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/delay.h>
24 #include <linux/crash_dump.h>
25 #include <linux/prefetch.h>
27 #include <trace/events/block.h>
29 #include <linux/blk-mq.h>
32 #include "blk-mq-tag.h"
36 static DEFINE_MUTEX(all_q_mutex);
37 static LIST_HEAD(all_q_list);
40 * Check if any of the ctx's have pending work in this hardware queue
42 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
44 return sbitmap_any_bit_set(&hctx->ctx_map);
48 * Mark this ctx as having pending work in this hardware queue
50 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
51 struct blk_mq_ctx *ctx)
53 if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
54 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
57 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
58 struct blk_mq_ctx *ctx)
60 sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
63 void blk_mq_freeze_queue_start(struct request_queue *q)
67 freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
68 if (freeze_depth == 1) {
69 percpu_ref_kill(&q->q_usage_counter);
70 blk_mq_run_hw_queues(q, false);
73 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
75 static void blk_mq_freeze_queue_wait(struct request_queue *q)
77 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
81 * Guarantee no request is in use, so we can change any data structure of
82 * the queue afterward.
84 void blk_freeze_queue(struct request_queue *q)
87 * In the !blk_mq case we are only calling this to kill the
88 * q_usage_counter, otherwise this increases the freeze depth
89 * and waits for it to return to zero. For this reason there is
90 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
91 * exported to drivers as the only user for unfreeze is blk_mq.
93 blk_mq_freeze_queue_start(q);
94 blk_mq_freeze_queue_wait(q);
97 void blk_mq_freeze_queue(struct request_queue *q)
100 * ...just an alias to keep freeze and unfreeze actions balanced
101 * in the blk_mq_* namespace
105 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
107 void blk_mq_unfreeze_queue(struct request_queue *q)
111 freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
112 WARN_ON_ONCE(freeze_depth < 0);
114 percpu_ref_reinit(&q->q_usage_counter);
115 wake_up_all(&q->mq_freeze_wq);
118 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
121 * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
124 * Note: this function does not prevent that the struct request end_io()
125 * callback function is invoked. Additionally, it is not prevented that
126 * new queue_rq() calls occur unless the queue has been stopped first.
128 void blk_mq_quiesce_queue(struct request_queue *q)
130 struct blk_mq_hw_ctx *hctx;
134 blk_mq_stop_hw_queues(q);
136 queue_for_each_hw_ctx(q, hctx, i) {
137 if (hctx->flags & BLK_MQ_F_BLOCKING)
138 synchronize_srcu(&hctx->queue_rq_srcu);
145 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
147 void blk_mq_wake_waiters(struct request_queue *q)
149 struct blk_mq_hw_ctx *hctx;
152 queue_for_each_hw_ctx(q, hctx, i)
153 if (blk_mq_hw_queue_mapped(hctx))
154 blk_mq_tag_wakeup_all(hctx->tags, true);
157 * If we are called because the queue has now been marked as
158 * dying, we need to ensure that processes currently waiting on
159 * the queue are notified as well.
161 wake_up_all(&q->mq_freeze_wq);
164 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
166 return blk_mq_has_free_tags(hctx->tags);
168 EXPORT_SYMBOL(blk_mq_can_queue);
170 void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
171 struct request *rq, unsigned int op)
173 INIT_LIST_HEAD(&rq->queuelist);
174 /* csd/requeue_work/fifo_time is initialized before use */
178 if (blk_queue_io_stat(q))
179 rq->rq_flags |= RQF_IO_STAT;
180 /* do not touch atomic flags, it needs atomic ops against the timer */
182 INIT_HLIST_NODE(&rq->hash);
183 RB_CLEAR_NODE(&rq->rb_node);
186 rq->start_time = jiffies;
187 #ifdef CONFIG_BLK_CGROUP
189 set_start_time_ns(rq);
190 rq->io_start_time_ns = 0;
192 rq->nr_phys_segments = 0;
193 #if defined(CONFIG_BLK_DEV_INTEGRITY)
194 rq->nr_integrity_segments = 0;
197 /* tag was already set */
207 INIT_LIST_HEAD(&rq->timeout_list);
211 rq->end_io_data = NULL;
214 ctx->rq_dispatched[op_is_sync(op)]++;
216 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init);
218 struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data,
224 tag = blk_mq_get_tag(data);
225 if (tag != BLK_MQ_TAG_FAIL) {
226 rq = data->hctx->tags->static_rqs[tag];
228 if (blk_mq_tag_busy(data->hctx)) {
229 rq->rq_flags = RQF_MQ_INFLIGHT;
230 atomic_inc(&data->hctx->nr_active);
234 data->hctx->tags->rqs[tag] = rq;
235 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op);
241 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request);
243 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
246 struct blk_mq_ctx *ctx;
247 struct blk_mq_hw_ctx *hctx;
249 struct blk_mq_alloc_data alloc_data;
252 ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
256 ctx = blk_mq_get_ctx(q);
257 hctx = blk_mq_map_queue(q, ctx->cpu);
258 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
259 rq = __blk_mq_alloc_request(&alloc_data, rw);
264 return ERR_PTR(-EWOULDBLOCK);
268 rq->__sector = (sector_t) -1;
269 rq->bio = rq->biotail = NULL;
272 EXPORT_SYMBOL(blk_mq_alloc_request);
274 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
275 unsigned int flags, unsigned int hctx_idx)
277 struct blk_mq_hw_ctx *hctx;
278 struct blk_mq_ctx *ctx;
280 struct blk_mq_alloc_data alloc_data;
284 * If the tag allocator sleeps we could get an allocation for a
285 * different hardware context. No need to complicate the low level
286 * allocator for this for the rare use case of a command tied to
289 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
290 return ERR_PTR(-EINVAL);
292 if (hctx_idx >= q->nr_hw_queues)
293 return ERR_PTR(-EIO);
295 ret = blk_queue_enter(q, true);
300 * Check if the hardware context is actually mapped to anything.
301 * If not tell the caller that it should skip this queue.
303 hctx = q->queue_hw_ctx[hctx_idx];
304 if (!blk_mq_hw_queue_mapped(hctx)) {
308 ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask));
310 blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
311 rq = __blk_mq_alloc_request(&alloc_data, rw);
323 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
325 void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
328 const int tag = rq->tag;
329 struct request_queue *q = rq->q;
331 if (rq->rq_flags & RQF_MQ_INFLIGHT)
332 atomic_dec(&hctx->nr_active);
334 wbt_done(q->rq_wb, &rq->issue_stat);
337 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
338 clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
339 blk_mq_put_tag(hctx, hctx->tags, ctx, tag);
343 static void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx,
346 struct blk_mq_ctx *ctx = rq->mq_ctx;
348 ctx->rq_completed[rq_is_sync(rq)]++;
349 __blk_mq_free_request(hctx, ctx, rq);
352 void blk_mq_free_request(struct request *rq)
354 blk_mq_free_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq);
356 EXPORT_SYMBOL_GPL(blk_mq_free_request);
358 inline void __blk_mq_end_request(struct request *rq, int error)
360 blk_account_io_done(rq);
363 wbt_done(rq->q->rq_wb, &rq->issue_stat);
364 rq->end_io(rq, error);
366 if (unlikely(blk_bidi_rq(rq)))
367 blk_mq_free_request(rq->next_rq);
368 blk_mq_free_request(rq);
371 EXPORT_SYMBOL(__blk_mq_end_request);
373 void blk_mq_end_request(struct request *rq, int error)
375 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
377 __blk_mq_end_request(rq, error);
379 EXPORT_SYMBOL(blk_mq_end_request);
381 static void __blk_mq_complete_request_remote(void *data)
383 struct request *rq = data;
385 rq->q->softirq_done_fn(rq);
388 static void blk_mq_ipi_complete_request(struct request *rq)
390 struct blk_mq_ctx *ctx = rq->mq_ctx;
394 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
395 rq->q->softirq_done_fn(rq);
400 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
401 shared = cpus_share_cache(cpu, ctx->cpu);
403 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
404 rq->csd.func = __blk_mq_complete_request_remote;
407 smp_call_function_single_async(ctx->cpu, &rq->csd);
409 rq->q->softirq_done_fn(rq);
414 static void blk_mq_stat_add(struct request *rq)
416 if (rq->rq_flags & RQF_STATS) {
418 * We could rq->mq_ctx here, but there's less of a risk
419 * of races if we have the completion event add the stats
420 * to the local software queue.
422 struct blk_mq_ctx *ctx;
424 ctx = __blk_mq_get_ctx(rq->q, raw_smp_processor_id());
425 blk_stat_add(&ctx->stat[rq_data_dir(rq)], rq);
429 static void __blk_mq_complete_request(struct request *rq)
431 struct request_queue *q = rq->q;
435 if (!q->softirq_done_fn)
436 blk_mq_end_request(rq, rq->errors);
438 blk_mq_ipi_complete_request(rq);
442 * blk_mq_complete_request - end I/O on a request
443 * @rq: the request being processed
446 * Ends all I/O on a request. It does not handle partial completions.
447 * The actual completion happens out-of-order, through a IPI handler.
449 void blk_mq_complete_request(struct request *rq, int error)
451 struct request_queue *q = rq->q;
453 if (unlikely(blk_should_fake_timeout(q)))
455 if (!blk_mark_rq_complete(rq)) {
457 __blk_mq_complete_request(rq);
460 EXPORT_SYMBOL(blk_mq_complete_request);
462 int blk_mq_request_started(struct request *rq)
464 return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
466 EXPORT_SYMBOL_GPL(blk_mq_request_started);
468 void blk_mq_start_request(struct request *rq)
470 struct request_queue *q = rq->q;
472 trace_block_rq_issue(q, rq);
474 rq->resid_len = blk_rq_bytes(rq);
475 if (unlikely(blk_bidi_rq(rq)))
476 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
478 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
479 blk_stat_set_issue_time(&rq->issue_stat);
480 rq->rq_flags |= RQF_STATS;
481 wbt_issue(q->rq_wb, &rq->issue_stat);
487 * Ensure that ->deadline is visible before set the started
488 * flag and clear the completed flag.
490 smp_mb__before_atomic();
493 * Mark us as started and clear complete. Complete might have been
494 * set if requeue raced with timeout, which then marked it as
495 * complete. So be sure to clear complete again when we start
496 * the request, otherwise we'll ignore the completion event.
498 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
499 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
500 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
501 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
503 if (q->dma_drain_size && blk_rq_bytes(rq)) {
505 * Make sure space for the drain appears. We know we can do
506 * this because max_hw_segments has been adjusted to be one
507 * fewer than the device can handle.
509 rq->nr_phys_segments++;
512 EXPORT_SYMBOL(blk_mq_start_request);
514 static void __blk_mq_requeue_request(struct request *rq)
516 struct request_queue *q = rq->q;
518 trace_block_rq_requeue(q, rq);
519 wbt_requeue(q->rq_wb, &rq->issue_stat);
521 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
522 if (q->dma_drain_size && blk_rq_bytes(rq))
523 rq->nr_phys_segments--;
527 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
529 __blk_mq_requeue_request(rq);
531 BUG_ON(blk_queued_rq(rq));
532 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
534 EXPORT_SYMBOL(blk_mq_requeue_request);
536 static void blk_mq_requeue_work(struct work_struct *work)
538 struct request_queue *q =
539 container_of(work, struct request_queue, requeue_work.work);
541 struct request *rq, *next;
544 spin_lock_irqsave(&q->requeue_lock, flags);
545 list_splice_init(&q->requeue_list, &rq_list);
546 spin_unlock_irqrestore(&q->requeue_lock, flags);
548 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
549 if (!(rq->rq_flags & RQF_SOFTBARRIER))
552 rq->rq_flags &= ~RQF_SOFTBARRIER;
553 list_del_init(&rq->queuelist);
554 blk_mq_insert_request(rq, true, false, false);
557 while (!list_empty(&rq_list)) {
558 rq = list_entry(rq_list.next, struct request, queuelist);
559 list_del_init(&rq->queuelist);
560 blk_mq_insert_request(rq, false, false, false);
563 blk_mq_run_hw_queues(q, false);
566 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
567 bool kick_requeue_list)
569 struct request_queue *q = rq->q;
573 * We abuse this flag that is otherwise used by the I/O scheduler to
574 * request head insertation from the workqueue.
576 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
578 spin_lock_irqsave(&q->requeue_lock, flags);
580 rq->rq_flags |= RQF_SOFTBARRIER;
581 list_add(&rq->queuelist, &q->requeue_list);
583 list_add_tail(&rq->queuelist, &q->requeue_list);
585 spin_unlock_irqrestore(&q->requeue_lock, flags);
587 if (kick_requeue_list)
588 blk_mq_kick_requeue_list(q);
590 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
592 void blk_mq_kick_requeue_list(struct request_queue *q)
594 kblockd_schedule_delayed_work(&q->requeue_work, 0);
596 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
598 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
601 kblockd_schedule_delayed_work(&q->requeue_work,
602 msecs_to_jiffies(msecs));
604 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
606 void blk_mq_abort_requeue_list(struct request_queue *q)
611 spin_lock_irqsave(&q->requeue_lock, flags);
612 list_splice_init(&q->requeue_list, &rq_list);
613 spin_unlock_irqrestore(&q->requeue_lock, flags);
615 while (!list_empty(&rq_list)) {
618 rq = list_first_entry(&rq_list, struct request, queuelist);
619 list_del_init(&rq->queuelist);
621 blk_mq_end_request(rq, rq->errors);
624 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
626 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
628 if (tag < tags->nr_tags) {
629 prefetch(tags->rqs[tag]);
630 return tags->rqs[tag];
635 EXPORT_SYMBOL(blk_mq_tag_to_rq);
637 struct blk_mq_timeout_data {
639 unsigned int next_set;
642 void blk_mq_rq_timed_out(struct request *req, bool reserved)
644 const struct blk_mq_ops *ops = req->q->mq_ops;
645 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
648 * We know that complete is set at this point. If STARTED isn't set
649 * anymore, then the request isn't active and the "timeout" should
650 * just be ignored. This can happen due to the bitflag ordering.
651 * Timeout first checks if STARTED is set, and if it is, assumes
652 * the request is active. But if we race with completion, then
653 * we both flags will get cleared. So check here again, and ignore
654 * a timeout event with a request that isn't active.
656 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
660 ret = ops->timeout(req, reserved);
664 __blk_mq_complete_request(req);
666 case BLK_EH_RESET_TIMER:
668 blk_clear_rq_complete(req);
670 case BLK_EH_NOT_HANDLED:
673 printk(KERN_ERR "block: bad eh return: %d\n", ret);
678 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
679 struct request *rq, void *priv, bool reserved)
681 struct blk_mq_timeout_data *data = priv;
683 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
685 * If a request wasn't started before the queue was
686 * marked dying, kill it here or it'll go unnoticed.
688 if (unlikely(blk_queue_dying(rq->q))) {
690 blk_mq_end_request(rq, rq->errors);
695 if (time_after_eq(jiffies, rq->deadline)) {
696 if (!blk_mark_rq_complete(rq))
697 blk_mq_rq_timed_out(rq, reserved);
698 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
699 data->next = rq->deadline;
704 static void blk_mq_timeout_work(struct work_struct *work)
706 struct request_queue *q =
707 container_of(work, struct request_queue, timeout_work);
708 struct blk_mq_timeout_data data = {
714 /* A deadlock might occur if a request is stuck requiring a
715 * timeout at the same time a queue freeze is waiting
716 * completion, since the timeout code would not be able to
717 * acquire the queue reference here.
719 * That's why we don't use blk_queue_enter here; instead, we use
720 * percpu_ref_tryget directly, because we need to be able to
721 * obtain a reference even in the short window between the queue
722 * starting to freeze, by dropping the first reference in
723 * blk_mq_freeze_queue_start, and the moment the last request is
724 * consumed, marked by the instant q_usage_counter reaches
727 if (!percpu_ref_tryget(&q->q_usage_counter))
730 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
733 data.next = blk_rq_timeout(round_jiffies_up(data.next));
734 mod_timer(&q->timeout, data.next);
736 struct blk_mq_hw_ctx *hctx;
738 queue_for_each_hw_ctx(q, hctx, i) {
739 /* the hctx may be unmapped, so check it here */
740 if (blk_mq_hw_queue_mapped(hctx))
741 blk_mq_tag_idle(hctx);
748 * Reverse check our software queue for entries that we could potentially
749 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
750 * too much time checking for merges.
752 static bool blk_mq_attempt_merge(struct request_queue *q,
753 struct blk_mq_ctx *ctx, struct bio *bio)
758 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
764 if (!blk_rq_merge_ok(rq, bio))
767 el_ret = blk_try_merge(rq, bio);
768 if (el_ret == ELEVATOR_BACK_MERGE) {
769 if (bio_attempt_back_merge(q, rq, bio)) {
774 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
775 if (bio_attempt_front_merge(q, rq, bio)) {
786 struct flush_busy_ctx_data {
787 struct blk_mq_hw_ctx *hctx;
788 struct list_head *list;
791 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
793 struct flush_busy_ctx_data *flush_data = data;
794 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
795 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
797 sbitmap_clear_bit(sb, bitnr);
798 spin_lock(&ctx->lock);
799 list_splice_tail_init(&ctx->rq_list, flush_data->list);
800 spin_unlock(&ctx->lock);
805 * Process software queues that have been marked busy, splicing them
806 * to the for-dispatch
808 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
810 struct flush_busy_ctx_data data = {
815 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
817 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
819 static inline unsigned int queued_to_index(unsigned int queued)
824 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
827 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list)
829 struct request_queue *q = hctx->queue;
831 LIST_HEAD(driver_list);
832 struct list_head *dptr;
833 int queued, ret = BLK_MQ_RQ_QUEUE_OK;
836 * Start off with dptr being NULL, so we start the first request
837 * immediately, even if we have more pending.
842 * Now process all the entries, sending them to the driver.
845 while (!list_empty(list)) {
846 struct blk_mq_queue_data bd;
848 rq = list_first_entry(list, struct request, queuelist);
849 list_del_init(&rq->queuelist);
853 bd.last = list_empty(list);
855 ret = q->mq_ops->queue_rq(hctx, &bd);
857 case BLK_MQ_RQ_QUEUE_OK:
860 case BLK_MQ_RQ_QUEUE_BUSY:
861 list_add(&rq->queuelist, list);
862 __blk_mq_requeue_request(rq);
865 pr_err("blk-mq: bad return on queue: %d\n", ret);
866 case BLK_MQ_RQ_QUEUE_ERROR:
868 blk_mq_end_request(rq, rq->errors);
872 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
876 * We've done the first request. If we have more than 1
877 * left in the list, set dptr to defer issue.
879 if (!dptr && list->next != list->prev)
883 hctx->dispatched[queued_to_index(queued)]++;
886 * Any items that need requeuing? Stuff them into hctx->dispatch,
887 * that is where we will continue on next queue run.
889 if (!list_empty(list)) {
890 spin_lock(&hctx->lock);
891 list_splice(list, &hctx->dispatch);
892 spin_unlock(&hctx->lock);
895 * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
896 * it's possible the queue is stopped and restarted again
897 * before this. Queue restart will dispatch requests. And since
898 * requests in rq_list aren't added into hctx->dispatch yet,
899 * the requests in rq_list might get lost.
901 * blk_mq_run_hw_queue() already checks the STOPPED bit
903 blk_mq_run_hw_queue(hctx, true);
906 return ret != BLK_MQ_RQ_QUEUE_BUSY;
910 * Run this hardware queue, pulling any software queues mapped to it in.
911 * Note that this function currently has various problems around ordering
912 * of IO. In particular, we'd like FIFO behaviour on handling existing
913 * items on the hctx->dispatch list. Ignore that for now.
915 static void blk_mq_process_rq_list(struct blk_mq_hw_ctx *hctx)
918 LIST_HEAD(driver_list);
920 if (unlikely(blk_mq_hctx_stopped(hctx)))
926 * Touch any software queue that has pending entries.
928 blk_mq_flush_busy_ctxs(hctx, &rq_list);
931 * If we have previous entries on our dispatch list, grab them
932 * and stuff them at the front for more fair dispatch.
934 if (!list_empty_careful(&hctx->dispatch)) {
935 spin_lock(&hctx->lock);
936 if (!list_empty(&hctx->dispatch))
937 list_splice_init(&hctx->dispatch, &rq_list);
938 spin_unlock(&hctx->lock);
941 blk_mq_dispatch_rq_list(hctx, &rq_list);
944 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
948 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
949 cpu_online(hctx->next_cpu));
951 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
953 blk_mq_process_rq_list(hctx);
956 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
957 blk_mq_process_rq_list(hctx);
958 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
963 * It'd be great if the workqueue API had a way to pass
964 * in a mask and had some smarts for more clever placement.
965 * For now we just round-robin here, switching for every
966 * BLK_MQ_CPU_WORK_BATCH queued items.
968 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
970 if (hctx->queue->nr_hw_queues == 1)
971 return WORK_CPU_UNBOUND;
973 if (--hctx->next_cpu_batch <= 0) {
976 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
977 if (next_cpu >= nr_cpu_ids)
978 next_cpu = cpumask_first(hctx->cpumask);
980 hctx->next_cpu = next_cpu;
981 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
984 return hctx->next_cpu;
987 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
989 if (unlikely(blk_mq_hctx_stopped(hctx) ||
990 !blk_mq_hw_queue_mapped(hctx)))
993 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
995 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
996 __blk_mq_run_hw_queue(hctx);
1004 kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work);
1007 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1009 struct blk_mq_hw_ctx *hctx;
1012 queue_for_each_hw_ctx(q, hctx, i) {
1013 if ((!blk_mq_hctx_has_pending(hctx) &&
1014 list_empty_careful(&hctx->dispatch)) ||
1015 blk_mq_hctx_stopped(hctx))
1018 blk_mq_run_hw_queue(hctx, async);
1021 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1024 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1025 * @q: request queue.
1027 * The caller is responsible for serializing this function against
1028 * blk_mq_{start,stop}_hw_queue().
1030 bool blk_mq_queue_stopped(struct request_queue *q)
1032 struct blk_mq_hw_ctx *hctx;
1035 queue_for_each_hw_ctx(q, hctx, i)
1036 if (blk_mq_hctx_stopped(hctx))
1041 EXPORT_SYMBOL(blk_mq_queue_stopped);
1043 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1045 cancel_work(&hctx->run_work);
1046 cancel_delayed_work(&hctx->delay_work);
1047 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1049 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1051 void blk_mq_stop_hw_queues(struct request_queue *q)
1053 struct blk_mq_hw_ctx *hctx;
1056 queue_for_each_hw_ctx(q, hctx, i)
1057 blk_mq_stop_hw_queue(hctx);
1059 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1061 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1063 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1065 blk_mq_run_hw_queue(hctx, false);
1067 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1069 void blk_mq_start_hw_queues(struct request_queue *q)
1071 struct blk_mq_hw_ctx *hctx;
1074 queue_for_each_hw_ctx(q, hctx, i)
1075 blk_mq_start_hw_queue(hctx);
1077 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1079 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1081 if (!blk_mq_hctx_stopped(hctx))
1084 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1085 blk_mq_run_hw_queue(hctx, async);
1087 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1089 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1091 struct blk_mq_hw_ctx *hctx;
1094 queue_for_each_hw_ctx(q, hctx, i)
1095 blk_mq_start_stopped_hw_queue(hctx, async);
1097 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1099 static void blk_mq_run_work_fn(struct work_struct *work)
1101 struct blk_mq_hw_ctx *hctx;
1103 hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
1105 __blk_mq_run_hw_queue(hctx);
1108 static void blk_mq_delay_work_fn(struct work_struct *work)
1110 struct blk_mq_hw_ctx *hctx;
1112 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1114 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1115 __blk_mq_run_hw_queue(hctx);
1118 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1120 if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1123 kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1124 &hctx->delay_work, msecs_to_jiffies(msecs));
1126 EXPORT_SYMBOL(blk_mq_delay_queue);
1128 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1132 struct blk_mq_ctx *ctx = rq->mq_ctx;
1134 trace_block_rq_insert(hctx->queue, rq);
1137 list_add(&rq->queuelist, &ctx->rq_list);
1139 list_add_tail(&rq->queuelist, &ctx->rq_list);
1142 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1145 struct blk_mq_ctx *ctx = rq->mq_ctx;
1147 __blk_mq_insert_req_list(hctx, rq, at_head);
1148 blk_mq_hctx_mark_pending(hctx, ctx);
1151 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1154 struct blk_mq_ctx *ctx = rq->mq_ctx;
1155 struct request_queue *q = rq->q;
1156 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
1158 spin_lock(&ctx->lock);
1159 __blk_mq_insert_request(hctx, rq, at_head);
1160 spin_unlock(&ctx->lock);
1163 blk_mq_run_hw_queue(hctx, async);
1166 static void blk_mq_insert_requests(struct request_queue *q,
1167 struct blk_mq_ctx *ctx,
1168 struct list_head *list,
1173 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
1175 trace_block_unplug(q, depth, !from_schedule);
1178 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1181 spin_lock(&ctx->lock);
1182 while (!list_empty(list)) {
1185 rq = list_first_entry(list, struct request, queuelist);
1186 BUG_ON(rq->mq_ctx != ctx);
1187 list_del_init(&rq->queuelist);
1188 __blk_mq_insert_req_list(hctx, rq, false);
1190 blk_mq_hctx_mark_pending(hctx, ctx);
1191 spin_unlock(&ctx->lock);
1193 blk_mq_run_hw_queue(hctx, from_schedule);
1196 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1198 struct request *rqa = container_of(a, struct request, queuelist);
1199 struct request *rqb = container_of(b, struct request, queuelist);
1201 return !(rqa->mq_ctx < rqb->mq_ctx ||
1202 (rqa->mq_ctx == rqb->mq_ctx &&
1203 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1206 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1208 struct blk_mq_ctx *this_ctx;
1209 struct request_queue *this_q;
1212 LIST_HEAD(ctx_list);
1215 list_splice_init(&plug->mq_list, &list);
1217 list_sort(NULL, &list, plug_ctx_cmp);
1223 while (!list_empty(&list)) {
1224 rq = list_entry_rq(list.next);
1225 list_del_init(&rq->queuelist);
1227 if (rq->mq_ctx != this_ctx) {
1229 blk_mq_insert_requests(this_q, this_ctx,
1234 this_ctx = rq->mq_ctx;
1240 list_add_tail(&rq->queuelist, &ctx_list);
1244 * If 'this_ctx' is set, we know we have entries to complete
1245 * on 'ctx_list'. Do those.
1248 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1253 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1255 init_request_from_bio(rq, bio);
1257 blk_account_io_start(rq, true);
1260 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1262 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1263 !blk_queue_nomerges(hctx->queue);
1266 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1267 struct blk_mq_ctx *ctx,
1268 struct request *rq, struct bio *bio)
1270 if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1271 blk_mq_bio_to_request(rq, bio);
1272 spin_lock(&ctx->lock);
1274 __blk_mq_insert_request(hctx, rq, false);
1275 spin_unlock(&ctx->lock);
1278 struct request_queue *q = hctx->queue;
1280 spin_lock(&ctx->lock);
1281 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1282 blk_mq_bio_to_request(rq, bio);
1286 spin_unlock(&ctx->lock);
1287 __blk_mq_free_request(hctx, ctx, rq);
1292 static struct request *blk_mq_map_request(struct request_queue *q,
1294 struct blk_mq_alloc_data *data)
1296 struct blk_mq_hw_ctx *hctx;
1297 struct blk_mq_ctx *ctx;
1300 blk_queue_enter_live(q);
1301 ctx = blk_mq_get_ctx(q);
1302 hctx = blk_mq_map_queue(q, ctx->cpu);
1304 trace_block_getrq(q, bio, bio->bi_opf);
1305 blk_mq_set_alloc_data(data, q, 0, ctx, hctx);
1306 rq = __blk_mq_alloc_request(data, bio->bi_opf);
1308 data->hctx->queued++;
1312 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1314 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1317 static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie)
1320 struct request_queue *q = rq->q;
1321 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
1322 struct blk_mq_queue_data bd = {
1327 blk_qc_t new_cookie = request_to_qc_t(hctx, rq);
1329 if (blk_mq_hctx_stopped(hctx))
1333 * For OK queue, we are done. For error, kill it. Any other
1334 * error (busy), just add it to our list as we previously
1337 ret = q->mq_ops->queue_rq(hctx, &bd);
1338 if (ret == BLK_MQ_RQ_QUEUE_OK) {
1339 *cookie = new_cookie;
1343 __blk_mq_requeue_request(rq);
1345 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1346 *cookie = BLK_QC_T_NONE;
1348 blk_mq_end_request(rq, rq->errors);
1353 blk_mq_insert_request(rq, false, true, true);
1357 * Multiple hardware queue variant. This will not use per-process plugs,
1358 * but will attempt to bypass the hctx queueing if we can go straight to
1359 * hardware for SYNC IO.
1361 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1363 const int is_sync = op_is_sync(bio->bi_opf);
1364 const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1365 struct blk_mq_alloc_data data;
1367 unsigned int request_count = 0, srcu_idx;
1368 struct blk_plug *plug;
1369 struct request *same_queue_rq = NULL;
1371 unsigned int wb_acct;
1373 blk_queue_bounce(q, &bio);
1375 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1377 return BLK_QC_T_NONE;
1380 blk_queue_split(q, &bio, q->bio_split);
1382 if (!is_flush_fua && !blk_queue_nomerges(q) &&
1383 blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1384 return BLK_QC_T_NONE;
1386 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1388 rq = blk_mq_map_request(q, bio, &data);
1389 if (unlikely(!rq)) {
1390 __wbt_done(q->rq_wb, wb_acct);
1391 return BLK_QC_T_NONE;
1394 wbt_track(&rq->issue_stat, wb_acct);
1396 cookie = request_to_qc_t(data.hctx, rq);
1398 if (unlikely(is_flush_fua)) {
1399 blk_mq_bio_to_request(rq, bio);
1400 blk_insert_flush(rq);
1404 plug = current->plug;
1406 * If the driver supports defer issued based on 'last', then
1407 * queue it up like normal since we can potentially save some
1410 if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1411 !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1412 struct request *old_rq = NULL;
1414 blk_mq_bio_to_request(rq, bio);
1417 * We do limited plugging. If the bio can be merged, do that.
1418 * Otherwise the existing request in the plug list will be
1419 * issued. So the plug list will have one request at most
1423 * The plug list might get flushed before this. If that
1424 * happens, same_queue_rq is invalid and plug list is
1427 if (same_queue_rq && !list_empty(&plug->mq_list)) {
1428 old_rq = same_queue_rq;
1429 list_del_init(&old_rq->queuelist);
1431 list_add_tail(&rq->queuelist, &plug->mq_list);
1432 } else /* is_sync */
1434 blk_mq_put_ctx(data.ctx);
1438 if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) {
1440 blk_mq_try_issue_directly(old_rq, &cookie);
1443 srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu);
1444 blk_mq_try_issue_directly(old_rq, &cookie);
1445 srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx);
1450 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1452 * For a SYNC request, send it to the hardware immediately. For
1453 * an ASYNC request, just ensure that we run it later on. The
1454 * latter allows for merging opportunities and more efficient
1458 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1460 blk_mq_put_ctx(data.ctx);
1466 * Single hardware queue variant. This will attempt to use any per-process
1467 * plug for merging and IO deferral.
1469 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1471 const int is_sync = op_is_sync(bio->bi_opf);
1472 const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1473 struct blk_plug *plug;
1474 unsigned int request_count = 0;
1475 struct blk_mq_alloc_data data;
1478 unsigned int wb_acct;
1480 blk_queue_bounce(q, &bio);
1482 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1484 return BLK_QC_T_NONE;
1487 blk_queue_split(q, &bio, q->bio_split);
1489 if (!is_flush_fua && !blk_queue_nomerges(q)) {
1490 if (blk_attempt_plug_merge(q, bio, &request_count, NULL))
1491 return BLK_QC_T_NONE;
1493 request_count = blk_plug_queued_count(q);
1495 wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1497 rq = blk_mq_map_request(q, bio, &data);
1498 if (unlikely(!rq)) {
1499 __wbt_done(q->rq_wb, wb_acct);
1500 return BLK_QC_T_NONE;
1503 wbt_track(&rq->issue_stat, wb_acct);
1505 cookie = request_to_qc_t(data.hctx, rq);
1507 if (unlikely(is_flush_fua)) {
1508 blk_mq_bio_to_request(rq, bio);
1509 blk_insert_flush(rq);
1514 * A task plug currently exists. Since this is completely lockless,
1515 * utilize that to temporarily store requests until the task is
1516 * either done or scheduled away.
1518 plug = current->plug;
1520 struct request *last = NULL;
1522 blk_mq_bio_to_request(rq, bio);
1525 * @request_count may become stale because of schedule
1526 * out, so check the list again.
1528 if (list_empty(&plug->mq_list))
1531 trace_block_plug(q);
1533 last = list_entry_rq(plug->mq_list.prev);
1535 blk_mq_put_ctx(data.ctx);
1537 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1538 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1539 blk_flush_plug_list(plug, false);
1540 trace_block_plug(q);
1543 list_add_tail(&rq->queuelist, &plug->mq_list);
1547 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1549 * For a SYNC request, send it to the hardware immediately. For
1550 * an ASYNC request, just ensure that we run it later on. The
1551 * latter allows for merging opportunities and more efficient
1555 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1558 blk_mq_put_ctx(data.ctx);
1562 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1563 unsigned int hctx_idx)
1567 if (tags->rqs && set->ops->exit_request) {
1570 for (i = 0; i < tags->nr_tags; i++) {
1571 struct request *rq = tags->static_rqs[i];
1575 set->ops->exit_request(set->driver_data, rq,
1577 tags->static_rqs[i] = NULL;
1581 while (!list_empty(&tags->page_list)) {
1582 page = list_first_entry(&tags->page_list, struct page, lru);
1583 list_del_init(&page->lru);
1585 * Remove kmemleak object previously allocated in
1586 * blk_mq_init_rq_map().
1588 kmemleak_free(page_address(page));
1589 __free_pages(page, page->private);
1593 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1597 kfree(tags->static_rqs);
1598 tags->static_rqs = NULL;
1600 blk_mq_free_tags(tags);
1603 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1604 unsigned int hctx_idx,
1605 unsigned int nr_tags,
1606 unsigned int reserved_tags)
1608 struct blk_mq_tags *tags;
1610 tags = blk_mq_init_tags(nr_tags, reserved_tags,
1612 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1616 tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1617 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1620 blk_mq_free_tags(tags);
1624 tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1625 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1627 if (!tags->static_rqs) {
1629 blk_mq_free_tags(tags);
1636 static size_t order_to_size(unsigned int order)
1638 return (size_t)PAGE_SIZE << order;
1641 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1642 unsigned int hctx_idx, unsigned int depth)
1644 unsigned int i, j, entries_per_page, max_order = 4;
1645 size_t rq_size, left;
1647 INIT_LIST_HEAD(&tags->page_list);
1650 * rq_size is the size of the request plus driver payload, rounded
1651 * to the cacheline size
1653 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1655 left = rq_size * depth;
1657 for (i = 0; i < depth; ) {
1658 int this_order = max_order;
1663 while (this_order && left < order_to_size(this_order - 1))
1667 page = alloc_pages_node(set->numa_node,
1668 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1674 if (order_to_size(this_order) < rq_size)
1681 page->private = this_order;
1682 list_add_tail(&page->lru, &tags->page_list);
1684 p = page_address(page);
1686 * Allow kmemleak to scan these pages as they contain pointers
1687 * to additional allocations like via ops->init_request().
1689 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1690 entries_per_page = order_to_size(this_order) / rq_size;
1691 to_do = min(entries_per_page, depth - i);
1692 left -= to_do * rq_size;
1693 for (j = 0; j < to_do; j++) {
1694 struct request *rq = p;
1696 tags->static_rqs[i] = rq;
1697 if (set->ops->init_request) {
1698 if (set->ops->init_request(set->driver_data,
1701 tags->static_rqs[i] = NULL;
1713 blk_mq_free_rqs(set, tags, hctx_idx);
1718 * 'cpu' is going away. splice any existing rq_list entries from this
1719 * software queue to the hw queue dispatch list, and ensure that it
1722 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1724 struct blk_mq_hw_ctx *hctx;
1725 struct blk_mq_ctx *ctx;
1728 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1729 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1731 spin_lock(&ctx->lock);
1732 if (!list_empty(&ctx->rq_list)) {
1733 list_splice_init(&ctx->rq_list, &tmp);
1734 blk_mq_hctx_clear_pending(hctx, ctx);
1736 spin_unlock(&ctx->lock);
1738 if (list_empty(&tmp))
1741 spin_lock(&hctx->lock);
1742 list_splice_tail_init(&tmp, &hctx->dispatch);
1743 spin_unlock(&hctx->lock);
1745 blk_mq_run_hw_queue(hctx, true);
1749 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1751 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1755 /* hctx->ctxs will be freed in queue's release handler */
1756 static void blk_mq_exit_hctx(struct request_queue *q,
1757 struct blk_mq_tag_set *set,
1758 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1760 unsigned flush_start_tag = set->queue_depth;
1762 blk_mq_tag_idle(hctx);
1764 if (set->ops->exit_request)
1765 set->ops->exit_request(set->driver_data,
1766 hctx->fq->flush_rq, hctx_idx,
1767 flush_start_tag + hctx_idx);
1769 if (set->ops->exit_hctx)
1770 set->ops->exit_hctx(hctx, hctx_idx);
1772 if (hctx->flags & BLK_MQ_F_BLOCKING)
1773 cleanup_srcu_struct(&hctx->queue_rq_srcu);
1775 blk_mq_remove_cpuhp(hctx);
1776 blk_free_flush_queue(hctx->fq);
1777 sbitmap_free(&hctx->ctx_map);
1780 static void blk_mq_exit_hw_queues(struct request_queue *q,
1781 struct blk_mq_tag_set *set, int nr_queue)
1783 struct blk_mq_hw_ctx *hctx;
1786 queue_for_each_hw_ctx(q, hctx, i) {
1789 blk_mq_exit_hctx(q, set, hctx, i);
1793 static void blk_mq_free_hw_queues(struct request_queue *q,
1794 struct blk_mq_tag_set *set)
1796 struct blk_mq_hw_ctx *hctx;
1799 queue_for_each_hw_ctx(q, hctx, i)
1800 free_cpumask_var(hctx->cpumask);
1803 static int blk_mq_init_hctx(struct request_queue *q,
1804 struct blk_mq_tag_set *set,
1805 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1808 unsigned flush_start_tag = set->queue_depth;
1810 node = hctx->numa_node;
1811 if (node == NUMA_NO_NODE)
1812 node = hctx->numa_node = set->numa_node;
1814 INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1815 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1816 spin_lock_init(&hctx->lock);
1817 INIT_LIST_HEAD(&hctx->dispatch);
1819 hctx->queue_num = hctx_idx;
1820 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1822 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1824 hctx->tags = set->tags[hctx_idx];
1827 * Allocate space for all possible cpus to avoid allocation at
1830 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1833 goto unregister_cpu_notifier;
1835 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1841 if (set->ops->init_hctx &&
1842 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1845 hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1849 if (set->ops->init_request &&
1850 set->ops->init_request(set->driver_data,
1851 hctx->fq->flush_rq, hctx_idx,
1852 flush_start_tag + hctx_idx, node))
1855 if (hctx->flags & BLK_MQ_F_BLOCKING)
1856 init_srcu_struct(&hctx->queue_rq_srcu);
1863 if (set->ops->exit_hctx)
1864 set->ops->exit_hctx(hctx, hctx_idx);
1866 sbitmap_free(&hctx->ctx_map);
1869 unregister_cpu_notifier:
1870 blk_mq_remove_cpuhp(hctx);
1874 static void blk_mq_init_cpu_queues(struct request_queue *q,
1875 unsigned int nr_hw_queues)
1879 for_each_possible_cpu(i) {
1880 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1881 struct blk_mq_hw_ctx *hctx;
1883 memset(__ctx, 0, sizeof(*__ctx));
1885 spin_lock_init(&__ctx->lock);
1886 INIT_LIST_HEAD(&__ctx->rq_list);
1888 blk_stat_init(&__ctx->stat[BLK_STAT_READ]);
1889 blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]);
1891 /* If the cpu isn't online, the cpu is mapped to first hctx */
1895 hctx = blk_mq_map_queue(q, i);
1898 * Set local node, IFF we have more than one hw queue. If
1899 * not, we remain on the home node of the device
1901 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1902 hctx->numa_node = local_memory_node(cpu_to_node(i));
1906 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
1910 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
1911 set->queue_depth, set->reserved_tags);
1912 if (!set->tags[hctx_idx])
1915 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
1920 blk_mq_free_rq_map(set->tags[hctx_idx]);
1921 set->tags[hctx_idx] = NULL;
1925 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
1926 unsigned int hctx_idx)
1928 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
1929 blk_mq_free_rq_map(set->tags[hctx_idx]);
1930 set->tags[hctx_idx] = NULL;
1933 static void blk_mq_map_swqueue(struct request_queue *q,
1934 const struct cpumask *online_mask)
1936 unsigned int i, hctx_idx;
1937 struct blk_mq_hw_ctx *hctx;
1938 struct blk_mq_ctx *ctx;
1939 struct blk_mq_tag_set *set = q->tag_set;
1942 * Avoid others reading imcomplete hctx->cpumask through sysfs
1944 mutex_lock(&q->sysfs_lock);
1946 queue_for_each_hw_ctx(q, hctx, i) {
1947 cpumask_clear(hctx->cpumask);
1952 * Map software to hardware queues
1954 for_each_possible_cpu(i) {
1955 /* If the cpu isn't online, the cpu is mapped to first hctx */
1956 if (!cpumask_test_cpu(i, online_mask))
1959 hctx_idx = q->mq_map[i];
1960 /* unmapped hw queue can be remapped after CPU topo changed */
1961 if (!set->tags[hctx_idx] &&
1962 !__blk_mq_alloc_rq_map(set, hctx_idx)) {
1964 * If tags initialization fail for some hctx,
1965 * that hctx won't be brought online. In this
1966 * case, remap the current ctx to hctx[0] which
1967 * is guaranteed to always have tags allocated
1972 ctx = per_cpu_ptr(q->queue_ctx, i);
1973 hctx = blk_mq_map_queue(q, i);
1975 cpumask_set_cpu(i, hctx->cpumask);
1976 ctx->index_hw = hctx->nr_ctx;
1977 hctx->ctxs[hctx->nr_ctx++] = ctx;
1980 mutex_unlock(&q->sysfs_lock);
1982 queue_for_each_hw_ctx(q, hctx, i) {
1984 * If no software queues are mapped to this hardware queue,
1985 * disable it and free the request entries.
1987 if (!hctx->nr_ctx) {
1988 /* Never unmap queue 0. We need it as a
1989 * fallback in case of a new remap fails
1992 if (i && set->tags[i])
1993 blk_mq_free_map_and_requests(set, i);
1999 hctx->tags = set->tags[i];
2000 WARN_ON(!hctx->tags);
2003 * Set the map size to the number of mapped software queues.
2004 * This is more accurate and more efficient than looping
2005 * over all possibly mapped software queues.
2007 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2010 * Initialize batch roundrobin counts
2012 hctx->next_cpu = cpumask_first(hctx->cpumask);
2013 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2017 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2019 struct blk_mq_hw_ctx *hctx;
2022 queue_for_each_hw_ctx(q, hctx, i) {
2024 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2026 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2030 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2032 struct request_queue *q;
2034 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2035 blk_mq_freeze_queue(q);
2036 queue_set_hctx_shared(q, shared);
2037 blk_mq_unfreeze_queue(q);
2041 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2043 struct blk_mq_tag_set *set = q->tag_set;
2045 mutex_lock(&set->tag_list_lock);
2046 list_del_init(&q->tag_set_list);
2047 if (list_is_singular(&set->tag_list)) {
2048 /* just transitioned to unshared */
2049 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2050 /* update existing queue */
2051 blk_mq_update_tag_set_depth(set, false);
2053 mutex_unlock(&set->tag_list_lock);
2056 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2057 struct request_queue *q)
2061 mutex_lock(&set->tag_list_lock);
2063 /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2064 if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2065 set->flags |= BLK_MQ_F_TAG_SHARED;
2066 /* update existing queue */
2067 blk_mq_update_tag_set_depth(set, true);
2069 if (set->flags & BLK_MQ_F_TAG_SHARED)
2070 queue_set_hctx_shared(q, true);
2071 list_add_tail(&q->tag_set_list, &set->tag_list);
2073 mutex_unlock(&set->tag_list_lock);
2077 * It is the actual release handler for mq, but we do it from
2078 * request queue's release handler for avoiding use-after-free
2079 * and headache because q->mq_kobj shouldn't have been introduced,
2080 * but we can't group ctx/kctx kobj without it.
2082 void blk_mq_release(struct request_queue *q)
2084 struct blk_mq_hw_ctx *hctx;
2087 /* hctx kobj stays in hctx */
2088 queue_for_each_hw_ctx(q, hctx, i) {
2097 kfree(q->queue_hw_ctx);
2099 /* ctx kobj stays in queue_ctx */
2100 free_percpu(q->queue_ctx);
2103 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2105 struct request_queue *uninit_q, *q;
2107 uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2109 return ERR_PTR(-ENOMEM);
2111 q = blk_mq_init_allocated_queue(set, uninit_q);
2113 blk_cleanup_queue(uninit_q);
2117 EXPORT_SYMBOL(blk_mq_init_queue);
2119 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2120 struct request_queue *q)
2123 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2125 blk_mq_sysfs_unregister(q);
2126 for (i = 0; i < set->nr_hw_queues; i++) {
2132 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2133 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2138 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2145 atomic_set(&hctxs[i]->nr_active, 0);
2146 hctxs[i]->numa_node = node;
2147 hctxs[i]->queue_num = i;
2149 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2150 free_cpumask_var(hctxs[i]->cpumask);
2155 blk_mq_hctx_kobj_init(hctxs[i]);
2157 for (j = i; j < q->nr_hw_queues; j++) {
2158 struct blk_mq_hw_ctx *hctx = hctxs[j];
2162 blk_mq_free_map_and_requests(set, j);
2163 blk_mq_exit_hctx(q, set, hctx, j);
2164 free_cpumask_var(hctx->cpumask);
2165 kobject_put(&hctx->kobj);
2172 q->nr_hw_queues = i;
2173 blk_mq_sysfs_register(q);
2176 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2177 struct request_queue *q)
2179 /* mark the queue as mq asap */
2180 q->mq_ops = set->ops;
2182 q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2186 q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2187 GFP_KERNEL, set->numa_node);
2188 if (!q->queue_hw_ctx)
2191 q->mq_map = set->mq_map;
2193 blk_mq_realloc_hw_ctxs(set, q);
2194 if (!q->nr_hw_queues)
2197 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2198 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2200 q->nr_queues = nr_cpu_ids;
2202 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2204 if (!(set->flags & BLK_MQ_F_SG_MERGE))
2205 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2207 q->sg_reserved_size = INT_MAX;
2209 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2210 INIT_LIST_HEAD(&q->requeue_list);
2211 spin_lock_init(&q->requeue_lock);
2213 if (q->nr_hw_queues > 1)
2214 blk_queue_make_request(q, blk_mq_make_request);
2216 blk_queue_make_request(q, blk_sq_make_request);
2219 * Do this after blk_queue_make_request() overrides it...
2221 q->nr_requests = set->queue_depth;
2224 * Default to classic polling
2228 if (set->ops->complete)
2229 blk_queue_softirq_done(q, set->ops->complete);
2231 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2234 mutex_lock(&all_q_mutex);
2236 list_add_tail(&q->all_q_node, &all_q_list);
2237 blk_mq_add_queue_tag_set(set, q);
2238 blk_mq_map_swqueue(q, cpu_online_mask);
2240 mutex_unlock(&all_q_mutex);
2246 kfree(q->queue_hw_ctx);
2248 free_percpu(q->queue_ctx);
2251 return ERR_PTR(-ENOMEM);
2253 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2255 void blk_mq_free_queue(struct request_queue *q)
2257 struct blk_mq_tag_set *set = q->tag_set;
2259 mutex_lock(&all_q_mutex);
2260 list_del_init(&q->all_q_node);
2261 mutex_unlock(&all_q_mutex);
2265 blk_mq_del_queue_tag_set(q);
2267 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2268 blk_mq_free_hw_queues(q, set);
2271 /* Basically redo blk_mq_init_queue with queue frozen */
2272 static void blk_mq_queue_reinit(struct request_queue *q,
2273 const struct cpumask *online_mask)
2275 WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2277 blk_mq_sysfs_unregister(q);
2280 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2281 * we should change hctx numa_node according to new topology (this
2282 * involves free and re-allocate memory, worthy doing?)
2285 blk_mq_map_swqueue(q, online_mask);
2287 blk_mq_sysfs_register(q);
2291 * New online cpumask which is going to be set in this hotplug event.
2292 * Declare this cpumasks as global as cpu-hotplug operation is invoked
2293 * one-by-one and dynamically allocating this could result in a failure.
2295 static struct cpumask cpuhp_online_new;
2297 static void blk_mq_queue_reinit_work(void)
2299 struct request_queue *q;
2301 mutex_lock(&all_q_mutex);
2303 * We need to freeze and reinit all existing queues. Freezing
2304 * involves synchronous wait for an RCU grace period and doing it
2305 * one by one may take a long time. Start freezing all queues in
2306 * one swoop and then wait for the completions so that freezing can
2307 * take place in parallel.
2309 list_for_each_entry(q, &all_q_list, all_q_node)
2310 blk_mq_freeze_queue_start(q);
2311 list_for_each_entry(q, &all_q_list, all_q_node)
2312 blk_mq_freeze_queue_wait(q);
2314 list_for_each_entry(q, &all_q_list, all_q_node)
2315 blk_mq_queue_reinit(q, &cpuhp_online_new);
2317 list_for_each_entry(q, &all_q_list, all_q_node)
2318 blk_mq_unfreeze_queue(q);
2320 mutex_unlock(&all_q_mutex);
2323 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2325 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2326 blk_mq_queue_reinit_work();
2331 * Before hotadded cpu starts handling requests, new mappings must be
2332 * established. Otherwise, these requests in hw queue might never be
2335 * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2336 * for CPU0, and ctx1 for CPU1).
2338 * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2339 * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2341 * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2342 * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2343 * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2346 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2348 cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2349 cpumask_set_cpu(cpu, &cpuhp_online_new);
2350 blk_mq_queue_reinit_work();
2354 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2358 for (i = 0; i < set->nr_hw_queues; i++)
2359 if (!__blk_mq_alloc_rq_map(set, i))
2366 blk_mq_free_rq_map(set->tags[i]);
2372 * Allocate the request maps associated with this tag_set. Note that this
2373 * may reduce the depth asked for, if memory is tight. set->queue_depth
2374 * will be updated to reflect the allocated depth.
2376 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2381 depth = set->queue_depth;
2383 err = __blk_mq_alloc_rq_maps(set);
2387 set->queue_depth >>= 1;
2388 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2392 } while (set->queue_depth);
2394 if (!set->queue_depth || err) {
2395 pr_err("blk-mq: failed to allocate request map\n");
2399 if (depth != set->queue_depth)
2400 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2401 depth, set->queue_depth);
2407 * Alloc a tag set to be associated with one or more request queues.
2408 * May fail with EINVAL for various error conditions. May adjust the
2409 * requested depth down, if if it too large. In that case, the set
2410 * value will be stored in set->queue_depth.
2412 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2416 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2418 if (!set->nr_hw_queues)
2420 if (!set->queue_depth)
2422 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2425 if (!set->ops->queue_rq)
2428 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2429 pr_info("blk-mq: reduced tag depth to %u\n",
2431 set->queue_depth = BLK_MQ_MAX_DEPTH;
2435 * If a crashdump is active, then we are potentially in a very
2436 * memory constrained environment. Limit us to 1 queue and
2437 * 64 tags to prevent using too much memory.
2439 if (is_kdump_kernel()) {
2440 set->nr_hw_queues = 1;
2441 set->queue_depth = min(64U, set->queue_depth);
2444 * There is no use for more h/w queues than cpus.
2446 if (set->nr_hw_queues > nr_cpu_ids)
2447 set->nr_hw_queues = nr_cpu_ids;
2449 set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2450 GFP_KERNEL, set->numa_node);
2455 set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2456 GFP_KERNEL, set->numa_node);
2460 if (set->ops->map_queues)
2461 ret = set->ops->map_queues(set);
2463 ret = blk_mq_map_queues(set);
2465 goto out_free_mq_map;
2467 ret = blk_mq_alloc_rq_maps(set);
2469 goto out_free_mq_map;
2471 mutex_init(&set->tag_list_lock);
2472 INIT_LIST_HEAD(&set->tag_list);
2484 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2486 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2490 for (i = 0; i < nr_cpu_ids; i++)
2491 blk_mq_free_map_and_requests(set, i);
2499 EXPORT_SYMBOL(blk_mq_free_tag_set);
2501 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2503 struct blk_mq_tag_set *set = q->tag_set;
2504 struct blk_mq_hw_ctx *hctx;
2507 if (!set || nr > set->queue_depth)
2511 queue_for_each_hw_ctx(q, hctx, i) {
2514 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2520 q->nr_requests = nr;
2525 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2527 struct request_queue *q;
2529 if (nr_hw_queues > nr_cpu_ids)
2530 nr_hw_queues = nr_cpu_ids;
2531 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2534 list_for_each_entry(q, &set->tag_list, tag_set_list)
2535 blk_mq_freeze_queue(q);
2537 set->nr_hw_queues = nr_hw_queues;
2538 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2539 blk_mq_realloc_hw_ctxs(set, q);
2541 if (q->nr_hw_queues > 1)
2542 blk_queue_make_request(q, blk_mq_make_request);
2544 blk_queue_make_request(q, blk_sq_make_request);
2546 blk_mq_queue_reinit(q, cpu_online_mask);
2549 list_for_each_entry(q, &set->tag_list, tag_set_list)
2550 blk_mq_unfreeze_queue(q);
2552 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2554 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2555 struct blk_mq_hw_ctx *hctx,
2558 struct blk_rq_stat stat[2];
2559 unsigned long ret = 0;
2562 * If stats collection isn't on, don't sleep but turn it on for
2565 if (!blk_stat_enable(q))
2569 * We don't have to do this once per IO, should optimize this
2570 * to just use the current window of stats until it changes
2572 memset(&stat, 0, sizeof(stat));
2573 blk_hctx_stat_get(hctx, stat);
2576 * As an optimistic guess, use half of the mean service time
2577 * for this type of request. We can (and should) make this smarter.
2578 * For instance, if the completion latencies are tight, we can
2579 * get closer than just half the mean. This is especially
2580 * important on devices where the completion latencies are longer
2583 if (req_op(rq) == REQ_OP_READ && stat[BLK_STAT_READ].nr_samples)
2584 ret = (stat[BLK_STAT_READ].mean + 1) / 2;
2585 else if (req_op(rq) == REQ_OP_WRITE && stat[BLK_STAT_WRITE].nr_samples)
2586 ret = (stat[BLK_STAT_WRITE].mean + 1) / 2;
2591 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2592 struct blk_mq_hw_ctx *hctx,
2595 struct hrtimer_sleeper hs;
2596 enum hrtimer_mode mode;
2600 if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2606 * -1: don't ever hybrid sleep
2607 * 0: use half of prev avg
2608 * >0: use this specific value
2610 if (q->poll_nsec == -1)
2612 else if (q->poll_nsec > 0)
2613 nsecs = q->poll_nsec;
2615 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2620 set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2623 * This will be replaced with the stats tracking code, using
2624 * 'avg_completion_time / 2' as the pre-sleep target.
2628 mode = HRTIMER_MODE_REL;
2629 hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2630 hrtimer_set_expires(&hs.timer, kt);
2632 hrtimer_init_sleeper(&hs, current);
2634 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2636 set_current_state(TASK_UNINTERRUPTIBLE);
2637 hrtimer_start_expires(&hs.timer, mode);
2640 hrtimer_cancel(&hs.timer);
2641 mode = HRTIMER_MODE_ABS;
2642 } while (hs.task && !signal_pending(current));
2644 __set_current_state(TASK_RUNNING);
2645 destroy_hrtimer_on_stack(&hs.timer);
2649 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2651 struct request_queue *q = hctx->queue;
2655 * If we sleep, have the caller restart the poll loop to reset
2656 * the state. Like for the other success return cases, the
2657 * caller is responsible for checking if the IO completed. If
2658 * the IO isn't complete, we'll get called again and will go
2659 * straight to the busy poll loop.
2661 if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2664 hctx->poll_considered++;
2666 state = current->state;
2667 while (!need_resched()) {
2670 hctx->poll_invoked++;
2672 ret = q->mq_ops->poll(hctx, rq->tag);
2674 hctx->poll_success++;
2675 set_current_state(TASK_RUNNING);
2679 if (signal_pending_state(state, current))
2680 set_current_state(TASK_RUNNING);
2682 if (current->state == TASK_RUNNING)
2692 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2694 struct blk_mq_hw_ctx *hctx;
2695 struct blk_plug *plug;
2698 if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2699 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2702 plug = current->plug;
2704 blk_flush_plug_list(plug, false);
2706 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2707 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2709 return __blk_mq_poll(hctx, rq);
2711 EXPORT_SYMBOL_GPL(blk_mq_poll);
2713 void blk_mq_disable_hotplug(void)
2715 mutex_lock(&all_q_mutex);
2718 void blk_mq_enable_hotplug(void)
2720 mutex_unlock(&all_q_mutex);
2723 static int __init blk_mq_init(void)
2725 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2726 blk_mq_hctx_notify_dead);
2728 cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2729 blk_mq_queue_reinit_prepare,
2730 blk_mq_queue_reinit_dead);
2733 subsys_initcall(blk_mq_init);