_sched_allow_merge(q, rq, bio))
289 merged = bio_attempt_back_merge(rq, bio,
292 case ELEVATOR_FRONT_MERGE:
293 if (blk_mq_sched_allow_merge(q, rq, bio))
294 merged = bio_attempt_front_merge(rq, bio,
297 case ELEVATOR_DISCARD_MERGE:
298 merged = bio_attempt_discard_merge(q, rq, bio);
309 EXPORT_SYMBOL_GPL(blk_mq_bio_list_merge);
312 * Reverse check our software queue for entries that we could potentially
313 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
314 * too much time checking for merges.
316 static bool blk_mq_attempt_merge(struct request_queue *q,
317 struct blk_mq_hw_ctx *hctx,
318 struct blk_mq_ctx *ctx, struct bio *bio,
319 unsigned int nr_segs)
321 enum hctx_type type = hctx->type;
323 lockdep_assert_held(&ctx->lock);
325 if (blk_mq_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) {
333 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio,
334 unsigned int nr_segs)
336 struct elevator_queue *e = q->elevator;
337 struct blk_mq_ctx *ctx;
338 struct blk_mq_hw_ctx *hctx;
342 if (e && e->type->ops.bio_merge)
343 return e->type->ops.bio_merge(q, bio, nr_segs);
345 ctx = blk_mq_get_ctx(q);
346 hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
348 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
349 !list_empty_careful(&ctx->rq_lists[type])) {
350 /* default per sw-queue merge */
351 spin_lock(&ctx->lock);
352 ret = blk_mq_attempt_merge(q, hctx, ctx, bio, nr_segs);
353 spin_unlock(&ctx->lock);
359 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
361 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
363 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
365 void blk_mq_sched_request_inserted(struct request *rq)
367 trace_block_rq_insert(rq->q, rq);
369 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
371 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
376 * dispatch flush and passthrough rq directly
378 * passthrough request has to be added to hctx->dispatch directly.
379 * For some reason, device may be in one situation which can't
380 * handle FS request, so STS_RESOURCE is always returned and the
381 * FS request will be added to hctx->dispatch. However passthrough
382 * request may be required at that time for fixing the problem. If
383 * passthrough request is added to scheduler queue, there isn't any
384 * chance to dispatch it given we prioritize requests in hctx->dispatch.
386 if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq))
390 rq->rq_flags |= RQF_SORTED;
395 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
396 bool run_queue, bool async)
398 struct request_queue *q = rq->q;
399 struct elevator_queue *e = q->elevator;
400 struct blk_mq_ctx *ctx = rq->mq_ctx;
401 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
403 /* flush rq in flush machinery need to be dispatched directly */
404 if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) {
405 blk_insert_flush(rq);
409 WARN_ON(e && (rq->tag != -1));
411 if (blk_mq_sched_bypass_insert(hctx, !!e, 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);
438 if (e && e->type->ops.insert_requests) {
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;
469 if (e && e->type->ops.insert_requests)
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 && !e && !run_queue_async) {
478 blk_mq_try_issue_list_directly(hctx, list);
479 if (list_empty(list))
482 blk_mq_insert_requests(hctx, ctx, list);
485 blk_mq_run_hw_queue(hctx, run_queue_async);
487 percpu_ref_put(&q->q_usage_counter);
490 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
491 struct blk_mq_hw_ctx *hctx,
492 unsigned int hctx_idx)
494 if (hctx->sched_tags) {
495 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
496 blk_mq_free_rq_map(hctx->sched_tags);
497 hctx->sched_tags = NULL;
501 static int blk_mq_sched_alloc_tags(struct request_queue *q,
502 struct blk_mq_hw_ctx *hctx,
503 unsigned int hctx_idx)
505 struct blk_mq_tag_set *set = q->tag_set;
508 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
510 if (!hctx->sched_tags)
513 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
515 blk_mq_sched_free_tags(set, hctx, hctx_idx);
520 /* called in queue's release handler, tagset has gone away */
521 static void blk_mq_sched_tags_teardown(struct request_queue *q)
523 struct blk_mq_hw_ctx *hctx;
526 queue_for_each_hw_ctx(q, hctx, i) {
527 if (hctx->sched_tags) {
528 blk_mq_free_rq_map(hctx->sched_tags);
529 hctx->sched_tags = NULL;
534 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
536 struct blk_mq_hw_ctx *hctx;
537 struct elevator_queue *eq;
543 q->nr_requests = q->tag_set->queue_depth;
548 * Default to double of smaller one between hw queue_depth and 128,
549 * since we don't split into sync/async like the old code did.
550 * Additionally, this is a per-hw queue depth.
552 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
555 queue_for_each_hw_ctx(q, hctx, i) {
556 ret = blk_mq_sched_alloc_tags(q, hctx, i);
561 ret = e->ops.init_sched(q, e);
565 blk_mq_debugfs_register_sched(q);
567 queue_for_each_hw_ctx(q, hctx, i) {
568 if (e->ops.init_hctx) {
569 ret = e->ops.init_hctx(hctx, i);
572 blk_mq_sched_free_requests(q);
573 blk_mq_exit_sched(q, eq);
574 kobject_put(&eq->kobj);
578 blk_mq_debugfs_register_sched_hctx(q, hctx);
584 blk_mq_sched_free_requests(q);
585 blk_mq_sched_tags_teardown(q);
591 * called in either blk_queue_cleanup or elevator_switch, tagset
592 * is required for freeing requests
594 void blk_mq_sched_free_requests(struct request_queue *q)
596 struct blk_mq_hw_ctx *hctx;
599 queue_for_each_hw_ctx(q, hctx, i) {
600 if (hctx->sched_tags)
601 blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i);
605 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
607 struct blk_mq_hw_ctx *hctx;
610 queue_for_each_hw_ctx(q, hctx, i) {
611 blk_mq_debugfs_unregister_sched_hctx(hctx);
612 if (e->type->ops.exit_hctx && hctx->sched_data) {
613 e->type->ops.exit_hctx(hctx, i);
614 hctx->sched_data = NULL;
617 blk_mq_debugfs_unregister_sched(q);
618 if (e->type->ops.exit_sched)
619 e->type->ops.exit_sched(e);
620 blk_mq_sched_tags_teardown(q);