2 * blk-mq scheduling framework
4 * Copyright (C) 2016 Jens Axboe
6 #include <linux/kernel.h>
7 #include <linux/module.h>
8 #include <linux/blk-mq.h>
10 #include <trace/events/block.h>
14 #include "blk-mq-debugfs.h"
15 #include "blk-mq-sched.h"
16 #include "blk-mq-tag.h"
19 void blk_mq_sched_free_hctx_data(struct request_queue *q,
20 void (*exit)(struct blk_mq_hw_ctx *))
22 struct blk_mq_hw_ctx *hctx;
25 queue_for_each_hw_ctx(q, hctx, i) {
26 if (exit && hctx->sched_data)
28 kfree(hctx->sched_data);
29 hctx->sched_data = NULL;
32 EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data);
34 void blk_mq_sched_assign_ioc(struct request *rq, struct bio *bio)
36 struct request_queue *q = rq->q;
37 struct io_context *ioc = rq_ioc(bio);
40 spin_lock_irq(q->queue_lock);
41 icq = ioc_lookup_icq(ioc, q);
42 spin_unlock_irq(q->queue_lock);
45 icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
49 get_io_context(icq->ioc);
54 * Mark a hardware queue as needing a restart. For shared queues, maintain
55 * a count of how many hardware queues are marked for restart.
57 static void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
59 if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
62 if (hctx->flags & BLK_MQ_F_TAG_SHARED) {
63 struct request_queue *q = hctx->queue;
65 if (!test_and_set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
66 atomic_inc(&q->shared_hctx_restart);
68 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
71 static bool blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx *hctx)
73 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
76 if (hctx->flags & BLK_MQ_F_TAG_SHARED) {
77 struct request_queue *q = hctx->queue;
79 if (test_and_clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
80 atomic_dec(&q->shared_hctx_restart);
82 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
84 if (blk_mq_hctx_has_pending(hctx)) {
85 blk_mq_run_hw_queue(hctx, true);
92 /* return true if hctx need to run again */
93 static bool 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.mq.has_work &&
104 !e->type->ops.mq.has_work(hctx))
107 ret = blk_mq_get_dispatch_budget(hctx);
108 if (ret == BLK_STS_RESOURCE)
111 rq = e->type->ops.mq.dispatch_request(hctx);
113 blk_mq_put_dispatch_budget(hctx);
115 } else if (ret != BLK_STS_OK) {
116 blk_mq_end_request(rq, ret);
121 * Now this rq owns the budget which has to be released
122 * if this rq won't be queued to driver via .queue_rq()
123 * in blk_mq_dispatch_rq_list().
125 list_add(&rq->queuelist, &rq_list);
126 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
131 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
132 struct blk_mq_ctx *ctx)
134 unsigned idx = ctx->index_hw;
136 if (++idx == hctx->nr_ctx)
139 return hctx->ctxs[idx];
142 /* return true if hctx need to run again */
143 static bool 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);
153 if (!sbitmap_any_bit_set(&hctx->ctx_map))
156 ret = blk_mq_get_dispatch_budget(hctx);
157 if (ret == BLK_STS_RESOURCE)
160 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
162 blk_mq_put_dispatch_budget(hctx);
164 } else if (ret != BLK_STS_OK) {
165 blk_mq_end_request(rq, ret);
170 * Now this rq owns the budget which has to be released
171 * if this rq won't be queued to driver via .queue_rq()
172 * in blk_mq_dispatch_rq_list().
174 list_add(&rq->queuelist, &rq_list);
176 /* round robin for fair dispatch */
177 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
179 } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
181 WRITE_ONCE(hctx->dispatch_from, ctx);
186 /* return true if hw queue need to be run again */
187 bool blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
189 struct request_queue *q = hctx->queue;
190 struct elevator_queue *e = q->elevator;
191 const bool has_sched_dispatch = e && e->type->ops.mq.dispatch_request;
193 bool run_queue = false;
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 run_queue = blk_mq_do_dispatch_sched(hctx);
231 run_queue = blk_mq_do_dispatch_ctx(hctx);
233 } else if (has_sched_dispatch) {
234 run_queue = blk_mq_do_dispatch_sched(hctx);
235 } else if (q->mq_ops->get_budget) {
237 * If we need to get budget before queuing request, we
238 * dequeue request one by one from sw queue for avoiding
239 * to mess up I/O merge when dispatch runs out of resource.
241 * TODO: get more budgets, and dequeue more requests in
244 run_queue = blk_mq_do_dispatch_ctx(hctx);
246 blk_mq_flush_busy_ctxs(hctx, &rq_list);
247 blk_mq_dispatch_rq_list(q, &rq_list, false);
250 if (run_queue && !blk_mq_sched_needs_restart(hctx) &&
251 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state)) {
252 blk_mq_sched_mark_restart_hctx(hctx);
259 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
260 struct request **merged_request)
264 switch (elv_merge(q, &rq, bio)) {
265 case ELEVATOR_BACK_MERGE:
266 if (!blk_mq_sched_allow_merge(q, rq, bio))
268 if (!bio_attempt_back_merge(q, rq, bio))
270 *merged_request = attempt_back_merge(q, rq);
271 if (!*merged_request)
272 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
274 case ELEVATOR_FRONT_MERGE:
275 if (!blk_mq_sched_allow_merge(q, rq, bio))
277 if (!bio_attempt_front_merge(q, rq, bio))
279 *merged_request = attempt_front_merge(q, rq);
280 if (!*merged_request)
281 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
287 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
290 * Reverse check our software queue for entries that we could potentially
291 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
292 * too much time checking for merges.
294 static bool blk_mq_attempt_merge(struct request_queue *q,
295 struct blk_mq_ctx *ctx, struct bio *bio)
300 lockdep_assert_held(&ctx->lock);
302 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
308 if (!blk_rq_merge_ok(rq, bio))
311 switch (blk_try_merge(rq, bio)) {
312 case ELEVATOR_BACK_MERGE:
313 if (blk_mq_sched_allow_merge(q, rq, bio))
314 merged = bio_attempt_back_merge(q, rq, bio);
316 case ELEVATOR_FRONT_MERGE:
317 if (blk_mq_sched_allow_merge(q, rq, bio))
318 merged = bio_attempt_front_merge(q, rq, bio);
320 case ELEVATOR_DISCARD_MERGE:
321 merged = bio_attempt_discard_merge(q, rq, bio);
335 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
337 struct elevator_queue *e = q->elevator;
338 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
339 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
342 if (e && e->type->ops.mq.bio_merge) {
344 return e->type->ops.mq.bio_merge(hctx, bio);
347 if (hctx->flags & BLK_MQ_F_SHOULD_MERGE) {
348 /* default per sw-queue merge */
349 spin_lock(&ctx->lock);
350 ret = blk_mq_attempt_merge(q, ctx, bio);
351 spin_unlock(&ctx->lock);
358 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
360 return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
362 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
364 void blk_mq_sched_request_inserted(struct request *rq)
366 trace_block_rq_insert(rq->q, rq);
368 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
370 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
374 rq->rq_flags |= RQF_SORTED;
379 * If we already have a real request tag, send directly to
382 spin_lock(&hctx->lock);
383 list_add(&rq->queuelist, &hctx->dispatch);
384 spin_unlock(&hctx->lock);
389 * list_for_each_entry_rcu_rr - iterate in a round-robin fashion over rcu list
391 * @skip: the list element that will not be examined. Iteration starts at
393 * @head: head of the list to examine. This list must have at least one
394 * element, namely @skip.
395 * @member: name of the list_head structure within typeof(*pos).
397 #define list_for_each_entry_rcu_rr(pos, skip, head, member) \
398 for ((pos) = (skip); \
399 (pos = (pos)->member.next != (head) ? list_entry_rcu( \
400 (pos)->member.next, typeof(*pos), member) : \
401 list_entry_rcu((pos)->member.next->next, typeof(*pos), member)), \
405 * Called after a driver tag has been freed to check whether a hctx needs to
406 * be restarted. Restarts @hctx if its tag set is not shared. Restarts hardware
407 * queues in a round-robin fashion if the tag set of @hctx is shared with other
410 void blk_mq_sched_restart(struct blk_mq_hw_ctx *const hctx)
412 struct blk_mq_tags *const tags = hctx->tags;
413 struct blk_mq_tag_set *const set = hctx->queue->tag_set;
414 struct request_queue *const queue = hctx->queue, *q;
415 struct blk_mq_hw_ctx *hctx2;
418 if (set->flags & BLK_MQ_F_TAG_SHARED) {
420 * If this is 0, then we know that no hardware queues
421 * have RESTART marked. We're done.
423 if (!atomic_read(&queue->shared_hctx_restart))
427 list_for_each_entry_rcu_rr(q, queue, &set->tag_list,
429 queue_for_each_hw_ctx(q, hctx2, i)
430 if (hctx2->tags == tags &&
431 blk_mq_sched_restart_hctx(hctx2))
434 j = hctx->queue_num + 1;
435 for (i = 0; i < queue->nr_hw_queues; i++, j++) {
436 if (j == queue->nr_hw_queues)
438 hctx2 = queue->queue_hw_ctx[j];
439 if (hctx2->tags == tags &&
440 blk_mq_sched_restart_hctx(hctx2))
446 blk_mq_sched_restart_hctx(hctx);
451 * Add flush/fua to the queue. If we fail getting a driver tag, then
452 * punt to the requeue list. Requeue will re-invoke us from a context
453 * that's safe to block from.
455 static void blk_mq_sched_insert_flush(struct blk_mq_hw_ctx *hctx,
456 struct request *rq, bool can_block)
458 if (blk_mq_get_driver_tag(rq, &hctx, can_block)) {
459 blk_insert_flush(rq);
460 blk_mq_run_hw_queue(hctx, true);
462 blk_mq_add_to_requeue_list(rq, false, true);
465 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
466 bool run_queue, bool async, bool can_block)
468 struct request_queue *q = rq->q;
469 struct elevator_queue *e = q->elevator;
470 struct blk_mq_ctx *ctx = rq->mq_ctx;
471 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
473 if (rq->tag == -1 && op_is_flush(rq->cmd_flags)) {
474 blk_mq_sched_insert_flush(hctx, rq, can_block);
478 if (e && blk_mq_sched_bypass_insert(hctx, rq))
481 if (e && e->type->ops.mq.insert_requests) {
484 list_add(&rq->queuelist, &list);
485 e->type->ops.mq.insert_requests(hctx, &list, at_head);
487 spin_lock(&ctx->lock);
488 __blk_mq_insert_request(hctx, rq, at_head);
489 spin_unlock(&ctx->lock);
494 blk_mq_run_hw_queue(hctx, async);
497 void blk_mq_sched_insert_requests(struct request_queue *q,
498 struct blk_mq_ctx *ctx,
499 struct list_head *list, bool run_queue_async)
501 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
502 struct elevator_queue *e = hctx->queue->elevator;
505 struct request *rq, *next;
508 * We bypass requests that already have a driver tag assigned,
509 * which should only be flushes. Flushes are only ever inserted
510 * as single requests, so we shouldn't ever hit the
511 * WARN_ON_ONCE() below (but let's handle it just in case).
513 list_for_each_entry_safe(rq, next, list, queuelist) {
514 if (WARN_ON_ONCE(rq->tag != -1)) {
515 list_del_init(&rq->queuelist);
516 blk_mq_sched_bypass_insert(hctx, rq);
521 if (e && e->type->ops.mq.insert_requests)
522 e->type->ops.mq.insert_requests(hctx, list, false);
524 blk_mq_insert_requests(hctx, ctx, list);
526 blk_mq_run_hw_queue(hctx, run_queue_async);
529 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
530 struct blk_mq_hw_ctx *hctx,
531 unsigned int hctx_idx)
533 if (hctx->sched_tags) {
534 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
535 blk_mq_free_rq_map(hctx->sched_tags);
536 hctx->sched_tags = NULL;
540 static int blk_mq_sched_alloc_tags(struct request_queue *q,
541 struct blk_mq_hw_ctx *hctx,
542 unsigned int hctx_idx)
544 struct blk_mq_tag_set *set = q->tag_set;
547 hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
549 if (!hctx->sched_tags)
552 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
554 blk_mq_sched_free_tags(set, hctx, hctx_idx);
559 static void blk_mq_sched_tags_teardown(struct request_queue *q)
561 struct blk_mq_tag_set *set = q->tag_set;
562 struct blk_mq_hw_ctx *hctx;
565 queue_for_each_hw_ctx(q, hctx, i)
566 blk_mq_sched_free_tags(set, hctx, i);
569 int blk_mq_sched_init_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
570 unsigned int hctx_idx)
572 struct elevator_queue *e = q->elevator;
578 ret = blk_mq_sched_alloc_tags(q, hctx, hctx_idx);
582 if (e->type->ops.mq.init_hctx) {
583 ret = e->type->ops.mq.init_hctx(hctx, hctx_idx);
585 blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
590 blk_mq_debugfs_register_sched_hctx(q, hctx);
595 void blk_mq_sched_exit_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
596 unsigned int hctx_idx)
598 struct elevator_queue *e = q->elevator;
603 blk_mq_debugfs_unregister_sched_hctx(hctx);
605 if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
606 e->type->ops.mq.exit_hctx(hctx, hctx_idx);
607 hctx->sched_data = NULL;
610 blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
613 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
615 struct blk_mq_hw_ctx *hctx;
616 struct elevator_queue *eq;
626 * Default to double of smaller one between hw queue_depth and 128,
627 * since we don't split into sync/async like the old code did.
628 * Additionally, this is a per-hw queue depth.
630 q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
633 queue_for_each_hw_ctx(q, hctx, i) {
634 ret = blk_mq_sched_alloc_tags(q, hctx, i);
639 ret = e->ops.mq.init_sched(q, e);
643 blk_mq_debugfs_register_sched(q);
645 queue_for_each_hw_ctx(q, hctx, i) {
646 if (e->ops.mq.init_hctx) {
647 ret = e->ops.mq.init_hctx(hctx, i);
650 blk_mq_exit_sched(q, eq);
651 kobject_put(&eq->kobj);
655 blk_mq_debugfs_register_sched_hctx(q, hctx);
661 blk_mq_sched_tags_teardown(q);
666 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
668 struct blk_mq_hw_ctx *hctx;
671 queue_for_each_hw_ctx(q, hctx, i) {
672 blk_mq_debugfs_unregister_sched_hctx(hctx);
673 if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
674 e->type->ops.mq.exit_hctx(hctx, i);
675 hctx->sched_data = NULL;
678 blk_mq_debugfs_unregister_sched(q);
679 if (e->type->ops.mq.exit_sched)
680 e->type->ops.mq.exit_sched(e);
681 blk_mq_sched_tags_teardown(q);
685 int blk_mq_sched_init(struct request_queue *q)
689 mutex_lock(&q->sysfs_lock);
690 ret = elevator_init(q, NULL);
691 mutex_unlock(&q->sysfs_lock);