1 // SPDX-License-Identifier: GPL-2.0
3 * Block multiqueue core code
5 * Copyright (C) 2013-2014 Jens Axboe
6 * Copyright (C) 2013-2014 Christoph Hellwig
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/backing-dev.h>
11 #include <linux/bio.h>
12 #include <linux/blkdev.h>
13 #include <linux/kmemleak.h>
15 #include <linux/init.h>
16 #include <linux/slab.h>
17 #include <linux/workqueue.h>
18 #include <linux/smp.h>
19 #include <linux/llist.h>
20 #include <linux/list_sort.h>
21 #include <linux/cpu.h>
22 #include <linux/cache.h>
23 #include <linux/sched/sysctl.h>
24 #include <linux/sched/topology.h>
25 #include <linux/sched/signal.h>
26 #include <linux/delay.h>
27 #include <linux/crash_dump.h>
28 #include <linux/prefetch.h>
29 #include <linux/blk-crypto.h>
31 #include <trace/events/block.h>
33 #include <linux/blk-mq.h>
34 #include <linux/t10-pi.h>
37 #include "blk-mq-debugfs.h"
38 #include "blk-mq-tag.h"
41 #include "blk-mq-sched.h"
42 #include "blk-rq-qos.h"
44 static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
46 static void blk_mq_poll_stats_start(struct request_queue *q);
47 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
49 static int blk_mq_poll_stats_bkt(const struct request *rq)
51 int ddir, sectors, bucket;
53 ddir = rq_data_dir(rq);
54 sectors = blk_rq_stats_sectors(rq);
56 bucket = ddir + 2 * ilog2(sectors);
60 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
61 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
67 * Check if any of the ctx, dispatch list or elevator
68 * have pending work in this hardware queue.
70 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
72 return !list_empty_careful(&hctx->dispatch) ||
73 sbitmap_any_bit_set(&hctx->ctx_map) ||
74 blk_mq_sched_has_work(hctx);
78 * Mark this ctx as having pending work in this hardware queue
80 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
81 struct blk_mq_ctx *ctx)
83 const int bit = ctx->index_hw[hctx->type];
85 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
86 sbitmap_set_bit(&hctx->ctx_map, bit);
89 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
90 struct blk_mq_ctx *ctx)
92 const int bit = ctx->index_hw[hctx->type];
94 sbitmap_clear_bit(&hctx->ctx_map, bit);
98 struct block_device *part;
99 unsigned int inflight[2];
102 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
103 struct request *rq, void *priv,
106 struct mq_inflight *mi = priv;
108 if ((!mi->part->bd_partno || rq->part == mi->part) &&
109 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
110 mi->inflight[rq_data_dir(rq)]++;
115 unsigned int blk_mq_in_flight(struct request_queue *q,
116 struct block_device *part)
118 struct mq_inflight mi = { .part = part };
120 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
122 return mi.inflight[0] + mi.inflight[1];
125 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
126 unsigned int inflight[2])
128 struct mq_inflight mi = { .part = part };
130 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
131 inflight[0] = mi.inflight[0];
132 inflight[1] = mi.inflight[1];
135 void blk_freeze_queue_start(struct request_queue *q)
137 mutex_lock(&q->mq_freeze_lock);
138 if (++q->mq_freeze_depth == 1) {
139 percpu_ref_kill(&q->q_usage_counter);
140 mutex_unlock(&q->mq_freeze_lock);
142 blk_mq_run_hw_queues(q, false);
144 mutex_unlock(&q->mq_freeze_lock);
147 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
149 void blk_mq_freeze_queue_wait(struct request_queue *q)
151 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
153 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
155 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
156 unsigned long timeout)
158 return wait_event_timeout(q->mq_freeze_wq,
159 percpu_ref_is_zero(&q->q_usage_counter),
162 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
165 * Guarantee no request is in use, so we can change any data structure of
166 * the queue afterward.
168 void blk_freeze_queue(struct request_queue *q)
171 * In the !blk_mq case we are only calling this to kill the
172 * q_usage_counter, otherwise this increases the freeze depth
173 * and waits for it to return to zero. For this reason there is
174 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
175 * exported to drivers as the only user for unfreeze is blk_mq.
177 blk_freeze_queue_start(q);
178 blk_mq_freeze_queue_wait(q);
181 void blk_mq_freeze_queue(struct request_queue *q)
184 * ...just an alias to keep freeze and unfreeze actions balanced
185 * in the blk_mq_* namespace
189 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
191 void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
193 mutex_lock(&q->mq_freeze_lock);
195 q->q_usage_counter.data->force_atomic = true;
196 q->mq_freeze_depth--;
197 WARN_ON_ONCE(q->mq_freeze_depth < 0);
198 if (!q->mq_freeze_depth) {
199 percpu_ref_resurrect(&q->q_usage_counter);
200 wake_up_all(&q->mq_freeze_wq);
202 mutex_unlock(&q->mq_freeze_lock);
205 void blk_mq_unfreeze_queue(struct request_queue *q)
207 __blk_mq_unfreeze_queue(q, false);
209 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
212 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
213 * mpt3sas driver such that this function can be removed.
215 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
217 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
219 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
222 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
225 * Note: this function does not prevent that the struct request end_io()
226 * callback function is invoked. Once this function is returned, we make
227 * sure no dispatch can happen until the queue is unquiesced via
228 * blk_mq_unquiesce_queue().
230 void blk_mq_quiesce_queue(struct request_queue *q)
232 struct blk_mq_hw_ctx *hctx;
236 blk_mq_quiesce_queue_nowait(q);
238 queue_for_each_hw_ctx(q, hctx, i) {
239 if (hctx->flags & BLK_MQ_F_BLOCKING)
240 synchronize_srcu(hctx->srcu);
247 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
250 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
253 * This function recovers queue into the state before quiescing
254 * which is done by blk_mq_quiesce_queue.
256 void blk_mq_unquiesce_queue(struct request_queue *q)
258 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
260 /* dispatch requests which are inserted during quiescing */
261 blk_mq_run_hw_queues(q, true);
263 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
265 void blk_mq_wake_waiters(struct request_queue *q)
267 struct blk_mq_hw_ctx *hctx;
270 queue_for_each_hw_ctx(q, hctx, i)
271 if (blk_mq_hw_queue_mapped(hctx))
272 blk_mq_tag_wakeup_all(hctx->tags, true);
276 * Only need start/end time stamping if we have iostat or
277 * blk stats enabled, or using an IO scheduler.
279 static inline bool blk_mq_need_time_stamp(struct request *rq)
281 return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
284 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
285 unsigned int tag, u64 alloc_time_ns)
287 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
288 struct request *rq = tags->static_rqs[tag];
290 if (data->q->elevator) {
291 rq->tag = BLK_MQ_NO_TAG;
292 rq->internal_tag = tag;
295 rq->internal_tag = BLK_MQ_NO_TAG;
298 /* csd/requeue_work/fifo_time is initialized before use */
300 rq->mq_ctx = data->ctx;
301 rq->mq_hctx = data->hctx;
303 rq->cmd_flags = data->cmd_flags;
304 if (data->flags & BLK_MQ_REQ_PM)
305 rq->rq_flags |= RQF_PM;
306 if (blk_queue_io_stat(data->q))
307 rq->rq_flags |= RQF_IO_STAT;
308 INIT_LIST_HEAD(&rq->queuelist);
309 INIT_HLIST_NODE(&rq->hash);
310 RB_CLEAR_NODE(&rq->rb_node);
313 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
314 rq->alloc_time_ns = alloc_time_ns;
316 if (blk_mq_need_time_stamp(rq))
317 rq->start_time_ns = ktime_get_ns();
319 rq->start_time_ns = 0;
320 rq->io_start_time_ns = 0;
321 rq->stats_sectors = 0;
322 rq->nr_phys_segments = 0;
323 #if defined(CONFIG_BLK_DEV_INTEGRITY)
324 rq->nr_integrity_segments = 0;
326 blk_crypto_rq_set_defaults(rq);
327 /* tag was already set */
328 WRITE_ONCE(rq->deadline, 0);
333 rq->end_io_data = NULL;
335 data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++;
336 refcount_set(&rq->ref, 1);
338 if (!op_is_flush(data->cmd_flags)) {
339 struct elevator_queue *e = data->q->elevator;
342 if (e && e->type->ops.prepare_request) {
343 if (e->type->icq_cache)
344 blk_mq_sched_assign_ioc(rq);
346 e->type->ops.prepare_request(rq);
347 rq->rq_flags |= RQF_ELVPRIV;
351 data->hctx->queued++;
355 static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data)
357 struct request_queue *q = data->q;
358 struct elevator_queue *e = q->elevator;
359 u64 alloc_time_ns = 0;
362 /* alloc_time includes depth and tag waits */
363 if (blk_queue_rq_alloc_time(q))
364 alloc_time_ns = ktime_get_ns();
366 if (data->cmd_flags & REQ_NOWAIT)
367 data->flags |= BLK_MQ_REQ_NOWAIT;
371 * Flush/passthrough requests are special and go directly to the
372 * dispatch list. Don't include reserved tags in the
373 * limiting, as it isn't useful.
375 if (!op_is_flush(data->cmd_flags) &&
376 !blk_op_is_passthrough(data->cmd_flags) &&
377 e->type->ops.limit_depth &&
378 !(data->flags & BLK_MQ_REQ_RESERVED))
379 e->type->ops.limit_depth(data->cmd_flags, data);
383 data->ctx = blk_mq_get_ctx(q);
384 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
386 blk_mq_tag_busy(data->hctx);
389 * Waiting allocations only fail because of an inactive hctx. In that
390 * case just retry the hctx assignment and tag allocation as CPU hotplug
391 * should have migrated us to an online CPU by now.
393 tag = blk_mq_get_tag(data);
394 if (tag == BLK_MQ_NO_TAG) {
395 if (data->flags & BLK_MQ_REQ_NOWAIT)
399 * Give up the CPU and sleep for a random short time to ensure
400 * that thread using a realtime scheduling class are migrated
401 * off the CPU, and thus off the hctx that is going away.
406 return blk_mq_rq_ctx_init(data, tag, alloc_time_ns);
409 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
410 blk_mq_req_flags_t flags)
412 struct blk_mq_alloc_data data = {
420 ret = blk_queue_enter(q, flags);
424 rq = __blk_mq_alloc_request(&data);
428 rq->__sector = (sector_t) -1;
429 rq->bio = rq->biotail = NULL;
433 return ERR_PTR(-EWOULDBLOCK);
435 EXPORT_SYMBOL(blk_mq_alloc_request);
437 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
438 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
440 struct blk_mq_alloc_data data = {
445 u64 alloc_time_ns = 0;
450 /* alloc_time includes depth and tag waits */
451 if (blk_queue_rq_alloc_time(q))
452 alloc_time_ns = ktime_get_ns();
455 * If the tag allocator sleeps we could get an allocation for a
456 * different hardware context. No need to complicate the low level
457 * allocator for this for the rare use case of a command tied to
460 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
461 return ERR_PTR(-EINVAL);
463 if (hctx_idx >= q->nr_hw_queues)
464 return ERR_PTR(-EIO);
466 ret = blk_queue_enter(q, flags);
471 * Check if the hardware context is actually mapped to anything.
472 * If not tell the caller that it should skip this queue.
475 data.hctx = q->queue_hw_ctx[hctx_idx];
476 if (!blk_mq_hw_queue_mapped(data.hctx))
478 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
479 if (cpu >= nr_cpu_ids)
481 data.ctx = __blk_mq_get_ctx(q, cpu);
484 blk_mq_tag_busy(data.hctx);
487 tag = blk_mq_get_tag(&data);
488 if (tag == BLK_MQ_NO_TAG)
490 return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns);
496 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
498 static void __blk_mq_free_request(struct request *rq)
500 struct request_queue *q = rq->q;
501 struct blk_mq_ctx *ctx = rq->mq_ctx;
502 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
503 const int sched_tag = rq->internal_tag;
505 blk_crypto_free_request(rq);
506 blk_pm_mark_last_busy(rq);
508 if (rq->tag != BLK_MQ_NO_TAG)
509 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
510 if (sched_tag != BLK_MQ_NO_TAG)
511 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
512 blk_mq_sched_restart(hctx);
516 void blk_mq_free_request(struct request *rq)
518 struct request_queue *q = rq->q;
519 struct elevator_queue *e = q->elevator;
520 struct blk_mq_ctx *ctx = rq->mq_ctx;
521 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
523 if (rq->rq_flags & RQF_ELVPRIV) {
524 if (e && e->type->ops.finish_request)
525 e->type->ops.finish_request(rq);
527 put_io_context(rq->elv.icq->ioc);
532 ctx->rq_completed[rq_is_sync(rq)]++;
533 if (rq->rq_flags & RQF_MQ_INFLIGHT)
534 __blk_mq_dec_active_requests(hctx);
536 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
537 laptop_io_completion(q->disk->bdi);
541 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
542 if (refcount_dec_and_test(&rq->ref))
543 __blk_mq_free_request(rq);
545 EXPORT_SYMBOL_GPL(blk_mq_free_request);
547 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
551 if (blk_mq_need_time_stamp(rq))
552 now = ktime_get_ns();
554 if (rq->rq_flags & RQF_STATS) {
555 blk_mq_poll_stats_start(rq->q);
556 blk_stat_add(rq, now);
559 blk_mq_sched_completed_request(rq, now);
561 blk_account_io_done(rq, now);
564 rq_qos_done(rq->q, rq);
565 rq->end_io(rq, error);
567 blk_mq_free_request(rq);
570 EXPORT_SYMBOL(__blk_mq_end_request);
572 void blk_mq_end_request(struct request *rq, blk_status_t error)
574 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
576 __blk_mq_end_request(rq, error);
578 EXPORT_SYMBOL(blk_mq_end_request);
580 static void blk_complete_reqs(struct llist_head *list)
582 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
583 struct request *rq, *next;
585 llist_for_each_entry_safe(rq, next, entry, ipi_list)
586 rq->q->mq_ops->complete(rq);
589 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
591 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
594 static int blk_softirq_cpu_dead(unsigned int cpu)
596 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
600 static void __blk_mq_complete_request_remote(void *data)
602 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
605 static inline bool blk_mq_complete_need_ipi(struct request *rq)
607 int cpu = raw_smp_processor_id();
609 if (!IS_ENABLED(CONFIG_SMP) ||
610 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
613 * With force threaded interrupts enabled, raising softirq from an SMP
614 * function call will always result in waking the ksoftirqd thread.
615 * This is probably worse than completing the request on a different
618 if (force_irqthreads())
621 /* same CPU or cache domain? Complete locally */
622 if (cpu == rq->mq_ctx->cpu ||
623 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
624 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
627 /* don't try to IPI to an offline CPU */
628 return cpu_online(rq->mq_ctx->cpu);
631 static void blk_mq_complete_send_ipi(struct request *rq)
633 struct llist_head *list;
636 cpu = rq->mq_ctx->cpu;
637 list = &per_cpu(blk_cpu_done, cpu);
638 if (llist_add(&rq->ipi_list, list)) {
639 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
640 smp_call_function_single_async(cpu, &rq->csd);
644 static void blk_mq_raise_softirq(struct request *rq)
646 struct llist_head *list;
649 list = this_cpu_ptr(&blk_cpu_done);
650 if (llist_add(&rq->ipi_list, list))
651 raise_softirq(BLOCK_SOFTIRQ);
655 bool blk_mq_complete_request_remote(struct request *rq)
657 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
660 * For a polled request, always complete locallly, it's pointless
661 * to redirect the completion.
663 if (rq->cmd_flags & REQ_HIPRI)
666 if (blk_mq_complete_need_ipi(rq)) {
667 blk_mq_complete_send_ipi(rq);
671 if (rq->q->nr_hw_queues == 1) {
672 blk_mq_raise_softirq(rq);
677 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
680 * blk_mq_complete_request - end I/O on a request
681 * @rq: the request being processed
684 * Complete a request by scheduling the ->complete_rq operation.
686 void blk_mq_complete_request(struct request *rq)
688 if (!blk_mq_complete_request_remote(rq))
689 rq->q->mq_ops->complete(rq);
691 EXPORT_SYMBOL(blk_mq_complete_request);
693 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
694 __releases(hctx->srcu)
696 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
699 srcu_read_unlock(hctx->srcu, srcu_idx);
702 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
703 __acquires(hctx->srcu)
705 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
706 /* shut up gcc false positive */
710 *srcu_idx = srcu_read_lock(hctx->srcu);
714 * blk_mq_start_request - Start processing a request
715 * @rq: Pointer to request to be started
717 * Function used by device drivers to notify the block layer that a request
718 * is going to be processed now, so blk layer can do proper initializations
719 * such as starting the timeout timer.
721 void blk_mq_start_request(struct request *rq)
723 struct request_queue *q = rq->q;
725 trace_block_rq_issue(rq);
727 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
728 rq->io_start_time_ns = ktime_get_ns();
729 rq->stats_sectors = blk_rq_sectors(rq);
730 rq->rq_flags |= RQF_STATS;
734 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
737 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
739 #ifdef CONFIG_BLK_DEV_INTEGRITY
740 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
741 q->integrity.profile->prepare_fn(rq);
744 EXPORT_SYMBOL(blk_mq_start_request);
746 static void __blk_mq_requeue_request(struct request *rq)
748 struct request_queue *q = rq->q;
750 blk_mq_put_driver_tag(rq);
752 trace_block_rq_requeue(rq);
753 rq_qos_requeue(q, rq);
755 if (blk_mq_request_started(rq)) {
756 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
757 rq->rq_flags &= ~RQF_TIMED_OUT;
761 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
763 __blk_mq_requeue_request(rq);
765 /* this request will be re-inserted to io scheduler queue */
766 blk_mq_sched_requeue_request(rq);
768 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
770 EXPORT_SYMBOL(blk_mq_requeue_request);
772 static void blk_mq_requeue_work(struct work_struct *work)
774 struct request_queue *q =
775 container_of(work, struct request_queue, requeue_work.work);
777 struct request *rq, *next;
779 spin_lock_irq(&q->requeue_lock);
780 list_splice_init(&q->requeue_list, &rq_list);
781 spin_unlock_irq(&q->requeue_lock);
783 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
784 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
787 rq->rq_flags &= ~RQF_SOFTBARRIER;
788 list_del_init(&rq->queuelist);
790 * If RQF_DONTPREP, rq has contained some driver specific
791 * data, so insert it to hctx dispatch list to avoid any
794 if (rq->rq_flags & RQF_DONTPREP)
795 blk_mq_request_bypass_insert(rq, false, false);
797 blk_mq_sched_insert_request(rq, true, false, false);
800 while (!list_empty(&rq_list)) {
801 rq = list_entry(rq_list.next, struct request, queuelist);
802 list_del_init(&rq->queuelist);
803 blk_mq_sched_insert_request(rq, false, false, false);
806 blk_mq_run_hw_queues(q, false);
809 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
810 bool kick_requeue_list)
812 struct request_queue *q = rq->q;
816 * We abuse this flag that is otherwise used by the I/O scheduler to
817 * request head insertion from the workqueue.
819 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
821 spin_lock_irqsave(&q->requeue_lock, flags);
823 rq->rq_flags |= RQF_SOFTBARRIER;
824 list_add(&rq->queuelist, &q->requeue_list);
826 list_add_tail(&rq->queuelist, &q->requeue_list);
828 spin_unlock_irqrestore(&q->requeue_lock, flags);
830 if (kick_requeue_list)
831 blk_mq_kick_requeue_list(q);
834 void blk_mq_kick_requeue_list(struct request_queue *q)
836 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
838 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
840 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
843 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
844 msecs_to_jiffies(msecs));
846 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
848 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
850 if (tag < tags->nr_tags) {
851 prefetch(tags->rqs[tag]);
852 return tags->rqs[tag];
857 EXPORT_SYMBOL(blk_mq_tag_to_rq);
859 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
860 void *priv, bool reserved)
863 * If we find a request that isn't idle and the queue matches,
864 * we know the queue is busy. Return false to stop the iteration.
866 if (blk_mq_request_started(rq) && rq->q == hctx->queue) {
876 bool blk_mq_queue_inflight(struct request_queue *q)
880 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
883 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
885 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
887 req->rq_flags |= RQF_TIMED_OUT;
888 if (req->q->mq_ops->timeout) {
889 enum blk_eh_timer_return ret;
891 ret = req->q->mq_ops->timeout(req, reserved);
892 if (ret == BLK_EH_DONE)
894 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
900 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
902 unsigned long deadline;
904 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
906 if (rq->rq_flags & RQF_TIMED_OUT)
909 deadline = READ_ONCE(rq->deadline);
910 if (time_after_eq(jiffies, deadline))
915 else if (time_after(*next, deadline))
920 void blk_mq_put_rq_ref(struct request *rq)
924 else if (refcount_dec_and_test(&rq->ref))
925 __blk_mq_free_request(rq);
928 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
929 struct request *rq, void *priv, bool reserved)
931 unsigned long *next = priv;
934 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
935 * be reallocated underneath the timeout handler's processing, then
936 * the expire check is reliable. If the request is not expired, then
937 * it was completed and reallocated as a new request after returning
938 * from blk_mq_check_expired().
940 if (blk_mq_req_expired(rq, next))
941 blk_mq_rq_timed_out(rq, reserved);
945 static void blk_mq_timeout_work(struct work_struct *work)
947 struct request_queue *q =
948 container_of(work, struct request_queue, timeout_work);
949 unsigned long next = 0;
950 struct blk_mq_hw_ctx *hctx;
953 /* A deadlock might occur if a request is stuck requiring a
954 * timeout at the same time a queue freeze is waiting
955 * completion, since the timeout code would not be able to
956 * acquire the queue reference here.
958 * That's why we don't use blk_queue_enter here; instead, we use
959 * percpu_ref_tryget directly, because we need to be able to
960 * obtain a reference even in the short window between the queue
961 * starting to freeze, by dropping the first reference in
962 * blk_freeze_queue_start, and the moment the last request is
963 * consumed, marked by the instant q_usage_counter reaches
966 if (!percpu_ref_tryget(&q->q_usage_counter))
969 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
972 mod_timer(&q->timeout, next);
975 * Request timeouts are handled as a forward rolling timer. If
976 * we end up here it means that no requests are pending and
977 * also that no request has been pending for a while. Mark
980 queue_for_each_hw_ctx(q, hctx, i) {
981 /* the hctx may be unmapped, so check it here */
982 if (blk_mq_hw_queue_mapped(hctx))
983 blk_mq_tag_idle(hctx);
989 struct flush_busy_ctx_data {
990 struct blk_mq_hw_ctx *hctx;
991 struct list_head *list;
994 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
996 struct flush_busy_ctx_data *flush_data = data;
997 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
998 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
999 enum hctx_type type = hctx->type;
1001 spin_lock(&ctx->lock);
1002 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1003 sbitmap_clear_bit(sb, bitnr);
1004 spin_unlock(&ctx->lock);
1009 * Process software queues that have been marked busy, splicing them
1010 * to the for-dispatch
1012 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1014 struct flush_busy_ctx_data data = {
1019 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1021 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1023 struct dispatch_rq_data {
1024 struct blk_mq_hw_ctx *hctx;
1028 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1031 struct dispatch_rq_data *dispatch_data = data;
1032 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1033 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1034 enum hctx_type type = hctx->type;
1036 spin_lock(&ctx->lock);
1037 if (!list_empty(&ctx->rq_lists[type])) {
1038 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1039 list_del_init(&dispatch_data->rq->queuelist);
1040 if (list_empty(&ctx->rq_lists[type]))
1041 sbitmap_clear_bit(sb, bitnr);
1043 spin_unlock(&ctx->lock);
1045 return !dispatch_data->rq;
1048 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1049 struct blk_mq_ctx *start)
1051 unsigned off = start ? start->index_hw[hctx->type] : 0;
1052 struct dispatch_rq_data data = {
1057 __sbitmap_for_each_set(&hctx->ctx_map, off,
1058 dispatch_rq_from_ctx, &data);
1063 static inline unsigned int queued_to_index(unsigned int queued)
1068 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1071 static bool __blk_mq_get_driver_tag(struct request *rq)
1073 struct sbitmap_queue *bt = rq->mq_hctx->tags->bitmap_tags;
1074 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1077 blk_mq_tag_busy(rq->mq_hctx);
1079 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1080 bt = rq->mq_hctx->tags->breserved_tags;
1083 if (!hctx_may_queue(rq->mq_hctx, bt))
1087 tag = __sbitmap_queue_get(bt);
1088 if (tag == BLK_MQ_NO_TAG)
1091 rq->tag = tag + tag_offset;
1095 bool blk_mq_get_driver_tag(struct request *rq)
1097 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1099 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq))
1102 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1103 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1104 rq->rq_flags |= RQF_MQ_INFLIGHT;
1105 __blk_mq_inc_active_requests(hctx);
1107 hctx->tags->rqs[rq->tag] = rq;
1111 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1112 int flags, void *key)
1114 struct blk_mq_hw_ctx *hctx;
1116 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1118 spin_lock(&hctx->dispatch_wait_lock);
1119 if (!list_empty(&wait->entry)) {
1120 struct sbitmap_queue *sbq;
1122 list_del_init(&wait->entry);
1123 sbq = hctx->tags->bitmap_tags;
1124 atomic_dec(&sbq->ws_active);
1126 spin_unlock(&hctx->dispatch_wait_lock);
1128 blk_mq_run_hw_queue(hctx, true);
1133 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1134 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1135 * restart. For both cases, take care to check the condition again after
1136 * marking us as waiting.
1138 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1141 struct sbitmap_queue *sbq = hctx->tags->bitmap_tags;
1142 struct wait_queue_head *wq;
1143 wait_queue_entry_t *wait;
1146 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1147 blk_mq_sched_mark_restart_hctx(hctx);
1150 * It's possible that a tag was freed in the window between the
1151 * allocation failure and adding the hardware queue to the wait
1154 * Don't clear RESTART here, someone else could have set it.
1155 * At most this will cost an extra queue run.
1157 return blk_mq_get_driver_tag(rq);
1160 wait = &hctx->dispatch_wait;
1161 if (!list_empty_careful(&wait->entry))
1164 wq = &bt_wait_ptr(sbq, hctx)->wait;
1166 spin_lock_irq(&wq->lock);
1167 spin_lock(&hctx->dispatch_wait_lock);
1168 if (!list_empty(&wait->entry)) {
1169 spin_unlock(&hctx->dispatch_wait_lock);
1170 spin_unlock_irq(&wq->lock);
1174 atomic_inc(&sbq->ws_active);
1175 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1176 __add_wait_queue(wq, wait);
1179 * It's possible that a tag was freed in the window between the
1180 * allocation failure and adding the hardware queue to the wait
1183 ret = blk_mq_get_driver_tag(rq);
1185 spin_unlock(&hctx->dispatch_wait_lock);
1186 spin_unlock_irq(&wq->lock);
1191 * We got a tag, remove ourselves from the wait queue to ensure
1192 * someone else gets the wakeup.
1194 list_del_init(&wait->entry);
1195 atomic_dec(&sbq->ws_active);
1196 spin_unlock(&hctx->dispatch_wait_lock);
1197 spin_unlock_irq(&wq->lock);
1202 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1203 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1205 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1206 * - EWMA is one simple way to compute running average value
1207 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1208 * - take 4 as factor for avoiding to get too small(0) result, and this
1209 * factor doesn't matter because EWMA decreases exponentially
1211 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1215 ewma = hctx->dispatch_busy;
1220 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1222 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1223 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1225 hctx->dispatch_busy = ewma;
1228 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1230 static void blk_mq_handle_dev_resource(struct request *rq,
1231 struct list_head *list)
1233 struct request *next =
1234 list_first_entry_or_null(list, struct request, queuelist);
1237 * If an I/O scheduler has been configured and we got a driver tag for
1238 * the next request already, free it.
1241 blk_mq_put_driver_tag(next);
1243 list_add(&rq->queuelist, list);
1244 __blk_mq_requeue_request(rq);
1247 static void blk_mq_handle_zone_resource(struct request *rq,
1248 struct list_head *zone_list)
1251 * If we end up here it is because we cannot dispatch a request to a
1252 * specific zone due to LLD level zone-write locking or other zone
1253 * related resource not being available. In this case, set the request
1254 * aside in zone_list for retrying it later.
1256 list_add(&rq->queuelist, zone_list);
1257 __blk_mq_requeue_request(rq);
1260 enum prep_dispatch {
1262 PREP_DISPATCH_NO_TAG,
1263 PREP_DISPATCH_NO_BUDGET,
1266 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1269 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1270 int budget_token = -1;
1273 budget_token = blk_mq_get_dispatch_budget(rq->q);
1274 if (budget_token < 0) {
1275 blk_mq_put_driver_tag(rq);
1276 return PREP_DISPATCH_NO_BUDGET;
1278 blk_mq_set_rq_budget_token(rq, budget_token);
1281 if (!blk_mq_get_driver_tag(rq)) {
1283 * The initial allocation attempt failed, so we need to
1284 * rerun the hardware queue when a tag is freed. The
1285 * waitqueue takes care of that. If the queue is run
1286 * before we add this entry back on the dispatch list,
1287 * we'll re-run it below.
1289 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1291 * All budgets not got from this function will be put
1292 * together during handling partial dispatch
1295 blk_mq_put_dispatch_budget(rq->q, budget_token);
1296 return PREP_DISPATCH_NO_TAG;
1300 return PREP_DISPATCH_OK;
1303 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1304 static void blk_mq_release_budgets(struct request_queue *q,
1305 struct list_head *list)
1309 list_for_each_entry(rq, list, queuelist) {
1310 int budget_token = blk_mq_get_rq_budget_token(rq);
1312 if (budget_token >= 0)
1313 blk_mq_put_dispatch_budget(q, budget_token);
1318 * Returns true if we did some work AND can potentially do more.
1320 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1321 unsigned int nr_budgets)
1323 enum prep_dispatch prep;
1324 struct request_queue *q = hctx->queue;
1325 struct request *rq, *nxt;
1327 blk_status_t ret = BLK_STS_OK;
1328 LIST_HEAD(zone_list);
1329 bool needs_resource = false;
1331 if (list_empty(list))
1335 * Now process all the entries, sending them to the driver.
1337 errors = queued = 0;
1339 struct blk_mq_queue_data bd;
1341 rq = list_first_entry(list, struct request, queuelist);
1343 WARN_ON_ONCE(hctx != rq->mq_hctx);
1344 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1345 if (prep != PREP_DISPATCH_OK)
1348 list_del_init(&rq->queuelist);
1353 * Flag last if we have no more requests, or if we have more
1354 * but can't assign a driver tag to it.
1356 if (list_empty(list))
1359 nxt = list_first_entry(list, struct request, queuelist);
1360 bd.last = !blk_mq_get_driver_tag(nxt);
1364 * once the request is queued to lld, no need to cover the
1369 ret = q->mq_ops->queue_rq(hctx, &bd);
1374 case BLK_STS_RESOURCE:
1375 needs_resource = true;
1377 case BLK_STS_DEV_RESOURCE:
1378 blk_mq_handle_dev_resource(rq, list);
1380 case BLK_STS_ZONE_RESOURCE:
1382 * Move the request to zone_list and keep going through
1383 * the dispatch list to find more requests the drive can
1386 blk_mq_handle_zone_resource(rq, &zone_list);
1387 needs_resource = true;
1391 blk_mq_end_request(rq, ret);
1393 } while (!list_empty(list));
1395 if (!list_empty(&zone_list))
1396 list_splice_tail_init(&zone_list, list);
1398 hctx->dispatched[queued_to_index(queued)]++;
1400 /* If we didn't flush the entire list, we could have told the driver
1401 * there was more coming, but that turned out to be a lie.
1403 if ((!list_empty(list) || errors || needs_resource ||
1404 ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued)
1405 q->mq_ops->commit_rqs(hctx);
1407 * Any items that need requeuing? Stuff them into hctx->dispatch,
1408 * that is where we will continue on next queue run.
1410 if (!list_empty(list)) {
1412 /* For non-shared tags, the RESTART check will suffice */
1413 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1414 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1417 blk_mq_release_budgets(q, list);
1419 spin_lock(&hctx->lock);
1420 list_splice_tail_init(list, &hctx->dispatch);
1421 spin_unlock(&hctx->lock);
1424 * Order adding requests to hctx->dispatch and checking
1425 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1426 * in blk_mq_sched_restart(). Avoid restart code path to
1427 * miss the new added requests to hctx->dispatch, meantime
1428 * SCHED_RESTART is observed here.
1433 * If SCHED_RESTART was set by the caller of this function and
1434 * it is no longer set that means that it was cleared by another
1435 * thread and hence that a queue rerun is needed.
1437 * If 'no_tag' is set, that means that we failed getting
1438 * a driver tag with an I/O scheduler attached. If our dispatch
1439 * waitqueue is no longer active, ensure that we run the queue
1440 * AFTER adding our entries back to the list.
1442 * If no I/O scheduler has been configured it is possible that
1443 * the hardware queue got stopped and restarted before requests
1444 * were pushed back onto the dispatch list. Rerun the queue to
1445 * avoid starvation. Notes:
1446 * - blk_mq_run_hw_queue() checks whether or not a queue has
1447 * been stopped before rerunning a queue.
1448 * - Some but not all block drivers stop a queue before
1449 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1452 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1453 * bit is set, run queue after a delay to avoid IO stalls
1454 * that could otherwise occur if the queue is idle. We'll do
1455 * similar if we couldn't get budget or couldn't lock a zone
1456 * and SCHED_RESTART is set.
1458 needs_restart = blk_mq_sched_needs_restart(hctx);
1459 if (prep == PREP_DISPATCH_NO_BUDGET)
1460 needs_resource = true;
1461 if (!needs_restart ||
1462 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1463 blk_mq_run_hw_queue(hctx, true);
1464 else if (needs_restart && needs_resource)
1465 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1467 blk_mq_update_dispatch_busy(hctx, true);
1470 blk_mq_update_dispatch_busy(hctx, false);
1472 return (queued + errors) != 0;
1476 * __blk_mq_run_hw_queue - Run a hardware queue.
1477 * @hctx: Pointer to the hardware queue to run.
1479 * Send pending requests to the hardware.
1481 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1486 * We can't run the queue inline with ints disabled. Ensure that
1487 * we catch bad users of this early.
1489 WARN_ON_ONCE(in_interrupt());
1491 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1493 hctx_lock(hctx, &srcu_idx);
1494 blk_mq_sched_dispatch_requests(hctx);
1495 hctx_unlock(hctx, srcu_idx);
1498 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1500 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1502 if (cpu >= nr_cpu_ids)
1503 cpu = cpumask_first(hctx->cpumask);
1508 * It'd be great if the workqueue API had a way to pass
1509 * in a mask and had some smarts for more clever placement.
1510 * For now we just round-robin here, switching for every
1511 * BLK_MQ_CPU_WORK_BATCH queued items.
1513 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1516 int next_cpu = hctx->next_cpu;
1518 if (hctx->queue->nr_hw_queues == 1)
1519 return WORK_CPU_UNBOUND;
1521 if (--hctx->next_cpu_batch <= 0) {
1523 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1525 if (next_cpu >= nr_cpu_ids)
1526 next_cpu = blk_mq_first_mapped_cpu(hctx);
1527 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1531 * Do unbound schedule if we can't find a online CPU for this hctx,
1532 * and it should only happen in the path of handling CPU DEAD.
1534 if (!cpu_online(next_cpu)) {
1541 * Make sure to re-select CPU next time once after CPUs
1542 * in hctx->cpumask become online again.
1544 hctx->next_cpu = next_cpu;
1545 hctx->next_cpu_batch = 1;
1546 return WORK_CPU_UNBOUND;
1549 hctx->next_cpu = next_cpu;
1554 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1555 * @hctx: Pointer to the hardware queue to run.
1556 * @async: If we want to run the queue asynchronously.
1557 * @msecs: Milliseconds of delay to wait before running the queue.
1559 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1560 * with a delay of @msecs.
1562 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1563 unsigned long msecs)
1565 if (unlikely(blk_mq_hctx_stopped(hctx)))
1568 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1569 int cpu = get_cpu();
1570 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1571 __blk_mq_run_hw_queue(hctx);
1579 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1580 msecs_to_jiffies(msecs));
1584 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1585 * @hctx: Pointer to the hardware queue to run.
1586 * @msecs: Milliseconds of delay to wait before running the queue.
1588 * Run a hardware queue asynchronously with a delay of @msecs.
1590 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1592 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1594 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1597 * blk_mq_run_hw_queue - Start to run a hardware queue.
1598 * @hctx: Pointer to the hardware queue to run.
1599 * @async: If we want to run the queue asynchronously.
1601 * Check if the request queue is not in a quiesced state and if there are
1602 * pending requests to be sent. If this is true, run the queue to send requests
1605 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1611 * When queue is quiesced, we may be switching io scheduler, or
1612 * updating nr_hw_queues, or other things, and we can't run queue
1613 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1615 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1618 hctx_lock(hctx, &srcu_idx);
1619 need_run = !blk_queue_quiesced(hctx->queue) &&
1620 blk_mq_hctx_has_pending(hctx);
1621 hctx_unlock(hctx, srcu_idx);
1624 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1626 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1629 * Is the request queue handled by an IO scheduler that does not respect
1630 * hardware queues when dispatching?
1632 static bool blk_mq_has_sqsched(struct request_queue *q)
1634 struct elevator_queue *e = q->elevator;
1636 if (e && e->type->ops.dispatch_request &&
1637 !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE))
1643 * Return prefered queue to dispatch from (if any) for non-mq aware IO
1646 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
1648 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
1650 * If the IO scheduler does not respect hardware queues when
1651 * dispatching, we just don't bother with multiple HW queues and
1652 * dispatch from hctx for the current CPU since running multiple queues
1653 * just causes lock contention inside the scheduler and pointless cache
1656 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, 0, ctx);
1658 if (!blk_mq_hctx_stopped(hctx))
1664 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
1665 * @q: Pointer to the request queue to run.
1666 * @async: If we want to run the queue asynchronously.
1668 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1670 struct blk_mq_hw_ctx *hctx, *sq_hctx;
1674 if (blk_mq_has_sqsched(q))
1675 sq_hctx = blk_mq_get_sq_hctx(q);
1676 queue_for_each_hw_ctx(q, hctx, i) {
1677 if (blk_mq_hctx_stopped(hctx))
1680 * Dispatch from this hctx either if there's no hctx preferred
1681 * by IO scheduler or if it has requests that bypass the
1684 if (!sq_hctx || sq_hctx == hctx ||
1685 !list_empty_careful(&hctx->dispatch))
1686 blk_mq_run_hw_queue(hctx, async);
1689 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1692 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1693 * @q: Pointer to the request queue to run.
1694 * @msecs: Milliseconds of delay to wait before running the queues.
1696 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1698 struct blk_mq_hw_ctx *hctx, *sq_hctx;
1702 if (blk_mq_has_sqsched(q))
1703 sq_hctx = blk_mq_get_sq_hctx(q);
1704 queue_for_each_hw_ctx(q, hctx, i) {
1705 if (blk_mq_hctx_stopped(hctx))
1708 * Dispatch from this hctx either if there's no hctx preferred
1709 * by IO scheduler or if it has requests that bypass the
1712 if (!sq_hctx || sq_hctx == hctx ||
1713 !list_empty_careful(&hctx->dispatch))
1714 blk_mq_delay_run_hw_queue(hctx, msecs);
1717 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1720 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1721 * @q: request queue.
1723 * The caller is responsible for serializing this function against
1724 * blk_mq_{start,stop}_hw_queue().
1726 bool blk_mq_queue_stopped(struct request_queue *q)
1728 struct blk_mq_hw_ctx *hctx;
1731 queue_for_each_hw_ctx(q, hctx, i)
1732 if (blk_mq_hctx_stopped(hctx))
1737 EXPORT_SYMBOL(blk_mq_queue_stopped);
1740 * This function is often used for pausing .queue_rq() by driver when
1741 * there isn't enough resource or some conditions aren't satisfied, and
1742 * BLK_STS_RESOURCE is usually returned.
1744 * We do not guarantee that dispatch can be drained or blocked
1745 * after blk_mq_stop_hw_queue() returns. Please use
1746 * blk_mq_quiesce_queue() for that requirement.
1748 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1750 cancel_delayed_work(&hctx->run_work);
1752 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1754 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1757 * This function is often used for pausing .queue_rq() by driver when
1758 * there isn't enough resource or some conditions aren't satisfied, and
1759 * BLK_STS_RESOURCE is usually returned.
1761 * We do not guarantee that dispatch can be drained or blocked
1762 * after blk_mq_stop_hw_queues() returns. Please use
1763 * blk_mq_quiesce_queue() for that requirement.
1765 void blk_mq_stop_hw_queues(struct request_queue *q)
1767 struct blk_mq_hw_ctx *hctx;
1770 queue_for_each_hw_ctx(q, hctx, i)
1771 blk_mq_stop_hw_queue(hctx);
1773 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1775 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1777 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1779 blk_mq_run_hw_queue(hctx, false);
1781 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1783 void blk_mq_start_hw_queues(struct request_queue *q)
1785 struct blk_mq_hw_ctx *hctx;
1788 queue_for_each_hw_ctx(q, hctx, i)
1789 blk_mq_start_hw_queue(hctx);
1791 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1793 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1795 if (!blk_mq_hctx_stopped(hctx))
1798 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1799 blk_mq_run_hw_queue(hctx, async);
1801 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1803 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1805 struct blk_mq_hw_ctx *hctx;
1808 queue_for_each_hw_ctx(q, hctx, i)
1809 blk_mq_start_stopped_hw_queue(hctx, async);
1811 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1813 static void blk_mq_run_work_fn(struct work_struct *work)
1815 struct blk_mq_hw_ctx *hctx;
1817 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1820 * If we are stopped, don't run the queue.
1822 if (blk_mq_hctx_stopped(hctx))
1825 __blk_mq_run_hw_queue(hctx);
1828 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1832 struct blk_mq_ctx *ctx = rq->mq_ctx;
1833 enum hctx_type type = hctx->type;
1835 lockdep_assert_held(&ctx->lock);
1837 trace_block_rq_insert(rq);
1840 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1842 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1845 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1848 struct blk_mq_ctx *ctx = rq->mq_ctx;
1850 lockdep_assert_held(&ctx->lock);
1852 __blk_mq_insert_req_list(hctx, rq, at_head);
1853 blk_mq_hctx_mark_pending(hctx, ctx);
1857 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1858 * @rq: Pointer to request to be inserted.
1859 * @at_head: true if the request should be inserted at the head of the list.
1860 * @run_queue: If we should run the hardware queue after inserting the request.
1862 * Should only be used carefully, when the caller knows we want to
1863 * bypass a potential IO scheduler on the target device.
1865 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1868 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1870 spin_lock(&hctx->lock);
1872 list_add(&rq->queuelist, &hctx->dispatch);
1874 list_add_tail(&rq->queuelist, &hctx->dispatch);
1875 spin_unlock(&hctx->lock);
1878 blk_mq_run_hw_queue(hctx, false);
1881 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1882 struct list_head *list)
1886 enum hctx_type type = hctx->type;
1889 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1892 list_for_each_entry(rq, list, queuelist) {
1893 BUG_ON(rq->mq_ctx != ctx);
1894 trace_block_rq_insert(rq);
1897 spin_lock(&ctx->lock);
1898 list_splice_tail_init(list, &ctx->rq_lists[type]);
1899 blk_mq_hctx_mark_pending(hctx, ctx);
1900 spin_unlock(&ctx->lock);
1903 static int plug_rq_cmp(void *priv, const struct list_head *a,
1904 const struct list_head *b)
1906 struct request *rqa = container_of(a, struct request, queuelist);
1907 struct request *rqb = container_of(b, struct request, queuelist);
1909 if (rqa->mq_ctx != rqb->mq_ctx)
1910 return rqa->mq_ctx > rqb->mq_ctx;
1911 if (rqa->mq_hctx != rqb->mq_hctx)
1912 return rqa->mq_hctx > rqb->mq_hctx;
1914 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1917 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1921 if (list_empty(&plug->mq_list))
1923 list_splice_init(&plug->mq_list, &list);
1925 if (plug->rq_count > 2 && plug->multiple_queues)
1926 list_sort(NULL, &list, plug_rq_cmp);
1931 struct list_head rq_list;
1932 struct request *rq, *head_rq = list_entry_rq(list.next);
1933 struct list_head *pos = &head_rq->queuelist; /* skip first */
1934 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1935 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1936 unsigned int depth = 1;
1938 list_for_each_continue(pos, &list) {
1939 rq = list_entry_rq(pos);
1941 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1946 list_cut_before(&rq_list, &list, pos);
1947 trace_block_unplug(head_rq->q, depth, !from_schedule);
1948 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1950 } while(!list_empty(&list));
1953 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1954 unsigned int nr_segs)
1958 if (bio->bi_opf & REQ_RAHEAD)
1959 rq->cmd_flags |= REQ_FAILFAST_MASK;
1961 rq->__sector = bio->bi_iter.bi_sector;
1962 rq->write_hint = bio->bi_write_hint;
1963 blk_rq_bio_prep(rq, bio, nr_segs);
1965 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
1966 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1969 blk_account_io_start(rq);
1972 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1974 blk_qc_t *cookie, bool last)
1976 struct request_queue *q = rq->q;
1977 struct blk_mq_queue_data bd = {
1981 blk_qc_t new_cookie;
1984 new_cookie = request_to_qc_t(hctx, rq);
1987 * For OK queue, we are done. For error, caller may kill it.
1988 * Any other error (busy), just add it to our list as we
1989 * previously would have done.
1991 ret = q->mq_ops->queue_rq(hctx, &bd);
1994 blk_mq_update_dispatch_busy(hctx, false);
1995 *cookie = new_cookie;
1997 case BLK_STS_RESOURCE:
1998 case BLK_STS_DEV_RESOURCE:
1999 blk_mq_update_dispatch_busy(hctx, true);
2000 __blk_mq_requeue_request(rq);
2003 blk_mq_update_dispatch_busy(hctx, false);
2004 *cookie = BLK_QC_T_NONE;
2011 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2014 bool bypass_insert, bool last)
2016 struct request_queue *q = rq->q;
2017 bool run_queue = true;
2021 * RCU or SRCU read lock is needed before checking quiesced flag.
2023 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2024 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2025 * and avoid driver to try to dispatch again.
2027 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2029 bypass_insert = false;
2033 if (q->elevator && !bypass_insert)
2036 budget_token = blk_mq_get_dispatch_budget(q);
2037 if (budget_token < 0)
2040 blk_mq_set_rq_budget_token(rq, budget_token);
2042 if (!blk_mq_get_driver_tag(rq)) {
2043 blk_mq_put_dispatch_budget(q, budget_token);
2047 return __blk_mq_issue_directly(hctx, rq, cookie, last);
2050 return BLK_STS_RESOURCE;
2052 blk_mq_sched_insert_request(rq, false, run_queue, false);
2058 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2059 * @hctx: Pointer of the associated hardware queue.
2060 * @rq: Pointer to request to be sent.
2061 * @cookie: Request queue cookie.
2063 * If the device has enough resources to accept a new request now, send the
2064 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2065 * we can try send it another time in the future. Requests inserted at this
2066 * queue have higher priority.
2068 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2069 struct request *rq, blk_qc_t *cookie)
2074 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2076 hctx_lock(hctx, &srcu_idx);
2078 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
2079 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2080 blk_mq_request_bypass_insert(rq, false, true);
2081 else if (ret != BLK_STS_OK)
2082 blk_mq_end_request(rq, ret);
2084 hctx_unlock(hctx, srcu_idx);
2087 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2091 blk_qc_t unused_cookie;
2092 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2094 hctx_lock(hctx, &srcu_idx);
2095 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
2096 hctx_unlock(hctx, srcu_idx);
2101 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2102 struct list_head *list)
2107 while (!list_empty(list)) {
2109 struct request *rq = list_first_entry(list, struct request,
2112 list_del_init(&rq->queuelist);
2113 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2114 if (ret != BLK_STS_OK) {
2116 if (ret == BLK_STS_RESOURCE ||
2117 ret == BLK_STS_DEV_RESOURCE) {
2118 blk_mq_request_bypass_insert(rq, false,
2122 blk_mq_end_request(rq, ret);
2128 * If we didn't flush the entire list, we could have told
2129 * the driver there was more coming, but that turned out to
2132 if ((!list_empty(list) || errors) &&
2133 hctx->queue->mq_ops->commit_rqs && queued)
2134 hctx->queue->mq_ops->commit_rqs(hctx);
2137 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2139 list_add_tail(&rq->queuelist, &plug->mq_list);
2141 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
2142 struct request *tmp;
2144 tmp = list_first_entry(&plug->mq_list, struct request,
2146 if (tmp->q != rq->q)
2147 plug->multiple_queues = true;
2152 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
2153 * queues. This is important for md arrays to benefit from merging
2156 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
2158 if (plug->multiple_queues)
2159 return BLK_MAX_REQUEST_COUNT * 2;
2160 return BLK_MAX_REQUEST_COUNT;
2164 * blk_mq_submit_bio - Create and send a request to block device.
2165 * @bio: Bio pointer.
2167 * Builds up a request structure from @q and @bio and send to the device. The
2168 * request may not be queued directly to hardware if:
2169 * * This request can be merged with another one
2170 * * We want to place request at plug queue for possible future merging
2171 * * There is an IO scheduler active at this queue
2173 * It will not queue the request if there is an error with the bio, or at the
2176 * Returns: Request queue cookie.
2178 blk_qc_t blk_mq_submit_bio(struct bio *bio)
2180 struct request_queue *q = bio->bi_bdev->bd_disk->queue;
2181 const int is_sync = op_is_sync(bio->bi_opf);
2182 const int is_flush_fua = op_is_flush(bio->bi_opf);
2183 struct blk_mq_alloc_data data = {
2187 struct blk_plug *plug;
2188 struct request *same_queue_rq = NULL;
2189 unsigned int nr_segs;
2194 blk_queue_bounce(q, &bio);
2195 __blk_queue_split(&bio, &nr_segs);
2199 if (!bio_integrity_prep(bio))
2202 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2203 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2206 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2209 rq_qos_throttle(q, bio);
2211 hipri = bio->bi_opf & REQ_HIPRI;
2213 data.cmd_flags = bio->bi_opf;
2214 rq = __blk_mq_alloc_request(&data);
2215 if (unlikely(!rq)) {
2216 rq_qos_cleanup(q, bio);
2217 if (bio->bi_opf & REQ_NOWAIT)
2218 bio_wouldblock_error(bio);
2222 trace_block_getrq(bio);
2224 rq_qos_track(q, rq, bio);
2226 cookie = request_to_qc_t(data.hctx, rq);
2228 blk_mq_bio_to_request(rq, bio, nr_segs);
2230 ret = blk_crypto_init_request(rq);
2231 if (ret != BLK_STS_OK) {
2232 bio->bi_status = ret;
2234 blk_mq_free_request(rq);
2235 return BLK_QC_T_NONE;
2238 plug = blk_mq_plug(q, bio);
2239 if (unlikely(is_flush_fua)) {
2240 /* Bypass scheduler for flush requests */
2241 blk_insert_flush(rq);
2242 blk_mq_run_hw_queue(data.hctx, true);
2243 } else if (plug && (q->nr_hw_queues == 1 ||
2244 blk_mq_is_sbitmap_shared(rq->mq_hctx->flags) ||
2245 q->mq_ops->commit_rqs || !blk_queue_nonrot(q))) {
2247 * Use plugging if we have a ->commit_rqs() hook as well, as
2248 * we know the driver uses bd->last in a smart fashion.
2250 * Use normal plugging if this disk is slow HDD, as sequential
2251 * IO may benefit a lot from plug merging.
2253 unsigned int request_count = plug->rq_count;
2254 struct request *last = NULL;
2257 trace_block_plug(q);
2259 last = list_entry_rq(plug->mq_list.prev);
2261 if (request_count >= blk_plug_max_rq_count(plug) || (last &&
2262 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2263 blk_flush_plug_list(plug, false);
2264 trace_block_plug(q);
2267 blk_add_rq_to_plug(plug, rq);
2268 } else if (q->elevator) {
2269 /* Insert the request at the IO scheduler queue */
2270 blk_mq_sched_insert_request(rq, false, true, true);
2271 } else if (plug && !blk_queue_nomerges(q)) {
2273 * We do limited plugging. If the bio can be merged, do that.
2274 * Otherwise the existing request in the plug list will be
2275 * issued. So the plug list will have one request at most
2276 * The plug list might get flushed before this. If that happens,
2277 * the plug list is empty, and same_queue_rq is invalid.
2279 if (list_empty(&plug->mq_list))
2280 same_queue_rq = NULL;
2281 if (same_queue_rq) {
2282 list_del_init(&same_queue_rq->queuelist);
2285 blk_add_rq_to_plug(plug, rq);
2286 trace_block_plug(q);
2288 if (same_queue_rq) {
2289 data.hctx = same_queue_rq->mq_hctx;
2290 trace_block_unplug(q, 1, true);
2291 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2294 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2295 !data.hctx->dispatch_busy) {
2297 * There is no scheduler and we can try to send directly
2300 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2303 blk_mq_sched_insert_request(rq, false, true, true);
2307 return BLK_QC_T_NONE;
2311 return BLK_QC_T_NONE;
2314 static size_t order_to_size(unsigned int order)
2316 return (size_t)PAGE_SIZE << order;
2319 /* called before freeing request pool in @tags */
2320 static void blk_mq_clear_rq_mapping(struct blk_mq_tag_set *set,
2321 struct blk_mq_tags *tags, unsigned int hctx_idx)
2323 struct blk_mq_tags *drv_tags = set->tags[hctx_idx];
2325 unsigned long flags;
2327 list_for_each_entry(page, &tags->page_list, lru) {
2328 unsigned long start = (unsigned long)page_address(page);
2329 unsigned long end = start + order_to_size(page->private);
2332 for (i = 0; i < set->queue_depth; i++) {
2333 struct request *rq = drv_tags->rqs[i];
2334 unsigned long rq_addr = (unsigned long)rq;
2336 if (rq_addr >= start && rq_addr < end) {
2337 WARN_ON_ONCE(refcount_read(&rq->ref) != 0);
2338 cmpxchg(&drv_tags->rqs[i], rq, NULL);
2344 * Wait until all pending iteration is done.
2346 * Request reference is cleared and it is guaranteed to be observed
2347 * after the ->lock is released.
2349 spin_lock_irqsave(&drv_tags->lock, flags);
2350 spin_unlock_irqrestore(&drv_tags->lock, flags);
2353 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2354 unsigned int hctx_idx)
2358 if (tags->rqs && set->ops->exit_request) {
2361 for (i = 0; i < tags->nr_tags; i++) {
2362 struct request *rq = tags->static_rqs[i];
2366 set->ops->exit_request(set, rq, hctx_idx);
2367 tags->static_rqs[i] = NULL;
2371 blk_mq_clear_rq_mapping(set, tags, hctx_idx);
2373 while (!list_empty(&tags->page_list)) {
2374 page = list_first_entry(&tags->page_list, struct page, lru);
2375 list_del_init(&page->lru);
2377 * Remove kmemleak object previously allocated in
2378 * blk_mq_alloc_rqs().
2380 kmemleak_free(page_address(page));
2381 __free_pages(page, page->private);
2385 void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags)
2389 kfree(tags->static_rqs);
2390 tags->static_rqs = NULL;
2392 blk_mq_free_tags(tags, flags);
2395 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2396 unsigned int hctx_idx,
2397 unsigned int nr_tags,
2398 unsigned int reserved_tags,
2401 struct blk_mq_tags *tags;
2404 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2405 if (node == NUMA_NO_NODE)
2406 node = set->numa_node;
2408 tags = blk_mq_init_tags(nr_tags, reserved_tags, node, flags);
2412 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2413 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2416 blk_mq_free_tags(tags, flags);
2420 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2421 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2423 if (!tags->static_rqs) {
2425 blk_mq_free_tags(tags, flags);
2432 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2433 unsigned int hctx_idx, int node)
2437 if (set->ops->init_request) {
2438 ret = set->ops->init_request(set, rq, hctx_idx, node);
2443 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2447 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2448 unsigned int hctx_idx, unsigned int depth)
2450 unsigned int i, j, entries_per_page, max_order = 4;
2451 size_t rq_size, left;
2454 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2455 if (node == NUMA_NO_NODE)
2456 node = set->numa_node;
2458 INIT_LIST_HEAD(&tags->page_list);
2461 * rq_size is the size of the request plus driver payload, rounded
2462 * to the cacheline size
2464 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2466 left = rq_size * depth;
2468 for (i = 0; i < depth; ) {
2469 int this_order = max_order;
2474 while (this_order && left < order_to_size(this_order - 1))
2478 page = alloc_pages_node(node,
2479 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2485 if (order_to_size(this_order) < rq_size)
2492 page->private = this_order;
2493 list_add_tail(&page->lru, &tags->page_list);
2495 p = page_address(page);
2497 * Allow kmemleak to scan these pages as they contain pointers
2498 * to additional allocations like via ops->init_request().
2500 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2501 entries_per_page = order_to_size(this_order) / rq_size;
2502 to_do = min(entries_per_page, depth - i);
2503 left -= to_do * rq_size;
2504 for (j = 0; j < to_do; j++) {
2505 struct request *rq = p;
2507 tags->static_rqs[i] = rq;
2508 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2509 tags->static_rqs[i] = NULL;
2520 blk_mq_free_rqs(set, tags, hctx_idx);
2524 struct rq_iter_data {
2525 struct blk_mq_hw_ctx *hctx;
2529 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2531 struct rq_iter_data *iter_data = data;
2533 if (rq->mq_hctx != iter_data->hctx)
2535 iter_data->has_rq = true;
2539 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2541 struct blk_mq_tags *tags = hctx->sched_tags ?
2542 hctx->sched_tags : hctx->tags;
2543 struct rq_iter_data data = {
2547 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2551 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2552 struct blk_mq_hw_ctx *hctx)
2554 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2556 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2561 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2563 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2564 struct blk_mq_hw_ctx, cpuhp_online);
2566 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2567 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2571 * Prevent new request from being allocated on the current hctx.
2573 * The smp_mb__after_atomic() Pairs with the implied barrier in
2574 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2575 * seen once we return from the tag allocator.
2577 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2578 smp_mb__after_atomic();
2581 * Try to grab a reference to the queue and wait for any outstanding
2582 * requests. If we could not grab a reference the queue has been
2583 * frozen and there are no requests.
2585 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2586 while (blk_mq_hctx_has_requests(hctx))
2588 percpu_ref_put(&hctx->queue->q_usage_counter);
2594 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2596 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2597 struct blk_mq_hw_ctx, cpuhp_online);
2599 if (cpumask_test_cpu(cpu, hctx->cpumask))
2600 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2605 * 'cpu' is going away. splice any existing rq_list entries from this
2606 * software queue to the hw queue dispatch list, and ensure that it
2609 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2611 struct blk_mq_hw_ctx *hctx;
2612 struct blk_mq_ctx *ctx;
2614 enum hctx_type type;
2616 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2617 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2620 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2623 spin_lock(&ctx->lock);
2624 if (!list_empty(&ctx->rq_lists[type])) {
2625 list_splice_init(&ctx->rq_lists[type], &tmp);
2626 blk_mq_hctx_clear_pending(hctx, ctx);
2628 spin_unlock(&ctx->lock);
2630 if (list_empty(&tmp))
2633 spin_lock(&hctx->lock);
2634 list_splice_tail_init(&tmp, &hctx->dispatch);
2635 spin_unlock(&hctx->lock);
2637 blk_mq_run_hw_queue(hctx, true);
2641 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2643 if (!(hctx->flags & BLK_MQ_F_STACKING))
2644 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2645 &hctx->cpuhp_online);
2646 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2651 * Before freeing hw queue, clearing the flush request reference in
2652 * tags->rqs[] for avoiding potential UAF.
2654 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
2655 unsigned int queue_depth, struct request *flush_rq)
2658 unsigned long flags;
2660 /* The hw queue may not be mapped yet */
2664 WARN_ON_ONCE(refcount_read(&flush_rq->ref) != 0);
2666 for (i = 0; i < queue_depth; i++)
2667 cmpxchg(&tags->rqs[i], flush_rq, NULL);
2670 * Wait until all pending iteration is done.
2672 * Request reference is cleared and it is guaranteed to be observed
2673 * after the ->lock is released.
2675 spin_lock_irqsave(&tags->lock, flags);
2676 spin_unlock_irqrestore(&tags->lock, flags);
2679 /* hctx->ctxs will be freed in queue's release handler */
2680 static void blk_mq_exit_hctx(struct request_queue *q,
2681 struct blk_mq_tag_set *set,
2682 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2684 struct request *flush_rq = hctx->fq->flush_rq;
2686 if (blk_mq_hw_queue_mapped(hctx))
2687 blk_mq_tag_idle(hctx);
2689 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
2690 set->queue_depth, flush_rq);
2691 if (set->ops->exit_request)
2692 set->ops->exit_request(set, flush_rq, hctx_idx);
2694 if (set->ops->exit_hctx)
2695 set->ops->exit_hctx(hctx, hctx_idx);
2697 blk_mq_remove_cpuhp(hctx);
2699 spin_lock(&q->unused_hctx_lock);
2700 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2701 spin_unlock(&q->unused_hctx_lock);
2704 static void blk_mq_exit_hw_queues(struct request_queue *q,
2705 struct blk_mq_tag_set *set, int nr_queue)
2707 struct blk_mq_hw_ctx *hctx;
2710 queue_for_each_hw_ctx(q, hctx, i) {
2713 blk_mq_debugfs_unregister_hctx(hctx);
2714 blk_mq_exit_hctx(q, set, hctx, i);
2718 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2720 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2722 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2723 __alignof__(struct blk_mq_hw_ctx)) !=
2724 sizeof(struct blk_mq_hw_ctx));
2726 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2727 hw_ctx_size += sizeof(struct srcu_struct);
2732 static int blk_mq_init_hctx(struct request_queue *q,
2733 struct blk_mq_tag_set *set,
2734 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2736 hctx->queue_num = hctx_idx;
2738 if (!(hctx->flags & BLK_MQ_F_STACKING))
2739 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2740 &hctx->cpuhp_online);
2741 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2743 hctx->tags = set->tags[hctx_idx];
2745 if (set->ops->init_hctx &&
2746 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2747 goto unregister_cpu_notifier;
2749 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2755 if (set->ops->exit_hctx)
2756 set->ops->exit_hctx(hctx, hctx_idx);
2757 unregister_cpu_notifier:
2758 blk_mq_remove_cpuhp(hctx);
2762 static struct blk_mq_hw_ctx *
2763 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2766 struct blk_mq_hw_ctx *hctx;
2767 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2769 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2771 goto fail_alloc_hctx;
2773 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2776 atomic_set(&hctx->nr_active, 0);
2777 if (node == NUMA_NO_NODE)
2778 node = set->numa_node;
2779 hctx->numa_node = node;
2781 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2782 spin_lock_init(&hctx->lock);
2783 INIT_LIST_HEAD(&hctx->dispatch);
2785 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
2787 INIT_LIST_HEAD(&hctx->hctx_list);
2790 * Allocate space for all possible cpus to avoid allocation at
2793 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2798 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2799 gfp, node, false, false))
2803 spin_lock_init(&hctx->dispatch_wait_lock);
2804 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2805 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2807 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2811 if (hctx->flags & BLK_MQ_F_BLOCKING)
2812 init_srcu_struct(hctx->srcu);
2813 blk_mq_hctx_kobj_init(hctx);
2818 sbitmap_free(&hctx->ctx_map);
2822 free_cpumask_var(hctx->cpumask);
2829 static void blk_mq_init_cpu_queues(struct request_queue *q,
2830 unsigned int nr_hw_queues)
2832 struct blk_mq_tag_set *set = q->tag_set;
2835 for_each_possible_cpu(i) {
2836 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2837 struct blk_mq_hw_ctx *hctx;
2841 spin_lock_init(&__ctx->lock);
2842 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2843 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2848 * Set local node, IFF we have more than one hw queue. If
2849 * not, we remain on the home node of the device
2851 for (j = 0; j < set->nr_maps; j++) {
2852 hctx = blk_mq_map_queue_type(q, j, i);
2853 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2854 hctx->numa_node = cpu_to_node(i);
2859 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2862 unsigned int flags = set->flags;
2865 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2866 set->queue_depth, set->reserved_tags, flags);
2867 if (!set->tags[hctx_idx])
2870 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2875 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2876 set->tags[hctx_idx] = NULL;
2880 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2881 unsigned int hctx_idx)
2883 unsigned int flags = set->flags;
2885 if (set->tags && set->tags[hctx_idx]) {
2886 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2887 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2888 set->tags[hctx_idx] = NULL;
2892 static void blk_mq_map_swqueue(struct request_queue *q)
2894 unsigned int i, j, hctx_idx;
2895 struct blk_mq_hw_ctx *hctx;
2896 struct blk_mq_ctx *ctx;
2897 struct blk_mq_tag_set *set = q->tag_set;
2899 queue_for_each_hw_ctx(q, hctx, i) {
2900 cpumask_clear(hctx->cpumask);
2902 hctx->dispatch_from = NULL;
2906 * Map software to hardware queues.
2908 * If the cpu isn't present, the cpu is mapped to first hctx.
2910 for_each_possible_cpu(i) {
2912 ctx = per_cpu_ptr(q->queue_ctx, i);
2913 for (j = 0; j < set->nr_maps; j++) {
2914 if (!set->map[j].nr_queues) {
2915 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2916 HCTX_TYPE_DEFAULT, i);
2919 hctx_idx = set->map[j].mq_map[i];
2920 /* unmapped hw queue can be remapped after CPU topo changed */
2921 if (!set->tags[hctx_idx] &&
2922 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2924 * If tags initialization fail for some hctx,
2925 * that hctx won't be brought online. In this
2926 * case, remap the current ctx to hctx[0] which
2927 * is guaranteed to always have tags allocated
2929 set->map[j].mq_map[i] = 0;
2932 hctx = blk_mq_map_queue_type(q, j, i);
2933 ctx->hctxs[j] = hctx;
2935 * If the CPU is already set in the mask, then we've
2936 * mapped this one already. This can happen if
2937 * devices share queues across queue maps.
2939 if (cpumask_test_cpu(i, hctx->cpumask))
2942 cpumask_set_cpu(i, hctx->cpumask);
2944 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2945 hctx->ctxs[hctx->nr_ctx++] = ctx;
2948 * If the nr_ctx type overflows, we have exceeded the
2949 * amount of sw queues we can support.
2951 BUG_ON(!hctx->nr_ctx);
2954 for (; j < HCTX_MAX_TYPES; j++)
2955 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2956 HCTX_TYPE_DEFAULT, i);
2959 queue_for_each_hw_ctx(q, hctx, i) {
2961 * If no software queues are mapped to this hardware queue,
2962 * disable it and free the request entries.
2964 if (!hctx->nr_ctx) {
2965 /* Never unmap queue 0. We need it as a
2966 * fallback in case of a new remap fails
2969 if (i && set->tags[i])
2970 blk_mq_free_map_and_requests(set, i);
2976 hctx->tags = set->tags[i];
2977 WARN_ON(!hctx->tags);
2980 * Set the map size to the number of mapped software queues.
2981 * This is more accurate and more efficient than looping
2982 * over all possibly mapped software queues.
2984 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2987 * Initialize batch roundrobin counts
2989 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2990 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2995 * Caller needs to ensure that we're either frozen/quiesced, or that
2996 * the queue isn't live yet.
2998 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3000 struct blk_mq_hw_ctx *hctx;
3003 queue_for_each_hw_ctx(q, hctx, i) {
3005 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3007 blk_mq_tag_idle(hctx);
3008 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3013 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3016 struct request_queue *q;
3018 lockdep_assert_held(&set->tag_list_lock);
3020 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3021 blk_mq_freeze_queue(q);
3022 queue_set_hctx_shared(q, shared);
3023 blk_mq_unfreeze_queue(q);
3027 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3029 struct blk_mq_tag_set *set = q->tag_set;
3031 mutex_lock(&set->tag_list_lock);
3032 list_del(&q->tag_set_list);
3033 if (list_is_singular(&set->tag_list)) {
3034 /* just transitioned to unshared */
3035 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3036 /* update existing queue */
3037 blk_mq_update_tag_set_shared(set, false);
3039 mutex_unlock(&set->tag_list_lock);
3040 INIT_LIST_HEAD(&q->tag_set_list);
3043 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3044 struct request_queue *q)
3046 mutex_lock(&set->tag_list_lock);
3049 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3051 if (!list_empty(&set->tag_list) &&
3052 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3053 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3054 /* update existing queue */
3055 blk_mq_update_tag_set_shared(set, true);
3057 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3058 queue_set_hctx_shared(q, true);
3059 list_add_tail(&q->tag_set_list, &set->tag_list);
3061 mutex_unlock(&set->tag_list_lock);
3064 /* All allocations will be freed in release handler of q->mq_kobj */
3065 static int blk_mq_alloc_ctxs(struct request_queue *q)
3067 struct blk_mq_ctxs *ctxs;
3070 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3074 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3075 if (!ctxs->queue_ctx)
3078 for_each_possible_cpu(cpu) {
3079 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3083 q->mq_kobj = &ctxs->kobj;
3084 q->queue_ctx = ctxs->queue_ctx;
3093 * It is the actual release handler for mq, but we do it from
3094 * request queue's release handler for avoiding use-after-free
3095 * and headache because q->mq_kobj shouldn't have been introduced,
3096 * but we can't group ctx/kctx kobj without it.
3098 void blk_mq_release(struct request_queue *q)
3100 struct blk_mq_hw_ctx *hctx, *next;
3103 queue_for_each_hw_ctx(q, hctx, i)
3104 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3106 /* all hctx are in .unused_hctx_list now */
3107 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3108 list_del_init(&hctx->hctx_list);
3109 kobject_put(&hctx->kobj);
3112 kfree(q->queue_hw_ctx);
3115 * release .mq_kobj and sw queue's kobject now because
3116 * both share lifetime with request queue.
3118 blk_mq_sysfs_deinit(q);
3121 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3124 struct request_queue *q;
3127 q = blk_alloc_queue(set->numa_node);
3129 return ERR_PTR(-ENOMEM);
3130 q->queuedata = queuedata;
3131 ret = blk_mq_init_allocated_queue(set, q);
3133 blk_cleanup_queue(q);
3134 return ERR_PTR(ret);
3139 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3141 return blk_mq_init_queue_data(set, NULL);
3143 EXPORT_SYMBOL(blk_mq_init_queue);
3145 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
3146 struct lock_class_key *lkclass)
3148 struct request_queue *q;
3149 struct gendisk *disk;
3151 q = blk_mq_init_queue_data(set, queuedata);
3155 disk = __alloc_disk_node(q, set->numa_node, lkclass);
3157 blk_cleanup_queue(q);
3158 return ERR_PTR(-ENOMEM);
3162 EXPORT_SYMBOL(__blk_mq_alloc_disk);
3164 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3165 struct blk_mq_tag_set *set, struct request_queue *q,
3166 int hctx_idx, int node)
3168 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3170 /* reuse dead hctx first */
3171 spin_lock(&q->unused_hctx_lock);
3172 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3173 if (tmp->numa_node == node) {
3179 list_del_init(&hctx->hctx_list);
3180 spin_unlock(&q->unused_hctx_lock);
3183 hctx = blk_mq_alloc_hctx(q, set, node);
3187 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3193 kobject_put(&hctx->kobj);
3198 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3199 struct request_queue *q)
3202 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3204 if (q->nr_hw_queues < set->nr_hw_queues) {
3205 struct blk_mq_hw_ctx **new_hctxs;
3207 new_hctxs = kcalloc_node(set->nr_hw_queues,
3208 sizeof(*new_hctxs), GFP_KERNEL,
3213 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3215 q->queue_hw_ctx = new_hctxs;
3220 /* protect against switching io scheduler */
3221 mutex_lock(&q->sysfs_lock);
3222 for (i = 0; i < set->nr_hw_queues; i++) {
3224 struct blk_mq_hw_ctx *hctx;
3226 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3228 * If the hw queue has been mapped to another numa node,
3229 * we need to realloc the hctx. If allocation fails, fallback
3230 * to use the previous one.
3232 if (hctxs[i] && (hctxs[i]->numa_node == node))
3235 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3238 blk_mq_exit_hctx(q, set, hctxs[i], i);
3242 pr_warn("Allocate new hctx on node %d fails,\
3243 fallback to previous one on node %d\n",
3244 node, hctxs[i]->numa_node);
3250 * Increasing nr_hw_queues fails. Free the newly allocated
3251 * hctxs and keep the previous q->nr_hw_queues.
3253 if (i != set->nr_hw_queues) {
3254 j = q->nr_hw_queues;
3258 end = q->nr_hw_queues;
3259 q->nr_hw_queues = set->nr_hw_queues;
3262 for (; j < end; j++) {
3263 struct blk_mq_hw_ctx *hctx = hctxs[j];
3267 blk_mq_free_map_and_requests(set, j);
3268 blk_mq_exit_hctx(q, set, hctx, j);
3272 mutex_unlock(&q->sysfs_lock);
3275 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3276 struct request_queue *q)
3278 /* mark the queue as mq asap */
3279 q->mq_ops = set->ops;
3281 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3282 blk_mq_poll_stats_bkt,
3283 BLK_MQ_POLL_STATS_BKTS, q);
3287 if (blk_mq_alloc_ctxs(q))
3290 /* init q->mq_kobj and sw queues' kobjects */
3291 blk_mq_sysfs_init(q);
3293 INIT_LIST_HEAD(&q->unused_hctx_list);
3294 spin_lock_init(&q->unused_hctx_lock);
3296 blk_mq_realloc_hw_ctxs(set, q);
3297 if (!q->nr_hw_queues)
3300 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3301 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3305 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3306 if (set->nr_maps > HCTX_TYPE_POLL &&
3307 set->map[HCTX_TYPE_POLL].nr_queues)
3308 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3310 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3311 INIT_LIST_HEAD(&q->requeue_list);
3312 spin_lock_init(&q->requeue_lock);
3314 q->nr_requests = set->queue_depth;
3317 * Default to classic polling
3319 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3321 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3322 blk_mq_add_queue_tag_set(set, q);
3323 blk_mq_map_swqueue(q);
3327 kfree(q->queue_hw_ctx);
3328 q->nr_hw_queues = 0;
3329 blk_mq_sysfs_deinit(q);
3331 blk_stat_free_callback(q->poll_cb);
3337 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3339 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3340 void blk_mq_exit_queue(struct request_queue *q)
3342 struct blk_mq_tag_set *set = q->tag_set;
3344 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
3345 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3346 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
3347 blk_mq_del_queue_tag_set(q);
3350 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3354 for (i = 0; i < set->nr_hw_queues; i++) {
3355 if (!__blk_mq_alloc_map_and_request(set, i))
3364 blk_mq_free_map_and_requests(set, i);
3370 * Allocate the request maps associated with this tag_set. Note that this
3371 * may reduce the depth asked for, if memory is tight. set->queue_depth
3372 * will be updated to reflect the allocated depth.
3374 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3379 depth = set->queue_depth;
3381 err = __blk_mq_alloc_rq_maps(set);
3385 set->queue_depth >>= 1;
3386 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3390 } while (set->queue_depth);
3392 if (!set->queue_depth || err) {
3393 pr_err("blk-mq: failed to allocate request map\n");
3397 if (depth != set->queue_depth)
3398 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3399 depth, set->queue_depth);
3404 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3407 * blk_mq_map_queues() and multiple .map_queues() implementations
3408 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3409 * number of hardware queues.
3411 if (set->nr_maps == 1)
3412 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3414 if (set->ops->map_queues && !is_kdump_kernel()) {
3418 * transport .map_queues is usually done in the following
3421 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3422 * mask = get_cpu_mask(queue)
3423 * for_each_cpu(cpu, mask)
3424 * set->map[x].mq_map[cpu] = queue;
3427 * When we need to remap, the table has to be cleared for
3428 * killing stale mapping since one CPU may not be mapped
3431 for (i = 0; i < set->nr_maps; i++)
3432 blk_mq_clear_mq_map(&set->map[i]);
3434 return set->ops->map_queues(set);
3436 BUG_ON(set->nr_maps > 1);
3437 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3441 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3442 int cur_nr_hw_queues, int new_nr_hw_queues)
3444 struct blk_mq_tags **new_tags;
3446 if (cur_nr_hw_queues >= new_nr_hw_queues)
3449 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3450 GFP_KERNEL, set->numa_node);
3455 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3456 sizeof(*set->tags));
3458 set->tags = new_tags;
3459 set->nr_hw_queues = new_nr_hw_queues;
3464 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
3465 int new_nr_hw_queues)
3467 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
3471 * Alloc a tag set to be associated with one or more request queues.
3472 * May fail with EINVAL for various error conditions. May adjust the
3473 * requested depth down, if it's too large. In that case, the set
3474 * value will be stored in set->queue_depth.
3476 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3480 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3482 if (!set->nr_hw_queues)
3484 if (!set->queue_depth)
3486 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3489 if (!set->ops->queue_rq)
3492 if (!set->ops->get_budget ^ !set->ops->put_budget)
3495 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3496 pr_info("blk-mq: reduced tag depth to %u\n",
3498 set->queue_depth = BLK_MQ_MAX_DEPTH;
3503 else if (set->nr_maps > HCTX_MAX_TYPES)
3507 * If a crashdump is active, then we are potentially in a very
3508 * memory constrained environment. Limit us to 1 queue and
3509 * 64 tags to prevent using too much memory.
3511 if (is_kdump_kernel()) {
3512 set->nr_hw_queues = 1;
3514 set->queue_depth = min(64U, set->queue_depth);
3517 * There is no use for more h/w queues than cpus if we just have
3520 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3521 set->nr_hw_queues = nr_cpu_ids;
3523 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
3527 for (i = 0; i < set->nr_maps; i++) {
3528 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3529 sizeof(set->map[i].mq_map[0]),
3530 GFP_KERNEL, set->numa_node);
3531 if (!set->map[i].mq_map)
3532 goto out_free_mq_map;
3533 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3536 ret = blk_mq_update_queue_map(set);
3538 goto out_free_mq_map;
3540 ret = blk_mq_alloc_map_and_requests(set);
3542 goto out_free_mq_map;
3544 if (blk_mq_is_sbitmap_shared(set->flags)) {
3545 atomic_set(&set->active_queues_shared_sbitmap, 0);
3547 if (blk_mq_init_shared_sbitmap(set)) {
3549 goto out_free_mq_rq_maps;
3553 mutex_init(&set->tag_list_lock);
3554 INIT_LIST_HEAD(&set->tag_list);
3558 out_free_mq_rq_maps:
3559 for (i = 0; i < set->nr_hw_queues; i++)
3560 blk_mq_free_map_and_requests(set, i);
3562 for (i = 0; i < set->nr_maps; i++) {
3563 kfree(set->map[i].mq_map);
3564 set->map[i].mq_map = NULL;
3570 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3572 /* allocate and initialize a tagset for a simple single-queue device */
3573 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
3574 const struct blk_mq_ops *ops, unsigned int queue_depth,
3575 unsigned int set_flags)
3577 memset(set, 0, sizeof(*set));
3579 set->nr_hw_queues = 1;
3581 set->queue_depth = queue_depth;
3582 set->numa_node = NUMA_NO_NODE;
3583 set->flags = set_flags;
3584 return blk_mq_alloc_tag_set(set);
3586 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
3588 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3592 for (i = 0; i < set->nr_hw_queues; i++)
3593 blk_mq_free_map_and_requests(set, i);
3595 if (blk_mq_is_sbitmap_shared(set->flags))
3596 blk_mq_exit_shared_sbitmap(set);
3598 for (j = 0; j < set->nr_maps; j++) {
3599 kfree(set->map[j].mq_map);
3600 set->map[j].mq_map = NULL;
3606 EXPORT_SYMBOL(blk_mq_free_tag_set);
3608 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3610 struct blk_mq_tag_set *set = q->tag_set;
3611 struct blk_mq_hw_ctx *hctx;
3617 if (q->nr_requests == nr)
3620 blk_mq_freeze_queue(q);
3621 blk_mq_quiesce_queue(q);
3624 queue_for_each_hw_ctx(q, hctx, i) {
3628 * If we're using an MQ scheduler, just update the scheduler
3629 * queue depth. This is similar to what the old code would do.
3631 if (!hctx->sched_tags) {
3632 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3634 if (!ret && blk_mq_is_sbitmap_shared(set->flags))
3635 blk_mq_tag_resize_shared_sbitmap(set, nr);
3637 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3639 if (blk_mq_is_sbitmap_shared(set->flags)) {
3640 hctx->sched_tags->bitmap_tags =
3641 &q->sched_bitmap_tags;
3642 hctx->sched_tags->breserved_tags =
3643 &q->sched_breserved_tags;
3648 if (q->elevator && q->elevator->type->ops.depth_updated)
3649 q->elevator->type->ops.depth_updated(hctx);
3652 q->nr_requests = nr;
3653 if (q->elevator && blk_mq_is_sbitmap_shared(set->flags))
3654 sbitmap_queue_resize(&q->sched_bitmap_tags,
3655 nr - set->reserved_tags);
3658 blk_mq_unquiesce_queue(q);
3659 blk_mq_unfreeze_queue(q);
3665 * request_queue and elevator_type pair.
3666 * It is just used by __blk_mq_update_nr_hw_queues to cache
3667 * the elevator_type associated with a request_queue.
3669 struct blk_mq_qe_pair {
3670 struct list_head node;
3671 struct request_queue *q;
3672 struct elevator_type *type;
3676 * Cache the elevator_type in qe pair list and switch the
3677 * io scheduler to 'none'
3679 static bool blk_mq_elv_switch_none(struct list_head *head,
3680 struct request_queue *q)
3682 struct blk_mq_qe_pair *qe;
3687 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3691 INIT_LIST_HEAD(&qe->node);
3693 qe->type = q->elevator->type;
3694 list_add(&qe->node, head);
3696 mutex_lock(&q->sysfs_lock);
3698 * After elevator_switch_mq, the previous elevator_queue will be
3699 * released by elevator_release. The reference of the io scheduler
3700 * module get by elevator_get will also be put. So we need to get
3701 * a reference of the io scheduler module here to prevent it to be
3704 __module_get(qe->type->elevator_owner);
3705 elevator_switch_mq(q, NULL);
3706 mutex_unlock(&q->sysfs_lock);
3711 static void blk_mq_elv_switch_back(struct list_head *head,
3712 struct request_queue *q)
3714 struct blk_mq_qe_pair *qe;
3715 struct elevator_type *t = NULL;
3717 list_for_each_entry(qe, head, node)
3726 list_del(&qe->node);
3729 mutex_lock(&q->sysfs_lock);
3730 elevator_switch_mq(q, t);
3731 mutex_unlock(&q->sysfs_lock);
3734 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3737 struct request_queue *q;
3739 int prev_nr_hw_queues;
3741 lockdep_assert_held(&set->tag_list_lock);
3743 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3744 nr_hw_queues = nr_cpu_ids;
3745 if (nr_hw_queues < 1)
3747 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3750 list_for_each_entry(q, &set->tag_list, tag_set_list)
3751 blk_mq_freeze_queue(q);
3753 * Switch IO scheduler to 'none', cleaning up the data associated
3754 * with the previous scheduler. We will switch back once we are done
3755 * updating the new sw to hw queue mappings.
3757 list_for_each_entry(q, &set->tag_list, tag_set_list)
3758 if (!blk_mq_elv_switch_none(&head, q))
3761 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3762 blk_mq_debugfs_unregister_hctxs(q);
3763 blk_mq_sysfs_unregister(q);
3766 prev_nr_hw_queues = set->nr_hw_queues;
3767 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3771 set->nr_hw_queues = nr_hw_queues;
3773 blk_mq_update_queue_map(set);
3774 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3775 blk_mq_realloc_hw_ctxs(set, q);
3776 if (q->nr_hw_queues != set->nr_hw_queues) {
3777 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3778 nr_hw_queues, prev_nr_hw_queues);
3779 set->nr_hw_queues = prev_nr_hw_queues;
3780 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3783 blk_mq_map_swqueue(q);
3787 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3788 blk_mq_sysfs_register(q);
3789 blk_mq_debugfs_register_hctxs(q);
3793 list_for_each_entry(q, &set->tag_list, tag_set_list)
3794 blk_mq_elv_switch_back(&head, q);
3796 list_for_each_entry(q, &set->tag_list, tag_set_list)
3797 blk_mq_unfreeze_queue(q);
3800 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3802 mutex_lock(&set->tag_list_lock);
3803 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3804 mutex_unlock(&set->tag_list_lock);
3806 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3808 /* Enable polling stats and return whether they were already enabled. */
3809 static bool blk_poll_stats_enable(struct request_queue *q)
3811 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3812 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3814 blk_stat_add_callback(q, q->poll_cb);
3818 static void blk_mq_poll_stats_start(struct request_queue *q)
3821 * We don't arm the callback if polling stats are not enabled or the
3822 * callback is already active.
3824 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3825 blk_stat_is_active(q->poll_cb))
3828 blk_stat_activate_msecs(q->poll_cb, 100);
3831 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3833 struct request_queue *q = cb->data;
3836 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3837 if (cb->stat[bucket].nr_samples)
3838 q->poll_stat[bucket] = cb->stat[bucket];
3842 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3845 unsigned long ret = 0;
3849 * If stats collection isn't on, don't sleep but turn it on for
3852 if (!blk_poll_stats_enable(q))
3856 * As an optimistic guess, use half of the mean service time
3857 * for this type of request. We can (and should) make this smarter.
3858 * For instance, if the completion latencies are tight, we can
3859 * get closer than just half the mean. This is especially
3860 * important on devices where the completion latencies are longer
3861 * than ~10 usec. We do use the stats for the relevant IO size
3862 * if available which does lead to better estimates.
3864 bucket = blk_mq_poll_stats_bkt(rq);
3868 if (q->poll_stat[bucket].nr_samples)
3869 ret = (q->poll_stat[bucket].mean + 1) / 2;
3874 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3877 struct hrtimer_sleeper hs;
3878 enum hrtimer_mode mode;
3882 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3886 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3888 * 0: use half of prev avg
3889 * >0: use this specific value
3891 if (q->poll_nsec > 0)
3892 nsecs = q->poll_nsec;
3894 nsecs = blk_mq_poll_nsecs(q, rq);
3899 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3902 * This will be replaced with the stats tracking code, using
3903 * 'avg_completion_time / 2' as the pre-sleep target.
3907 mode = HRTIMER_MODE_REL;
3908 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3909 hrtimer_set_expires(&hs.timer, kt);
3912 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3914 set_current_state(TASK_UNINTERRUPTIBLE);
3915 hrtimer_sleeper_start_expires(&hs, mode);
3918 hrtimer_cancel(&hs.timer);
3919 mode = HRTIMER_MODE_ABS;
3920 } while (hs.task && !signal_pending(current));
3922 __set_current_state(TASK_RUNNING);
3923 destroy_hrtimer_on_stack(&hs.timer);
3927 static bool blk_mq_poll_hybrid(struct request_queue *q,
3928 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3932 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3935 if (!blk_qc_t_is_internal(cookie))
3936 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3938 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3940 * With scheduling, if the request has completed, we'll
3941 * get a NULL return here, as we clear the sched tag when
3942 * that happens. The request still remains valid, like always,
3943 * so we should be safe with just the NULL check.
3949 return blk_mq_poll_hybrid_sleep(q, rq);
3953 * blk_poll - poll for IO completions
3955 * @cookie: cookie passed back at IO submission time
3956 * @spin: whether to spin for completions
3959 * Poll for completions on the passed in queue. Returns number of
3960 * completed entries found. If @spin is true, then blk_poll will continue
3961 * looping until at least one completion is found, unless the task is
3962 * otherwise marked running (or we need to reschedule).
3964 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3966 struct blk_mq_hw_ctx *hctx;
3969 if (!blk_qc_t_valid(cookie) ||
3970 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3974 blk_flush_plug_list(current->plug, false);
3976 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3979 * If we sleep, have the caller restart the poll loop to reset
3980 * the state. Like for the other success return cases, the
3981 * caller is responsible for checking if the IO completed. If
3982 * the IO isn't complete, we'll get called again and will go
3983 * straight to the busy poll loop. If specified not to spin,
3984 * we also should not sleep.
3986 if (spin && blk_mq_poll_hybrid(q, hctx, cookie))
3989 hctx->poll_considered++;
3991 state = get_current_state();
3995 hctx->poll_invoked++;
3997 ret = q->mq_ops->poll(hctx);
3999 hctx->poll_success++;
4000 __set_current_state(TASK_RUNNING);
4004 if (signal_pending_state(state, current))
4005 __set_current_state(TASK_RUNNING);
4007 if (task_is_running(current))
4009 if (ret < 0 || !spin)
4012 } while (!need_resched());
4014 __set_current_state(TASK_RUNNING);
4017 EXPORT_SYMBOL_GPL(blk_poll);
4019 unsigned int blk_mq_rq_cpu(struct request *rq)
4021 return rq->mq_ctx->cpu;
4023 EXPORT_SYMBOL(blk_mq_rq_cpu);
4025 void blk_mq_cancel_work_sync(struct request_queue *q)
4027 if (queue_is_mq(q)) {
4028 struct blk_mq_hw_ctx *hctx;
4031 cancel_delayed_work_sync(&q->requeue_work);
4033 queue_for_each_hw_ctx(q, hctx, i)
4034 cancel_delayed_work_sync(&hctx->run_work);
4038 static int __init blk_mq_init(void)
4042 for_each_possible_cpu(i)
4043 init_llist_head(&per_cpu(blk_cpu_done, i));
4044 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4046 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4047 "block/softirq:dead", NULL,
4048 blk_softirq_cpu_dead);
4049 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4050 blk_mq_hctx_notify_dead);
4051 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4052 blk_mq_hctx_notify_online,
4053 blk_mq_hctx_notify_offline);
4056 subsys_initcall(blk_mq_init);