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)) ||
461 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
462 return ERR_PTR(-EINVAL);
464 if (hctx_idx >= q->nr_hw_queues)
465 return ERR_PTR(-EIO);
467 ret = blk_queue_enter(q, flags);
472 * Check if the hardware context is actually mapped to anything.
473 * If not tell the caller that it should skip this queue.
476 data.hctx = q->queue_hw_ctx[hctx_idx];
477 if (!blk_mq_hw_queue_mapped(data.hctx))
479 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
480 if (cpu >= nr_cpu_ids)
482 data.ctx = __blk_mq_get_ctx(q, cpu);
485 blk_mq_tag_busy(data.hctx);
488 tag = blk_mq_get_tag(&data);
489 if (tag == BLK_MQ_NO_TAG)
491 return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns);
497 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
499 static void __blk_mq_free_request(struct request *rq)
501 struct request_queue *q = rq->q;
502 struct blk_mq_ctx *ctx = rq->mq_ctx;
503 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
504 const int sched_tag = rq->internal_tag;
506 blk_crypto_free_request(rq);
507 blk_pm_mark_last_busy(rq);
509 if (rq->tag != BLK_MQ_NO_TAG)
510 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
511 if (sched_tag != BLK_MQ_NO_TAG)
512 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
513 blk_mq_sched_restart(hctx);
517 void blk_mq_free_request(struct request *rq)
519 struct request_queue *q = rq->q;
520 struct elevator_queue *e = q->elevator;
521 struct blk_mq_ctx *ctx = rq->mq_ctx;
522 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
524 if (rq->rq_flags & RQF_ELVPRIV) {
525 if (e && e->type->ops.finish_request)
526 e->type->ops.finish_request(rq);
528 put_io_context(rq->elv.icq->ioc);
533 ctx->rq_completed[rq_is_sync(rq)]++;
534 if (rq->rq_flags & RQF_MQ_INFLIGHT)
535 __blk_mq_dec_active_requests(hctx);
537 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
538 laptop_io_completion(q->disk->bdi);
542 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
543 if (refcount_dec_and_test(&rq->ref))
544 __blk_mq_free_request(rq);
546 EXPORT_SYMBOL_GPL(blk_mq_free_request);
548 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
552 if (blk_mq_need_time_stamp(rq))
553 now = ktime_get_ns();
555 if (rq->rq_flags & RQF_STATS) {
556 blk_mq_poll_stats_start(rq->q);
557 blk_stat_add(rq, now);
560 blk_mq_sched_completed_request(rq, now);
562 blk_account_io_done(rq, now);
565 rq_qos_done(rq->q, rq);
566 rq->end_io(rq, error);
568 blk_mq_free_request(rq);
571 EXPORT_SYMBOL(__blk_mq_end_request);
573 void blk_mq_end_request(struct request *rq, blk_status_t error)
575 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
577 __blk_mq_end_request(rq, error);
579 EXPORT_SYMBOL(blk_mq_end_request);
581 static void blk_complete_reqs(struct llist_head *list)
583 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
584 struct request *rq, *next;
586 llist_for_each_entry_safe(rq, next, entry, ipi_list)
587 rq->q->mq_ops->complete(rq);
590 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
592 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
595 static int blk_softirq_cpu_dead(unsigned int cpu)
597 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
601 static void __blk_mq_complete_request_remote(void *data)
603 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
606 static inline bool blk_mq_complete_need_ipi(struct request *rq)
608 int cpu = raw_smp_processor_id();
610 if (!IS_ENABLED(CONFIG_SMP) ||
611 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
614 * With force threaded interrupts enabled, raising softirq from an SMP
615 * function call will always result in waking the ksoftirqd thread.
616 * This is probably worse than completing the request on a different
619 if (force_irqthreads())
622 /* same CPU or cache domain? Complete locally */
623 if (cpu == rq->mq_ctx->cpu ||
624 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
625 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
628 /* don't try to IPI to an offline CPU */
629 return cpu_online(rq->mq_ctx->cpu);
632 static void blk_mq_complete_send_ipi(struct request *rq)
634 struct llist_head *list;
637 cpu = rq->mq_ctx->cpu;
638 list = &per_cpu(blk_cpu_done, cpu);
639 if (llist_add(&rq->ipi_list, list)) {
640 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
641 smp_call_function_single_async(cpu, &rq->csd);
645 static void blk_mq_raise_softirq(struct request *rq)
647 struct llist_head *list;
650 list = this_cpu_ptr(&blk_cpu_done);
651 if (llist_add(&rq->ipi_list, list))
652 raise_softirq(BLOCK_SOFTIRQ);
656 bool blk_mq_complete_request_remote(struct request *rq)
658 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
661 * For a polled request, always complete locallly, it's pointless
662 * to redirect the completion.
664 if (rq->cmd_flags & REQ_HIPRI)
667 if (blk_mq_complete_need_ipi(rq)) {
668 blk_mq_complete_send_ipi(rq);
672 if (rq->q->nr_hw_queues == 1) {
673 blk_mq_raise_softirq(rq);
678 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
681 * blk_mq_complete_request - end I/O on a request
682 * @rq: the request being processed
685 * Complete a request by scheduling the ->complete_rq operation.
687 void blk_mq_complete_request(struct request *rq)
689 if (!blk_mq_complete_request_remote(rq))
690 rq->q->mq_ops->complete(rq);
692 EXPORT_SYMBOL(blk_mq_complete_request);
694 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
695 __releases(hctx->srcu)
697 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
700 srcu_read_unlock(hctx->srcu, srcu_idx);
703 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
704 __acquires(hctx->srcu)
706 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
707 /* shut up gcc false positive */
711 *srcu_idx = srcu_read_lock(hctx->srcu);
715 * blk_mq_start_request - Start processing a request
716 * @rq: Pointer to request to be started
718 * Function used by device drivers to notify the block layer that a request
719 * is going to be processed now, so blk layer can do proper initializations
720 * such as starting the timeout timer.
722 void blk_mq_start_request(struct request *rq)
724 struct request_queue *q = rq->q;
726 trace_block_rq_issue(rq);
728 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
729 rq->io_start_time_ns = ktime_get_ns();
730 rq->stats_sectors = blk_rq_sectors(rq);
731 rq->rq_flags |= RQF_STATS;
735 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
738 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
740 #ifdef CONFIG_BLK_DEV_INTEGRITY
741 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
742 q->integrity.profile->prepare_fn(rq);
745 EXPORT_SYMBOL(blk_mq_start_request);
747 static void __blk_mq_requeue_request(struct request *rq)
749 struct request_queue *q = rq->q;
751 blk_mq_put_driver_tag(rq);
753 trace_block_rq_requeue(rq);
754 rq_qos_requeue(q, rq);
756 if (blk_mq_request_started(rq)) {
757 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
758 rq->rq_flags &= ~RQF_TIMED_OUT;
762 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
764 __blk_mq_requeue_request(rq);
766 /* this request will be re-inserted to io scheduler queue */
767 blk_mq_sched_requeue_request(rq);
769 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
771 EXPORT_SYMBOL(blk_mq_requeue_request);
773 static void blk_mq_requeue_work(struct work_struct *work)
775 struct request_queue *q =
776 container_of(work, struct request_queue, requeue_work.work);
778 struct request *rq, *next;
780 spin_lock_irq(&q->requeue_lock);
781 list_splice_init(&q->requeue_list, &rq_list);
782 spin_unlock_irq(&q->requeue_lock);
784 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
785 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
788 rq->rq_flags &= ~RQF_SOFTBARRIER;
789 list_del_init(&rq->queuelist);
791 * If RQF_DONTPREP, rq has contained some driver specific
792 * data, so insert it to hctx dispatch list to avoid any
795 if (rq->rq_flags & RQF_DONTPREP)
796 blk_mq_request_bypass_insert(rq, false, false);
798 blk_mq_sched_insert_request(rq, true, false, false);
801 while (!list_empty(&rq_list)) {
802 rq = list_entry(rq_list.next, struct request, queuelist);
803 list_del_init(&rq->queuelist);
804 blk_mq_sched_insert_request(rq, false, false, false);
807 blk_mq_run_hw_queues(q, false);
810 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
811 bool kick_requeue_list)
813 struct request_queue *q = rq->q;
817 * We abuse this flag that is otherwise used by the I/O scheduler to
818 * request head insertion from the workqueue.
820 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
822 spin_lock_irqsave(&q->requeue_lock, flags);
824 rq->rq_flags |= RQF_SOFTBARRIER;
825 list_add(&rq->queuelist, &q->requeue_list);
827 list_add_tail(&rq->queuelist, &q->requeue_list);
829 spin_unlock_irqrestore(&q->requeue_lock, flags);
831 if (kick_requeue_list)
832 blk_mq_kick_requeue_list(q);
835 void blk_mq_kick_requeue_list(struct request_queue *q)
837 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
839 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
841 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
844 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
845 msecs_to_jiffies(msecs));
847 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
849 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
851 if (tag < tags->nr_tags) {
852 prefetch(tags->rqs[tag]);
853 return tags->rqs[tag];
858 EXPORT_SYMBOL(blk_mq_tag_to_rq);
860 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
861 void *priv, bool reserved)
864 * If we find a request that isn't idle and the queue matches,
865 * we know the queue is busy. Return false to stop the iteration.
867 if (blk_mq_request_started(rq) && rq->q == hctx->queue) {
877 bool blk_mq_queue_inflight(struct request_queue *q)
881 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
884 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
886 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
888 req->rq_flags |= RQF_TIMED_OUT;
889 if (req->q->mq_ops->timeout) {
890 enum blk_eh_timer_return ret;
892 ret = req->q->mq_ops->timeout(req, reserved);
893 if (ret == BLK_EH_DONE)
895 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
901 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
903 unsigned long deadline;
905 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
907 if (rq->rq_flags & RQF_TIMED_OUT)
910 deadline = READ_ONCE(rq->deadline);
911 if (time_after_eq(jiffies, deadline))
916 else if (time_after(*next, deadline))
921 void blk_mq_put_rq_ref(struct request *rq)
925 else if (refcount_dec_and_test(&rq->ref))
926 __blk_mq_free_request(rq);
929 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
930 struct request *rq, void *priv, bool reserved)
932 unsigned long *next = priv;
935 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
936 * be reallocated underneath the timeout handler's processing, then
937 * the expire check is reliable. If the request is not expired, then
938 * it was completed and reallocated as a new request after returning
939 * from blk_mq_check_expired().
941 if (blk_mq_req_expired(rq, next))
942 blk_mq_rq_timed_out(rq, reserved);
946 static void blk_mq_timeout_work(struct work_struct *work)
948 struct request_queue *q =
949 container_of(work, struct request_queue, timeout_work);
950 unsigned long next = 0;
951 struct blk_mq_hw_ctx *hctx;
954 /* A deadlock might occur if a request is stuck requiring a
955 * timeout at the same time a queue freeze is waiting
956 * completion, since the timeout code would not be able to
957 * acquire the queue reference here.
959 * That's why we don't use blk_queue_enter here; instead, we use
960 * percpu_ref_tryget directly, because we need to be able to
961 * obtain a reference even in the short window between the queue
962 * starting to freeze, by dropping the first reference in
963 * blk_freeze_queue_start, and the moment the last request is
964 * consumed, marked by the instant q_usage_counter reaches
967 if (!percpu_ref_tryget(&q->q_usage_counter))
970 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
973 mod_timer(&q->timeout, next);
976 * Request timeouts are handled as a forward rolling timer. If
977 * we end up here it means that no requests are pending and
978 * also that no request has been pending for a while. Mark
981 queue_for_each_hw_ctx(q, hctx, i) {
982 /* the hctx may be unmapped, so check it here */
983 if (blk_mq_hw_queue_mapped(hctx))
984 blk_mq_tag_idle(hctx);
990 struct flush_busy_ctx_data {
991 struct blk_mq_hw_ctx *hctx;
992 struct list_head *list;
995 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
997 struct flush_busy_ctx_data *flush_data = data;
998 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
999 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1000 enum hctx_type type = hctx->type;
1002 spin_lock(&ctx->lock);
1003 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1004 sbitmap_clear_bit(sb, bitnr);
1005 spin_unlock(&ctx->lock);
1010 * Process software queues that have been marked busy, splicing them
1011 * to the for-dispatch
1013 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1015 struct flush_busy_ctx_data data = {
1020 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1022 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1024 struct dispatch_rq_data {
1025 struct blk_mq_hw_ctx *hctx;
1029 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1032 struct dispatch_rq_data *dispatch_data = data;
1033 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1034 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1035 enum hctx_type type = hctx->type;
1037 spin_lock(&ctx->lock);
1038 if (!list_empty(&ctx->rq_lists[type])) {
1039 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1040 list_del_init(&dispatch_data->rq->queuelist);
1041 if (list_empty(&ctx->rq_lists[type]))
1042 sbitmap_clear_bit(sb, bitnr);
1044 spin_unlock(&ctx->lock);
1046 return !dispatch_data->rq;
1049 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1050 struct blk_mq_ctx *start)
1052 unsigned off = start ? start->index_hw[hctx->type] : 0;
1053 struct dispatch_rq_data data = {
1058 __sbitmap_for_each_set(&hctx->ctx_map, off,
1059 dispatch_rq_from_ctx, &data);
1064 static inline unsigned int queued_to_index(unsigned int queued)
1069 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1072 static bool __blk_mq_get_driver_tag(struct request *rq)
1074 struct sbitmap_queue *bt = rq->mq_hctx->tags->bitmap_tags;
1075 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1078 blk_mq_tag_busy(rq->mq_hctx);
1080 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1081 bt = rq->mq_hctx->tags->breserved_tags;
1084 if (!hctx_may_queue(rq->mq_hctx, bt))
1088 tag = __sbitmap_queue_get(bt);
1089 if (tag == BLK_MQ_NO_TAG)
1092 rq->tag = tag + tag_offset;
1096 bool blk_mq_get_driver_tag(struct request *rq)
1098 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1100 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq))
1103 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1104 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1105 rq->rq_flags |= RQF_MQ_INFLIGHT;
1106 __blk_mq_inc_active_requests(hctx);
1108 hctx->tags->rqs[rq->tag] = rq;
1112 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1113 int flags, void *key)
1115 struct blk_mq_hw_ctx *hctx;
1117 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1119 spin_lock(&hctx->dispatch_wait_lock);
1120 if (!list_empty(&wait->entry)) {
1121 struct sbitmap_queue *sbq;
1123 list_del_init(&wait->entry);
1124 sbq = hctx->tags->bitmap_tags;
1125 atomic_dec(&sbq->ws_active);
1127 spin_unlock(&hctx->dispatch_wait_lock);
1129 blk_mq_run_hw_queue(hctx, true);
1134 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1135 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1136 * restart. For both cases, take care to check the condition again after
1137 * marking us as waiting.
1139 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1142 struct sbitmap_queue *sbq = hctx->tags->bitmap_tags;
1143 struct wait_queue_head *wq;
1144 wait_queue_entry_t *wait;
1147 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1148 blk_mq_sched_mark_restart_hctx(hctx);
1151 * It's possible that a tag was freed in the window between the
1152 * allocation failure and adding the hardware queue to the wait
1155 * Don't clear RESTART here, someone else could have set it.
1156 * At most this will cost an extra queue run.
1158 return blk_mq_get_driver_tag(rq);
1161 wait = &hctx->dispatch_wait;
1162 if (!list_empty_careful(&wait->entry))
1165 wq = &bt_wait_ptr(sbq, hctx)->wait;
1167 spin_lock_irq(&wq->lock);
1168 spin_lock(&hctx->dispatch_wait_lock);
1169 if (!list_empty(&wait->entry)) {
1170 spin_unlock(&hctx->dispatch_wait_lock);
1171 spin_unlock_irq(&wq->lock);
1175 atomic_inc(&sbq->ws_active);
1176 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1177 __add_wait_queue(wq, wait);
1180 * It's possible that a tag was freed in the window between the
1181 * allocation failure and adding the hardware queue to the wait
1184 ret = blk_mq_get_driver_tag(rq);
1186 spin_unlock(&hctx->dispatch_wait_lock);
1187 spin_unlock_irq(&wq->lock);
1192 * We got a tag, remove ourselves from the wait queue to ensure
1193 * someone else gets the wakeup.
1195 list_del_init(&wait->entry);
1196 atomic_dec(&sbq->ws_active);
1197 spin_unlock(&hctx->dispatch_wait_lock);
1198 spin_unlock_irq(&wq->lock);
1203 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1204 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1206 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1207 * - EWMA is one simple way to compute running average value
1208 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1209 * - take 4 as factor for avoiding to get too small(0) result, and this
1210 * factor doesn't matter because EWMA decreases exponentially
1212 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1216 ewma = hctx->dispatch_busy;
1221 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1223 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1224 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1226 hctx->dispatch_busy = ewma;
1229 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1231 static void blk_mq_handle_dev_resource(struct request *rq,
1232 struct list_head *list)
1234 struct request *next =
1235 list_first_entry_or_null(list, struct request, queuelist);
1238 * If an I/O scheduler has been configured and we got a driver tag for
1239 * the next request already, free it.
1242 blk_mq_put_driver_tag(next);
1244 list_add(&rq->queuelist, list);
1245 __blk_mq_requeue_request(rq);
1248 static void blk_mq_handle_zone_resource(struct request *rq,
1249 struct list_head *zone_list)
1252 * If we end up here it is because we cannot dispatch a request to a
1253 * specific zone due to LLD level zone-write locking or other zone
1254 * related resource not being available. In this case, set the request
1255 * aside in zone_list for retrying it later.
1257 list_add(&rq->queuelist, zone_list);
1258 __blk_mq_requeue_request(rq);
1261 enum prep_dispatch {
1263 PREP_DISPATCH_NO_TAG,
1264 PREP_DISPATCH_NO_BUDGET,
1267 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1270 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1271 int budget_token = -1;
1274 budget_token = blk_mq_get_dispatch_budget(rq->q);
1275 if (budget_token < 0) {
1276 blk_mq_put_driver_tag(rq);
1277 return PREP_DISPATCH_NO_BUDGET;
1279 blk_mq_set_rq_budget_token(rq, budget_token);
1282 if (!blk_mq_get_driver_tag(rq)) {
1284 * The initial allocation attempt failed, so we need to
1285 * rerun the hardware queue when a tag is freed. The
1286 * waitqueue takes care of that. If the queue is run
1287 * before we add this entry back on the dispatch list,
1288 * we'll re-run it below.
1290 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1292 * All budgets not got from this function will be put
1293 * together during handling partial dispatch
1296 blk_mq_put_dispatch_budget(rq->q, budget_token);
1297 return PREP_DISPATCH_NO_TAG;
1301 return PREP_DISPATCH_OK;
1304 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1305 static void blk_mq_release_budgets(struct request_queue *q,
1306 struct list_head *list)
1310 list_for_each_entry(rq, list, queuelist) {
1311 int budget_token = blk_mq_get_rq_budget_token(rq);
1313 if (budget_token >= 0)
1314 blk_mq_put_dispatch_budget(q, budget_token);
1319 * Returns true if we did some work AND can potentially do more.
1321 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1322 unsigned int nr_budgets)
1324 enum prep_dispatch prep;
1325 struct request_queue *q = hctx->queue;
1326 struct request *rq, *nxt;
1328 blk_status_t ret = BLK_STS_OK;
1329 LIST_HEAD(zone_list);
1330 bool needs_resource = false;
1332 if (list_empty(list))
1336 * Now process all the entries, sending them to the driver.
1338 errors = queued = 0;
1340 struct blk_mq_queue_data bd;
1342 rq = list_first_entry(list, struct request, queuelist);
1344 WARN_ON_ONCE(hctx != rq->mq_hctx);
1345 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1346 if (prep != PREP_DISPATCH_OK)
1349 list_del_init(&rq->queuelist);
1354 * Flag last if we have no more requests, or if we have more
1355 * but can't assign a driver tag to it.
1357 if (list_empty(list))
1360 nxt = list_first_entry(list, struct request, queuelist);
1361 bd.last = !blk_mq_get_driver_tag(nxt);
1365 * once the request is queued to lld, no need to cover the
1370 ret = q->mq_ops->queue_rq(hctx, &bd);
1375 case BLK_STS_RESOURCE:
1376 needs_resource = true;
1378 case BLK_STS_DEV_RESOURCE:
1379 blk_mq_handle_dev_resource(rq, list);
1381 case BLK_STS_ZONE_RESOURCE:
1383 * Move the request to zone_list and keep going through
1384 * the dispatch list to find more requests the drive can
1387 blk_mq_handle_zone_resource(rq, &zone_list);
1388 needs_resource = true;
1392 blk_mq_end_request(rq, ret);
1394 } while (!list_empty(list));
1396 if (!list_empty(&zone_list))
1397 list_splice_tail_init(&zone_list, list);
1399 hctx->dispatched[queued_to_index(queued)]++;
1401 /* If we didn't flush the entire list, we could have told the driver
1402 * there was more coming, but that turned out to be a lie.
1404 if ((!list_empty(list) || errors || needs_resource ||
1405 ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued)
1406 q->mq_ops->commit_rqs(hctx);
1408 * Any items that need requeuing? Stuff them into hctx->dispatch,
1409 * that is where we will continue on next queue run.
1411 if (!list_empty(list)) {
1413 /* For non-shared tags, the RESTART check will suffice */
1414 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1415 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1418 blk_mq_release_budgets(q, list);
1420 spin_lock(&hctx->lock);
1421 list_splice_tail_init(list, &hctx->dispatch);
1422 spin_unlock(&hctx->lock);
1425 * Order adding requests to hctx->dispatch and checking
1426 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1427 * in blk_mq_sched_restart(). Avoid restart code path to
1428 * miss the new added requests to hctx->dispatch, meantime
1429 * SCHED_RESTART is observed here.
1434 * If SCHED_RESTART was set by the caller of this function and
1435 * it is no longer set that means that it was cleared by another
1436 * thread and hence that a queue rerun is needed.
1438 * If 'no_tag' is set, that means that we failed getting
1439 * a driver tag with an I/O scheduler attached. If our dispatch
1440 * waitqueue is no longer active, ensure that we run the queue
1441 * AFTER adding our entries back to the list.
1443 * If no I/O scheduler has been configured it is possible that
1444 * the hardware queue got stopped and restarted before requests
1445 * were pushed back onto the dispatch list. Rerun the queue to
1446 * avoid starvation. Notes:
1447 * - blk_mq_run_hw_queue() checks whether or not a queue has
1448 * been stopped before rerunning a queue.
1449 * - Some but not all block drivers stop a queue before
1450 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1453 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1454 * bit is set, run queue after a delay to avoid IO stalls
1455 * that could otherwise occur if the queue is idle. We'll do
1456 * similar if we couldn't get budget or couldn't lock a zone
1457 * and SCHED_RESTART is set.
1459 needs_restart = blk_mq_sched_needs_restart(hctx);
1460 if (prep == PREP_DISPATCH_NO_BUDGET)
1461 needs_resource = true;
1462 if (!needs_restart ||
1463 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1464 blk_mq_run_hw_queue(hctx, true);
1465 else if (needs_restart && needs_resource)
1466 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1468 blk_mq_update_dispatch_busy(hctx, true);
1471 blk_mq_update_dispatch_busy(hctx, false);
1473 return (queued + errors) != 0;
1477 * __blk_mq_run_hw_queue - Run a hardware queue.
1478 * @hctx: Pointer to the hardware queue to run.
1480 * Send pending requests to the hardware.
1482 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1487 * We can't run the queue inline with ints disabled. Ensure that
1488 * we catch bad users of this early.
1490 WARN_ON_ONCE(in_interrupt());
1492 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1494 hctx_lock(hctx, &srcu_idx);
1495 blk_mq_sched_dispatch_requests(hctx);
1496 hctx_unlock(hctx, srcu_idx);
1499 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1501 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1503 if (cpu >= nr_cpu_ids)
1504 cpu = cpumask_first(hctx->cpumask);
1509 * It'd be great if the workqueue API had a way to pass
1510 * in a mask and had some smarts for more clever placement.
1511 * For now we just round-robin here, switching for every
1512 * BLK_MQ_CPU_WORK_BATCH queued items.
1514 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1517 int next_cpu = hctx->next_cpu;
1519 if (hctx->queue->nr_hw_queues == 1)
1520 return WORK_CPU_UNBOUND;
1522 if (--hctx->next_cpu_batch <= 0) {
1524 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1526 if (next_cpu >= nr_cpu_ids)
1527 next_cpu = blk_mq_first_mapped_cpu(hctx);
1528 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1532 * Do unbound schedule if we can't find a online CPU for this hctx,
1533 * and it should only happen in the path of handling CPU DEAD.
1535 if (!cpu_online(next_cpu)) {
1542 * Make sure to re-select CPU next time once after CPUs
1543 * in hctx->cpumask become online again.
1545 hctx->next_cpu = next_cpu;
1546 hctx->next_cpu_batch = 1;
1547 return WORK_CPU_UNBOUND;
1550 hctx->next_cpu = next_cpu;
1555 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1556 * @hctx: Pointer to the hardware queue to run.
1557 * @async: If we want to run the queue asynchronously.
1558 * @msecs: Milliseconds of delay to wait before running the queue.
1560 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1561 * with a delay of @msecs.
1563 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1564 unsigned long msecs)
1566 if (unlikely(blk_mq_hctx_stopped(hctx)))
1569 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1570 int cpu = get_cpu();
1571 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1572 __blk_mq_run_hw_queue(hctx);
1580 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1581 msecs_to_jiffies(msecs));
1585 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1586 * @hctx: Pointer to the hardware queue to run.
1587 * @msecs: Milliseconds of delay to wait before running the queue.
1589 * Run a hardware queue asynchronously with a delay of @msecs.
1591 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1593 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1595 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1598 * blk_mq_run_hw_queue - Start to run a hardware queue.
1599 * @hctx: Pointer to the hardware queue to run.
1600 * @async: If we want to run the queue asynchronously.
1602 * Check if the request queue is not in a quiesced state and if there are
1603 * pending requests to be sent. If this is true, run the queue to send requests
1606 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1612 * When queue is quiesced, we may be switching io scheduler, or
1613 * updating nr_hw_queues, or other things, and we can't run queue
1614 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1616 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1619 hctx_lock(hctx, &srcu_idx);
1620 need_run = !blk_queue_quiesced(hctx->queue) &&
1621 blk_mq_hctx_has_pending(hctx);
1622 hctx_unlock(hctx, srcu_idx);
1625 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1627 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1630 * Is the request queue handled by an IO scheduler that does not respect
1631 * hardware queues when dispatching?
1633 static bool blk_mq_has_sqsched(struct request_queue *q)
1635 struct elevator_queue *e = q->elevator;
1637 if (e && e->type->ops.dispatch_request &&
1638 !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE))
1644 * Return prefered queue to dispatch from (if any) for non-mq aware IO
1647 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
1649 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
1651 * If the IO scheduler does not respect hardware queues when
1652 * dispatching, we just don't bother with multiple HW queues and
1653 * dispatch from hctx for the current CPU since running multiple queues
1654 * just causes lock contention inside the scheduler and pointless cache
1657 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, 0, ctx);
1659 if (!blk_mq_hctx_stopped(hctx))
1665 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
1666 * @q: Pointer to the request queue to run.
1667 * @async: If we want to run the queue asynchronously.
1669 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1671 struct blk_mq_hw_ctx *hctx, *sq_hctx;
1675 if (blk_mq_has_sqsched(q))
1676 sq_hctx = blk_mq_get_sq_hctx(q);
1677 queue_for_each_hw_ctx(q, hctx, i) {
1678 if (blk_mq_hctx_stopped(hctx))
1681 * Dispatch from this hctx either if there's no hctx preferred
1682 * by IO scheduler or if it has requests that bypass the
1685 if (!sq_hctx || sq_hctx == hctx ||
1686 !list_empty_careful(&hctx->dispatch))
1687 blk_mq_run_hw_queue(hctx, async);
1690 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1693 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1694 * @q: Pointer to the request queue to run.
1695 * @msecs: Milliseconds of delay to wait before running the queues.
1697 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1699 struct blk_mq_hw_ctx *hctx, *sq_hctx;
1703 if (blk_mq_has_sqsched(q))
1704 sq_hctx = blk_mq_get_sq_hctx(q);
1705 queue_for_each_hw_ctx(q, hctx, i) {
1706 if (blk_mq_hctx_stopped(hctx))
1709 * Dispatch from this hctx either if there's no hctx preferred
1710 * by IO scheduler or if it has requests that bypass the
1713 if (!sq_hctx || sq_hctx == hctx ||
1714 !list_empty_careful(&hctx->dispatch))
1715 blk_mq_delay_run_hw_queue(hctx, msecs);
1718 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1721 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1722 * @q: request queue.
1724 * The caller is responsible for serializing this function against
1725 * blk_mq_{start,stop}_hw_queue().
1727 bool blk_mq_queue_stopped(struct request_queue *q)
1729 struct blk_mq_hw_ctx *hctx;
1732 queue_for_each_hw_ctx(q, hctx, i)
1733 if (blk_mq_hctx_stopped(hctx))
1738 EXPORT_SYMBOL(blk_mq_queue_stopped);
1741 * This function is often used for pausing .queue_rq() by driver when
1742 * there isn't enough resource or some conditions aren't satisfied, and
1743 * BLK_STS_RESOURCE is usually returned.
1745 * We do not guarantee that dispatch can be drained or blocked
1746 * after blk_mq_stop_hw_queue() returns. Please use
1747 * blk_mq_quiesce_queue() for that requirement.
1749 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1751 cancel_delayed_work(&hctx->run_work);
1753 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1755 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1758 * This function is often used for pausing .queue_rq() by driver when
1759 * there isn't enough resource or some conditions aren't satisfied, and
1760 * BLK_STS_RESOURCE is usually returned.
1762 * We do not guarantee that dispatch can be drained or blocked
1763 * after blk_mq_stop_hw_queues() returns. Please use
1764 * blk_mq_quiesce_queue() for that requirement.
1766 void blk_mq_stop_hw_queues(struct request_queue *q)
1768 struct blk_mq_hw_ctx *hctx;
1771 queue_for_each_hw_ctx(q, hctx, i)
1772 blk_mq_stop_hw_queue(hctx);
1774 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1776 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1778 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1780 blk_mq_run_hw_queue(hctx, false);
1782 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1784 void blk_mq_start_hw_queues(struct request_queue *q)
1786 struct blk_mq_hw_ctx *hctx;
1789 queue_for_each_hw_ctx(q, hctx, i)
1790 blk_mq_start_hw_queue(hctx);
1792 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1794 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1796 if (!blk_mq_hctx_stopped(hctx))
1799 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1800 blk_mq_run_hw_queue(hctx, async);
1802 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1804 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1806 struct blk_mq_hw_ctx *hctx;
1809 queue_for_each_hw_ctx(q, hctx, i)
1810 blk_mq_start_stopped_hw_queue(hctx, async);
1812 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1814 static void blk_mq_run_work_fn(struct work_struct *work)
1816 struct blk_mq_hw_ctx *hctx;
1818 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1821 * If we are stopped, don't run the queue.
1823 if (blk_mq_hctx_stopped(hctx))
1826 __blk_mq_run_hw_queue(hctx);
1829 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1833 struct blk_mq_ctx *ctx = rq->mq_ctx;
1834 enum hctx_type type = hctx->type;
1836 lockdep_assert_held(&ctx->lock);
1838 trace_block_rq_insert(rq);
1841 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1843 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1846 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1849 struct blk_mq_ctx *ctx = rq->mq_ctx;
1851 lockdep_assert_held(&ctx->lock);
1853 __blk_mq_insert_req_list(hctx, rq, at_head);
1854 blk_mq_hctx_mark_pending(hctx, ctx);
1858 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1859 * @rq: Pointer to request to be inserted.
1860 * @at_head: true if the request should be inserted at the head of the list.
1861 * @run_queue: If we should run the hardware queue after inserting the request.
1863 * Should only be used carefully, when the caller knows we want to
1864 * bypass a potential IO scheduler on the target device.
1866 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1869 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1871 spin_lock(&hctx->lock);
1873 list_add(&rq->queuelist, &hctx->dispatch);
1875 list_add_tail(&rq->queuelist, &hctx->dispatch);
1876 spin_unlock(&hctx->lock);
1879 blk_mq_run_hw_queue(hctx, false);
1882 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1883 struct list_head *list)
1887 enum hctx_type type = hctx->type;
1890 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1893 list_for_each_entry(rq, list, queuelist) {
1894 BUG_ON(rq->mq_ctx != ctx);
1895 trace_block_rq_insert(rq);
1898 spin_lock(&ctx->lock);
1899 list_splice_tail_init(list, &ctx->rq_lists[type]);
1900 blk_mq_hctx_mark_pending(hctx, ctx);
1901 spin_unlock(&ctx->lock);
1904 static int plug_rq_cmp(void *priv, const struct list_head *a,
1905 const struct list_head *b)
1907 struct request *rqa = container_of(a, struct request, queuelist);
1908 struct request *rqb = container_of(b, struct request, queuelist);
1910 if (rqa->mq_ctx != rqb->mq_ctx)
1911 return rqa->mq_ctx > rqb->mq_ctx;
1912 if (rqa->mq_hctx != rqb->mq_hctx)
1913 return rqa->mq_hctx > rqb->mq_hctx;
1915 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1918 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1922 if (list_empty(&plug->mq_list))
1924 list_splice_init(&plug->mq_list, &list);
1926 if (plug->rq_count > 2 && plug->multiple_queues)
1927 list_sort(NULL, &list, plug_rq_cmp);
1932 struct list_head rq_list;
1933 struct request *rq, *head_rq = list_entry_rq(list.next);
1934 struct list_head *pos = &head_rq->queuelist; /* skip first */
1935 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1936 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1937 unsigned int depth = 1;
1939 list_for_each_continue(pos, &list) {
1940 rq = list_entry_rq(pos);
1942 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1947 list_cut_before(&rq_list, &list, pos);
1948 trace_block_unplug(head_rq->q, depth, !from_schedule);
1949 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1951 } while(!list_empty(&list));
1954 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1955 unsigned int nr_segs)
1959 if (bio->bi_opf & REQ_RAHEAD)
1960 rq->cmd_flags |= REQ_FAILFAST_MASK;
1962 rq->__sector = bio->bi_iter.bi_sector;
1963 rq->write_hint = bio->bi_write_hint;
1964 blk_rq_bio_prep(rq, bio, nr_segs);
1966 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
1967 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1970 blk_account_io_start(rq);
1973 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1975 blk_qc_t *cookie, bool last)
1977 struct request_queue *q = rq->q;
1978 struct blk_mq_queue_data bd = {
1982 blk_qc_t new_cookie;
1985 new_cookie = request_to_qc_t(hctx, rq);
1988 * For OK queue, we are done. For error, caller may kill it.
1989 * Any other error (busy), just add it to our list as we
1990 * previously would have done.
1992 ret = q->mq_ops->queue_rq(hctx, &bd);
1995 blk_mq_update_dispatch_busy(hctx, false);
1996 *cookie = new_cookie;
1998 case BLK_STS_RESOURCE:
1999 case BLK_STS_DEV_RESOURCE:
2000 blk_mq_update_dispatch_busy(hctx, true);
2001 __blk_mq_requeue_request(rq);
2004 blk_mq_update_dispatch_busy(hctx, false);
2005 *cookie = BLK_QC_T_NONE;
2012 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2015 bool bypass_insert, bool last)
2017 struct request_queue *q = rq->q;
2018 bool run_queue = true;
2022 * RCU or SRCU read lock is needed before checking quiesced flag.
2024 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2025 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2026 * and avoid driver to try to dispatch again.
2028 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2030 bypass_insert = false;
2034 if (q->elevator && !bypass_insert)
2037 budget_token = blk_mq_get_dispatch_budget(q);
2038 if (budget_token < 0)
2041 blk_mq_set_rq_budget_token(rq, budget_token);
2043 if (!blk_mq_get_driver_tag(rq)) {
2044 blk_mq_put_dispatch_budget(q, budget_token);
2048 return __blk_mq_issue_directly(hctx, rq, cookie, last);
2051 return BLK_STS_RESOURCE;
2053 blk_mq_sched_insert_request(rq, false, run_queue, false);
2059 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2060 * @hctx: Pointer of the associated hardware queue.
2061 * @rq: Pointer to request to be sent.
2062 * @cookie: Request queue cookie.
2064 * If the device has enough resources to accept a new request now, send the
2065 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2066 * we can try send it another time in the future. Requests inserted at this
2067 * queue have higher priority.
2069 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2070 struct request *rq, blk_qc_t *cookie)
2075 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2077 hctx_lock(hctx, &srcu_idx);
2079 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
2080 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2081 blk_mq_request_bypass_insert(rq, false, true);
2082 else if (ret != BLK_STS_OK)
2083 blk_mq_end_request(rq, ret);
2085 hctx_unlock(hctx, srcu_idx);
2088 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2092 blk_qc_t unused_cookie;
2093 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2095 hctx_lock(hctx, &srcu_idx);
2096 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
2097 hctx_unlock(hctx, srcu_idx);
2102 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2103 struct list_head *list)
2108 while (!list_empty(list)) {
2110 struct request *rq = list_first_entry(list, struct request,
2113 list_del_init(&rq->queuelist);
2114 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2115 if (ret != BLK_STS_OK) {
2117 if (ret == BLK_STS_RESOURCE ||
2118 ret == BLK_STS_DEV_RESOURCE) {
2119 blk_mq_request_bypass_insert(rq, false,
2123 blk_mq_end_request(rq, ret);
2129 * If we didn't flush the entire list, we could have told
2130 * the driver there was more coming, but that turned out to
2133 if ((!list_empty(list) || errors) &&
2134 hctx->queue->mq_ops->commit_rqs && queued)
2135 hctx->queue->mq_ops->commit_rqs(hctx);
2138 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2140 list_add_tail(&rq->queuelist, &plug->mq_list);
2142 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
2143 struct request *tmp;
2145 tmp = list_first_entry(&plug->mq_list, struct request,
2147 if (tmp->q != rq->q)
2148 plug->multiple_queues = true;
2153 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
2154 * queues. This is important for md arrays to benefit from merging
2157 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
2159 if (plug->multiple_queues)
2160 return BLK_MAX_REQUEST_COUNT * 2;
2161 return BLK_MAX_REQUEST_COUNT;
2165 * blk_mq_submit_bio - Create and send a request to block device.
2166 * @bio: Bio pointer.
2168 * Builds up a request structure from @q and @bio and send to the device. The
2169 * request may not be queued directly to hardware if:
2170 * * This request can be merged with another one
2171 * * We want to place request at plug queue for possible future merging
2172 * * There is an IO scheduler active at this queue
2174 * It will not queue the request if there is an error with the bio, or at the
2177 * Returns: Request queue cookie.
2179 blk_qc_t blk_mq_submit_bio(struct bio *bio)
2181 struct request_queue *q = bio->bi_bdev->bd_disk->queue;
2182 const int is_sync = op_is_sync(bio->bi_opf);
2183 const int is_flush_fua = op_is_flush(bio->bi_opf);
2184 struct blk_mq_alloc_data data = {
2188 struct blk_plug *plug;
2189 struct request *same_queue_rq = NULL;
2190 unsigned int nr_segs;
2195 blk_queue_bounce(q, &bio);
2196 __blk_queue_split(&bio, &nr_segs);
2200 if (!bio_integrity_prep(bio))
2203 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2204 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2207 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2210 rq_qos_throttle(q, bio);
2212 hipri = bio->bi_opf & REQ_HIPRI;
2214 data.cmd_flags = bio->bi_opf;
2215 rq = __blk_mq_alloc_request(&data);
2216 if (unlikely(!rq)) {
2217 rq_qos_cleanup(q, bio);
2218 if (bio->bi_opf & REQ_NOWAIT)
2219 bio_wouldblock_error(bio);
2223 trace_block_getrq(bio);
2225 rq_qos_track(q, rq, bio);
2227 cookie = request_to_qc_t(data.hctx, rq);
2229 blk_mq_bio_to_request(rq, bio, nr_segs);
2231 ret = blk_crypto_rq_get_keyslot(rq);
2232 if (ret != BLK_STS_OK) {
2233 bio->bi_status = ret;
2235 blk_mq_free_request(rq);
2236 return BLK_QC_T_NONE;
2239 plug = blk_mq_plug(q, bio);
2240 if (unlikely(is_flush_fua)) {
2241 /* Bypass scheduler for flush requests */
2242 blk_insert_flush(rq);
2243 blk_mq_run_hw_queue(data.hctx, true);
2244 } else if (plug && (q->nr_hw_queues == 1 ||
2245 blk_mq_is_sbitmap_shared(rq->mq_hctx->flags) ||
2246 q->mq_ops->commit_rqs || !blk_queue_nonrot(q))) {
2248 * Use plugging if we have a ->commit_rqs() hook as well, as
2249 * we know the driver uses bd->last in a smart fashion.
2251 * Use normal plugging if this disk is slow HDD, as sequential
2252 * IO may benefit a lot from plug merging.
2254 unsigned int request_count = plug->rq_count;
2255 struct request *last = NULL;
2258 trace_block_plug(q);
2260 last = list_entry_rq(plug->mq_list.prev);
2262 if (request_count >= blk_plug_max_rq_count(plug) || (last &&
2263 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2264 blk_flush_plug_list(plug, false);
2265 trace_block_plug(q);
2268 blk_add_rq_to_plug(plug, rq);
2269 } else if (q->elevator) {
2270 /* Insert the request at the IO scheduler queue */
2271 blk_mq_sched_insert_request(rq, false, true, true);
2272 } else if (plug && !blk_queue_nomerges(q)) {
2274 * We do limited plugging. If the bio can be merged, do that.
2275 * Otherwise the existing request in the plug list will be
2276 * issued. So the plug list will have one request at most
2277 * The plug list might get flushed before this. If that happens,
2278 * the plug list is empty, and same_queue_rq is invalid.
2280 if (list_empty(&plug->mq_list))
2281 same_queue_rq = NULL;
2282 if (same_queue_rq) {
2283 list_del_init(&same_queue_rq->queuelist);
2286 blk_add_rq_to_plug(plug, rq);
2287 trace_block_plug(q);
2289 if (same_queue_rq) {
2290 data.hctx = same_queue_rq->mq_hctx;
2291 trace_block_unplug(q, 1, true);
2292 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2295 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2296 !data.hctx->dispatch_busy) {
2298 * There is no scheduler and we can try to send directly
2301 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2304 blk_mq_sched_insert_request(rq, false, true, true);
2308 return BLK_QC_T_NONE;
2312 return BLK_QC_T_NONE;
2315 static size_t order_to_size(unsigned int order)
2317 return (size_t)PAGE_SIZE << order;
2320 /* called before freeing request pool in @tags */
2321 static void blk_mq_clear_rq_mapping(struct blk_mq_tag_set *set,
2322 struct blk_mq_tags *tags, unsigned int hctx_idx)
2324 struct blk_mq_tags *drv_tags = set->tags[hctx_idx];
2326 unsigned long flags;
2328 list_for_each_entry(page, &tags->page_list, lru) {
2329 unsigned long start = (unsigned long)page_address(page);
2330 unsigned long end = start + order_to_size(page->private);
2333 for (i = 0; i < set->queue_depth; i++) {
2334 struct request *rq = drv_tags->rqs[i];
2335 unsigned long rq_addr = (unsigned long)rq;
2337 if (rq_addr >= start && rq_addr < end) {
2338 WARN_ON_ONCE(refcount_read(&rq->ref) != 0);
2339 cmpxchg(&drv_tags->rqs[i], rq, NULL);
2345 * Wait until all pending iteration is done.
2347 * Request reference is cleared and it is guaranteed to be observed
2348 * after the ->lock is released.
2350 spin_lock_irqsave(&drv_tags->lock, flags);
2351 spin_unlock_irqrestore(&drv_tags->lock, flags);
2354 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2355 unsigned int hctx_idx)
2359 if (tags->rqs && set->ops->exit_request) {
2362 for (i = 0; i < tags->nr_tags; i++) {
2363 struct request *rq = tags->static_rqs[i];
2367 set->ops->exit_request(set, rq, hctx_idx);
2368 tags->static_rqs[i] = NULL;
2372 blk_mq_clear_rq_mapping(set, tags, hctx_idx);
2374 while (!list_empty(&tags->page_list)) {
2375 page = list_first_entry(&tags->page_list, struct page, lru);
2376 list_del_init(&page->lru);
2378 * Remove kmemleak object previously allocated in
2379 * blk_mq_alloc_rqs().
2381 kmemleak_free(page_address(page));
2382 __free_pages(page, page->private);
2386 void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags)
2390 kfree(tags->static_rqs);
2391 tags->static_rqs = NULL;
2393 blk_mq_free_tags(tags, flags);
2396 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2397 unsigned int hctx_idx,
2398 unsigned int nr_tags,
2399 unsigned int reserved_tags,
2402 struct blk_mq_tags *tags;
2405 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2406 if (node == NUMA_NO_NODE)
2407 node = set->numa_node;
2409 tags = blk_mq_init_tags(nr_tags, reserved_tags, node, flags);
2413 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2414 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2417 blk_mq_free_tags(tags, flags);
2421 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2422 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2424 if (!tags->static_rqs) {
2426 blk_mq_free_tags(tags, flags);
2433 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2434 unsigned int hctx_idx, int node)
2438 if (set->ops->init_request) {
2439 ret = set->ops->init_request(set, rq, hctx_idx, node);
2444 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2448 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2449 unsigned int hctx_idx, unsigned int depth)
2451 unsigned int i, j, entries_per_page, max_order = 4;
2452 size_t rq_size, left;
2455 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2456 if (node == NUMA_NO_NODE)
2457 node = set->numa_node;
2459 INIT_LIST_HEAD(&tags->page_list);
2462 * rq_size is the size of the request plus driver payload, rounded
2463 * to the cacheline size
2465 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2467 left = rq_size * depth;
2469 for (i = 0; i < depth; ) {
2470 int this_order = max_order;
2475 while (this_order && left < order_to_size(this_order - 1))
2479 page = alloc_pages_node(node,
2480 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2486 if (order_to_size(this_order) < rq_size)
2493 page->private = this_order;
2494 list_add_tail(&page->lru, &tags->page_list);
2496 p = page_address(page);
2498 * Allow kmemleak to scan these pages as they contain pointers
2499 * to additional allocations like via ops->init_request().
2501 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2502 entries_per_page = order_to_size(this_order) / rq_size;
2503 to_do = min(entries_per_page, depth - i);
2504 left -= to_do * rq_size;
2505 for (j = 0; j < to_do; j++) {
2506 struct request *rq = p;
2508 tags->static_rqs[i] = rq;
2509 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2510 tags->static_rqs[i] = NULL;
2521 blk_mq_free_rqs(set, tags, hctx_idx);
2525 struct rq_iter_data {
2526 struct blk_mq_hw_ctx *hctx;
2530 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2532 struct rq_iter_data *iter_data = data;
2534 if (rq->mq_hctx != iter_data->hctx)
2536 iter_data->has_rq = true;
2540 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2542 struct blk_mq_tags *tags = hctx->sched_tags ?
2543 hctx->sched_tags : hctx->tags;
2544 struct rq_iter_data data = {
2548 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2552 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2553 struct blk_mq_hw_ctx *hctx)
2555 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2557 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2562 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2564 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2565 struct blk_mq_hw_ctx, cpuhp_online);
2567 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2568 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2572 * Prevent new request from being allocated on the current hctx.
2574 * The smp_mb__after_atomic() Pairs with the implied barrier in
2575 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2576 * seen once we return from the tag allocator.
2578 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2579 smp_mb__after_atomic();
2582 * Try to grab a reference to the queue and wait for any outstanding
2583 * requests. If we could not grab a reference the queue has been
2584 * frozen and there are no requests.
2586 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2587 while (blk_mq_hctx_has_requests(hctx))
2589 percpu_ref_put(&hctx->queue->q_usage_counter);
2595 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2597 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2598 struct blk_mq_hw_ctx, cpuhp_online);
2600 if (cpumask_test_cpu(cpu, hctx->cpumask))
2601 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2606 * 'cpu' is going away. splice any existing rq_list entries from this
2607 * software queue to the hw queue dispatch list, and ensure that it
2610 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2612 struct blk_mq_hw_ctx *hctx;
2613 struct blk_mq_ctx *ctx;
2615 enum hctx_type type;
2617 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2618 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2621 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2624 spin_lock(&ctx->lock);
2625 if (!list_empty(&ctx->rq_lists[type])) {
2626 list_splice_init(&ctx->rq_lists[type], &tmp);
2627 blk_mq_hctx_clear_pending(hctx, ctx);
2629 spin_unlock(&ctx->lock);
2631 if (list_empty(&tmp))
2634 spin_lock(&hctx->lock);
2635 list_splice_tail_init(&tmp, &hctx->dispatch);
2636 spin_unlock(&hctx->lock);
2638 blk_mq_run_hw_queue(hctx, true);
2642 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2644 if (!(hctx->flags & BLK_MQ_F_STACKING))
2645 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2646 &hctx->cpuhp_online);
2647 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2652 * Before freeing hw queue, clearing the flush request reference in
2653 * tags->rqs[] for avoiding potential UAF.
2655 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
2656 unsigned int queue_depth, struct request *flush_rq)
2659 unsigned long flags;
2661 /* The hw queue may not be mapped yet */
2665 WARN_ON_ONCE(refcount_read(&flush_rq->ref) != 0);
2667 for (i = 0; i < queue_depth; i++)
2668 cmpxchg(&tags->rqs[i], flush_rq, NULL);
2671 * Wait until all pending iteration is done.
2673 * Request reference is cleared and it is guaranteed to be observed
2674 * after the ->lock is released.
2676 spin_lock_irqsave(&tags->lock, flags);
2677 spin_unlock_irqrestore(&tags->lock, flags);
2680 /* hctx->ctxs will be freed in queue's release handler */
2681 static void blk_mq_exit_hctx(struct request_queue *q,
2682 struct blk_mq_tag_set *set,
2683 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2685 struct request *flush_rq = hctx->fq->flush_rq;
2687 if (blk_mq_hw_queue_mapped(hctx))
2688 blk_mq_tag_idle(hctx);
2690 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
2691 set->queue_depth, flush_rq);
2692 if (set->ops->exit_request)
2693 set->ops->exit_request(set, flush_rq, hctx_idx);
2695 if (set->ops->exit_hctx)
2696 set->ops->exit_hctx(hctx, hctx_idx);
2698 blk_mq_remove_cpuhp(hctx);
2700 spin_lock(&q->unused_hctx_lock);
2701 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2702 spin_unlock(&q->unused_hctx_lock);
2705 static void blk_mq_exit_hw_queues(struct request_queue *q,
2706 struct blk_mq_tag_set *set, int nr_queue)
2708 struct blk_mq_hw_ctx *hctx;
2711 queue_for_each_hw_ctx(q, hctx, i) {
2714 blk_mq_debugfs_unregister_hctx(hctx);
2715 blk_mq_exit_hctx(q, set, hctx, i);
2719 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2721 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2723 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2724 __alignof__(struct blk_mq_hw_ctx)) !=
2725 sizeof(struct blk_mq_hw_ctx));
2727 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2728 hw_ctx_size += sizeof(struct srcu_struct);
2733 static int blk_mq_init_hctx(struct request_queue *q,
2734 struct blk_mq_tag_set *set,
2735 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2737 hctx->queue_num = hctx_idx;
2739 if (!(hctx->flags & BLK_MQ_F_STACKING))
2740 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2741 &hctx->cpuhp_online);
2742 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2744 hctx->tags = set->tags[hctx_idx];
2746 if (set->ops->init_hctx &&
2747 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2748 goto unregister_cpu_notifier;
2750 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2756 if (set->ops->exit_hctx)
2757 set->ops->exit_hctx(hctx, hctx_idx);
2758 unregister_cpu_notifier:
2759 blk_mq_remove_cpuhp(hctx);
2763 static struct blk_mq_hw_ctx *
2764 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2767 struct blk_mq_hw_ctx *hctx;
2768 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2770 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2772 goto fail_alloc_hctx;
2774 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2777 atomic_set(&hctx->nr_active, 0);
2778 if (node == NUMA_NO_NODE)
2779 node = set->numa_node;
2780 hctx->numa_node = node;
2782 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2783 spin_lock_init(&hctx->lock);
2784 INIT_LIST_HEAD(&hctx->dispatch);
2786 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
2788 INIT_LIST_HEAD(&hctx->hctx_list);
2791 * Allocate space for all possible cpus to avoid allocation at
2794 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2799 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2800 gfp, node, false, false))
2804 spin_lock_init(&hctx->dispatch_wait_lock);
2805 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2806 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2808 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2812 if (hctx->flags & BLK_MQ_F_BLOCKING)
2813 init_srcu_struct(hctx->srcu);
2814 blk_mq_hctx_kobj_init(hctx);
2819 sbitmap_free(&hctx->ctx_map);
2823 free_cpumask_var(hctx->cpumask);
2830 static void blk_mq_init_cpu_queues(struct request_queue *q,
2831 unsigned int nr_hw_queues)
2833 struct blk_mq_tag_set *set = q->tag_set;
2836 for_each_possible_cpu(i) {
2837 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2838 struct blk_mq_hw_ctx *hctx;
2842 spin_lock_init(&__ctx->lock);
2843 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2844 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2849 * Set local node, IFF we have more than one hw queue. If
2850 * not, we remain on the home node of the device
2852 for (j = 0; j < set->nr_maps; j++) {
2853 hctx = blk_mq_map_queue_type(q, j, i);
2854 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2855 hctx->numa_node = cpu_to_node(i);
2860 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2863 unsigned int flags = set->flags;
2866 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2867 set->queue_depth, set->reserved_tags, flags);
2868 if (!set->tags[hctx_idx])
2871 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2876 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2877 set->tags[hctx_idx] = NULL;
2881 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2882 unsigned int hctx_idx)
2884 unsigned int flags = set->flags;
2886 if (set->tags && set->tags[hctx_idx]) {
2887 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2888 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2889 set->tags[hctx_idx] = NULL;
2893 static void blk_mq_map_swqueue(struct request_queue *q)
2895 unsigned int i, j, hctx_idx;
2896 struct blk_mq_hw_ctx *hctx;
2897 struct blk_mq_ctx *ctx;
2898 struct blk_mq_tag_set *set = q->tag_set;
2900 queue_for_each_hw_ctx(q, hctx, i) {
2901 cpumask_clear(hctx->cpumask);
2903 hctx->dispatch_from = NULL;
2907 * Map software to hardware queues.
2909 * If the cpu isn't present, the cpu is mapped to first hctx.
2911 for_each_possible_cpu(i) {
2913 ctx = per_cpu_ptr(q->queue_ctx, i);
2914 for (j = 0; j < set->nr_maps; j++) {
2915 if (!set->map[j].nr_queues) {
2916 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2917 HCTX_TYPE_DEFAULT, i);
2920 hctx_idx = set->map[j].mq_map[i];
2921 /* unmapped hw queue can be remapped after CPU topo changed */
2922 if (!set->tags[hctx_idx] &&
2923 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2925 * If tags initialization fail for some hctx,
2926 * that hctx won't be brought online. In this
2927 * case, remap the current ctx to hctx[0] which
2928 * is guaranteed to always have tags allocated
2930 set->map[j].mq_map[i] = 0;
2933 hctx = blk_mq_map_queue_type(q, j, i);
2934 ctx->hctxs[j] = hctx;
2936 * If the CPU is already set in the mask, then we've
2937 * mapped this one already. This can happen if
2938 * devices share queues across queue maps.
2940 if (cpumask_test_cpu(i, hctx->cpumask))
2943 cpumask_set_cpu(i, hctx->cpumask);
2945 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2946 hctx->ctxs[hctx->nr_ctx++] = ctx;
2949 * If the nr_ctx type overflows, we have exceeded the
2950 * amount of sw queues we can support.
2952 BUG_ON(!hctx->nr_ctx);
2955 for (; j < HCTX_MAX_TYPES; j++)
2956 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2957 HCTX_TYPE_DEFAULT, i);
2960 queue_for_each_hw_ctx(q, hctx, i) {
2962 * If no software queues are mapped to this hardware queue,
2963 * disable it and free the request entries.
2965 if (!hctx->nr_ctx) {
2966 /* Never unmap queue 0. We need it as a
2967 * fallback in case of a new remap fails
2970 if (i && set->tags[i])
2971 blk_mq_free_map_and_requests(set, i);
2977 hctx->tags = set->tags[i];
2978 WARN_ON(!hctx->tags);
2981 * Set the map size to the number of mapped software queues.
2982 * This is more accurate and more efficient than looping
2983 * over all possibly mapped software queues.
2985 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2988 * Initialize batch roundrobin counts
2990 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2991 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2996 * Caller needs to ensure that we're either frozen/quiesced, or that
2997 * the queue isn't live yet.
2999 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3001 struct blk_mq_hw_ctx *hctx;
3004 queue_for_each_hw_ctx(q, hctx, i) {
3006 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3008 blk_mq_tag_idle(hctx);
3009 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3014 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3017 struct request_queue *q;
3019 lockdep_assert_held(&set->tag_list_lock);
3021 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3022 blk_mq_freeze_queue(q);
3023 queue_set_hctx_shared(q, shared);
3024 blk_mq_unfreeze_queue(q);
3028 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3030 struct blk_mq_tag_set *set = q->tag_set;
3032 mutex_lock(&set->tag_list_lock);
3033 list_del(&q->tag_set_list);
3034 if (list_is_singular(&set->tag_list)) {
3035 /* just transitioned to unshared */
3036 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3037 /* update existing queue */
3038 blk_mq_update_tag_set_shared(set, false);
3040 mutex_unlock(&set->tag_list_lock);
3041 INIT_LIST_HEAD(&q->tag_set_list);
3044 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3045 struct request_queue *q)
3047 mutex_lock(&set->tag_list_lock);
3050 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3052 if (!list_empty(&set->tag_list) &&
3053 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3054 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3055 /* update existing queue */
3056 blk_mq_update_tag_set_shared(set, true);
3058 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3059 queue_set_hctx_shared(q, true);
3060 list_add_tail(&q->tag_set_list, &set->tag_list);
3062 mutex_unlock(&set->tag_list_lock);
3065 /* All allocations will be freed in release handler of q->mq_kobj */
3066 static int blk_mq_alloc_ctxs(struct request_queue *q)
3068 struct blk_mq_ctxs *ctxs;
3071 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3075 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3076 if (!ctxs->queue_ctx)
3079 for_each_possible_cpu(cpu) {
3080 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3084 q->mq_kobj = &ctxs->kobj;
3085 q->queue_ctx = ctxs->queue_ctx;
3094 * It is the actual release handler for mq, but we do it from
3095 * request queue's release handler for avoiding use-after-free
3096 * and headache because q->mq_kobj shouldn't have been introduced,
3097 * but we can't group ctx/kctx kobj without it.
3099 void blk_mq_release(struct request_queue *q)
3101 struct blk_mq_hw_ctx *hctx, *next;
3104 queue_for_each_hw_ctx(q, hctx, i)
3105 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3107 /* all hctx are in .unused_hctx_list now */
3108 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3109 list_del_init(&hctx->hctx_list);
3110 kobject_put(&hctx->kobj);
3113 kfree(q->queue_hw_ctx);
3116 * release .mq_kobj and sw queue's kobject now because
3117 * both share lifetime with request queue.
3119 blk_mq_sysfs_deinit(q);
3122 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3125 struct request_queue *q;
3128 q = blk_alloc_queue(set->numa_node);
3130 return ERR_PTR(-ENOMEM);
3131 q->queuedata = queuedata;
3132 ret = blk_mq_init_allocated_queue(set, q);
3134 blk_cleanup_queue(q);
3135 return ERR_PTR(ret);
3140 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3142 return blk_mq_init_queue_data(set, NULL);
3144 EXPORT_SYMBOL(blk_mq_init_queue);
3146 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
3147 struct lock_class_key *lkclass)
3149 struct request_queue *q;
3150 struct gendisk *disk;
3152 q = blk_mq_init_queue_data(set, queuedata);
3156 disk = __alloc_disk_node(q, set->numa_node, lkclass);
3158 blk_cleanup_queue(q);
3159 return ERR_PTR(-ENOMEM);
3163 EXPORT_SYMBOL(__blk_mq_alloc_disk);
3165 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3166 struct blk_mq_tag_set *set, struct request_queue *q,
3167 int hctx_idx, int node)
3169 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3171 /* reuse dead hctx first */
3172 spin_lock(&q->unused_hctx_lock);
3173 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3174 if (tmp->numa_node == node) {
3180 list_del_init(&hctx->hctx_list);
3181 spin_unlock(&q->unused_hctx_lock);
3184 hctx = blk_mq_alloc_hctx(q, set, node);
3188 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3194 kobject_put(&hctx->kobj);
3199 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3200 struct request_queue *q)
3203 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3205 if (q->nr_hw_queues < set->nr_hw_queues) {
3206 struct blk_mq_hw_ctx **new_hctxs;
3208 new_hctxs = kcalloc_node(set->nr_hw_queues,
3209 sizeof(*new_hctxs), GFP_KERNEL,
3214 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3216 q->queue_hw_ctx = new_hctxs;
3221 /* protect against switching io scheduler */
3222 mutex_lock(&q->sysfs_lock);
3223 for (i = 0; i < set->nr_hw_queues; i++) {
3225 struct blk_mq_hw_ctx *hctx;
3227 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3229 * If the hw queue has been mapped to another numa node,
3230 * we need to realloc the hctx. If allocation fails, fallback
3231 * to use the previous one.
3233 if (hctxs[i] && (hctxs[i]->numa_node == node))
3236 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3239 blk_mq_exit_hctx(q, set, hctxs[i], i);
3243 pr_warn("Allocate new hctx on node %d fails,\
3244 fallback to previous one on node %d\n",
3245 node, hctxs[i]->numa_node);
3251 * Increasing nr_hw_queues fails. Free the newly allocated
3252 * hctxs and keep the previous q->nr_hw_queues.
3254 if (i != set->nr_hw_queues) {
3255 j = q->nr_hw_queues;
3259 end = q->nr_hw_queues;
3260 q->nr_hw_queues = set->nr_hw_queues;
3263 for (; j < end; j++) {
3264 struct blk_mq_hw_ctx *hctx = hctxs[j];
3268 blk_mq_free_map_and_requests(set, j);
3269 blk_mq_exit_hctx(q, set, hctx, j);
3273 mutex_unlock(&q->sysfs_lock);
3276 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3277 struct request_queue *q)
3279 /* mark the queue as mq asap */
3280 q->mq_ops = set->ops;
3282 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3283 blk_mq_poll_stats_bkt,
3284 BLK_MQ_POLL_STATS_BKTS, q);
3288 if (blk_mq_alloc_ctxs(q))
3291 /* init q->mq_kobj and sw queues' kobjects */
3292 blk_mq_sysfs_init(q);
3294 INIT_LIST_HEAD(&q->unused_hctx_list);
3295 spin_lock_init(&q->unused_hctx_lock);
3297 blk_mq_realloc_hw_ctxs(set, q);
3298 if (!q->nr_hw_queues)
3301 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3302 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3306 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3307 if (set->nr_maps > HCTX_TYPE_POLL &&
3308 set->map[HCTX_TYPE_POLL].nr_queues)
3309 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3311 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3312 INIT_LIST_HEAD(&q->requeue_list);
3313 spin_lock_init(&q->requeue_lock);
3315 q->nr_requests = set->queue_depth;
3318 * Default to classic polling
3320 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3322 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3323 blk_mq_add_queue_tag_set(set, q);
3324 blk_mq_map_swqueue(q);
3328 kfree(q->queue_hw_ctx);
3329 q->nr_hw_queues = 0;
3330 blk_mq_sysfs_deinit(q);
3332 blk_stat_free_callback(q->poll_cb);
3338 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3340 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3341 void blk_mq_exit_queue(struct request_queue *q)
3343 struct blk_mq_tag_set *set = q->tag_set;
3345 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
3346 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3347 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
3348 blk_mq_del_queue_tag_set(q);
3351 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3355 for (i = 0; i < set->nr_hw_queues; i++) {
3356 if (!__blk_mq_alloc_map_and_request(set, i))
3365 blk_mq_free_map_and_requests(set, i);
3371 * Allocate the request maps associated with this tag_set. Note that this
3372 * may reduce the depth asked for, if memory is tight. set->queue_depth
3373 * will be updated to reflect the allocated depth.
3375 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3380 depth = set->queue_depth;
3382 err = __blk_mq_alloc_rq_maps(set);
3386 set->queue_depth >>= 1;
3387 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3391 } while (set->queue_depth);
3393 if (!set->queue_depth || err) {
3394 pr_err("blk-mq: failed to allocate request map\n");
3398 if (depth != set->queue_depth)
3399 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3400 depth, set->queue_depth);
3405 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3408 * blk_mq_map_queues() and multiple .map_queues() implementations
3409 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3410 * number of hardware queues.
3412 if (set->nr_maps == 1)
3413 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3415 if (set->ops->map_queues && !is_kdump_kernel()) {
3419 * transport .map_queues is usually done in the following
3422 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3423 * mask = get_cpu_mask(queue)
3424 * for_each_cpu(cpu, mask)
3425 * set->map[x].mq_map[cpu] = queue;
3428 * When we need to remap, the table has to be cleared for
3429 * killing stale mapping since one CPU may not be mapped
3432 for (i = 0; i < set->nr_maps; i++)
3433 blk_mq_clear_mq_map(&set->map[i]);
3435 return set->ops->map_queues(set);
3437 BUG_ON(set->nr_maps > 1);
3438 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3442 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3443 int cur_nr_hw_queues, int new_nr_hw_queues)
3445 struct blk_mq_tags **new_tags;
3447 if (cur_nr_hw_queues >= new_nr_hw_queues)
3450 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3451 GFP_KERNEL, set->numa_node);
3456 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3457 sizeof(*set->tags));
3459 set->tags = new_tags;
3460 set->nr_hw_queues = new_nr_hw_queues;
3465 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
3466 int new_nr_hw_queues)
3468 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
3472 * Alloc a tag set to be associated with one or more request queues.
3473 * May fail with EINVAL for various error conditions. May adjust the
3474 * requested depth down, if it's too large. In that case, the set
3475 * value will be stored in set->queue_depth.
3477 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3481 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3483 if (!set->nr_hw_queues)
3485 if (!set->queue_depth)
3487 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3490 if (!set->ops->queue_rq)
3493 if (!set->ops->get_budget ^ !set->ops->put_budget)
3496 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3497 pr_info("blk-mq: reduced tag depth to %u\n",
3499 set->queue_depth = BLK_MQ_MAX_DEPTH;
3504 else if (set->nr_maps > HCTX_MAX_TYPES)
3508 * If a crashdump is active, then we are potentially in a very
3509 * memory constrained environment. Limit us to 1 queue and
3510 * 64 tags to prevent using too much memory.
3512 if (is_kdump_kernel()) {
3513 set->nr_hw_queues = 1;
3515 set->queue_depth = min(64U, set->queue_depth);
3518 * There is no use for more h/w queues than cpus if we just have
3521 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3522 set->nr_hw_queues = nr_cpu_ids;
3524 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
3528 for (i = 0; i < set->nr_maps; i++) {
3529 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3530 sizeof(set->map[i].mq_map[0]),
3531 GFP_KERNEL, set->numa_node);
3532 if (!set->map[i].mq_map)
3533 goto out_free_mq_map;
3534 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3537 ret = blk_mq_update_queue_map(set);
3539 goto out_free_mq_map;
3541 ret = blk_mq_alloc_map_and_requests(set);
3543 goto out_free_mq_map;
3545 if (blk_mq_is_sbitmap_shared(set->flags)) {
3546 atomic_set(&set->active_queues_shared_sbitmap, 0);
3548 if (blk_mq_init_shared_sbitmap(set)) {
3550 goto out_free_mq_rq_maps;
3554 mutex_init(&set->tag_list_lock);
3555 INIT_LIST_HEAD(&set->tag_list);
3559 out_free_mq_rq_maps:
3560 for (i = 0; i < set->nr_hw_queues; i++)
3561 blk_mq_free_map_and_requests(set, i);
3563 for (i = 0; i < set->nr_maps; i++) {
3564 kfree(set->map[i].mq_map);
3565 set->map[i].mq_map = NULL;
3571 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3573 /* allocate and initialize a tagset for a simple single-queue device */
3574 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
3575 const struct blk_mq_ops *ops, unsigned int queue_depth,
3576 unsigned int set_flags)
3578 memset(set, 0, sizeof(*set));
3580 set->nr_hw_queues = 1;
3582 set->queue_depth = queue_depth;
3583 set->numa_node = NUMA_NO_NODE;
3584 set->flags = set_flags;
3585 return blk_mq_alloc_tag_set(set);
3587 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
3589 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3593 for (i = 0; i < set->nr_hw_queues; i++)
3594 blk_mq_free_map_and_requests(set, i);
3596 if (blk_mq_is_sbitmap_shared(set->flags))
3597 blk_mq_exit_shared_sbitmap(set);
3599 for (j = 0; j < set->nr_maps; j++) {
3600 kfree(set->map[j].mq_map);
3601 set->map[j].mq_map = NULL;
3607 EXPORT_SYMBOL(blk_mq_free_tag_set);
3609 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3611 struct blk_mq_tag_set *set = q->tag_set;
3612 struct blk_mq_hw_ctx *hctx;
3618 if (q->nr_requests == nr)
3621 blk_mq_freeze_queue(q);
3622 blk_mq_quiesce_queue(q);
3625 queue_for_each_hw_ctx(q, hctx, i) {
3629 * If we're using an MQ scheduler, just update the scheduler
3630 * queue depth. This is similar to what the old code would do.
3632 if (!hctx->sched_tags) {
3633 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3635 if (!ret && blk_mq_is_sbitmap_shared(set->flags))
3636 blk_mq_tag_resize_shared_sbitmap(set, nr);
3638 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3640 if (blk_mq_is_sbitmap_shared(set->flags)) {
3641 hctx->sched_tags->bitmap_tags =
3642 &q->sched_bitmap_tags;
3643 hctx->sched_tags->breserved_tags =
3644 &q->sched_breserved_tags;
3649 if (q->elevator && q->elevator->type->ops.depth_updated)
3650 q->elevator->type->ops.depth_updated(hctx);
3653 q->nr_requests = nr;
3654 if (q->elevator && blk_mq_is_sbitmap_shared(set->flags))
3655 sbitmap_queue_resize(&q->sched_bitmap_tags,
3656 nr - set->reserved_tags);
3659 blk_mq_unquiesce_queue(q);
3660 blk_mq_unfreeze_queue(q);
3666 * request_queue and elevator_type pair.
3667 * It is just used by __blk_mq_update_nr_hw_queues to cache
3668 * the elevator_type associated with a request_queue.
3670 struct blk_mq_qe_pair {
3671 struct list_head node;
3672 struct request_queue *q;
3673 struct elevator_type *type;
3677 * Cache the elevator_type in qe pair list and switch the
3678 * io scheduler to 'none'
3680 static bool blk_mq_elv_switch_none(struct list_head *head,
3681 struct request_queue *q)
3683 struct blk_mq_qe_pair *qe;
3688 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3692 INIT_LIST_HEAD(&qe->node);
3694 qe->type = q->elevator->type;
3695 list_add(&qe->node, head);
3697 mutex_lock(&q->sysfs_lock);
3699 * After elevator_switch_mq, the previous elevator_queue will be
3700 * released by elevator_release. The reference of the io scheduler
3701 * module get by elevator_get will also be put. So we need to get
3702 * a reference of the io scheduler module here to prevent it to be
3705 __module_get(qe->type->elevator_owner);
3706 elevator_switch_mq(q, NULL);
3707 mutex_unlock(&q->sysfs_lock);
3712 static void blk_mq_elv_switch_back(struct list_head *head,
3713 struct request_queue *q)
3715 struct blk_mq_qe_pair *qe;
3716 struct elevator_type *t = NULL;
3718 list_for_each_entry(qe, head, node)
3727 list_del(&qe->node);
3730 mutex_lock(&q->sysfs_lock);
3731 elevator_switch_mq(q, t);
3732 mutex_unlock(&q->sysfs_lock);
3735 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3738 struct request_queue *q;
3740 int prev_nr_hw_queues;
3742 lockdep_assert_held(&set->tag_list_lock);
3744 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3745 nr_hw_queues = nr_cpu_ids;
3746 if (nr_hw_queues < 1)
3748 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3751 list_for_each_entry(q, &set->tag_list, tag_set_list)
3752 blk_mq_freeze_queue(q);
3754 * Switch IO scheduler to 'none', cleaning up the data associated
3755 * with the previous scheduler. We will switch back once we are done
3756 * updating the new sw to hw queue mappings.
3758 list_for_each_entry(q, &set->tag_list, tag_set_list)
3759 if (!blk_mq_elv_switch_none(&head, q))
3762 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3763 blk_mq_debugfs_unregister_hctxs(q);
3764 blk_mq_sysfs_unregister(q);
3767 prev_nr_hw_queues = set->nr_hw_queues;
3768 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3772 set->nr_hw_queues = nr_hw_queues;
3774 blk_mq_update_queue_map(set);
3775 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3776 blk_mq_realloc_hw_ctxs(set, q);
3777 if (q->nr_hw_queues != set->nr_hw_queues) {
3778 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3779 nr_hw_queues, prev_nr_hw_queues);
3780 set->nr_hw_queues = prev_nr_hw_queues;
3781 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3784 blk_mq_map_swqueue(q);
3788 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3789 blk_mq_sysfs_register(q);
3790 blk_mq_debugfs_register_hctxs(q);
3794 list_for_each_entry(q, &set->tag_list, tag_set_list)
3795 blk_mq_elv_switch_back(&head, q);
3797 list_for_each_entry(q, &set->tag_list, tag_set_list)
3798 blk_mq_unfreeze_queue(q);
3801 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3803 mutex_lock(&set->tag_list_lock);
3804 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3805 mutex_unlock(&set->tag_list_lock);
3807 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3809 /* Enable polling stats and return whether they were already enabled. */
3810 static bool blk_poll_stats_enable(struct request_queue *q)
3812 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3813 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3815 blk_stat_add_callback(q, q->poll_cb);
3819 static void blk_mq_poll_stats_start(struct request_queue *q)
3822 * We don't arm the callback if polling stats are not enabled or the
3823 * callback is already active.
3825 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3826 blk_stat_is_active(q->poll_cb))
3829 blk_stat_activate_msecs(q->poll_cb, 100);
3832 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3834 struct request_queue *q = cb->data;
3837 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3838 if (cb->stat[bucket].nr_samples)
3839 q->poll_stat[bucket] = cb->stat[bucket];
3843 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3846 unsigned long ret = 0;
3850 * If stats collection isn't on, don't sleep but turn it on for
3853 if (!blk_poll_stats_enable(q))
3857 * As an optimistic guess, use half of the mean service time
3858 * for this type of request. We can (and should) make this smarter.
3859 * For instance, if the completion latencies are tight, we can
3860 * get closer than just half the mean. This is especially
3861 * important on devices where the completion latencies are longer
3862 * than ~10 usec. We do use the stats for the relevant IO size
3863 * if available which does lead to better estimates.
3865 bucket = blk_mq_poll_stats_bkt(rq);
3869 if (q->poll_stat[bucket].nr_samples)
3870 ret = (q->poll_stat[bucket].mean + 1) / 2;
3875 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3878 struct hrtimer_sleeper hs;
3879 enum hrtimer_mode mode;
3883 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3887 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3889 * 0: use half of prev avg
3890 * >0: use this specific value
3892 if (q->poll_nsec > 0)
3893 nsecs = q->poll_nsec;
3895 nsecs = blk_mq_poll_nsecs(q, rq);
3900 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3903 * This will be replaced with the stats tracking code, using
3904 * 'avg_completion_time / 2' as the pre-sleep target.
3908 mode = HRTIMER_MODE_REL;
3909 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3910 hrtimer_set_expires(&hs.timer, kt);
3913 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3915 set_current_state(TASK_UNINTERRUPTIBLE);
3916 hrtimer_sleeper_start_expires(&hs, mode);
3919 hrtimer_cancel(&hs.timer);
3920 mode = HRTIMER_MODE_ABS;
3921 } while (hs.task && !signal_pending(current));
3923 __set_current_state(TASK_RUNNING);
3924 destroy_hrtimer_on_stack(&hs.timer);
3928 static bool blk_mq_poll_hybrid(struct request_queue *q,
3929 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3933 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3936 if (!blk_qc_t_is_internal(cookie))
3937 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3939 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3941 * With scheduling, if the request has completed, we'll
3942 * get a NULL return here, as we clear the sched tag when
3943 * that happens. The request still remains valid, like always,
3944 * so we should be safe with just the NULL check.
3950 return blk_mq_poll_hybrid_sleep(q, rq);
3954 * blk_poll - poll for IO completions
3956 * @cookie: cookie passed back at IO submission time
3957 * @spin: whether to spin for completions
3960 * Poll for completions on the passed in queue. Returns number of
3961 * completed entries found. If @spin is true, then blk_poll will continue
3962 * looping until at least one completion is found, unless the task is
3963 * otherwise marked running (or we need to reschedule).
3965 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3967 struct blk_mq_hw_ctx *hctx;
3970 if (!blk_qc_t_valid(cookie) ||
3971 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3975 blk_flush_plug_list(current->plug, false);
3977 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3980 * If we sleep, have the caller restart the poll loop to reset
3981 * the state. Like for the other success return cases, the
3982 * caller is responsible for checking if the IO completed. If
3983 * the IO isn't complete, we'll get called again and will go
3984 * straight to the busy poll loop. If specified not to spin,
3985 * we also should not sleep.
3987 if (spin && blk_mq_poll_hybrid(q, hctx, cookie))
3990 hctx->poll_considered++;
3992 state = get_current_state();
3996 hctx->poll_invoked++;
3998 ret = q->mq_ops->poll(hctx);
4000 hctx->poll_success++;
4001 __set_current_state(TASK_RUNNING);
4005 if (signal_pending_state(state, current))
4006 __set_current_state(TASK_RUNNING);
4008 if (task_is_running(current))
4010 if (ret < 0 || !spin)
4013 } while (!need_resched());
4015 __set_current_state(TASK_RUNNING);
4018 EXPORT_SYMBOL_GPL(blk_poll);
4020 unsigned int blk_mq_rq_cpu(struct request *rq)
4022 return rq->mq_ctx->cpu;
4024 EXPORT_SYMBOL(blk_mq_rq_cpu);
4026 void blk_mq_cancel_work_sync(struct request_queue *q)
4028 if (queue_is_mq(q)) {
4029 struct blk_mq_hw_ctx *hctx;
4032 cancel_delayed_work_sync(&q->requeue_work);
4034 queue_for_each_hw_ctx(q, hctx, i)
4035 cancel_delayed_work_sync(&hctx->run_work);
4039 static int __init blk_mq_init(void)
4043 for_each_possible_cpu(i)
4044 init_llist_head(&per_cpu(blk_cpu_done, i));
4045 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4047 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4048 "block/softirq:dead", NULL,
4049 blk_softirq_cpu_dead);
4050 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4051 blk_mq_hctx_notify_dead);
4052 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4053 blk_mq_hctx_notify_online,
4054 blk_mq_hctx_notify_offline);
4057 subsys_initcall(blk_mq_init);