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_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_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_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);
330 return ERR_PTR(-EWOULDBLOCK);
334 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
336 void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
339 const int sched_tag = rq->internal_tag;
340 struct request_queue *q = rq->q;
342 if (rq->rq_flags & RQF_MQ_INFLIGHT)
343 atomic_dec(&hctx->nr_active);
345 wbt_done(q->rq_wb, &rq->issue_stat);
348 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
349 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
351 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
353 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
354 blk_mq_sched_restart(hctx);
358 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx,
361 struct blk_mq_ctx *ctx = rq->mq_ctx;
363 ctx->rq_completed[rq_is_sync(rq)]++;
364 __blk_mq_finish_request(hctx, ctx, rq);
367 void blk_mq_finish_request(struct request *rq)
369 blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq);
371 EXPORT_SYMBOL_GPL(blk_mq_finish_request);
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;
447 if (rq->internal_tag != -1)
448 blk_mq_sched_completed_request(rq);
452 if (!q->softirq_done_fn)
453 blk_mq_end_request(rq, rq->errors);
455 blk_mq_ipi_complete_request(rq);
459 * blk_mq_complete_request - end I/O on a request
460 * @rq: the request being processed
463 * Ends all I/O on a request. It does not handle partial completions.
464 * The actual completion happens out-of-order, through a IPI handler.
466 void blk_mq_complete_request(struct request *rq, int error)
468 struct request_queue *q = rq->q;
470 if (unlikely(blk_should_fake_timeout(q)))
472 if (!blk_mark_rq_complete(rq)) {
474 __blk_mq_complete_request(rq);
477 EXPORT_SYMBOL(blk_mq_complete_request);
479 int blk_mq_request_started(struct request *rq)
481 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
483 EXPORT_SYMBOL_GPL(blk_mq_request_started);
485 void blk_mq_start_request(struct request *rq)
487 struct request_queue *q = rq->q;
489 blk_mq_sched_started_request(rq);
491 trace_block_rq_issue(q, rq);
493 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
494 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
495 rq->rq_flags |= RQF_STATS;
496 wbt_issue(q->rq_wb, &rq->issue_stat);
502 * Ensure that ->deadline is visible before set the started
503 * flag and clear the completed flag.
505 smp_mb__before_atomic();
508 * Mark us as started and clear complete. Complete might have been
509 * set if requeue raced with timeout, which then marked it as
510 * complete. So be sure to clear complete again when we start
511 * the request, otherwise we'll ignore the completion event.
513 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
514 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
515 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
516 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
518 if (q->dma_drain_size && blk_rq_bytes(rq)) {
520 * Make sure space for the drain appears. We know we can do
521 * this because max_hw_segments has been adjusted to be one
522 * fewer than the device can handle.
524 rq->nr_phys_segments++;
527 EXPORT_SYMBOL(blk_mq_start_request);
530 * When we reach here because queue is busy, REQ_ATOM_COMPLETE
531 * flag isn't set yet, so there may be race with timeout handler,
532 * but given rq->deadline is just set in .queue_rq() under
533 * this situation, the race won't be possible in reality because
534 * rq->timeout should be set as big enough to cover the window
535 * between blk_mq_start_request() called from .queue_rq() and
536 * clearing REQ_ATOM_STARTED here.
538 static void __blk_mq_requeue_request(struct request *rq)
540 struct request_queue *q = rq->q;
542 trace_block_rq_requeue(q, rq);
543 wbt_requeue(q->rq_wb, &rq->issue_stat);
544 blk_mq_sched_requeue_request(rq);
546 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
547 if (q->dma_drain_size && blk_rq_bytes(rq))
548 rq->nr_phys_segments--;
552 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
554 __blk_mq_requeue_request(rq);
556 BUG_ON(blk_queued_rq(rq));
557 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
559 EXPORT_SYMBOL(blk_mq_requeue_request);
561 static void blk_mq_requeue_work(struct work_struct *work)
563 struct request_queue *q =
564 container_of(work, struct request_queue, requeue_work.work);
566 struct request *rq, *next;
569 spin_lock_irqsave(&q->requeue_lock, flags);
570 list_splice_init(&q->requeue_list, &rq_list);
571 spin_unlock_irqrestore(&q->requeue_lock, flags);
573 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
574 if (!(rq->rq_flags & RQF_SOFTBARRIER))
577 rq->rq_flags &= ~RQF_SOFTBARRIER;
578 list_del_init(&rq->queuelist);
579 blk_mq_sched_insert_request(rq, true, false, false, true);
582 while (!list_empty(&rq_list)) {
583 rq = list_entry(rq_list.next, struct request, queuelist);
584 list_del_init(&rq->queuelist);
585 blk_mq_sched_insert_request(rq, false, false, false, true);
588 blk_mq_run_hw_queues(q, false);
591 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
592 bool kick_requeue_list)
594 struct request_queue *q = rq->q;
598 * We abuse this flag that is otherwise used by the I/O scheduler to
599 * request head insertation from the workqueue.
601 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
603 spin_lock_irqsave(&q->requeue_lock, flags);
605 rq->rq_flags |= RQF_SOFTBARRIER;
606 list_add(&rq->queuelist, &q->requeue_list);
608 list_add_tail(&rq->queuelist, &q->requeue_list);
610 spin_unlock_irqrestore(&q->requeue_lock, flags);
612 if (kick_requeue_list)
613 blk_mq_kick_requeue_list(q);
615 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
617 void blk_mq_kick_requeue_list(struct request_queue *q)
619 kblockd_schedule_delayed_work(&q->requeue_work, 0);
621 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
623 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
626 kblockd_schedule_delayed_work(&q->requeue_work,
627 msecs_to_jiffies(msecs));
629 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
631 void blk_mq_abort_requeue_list(struct request_queue *q)
636 spin_lock_irqsave(&q->requeue_lock, flags);
637 list_splice_init(&q->requeue_list, &rq_list);
638 spin_unlock_irqrestore(&q->requeue_lock, flags);
640 while (!list_empty(&rq_list)) {
643 rq = list_first_entry(&rq_list, struct request, queuelist);
644 list_del_init(&rq->queuelist);
646 blk_mq_end_request(rq, rq->errors);
649 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
651 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
653 if (tag < tags->nr_tags) {
654 prefetch(tags->rqs[tag]);
655 return tags->rqs[tag];
660 EXPORT_SYMBOL(blk_mq_tag_to_rq);
662 struct blk_mq_timeout_data {
664 unsigned int next_set;
667 void blk_mq_rq_timed_out(struct request *req, bool reserved)
669 const struct blk_mq_ops *ops = req->q->mq_ops;
670 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
673 * We know that complete is set at this point. If STARTED isn't set
674 * anymore, then the request isn't active and the "timeout" should
675 * just be ignored. This can happen due to the bitflag ordering.
676 * Timeout first checks if STARTED is set, and if it is, assumes
677 * the request is active. But if we race with completion, then
678 * both flags will get cleared. So check here again, and ignore
679 * a timeout event with a request that isn't active.
681 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
685 ret = ops->timeout(req, reserved);
689 __blk_mq_complete_request(req);
691 case BLK_EH_RESET_TIMER:
693 blk_clear_rq_complete(req);
695 case BLK_EH_NOT_HANDLED:
698 printk(KERN_ERR "block: bad eh return: %d\n", ret);
703 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
704 struct request *rq, void *priv, bool reserved)
706 struct blk_mq_timeout_data *data = priv;
708 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
712 * The rq being checked may have been freed and reallocated
713 * out already here, we avoid this race by checking rq->deadline
714 * and REQ_ATOM_COMPLETE flag together:
716 * - if rq->deadline is observed as new value because of
717 * reusing, the rq won't be timed out because of timing.
718 * - if rq->deadline is observed as previous value,
719 * REQ_ATOM_COMPLETE flag won't be cleared in reuse path
720 * because we put a barrier between setting rq->deadline
721 * and clearing the flag in blk_mq_start_request(), so
722 * this rq won't be timed out too.
724 if (time_after_eq(jiffies, rq->deadline)) {
725 if (!blk_mark_rq_complete(rq))
726 blk_mq_rq_timed_out(rq, reserved);
727 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
728 data->next = rq->deadline;
733 static void blk_mq_timeout_work(struct work_struct *work)
735 struct request_queue *q =
736 container_of(work, struct request_queue, timeout_work);
737 struct blk_mq_timeout_data data = {
743 /* A deadlock might occur if a request is stuck requiring a
744 * timeout at the same time a queue freeze is waiting
745 * completion, since the timeout code would not be able to
746 * acquire the queue reference here.
748 * That's why we don't use blk_queue_enter here; instead, we use
749 * percpu_ref_tryget directly, because we need to be able to
750 * obtain a reference even in the short window between the queue
751 * starting to freeze, by dropping the first reference in
752 * blk_freeze_queue_start, and the moment the last request is
753 * consumed, marked by the instant q_usage_counter reaches
756 if (!percpu_ref_tryget(&q->q_usage_counter))
759 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
762 data.next = blk_rq_timeout(round_jiffies_up(data.next));
763 mod_timer(&q->timeout, data.next);
765 struct blk_mq_hw_ctx *hctx;
767 queue_for_each_hw_ctx(q, hctx, i) {
768 /* the hctx may be unmapped, so check it here */
769 if (blk_mq_hw_queue_mapped(hctx))
770 blk_mq_tag_idle(hctx);
777 * Reverse check our software queue for entries that we could potentially
778 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
779 * too much time checking for merges.
781 static bool blk_mq_attempt_merge(struct request_queue *q,
782 struct blk_mq_ctx *ctx, struct bio *bio)
787 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
793 if (!blk_rq_merge_ok(rq, bio))
796 switch (blk_try_merge(rq, bio)) {
797 case ELEVATOR_BACK_MERGE:
798 if (blk_mq_sched_allow_merge(q, rq, bio))
799 merged = bio_attempt_back_merge(q, rq, bio);
801 case ELEVATOR_FRONT_MERGE:
802 if (blk_mq_sched_allow_merge(q, rq, bio))
803 merged = bio_attempt_front_merge(q, rq, bio);
805 case ELEVATOR_DISCARD_MERGE:
806 merged = bio_attempt_discard_merge(q, rq, bio);
820 struct flush_busy_ctx_data {
821 struct blk_mq_hw_ctx *hctx;
822 struct list_head *list;
825 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
827 struct flush_busy_ctx_data *flush_data = data;
828 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
829 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
831 sbitmap_clear_bit(sb, bitnr);
832 spin_lock(&ctx->lock);
833 list_splice_tail_init(&ctx->rq_list, flush_data->list);
834 spin_unlock(&ctx->lock);
839 * Process software queues that have been marked busy, splicing them
840 * to the for-dispatch
842 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
844 struct flush_busy_ctx_data data = {
849 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
851 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
853 static inline unsigned int queued_to_index(unsigned int queued)
858 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
861 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
864 struct blk_mq_alloc_data data = {
866 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
867 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
873 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
874 data.flags |= BLK_MQ_REQ_RESERVED;
876 rq->tag = blk_mq_get_tag(&data);
878 if (blk_mq_tag_busy(data.hctx)) {
879 rq->rq_flags |= RQF_MQ_INFLIGHT;
880 atomic_inc(&data.hctx->nr_active);
882 data.hctx->tags->rqs[rq->tag] = rq;
888 return rq->tag != -1;
891 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
894 blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
897 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
898 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
899 atomic_dec(&hctx->nr_active);
903 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
906 if (rq->tag == -1 || rq->internal_tag == -1)
909 __blk_mq_put_driver_tag(hctx, rq);
912 static void blk_mq_put_driver_tag(struct request *rq)
914 struct blk_mq_hw_ctx *hctx;
916 if (rq->tag == -1 || rq->internal_tag == -1)
919 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
920 __blk_mq_put_driver_tag(hctx, rq);
924 * If we fail getting a driver tag because all the driver tags are already
925 * assigned and on the dispatch list, BUT the first entry does not have a
926 * tag, then we could deadlock. For that case, move entries with assigned
927 * driver tags to the front, leaving the set of tagged requests in the
928 * same order, and the untagged set in the same order.
930 static bool reorder_tags_to_front(struct list_head *list)
932 struct request *rq, *tmp, *first = NULL;
934 list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
938 list_move(&rq->queuelist, list);
944 return first != NULL;
947 static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags,
950 struct blk_mq_hw_ctx *hctx;
952 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
954 list_del(&wait->task_list);
955 clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
956 blk_mq_run_hw_queue(hctx, true);
960 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
962 struct sbq_wait_state *ws;
965 * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
966 * The thread which wins the race to grab this bit adds the hardware
967 * queue to the wait queue.
969 if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
970 test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
973 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
974 ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
977 * As soon as this returns, it's no longer safe to fiddle with
978 * hctx->dispatch_wait, since a completion can wake up the wait queue
979 * and unlock the bit.
981 add_wait_queue(&ws->wait, &hctx->dispatch_wait);
985 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
987 struct blk_mq_hw_ctx *hctx;
989 int errors, queued, ret = BLK_MQ_RQ_QUEUE_OK;
991 if (list_empty(list))
995 * Now process all the entries, sending them to the driver.
999 struct blk_mq_queue_data bd;
1001 rq = list_first_entry(list, struct request, queuelist);
1002 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
1003 if (!queued && reorder_tags_to_front(list))
1007 * The initial allocation attempt failed, so we need to
1008 * rerun the hardware queue when a tag is freed.
1010 if (!blk_mq_dispatch_wait_add(hctx))
1014 * It's possible that a tag was freed in the window
1015 * between the allocation failure and adding the
1016 * hardware queue to the wait queue.
1018 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1022 list_del_init(&rq->queuelist);
1027 * Flag last if we have no more requests, or if we have more
1028 * but can't assign a driver tag to it.
1030 if (list_empty(list))
1033 struct request *nxt;
1035 nxt = list_first_entry(list, struct request, queuelist);
1036 bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1039 ret = q->mq_ops->queue_rq(hctx, &bd);
1041 case BLK_MQ_RQ_QUEUE_OK:
1044 case BLK_MQ_RQ_QUEUE_BUSY:
1045 blk_mq_put_driver_tag_hctx(hctx, rq);
1046 list_add(&rq->queuelist, list);
1047 __blk_mq_requeue_request(rq);
1050 pr_err("blk-mq: bad return on queue: %d\n", ret);
1051 case BLK_MQ_RQ_QUEUE_ERROR:
1054 blk_mq_end_request(rq, rq->errors);
1058 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
1060 } while (!list_empty(list));
1062 hctx->dispatched[queued_to_index(queued)]++;
1065 * Any items that need requeuing? Stuff them into hctx->dispatch,
1066 * that is where we will continue on next queue run.
1068 if (!list_empty(list)) {
1070 * If an I/O scheduler has been configured and we got a driver
1071 * tag for the next request already, free it again.
1073 rq = list_first_entry(list, struct request, queuelist);
1074 blk_mq_put_driver_tag(rq);
1076 spin_lock(&hctx->lock);
1077 list_splice_init(list, &hctx->dispatch);
1078 spin_unlock(&hctx->lock);
1081 * If SCHED_RESTART was set by the caller of this function and
1082 * it is no longer set that means that it was cleared by another
1083 * thread and hence that a queue rerun is needed.
1085 * If TAG_WAITING is set that means that an I/O scheduler has
1086 * been configured and another thread is waiting for a driver
1087 * tag. To guarantee fairness, do not rerun this hardware queue
1088 * but let the other thread grab the driver tag.
1090 * If no I/O scheduler has been configured it is possible that
1091 * the hardware queue got stopped and restarted before requests
1092 * were pushed back onto the dispatch list. Rerun the queue to
1093 * avoid starvation. Notes:
1094 * - blk_mq_run_hw_queue() checks whether or not a queue has
1095 * been stopped before rerunning a queue.
1096 * - Some but not all block drivers stop a queue before
1097 * returning BLK_MQ_RQ_QUEUE_BUSY. Two exceptions are scsi-mq
1100 if (!blk_mq_sched_needs_restart(hctx) &&
1101 !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1102 blk_mq_run_hw_queue(hctx, true);
1105 return (queued + errors) != 0;
1108 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1112 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1113 cpu_online(hctx->next_cpu));
1115 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1117 blk_mq_sched_dispatch_requests(hctx);
1122 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1123 blk_mq_sched_dispatch_requests(hctx);
1124 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1129 * It'd be great if the workqueue API had a way to pass
1130 * in a mask and had some smarts for more clever placement.
1131 * For now we just round-robin here, switching for every
1132 * BLK_MQ_CPU_WORK_BATCH queued items.
1134 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1136 if (hctx->queue->nr_hw_queues == 1)
1137 return WORK_CPU_UNBOUND;
1139 if (--hctx->next_cpu_batch <= 0) {
1142 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1143 if (next_cpu >= nr_cpu_ids)
1144 next_cpu = cpumask_first(hctx->cpumask);
1146 hctx->next_cpu = next_cpu;
1147 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1150 return hctx->next_cpu;
1153 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1154 unsigned long msecs)
1156 if (unlikely(blk_mq_hctx_stopped(hctx) ||
1157 !blk_mq_hw_queue_mapped(hctx)))
1160 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1161 int cpu = get_cpu();
1162 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1163 __blk_mq_run_hw_queue(hctx);
1172 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx),
1175 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1176 &hctx->delayed_run_work,
1177 msecs_to_jiffies(msecs));
1180 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1182 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1184 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1186 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1188 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1190 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1192 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1194 struct blk_mq_hw_ctx *hctx;
1197 queue_for_each_hw_ctx(q, hctx, i) {
1198 if (!blk_mq_hctx_has_pending(hctx) ||
1199 blk_mq_hctx_stopped(hctx))
1202 blk_mq_run_hw_queue(hctx, async);
1205 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1208 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1209 * @q: request queue.
1211 * The caller is responsible for serializing this function against
1212 * blk_mq_{start,stop}_hw_queue().
1214 bool blk_mq_queue_stopped(struct request_queue *q)
1216 struct blk_mq_hw_ctx *hctx;
1219 queue_for_each_hw_ctx(q, hctx, i)
1220 if (blk_mq_hctx_stopped(hctx))
1225 EXPORT_SYMBOL(blk_mq_queue_stopped);
1227 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1229 cancel_work(&hctx->run_work);
1230 cancel_delayed_work(&hctx->delay_work);
1231 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1233 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1235 void blk_mq_stop_hw_queues(struct request_queue *q)
1237 struct blk_mq_hw_ctx *hctx;
1240 queue_for_each_hw_ctx(q, hctx, i)
1241 blk_mq_stop_hw_queue(hctx);
1243 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1245 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1247 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1249 blk_mq_run_hw_queue(hctx, false);
1251 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1253 void blk_mq_start_hw_queues(struct request_queue *q)
1255 struct blk_mq_hw_ctx *hctx;
1258 queue_for_each_hw_ctx(q, hctx, i)
1259 blk_mq_start_hw_queue(hctx);
1261 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1263 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1265 if (!blk_mq_hctx_stopped(hctx))
1268 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1269 blk_mq_run_hw_queue(hctx, async);
1271 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1273 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1275 struct blk_mq_hw_ctx *hctx;
1278 queue_for_each_hw_ctx(q, hctx, i)
1279 blk_mq_start_stopped_hw_queue(hctx, async);
1281 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1283 static void blk_mq_run_work_fn(struct work_struct *work)
1285 struct blk_mq_hw_ctx *hctx;
1287 hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
1289 __blk_mq_run_hw_queue(hctx);
1292 static void blk_mq_delayed_run_work_fn(struct work_struct *work)
1294 struct blk_mq_hw_ctx *hctx;
1296 hctx = container_of(work, struct blk_mq_hw_ctx, delayed_run_work.work);
1298 __blk_mq_run_hw_queue(hctx);
1301 static void blk_mq_delay_work_fn(struct work_struct *work)
1303 struct blk_mq_hw_ctx *hctx;
1305 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1307 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1308 __blk_mq_run_hw_queue(hctx);
1311 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1313 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1316 blk_mq_stop_hw_queue(hctx);
1317 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1318 &hctx->delay_work, msecs_to_jiffies(msecs));
1320 EXPORT_SYMBOL(blk_mq_delay_queue);
1322 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1326 struct blk_mq_ctx *ctx = rq->mq_ctx;
1328 trace_block_rq_insert(hctx->queue, rq);
1331 list_add(&rq->queuelist, &ctx->rq_list);
1333 list_add_tail(&rq->queuelist, &ctx->rq_list);
1336 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1339 struct blk_mq_ctx *ctx = rq->mq_ctx;
1341 __blk_mq_insert_req_list(hctx, rq, at_head);
1342 blk_mq_hctx_mark_pending(hctx, ctx);
1345 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1346 struct list_head *list)
1350 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1353 spin_lock(&ctx->lock);
1354 while (!list_empty(list)) {
1357 rq = list_first_entry(list, struct request, queuelist);
1358 BUG_ON(rq->mq_ctx != ctx);
1359 list_del_init(&rq->queuelist);
1360 __blk_mq_insert_req_list(hctx, rq, false);
1362 blk_mq_hctx_mark_pending(hctx, ctx);
1363 spin_unlock(&ctx->lock);
1366 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1368 struct request *rqa = container_of(a, struct request, queuelist);
1369 struct request *rqb = container_of(b, struct request, queuelist);
1371 return !(rqa->mq_ctx < rqb->mq_ctx ||
1372 (rqa->mq_ctx == rqb->mq_ctx &&
1373 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1376 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1378 struct blk_mq_ctx *this_ctx;
1379 struct request_queue *this_q;
1382 LIST_HEAD(ctx_list);
1385 list_splice_init(&plug->mq_list, &list);
1387 list_sort(NULL, &list, plug_ctx_cmp);
1393 while (!list_empty(&list)) {
1394 rq = list_entry_rq(list.next);
1395 list_del_init(&rq->queuelist);
1397 if (rq->mq_ctx != this_ctx) {
1399 trace_block_unplug(this_q, depth, from_schedule);
1400 blk_mq_sched_insert_requests(this_q, this_ctx,
1405 this_ctx = rq->mq_ctx;
1411 list_add_tail(&rq->queuelist, &ctx_list);
1415 * If 'this_ctx' is set, we know we have entries to complete
1416 * on 'ctx_list'. Do those.
1419 trace_block_unplug(this_q, depth, from_schedule);
1420 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1425 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1427 init_request_from_bio(rq, bio);
1429 blk_account_io_start(rq, true);
1432 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1434 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1435 !blk_queue_nomerges(hctx->queue);
1438 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1439 struct blk_mq_ctx *ctx,
1440 struct request *rq, struct bio *bio)
1442 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1443 blk_mq_bio_to_request(rq, bio);
1444 spin_lock(&ctx->lock);
1446 __blk_mq_insert_request(hctx, rq, false);
1447 spin_unlock(&ctx->lock);
1450 struct request_queue *q = hctx->queue;
1452 spin_lock(&ctx->lock);
1453 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1454 blk_mq_bio_to_request(rq, bio);
1458 spin_unlock(&ctx->lock);
1459 __blk_mq_finish_request(hctx, ctx, rq);
1464 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1467 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1469 return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1472 static void __blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie,
1475 struct request_queue *q = rq->q;
1476 struct blk_mq_queue_data bd = {
1480 struct blk_mq_hw_ctx *hctx;
1481 blk_qc_t new_cookie;
1487 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1490 new_cookie = request_to_qc_t(hctx, rq);
1493 * For OK queue, we are done. For error, kill it. Any other
1494 * error (busy), just add it to our list as we previously
1497 ret = q->mq_ops->queue_rq(hctx, &bd);
1498 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1499 *cookie = new_cookie;
1503 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1504 *cookie = BLK_QC_T_NONE;
1506 blk_mq_end_request(rq, rq->errors);
1510 __blk_mq_requeue_request(rq);
1512 blk_mq_sched_insert_request(rq, false, true, false, may_sleep);
1515 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1516 struct request *rq, blk_qc_t *cookie)
1518 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1520 __blk_mq_try_issue_directly(rq, cookie, false);
1523 unsigned int srcu_idx;
1527 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1528 __blk_mq_try_issue_directly(rq, cookie, true);
1529 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1533 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1535 const int is_sync = op_is_sync(bio->bi_opf);
1536 const int is_flush_fua = op_is_flush(bio->bi_opf);
1537 struct blk_mq_alloc_data data = { .flags = 0 };
1539 unsigned int request_count = 0;
1540 struct blk_plug *plug;
1541 struct request *same_queue_rq = NULL;
1543 unsigned int wb_acct;
1545 blk_queue_bounce(q, &bio);
1547 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1549 return BLK_QC_T_NONE;
1552 blk_queue_split(q, &bio, q->bio_split);
1554 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1555 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1556 return BLK_QC_T_NONE;
1558 if (blk_mq_sched_bio_merge(q, bio))
1559 return BLK_QC_T_NONE;
1561 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1563 trace_block_getrq(q, bio, bio->bi_opf);
1565 rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1566 if (unlikely(!rq)) {
1567 __wbt_done(q->rq_wb, wb_acct);
1568 return BLK_QC_T_NONE;
1571 wbt_track(&rq->issue_stat, wb_acct);
1573 cookie = request_to_qc_t(data.hctx, rq);
1575 plug = current->plug;
1576 if (unlikely(is_flush_fua)) {
1577 blk_mq_bio_to_request(rq, bio);
1579 blk_mq_sched_insert_request(rq, false, true, true,
1582 blk_insert_flush(rq);
1583 blk_mq_run_hw_queue(data.hctx, true);
1585 } else if (plug && q->nr_hw_queues == 1) {
1586 struct request *last = NULL;
1588 blk_mq_bio_to_request(rq, bio);
1591 * @request_count may become stale because of schedule
1592 * out, so check the list again.
1594 if (list_empty(&plug->mq_list))
1596 else if (blk_queue_nomerges(q))
1597 request_count = blk_plug_queued_count(q);
1600 trace_block_plug(q);
1602 last = list_entry_rq(plug->mq_list.prev);
1604 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1605 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1606 blk_flush_plug_list(plug, false);
1607 trace_block_plug(q);
1610 list_add_tail(&rq->queuelist, &plug->mq_list);
1611 } else if (plug && !blk_queue_nomerges(q)) {
1612 blk_mq_bio_to_request(rq, bio);
1615 * We do limited plugging. If the bio can be merged, do that.
1616 * Otherwise the existing request in the plug list will be
1617 * issued. So the plug list will have one request at most
1618 * The plug list might get flushed before this. If that happens,
1619 * the plug list is empty, and same_queue_rq is invalid.
1621 if (list_empty(&plug->mq_list))
1622 same_queue_rq = NULL;
1624 list_del_init(&same_queue_rq->queuelist);
1625 list_add_tail(&rq->queuelist, &plug->mq_list);
1627 blk_mq_put_ctx(data.ctx);
1630 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1634 } else if (q->nr_hw_queues > 1 && is_sync) {
1635 blk_mq_put_ctx(data.ctx);
1636 blk_mq_bio_to_request(rq, bio);
1637 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1639 } else if (q->elevator) {
1640 blk_mq_bio_to_request(rq, bio);
1641 blk_mq_sched_insert_request(rq, false, true, true, true);
1642 } else if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio))
1643 blk_mq_run_hw_queue(data.hctx, true);
1645 blk_mq_put_ctx(data.ctx);
1649 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1650 unsigned int hctx_idx)
1654 if (tags->rqs && set->ops->exit_request) {
1657 for (i = 0; i < tags->nr_tags; i++) {
1658 struct request *rq = tags->static_rqs[i];
1662 set->ops->exit_request(set->driver_data, rq,
1664 tags->static_rqs[i] = NULL;
1668 while (!list_empty(&tags->page_list)) {
1669 page = list_first_entry(&tags->page_list, struct page, lru);
1670 list_del_init(&page->lru);
1672 * Remove kmemleak object previously allocated in
1673 * blk_mq_init_rq_map().
1675 kmemleak_free(page_address(page));
1676 __free_pages(page, page->private);
1680 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1684 kfree(tags->static_rqs);
1685 tags->static_rqs = NULL;
1687 blk_mq_free_tags(tags);
1690 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1691 unsigned int hctx_idx,
1692 unsigned int nr_tags,
1693 unsigned int reserved_tags)
1695 struct blk_mq_tags *tags;
1698 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1699 if (node == NUMA_NO_NODE)
1700 node = set->numa_node;
1702 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1703 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1707 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1708 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1711 blk_mq_free_tags(tags);
1715 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1716 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1718 if (!tags->static_rqs) {
1720 blk_mq_free_tags(tags);
1727 static size_t order_to_size(unsigned int order)
1729 return (size_t)PAGE_SIZE << order;
1732 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1733 unsigned int hctx_idx, unsigned int depth)
1735 unsigned int i, j, entries_per_page, max_order = 4;
1736 size_t rq_size, left;
1739 node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1740 if (node == NUMA_NO_NODE)
1741 node = set->numa_node;
1743 INIT_LIST_HEAD(&tags->page_list);
1746 * rq_size is the size of the request plus driver payload, rounded
1747 * to the cacheline size
1749 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1751 left = rq_size * depth;
1753 for (i = 0; i < depth; ) {
1754 int this_order = max_order;
1759 while (this_order && left < order_to_size(this_order - 1))
1763 page = alloc_pages_node(node,
1764 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1770 if (order_to_size(this_order) < rq_size)
1777 page->private = this_order;
1778 list_add_tail(&page->lru, &tags->page_list);
1780 p = page_address(page);
1782 * Allow kmemleak to scan these pages as they contain pointers
1783 * to additional allocations like via ops->init_request().
1785 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1786 entries_per_page = order_to_size(this_order) / rq_size;
1787 to_do = min(entries_per_page, depth - i);
1788 left -= to_do * rq_size;
1789 for (j = 0; j < to_do; j++) {
1790 struct request *rq = p;
1792 tags->static_rqs[i] = rq;
1793 if (set->ops->init_request) {
1794 if (set->ops->init_request(set->driver_data,
1797 tags->static_rqs[i] = NULL;
1809 blk_mq_free_rqs(set, tags, hctx_idx);
1814 * 'cpu' is going away. splice any existing rq_list entries from this
1815 * software queue to the hw queue dispatch list, and ensure that it
1818 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1820 struct blk_mq_hw_ctx *hctx;
1821 struct blk_mq_ctx *ctx;
1824 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1825 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1827 spin_lock(&ctx->lock);
1828 if (!list_empty(&ctx->rq_list)) {
1829 list_splice_init(&ctx->rq_list, &tmp);
1830 blk_mq_hctx_clear_pending(hctx, ctx);
1832 spin_unlock(&ctx->lock);
1834 if (list_empty(&tmp))
1837 spin_lock(&hctx->lock);
1838 list_splice_tail_init(&tmp, &hctx->dispatch);
1839 spin_unlock(&hctx->lock);
1841 blk_mq_run_hw_queue(hctx, true);
1845 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1847 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1851 /* hctx->ctxs will be freed in queue's release handler */
1852 static void blk_mq_exit_hctx(struct request_queue *q,
1853 struct blk_mq_tag_set *set,
1854 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1856 unsigned flush_start_tag = set->queue_depth;
1858 blk_mq_tag_idle(hctx);
1860 if (set->ops->exit_request)
1861 set->ops->exit_request(set->driver_data,
1862 hctx->fq->flush_rq, hctx_idx,
1863 flush_start_tag + hctx_idx);
1865 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1867 if (set->ops->exit_hctx)
1868 set->ops->exit_hctx(hctx, hctx_idx);
1870 if (hctx->flags & BLK_MQ_F_BLOCKING)
1871 cleanup_srcu_struct(&hctx->queue_rq_srcu);
1873 blk_mq_remove_cpuhp(hctx);
1874 blk_free_flush_queue(hctx->fq);
1875 sbitmap_free(&hctx->ctx_map);
1878 static void blk_mq_exit_hw_queues(struct request_queue *q,
1879 struct blk_mq_tag_set *set, int nr_queue)
1881 struct blk_mq_hw_ctx *hctx;
1884 queue_for_each_hw_ctx(q, hctx, i) {
1887 blk_mq_exit_hctx(q, set, hctx, i);
1891 static int blk_mq_init_hctx(struct request_queue *q,
1892 struct blk_mq_tag_set *set,
1893 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1896 unsigned flush_start_tag = set->queue_depth;
1898 node = hctx->numa_node;
1899 if (node == NUMA_NO_NODE)
1900 node = hctx->numa_node = set->numa_node;
1902 INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1903 INIT_DELAYED_WORK(&hctx->delayed_run_work, blk_mq_delayed_run_work_fn);
1904 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1905 spin_lock_init(&hctx->lock);
1906 INIT_LIST_HEAD(&hctx->dispatch);
1908 hctx->queue_num = hctx_idx;
1909 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1911 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1913 hctx->tags = set->tags[hctx_idx];
1916 * Allocate space for all possible cpus to avoid allocation at
1919 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1922 goto unregister_cpu_notifier;
1924 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1930 if (set->ops->init_hctx &&
1931 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1934 if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
1937 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1939 goto sched_exit_hctx;
1941 if (set->ops->init_request &&
1942 set->ops->init_request(set->driver_data,
1943 hctx->fq->flush_rq, hctx_idx,
1944 flush_start_tag + hctx_idx, node))
1947 if (hctx->flags & BLK_MQ_F_BLOCKING)
1948 init_srcu_struct(&hctx->queue_rq_srcu);
1955 blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1957 if (set->ops->exit_hctx)
1958 set->ops->exit_hctx(hctx, hctx_idx);
1960 sbitmap_free(&hctx->ctx_map);
1963 unregister_cpu_notifier:
1964 blk_mq_remove_cpuhp(hctx);
1968 static void blk_mq_init_cpu_queues(struct request_queue *q,
1969 unsigned int nr_hw_queues)
1973 for_each_possible_cpu(i) {
1974 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1975 struct blk_mq_hw_ctx *hctx;
1978 spin_lock_init(&__ctx->lock);
1979 INIT_LIST_HEAD(&__ctx->rq_list);
1982 /* If the cpu isn't online, the cpu is mapped to first hctx */
1986 hctx = blk_mq_map_queue(q, i);
1989 * Set local node, IFF we have more than one hw queue. If
1990 * not, we remain on the home node of the device
1992 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1993 hctx->numa_node = local_memory_node(cpu_to_node(i));
1997 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2001 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2002 set->queue_depth, set->reserved_tags);
2003 if (!set->tags[hctx_idx])
2006 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2011 blk_mq_free_rq_map(set->tags[hctx_idx]);
2012 set->tags[hctx_idx] = NULL;
2016 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2017 unsigned int hctx_idx)
2019 if (set->tags[hctx_idx]) {
2020 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2021 blk_mq_free_rq_map(set->tags[hctx_idx]);
2022 set->tags[hctx_idx] = NULL;
2026 static void blk_mq_map_swqueue(struct request_queue *q,
2027 const struct cpumask *online_mask)
2029 unsigned int i, hctx_idx;
2030 struct blk_mq_hw_ctx *hctx;
2031 struct blk_mq_ctx *ctx;
2032 struct blk_mq_tag_set *set = q->tag_set;
2035 * Avoid others reading imcomplete hctx->cpumask through sysfs
2037 mutex_lock(&q->sysfs_lock);
2039 queue_for_each_hw_ctx(q, hctx, i) {
2040 cpumask_clear(hctx->cpumask);
2045 * Map software to hardware queues
2047 for_each_possible_cpu(i) {
2048 /* If the cpu isn't online, the cpu is mapped to first hctx */
2049 if (!cpumask_test_cpu(i, online_mask))
2052 hctx_idx = q->mq_map[i];
2053 /* unmapped hw queue can be remapped after CPU topo changed */
2054 if (!set->tags[hctx_idx] &&
2055 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2057 * If tags initialization fail for some hctx,
2058 * that hctx won't be brought online. In this
2059 * case, remap the current ctx to hctx[0] which
2060 * is guaranteed to always have tags allocated
2065 ctx = per_cpu_ptr(q->queue_ctx, i);
2066 hctx = blk_mq_map_queue(q, i);
2068 cpumask_set_cpu(i, hctx->cpumask);
2069 ctx->index_hw = hctx->nr_ctx;
2070 hctx->ctxs[hctx->nr_ctx++] = ctx;
2073 mutex_unlock(&q->sysfs_lock);
2075 queue_for_each_hw_ctx(q, hctx, i) {
2077 * If no software queues are mapped to this hardware queue,
2078 * disable it and free the request entries.
2080 if (!hctx->nr_ctx) {
2081 /* Never unmap queue 0. We need it as a
2082 * fallback in case of a new remap fails
2085 if (i && set->tags[i])
2086 blk_mq_free_map_and_requests(set, i);
2092 hctx->tags = set->tags[i];
2093 WARN_ON(!hctx->tags);
2096 * Set the map size to the number of mapped software queues.
2097 * This is more accurate and more efficient than looping
2098 * over all possibly mapped software queues.
2100 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2103 * Initialize batch roundrobin counts
2105 hctx->next_cpu = cpumask_first(hctx->cpumask);
2106 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2110 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2112 struct blk_mq_hw_ctx *hctx;
2115 queue_for_each_hw_ctx(q, hctx, i) {
2117 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2119 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2123 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2125 struct request_queue *q;
2127 lockdep_assert_held(&set->tag_list_lock);
2129 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2130 blk_mq_freeze_queue(q);
2131 queue_set_hctx_shared(q, shared);
2132 blk_mq_unfreeze_queue(q);
2136 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2138 struct blk_mq_tag_set *set = q->tag_set;
2140 mutex_lock(&set->tag_list_lock);
2141 list_del_rcu(&q->tag_set_list);
2142 INIT_LIST_HEAD(&q->tag_set_list);
2143 if (list_is_singular(&set->tag_list)) {
2144 /* just transitioned to unshared */
2145 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2146 /* update existing queue */
2147 blk_mq_update_tag_set_depth(set, false);
2149 mutex_unlock(&set->tag_list_lock);
2154 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2155 struct request_queue *q)
2159 mutex_lock(&set->tag_list_lock);
2161 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2162 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2163 set->flags |= BLK_MQ_F_TAG_SHARED;
2164 /* update existing queue */
2165 blk_mq_update_tag_set_depth(set, true);
2167 if (set->flags & BLK_MQ_F_TAG_SHARED)
2168 queue_set_hctx_shared(q, true);
2169 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2171 mutex_unlock(&set->tag_list_lock);
2175 * It is the actual release handler for mq, but we do it from
2176 * request queue's release handler for avoiding use-after-free
2177 * and headache because q->mq_kobj shouldn't have been introduced,
2178 * but we can't group ctx/kctx kobj without it.
2180 void blk_mq_release(struct request_queue *q)
2182 struct blk_mq_hw_ctx *hctx;
2185 /* hctx kobj stays in hctx */
2186 queue_for_each_hw_ctx(q, hctx, i) {
2189 kobject_put(&hctx->kobj);
2194 kfree(q->queue_hw_ctx);
2197 * release .mq_kobj and sw queue's kobject now because
2198 * both share lifetime with request queue.
2200 blk_mq_sysfs_deinit(q);
2202 free_percpu(q->queue_ctx);
2205 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2207 struct request_queue *uninit_q, *q;
2209 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2211 return ERR_PTR(-ENOMEM);
2213 q = blk_mq_init_allocated_queue(set, uninit_q);
2215 blk_cleanup_queue(uninit_q);
2219 EXPORT_SYMBOL(blk_mq_init_queue);
2221 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2222 struct request_queue *q)
2225 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2227 blk_mq_sysfs_unregister(q);
2228 for (i = 0; i < set->nr_hw_queues; i++) {
2234 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2235 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2240 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2247 atomic_set(&hctxs[i]->nr_active, 0);
2248 hctxs[i]->numa_node = node;
2249 hctxs[i]->queue_num = i;
2251 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2252 free_cpumask_var(hctxs[i]->cpumask);
2257 blk_mq_hctx_kobj_init(hctxs[i]);
2259 for (j = i; j < q->nr_hw_queues; j++) {
2260 struct blk_mq_hw_ctx *hctx = hctxs[j];
2264 blk_mq_free_map_and_requests(set, j);
2265 blk_mq_exit_hctx(q, set, hctx, j);
2266 kobject_put(&hctx->kobj);
2271 q->nr_hw_queues = i;
2272 blk_mq_sysfs_register(q);
2275 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2276 struct request_queue *q)
2278 /* mark the queue as mq asap */
2279 q->mq_ops = set->ops;
2281 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2282 blk_stat_rq_ddir, 2, q);
2286 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2290 /* init q->mq_kobj and sw queues' kobjects */
2291 blk_mq_sysfs_init(q);
2293 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2294 GFP_KERNEL, set->numa_node);
2295 if (!q->queue_hw_ctx)
2298 q->mq_map = set->mq_map;
2300 blk_mq_realloc_hw_ctxs(set, q);
2301 if (!q->nr_hw_queues)
2304 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2305 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2307 q->nr_queues = nr_cpu_ids;
2309 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2311 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2312 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2314 q->sg_reserved_size = INT_MAX;
2316 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2317 INIT_LIST_HEAD(&q->requeue_list);
2318 spin_lock_init(&q->requeue_lock);
2320 blk_queue_make_request(q, blk_mq_make_request);
2323 * Do this after blk_queue_make_request() overrides it...
2325 q->nr_requests = set->queue_depth;
2328 * Default to classic polling
2332 if (set->ops->complete)
2333 blk_queue_softirq_done(q, set->ops->complete);
2335 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2338 mutex_lock(&all_q_mutex);
2340 list_add_tail(&q->all_q_node, &all_q_list);
2341 blk_mq_add_queue_tag_set(set, q);
2342 blk_mq_map_swqueue(q, cpu_online_mask);
2344 mutex_unlock(&all_q_mutex);
2347 if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2350 ret = blk_mq_sched_init(q);
2352 return ERR_PTR(ret);
2358 kfree(q->queue_hw_ctx);
2360 free_percpu(q->queue_ctx);
2363 return ERR_PTR(-ENOMEM);
2365 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2367 void blk_mq_free_queue(struct request_queue *q)
2369 struct blk_mq_tag_set *set = q->tag_set;
2371 mutex_lock(&all_q_mutex);
2372 list_del_init(&q->all_q_node);
2373 mutex_unlock(&all_q_mutex);
2375 blk_mq_del_queue_tag_set(q);
2377 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2380 /* Basically redo blk_mq_init_queue with queue frozen */
2381 static void blk_mq_queue_reinit(struct request_queue *q,
2382 const struct cpumask *online_mask)
2384 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2386 blk_mq_sysfs_unregister(q);
2389 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2390 * we should change hctx numa_node according to new topology (this
2391 * involves free and re-allocate memory, worthy doing?)
2394 blk_mq_map_swqueue(q, online_mask);
2396 blk_mq_sysfs_register(q);
2400 * New online cpumask which is going to be set in this hotplug event.
2401 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2402 * one-by-one and dynamically allocating this could result in a failure.
2404 static struct cpumask cpuhp_online_new;
2406 static void blk_mq_queue_reinit_work(void)
2408 struct request_queue *q;
2410 mutex_lock(&all_q_mutex);
2412 * We need to freeze and reinit all existing queues. Freezing
2413 * involves synchronous wait for an RCU grace period and doing it
2414 * one by one may take a long time. Start freezing all queues in
2415 * one swoop and then wait for the completions so that freezing can
2416 * take place in parallel.
2418 list_for_each_entry(q, &all_q_list, all_q_node)
2419 blk_freeze_queue_start(q);
2420 list_for_each_entry(q, &all_q_list, all_q_node)
2421 blk_mq_freeze_queue_wait(q);
2423 list_for_each_entry(q, &all_q_list, all_q_node)
2424 blk_mq_queue_reinit(q, &cpuhp_online_new);
2426 list_for_each_entry(q, &all_q_list, all_q_node)
2427 blk_mq_unfreeze_queue(q);
2429 mutex_unlock(&all_q_mutex);
2432 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2434 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2435 blk_mq_queue_reinit_work();
2440 * Before hotadded cpu starts handling requests, new mappings must be
2441 * established. Otherwise, these requests in hw queue might never be
2444 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2445 * for CPU0, and ctx1 for CPU1).
2447 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2448 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2450 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2451 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2452 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2455 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2457 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2458 cpumask_set_cpu(cpu, &cpuhp_online_new);
2459 blk_mq_queue_reinit_work();
2463 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2467 for (i = 0; i < set->nr_hw_queues; i++)
2468 if (!__blk_mq_alloc_rq_map(set, i))
2475 blk_mq_free_rq_map(set->tags[i]);
2481 * Allocate the request maps associated with this tag_set. Note that this
2482 * may reduce the depth asked for, if memory is tight. set->queue_depth
2483 * will be updated to reflect the allocated depth.
2485 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2490 depth = set->queue_depth;
2492 err = __blk_mq_alloc_rq_maps(set);
2496 set->queue_depth >>= 1;
2497 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2501 } while (set->queue_depth);
2503 if (!set->queue_depth || err) {
2504 pr_err("blk-mq: failed to allocate request map\n");
2508 if (depth != set->queue_depth)
2509 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2510 depth, set->queue_depth);
2515 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2517 if (set->ops->map_queues)
2518 return set->ops->map_queues(set);
2520 return blk_mq_map_queues(set);
2524 * Alloc a tag set to be associated with one or more request queues.
2525 * May fail with EINVAL for various error conditions. May adjust the
2526 * requested depth down, if if it too large. In that case, the set
2527 * value will be stored in set->queue_depth.
2529 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2533 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2535 if (!set->nr_hw_queues)
2537 if (!set->queue_depth)
2539 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2542 if (!set->ops->queue_rq)
2545 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2546 pr_info("blk-mq: reduced tag depth to %u\n",
2548 set->queue_depth = BLK_MQ_MAX_DEPTH;
2552 * If a crashdump is active, then we are potentially in a very
2553 * memory constrained environment. Limit us to 1 queue and
2554 * 64 tags to prevent using too much memory.
2556 if (is_kdump_kernel()) {
2557 set->nr_hw_queues = 1;
2558 set->queue_depth = min(64U, set->queue_depth);
2561 * There is no use for more h/w queues than cpus.
2563 if (set->nr_hw_queues > nr_cpu_ids)
2564 set->nr_hw_queues = nr_cpu_ids;
2566 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2567 GFP_KERNEL, set->numa_node);
2572 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2573 GFP_KERNEL, set->numa_node);
2577 ret = blk_mq_update_queue_map(set);
2579 goto out_free_mq_map;
2581 ret = blk_mq_alloc_rq_maps(set);
2583 goto out_free_mq_map;
2585 mutex_init(&set->tag_list_lock);
2586 INIT_LIST_HEAD(&set->tag_list);
2598 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2600 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2604 for (i = 0; i < nr_cpu_ids; i++)
2605 blk_mq_free_map_and_requests(set, i);
2613 EXPORT_SYMBOL(blk_mq_free_tag_set);
2615 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2617 struct blk_mq_tag_set *set = q->tag_set;
2618 struct blk_mq_hw_ctx *hctx;
2624 blk_mq_freeze_queue(q);
2625 blk_mq_quiesce_queue(q);
2628 queue_for_each_hw_ctx(q, hctx, i) {
2632 * If we're using an MQ scheduler, just update the scheduler
2633 * queue depth. This is similar to what the old code would do.
2635 if (!hctx->sched_tags) {
2636 ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2637 min(nr, set->queue_depth),
2640 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2648 q->nr_requests = nr;
2650 blk_mq_unfreeze_queue(q);
2651 blk_mq_start_stopped_hw_queues(q, true);
2656 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2658 struct request_queue *q;
2660 lockdep_assert_held(&set->tag_list_lock);
2662 if (nr_hw_queues > nr_cpu_ids)
2663 nr_hw_queues = nr_cpu_ids;
2664 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2667 list_for_each_entry(q, &set->tag_list, tag_set_list)
2668 blk_mq_freeze_queue(q);
2670 set->nr_hw_queues = nr_hw_queues;
2671 blk_mq_update_queue_map(set);
2672 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2673 blk_mq_realloc_hw_ctxs(set, q);
2674 blk_mq_queue_reinit(q, cpu_online_mask);
2677 list_for_each_entry(q, &set->tag_list, tag_set_list)
2678 blk_mq_unfreeze_queue(q);
2680 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2682 /* Enable polling stats and return whether they were already enabled. */
2683 static bool blk_poll_stats_enable(struct request_queue *q)
2685 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2686 test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2688 blk_stat_add_callback(q, q->poll_cb);
2692 static void blk_mq_poll_stats_start(struct request_queue *q)
2695 * We don't arm the callback if polling stats are not enabled or the
2696 * callback is already active.
2698 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2699 blk_stat_is_active(q->poll_cb))
2702 blk_stat_activate_msecs(q->poll_cb, 100);
2705 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2707 struct request_queue *q = cb->data;
2709 if (cb->stat[READ].nr_samples)
2710 q->poll_stat[READ] = cb->stat[READ];
2711 if (cb->stat[WRITE].nr_samples)
2712 q->poll_stat[WRITE] = cb->stat[WRITE];
2715 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2716 struct blk_mq_hw_ctx *hctx,
2719 unsigned long ret = 0;
2722 * If stats collection isn't on, don't sleep but turn it on for
2725 if (!blk_poll_stats_enable(q))
2729 * As an optimistic guess, use half of the mean service time
2730 * for this type of request. We can (and should) make this smarter.
2731 * For instance, if the completion latencies are tight, we can
2732 * get closer than just half the mean. This is especially
2733 * important on devices where the completion latencies are longer
2736 if (req_op(rq) == REQ_OP_READ && q->poll_stat[READ].nr_samples)
2737 ret = (q->poll_stat[READ].mean + 1) / 2;
2738 else if (req_op(rq) == REQ_OP_WRITE && q->poll_stat[WRITE].nr_samples)
2739 ret = (q->poll_stat[WRITE].mean + 1) / 2;
2744 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2745 struct blk_mq_hw_ctx *hctx,
2748 struct hrtimer_sleeper hs;
2749 enum hrtimer_mode mode;
2753 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2759 * -1: don't ever hybrid sleep
2760 * 0: use half of prev avg
2761 * >0: use this specific value
2763 if (q->poll_nsec == -1)
2765 else if (q->poll_nsec > 0)
2766 nsecs = q->poll_nsec;
2768 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2773 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2776 * This will be replaced with the stats tracking code, using
2777 * 'avg_completion_time / 2' as the pre-sleep target.
2781 mode = HRTIMER_MODE_REL;
2782 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2783 hrtimer_set_expires(&hs.timer, kt);
2785 hrtimer_init_sleeper(&hs, current);
2787 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2789 set_current_state(TASK_UNINTERRUPTIBLE);
2790 hrtimer_start_expires(&hs.timer, mode);
2793 hrtimer_cancel(&hs.timer);
2794 mode = HRTIMER_MODE_ABS;
2795 } while (hs.task && !signal_pending(current));
2797 __set_current_state(TASK_RUNNING);
2798 destroy_hrtimer_on_stack(&hs.timer);
2802 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2804 struct request_queue *q = hctx->queue;
2808 * If we sleep, have the caller restart the poll loop to reset
2809 * the state. Like for the other success return cases, the
2810 * caller is responsible for checking if the IO completed. If
2811 * the IO isn't complete, we'll get called again and will go
2812 * straight to the busy poll loop.
2814 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2817 hctx->poll_considered++;
2819 state = current->state;
2820 while (!need_resched()) {
2823 hctx->poll_invoked++;
2825 ret = q->mq_ops->poll(hctx, rq->tag);
2827 hctx->poll_success++;
2828 set_current_state(TASK_RUNNING);
2832 if (signal_pending_state(state, current))
2833 set_current_state(TASK_RUNNING);
2835 if (current->state == TASK_RUNNING)
2845 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2847 struct blk_mq_hw_ctx *hctx;
2848 struct blk_plug *plug;
2851 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2852 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2855 plug = current->plug;
2857 blk_flush_plug_list(plug, false);
2859 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2860 if (!blk_qc_t_is_internal(cookie))
2861 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2863 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2865 return __blk_mq_poll(hctx, rq);
2867 EXPORT_SYMBOL_GPL(blk_mq_poll);
2869 void blk_mq_disable_hotplug(void)
2871 mutex_lock(&all_q_mutex);
2874 void blk_mq_enable_hotplug(void)
2876 mutex_unlock(&all_q_mutex);
2879 static int __init blk_mq_init(void)
2881 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2882 blk_mq_hctx_notify_dead);
2884 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2885 blk_mq_queue_reinit_prepare,
2886 blk_mq_queue_reinit_dead);
2889 subsys_initcall(blk_mq_init);