2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/kmemleak.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/workqueue.h>
17 #include <linux/smp.h>
18 #include <linux/llist.h>
19 #include <linux/list_sort.h>
20 #include <linux/cpu.h>
21 #include <linux/cache.h>
22 #include <linux/sched/sysctl.h>
23 #include <linux/sched/topology.h>
24 #include <linux/sched/signal.h>
25 #include <linux/delay.h>
26 #include <linux/crash_dump.h>
27 #include <linux/prefetch.h>
29 #include <trace/events/block.h>
31 #include <linux/blk-mq.h>
34 #include "blk-mq-tag.h"
37 #include "blk-mq-sched.h"
39 static DEFINE_MUTEX(all_q_mutex);
40 static LIST_HEAD(all_q_list);
42 static void blk_mq_poll_stats_start(struct request_queue *q);
43 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
46 * Check if any of the ctx's have pending work in this hardware queue
48 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
50 return sbitmap_any_bit_set(&hctx->ctx_map) ||
51 !list_empty_careful(&hctx->dispatch) ||
52 blk_mq_sched_has_work(hctx);
56 * Mark this ctx as having pending work in this hardware queue
58 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
59 struct blk_mq_ctx *ctx)
61 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
62 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
65 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
66 struct blk_mq_ctx *ctx)
68 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
71 void blk_mq_freeze_queue_start(struct request_queue *q)
75 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
76 if (freeze_depth == 1) {
77 percpu_ref_kill(&q->q_usage_counter);
78 blk_mq_run_hw_queues(q, false);
81 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
83 void blk_mq_freeze_queue_wait(struct request_queue *q)
85 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
87 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
89 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
90 unsigned long timeout)
92 return wait_event_timeout(q->mq_freeze_wq,
93 percpu_ref_is_zero(&q->q_usage_counter),
96 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
99 * Guarantee no request is in use, so we can change any data structure of
100 * the queue afterward.
102 void blk_freeze_queue(struct request_queue *q)
105 * In the !blk_mq case we are only calling this to kill the
106 * q_usage_counter, otherwise this increases the freeze depth
107 * and waits for it to return to zero. For this reason there is
108 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
109 * exported to drivers as the only user for unfreeze is blk_mq.
111 blk_mq_freeze_queue_start(q);
112 blk_mq_freeze_queue_wait(q);
115 void blk_mq_freeze_queue(struct request_queue *q)
118 * ...just an alias to keep freeze and unfreeze actions balanced
119 * in the blk_mq_* namespace
123 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
125 void blk_mq_unfreeze_queue(struct request_queue *q)
129 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
130 WARN_ON_ONCE(freeze_depth < 0);
132 percpu_ref_reinit(&q->q_usage_counter);
133 wake_up_all(&q->mq_freeze_wq);
136 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
139 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
142 * Note: this function does not prevent that the struct request end_io()
143 * callback function is invoked. Additionally, it is not prevented that
144 * new queue_rq() calls occur unless the queue has been stopped first.
146 void blk_mq_quiesce_queue(struct request_queue *q)
148 struct blk_mq_hw_ctx *hctx;
152 blk_mq_stop_hw_queues(q);
154 queue_for_each_hw_ctx(q, hctx, i) {
155 if (hctx->flags & BLK_MQ_F_BLOCKING)
156 synchronize_srcu(&hctx->queue_rq_srcu);
163 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
165 void blk_mq_wake_waiters(struct request_queue *q)
167 struct blk_mq_hw_ctx *hctx;
170 queue_for_each_hw_ctx(q, hctx, i)
171 if (blk_mq_hw_queue_mapped(hctx))
172 blk_mq_tag_wakeup_all(hctx->tags, true);
175 * If we are called because the queue has now been marked as
176 * dying, we need to ensure that processes currently waiting on
177 * the queue are notified as well.
179 wake_up_all(&q->mq_freeze_wq);
182 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
184 return blk_mq_has_free_tags(hctx->tags);
186 EXPORT_SYMBOL(blk_mq_can_queue);
188 void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
189 struct request *rq, unsigned int op)
191 INIT_LIST_HEAD(&rq->queuelist);
192 /* csd/requeue_work/fifo_time is initialized before use */
196 if (blk_queue_io_stat(q))
197 rq->rq_flags |= RQF_IO_STAT;
198 /* do not touch atomic flags, it needs atomic ops against the timer */
200 INIT_HLIST_NODE(&rq->hash);
201 RB_CLEAR_NODE(&rq->rb_node);
204 rq->start_time = jiffies;
205 #ifdef CONFIG_BLK_CGROUP
207 set_start_time_ns(rq);
208 rq->io_start_time_ns = 0;
210 rq->nr_phys_segments = 0;
211 #if defined(CONFIG_BLK_DEV_INTEGRITY)
212 rq->nr_integrity_segments = 0;
215 /* tag was already set */
219 INIT_LIST_HEAD(&rq->timeout_list);
223 rq->end_io_data = NULL;
226 ctx->rq_dispatched[op_is_sync(op)]++;
228 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init);
230 struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data,
236 tag = blk_mq_get_tag(data);
237 if (tag != BLK_MQ_TAG_FAIL) {
238 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
240 rq = tags->static_rqs[tag];
242 if (data->flags & BLK_MQ_REQ_INTERNAL) {
244 rq->internal_tag = tag;
246 if (blk_mq_tag_busy(data->hctx)) {
247 rq->rq_flags = RQF_MQ_INFLIGHT;
248 atomic_inc(&data->hctx->nr_active);
251 rq->internal_tag = -1;
252 data->hctx->tags->rqs[rq->tag] = rq;
255 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op);
261 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request);
263 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
266 struct blk_mq_alloc_data alloc_data = { .flags = flags };
270 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
274 rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data);
276 blk_mq_put_ctx(alloc_data.ctx);
280 return ERR_PTR(-EWOULDBLOCK);
283 rq->__sector = (sector_t) -1;
284 rq->bio = rq->biotail = NULL;
287 EXPORT_SYMBOL(blk_mq_alloc_request);
289 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
290 unsigned int flags, unsigned int hctx_idx)
292 struct blk_mq_alloc_data alloc_data = { .flags = flags };
298 * If the tag allocator sleeps we could get an allocation for a
299 * different hardware context. No need to complicate the low level
300 * allocator for this for the rare use case of a command tied to
303 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
304 return ERR_PTR(-EINVAL);
306 if (hctx_idx >= q->nr_hw_queues)
307 return ERR_PTR(-EIO);
309 ret = blk_queue_enter(q, true);
314 * Check if the hardware context is actually mapped to anything.
315 * If not tell the caller that it should skip this queue.
317 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
318 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
320 return ERR_PTR(-EXDEV);
322 cpu = cpumask_first(alloc_data.hctx->cpumask);
323 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
325 rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data);
327 blk_mq_put_ctx(alloc_data.ctx);
331 return ERR_PTR(-EWOULDBLOCK);
335 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
337 void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
340 const int sched_tag = rq->internal_tag;
341 struct request_queue *q = rq->q;
343 if (rq->rq_flags & RQF_MQ_INFLIGHT)
344 atomic_dec(&hctx->nr_active);
346 wbt_done(q->rq_wb, &rq->issue_stat);
349 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
350 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
352 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
354 blk_mq_sched_completed_request(hctx, rq);
355 blk_mq_sched_restart_queues(hctx);
359 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx,
362 struct blk_mq_ctx *ctx = rq->mq_ctx;
364 ctx->rq_completed[rq_is_sync(rq)]++;
365 __blk_mq_finish_request(hctx, ctx, rq);
368 void blk_mq_finish_request(struct request *rq)
370 blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq);
373 void blk_mq_free_request(struct request *rq)
375 blk_mq_sched_put_request(rq);
377 EXPORT_SYMBOL_GPL(blk_mq_free_request);
379 inline void __blk_mq_end_request(struct request *rq, int error)
381 blk_account_io_done(rq);
384 wbt_done(rq->q->rq_wb, &rq->issue_stat);
385 rq->end_io(rq, error);
387 if (unlikely(blk_bidi_rq(rq)))
388 blk_mq_free_request(rq->next_rq);
389 blk_mq_free_request(rq);
392 EXPORT_SYMBOL(__blk_mq_end_request);
394 void blk_mq_end_request(struct request *rq, int error)
396 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
398 __blk_mq_end_request(rq, error);
400 EXPORT_SYMBOL(blk_mq_end_request);
402 static void __blk_mq_complete_request_remote(void *data)
404 struct request *rq = data;
406 rq->q->softirq_done_fn(rq);
409 static void blk_mq_ipi_complete_request(struct request *rq)
411 struct blk_mq_ctx *ctx = rq->mq_ctx;
415 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
416 rq->q->softirq_done_fn(rq);
421 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
422 shared = cpus_share_cache(cpu, ctx->cpu);
424 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
425 rq->csd.func = __blk_mq_complete_request_remote;
428 smp_call_function_single_async(ctx->cpu, &rq->csd);
430 rq->q->softirq_done_fn(rq);
435 static void blk_mq_stat_add(struct request *rq)
437 if (rq->rq_flags & RQF_STATS) {
438 blk_mq_poll_stats_start(rq->q);
443 static void __blk_mq_complete_request(struct request *rq)
445 struct request_queue *q = rq->q;
449 if (!q->softirq_done_fn)
450 blk_mq_end_request(rq, rq->errors);
452 blk_mq_ipi_complete_request(rq);
456 * blk_mq_complete_request - end I/O on a request
457 * @rq: the request being processed
460 * Ends all I/O on a request. It does not handle partial completions.
461 * The actual completion happens out-of-order, through a IPI handler.
463 void blk_mq_complete_request(struct request *rq, int error)
465 struct request_queue *q = rq->q;
467 if (unlikely(blk_should_fake_timeout(q)))
469 if (!blk_mark_rq_complete(rq)) {
471 __blk_mq_complete_request(rq);
474 EXPORT_SYMBOL(blk_mq_complete_request);
476 int blk_mq_request_started(struct request *rq)
478 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
480 EXPORT_SYMBOL_GPL(blk_mq_request_started);
482 void blk_mq_start_request(struct request *rq)
484 struct request_queue *q = rq->q;
486 blk_mq_sched_started_request(rq);
488 trace_block_rq_issue(q, rq);
490 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
491 blk_stat_set_issue_time(&rq->issue_stat);
492 rq->rq_flags |= RQF_STATS;
493 wbt_issue(q->rq_wb, &rq->issue_stat);
499 * Ensure that ->deadline is visible before set the started
500 * flag and clear the completed flag.
502 smp_mb__before_atomic();
505 * Mark us as started and clear complete. Complete might have been
506 * set if requeue raced with timeout, which then marked it as
507 * complete. So be sure to clear complete again when we start
508 * the request, otherwise we'll ignore the completion event.
510 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
511 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
512 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
513 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
515 if (q->dma_drain_size && blk_rq_bytes(rq)) {
517 * Make sure space for the drain appears. We know we can do
518 * this because max_hw_segments has been adjusted to be one
519 * fewer than the device can handle.
521 rq->nr_phys_segments++;
524 EXPORT_SYMBOL(blk_mq_start_request);
526 static void __blk_mq_requeue_request(struct request *rq)
528 struct request_queue *q = rq->q;
530 trace_block_rq_requeue(q, rq);
531 wbt_requeue(q->rq_wb, &rq->issue_stat);
532 blk_mq_sched_requeue_request(rq);
534 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
535 if (q->dma_drain_size && blk_rq_bytes(rq))
536 rq->nr_phys_segments--;
540 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
542 __blk_mq_requeue_request(rq);
544 BUG_ON(blk_queued_rq(rq));
545 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
547 EXPORT_SYMBOL(blk_mq_requeue_request);
549 static void blk_mq_requeue_work(struct work_struct *work)
551 struct request_queue *q =
552 container_of(work, struct request_queue, requeue_work.work);
554 struct request *rq, *next;
557 spin_lock_irqsave(&q->requeue_lock, flags);
558 list_splice_init(&q->requeue_list, &rq_list);
559 spin_unlock_irqrestore(&q->requeue_lock, flags);
561 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
562 if (!(rq->rq_flags & RQF_SOFTBARRIER))
565 rq->rq_flags &= ~RQF_SOFTBARRIER;
566 list_del_init(&rq->queuelist);
567 blk_mq_sched_insert_request(rq, true, false, false, true);
570 while (!list_empty(&rq_list)) {
571 rq = list_entry(rq_list.next, struct request, queuelist);
572 list_del_init(&rq->queuelist);
573 blk_mq_sched_insert_request(rq, false, false, false, true);
576 blk_mq_run_hw_queues(q, false);
579 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
580 bool kick_requeue_list)
582 struct request_queue *q = rq->q;
586 * We abuse this flag that is otherwise used by the I/O scheduler to
587 * request head insertation from the workqueue.
589 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
591 spin_lock_irqsave(&q->requeue_lock, flags);
593 rq->rq_flags |= RQF_SOFTBARRIER;
594 list_add(&rq->queuelist, &q->requeue_list);
596 list_add_tail(&rq->queuelist, &q->requeue_list);
598 spin_unlock_irqrestore(&q->requeue_lock, flags);
600 if (kick_requeue_list)
601 blk_mq_kick_requeue_list(q);
603 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
605 void blk_mq_kick_requeue_list(struct request_queue *q)
607 kblockd_schedule_delayed_work(&q->requeue_work, 0);
609 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
611 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
614 kblockd_schedule_delayed_work(&q->requeue_work,
615 msecs_to_jiffies(msecs));
617 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
619 void blk_mq_abort_requeue_list(struct request_queue *q)
624 spin_lock_irqsave(&q->requeue_lock, flags);
625 list_splice_init(&q->requeue_list, &rq_list);
626 spin_unlock_irqrestore(&q->requeue_lock, flags);
628 while (!list_empty(&rq_list)) {
631 rq = list_first_entry(&rq_list, struct request, queuelist);
632 list_del_init(&rq->queuelist);
634 blk_mq_end_request(rq, rq->errors);
637 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
639 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
641 if (tag < tags->nr_tags) {
642 prefetch(tags->rqs[tag]);
643 return tags->rqs[tag];
648 EXPORT_SYMBOL(blk_mq_tag_to_rq);
650 struct blk_mq_timeout_data {
652 unsigned int next_set;
655 void blk_mq_rq_timed_out(struct request *req, bool reserved)
657 const struct blk_mq_ops *ops = req->q->mq_ops;
658 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
661 * We know that complete is set at this point. If STARTED isn't set
662 * anymore, then the request isn't active and the "timeout" should
663 * just be ignored. This can happen due to the bitflag ordering.
664 * Timeout first checks if STARTED is set, and if it is, assumes
665 * the request is active. But if we race with completion, then
666 * we both flags will get cleared. So check here again, and ignore
667 * a timeout event with a request that isn't active.
669 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
673 ret = ops->timeout(req, reserved);
677 __blk_mq_complete_request(req);
679 case BLK_EH_RESET_TIMER:
681 blk_clear_rq_complete(req);
683 case BLK_EH_NOT_HANDLED:
686 printk(KERN_ERR "block: bad eh return: %d\n", ret);
691 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
692 struct request *rq, void *priv, bool reserved)
694 struct blk_mq_timeout_data *data = priv;
696 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
698 * If a request wasn't started before the queue was
699 * marked dying, kill it here or it'll go unnoticed.
701 if (unlikely(blk_queue_dying(rq->q))) {
703 blk_mq_end_request(rq, rq->errors);
708 if (time_after_eq(jiffies, rq->deadline)) {
709 if (!blk_mark_rq_complete(rq))
710 blk_mq_rq_timed_out(rq, reserved);
711 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
712 data->next = rq->deadline;
717 static void blk_mq_timeout_work(struct work_struct *work)
719 struct request_queue *q =
720 container_of(work, struct request_queue, timeout_work);
721 struct blk_mq_timeout_data data = {
727 /* A deadlock might occur if a request is stuck requiring a
728 * timeout at the same time a queue freeze is waiting
729 * completion, since the timeout code would not be able to
730 * acquire the queue reference here.
732 * That's why we don't use blk_queue_enter here; instead, we use
733 * percpu_ref_tryget directly, because we need to be able to
734 * obtain a reference even in the short window between the queue
735 * starting to freeze, by dropping the first reference in
736 * blk_mq_freeze_queue_start, and the moment the last request is
737 * consumed, marked by the instant q_usage_counter reaches
740 if (!percpu_ref_tryget(&q->q_usage_counter))
743 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
746 data.next = blk_rq_timeout(round_jiffies_up(data.next));
747 mod_timer(&q->timeout, data.next);
749 struct blk_mq_hw_ctx *hctx;
751 queue_for_each_hw_ctx(q, hctx, i) {
752 /* the hctx may be unmapped, so check it here */
753 if (blk_mq_hw_queue_mapped(hctx))
754 blk_mq_tag_idle(hctx);
761 * Reverse check our software queue for entries that we could potentially
762 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
763 * too much time checking for merges.
765 static bool blk_mq_attempt_merge(struct request_queue *q,
766 struct blk_mq_ctx *ctx, struct bio *bio)
771 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
777 if (!blk_rq_merge_ok(rq, bio))
780 switch (blk_try_merge(rq, bio)) {
781 case ELEVATOR_BACK_MERGE:
782 if (blk_mq_sched_allow_merge(q, rq, bio))
783 merged = bio_attempt_back_merge(q, rq, bio);
785 case ELEVATOR_FRONT_MERGE:
786 if (blk_mq_sched_allow_merge(q, rq, bio))
787 merged = bio_attempt_front_merge(q, rq, bio);
789 case ELEVATOR_DISCARD_MERGE:
790 merged = bio_attempt_discard_merge(q, rq, bio);
804 struct flush_busy_ctx_data {
805 struct blk_mq_hw_ctx *hctx;
806 struct list_head *list;
809 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
811 struct flush_busy_ctx_data *flush_data = data;
812 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
813 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
815 sbitmap_clear_bit(sb, bitnr);
816 spin_lock(&ctx->lock);
817 list_splice_tail_init(&ctx->rq_list, flush_data->list);
818 spin_unlock(&ctx->lock);
823 * Process software queues that have been marked busy, splicing them
824 * to the for-dispatch
826 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
828 struct flush_busy_ctx_data data = {
833 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
835 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
837 static inline unsigned int queued_to_index(unsigned int queued)
842 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
845 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
848 struct blk_mq_alloc_data data = {
850 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
851 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
861 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
862 data.flags |= BLK_MQ_REQ_RESERVED;
864 rq->tag = blk_mq_get_tag(&data);
866 if (blk_mq_tag_busy(data.hctx)) {
867 rq->rq_flags |= RQF_MQ_INFLIGHT;
868 atomic_inc(&data.hctx->nr_active);
870 data.hctx->tags->rqs[rq->tag] = rq;
877 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
880 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
883 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
884 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
885 atomic_dec(&hctx->nr_active);
889 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
892 if (rq->tag == -1 || rq->internal_tag == -1)
895 __blk_mq_put_driver_tag(hctx, rq);
898 static void blk_mq_put_driver_tag(struct request *rq)
900 struct blk_mq_hw_ctx *hctx;
902 if (rq->tag == -1 || rq->internal_tag == -1)
905 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
906 __blk_mq_put_driver_tag(hctx, rq);
910 * If we fail getting a driver tag because all the driver tags are already
911 * assigned and on the dispatch list, BUT the first entry does not have a
912 * tag, then we could deadlock. For that case, move entries with assigned
913 * driver tags to the front, leaving the set of tagged requests in the
914 * same order, and the untagged set in the same order.
916 static bool reorder_tags_to_front(struct list_head *list)
918 struct request *rq, *tmp, *first = NULL;
920 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
924 list_move(&rq->queuelist, list);
930 return first != NULL;
933 static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags,
936 struct blk_mq_hw_ctx *hctx;
938 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
940 list_del(&wait->task_list);
941 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
942 blk_mq_run_hw_queue(hctx, true);
946 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
948 struct sbq_wait_state *ws;
951 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
952 * The thread which wins the race to grab this bit adds the hardware
953 * queue to the wait queue.
955 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
956 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
959 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
960 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
963 * As soon as this returns, it's no longer safe to fiddle with
964 * hctx->dispatch_wait, since a completion can wake up the wait queue
965 * and unlock the bit.
967 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
971 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list)
973 struct request_queue *q = hctx->queue;
975 LIST_HEAD(driver_list);
976 struct list_head *dptr;
977 int queued, ret = BLK_MQ_RQ_QUEUE_OK;
980 * Start off with dptr being NULL, so we start the first request
981 * immediately, even if we have more pending.
986 * Now process all the entries, sending them to the driver.
989 while (!list_empty(list)) {
990 struct blk_mq_queue_data bd;
992 rq = list_first_entry(list, struct request, queuelist);
993 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
994 if (!queued && reorder_tags_to_front(list))
998 * The initial allocation attempt failed, so we need to
999 * rerun the hardware queue when a tag is freed.
1001 if (blk_mq_dispatch_wait_add(hctx)) {
1003 * It's possible that a tag was freed in the
1004 * window between the allocation failure and
1005 * adding the hardware queue to the wait queue.
1007 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1014 list_del_init(&rq->queuelist);
1020 * Flag last if we have no more requests, or if we have more
1021 * but can't assign a driver tag to it.
1023 if (list_empty(list))
1026 struct request *nxt;
1028 nxt = list_first_entry(list, struct request, queuelist);
1029 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1032 ret = q->mq_ops->queue_rq(hctx, &bd);
1034 case BLK_MQ_RQ_QUEUE_OK:
1037 case BLK_MQ_RQ_QUEUE_BUSY:
1038 blk_mq_put_driver_tag_hctx(hctx, rq);
1039 list_add(&rq->queuelist, list);
1040 __blk_mq_requeue_request(rq);
1043 pr_err("blk-mq: bad return on queue: %d\n", ret);
1044 case BLK_MQ_RQ_QUEUE_ERROR:
1046 blk_mq_end_request(rq, rq->errors);
1050 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
1054 * We've done the first request. If we have more than 1
1055 * left in the list, set dptr to defer issue.
1057 if (!dptr && list->next != list->prev)
1058 dptr = &driver_list;
1061 hctx->dispatched[queued_to_index(queued)]++;
1064 * Any items that need requeuing? Stuff them into hctx->dispatch,
1065 * that is where we will continue on next queue run.
1067 if (!list_empty(list)) {
1069 * If we got a driver tag for the next request already,
1072 rq = list_first_entry(list, struct request, queuelist);
1073 blk_mq_put_driver_tag(rq);
1075 spin_lock(&hctx->lock);
1076 list_splice_init(list, &hctx->dispatch);
1077 spin_unlock(&hctx->lock);
1080 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
1081 * it's possible the queue is stopped and restarted again
1082 * before this. Queue restart will dispatch requests. And since
1083 * requests in rq_list aren't added into hctx->dispatch yet,
1084 * the requests in rq_list might get lost.
1086 * blk_mq_run_hw_queue() already checks the STOPPED bit
1088 * If RESTART or TAG_WAITING is set, then let completion restart
1089 * the queue instead of potentially looping here.
1091 if (!blk_mq_sched_needs_restart(hctx) &&
1092 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1093 blk_mq_run_hw_queue(hctx, true);
1099 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1103 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1104 cpu_online(hctx->next_cpu));
1106 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1108 blk_mq_sched_dispatch_requests(hctx);
1111 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1112 blk_mq_sched_dispatch_requests(hctx);
1113 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1118 * It'd be great if the workqueue API had a way to pass
1119 * in a mask and had some smarts for more clever placement.
1120 * For now we just round-robin here, switching for every
1121 * BLK_MQ_CPU_WORK_BATCH queued items.
1123 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1125 if (hctx->queue->nr_hw_queues == 1)
1126 return WORK_CPU_UNBOUND;
1128 if (--hctx->next_cpu_batch <= 0) {
1131 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1132 if (next_cpu >= nr_cpu_ids)
1133 next_cpu = cpumask_first(hctx->cpumask);
1135 hctx->next_cpu = next_cpu;
1136 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1139 return hctx->next_cpu;
1142 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1144 if (unlikely(blk_mq_hctx_stopped(hctx) ||
1145 !blk_mq_hw_queue_mapped(hctx)))
1148 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1149 int cpu = get_cpu();
1150 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1151 __blk_mq_run_hw_queue(hctx);
1159 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work);
1162 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1164 struct blk_mq_hw_ctx *hctx;
1167 queue_for_each_hw_ctx(q, hctx, i) {
1168 if (!blk_mq_hctx_has_pending(hctx) ||
1169 blk_mq_hctx_stopped(hctx))
1172 blk_mq_run_hw_queue(hctx, async);
1175 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1178 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1179 * @q: request queue.
1181 * The caller is responsible for serializing this function against
1182 * blk_mq_{start,stop}_hw_queue().
1184 bool blk_mq_queue_stopped(struct request_queue *q)
1186 struct blk_mq_hw_ctx *hctx;
1189 queue_for_each_hw_ctx(q, hctx, i)
1190 if (blk_mq_hctx_stopped(hctx))
1195 EXPORT_SYMBOL(blk_mq_queue_stopped);
1197 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1199 cancel_work(&hctx->run_work);
1200 cancel_delayed_work(&hctx->delay_work);
1201 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1203 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1205 void blk_mq_stop_hw_queues(struct request_queue *q)
1207 struct blk_mq_hw_ctx *hctx;
1210 queue_for_each_hw_ctx(q, hctx, i)
1211 blk_mq_stop_hw_queue(hctx);
1213 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1215 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1217 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1219 blk_mq_run_hw_queue(hctx, false);
1221 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1223 void blk_mq_start_hw_queues(struct request_queue *q)
1225 struct blk_mq_hw_ctx *hctx;
1228 queue_for_each_hw_ctx(q, hctx, i)
1229 blk_mq_start_hw_queue(hctx);
1231 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1233 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1235 if (!blk_mq_hctx_stopped(hctx))
1238 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1239 blk_mq_run_hw_queue(hctx, async);
1241 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1243 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1245 struct blk_mq_hw_ctx *hctx;
1248 queue_for_each_hw_ctx(q, hctx, i)
1249 blk_mq_start_stopped_hw_queue(hctx, async);
1251 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1253 static void blk_mq_run_work_fn(struct work_struct *work)
1255 struct blk_mq_hw_ctx *hctx;
1257 hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
1259 __blk_mq_run_hw_queue(hctx);
1262 static void blk_mq_delay_work_fn(struct work_struct *work)
1264 struct blk_mq_hw_ctx *hctx;
1266 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1268 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1269 __blk_mq_run_hw_queue(hctx);
1272 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1274 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1277 blk_mq_stop_hw_queue(hctx);
1278 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1279 &hctx->delay_work, msecs_to_jiffies(msecs));
1281 EXPORT_SYMBOL(blk_mq_delay_queue);
1283 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1287 struct blk_mq_ctx *ctx = rq->mq_ctx;
1289 trace_block_rq_insert(hctx->queue, rq);
1292 list_add(&rq->queuelist, &ctx->rq_list);
1294 list_add_tail(&rq->queuelist, &ctx->rq_list);
1297 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1300 struct blk_mq_ctx *ctx = rq->mq_ctx;
1302 __blk_mq_insert_req_list(hctx, rq, at_head);
1303 blk_mq_hctx_mark_pending(hctx, ctx);
1306 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1307 struct list_head *list)
1311 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1314 spin_lock(&ctx->lock);
1315 while (!list_empty(list)) {
1318 rq = list_first_entry(list, struct request, queuelist);
1319 BUG_ON(rq->mq_ctx != ctx);
1320 list_del_init(&rq->queuelist);
1321 __blk_mq_insert_req_list(hctx, rq, false);
1323 blk_mq_hctx_mark_pending(hctx, ctx);
1324 spin_unlock(&ctx->lock);
1327 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1329 struct request *rqa = container_of(a, struct request, queuelist);
1330 struct request *rqb = container_of(b, struct request, queuelist);
1332 return !(rqa->mq_ctx < rqb->mq_ctx ||
1333 (rqa->mq_ctx == rqb->mq_ctx &&
1334 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1337 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1339 struct blk_mq_ctx *this_ctx;
1340 struct request_queue *this_q;
1343 LIST_HEAD(ctx_list);
1346 list_splice_init(&plug->mq_list, &list);
1348 list_sort(NULL, &list, plug_ctx_cmp);
1354 while (!list_empty(&list)) {
1355 rq = list_entry_rq(list.next);
1356 list_del_init(&rq->queuelist);
1358 if (rq->mq_ctx != this_ctx) {
1360 trace_block_unplug(this_q, depth, from_schedule);
1361 blk_mq_sched_insert_requests(this_q, this_ctx,
1366 this_ctx = rq->mq_ctx;
1372 list_add_tail(&rq->queuelist, &ctx_list);
1376 * If 'this_ctx' is set, we know we have entries to complete
1377 * on 'ctx_list'. Do those.
1380 trace_block_unplug(this_q, depth, from_schedule);
1381 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1386 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1388 init_request_from_bio(rq, bio);
1390 blk_account_io_start(rq, true);
1393 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1395 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1396 !blk_queue_nomerges(hctx->queue);
1399 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1400 struct blk_mq_ctx *ctx,
1401 struct request *rq, struct bio *bio)
1403 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1404 blk_mq_bio_to_request(rq, bio);
1405 spin_lock(&ctx->lock);
1407 __blk_mq_insert_request(hctx, rq, false);
1408 spin_unlock(&ctx->lock);
1411 struct request_queue *q = hctx->queue;
1413 spin_lock(&ctx->lock);
1414 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1415 blk_mq_bio_to_request(rq, bio);
1419 spin_unlock(&ctx->lock);
1420 __blk_mq_finish_request(hctx, ctx, rq);
1425 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1428 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1430 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1433 static void __blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie,
1436 struct request_queue *q = rq->q;
1437 struct blk_mq_queue_data bd = {
1442 struct blk_mq_hw_ctx *hctx;
1443 blk_qc_t new_cookie;
1449 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1452 new_cookie = request_to_qc_t(hctx, rq);
1455 * For OK queue, we are done. For error, kill it. Any other
1456 * error (busy), just add it to our list as we previously
1459 ret = q->mq_ops->queue_rq(hctx, &bd);
1460 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1461 *cookie = new_cookie;
1465 __blk_mq_requeue_request(rq);
1467 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1468 *cookie = BLK_QC_T_NONE;
1470 blk_mq_end_request(rq, rq->errors);
1475 blk_mq_sched_insert_request(rq, false, true, false, may_sleep);
1478 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1479 struct request *rq, blk_qc_t *cookie)
1481 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1483 __blk_mq_try_issue_directly(rq, cookie, false);
1486 unsigned int srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1487 __blk_mq_try_issue_directly(rq, cookie, true);
1488 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1492 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1494 const int is_sync = op_is_sync(bio->bi_opf);
1495 const int is_flush_fua = op_is_flush(bio->bi_opf);
1496 struct blk_mq_alloc_data data = { .flags = 0 };
1498 unsigned int request_count = 0;
1499 struct blk_plug *plug;
1500 struct request *same_queue_rq = NULL;
1502 unsigned int wb_acct;
1504 blk_queue_bounce(q, &bio);
1506 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1508 return BLK_QC_T_NONE;
1511 blk_queue_split(q, &bio, q->bio_split);
1513 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1514 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1515 return BLK_QC_T_NONE;
1517 if (blk_mq_sched_bio_merge(q, bio))
1518 return BLK_QC_T_NONE;
1520 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1522 trace_block_getrq(q, bio, bio->bi_opf);
1524 rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1525 if (unlikely(!rq)) {
1526 __wbt_done(q->rq_wb, wb_acct);
1527 return BLK_QC_T_NONE;
1530 wbt_track(&rq->issue_stat, wb_acct);
1532 cookie = request_to_qc_t(data.hctx, rq);
1534 plug = current->plug;
1535 if (unlikely(is_flush_fua)) {
1536 blk_mq_bio_to_request(rq, bio);
1538 blk_mq_sched_insert_request(rq, false, true, true,
1541 blk_insert_flush(rq);
1542 blk_mq_run_hw_queue(data.hctx, true);
1544 } else if (plug && q->nr_hw_queues == 1) {
1545 struct request *last = NULL;
1547 blk_mq_bio_to_request(rq, bio);
1550 * @request_count may become stale because of schedule
1551 * out, so check the list again.
1553 if (list_empty(&plug->mq_list))
1555 else if (blk_queue_nomerges(q))
1556 request_count = blk_plug_queued_count(q);
1559 trace_block_plug(q);
1561 last = list_entry_rq(plug->mq_list.prev);
1563 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1564 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1565 blk_flush_plug_list(plug, false);
1566 trace_block_plug(q);
1569 list_add_tail(&rq->queuelist, &plug->mq_list);
1570 } else if (plug && !blk_queue_nomerges(q)) {
1571 blk_mq_bio_to_request(rq, bio);
1574 * We do limited plugging. If the bio can be merged, do that.
1575 * Otherwise the existing request in the plug list will be
1576 * issued. So the plug list will have one request at most
1577 * The plug list might get flushed before this. If that happens,
1578 * the plug list is empty, and same_queue_rq is invalid.
1580 if (list_empty(&plug->mq_list))
1581 same_queue_rq = NULL;
1583 list_del_init(&same_queue_rq->queuelist);
1584 list_add_tail(&rq->queuelist, &plug->mq_list);
1587 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1589 } else if (q->nr_hw_queues > 1 && is_sync) {
1590 blk_mq_bio_to_request(rq, bio);
1591 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1592 } else if (q->elevator) {
1593 blk_mq_bio_to_request(rq, bio);
1594 blk_mq_sched_insert_request(rq, false, true, true, true);
1595 } else if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1596 blk_mq_run_hw_queue(data.hctx, true);
1599 blk_mq_put_ctx(data.ctx);
1603 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1604 unsigned int hctx_idx)
1608 if (tags->rqs && set->ops->exit_request) {
1611 for (i = 0; i < tags->nr_tags; i++) {
1612 struct request *rq = tags->static_rqs[i];
1616 set->ops->exit_request(set->driver_data, rq,
1618 tags->static_rqs[i] = NULL;
1622 while (!list_empty(&tags->page_list)) {
1623 page = list_first_entry(&tags->page_list, struct page, lru);
1624 list_del_init(&page->lru);
1626 * Remove kmemleak object previously allocated in
1627 * blk_mq_init_rq_map().
1629 kmemleak_free(page_address(page));
1630 __free_pages(page, page->private);
1634 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1638 kfree(tags->static_rqs);
1639 tags->static_rqs = NULL;
1641 blk_mq_free_tags(tags);
1644 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1645 unsigned int hctx_idx,
1646 unsigned int nr_tags,
1647 unsigned int reserved_tags)
1649 struct blk_mq_tags *tags;
1652 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1653 if (node == NUMA_NO_NODE)
1654 node = set->numa_node;
1656 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1657 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1661 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1662 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1665 blk_mq_free_tags(tags);
1669 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1670 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1672 if (!tags->static_rqs) {
1674 blk_mq_free_tags(tags);
1681 static size_t order_to_size(unsigned int order)
1683 return (size_t)PAGE_SIZE << order;
1686 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1687 unsigned int hctx_idx, unsigned int depth)
1689 unsigned int i, j, entries_per_page, max_order = 4;
1690 size_t rq_size, left;
1693 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1694 if (node == NUMA_NO_NODE)
1695 node = set->numa_node;
1697 INIT_LIST_HEAD(&tags->page_list);
1700 * rq_size is the size of the request plus driver payload, rounded
1701 * to the cacheline size
1703 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1705 left = rq_size * depth;
1707 for (i = 0; i < depth; ) {
1708 int this_order = max_order;
1713 while (this_order && left < order_to_size(this_order - 1))
1717 page = alloc_pages_node(node,
1718 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1724 if (order_to_size(this_order) < rq_size)
1731 page->private = this_order;
1732 list_add_tail(&page->lru, &tags->page_list);
1734 p = page_address(page);
1736 * Allow kmemleak to scan these pages as they contain pointers
1737 * to additional allocations like via ops->init_request().
1739 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1740 entries_per_page = order_to_size(this_order) / rq_size;
1741 to_do = min(entries_per_page, depth - i);
1742 left -= to_do * rq_size;
1743 for (j = 0; j < to_do; j++) {
1744 struct request *rq = p;
1746 tags->static_rqs[i] = rq;
1747 if (set->ops->init_request) {
1748 if (set->ops->init_request(set->driver_data,
1751 tags->static_rqs[i] = NULL;
1763 blk_mq_free_rqs(set, tags, hctx_idx);
1768 * 'cpu' is going away. splice any existing rq_list entries from this
1769 * software queue to the hw queue dispatch list, and ensure that it
1772 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1774 struct blk_mq_hw_ctx *hctx;
1775 struct blk_mq_ctx *ctx;
1778 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1779 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1781 spin_lock(&ctx->lock);
1782 if (!list_empty(&ctx->rq_list)) {
1783 list_splice_init(&ctx->rq_list, &tmp);
1784 blk_mq_hctx_clear_pending(hctx, ctx);
1786 spin_unlock(&ctx->lock);
1788 if (list_empty(&tmp))
1791 spin_lock(&hctx->lock);
1792 list_splice_tail_init(&tmp, &hctx->dispatch);
1793 spin_unlock(&hctx->lock);
1795 blk_mq_run_hw_queue(hctx, true);
1799 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1801 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1805 /* hctx->ctxs will be freed in queue's release handler */
1806 static void blk_mq_exit_hctx(struct request_queue *q,
1807 struct blk_mq_tag_set *set,
1808 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1810 unsigned flush_start_tag = set->queue_depth;
1812 blk_mq_tag_idle(hctx);
1814 if (set->ops->exit_request)
1815 set->ops->exit_request(set->driver_data,
1816 hctx->fq->flush_rq, hctx_idx,
1817 flush_start_tag + hctx_idx);
1819 if (set->ops->exit_hctx)
1820 set->ops->exit_hctx(hctx, hctx_idx);
1822 if (hctx->flags & BLK_MQ_F_BLOCKING)
1823 cleanup_srcu_struct(&hctx->queue_rq_srcu);
1825 blk_mq_remove_cpuhp(hctx);
1826 blk_free_flush_queue(hctx->fq);
1827 sbitmap_free(&hctx->ctx_map);
1830 static void blk_mq_exit_hw_queues(struct request_queue *q,
1831 struct blk_mq_tag_set *set, int nr_queue)
1833 struct blk_mq_hw_ctx *hctx;
1836 queue_for_each_hw_ctx(q, hctx, i) {
1839 blk_mq_exit_hctx(q, set, hctx, i);
1843 static int blk_mq_init_hctx(struct request_queue *q,
1844 struct blk_mq_tag_set *set,
1845 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1848 unsigned flush_start_tag = set->queue_depth;
1850 node = hctx->numa_node;
1851 if (node == NUMA_NO_NODE)
1852 node = hctx->numa_node = set->numa_node;
1854 INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1855 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1856 spin_lock_init(&hctx->lock);
1857 INIT_LIST_HEAD(&hctx->dispatch);
1859 hctx->queue_num = hctx_idx;
1860 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1862 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1864 hctx->tags = set->tags[hctx_idx];
1867 * Allocate space for all possible cpus to avoid allocation at
1870 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1873 goto unregister_cpu_notifier;
1875 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1881 if (set->ops->init_hctx &&
1882 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1885 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1889 if (set->ops->init_request &&
1890 set->ops->init_request(set->driver_data,
1891 hctx->fq->flush_rq, hctx_idx,
1892 flush_start_tag + hctx_idx, node))
1895 if (hctx->flags & BLK_MQ_F_BLOCKING)
1896 init_srcu_struct(&hctx->queue_rq_srcu);
1903 if (set->ops->exit_hctx)
1904 set->ops->exit_hctx(hctx, hctx_idx);
1906 sbitmap_free(&hctx->ctx_map);
1909 unregister_cpu_notifier:
1910 blk_mq_remove_cpuhp(hctx);
1914 static void blk_mq_init_cpu_queues(struct request_queue *q,
1915 unsigned int nr_hw_queues)
1919 for_each_possible_cpu(i) {
1920 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1921 struct blk_mq_hw_ctx *hctx;
1924 spin_lock_init(&__ctx->lock);
1925 INIT_LIST_HEAD(&__ctx->rq_list);
1928 /* If the cpu isn't online, the cpu is mapped to first hctx */
1932 hctx = blk_mq_map_queue(q, i);
1935 * Set local node, IFF we have more than one hw queue. If
1936 * not, we remain on the home node of the device
1938 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1939 hctx->numa_node = local_memory_node(cpu_to_node(i));
1943 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
1947 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
1948 set->queue_depth, set->reserved_tags);
1949 if (!set->tags[hctx_idx])
1952 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
1957 blk_mq_free_rq_map(set->tags[hctx_idx]);
1958 set->tags[hctx_idx] = NULL;
1962 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
1963 unsigned int hctx_idx)
1965 if (set->tags[hctx_idx]) {
1966 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
1967 blk_mq_free_rq_map(set->tags[hctx_idx]);
1968 set->tags[hctx_idx] = NULL;
1972 static void blk_mq_map_swqueue(struct request_queue *q,
1973 const struct cpumask *online_mask)
1975 unsigned int i, hctx_idx;
1976 struct blk_mq_hw_ctx *hctx;
1977 struct blk_mq_ctx *ctx;
1978 struct blk_mq_tag_set *set = q->tag_set;
1981 * Avoid others reading imcomplete hctx->cpumask through sysfs
1983 mutex_lock(&q->sysfs_lock);
1985 queue_for_each_hw_ctx(q, hctx, i) {
1986 cpumask_clear(hctx->cpumask);
1991 * Map software to hardware queues
1993 for_each_possible_cpu(i) {
1994 /* If the cpu isn't online, the cpu is mapped to first hctx */
1995 if (!cpumask_test_cpu(i, online_mask))
1998 hctx_idx = q->mq_map[i];
1999 /* unmapped hw queue can be remapped after CPU topo changed */
2000 if (!set->tags[hctx_idx] &&
2001 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2003 * If tags initialization fail for some hctx,
2004 * that hctx won't be brought online. In this
2005 * case, remap the current ctx to hctx[0] which
2006 * is guaranteed to always have tags allocated
2011 ctx = per_cpu_ptr(q->queue_ctx, i);
2012 hctx = blk_mq_map_queue(q, i);
2014 cpumask_set_cpu(i, hctx->cpumask);
2015 ctx->index_hw = hctx->nr_ctx;
2016 hctx->ctxs[hctx->nr_ctx++] = ctx;
2019 mutex_unlock(&q->sysfs_lock);
2021 queue_for_each_hw_ctx(q, hctx, i) {
2023 * If no software queues are mapped to this hardware queue,
2024 * disable it and free the request entries.
2026 if (!hctx->nr_ctx) {
2027 /* Never unmap queue 0. We need it as a
2028 * fallback in case of a new remap fails
2031 if (i && set->tags[i])
2032 blk_mq_free_map_and_requests(set, i);
2038 hctx->tags = set->tags[i];
2039 WARN_ON(!hctx->tags);
2042 * Set the map size to the number of mapped software queues.
2043 * This is more accurate and more efficient than looping
2044 * over all possibly mapped software queues.
2046 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2049 * Initialize batch roundrobin counts
2051 hctx->next_cpu = cpumask_first(hctx->cpumask);
2052 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2056 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2058 struct blk_mq_hw_ctx *hctx;
2061 queue_for_each_hw_ctx(q, hctx, i) {
2063 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2065 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2069 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2071 struct request_queue *q;
2073 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2074 blk_mq_freeze_queue(q);
2075 queue_set_hctx_shared(q, shared);
2076 blk_mq_unfreeze_queue(q);
2080 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2082 struct blk_mq_tag_set *set = q->tag_set;
2084 mutex_lock(&set->tag_list_lock);
2085 list_del_init(&q->tag_set_list);
2086 if (list_is_singular(&set->tag_list)) {
2087 /* just transitioned to unshared */
2088 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2089 /* update existing queue */
2090 blk_mq_update_tag_set_depth(set, false);
2092 mutex_unlock(&set->tag_list_lock);
2095 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2096 struct request_queue *q)
2100 mutex_lock(&set->tag_list_lock);
2102 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2103 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2104 set->flags |= BLK_MQ_F_TAG_SHARED;
2105 /* update existing queue */
2106 blk_mq_update_tag_set_depth(set, true);
2108 if (set->flags & BLK_MQ_F_TAG_SHARED)
2109 queue_set_hctx_shared(q, true);
2110 list_add_tail(&q->tag_set_list, &set->tag_list);
2112 mutex_unlock(&set->tag_list_lock);
2116 * It is the actual release handler for mq, but we do it from
2117 * request queue's release handler for avoiding use-after-free
2118 * and headache because q->mq_kobj shouldn't have been introduced,
2119 * but we can't group ctx/kctx kobj without it.
2121 void blk_mq_release(struct request_queue *q)
2123 struct blk_mq_hw_ctx *hctx;
2126 blk_mq_sched_teardown(q);
2128 /* hctx kobj stays in hctx */
2129 queue_for_each_hw_ctx(q, hctx, i) {
2132 kobject_put(&hctx->kobj);
2137 kfree(q->queue_hw_ctx);
2140 * release .mq_kobj and sw queue's kobject now because
2141 * both share lifetime with request queue.
2143 blk_mq_sysfs_deinit(q);
2145 free_percpu(q->queue_ctx);
2148 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2150 struct request_queue *uninit_q, *q;
2152 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2154 return ERR_PTR(-ENOMEM);
2156 q = blk_mq_init_allocated_queue(set, uninit_q);
2158 blk_cleanup_queue(uninit_q);
2162 EXPORT_SYMBOL(blk_mq_init_queue);
2164 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2165 struct request_queue *q)
2168 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2170 blk_mq_sysfs_unregister(q);
2171 for (i = 0; i < set->nr_hw_queues; i++) {
2177 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2178 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2183 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2190 atomic_set(&hctxs[i]->nr_active, 0);
2191 hctxs[i]->numa_node = node;
2192 hctxs[i]->queue_num = i;
2194 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2195 free_cpumask_var(hctxs[i]->cpumask);
2200 blk_mq_hctx_kobj_init(hctxs[i]);
2202 for (j = i; j < q->nr_hw_queues; j++) {
2203 struct blk_mq_hw_ctx *hctx = hctxs[j];
2207 blk_mq_free_map_and_requests(set, j);
2208 blk_mq_exit_hctx(q, set, hctx, j);
2209 kobject_put(&hctx->kobj);
2214 q->nr_hw_queues = i;
2215 blk_mq_sysfs_register(q);
2218 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2219 struct request_queue *q)
2221 /* mark the queue as mq asap */
2222 q->mq_ops = set->ops;
2224 q->stats = blk_alloc_queue_stats();
2228 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2229 blk_stat_rq_ddir, 2, q);
2233 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2237 /* init q->mq_kobj and sw queues' kobjects */
2238 blk_mq_sysfs_init(q);
2240 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2241 GFP_KERNEL, set->numa_node);
2242 if (!q->queue_hw_ctx)
2245 q->mq_map = set->mq_map;
2247 blk_mq_realloc_hw_ctxs(set, q);
2248 if (!q->nr_hw_queues)
2251 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2252 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2254 q->nr_queues = nr_cpu_ids;
2256 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2258 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2259 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2261 q->sg_reserved_size = INT_MAX;
2263 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2264 INIT_LIST_HEAD(&q->requeue_list);
2265 spin_lock_init(&q->requeue_lock);
2267 blk_queue_make_request(q, blk_mq_make_request);
2270 * Do this after blk_queue_make_request() overrides it...
2272 q->nr_requests = set->queue_depth;
2275 * Default to classic polling
2279 if (set->ops->complete)
2280 blk_queue_softirq_done(q, set->ops->complete);
2282 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2285 mutex_lock(&all_q_mutex);
2287 list_add_tail(&q->all_q_node, &all_q_list);
2288 blk_mq_add_queue_tag_set(set, q);
2289 blk_mq_map_swqueue(q, cpu_online_mask);
2291 mutex_unlock(&all_q_mutex);
2294 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2297 ret = blk_mq_sched_init(q);
2299 return ERR_PTR(ret);
2305 kfree(q->queue_hw_ctx);
2307 free_percpu(q->queue_ctx);
2310 return ERR_PTR(-ENOMEM);
2312 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2314 void blk_mq_free_queue(struct request_queue *q)
2316 struct blk_mq_tag_set *set = q->tag_set;
2318 mutex_lock(&all_q_mutex);
2319 list_del_init(&q->all_q_node);
2320 mutex_unlock(&all_q_mutex);
2322 blk_mq_del_queue_tag_set(q);
2324 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2327 /* Basically redo blk_mq_init_queue with queue frozen */
2328 static void blk_mq_queue_reinit(struct request_queue *q,
2329 const struct cpumask *online_mask)
2331 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2333 blk_mq_sysfs_unregister(q);
2336 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2337 * we should change hctx numa_node according to new topology (this
2338 * involves free and re-allocate memory, worthy doing?)
2341 blk_mq_map_swqueue(q, online_mask);
2343 blk_mq_sysfs_register(q);
2347 * New online cpumask which is going to be set in this hotplug event.
2348 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2349 * one-by-one and dynamically allocating this could result in a failure.
2351 static struct cpumask cpuhp_online_new;
2353 static void blk_mq_queue_reinit_work(void)
2355 struct request_queue *q;
2357 mutex_lock(&all_q_mutex);
2359 * We need to freeze and reinit all existing queues. Freezing
2360 * involves synchronous wait for an RCU grace period and doing it
2361 * one by one may take a long time. Start freezing all queues in
2362 * one swoop and then wait for the completions so that freezing can
2363 * take place in parallel.
2365 list_for_each_entry(q, &all_q_list, all_q_node)
2366 blk_mq_freeze_queue_start(q);
2367 list_for_each_entry(q, &all_q_list, all_q_node)
2368 blk_mq_freeze_queue_wait(q);
2370 list_for_each_entry(q, &all_q_list, all_q_node)
2371 blk_mq_queue_reinit(q, &cpuhp_online_new);
2373 list_for_each_entry(q, &all_q_list, all_q_node)
2374 blk_mq_unfreeze_queue(q);
2376 mutex_unlock(&all_q_mutex);
2379 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2381 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2382 blk_mq_queue_reinit_work();
2387 * Before hotadded cpu starts handling requests, new mappings must be
2388 * established. Otherwise, these requests in hw queue might never be
2391 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2392 * for CPU0, and ctx1 for CPU1).
2394 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2395 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2397 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2398 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2399 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2402 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2404 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2405 cpumask_set_cpu(cpu, &cpuhp_online_new);
2406 blk_mq_queue_reinit_work();
2410 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2414 for (i = 0; i < set->nr_hw_queues; i++)
2415 if (!__blk_mq_alloc_rq_map(set, i))
2422 blk_mq_free_rq_map(set->tags[i]);
2428 * Allocate the request maps associated with this tag_set. Note that this
2429 * may reduce the depth asked for, if memory is tight. set->queue_depth
2430 * will be updated to reflect the allocated depth.
2432 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2437 depth = set->queue_depth;
2439 err = __blk_mq_alloc_rq_maps(set);
2443 set->queue_depth >>= 1;
2444 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2448 } while (set->queue_depth);
2450 if (!set->queue_depth || err) {
2451 pr_err("blk-mq: failed to allocate request map\n");
2455 if (depth != set->queue_depth)
2456 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2457 depth, set->queue_depth);
2463 * Alloc a tag set to be associated with one or more request queues.
2464 * May fail with EINVAL for various error conditions. May adjust the
2465 * requested depth down, if if it too large. In that case, the set
2466 * value will be stored in set->queue_depth.
2468 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2472 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2474 if (!set->nr_hw_queues)
2476 if (!set->queue_depth)
2478 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2481 if (!set->ops->queue_rq)
2484 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2485 pr_info("blk-mq: reduced tag depth to %u\n",
2487 set->queue_depth = BLK_MQ_MAX_DEPTH;
2491 * If a crashdump is active, then we are potentially in a very
2492 * memory constrained environment. Limit us to 1 queue and
2493 * 64 tags to prevent using too much memory.
2495 if (is_kdump_kernel()) {
2496 set->nr_hw_queues = 1;
2497 set->queue_depth = min(64U, set->queue_depth);
2500 * There is no use for more h/w queues than cpus.
2502 if (set->nr_hw_queues > nr_cpu_ids)
2503 set->nr_hw_queues = nr_cpu_ids;
2505 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2506 GFP_KERNEL, set->numa_node);
2511 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2512 GFP_KERNEL, set->numa_node);
2516 if (set->ops->map_queues)
2517 ret = set->ops->map_queues(set);
2519 ret = blk_mq_map_queues(set);
2521 goto out_free_mq_map;
2523 ret = blk_mq_alloc_rq_maps(set);
2525 goto out_free_mq_map;
2527 mutex_init(&set->tag_list_lock);
2528 INIT_LIST_HEAD(&set->tag_list);
2540 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2542 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2546 for (i = 0; i < nr_cpu_ids; i++)
2547 blk_mq_free_map_and_requests(set, i);
2555 EXPORT_SYMBOL(blk_mq_free_tag_set);
2557 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2559 struct blk_mq_tag_set *set = q->tag_set;
2560 struct blk_mq_hw_ctx *hctx;
2566 blk_mq_freeze_queue(q);
2567 blk_mq_quiesce_queue(q);
2570 queue_for_each_hw_ctx(q, hctx, i) {
2574 * If we're using an MQ scheduler, just update the scheduler
2575 * queue depth. This is similar to what the old code would do.
2577 if (!hctx->sched_tags) {
2578 ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2579 min(nr, set->queue_depth),
2582 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2590 q->nr_requests = nr;
2592 blk_mq_unfreeze_queue(q);
2593 blk_mq_start_stopped_hw_queues(q, true);
2598 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2600 struct request_queue *q;
2602 if (nr_hw_queues > nr_cpu_ids)
2603 nr_hw_queues = nr_cpu_ids;
2604 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2607 list_for_each_entry(q, &set->tag_list, tag_set_list)
2608 blk_mq_freeze_queue(q);
2610 set->nr_hw_queues = nr_hw_queues;
2611 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2612 blk_mq_realloc_hw_ctxs(set, q);
2613 blk_mq_queue_reinit(q, cpu_online_mask);
2616 list_for_each_entry(q, &set->tag_list, tag_set_list)
2617 blk_mq_unfreeze_queue(q);
2619 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2621 /* Enable polling stats and return whether they were already enabled. */
2622 static bool blk_poll_stats_enable(struct request_queue *q)
2624 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2625 test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2627 blk_stat_add_callback(q, q->poll_cb);
2631 static void blk_mq_poll_stats_start(struct request_queue *q)
2634 * We don't arm the callback if polling stats are not enabled or the
2635 * callback is already active.
2637 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2638 blk_stat_is_active(q->poll_cb))
2641 blk_stat_activate_msecs(q->poll_cb, 100);
2644 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2646 struct request_queue *q = cb->data;
2648 if (cb->stat[READ].nr_samples)
2649 q->poll_stat[READ] = cb->stat[READ];
2650 if (cb->stat[WRITE].nr_samples)
2651 q->poll_stat[WRITE] = cb->stat[WRITE];
2654 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2655 struct blk_mq_hw_ctx *hctx,
2658 unsigned long ret = 0;
2661 * If stats collection isn't on, don't sleep but turn it on for
2664 if (!blk_poll_stats_enable(q))
2668 * As an optimistic guess, use half of the mean service time
2669 * for this type of request. We can (and should) make this smarter.
2670 * For instance, if the completion latencies are tight, we can
2671 * get closer than just half the mean. This is especially
2672 * important on devices where the completion latencies are longer
2675 if (req_op(rq) == REQ_OP_READ && q->poll_stat[READ].nr_samples)
2676 ret = (q->poll_stat[READ].mean + 1) / 2;
2677 else if (req_op(rq) == REQ_OP_WRITE && q->poll_stat[WRITE].nr_samples)
2678 ret = (q->poll_stat[WRITE].mean + 1) / 2;
2683 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2684 struct blk_mq_hw_ctx *hctx,
2687 struct hrtimer_sleeper hs;
2688 enum hrtimer_mode mode;
2692 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2698 * -1: don't ever hybrid sleep
2699 * 0: use half of prev avg
2700 * >0: use this specific value
2702 if (q->poll_nsec == -1)
2704 else if (q->poll_nsec > 0)
2705 nsecs = q->poll_nsec;
2707 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2712 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2715 * This will be replaced with the stats tracking code, using
2716 * 'avg_completion_time / 2' as the pre-sleep target.
2720 mode = HRTIMER_MODE_REL;
2721 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2722 hrtimer_set_expires(&hs.timer, kt);
2724 hrtimer_init_sleeper(&hs, current);
2726 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2728 set_current_state(TASK_UNINTERRUPTIBLE);
2729 hrtimer_start_expires(&hs.timer, mode);
2732 hrtimer_cancel(&hs.timer);
2733 mode = HRTIMER_MODE_ABS;
2734 } while (hs.task && !signal_pending(current));
2736 __set_current_state(TASK_RUNNING);
2737 destroy_hrtimer_on_stack(&hs.timer);
2741 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2743 struct request_queue *q = hctx->queue;
2747 * If we sleep, have the caller restart the poll loop to reset
2748 * the state. Like for the other success return cases, the
2749 * caller is responsible for checking if the IO completed. If
2750 * the IO isn't complete, we'll get called again and will go
2751 * straight to the busy poll loop.
2753 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2756 hctx->poll_considered++;
2758 state = current->state;
2759 while (!need_resched()) {
2762 hctx->poll_invoked++;
2764 ret = q->mq_ops->poll(hctx, rq->tag);
2766 hctx->poll_success++;
2767 set_current_state(TASK_RUNNING);
2771 if (signal_pending_state(state, current))
2772 set_current_state(TASK_RUNNING);
2774 if (current->state == TASK_RUNNING)
2784 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2786 struct blk_mq_hw_ctx *hctx;
2787 struct blk_plug *plug;
2790 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2791 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2794 plug = current->plug;
2796 blk_flush_plug_list(plug, false);
2798 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2799 if (!blk_qc_t_is_internal(cookie))
2800 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2802 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2804 return __blk_mq_poll(hctx, rq);
2806 EXPORT_SYMBOL_GPL(blk_mq_poll);
2808 void blk_mq_disable_hotplug(void)
2810 mutex_lock(&all_q_mutex);
2813 void blk_mq_enable_hotplug(void)
2815 mutex_unlock(&all_q_mutex);
2818 static int __init blk_mq_init(void)
2820 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2821 blk_mq_hctx_notify_dead);
2823 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2824 blk_mq_queue_reinit_prepare,
2825 blk_mq_queue_reinit_dead);
2828 subsys_initcall(blk_mq_init);