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 static void 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;
99 if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
102 if (!blk_mq_get_dispatch_budget(hctx))
105 rq = e->type->ops.dispatch_request(hctx);
107 blk_mq_put_dispatch_budget(hctx);
109 * We're releasing without dispatching. Holding the
110 * budget could have blocked any "hctx"s with the
111 * same queue and if we didn't dispatch then there's
112 * no guarantee anyone will kick the queue. Kick it
115 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
120 * Now this rq owns the budget which has to be released
121 * if this rq won't be queued to driver via .queue_rq()
122 * in blk_mq_dispatch_rq_list().
124 list_add(&rq->queuelist, &rq_list);
125 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
128 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
129 struct blk_mq_ctx *ctx)
131 unsigned short idx = ctx->index_hw[hctx->type];
133 if (++idx == hctx->nr_ctx)
136 return hctx->ctxs[idx];
140 * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
141 * its queue by itself in its completion handler, so we don't need to
142 * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
144 static void blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
146 struct request_queue *q = hctx->queue;
148 struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
153 if (!sbitmap_any_bit_set(&hctx->ctx_map))
156 if (!blk_mq_get_dispatch_budget(hctx))
159 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
161 blk_mq_put_dispatch_budget(hctx);
163 * We're releasing without dispatching. Holding the
164 * budget could have blocked any "hctx"s with the
165 * same queue and if we didn't dispatch then there's
166 * no guarantee anyone will kick the queue. Kick it
169 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
174 * Now this rq owns the budget which has to be released
175 * if this rq won't be queued to driver via .queue_rq()
176 * in blk_mq_dispatch_rq_list().
178 list_add(&rq->queuelist, &rq_list);
180 /* round robin for fair dispatch */
181 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
183 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
185 WRITE_ONCE(hctx->dispatch_from, ctx);
188 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
190 struct request_queue *q = hctx->queue;
191 struct elevator_queue *e = q->elevator;
192 const bool has_sched_dispatch = e && e->type->ops.dispatch_request;
195 /* RCU or SRCU read lock is needed before checking quiesced flag */
196 if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
202 * If we have previous entries on our dispatch list, grab them first for
203 * more fair dispatch.
205 if (!list_empty_careful(&hctx->dispatch)) {
206 spin_lock(&hctx->lock);
207 if (!list_empty(&hctx->dispatch))
208 list_splice_init(&hctx->dispatch, &rq_list);
209 spin_unlock(&hctx->lock);
213 * Only ask the scheduler for requests, if we didn't have residual
214 * requests from the dispatch list. This is to avoid the case where
215 * we only ever dispatch a fraction of the requests available because
216 * of low device queue depth. Once we pull requests out of the IO
217 * scheduler, we can no longer merge or sort them. So it's best to
218 * leave them there for as long as we can. Mark the hw queue as
219 * needing a restart in that case.
221 * We want to dispatch from the scheduler if there was nothing
222 * on the dispatch list or we were able to dispatch from the
225 if (!list_empty(&rq_list)) {
226 blk_mq_sched_mark_restart_hctx(hctx);
227 if (blk_mq_dispatch_rq_list(q, &rq_list, false)) {
228 if (has_sched_dispatch)
229 blk_mq_do_dispatch_sched(hctx);
231 blk_mq_do_dispatch_ctx(hctx);
233 } else if (has_sched_dispatch) {
234 blk_mq_do_dispatch_sched(hctx);
235 } else if (hctx->dispatch_busy) {
236 /* dequeue request one by one from sw queue if queue is busy */
237 blk_mq_do_dispatch_ctx(hctx);
239 blk_mq_flush_busy_ctxs(hctx, &rq_list);
240 blk_mq_dispatch_rq_list(q, &rq_list, false);
244 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
245 unsigned int nr_segs, struct request **merged_request)
249 switch (elv_merge(q, &rq, bio)) {
250 case ELEVATOR_BACK_MERGE:
251 if (!blk_mq_sched_allow_merge(q, rq, bio))
253 if (!bio_attempt_back_merge(rq, bio, nr_segs))
255 *merged_request = attempt_back_merge(q, rq);
256 if (!*merged_request)
257 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
259 case ELEVATOR_FRONT_MERGE:
260 if (!blk_mq_sched_allow_merge(q, rq, bio))
262 if (!bio_attempt_front_merge(rq, bio, nr_segs))
264 *merged_request = attempt_front_merge(q, rq);
265 if (!*merged_request)
266 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
268 case ELEVATOR_DISCARD_MERGE:
269 return bio_attempt_discard_merge(q, rq, bio);
274 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
277 * Iterate list of requests and see if we can merge this bio with any
280 bool blk_mq_bio_list_merge(struct request_queue *q, struct list_head *list,
281 struct bio *bio, unsigned int nr_segs)
286 list_for_each_entry_reverse(rq, list, queuelist) {
292 if (!blk_rq_merge_ok(rq, bio))
295 switch (blk_try_merge(rq, bio)) {
296 case ELEVATOR_BACK_MERGE:
297 if (blk_mq_sched_allow_merge(q, rq, bio))
298 merged = bio_attempt_back_merge(rq, bio,
301 case ELEVATOR_FRONT_MERGE:
302 if (blk_mq_sched_allow_merge(q, rq, bio))
303 merged = bio_attempt_front_merge(rq, bio,
306 case ELEVATOR_DISCARD_MERGE:
307 merged = bio_attempt_discard_merge(q, rq, bio);
318 EXPORT_SYMBOL_GPL(blk_mq_bio_list_merge);
321 * Reverse check our software queue for entries that we could potentially
322 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
323 * too much time checking for merges.
325 static bool blk_mq_attempt_merge(struct request_queue *q,
326 struct blk_mq_hw_ctx *hctx,
327 struct blk_mq_ctx *ctx, struct bio *bio,
328 unsigned int nr_segs)
330 enum hctx_type type = hctx->type;
332 lockdep_assert_held(&ctx->lock);
334 if (blk_mq_bio_list_merge(q, &ctx->rq_lists[type], bio, nr_segs)) {
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 = blk_mq_get_ctx(q);
347 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
351 if (e && e->type->ops.bio_merge)
352 return e->type->ops.bio_merge(hctx, bio, nr_segs);
355 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
356 !list_empty_careful(&ctx->rq_lists[type])) {
357 /* default per sw-queue merge */
358 spin_lock(&ctx->lock);
359 ret = blk_mq_attempt_merge(q, hctx, ctx, bio, nr_segs);
360 spin_unlock(&ctx->lock);
366 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
368 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
370 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
372 void blk_mq_sched_request_inserted(struct request *rq)
374 trace_block_rq_insert(rq->q, rq);
376 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
378 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
383 * dispatch flush and passthrough rq directly
385 * passthrough request has to be added to hctx->dispatch directly.
386 * For some reason, device may be in one situation which can't
387 * handle FS request, so STS_RESOURCE is always returned and the
388 * FS request will be added to hctx->dispatch. However passthrough
389 * request may be required at that time for fixing the problem. If
390 * passthrough request is added to scheduler queue, there isn't any
391 * chance to dispatch it given we prioritize requests in hctx->dispatch.
393 if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq))
397 rq->rq_flags |= RQF_SORTED;
402 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
403 bool run_queue, bool async)
405 struct request_queue *q = rq->q;
406 struct elevator_queue *e = q->elevator;
407 struct blk_mq_ctx *ctx = rq->mq_ctx;
408 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
410 /* flush rq in flush machinery need to be dispatched directly */
411 if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) {
412 blk_insert_flush(rq);
416 WARN_ON(e && (rq->tag != -1));
418 if (blk_mq_sched_bypass_insert(hctx, !!e, rq)) {
420 * Firstly normal IO request is inserted to scheduler queue or
421 * sw queue, meantime we add flush request to dispatch queue(
422 * hctx->dispatch) directly and there is at most one in-flight
423 * flush request for each hw queue, so it doesn't matter to add
424 * flush request to tail or front of the dispatch queue.
426 * Secondly in case of NCQ, flush request belongs to non-NCQ
427 * command, and queueing it will fail when there is any
428 * in-flight normal IO request(NCQ command). When adding flush
429 * rq to the front of hctx->dispatch, it is easier to introduce
430 * extra time to flush rq's latency because of S_SCHED_RESTART
431 * compared with adding to the tail of dispatch queue, then
432 * chance of flush merge is increased, and less flush requests
433 * will be issued to controller. It is observed that ~10% time
434 * is saved in blktests block/004 on disk attached to AHCI/NCQ
435 * drive when adding flush rq to the front of hctx->dispatch.
437 * Simply queue flush rq to the front of hctx->dispatch so that
438 * intensive flush workloads can benefit in case of NCQ HW.
440 at_head = (rq->rq_flags & RQF_FLUSH_SEQ) ? true : at_head;
441 blk_mq_request_bypass_insert(rq, at_head, false);
445 if (e && e->type->ops.insert_requests) {
448 list_add(&rq->queuelist, &list);
449 e->type->ops.insert_requests(hctx, &list, at_head);
451 spin_lock(&ctx->lock);
452 __blk_mq_insert_request(hctx, rq, at_head);
453 spin_unlock(&ctx->lock);
458 blk_mq_run_hw_queue(hctx, async);
461 void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
462 struct blk_mq_ctx *ctx,
463 struct list_head *list, bool run_queue_async)
465 struct elevator_queue *e;
466 struct request_queue *q = hctx->queue;
469 * blk_mq_sched_insert_requests() is called from flush plug
470 * context only, and hold one usage counter to prevent queue
471 * from being released.
473 percpu_ref_get(&q->q_usage_counter);
475 e = hctx->queue->elevator;
476 if (e && e->type->ops.insert_requests)
477 e->type->ops.insert_requests(hctx, list, false);
480 * try to issue requests directly if the hw queue isn't
481 * busy in case of 'none' scheduler, and this way may save
482 * us one extra enqueue & dequeue to sw queue.
484 if (!hctx->dispatch_busy && !e && !run_queue_async) {
485 blk_mq_try_issue_list_directly(hctx, list);
486 if (list_empty(list))
489 blk_mq_insert_requests(hctx, ctx, list);
492 blk_mq_run_hw_queue(hctx, run_queue_async);
494 percpu_ref_put(&q->q_usage_counter);
497 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
498 struct blk_mq_hw_ctx *hctx,
499 unsigned int hctx_idx)
501 if (hctx->sched_tags) {
502 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
503 blk_mq_free_rq_map(hctx->sched_tags);
504 hctx->sched_tags = NULL;
508 static int blk_mq_sched_alloc_tags(struct request_queue *q,
509 struct blk_mq_hw_ctx *hctx,
510 unsigned int hctx_idx)
512 struct blk_mq_tag_set *set = q->tag_set;
515 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
517 if (!hctx->sched_tags)
520 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
522 blk_mq_sched_free_tags(set, hctx, hctx_idx);
527 /* called in queue's release handler, tagset has gone away */
528 static void blk_mq_sched_tags_teardown(struct request_queue *q)
530 struct blk_mq_hw_ctx *hctx;
533 queue_for_each_hw_ctx(q, hctx, i) {
534 if (hctx->sched_tags) {
535 blk_mq_free_rq_map(hctx->sched_tags);
536 hctx->sched_tags = NULL;
541 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
543 struct blk_mq_hw_ctx *hctx;
544 struct elevator_queue *eq;
550 q->nr_requests = q->tag_set->queue_depth;
555 * Default to double of smaller one between hw queue_depth and 128,
556 * since we don't split into sync/async like the old code did.
557 * Additionally, this is a per-hw queue depth.
559 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
562 queue_for_each_hw_ctx(q, hctx, i) {
563 ret = blk_mq_sched_alloc_tags(q, hctx, i);
568 ret = e->ops.init_sched(q, e);
572 blk_mq_debugfs_register_sched(q);
574 queue_for_each_hw_ctx(q, hctx, i) {
575 if (e->ops.init_hctx) {
576 ret = e->ops.init_hctx(hctx, i);
579 blk_mq_sched_free_requests(q);
580 blk_mq_exit_sched(q, eq);
581 kobject_put(&eq->kobj);
585 blk_mq_debugfs_register_sched_hctx(q, hctx);
591 blk_mq_sched_free_requests(q);
592 blk_mq_sched_tags_teardown(q);
598 * called in either blk_queue_cleanup or elevator_switch, tagset
599 * is required for freeing requests
601 void blk_mq_sched_free_requests(struct request_queue *q)
603 struct blk_mq_hw_ctx *hctx;
606 queue_for_each_hw_ctx(q, hctx, i) {
607 if (hctx->sched_tags)
608 blk_mq_free_rqs(q->tag_set, hctx->sched_tags, i);
612 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
614 struct blk_mq_hw_ctx *hctx;
617 queue_for_each_hw_ctx(q, hctx, i) {
618 blk_mq_debugfs_unregister_sched_hctx(hctx);
619 if (e->type->ops.exit_hctx && hctx->sched_data) {
620 e->type->ops.exit_hctx(hctx, i);
621 hctx->sched_data = NULL;
624 blk_mq_debugfs_unregister_sched(q);
625 if (e->type->ops.exit_sched)
626 e->type->ops.exit_sched(e);
627 blk_mq_sched_tags_teardown(q);