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)
193 mutex_lock(&q->mq_freeze_lock);
194 q->mq_freeze_depth--;
195 WARN_ON_ONCE(q->mq_freeze_depth < 0);
196 if (!q->mq_freeze_depth) {
197 percpu_ref_resurrect(&q->q_usage_counter);
198 wake_up_all(&q->mq_freeze_wq);
200 mutex_unlock(&q->mq_freeze_lock);
202 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
205 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
206 * mpt3sas driver such that this function can be removed.
208 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
210 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
212 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
215 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
218 * Note: this function does not prevent that the struct request end_io()
219 * callback function is invoked. Once this function is returned, we make
220 * sure no dispatch can happen until the queue is unquiesced via
221 * blk_mq_unquiesce_queue().
223 void blk_mq_quiesce_queue(struct request_queue *q)
225 struct blk_mq_hw_ctx *hctx;
229 blk_mq_quiesce_queue_nowait(q);
231 queue_for_each_hw_ctx(q, hctx, i) {
232 if (hctx->flags & BLK_MQ_F_BLOCKING)
233 synchronize_srcu(hctx->srcu);
240 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
243 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
246 * This function recovers queue into the state before quiescing
247 * which is done by blk_mq_quiesce_queue.
249 void blk_mq_unquiesce_queue(struct request_queue *q)
251 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
253 /* dispatch requests which are inserted during quiescing */
254 blk_mq_run_hw_queues(q, true);
256 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
258 void blk_mq_wake_waiters(struct request_queue *q)
260 struct blk_mq_hw_ctx *hctx;
263 queue_for_each_hw_ctx(q, hctx, i)
264 if (blk_mq_hw_queue_mapped(hctx))
265 blk_mq_tag_wakeup_all(hctx->tags, true);
269 * Only need start/end time stamping if we have iostat or
270 * blk stats enabled, or using an IO scheduler.
272 static inline bool blk_mq_need_time_stamp(struct request *rq)
274 return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
277 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
278 unsigned int tag, u64 alloc_time_ns)
280 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
281 struct request *rq = tags->static_rqs[tag];
283 if (data->q->elevator) {
284 rq->tag = BLK_MQ_NO_TAG;
285 rq->internal_tag = tag;
288 rq->internal_tag = BLK_MQ_NO_TAG;
291 /* csd/requeue_work/fifo_time is initialized before use */
293 rq->mq_ctx = data->ctx;
294 rq->mq_hctx = data->hctx;
296 rq->cmd_flags = data->cmd_flags;
297 if (data->flags & BLK_MQ_REQ_PM)
298 rq->rq_flags |= RQF_PM;
299 if (blk_queue_io_stat(data->q))
300 rq->rq_flags |= RQF_IO_STAT;
301 INIT_LIST_HEAD(&rq->queuelist);
302 INIT_HLIST_NODE(&rq->hash);
303 RB_CLEAR_NODE(&rq->rb_node);
306 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
307 rq->alloc_time_ns = alloc_time_ns;
309 if (blk_mq_need_time_stamp(rq))
310 rq->start_time_ns = ktime_get_ns();
312 rq->start_time_ns = 0;
313 rq->io_start_time_ns = 0;
314 rq->stats_sectors = 0;
315 rq->nr_phys_segments = 0;
316 #if defined(CONFIG_BLK_DEV_INTEGRITY)
317 rq->nr_integrity_segments = 0;
319 blk_crypto_rq_set_defaults(rq);
320 /* tag was already set */
321 WRITE_ONCE(rq->deadline, 0);
326 rq->end_io_data = NULL;
328 data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++;
329 refcount_set(&rq->ref, 1);
331 if (!op_is_flush(data->cmd_flags)) {
332 struct elevator_queue *e = data->q->elevator;
335 if (e && e->type->ops.prepare_request) {
336 if (e->type->icq_cache)
337 blk_mq_sched_assign_ioc(rq);
339 e->type->ops.prepare_request(rq);
340 rq->rq_flags |= RQF_ELVPRIV;
344 data->hctx->queued++;
348 static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data)
350 struct request_queue *q = data->q;
351 struct elevator_queue *e = q->elevator;
352 u64 alloc_time_ns = 0;
355 /* alloc_time includes depth and tag waits */
356 if (blk_queue_rq_alloc_time(q))
357 alloc_time_ns = ktime_get_ns();
359 if (data->cmd_flags & REQ_NOWAIT)
360 data->flags |= BLK_MQ_REQ_NOWAIT;
364 * Flush requests are special and go directly to the
365 * dispatch list. Don't include reserved tags in the
366 * limiting, as it isn't useful.
368 if (!op_is_flush(data->cmd_flags) &&
369 e->type->ops.limit_depth &&
370 !(data->flags & BLK_MQ_REQ_RESERVED))
371 e->type->ops.limit_depth(data->cmd_flags, data);
375 data->ctx = blk_mq_get_ctx(q);
376 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
378 blk_mq_tag_busy(data->hctx);
381 * Waiting allocations only fail because of an inactive hctx. In that
382 * case just retry the hctx assignment and tag allocation as CPU hotplug
383 * should have migrated us to an online CPU by now.
385 tag = blk_mq_get_tag(data);
386 if (tag == BLK_MQ_NO_TAG) {
387 if (data->flags & BLK_MQ_REQ_NOWAIT)
391 * Give up the CPU and sleep for a random short time to ensure
392 * that thread using a realtime scheduling class are migrated
393 * off the CPU, and thus off the hctx that is going away.
398 return blk_mq_rq_ctx_init(data, tag, alloc_time_ns);
401 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
402 blk_mq_req_flags_t flags)
404 struct blk_mq_alloc_data data = {
412 ret = blk_queue_enter(q, flags);
416 rq = __blk_mq_alloc_request(&data);
420 rq->__sector = (sector_t) -1;
421 rq->bio = rq->biotail = NULL;
425 return ERR_PTR(-EWOULDBLOCK);
427 EXPORT_SYMBOL(blk_mq_alloc_request);
429 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
430 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
432 struct blk_mq_alloc_data data = {
437 u64 alloc_time_ns = 0;
442 /* alloc_time includes depth and tag waits */
443 if (blk_queue_rq_alloc_time(q))
444 alloc_time_ns = ktime_get_ns();
447 * If the tag allocator sleeps we could get an allocation for a
448 * different hardware context. No need to complicate the low level
449 * allocator for this for the rare use case of a command tied to
452 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
453 return ERR_PTR(-EINVAL);
455 if (hctx_idx >= q->nr_hw_queues)
456 return ERR_PTR(-EIO);
458 ret = blk_queue_enter(q, flags);
463 * Check if the hardware context is actually mapped to anything.
464 * If not tell the caller that it should skip this queue.
467 data.hctx = q->queue_hw_ctx[hctx_idx];
468 if (!blk_mq_hw_queue_mapped(data.hctx))
470 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
471 data.ctx = __blk_mq_get_ctx(q, cpu);
474 blk_mq_tag_busy(data.hctx);
477 tag = blk_mq_get_tag(&data);
478 if (tag == BLK_MQ_NO_TAG)
480 return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns);
486 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
488 static void __blk_mq_free_request(struct request *rq)
490 struct request_queue *q = rq->q;
491 struct blk_mq_ctx *ctx = rq->mq_ctx;
492 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
493 const int sched_tag = rq->internal_tag;
495 blk_crypto_free_request(rq);
496 blk_pm_mark_last_busy(rq);
498 if (rq->tag != BLK_MQ_NO_TAG)
499 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
500 if (sched_tag != BLK_MQ_NO_TAG)
501 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
502 blk_mq_sched_restart(hctx);
506 void blk_mq_free_request(struct request *rq)
508 struct request_queue *q = rq->q;
509 struct elevator_queue *e = q->elevator;
510 struct blk_mq_ctx *ctx = rq->mq_ctx;
511 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
513 if (rq->rq_flags & RQF_ELVPRIV) {
514 if (e && e->type->ops.finish_request)
515 e->type->ops.finish_request(rq);
517 put_io_context(rq->elv.icq->ioc);
522 ctx->rq_completed[rq_is_sync(rq)]++;
523 if (rq->rq_flags & RQF_MQ_INFLIGHT)
524 __blk_mq_dec_active_requests(hctx);
526 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
527 laptop_io_completion(q->backing_dev_info);
531 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
532 if (refcount_dec_and_test(&rq->ref))
533 __blk_mq_free_request(rq);
535 EXPORT_SYMBOL_GPL(blk_mq_free_request);
537 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
541 if (blk_mq_need_time_stamp(rq))
542 now = ktime_get_ns();
544 if (rq->rq_flags & RQF_STATS) {
545 blk_mq_poll_stats_start(rq->q);
546 blk_stat_add(rq, now);
549 blk_mq_sched_completed_request(rq, now);
551 blk_account_io_done(rq, now);
554 rq_qos_done(rq->q, rq);
555 rq->end_io(rq, error);
557 blk_mq_free_request(rq);
560 EXPORT_SYMBOL(__blk_mq_end_request);
562 void blk_mq_end_request(struct request *rq, blk_status_t error)
564 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
566 __blk_mq_end_request(rq, error);
568 EXPORT_SYMBOL(blk_mq_end_request);
570 static void blk_complete_reqs(struct llist_head *list)
572 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
573 struct request *rq, *next;
575 llist_for_each_entry_safe(rq, next, entry, ipi_list)
576 rq->q->mq_ops->complete(rq);
579 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
581 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
584 static int blk_softirq_cpu_dead(unsigned int cpu)
586 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
590 static void __blk_mq_complete_request_remote(void *data)
592 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
595 static inline bool blk_mq_complete_need_ipi(struct request *rq)
597 int cpu = raw_smp_processor_id();
599 if (!IS_ENABLED(CONFIG_SMP) ||
600 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
603 * With force threaded interrupts enabled, raising softirq from an SMP
604 * function call will always result in waking the ksoftirqd thread.
605 * This is probably worse than completing the request on a different
608 if (force_irqthreads)
611 /* same CPU or cache domain? Complete locally */
612 if (cpu == rq->mq_ctx->cpu ||
613 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
614 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
617 /* don't try to IPI to an offline CPU */
618 return cpu_online(rq->mq_ctx->cpu);
621 static void blk_mq_complete_send_ipi(struct request *rq)
623 struct llist_head *list;
626 cpu = rq->mq_ctx->cpu;
627 list = &per_cpu(blk_cpu_done, cpu);
628 if (llist_add(&rq->ipi_list, list)) {
629 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
630 smp_call_function_single_async(cpu, &rq->csd);
634 static void blk_mq_raise_softirq(struct request *rq)
636 struct llist_head *list;
639 list = this_cpu_ptr(&blk_cpu_done);
640 if (llist_add(&rq->ipi_list, list))
641 raise_softirq(BLOCK_SOFTIRQ);
645 bool blk_mq_complete_request_remote(struct request *rq)
647 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
650 * For a polled request, always complete locallly, it's pointless
651 * to redirect the completion.
653 if (rq->cmd_flags & REQ_HIPRI)
656 if (blk_mq_complete_need_ipi(rq)) {
657 blk_mq_complete_send_ipi(rq);
661 if (rq->q->nr_hw_queues == 1) {
662 blk_mq_raise_softirq(rq);
667 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
670 * blk_mq_complete_request - end I/O on a request
671 * @rq: the request being processed
674 * Complete a request by scheduling the ->complete_rq operation.
676 void blk_mq_complete_request(struct request *rq)
678 if (!blk_mq_complete_request_remote(rq))
679 rq->q->mq_ops->complete(rq);
681 EXPORT_SYMBOL(blk_mq_complete_request);
683 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
684 __releases(hctx->srcu)
686 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
689 srcu_read_unlock(hctx->srcu, srcu_idx);
692 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
693 __acquires(hctx->srcu)
695 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
696 /* shut up gcc false positive */
700 *srcu_idx = srcu_read_lock(hctx->srcu);
704 * blk_mq_start_request - Start processing a request
705 * @rq: Pointer to request to be started
707 * Function used by device drivers to notify the block layer that a request
708 * is going to be processed now, so blk layer can do proper initializations
709 * such as starting the timeout timer.
711 void blk_mq_start_request(struct request *rq)
713 struct request_queue *q = rq->q;
715 trace_block_rq_issue(rq);
717 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
718 rq->io_start_time_ns = ktime_get_ns();
719 rq->stats_sectors = blk_rq_sectors(rq);
720 rq->rq_flags |= RQF_STATS;
724 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
727 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
729 #ifdef CONFIG_BLK_DEV_INTEGRITY
730 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
731 q->integrity.profile->prepare_fn(rq);
734 EXPORT_SYMBOL(blk_mq_start_request);
736 static void __blk_mq_requeue_request(struct request *rq)
738 struct request_queue *q = rq->q;
740 blk_mq_put_driver_tag(rq);
742 trace_block_rq_requeue(rq);
743 rq_qos_requeue(q, rq);
745 if (blk_mq_request_started(rq)) {
746 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
747 rq->rq_flags &= ~RQF_TIMED_OUT;
751 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
753 __blk_mq_requeue_request(rq);
755 /* this request will be re-inserted to io scheduler queue */
756 blk_mq_sched_requeue_request(rq);
758 BUG_ON(!list_empty(&rq->queuelist));
759 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
761 EXPORT_SYMBOL(blk_mq_requeue_request);
763 static void blk_mq_requeue_work(struct work_struct *work)
765 struct request_queue *q =
766 container_of(work, struct request_queue, requeue_work.work);
768 struct request *rq, *next;
770 spin_lock_irq(&q->requeue_lock);
771 list_splice_init(&q->requeue_list, &rq_list);
772 spin_unlock_irq(&q->requeue_lock);
774 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
775 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
778 rq->rq_flags &= ~RQF_SOFTBARRIER;
779 list_del_init(&rq->queuelist);
781 * If RQF_DONTPREP, rq has contained some driver specific
782 * data, so insert it to hctx dispatch list to avoid any
785 if (rq->rq_flags & RQF_DONTPREP)
786 blk_mq_request_bypass_insert(rq, false, false);
788 blk_mq_sched_insert_request(rq, true, false, false);
791 while (!list_empty(&rq_list)) {
792 rq = list_entry(rq_list.next, struct request, queuelist);
793 list_del_init(&rq->queuelist);
794 blk_mq_sched_insert_request(rq, false, false, false);
797 blk_mq_run_hw_queues(q, false);
800 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
801 bool kick_requeue_list)
803 struct request_queue *q = rq->q;
807 * We abuse this flag that is otherwise used by the I/O scheduler to
808 * request head insertion from the workqueue.
810 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
812 spin_lock_irqsave(&q->requeue_lock, flags);
814 rq->rq_flags |= RQF_SOFTBARRIER;
815 list_add(&rq->queuelist, &q->requeue_list);
817 list_add_tail(&rq->queuelist, &q->requeue_list);
819 spin_unlock_irqrestore(&q->requeue_lock, flags);
821 if (kick_requeue_list)
822 blk_mq_kick_requeue_list(q);
825 void blk_mq_kick_requeue_list(struct request_queue *q)
827 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
829 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
831 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
834 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
835 msecs_to_jiffies(msecs));
837 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
839 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
841 if (tag < tags->nr_tags) {
842 prefetch(tags->rqs[tag]);
843 return tags->rqs[tag];
848 EXPORT_SYMBOL(blk_mq_tag_to_rq);
850 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
851 void *priv, bool reserved)
854 * If we find a request that isn't idle and the queue matches,
855 * we know the queue is busy. Return false to stop the iteration.
857 if (blk_mq_request_started(rq) && rq->q == hctx->queue) {
867 bool blk_mq_queue_inflight(struct request_queue *q)
871 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
874 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
876 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
878 req->rq_flags |= RQF_TIMED_OUT;
879 if (req->q->mq_ops->timeout) {
880 enum blk_eh_timer_return ret;
882 ret = req->q->mq_ops->timeout(req, reserved);
883 if (ret == BLK_EH_DONE)
885 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
891 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
893 unsigned long deadline;
895 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
897 if (rq->rq_flags & RQF_TIMED_OUT)
900 deadline = READ_ONCE(rq->deadline);
901 if (time_after_eq(jiffies, deadline))
906 else if (time_after(*next, deadline))
911 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
912 struct request *rq, void *priv, bool reserved)
914 unsigned long *next = priv;
917 * Just do a quick check if it is expired before locking the request in
918 * so we're not unnecessarilly synchronizing across CPUs.
920 if (!blk_mq_req_expired(rq, next))
924 * We have reason to believe the request may be expired. Take a
925 * reference on the request to lock this request lifetime into its
926 * currently allocated context to prevent it from being reallocated in
927 * the event the completion by-passes this timeout handler.
929 * If the reference was already released, then the driver beat the
930 * timeout handler to posting a natural completion.
932 if (!refcount_inc_not_zero(&rq->ref))
936 * The request is now locked and cannot be reallocated underneath the
937 * timeout handler's processing. Re-verify this exact request is truly
938 * expired; if it is not expired, then the request was completed and
939 * reallocated as a new request.
941 if (blk_mq_req_expired(rq, next))
942 blk_mq_rq_timed_out(rq, reserved);
944 if (is_flush_rq(rq, hctx))
946 else if (refcount_dec_and_test(&rq->ref))
947 __blk_mq_free_request(rq);
952 static void blk_mq_timeout_work(struct work_struct *work)
954 struct request_queue *q =
955 container_of(work, struct request_queue, timeout_work);
956 unsigned long next = 0;
957 struct blk_mq_hw_ctx *hctx;
960 /* A deadlock might occur if a request is stuck requiring a
961 * timeout at the same time a queue freeze is waiting
962 * completion, since the timeout code would not be able to
963 * acquire the queue reference here.
965 * That's why we don't use blk_queue_enter here; instead, we use
966 * percpu_ref_tryget directly, because we need to be able to
967 * obtain a reference even in the short window between the queue
968 * starting to freeze, by dropping the first reference in
969 * blk_freeze_queue_start, and the moment the last request is
970 * consumed, marked by the instant q_usage_counter reaches
973 if (!percpu_ref_tryget(&q->q_usage_counter))
976 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
979 mod_timer(&q->timeout, next);
982 * Request timeouts are handled as a forward rolling timer. If
983 * we end up here it means that no requests are pending and
984 * also that no request has been pending for a while. Mark
987 queue_for_each_hw_ctx(q, hctx, i) {
988 /* the hctx may be unmapped, so check it here */
989 if (blk_mq_hw_queue_mapped(hctx))
990 blk_mq_tag_idle(hctx);
996 struct flush_busy_ctx_data {
997 struct blk_mq_hw_ctx *hctx;
998 struct list_head *list;
1001 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1003 struct flush_busy_ctx_data *flush_data = data;
1004 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1005 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1006 enum hctx_type type = hctx->type;
1008 spin_lock(&ctx->lock);
1009 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1010 sbitmap_clear_bit(sb, bitnr);
1011 spin_unlock(&ctx->lock);
1016 * Process software queues that have been marked busy, splicing them
1017 * to the for-dispatch
1019 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1021 struct flush_busy_ctx_data data = {
1026 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1028 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1030 struct dispatch_rq_data {
1031 struct blk_mq_hw_ctx *hctx;
1035 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1038 struct dispatch_rq_data *dispatch_data = data;
1039 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1040 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1041 enum hctx_type type = hctx->type;
1043 spin_lock(&ctx->lock);
1044 if (!list_empty(&ctx->rq_lists[type])) {
1045 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1046 list_del_init(&dispatch_data->rq->queuelist);
1047 if (list_empty(&ctx->rq_lists[type]))
1048 sbitmap_clear_bit(sb, bitnr);
1050 spin_unlock(&ctx->lock);
1052 return !dispatch_data->rq;
1055 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1056 struct blk_mq_ctx *start)
1058 unsigned off = start ? start->index_hw[hctx->type] : 0;
1059 struct dispatch_rq_data data = {
1064 __sbitmap_for_each_set(&hctx->ctx_map, off,
1065 dispatch_rq_from_ctx, &data);
1070 static inline unsigned int queued_to_index(unsigned int queued)
1075 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1078 static bool __blk_mq_get_driver_tag(struct request *rq)
1080 struct sbitmap_queue *bt = rq->mq_hctx->tags->bitmap_tags;
1081 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1084 blk_mq_tag_busy(rq->mq_hctx);
1086 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1087 bt = rq->mq_hctx->tags->breserved_tags;
1090 if (!hctx_may_queue(rq->mq_hctx, bt))
1094 tag = __sbitmap_queue_get(bt);
1095 if (tag == BLK_MQ_NO_TAG)
1098 rq->tag = tag + tag_offset;
1102 static bool blk_mq_get_driver_tag(struct request *rq)
1104 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1106 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq))
1109 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1110 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1111 rq->rq_flags |= RQF_MQ_INFLIGHT;
1112 __blk_mq_inc_active_requests(hctx);
1114 hctx->tags->rqs[rq->tag] = rq;
1118 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1119 int flags, void *key)
1121 struct blk_mq_hw_ctx *hctx;
1123 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1125 spin_lock(&hctx->dispatch_wait_lock);
1126 if (!list_empty(&wait->entry)) {
1127 struct sbitmap_queue *sbq;
1129 list_del_init(&wait->entry);
1130 sbq = hctx->tags->bitmap_tags;
1131 atomic_dec(&sbq->ws_active);
1133 spin_unlock(&hctx->dispatch_wait_lock);
1135 blk_mq_run_hw_queue(hctx, true);
1140 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1141 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1142 * restart. For both cases, take care to check the condition again after
1143 * marking us as waiting.
1145 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1148 struct sbitmap_queue *sbq = hctx->tags->bitmap_tags;
1149 struct wait_queue_head *wq;
1150 wait_queue_entry_t *wait;
1153 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1154 blk_mq_sched_mark_restart_hctx(hctx);
1157 * It's possible that a tag was freed in the window between the
1158 * allocation failure and adding the hardware queue to the wait
1161 * Don't clear RESTART here, someone else could have set it.
1162 * At most this will cost an extra queue run.
1164 return blk_mq_get_driver_tag(rq);
1167 wait = &hctx->dispatch_wait;
1168 if (!list_empty_careful(&wait->entry))
1171 wq = &bt_wait_ptr(sbq, hctx)->wait;
1173 spin_lock_irq(&wq->lock);
1174 spin_lock(&hctx->dispatch_wait_lock);
1175 if (!list_empty(&wait->entry)) {
1176 spin_unlock(&hctx->dispatch_wait_lock);
1177 spin_unlock_irq(&wq->lock);
1181 atomic_inc(&sbq->ws_active);
1182 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1183 __add_wait_queue(wq, wait);
1186 * It's possible that a tag was freed in the window between the
1187 * allocation failure and adding the hardware queue to the wait
1190 ret = blk_mq_get_driver_tag(rq);
1192 spin_unlock(&hctx->dispatch_wait_lock);
1193 spin_unlock_irq(&wq->lock);
1198 * We got a tag, remove ourselves from the wait queue to ensure
1199 * someone else gets the wakeup.
1201 list_del_init(&wait->entry);
1202 atomic_dec(&sbq->ws_active);
1203 spin_unlock(&hctx->dispatch_wait_lock);
1204 spin_unlock_irq(&wq->lock);
1209 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1210 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1212 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1213 * - EWMA is one simple way to compute running average value
1214 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1215 * - take 4 as factor for avoiding to get too small(0) result, and this
1216 * factor doesn't matter because EWMA decreases exponentially
1218 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1222 if (hctx->queue->elevator)
1225 ewma = hctx->dispatch_busy;
1230 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1232 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1233 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1235 hctx->dispatch_busy = ewma;
1238 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1240 static void blk_mq_handle_dev_resource(struct request *rq,
1241 struct list_head *list)
1243 struct request *next =
1244 list_first_entry_or_null(list, struct request, queuelist);
1247 * If an I/O scheduler has been configured and we got a driver tag for
1248 * the next request already, free it.
1251 blk_mq_put_driver_tag(next);
1253 list_add(&rq->queuelist, list);
1254 __blk_mq_requeue_request(rq);
1257 static void blk_mq_handle_zone_resource(struct request *rq,
1258 struct list_head *zone_list)
1261 * If we end up here it is because we cannot dispatch a request to a
1262 * specific zone due to LLD level zone-write locking or other zone
1263 * related resource not being available. In this case, set the request
1264 * aside in zone_list for retrying it later.
1266 list_add(&rq->queuelist, zone_list);
1267 __blk_mq_requeue_request(rq);
1270 enum prep_dispatch {
1272 PREP_DISPATCH_NO_TAG,
1273 PREP_DISPATCH_NO_BUDGET,
1276 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1279 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1280 int budget_token = -1;
1283 budget_token = blk_mq_get_dispatch_budget(rq->q);
1284 if (budget_token < 0) {
1285 blk_mq_put_driver_tag(rq);
1286 return PREP_DISPATCH_NO_BUDGET;
1288 blk_mq_set_rq_budget_token(rq, budget_token);
1291 if (!blk_mq_get_driver_tag(rq)) {
1293 * The initial allocation attempt failed, so we need to
1294 * rerun the hardware queue when a tag is freed. The
1295 * waitqueue takes care of that. If the queue is run
1296 * before we add this entry back on the dispatch list,
1297 * we'll re-run it below.
1299 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1301 * All budgets not got from this function will be put
1302 * together during handling partial dispatch
1305 blk_mq_put_dispatch_budget(rq->q, budget_token);
1306 return PREP_DISPATCH_NO_TAG;
1310 return PREP_DISPATCH_OK;
1313 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1314 static void blk_mq_release_budgets(struct request_queue *q,
1315 struct list_head *list)
1319 list_for_each_entry(rq, list, queuelist) {
1320 int budget_token = blk_mq_get_rq_budget_token(rq);
1322 if (budget_token >= 0)
1323 blk_mq_put_dispatch_budget(q, budget_token);
1328 * Returns true if we did some work AND can potentially do more.
1330 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1331 unsigned int nr_budgets)
1333 enum prep_dispatch prep;
1334 struct request_queue *q = hctx->queue;
1335 struct request *rq, *nxt;
1337 blk_status_t ret = BLK_STS_OK;
1338 LIST_HEAD(zone_list);
1340 if (list_empty(list))
1344 * Now process all the entries, sending them to the driver.
1346 errors = queued = 0;
1348 struct blk_mq_queue_data bd;
1350 rq = list_first_entry(list, struct request, queuelist);
1352 WARN_ON_ONCE(hctx != rq->mq_hctx);
1353 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1354 if (prep != PREP_DISPATCH_OK)
1357 list_del_init(&rq->queuelist);
1362 * Flag last if we have no more requests, or if we have more
1363 * but can't assign a driver tag to it.
1365 if (list_empty(list))
1368 nxt = list_first_entry(list, struct request, queuelist);
1369 bd.last = !blk_mq_get_driver_tag(nxt);
1373 * once the request is queued to lld, no need to cover the
1378 ret = q->mq_ops->queue_rq(hctx, &bd);
1383 case BLK_STS_RESOURCE:
1384 case BLK_STS_DEV_RESOURCE:
1385 blk_mq_handle_dev_resource(rq, list);
1387 case BLK_STS_ZONE_RESOURCE:
1389 * Move the request to zone_list and keep going through
1390 * the dispatch list to find more requests the drive can
1393 blk_mq_handle_zone_resource(rq, &zone_list);
1397 blk_mq_end_request(rq, ret);
1399 } while (!list_empty(list));
1401 if (!list_empty(&zone_list))
1402 list_splice_tail_init(&zone_list, list);
1404 hctx->dispatched[queued_to_index(queued)]++;
1406 /* If we didn't flush the entire list, we could have told the driver
1407 * there was more coming, but that turned out to be a lie.
1409 if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued)
1410 q->mq_ops->commit_rqs(hctx);
1412 * Any items that need requeuing? Stuff them into hctx->dispatch,
1413 * that is where we will continue on next queue run.
1415 if (!list_empty(list)) {
1417 /* For non-shared tags, the RESTART check will suffice */
1418 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1419 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1420 bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET;
1423 blk_mq_release_budgets(q, list);
1425 spin_lock(&hctx->lock);
1426 list_splice_tail_init(list, &hctx->dispatch);
1427 spin_unlock(&hctx->lock);
1430 * Order adding requests to hctx->dispatch and checking
1431 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1432 * in blk_mq_sched_restart(). Avoid restart code path to
1433 * miss the new added requests to hctx->dispatch, meantime
1434 * SCHED_RESTART is observed here.
1439 * If SCHED_RESTART was set by the caller of this function and
1440 * it is no longer set that means that it was cleared by another
1441 * thread and hence that a queue rerun is needed.
1443 * If 'no_tag' is set, that means that we failed getting
1444 * a driver tag with an I/O scheduler attached. If our dispatch
1445 * waitqueue is no longer active, ensure that we run the queue
1446 * AFTER adding our entries back to the list.
1448 * If no I/O scheduler has been configured it is possible that
1449 * the hardware queue got stopped and restarted before requests
1450 * were pushed back onto the dispatch list. Rerun the queue to
1451 * avoid starvation. Notes:
1452 * - blk_mq_run_hw_queue() checks whether or not a queue has
1453 * been stopped before rerunning a queue.
1454 * - Some but not all block drivers stop a queue before
1455 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1458 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1459 * bit is set, run queue after a delay to avoid IO stalls
1460 * that could otherwise occur if the queue is idle. We'll do
1461 * similar if we couldn't get budget and SCHED_RESTART is set.
1463 needs_restart = blk_mq_sched_needs_restart(hctx);
1464 if (!needs_restart ||
1465 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1466 blk_mq_run_hw_queue(hctx, true);
1467 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1469 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1471 blk_mq_update_dispatch_busy(hctx, true);
1474 blk_mq_update_dispatch_busy(hctx, false);
1476 return (queued + errors) != 0;
1480 * __blk_mq_run_hw_queue - Run a hardware queue.
1481 * @hctx: Pointer to the hardware queue to run.
1483 * Send pending requests to the hardware.
1485 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1490 * We can't run the queue inline with ints disabled. Ensure that
1491 * we catch bad users of this early.
1493 WARN_ON_ONCE(in_interrupt());
1495 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1497 hctx_lock(hctx, &srcu_idx);
1498 blk_mq_sched_dispatch_requests(hctx);
1499 hctx_unlock(hctx, srcu_idx);
1502 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1504 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1506 if (cpu >= nr_cpu_ids)
1507 cpu = cpumask_first(hctx->cpumask);
1512 * It'd be great if the workqueue API had a way to pass
1513 * in a mask and had some smarts for more clever placement.
1514 * For now we just round-robin here, switching for every
1515 * BLK_MQ_CPU_WORK_BATCH queued items.
1517 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1520 int next_cpu = hctx->next_cpu;
1522 if (hctx->queue->nr_hw_queues == 1)
1523 return WORK_CPU_UNBOUND;
1525 if (--hctx->next_cpu_batch <= 0) {
1527 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1529 if (next_cpu >= nr_cpu_ids)
1530 next_cpu = blk_mq_first_mapped_cpu(hctx);
1531 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1535 * Do unbound schedule if we can't find a online CPU for this hctx,
1536 * and it should only happen in the path of handling CPU DEAD.
1538 if (!cpu_online(next_cpu)) {
1545 * Make sure to re-select CPU next time once after CPUs
1546 * in hctx->cpumask become online again.
1548 hctx->next_cpu = next_cpu;
1549 hctx->next_cpu_batch = 1;
1550 return WORK_CPU_UNBOUND;
1553 hctx->next_cpu = next_cpu;
1558 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1559 * @hctx: Pointer to the hardware queue to run.
1560 * @async: If we want to run the queue asynchronously.
1561 * @msecs: Milliseconds of delay to wait before running the queue.
1563 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1564 * with a delay of @msecs.
1566 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1567 unsigned long msecs)
1569 if (unlikely(blk_mq_hctx_stopped(hctx)))
1572 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1573 int cpu = get_cpu();
1574 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1575 __blk_mq_run_hw_queue(hctx);
1583 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1584 msecs_to_jiffies(msecs));
1588 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1589 * @hctx: Pointer to the hardware queue to run.
1590 * @msecs: Milliseconds of delay to wait before running the queue.
1592 * Run a hardware queue asynchronously with a delay of @msecs.
1594 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1596 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1598 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1601 * blk_mq_run_hw_queue - Start to run a hardware queue.
1602 * @hctx: Pointer to the hardware queue to run.
1603 * @async: If we want to run the queue asynchronously.
1605 * Check if the request queue is not in a quiesced state and if there are
1606 * pending requests to be sent. If this is true, run the queue to send requests
1609 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1615 * When queue is quiesced, we may be switching io scheduler, or
1616 * updating nr_hw_queues, or other things, and we can't run queue
1617 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1619 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1622 hctx_lock(hctx, &srcu_idx);
1623 need_run = !blk_queue_quiesced(hctx->queue) &&
1624 blk_mq_hctx_has_pending(hctx);
1625 hctx_unlock(hctx, srcu_idx);
1628 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1630 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1633 * Is the request queue handled by an IO scheduler that does not respect
1634 * hardware queues when dispatching?
1636 static bool blk_mq_has_sqsched(struct request_queue *q)
1638 struct elevator_queue *e = q->elevator;
1640 if (e && e->type->ops.dispatch_request &&
1641 !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE))
1647 * Return prefered queue to dispatch from (if any) for non-mq aware IO
1650 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
1652 struct blk_mq_hw_ctx *hctx;
1655 * If the IO scheduler does not respect hardware queues when
1656 * dispatching, we just don't bother with multiple HW queues and
1657 * dispatch from hctx for the current CPU since running multiple queues
1658 * just causes lock contention inside the scheduler and pointless cache
1661 hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT,
1662 raw_smp_processor_id());
1663 if (!blk_mq_hctx_stopped(hctx))
1669 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
1670 * @q: Pointer to the request queue to run.
1671 * @async: If we want to run the queue asynchronously.
1673 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1675 struct blk_mq_hw_ctx *hctx, *sq_hctx;
1679 if (blk_mq_has_sqsched(q))
1680 sq_hctx = blk_mq_get_sq_hctx(q);
1681 queue_for_each_hw_ctx(q, hctx, i) {
1682 if (blk_mq_hctx_stopped(hctx))
1685 * Dispatch from this hctx either if there's no hctx preferred
1686 * by IO scheduler or if it has requests that bypass the
1689 if (!sq_hctx || sq_hctx == hctx ||
1690 !list_empty_careful(&hctx->dispatch))
1691 blk_mq_run_hw_queue(hctx, async);
1694 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1697 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1698 * @q: Pointer to the request queue to run.
1699 * @msecs: Milliseconds of delay to wait before running the queues.
1701 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1703 struct blk_mq_hw_ctx *hctx, *sq_hctx;
1707 if (blk_mq_has_sqsched(q))
1708 sq_hctx = blk_mq_get_sq_hctx(q);
1709 queue_for_each_hw_ctx(q, hctx, i) {
1710 if (blk_mq_hctx_stopped(hctx))
1713 * Dispatch from this hctx either if there's no hctx preferred
1714 * by IO scheduler or if it has requests that bypass the
1717 if (!sq_hctx || sq_hctx == hctx ||
1718 !list_empty_careful(&hctx->dispatch))
1719 blk_mq_delay_run_hw_queue(hctx, msecs);
1722 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1725 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1726 * @q: request queue.
1728 * The caller is responsible for serializing this function against
1729 * blk_mq_{start,stop}_hw_queue().
1731 bool blk_mq_queue_stopped(struct request_queue *q)
1733 struct blk_mq_hw_ctx *hctx;
1736 queue_for_each_hw_ctx(q, hctx, i)
1737 if (blk_mq_hctx_stopped(hctx))
1742 EXPORT_SYMBOL(blk_mq_queue_stopped);
1745 * This function is often used for pausing .queue_rq() by driver when
1746 * there isn't enough resource or some conditions aren't satisfied, and
1747 * BLK_STS_RESOURCE is usually returned.
1749 * We do not guarantee that dispatch can be drained or blocked
1750 * after blk_mq_stop_hw_queue() returns. Please use
1751 * blk_mq_quiesce_queue() for that requirement.
1753 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1755 cancel_delayed_work(&hctx->run_work);
1757 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1759 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1762 * This function is often used for pausing .queue_rq() by driver when
1763 * there isn't enough resource or some conditions aren't satisfied, and
1764 * BLK_STS_RESOURCE is usually returned.
1766 * We do not guarantee that dispatch can be drained or blocked
1767 * after blk_mq_stop_hw_queues() returns. Please use
1768 * blk_mq_quiesce_queue() for that requirement.
1770 void blk_mq_stop_hw_queues(struct request_queue *q)
1772 struct blk_mq_hw_ctx *hctx;
1775 queue_for_each_hw_ctx(q, hctx, i)
1776 blk_mq_stop_hw_queue(hctx);
1778 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1780 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1782 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1784 blk_mq_run_hw_queue(hctx, false);
1786 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1788 void blk_mq_start_hw_queues(struct request_queue *q)
1790 struct blk_mq_hw_ctx *hctx;
1793 queue_for_each_hw_ctx(q, hctx, i)
1794 blk_mq_start_hw_queue(hctx);
1796 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1798 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1800 if (!blk_mq_hctx_stopped(hctx))
1803 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1804 blk_mq_run_hw_queue(hctx, async);
1806 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1808 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1810 struct blk_mq_hw_ctx *hctx;
1813 queue_for_each_hw_ctx(q, hctx, i)
1814 blk_mq_start_stopped_hw_queue(hctx, async);
1816 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1818 static void blk_mq_run_work_fn(struct work_struct *work)
1820 struct blk_mq_hw_ctx *hctx;
1822 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1825 * If we are stopped, don't run the queue.
1827 if (blk_mq_hctx_stopped(hctx))
1830 __blk_mq_run_hw_queue(hctx);
1833 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1837 struct blk_mq_ctx *ctx = rq->mq_ctx;
1838 enum hctx_type type = hctx->type;
1840 lockdep_assert_held(&ctx->lock);
1842 trace_block_rq_insert(rq);
1845 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1847 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1850 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1853 struct blk_mq_ctx *ctx = rq->mq_ctx;
1855 lockdep_assert_held(&ctx->lock);
1857 __blk_mq_insert_req_list(hctx, rq, at_head);
1858 blk_mq_hctx_mark_pending(hctx, ctx);
1862 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1863 * @rq: Pointer to request to be inserted.
1864 * @at_head: true if the request should be inserted at the head of the list.
1865 * @run_queue: If we should run the hardware queue after inserting the request.
1867 * Should only be used carefully, when the caller knows we want to
1868 * bypass a potential IO scheduler on the target device.
1870 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1873 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1875 spin_lock(&hctx->lock);
1877 list_add(&rq->queuelist, &hctx->dispatch);
1879 list_add_tail(&rq->queuelist, &hctx->dispatch);
1880 spin_unlock(&hctx->lock);
1883 blk_mq_run_hw_queue(hctx, false);
1886 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1887 struct list_head *list)
1891 enum hctx_type type = hctx->type;
1894 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1897 list_for_each_entry(rq, list, queuelist) {
1898 BUG_ON(rq->mq_ctx != ctx);
1899 trace_block_rq_insert(rq);
1902 spin_lock(&ctx->lock);
1903 list_splice_tail_init(list, &ctx->rq_lists[type]);
1904 blk_mq_hctx_mark_pending(hctx, ctx);
1905 spin_unlock(&ctx->lock);
1908 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1910 struct request *rqa = container_of(a, struct request, queuelist);
1911 struct request *rqb = container_of(b, struct request, queuelist);
1913 if (rqa->mq_ctx != rqb->mq_ctx)
1914 return rqa->mq_ctx > rqb->mq_ctx;
1915 if (rqa->mq_hctx != rqb->mq_hctx)
1916 return rqa->mq_hctx > rqb->mq_hctx;
1918 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1921 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1925 if (list_empty(&plug->mq_list))
1927 list_splice_init(&plug->mq_list, &list);
1929 if (plug->rq_count > 2 && plug->multiple_queues)
1930 list_sort(NULL, &list, plug_rq_cmp);
1935 struct list_head rq_list;
1936 struct request *rq, *head_rq = list_entry_rq(list.next);
1937 struct list_head *pos = &head_rq->queuelist; /* skip first */
1938 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1939 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1940 unsigned int depth = 1;
1942 list_for_each_continue(pos, &list) {
1943 rq = list_entry_rq(pos);
1945 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1950 list_cut_before(&rq_list, &list, pos);
1951 trace_block_unplug(head_rq->q, depth, !from_schedule);
1952 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1954 } while(!list_empty(&list));
1957 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1958 unsigned int nr_segs)
1962 if (bio->bi_opf & REQ_RAHEAD)
1963 rq->cmd_flags |= REQ_FAILFAST_MASK;
1965 rq->__sector = bio->bi_iter.bi_sector;
1966 rq->write_hint = bio->bi_write_hint;
1967 blk_rq_bio_prep(rq, bio, nr_segs);
1969 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
1970 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1973 blk_account_io_start(rq);
1976 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1978 blk_qc_t *cookie, bool last)
1980 struct request_queue *q = rq->q;
1981 struct blk_mq_queue_data bd = {
1985 blk_qc_t new_cookie;
1988 new_cookie = request_to_qc_t(hctx, rq);
1991 * For OK queue, we are done. For error, caller may kill it.
1992 * Any other error (busy), just add it to our list as we
1993 * previously would have done.
1995 ret = q->mq_ops->queue_rq(hctx, &bd);
1998 blk_mq_update_dispatch_busy(hctx, false);
1999 *cookie = new_cookie;
2001 case BLK_STS_RESOURCE:
2002 case BLK_STS_DEV_RESOURCE:
2003 blk_mq_update_dispatch_busy(hctx, true);
2004 __blk_mq_requeue_request(rq);
2007 blk_mq_update_dispatch_busy(hctx, false);
2008 *cookie = BLK_QC_T_NONE;
2015 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2018 bool bypass_insert, bool last)
2020 struct request_queue *q = rq->q;
2021 bool run_queue = true;
2025 * RCU or SRCU read lock is needed before checking quiesced flag.
2027 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2028 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2029 * and avoid driver to try to dispatch again.
2031 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2033 bypass_insert = false;
2037 if (q->elevator && !bypass_insert)
2040 budget_token = blk_mq_get_dispatch_budget(q);
2041 if (budget_token < 0)
2044 blk_mq_set_rq_budget_token(rq, budget_token);
2046 if (!blk_mq_get_driver_tag(rq)) {
2047 blk_mq_put_dispatch_budget(q, budget_token);
2051 return __blk_mq_issue_directly(hctx, rq, cookie, last);
2054 return BLK_STS_RESOURCE;
2056 blk_mq_sched_insert_request(rq, false, run_queue, false);
2062 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2063 * @hctx: Pointer of the associated hardware queue.
2064 * @rq: Pointer to request to be sent.
2065 * @cookie: Request queue cookie.
2067 * If the device has enough resources to accept a new request now, send the
2068 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2069 * we can try send it another time in the future. Requests inserted at this
2070 * queue have higher priority.
2072 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2073 struct request *rq, blk_qc_t *cookie)
2078 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2080 hctx_lock(hctx, &srcu_idx);
2082 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
2083 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2084 blk_mq_request_bypass_insert(rq, false, true);
2085 else if (ret != BLK_STS_OK)
2086 blk_mq_end_request(rq, ret);
2088 hctx_unlock(hctx, srcu_idx);
2091 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2095 blk_qc_t unused_cookie;
2096 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2098 hctx_lock(hctx, &srcu_idx);
2099 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
2100 hctx_unlock(hctx, srcu_idx);
2105 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2106 struct list_head *list)
2111 while (!list_empty(list)) {
2113 struct request *rq = list_first_entry(list, struct request,
2116 list_del_init(&rq->queuelist);
2117 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2118 if (ret != BLK_STS_OK) {
2119 if (ret == BLK_STS_RESOURCE ||
2120 ret == BLK_STS_DEV_RESOURCE) {
2121 blk_mq_request_bypass_insert(rq, false,
2125 blk_mq_end_request(rq, ret);
2132 * If we didn't flush the entire list, we could have told
2133 * the driver there was more coming, but that turned out to
2136 if ((!list_empty(list) || errors) &&
2137 hctx->queue->mq_ops->commit_rqs && queued)
2138 hctx->queue->mq_ops->commit_rqs(hctx);
2141 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2143 list_add_tail(&rq->queuelist, &plug->mq_list);
2145 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
2146 struct request *tmp;
2148 tmp = list_first_entry(&plug->mq_list, struct request,
2150 if (tmp->q != rq->q)
2151 plug->multiple_queues = true;
2156 * blk_mq_submit_bio - Create and send a request to block device.
2157 * @bio: Bio pointer.
2159 * Builds up a request structure from @q and @bio and send to the device. The
2160 * request may not be queued directly to hardware if:
2161 * * This request can be merged with another one
2162 * * We want to place request at plug queue for possible future merging
2163 * * There is an IO scheduler active at this queue
2165 * It will not queue the request if there is an error with the bio, or at the
2168 * Returns: Request queue cookie.
2170 blk_qc_t blk_mq_submit_bio(struct bio *bio)
2172 struct request_queue *q = bio->bi_bdev->bd_disk->queue;
2173 const int is_sync = op_is_sync(bio->bi_opf);
2174 const int is_flush_fua = op_is_flush(bio->bi_opf);
2175 struct blk_mq_alloc_data data = {
2179 struct blk_plug *plug;
2180 struct request *same_queue_rq = NULL;
2181 unsigned int nr_segs;
2186 blk_queue_bounce(q, &bio);
2187 __blk_queue_split(&bio, &nr_segs);
2189 if (!bio_integrity_prep(bio))
2192 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2193 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2196 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2199 rq_qos_throttle(q, bio);
2201 hipri = bio->bi_opf & REQ_HIPRI;
2203 data.cmd_flags = bio->bi_opf;
2204 rq = __blk_mq_alloc_request(&data);
2205 if (unlikely(!rq)) {
2206 rq_qos_cleanup(q, bio);
2207 if (bio->bi_opf & REQ_NOWAIT)
2208 bio_wouldblock_error(bio);
2212 trace_block_getrq(bio);
2214 rq_qos_track(q, rq, bio);
2216 cookie = request_to_qc_t(data.hctx, rq);
2218 blk_mq_bio_to_request(rq, bio, nr_segs);
2220 ret = blk_crypto_init_request(rq);
2221 if (ret != BLK_STS_OK) {
2222 bio->bi_status = ret;
2224 blk_mq_free_request(rq);
2225 return BLK_QC_T_NONE;
2228 plug = blk_mq_plug(q, bio);
2229 if (unlikely(is_flush_fua)) {
2230 /* Bypass scheduler for flush requests */
2231 blk_insert_flush(rq);
2232 blk_mq_run_hw_queue(data.hctx, true);
2233 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2234 !blk_queue_nonrot(q))) {
2236 * Use plugging if we have a ->commit_rqs() hook as well, as
2237 * we know the driver uses bd->last in a smart fashion.
2239 * Use normal plugging if this disk is slow HDD, as sequential
2240 * IO may benefit a lot from plug merging.
2242 unsigned int request_count = plug->rq_count;
2243 struct request *last = NULL;
2246 trace_block_plug(q);
2248 last = list_entry_rq(plug->mq_list.prev);
2250 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2251 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2252 blk_flush_plug_list(plug, false);
2253 trace_block_plug(q);
2256 blk_add_rq_to_plug(plug, rq);
2257 } else if (q->elevator) {
2258 /* Insert the request at the IO scheduler queue */
2259 blk_mq_sched_insert_request(rq, false, true, true);
2260 } else if (plug && !blk_queue_nomerges(q)) {
2262 * We do limited plugging. If the bio can be merged, do that.
2263 * Otherwise the existing request in the plug list will be
2264 * issued. So the plug list will have one request at most
2265 * The plug list might get flushed before this. If that happens,
2266 * the plug list is empty, and same_queue_rq is invalid.
2268 if (list_empty(&plug->mq_list))
2269 same_queue_rq = NULL;
2270 if (same_queue_rq) {
2271 list_del_init(&same_queue_rq->queuelist);
2274 blk_add_rq_to_plug(plug, rq);
2275 trace_block_plug(q);
2277 if (same_queue_rq) {
2278 data.hctx = same_queue_rq->mq_hctx;
2279 trace_block_unplug(q, 1, true);
2280 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2283 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2284 !data.hctx->dispatch_busy) {
2286 * There is no scheduler and we can try to send directly
2289 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2292 blk_mq_sched_insert_request(rq, false, true, true);
2296 return BLK_QC_T_NONE;
2300 return BLK_QC_T_NONE;
2303 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2304 unsigned int hctx_idx)
2308 if (tags->rqs && set->ops->exit_request) {
2311 for (i = 0; i < tags->nr_tags; i++) {
2312 struct request *rq = tags->static_rqs[i];
2316 set->ops->exit_request(set, rq, hctx_idx);
2317 tags->static_rqs[i] = NULL;
2321 while (!list_empty(&tags->page_list)) {
2322 page = list_first_entry(&tags->page_list, struct page, lru);
2323 list_del_init(&page->lru);
2325 * Remove kmemleak object previously allocated in
2326 * blk_mq_alloc_rqs().
2328 kmemleak_free(page_address(page));
2329 __free_pages(page, page->private);
2333 void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags)
2337 kfree(tags->static_rqs);
2338 tags->static_rqs = NULL;
2340 blk_mq_free_tags(tags, flags);
2343 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2344 unsigned int hctx_idx,
2345 unsigned int nr_tags,
2346 unsigned int reserved_tags,
2349 struct blk_mq_tags *tags;
2352 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2353 if (node == NUMA_NO_NODE)
2354 node = set->numa_node;
2356 tags = blk_mq_init_tags(nr_tags, reserved_tags, node, flags);
2360 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2361 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2364 blk_mq_free_tags(tags, flags);
2368 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2369 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2371 if (!tags->static_rqs) {
2373 blk_mq_free_tags(tags, flags);
2380 static size_t order_to_size(unsigned int order)
2382 return (size_t)PAGE_SIZE << order;
2385 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2386 unsigned int hctx_idx, int node)
2390 if (set->ops->init_request) {
2391 ret = set->ops->init_request(set, rq, hctx_idx, node);
2396 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2400 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2401 unsigned int hctx_idx, unsigned int depth)
2403 unsigned int i, j, entries_per_page, max_order = 4;
2404 size_t rq_size, left;
2407 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2408 if (node == NUMA_NO_NODE)
2409 node = set->numa_node;
2411 INIT_LIST_HEAD(&tags->page_list);
2414 * rq_size is the size of the request plus driver payload, rounded
2415 * to the cacheline size
2417 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2419 left = rq_size * depth;
2421 for (i = 0; i < depth; ) {
2422 int this_order = max_order;
2427 while (this_order && left < order_to_size(this_order - 1))
2431 page = alloc_pages_node(node,
2432 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2438 if (order_to_size(this_order) < rq_size)
2445 page->private = this_order;
2446 list_add_tail(&page->lru, &tags->page_list);
2448 p = page_address(page);
2450 * Allow kmemleak to scan these pages as they contain pointers
2451 * to additional allocations like via ops->init_request().
2453 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2454 entries_per_page = order_to_size(this_order) / rq_size;
2455 to_do = min(entries_per_page, depth - i);
2456 left -= to_do * rq_size;
2457 for (j = 0; j < to_do; j++) {
2458 struct request *rq = p;
2460 tags->static_rqs[i] = rq;
2461 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2462 tags->static_rqs[i] = NULL;
2473 blk_mq_free_rqs(set, tags, hctx_idx);
2477 struct rq_iter_data {
2478 struct blk_mq_hw_ctx *hctx;
2482 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2484 struct rq_iter_data *iter_data = data;
2486 if (rq->mq_hctx != iter_data->hctx)
2488 iter_data->has_rq = true;
2492 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2494 struct blk_mq_tags *tags = hctx->sched_tags ?
2495 hctx->sched_tags : hctx->tags;
2496 struct rq_iter_data data = {
2500 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2504 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2505 struct blk_mq_hw_ctx *hctx)
2507 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2509 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2514 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2516 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2517 struct blk_mq_hw_ctx, cpuhp_online);
2519 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2520 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2524 * Prevent new request from being allocated on the current hctx.
2526 * The smp_mb__after_atomic() Pairs with the implied barrier in
2527 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2528 * seen once we return from the tag allocator.
2530 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2531 smp_mb__after_atomic();
2534 * Try to grab a reference to the queue and wait for any outstanding
2535 * requests. If we could not grab a reference the queue has been
2536 * frozen and there are no requests.
2538 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2539 while (blk_mq_hctx_has_requests(hctx))
2541 percpu_ref_put(&hctx->queue->q_usage_counter);
2547 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2549 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2550 struct blk_mq_hw_ctx, cpuhp_online);
2552 if (cpumask_test_cpu(cpu, hctx->cpumask))
2553 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2558 * 'cpu' is going away. splice any existing rq_list entries from this
2559 * software queue to the hw queue dispatch list, and ensure that it
2562 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2564 struct blk_mq_hw_ctx *hctx;
2565 struct blk_mq_ctx *ctx;
2567 enum hctx_type type;
2569 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2570 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2573 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2576 spin_lock(&ctx->lock);
2577 if (!list_empty(&ctx->rq_lists[type])) {
2578 list_splice_init(&ctx->rq_lists[type], &tmp);
2579 blk_mq_hctx_clear_pending(hctx, ctx);
2581 spin_unlock(&ctx->lock);
2583 if (list_empty(&tmp))
2586 spin_lock(&hctx->lock);
2587 list_splice_tail_init(&tmp, &hctx->dispatch);
2588 spin_unlock(&hctx->lock);
2590 blk_mq_run_hw_queue(hctx, true);
2594 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2596 if (!(hctx->flags & BLK_MQ_F_STACKING))
2597 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2598 &hctx->cpuhp_online);
2599 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2603 /* hctx->ctxs will be freed in queue's release handler */
2604 static void blk_mq_exit_hctx(struct request_queue *q,
2605 struct blk_mq_tag_set *set,
2606 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2608 if (blk_mq_hw_queue_mapped(hctx))
2609 blk_mq_tag_idle(hctx);
2611 if (set->ops->exit_request)
2612 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2614 if (set->ops->exit_hctx)
2615 set->ops->exit_hctx(hctx, hctx_idx);
2617 blk_mq_remove_cpuhp(hctx);
2619 spin_lock(&q->unused_hctx_lock);
2620 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2621 spin_unlock(&q->unused_hctx_lock);
2624 static void blk_mq_exit_hw_queues(struct request_queue *q,
2625 struct blk_mq_tag_set *set, int nr_queue)
2627 struct blk_mq_hw_ctx *hctx;
2630 queue_for_each_hw_ctx(q, hctx, i) {
2633 blk_mq_debugfs_unregister_hctx(hctx);
2634 blk_mq_exit_hctx(q, set, hctx, i);
2638 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2640 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2642 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2643 __alignof__(struct blk_mq_hw_ctx)) !=
2644 sizeof(struct blk_mq_hw_ctx));
2646 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2647 hw_ctx_size += sizeof(struct srcu_struct);
2652 static int blk_mq_init_hctx(struct request_queue *q,
2653 struct blk_mq_tag_set *set,
2654 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2656 hctx->queue_num = hctx_idx;
2658 if (!(hctx->flags & BLK_MQ_F_STACKING))
2659 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2660 &hctx->cpuhp_online);
2661 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2663 hctx->tags = set->tags[hctx_idx];
2665 if (set->ops->init_hctx &&
2666 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2667 goto unregister_cpu_notifier;
2669 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2675 if (set->ops->exit_hctx)
2676 set->ops->exit_hctx(hctx, hctx_idx);
2677 unregister_cpu_notifier:
2678 blk_mq_remove_cpuhp(hctx);
2682 static struct blk_mq_hw_ctx *
2683 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2686 struct blk_mq_hw_ctx *hctx;
2687 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2689 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2691 goto fail_alloc_hctx;
2693 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2696 atomic_set(&hctx->nr_active, 0);
2697 if (node == NUMA_NO_NODE)
2698 node = set->numa_node;
2699 hctx->numa_node = node;
2701 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2702 spin_lock_init(&hctx->lock);
2703 INIT_LIST_HEAD(&hctx->dispatch);
2705 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
2707 INIT_LIST_HEAD(&hctx->hctx_list);
2710 * Allocate space for all possible cpus to avoid allocation at
2713 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2718 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2719 gfp, node, false, false))
2723 spin_lock_init(&hctx->dispatch_wait_lock);
2724 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2725 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2727 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2731 if (hctx->flags & BLK_MQ_F_BLOCKING)
2732 init_srcu_struct(hctx->srcu);
2733 blk_mq_hctx_kobj_init(hctx);
2738 sbitmap_free(&hctx->ctx_map);
2742 free_cpumask_var(hctx->cpumask);
2749 static void blk_mq_init_cpu_queues(struct request_queue *q,
2750 unsigned int nr_hw_queues)
2752 struct blk_mq_tag_set *set = q->tag_set;
2755 for_each_possible_cpu(i) {
2756 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2757 struct blk_mq_hw_ctx *hctx;
2761 spin_lock_init(&__ctx->lock);
2762 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2763 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2768 * Set local node, IFF we have more than one hw queue. If
2769 * not, we remain on the home node of the device
2771 for (j = 0; j < set->nr_maps; j++) {
2772 hctx = blk_mq_map_queue_type(q, j, i);
2773 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2774 hctx->numa_node = cpu_to_node(i);
2779 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2782 unsigned int flags = set->flags;
2785 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2786 set->queue_depth, set->reserved_tags, flags);
2787 if (!set->tags[hctx_idx])
2790 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2795 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2796 set->tags[hctx_idx] = NULL;
2800 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2801 unsigned int hctx_idx)
2803 unsigned int flags = set->flags;
2805 if (set->tags && set->tags[hctx_idx]) {
2806 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2807 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2808 set->tags[hctx_idx] = NULL;
2812 static void blk_mq_map_swqueue(struct request_queue *q)
2814 unsigned int i, j, hctx_idx;
2815 struct blk_mq_hw_ctx *hctx;
2816 struct blk_mq_ctx *ctx;
2817 struct blk_mq_tag_set *set = q->tag_set;
2819 queue_for_each_hw_ctx(q, hctx, i) {
2820 cpumask_clear(hctx->cpumask);
2822 hctx->dispatch_from = NULL;
2826 * Map software to hardware queues.
2828 * If the cpu isn't present, the cpu is mapped to first hctx.
2830 for_each_possible_cpu(i) {
2832 ctx = per_cpu_ptr(q->queue_ctx, i);
2833 for (j = 0; j < set->nr_maps; j++) {
2834 if (!set->map[j].nr_queues) {
2835 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2836 HCTX_TYPE_DEFAULT, i);
2839 hctx_idx = set->map[j].mq_map[i];
2840 /* unmapped hw queue can be remapped after CPU topo changed */
2841 if (!set->tags[hctx_idx] &&
2842 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2844 * If tags initialization fail for some hctx,
2845 * that hctx won't be brought online. In this
2846 * case, remap the current ctx to hctx[0] which
2847 * is guaranteed to always have tags allocated
2849 set->map[j].mq_map[i] = 0;
2852 hctx = blk_mq_map_queue_type(q, j, i);
2853 ctx->hctxs[j] = hctx;
2855 * If the CPU is already set in the mask, then we've
2856 * mapped this one already. This can happen if
2857 * devices share queues across queue maps.
2859 if (cpumask_test_cpu(i, hctx->cpumask))
2862 cpumask_set_cpu(i, hctx->cpumask);
2864 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2865 hctx->ctxs[hctx->nr_ctx++] = ctx;
2868 * If the nr_ctx type overflows, we have exceeded the
2869 * amount of sw queues we can support.
2871 BUG_ON(!hctx->nr_ctx);
2874 for (; j < HCTX_MAX_TYPES; j++)
2875 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2876 HCTX_TYPE_DEFAULT, i);
2879 queue_for_each_hw_ctx(q, hctx, i) {
2881 * If no software queues are mapped to this hardware queue,
2882 * disable it and free the request entries.
2884 if (!hctx->nr_ctx) {
2885 /* Never unmap queue 0. We need it as a
2886 * fallback in case of a new remap fails
2889 if (i && set->tags[i])
2890 blk_mq_free_map_and_requests(set, i);
2896 hctx->tags = set->tags[i];
2897 WARN_ON(!hctx->tags);
2900 * Set the map size to the number of mapped software queues.
2901 * This is more accurate and more efficient than looping
2902 * over all possibly mapped software queues.
2904 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2907 * Initialize batch roundrobin counts
2909 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2910 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2915 * Caller needs to ensure that we're either frozen/quiesced, or that
2916 * the queue isn't live yet.
2918 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2920 struct blk_mq_hw_ctx *hctx;
2923 queue_for_each_hw_ctx(q, hctx, i) {
2925 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
2927 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
2931 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
2934 struct request_queue *q;
2936 lockdep_assert_held(&set->tag_list_lock);
2938 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2939 blk_mq_freeze_queue(q);
2940 queue_set_hctx_shared(q, shared);
2941 blk_mq_unfreeze_queue(q);
2945 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2947 struct blk_mq_tag_set *set = q->tag_set;
2949 mutex_lock(&set->tag_list_lock);
2950 list_del(&q->tag_set_list);
2951 if (list_is_singular(&set->tag_list)) {
2952 /* just transitioned to unshared */
2953 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
2954 /* update existing queue */
2955 blk_mq_update_tag_set_shared(set, false);
2957 mutex_unlock(&set->tag_list_lock);
2958 INIT_LIST_HEAD(&q->tag_set_list);
2961 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2962 struct request_queue *q)
2964 mutex_lock(&set->tag_list_lock);
2967 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2969 if (!list_empty(&set->tag_list) &&
2970 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2971 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
2972 /* update existing queue */
2973 blk_mq_update_tag_set_shared(set, true);
2975 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
2976 queue_set_hctx_shared(q, true);
2977 list_add_tail(&q->tag_set_list, &set->tag_list);
2979 mutex_unlock(&set->tag_list_lock);
2982 /* All allocations will be freed in release handler of q->mq_kobj */
2983 static int blk_mq_alloc_ctxs(struct request_queue *q)
2985 struct blk_mq_ctxs *ctxs;
2988 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2992 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2993 if (!ctxs->queue_ctx)
2996 for_each_possible_cpu(cpu) {
2997 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3001 q->mq_kobj = &ctxs->kobj;
3002 q->queue_ctx = ctxs->queue_ctx;
3011 * It is the actual release handler for mq, but we do it from
3012 * request queue's release handler for avoiding use-after-free
3013 * and headache because q->mq_kobj shouldn't have been introduced,
3014 * but we can't group ctx/kctx kobj without it.
3016 void blk_mq_release(struct request_queue *q)
3018 struct blk_mq_hw_ctx *hctx, *next;
3021 queue_for_each_hw_ctx(q, hctx, i)
3022 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3024 /* all hctx are in .unused_hctx_list now */
3025 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3026 list_del_init(&hctx->hctx_list);
3027 kobject_put(&hctx->kobj);
3030 kfree(q->queue_hw_ctx);
3033 * release .mq_kobj and sw queue's kobject now because
3034 * both share lifetime with request queue.
3036 blk_mq_sysfs_deinit(q);
3039 struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3042 struct request_queue *uninit_q, *q;
3044 uninit_q = blk_alloc_queue(set->numa_node);
3046 return ERR_PTR(-ENOMEM);
3047 uninit_q->queuedata = queuedata;
3050 * Initialize the queue without an elevator. device_add_disk() will do
3051 * the initialization.
3053 q = blk_mq_init_allocated_queue(set, uninit_q, false);
3055 blk_cleanup_queue(uninit_q);
3059 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data);
3061 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3063 return blk_mq_init_queue_data(set, NULL);
3065 EXPORT_SYMBOL(blk_mq_init_queue);
3068 * Helper for setting up a queue with mq ops, given queue depth, and
3069 * the passed in mq ops flags.
3071 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
3072 const struct blk_mq_ops *ops,
3073 unsigned int queue_depth,
3074 unsigned int set_flags)
3076 struct request_queue *q;
3079 memset(set, 0, sizeof(*set));
3081 set->nr_hw_queues = 1;
3083 set->queue_depth = queue_depth;
3084 set->numa_node = NUMA_NO_NODE;
3085 set->flags = set_flags;
3087 ret = blk_mq_alloc_tag_set(set);
3089 return ERR_PTR(ret);
3091 q = blk_mq_init_queue(set);
3093 blk_mq_free_tag_set(set);
3099 EXPORT_SYMBOL(blk_mq_init_sq_queue);
3101 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3102 struct blk_mq_tag_set *set, struct request_queue *q,
3103 int hctx_idx, int node)
3105 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3107 /* reuse dead hctx first */
3108 spin_lock(&q->unused_hctx_lock);
3109 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3110 if (tmp->numa_node == node) {
3116 list_del_init(&hctx->hctx_list);
3117 spin_unlock(&q->unused_hctx_lock);
3120 hctx = blk_mq_alloc_hctx(q, set, node);
3124 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3130 kobject_put(&hctx->kobj);
3135 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3136 struct request_queue *q)
3139 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3141 if (q->nr_hw_queues < set->nr_hw_queues) {
3142 struct blk_mq_hw_ctx **new_hctxs;
3144 new_hctxs = kcalloc_node(set->nr_hw_queues,
3145 sizeof(*new_hctxs), GFP_KERNEL,
3150 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3152 q->queue_hw_ctx = new_hctxs;
3157 /* protect against switching io scheduler */
3158 mutex_lock(&q->sysfs_lock);
3159 for (i = 0; i < set->nr_hw_queues; i++) {
3161 struct blk_mq_hw_ctx *hctx;
3163 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3165 * If the hw queue has been mapped to another numa node,
3166 * we need to realloc the hctx. If allocation fails, fallback
3167 * to use the previous one.
3169 if (hctxs[i] && (hctxs[i]->numa_node == node))
3172 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3175 blk_mq_exit_hctx(q, set, hctxs[i], i);
3179 pr_warn("Allocate new hctx on node %d fails,\
3180 fallback to previous one on node %d\n",
3181 node, hctxs[i]->numa_node);
3187 * Increasing nr_hw_queues fails. Free the newly allocated
3188 * hctxs and keep the previous q->nr_hw_queues.
3190 if (i != set->nr_hw_queues) {
3191 j = q->nr_hw_queues;
3195 end = q->nr_hw_queues;
3196 q->nr_hw_queues = set->nr_hw_queues;
3199 for (; j < end; j++) {
3200 struct blk_mq_hw_ctx *hctx = hctxs[j];
3204 blk_mq_free_map_and_requests(set, j);
3205 blk_mq_exit_hctx(q, set, hctx, j);
3209 mutex_unlock(&q->sysfs_lock);
3212 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3213 struct request_queue *q,
3216 /* mark the queue as mq asap */
3217 q->mq_ops = set->ops;
3219 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3220 blk_mq_poll_stats_bkt,
3221 BLK_MQ_POLL_STATS_BKTS, q);
3225 if (blk_mq_alloc_ctxs(q))
3228 /* init q->mq_kobj and sw queues' kobjects */
3229 blk_mq_sysfs_init(q);
3231 INIT_LIST_HEAD(&q->unused_hctx_list);
3232 spin_lock_init(&q->unused_hctx_lock);
3234 blk_mq_realloc_hw_ctxs(set, q);
3235 if (!q->nr_hw_queues)
3238 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3239 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3243 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3244 if (set->nr_maps > HCTX_TYPE_POLL &&
3245 set->map[HCTX_TYPE_POLL].nr_queues)
3246 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3248 q->sg_reserved_size = INT_MAX;
3250 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3251 INIT_LIST_HEAD(&q->requeue_list);
3252 spin_lock_init(&q->requeue_lock);
3254 q->nr_requests = set->queue_depth;
3257 * Default to classic polling
3259 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3261 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3262 blk_mq_add_queue_tag_set(set, q);
3263 blk_mq_map_swqueue(q);
3266 elevator_init_mq(q);
3271 kfree(q->queue_hw_ctx);
3272 q->nr_hw_queues = 0;
3273 blk_mq_sysfs_deinit(q);
3275 blk_stat_free_callback(q->poll_cb);
3279 return ERR_PTR(-ENOMEM);
3281 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3283 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3284 void blk_mq_exit_queue(struct request_queue *q)
3286 struct blk_mq_tag_set *set = q->tag_set;
3288 blk_mq_del_queue_tag_set(q);
3289 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3292 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3296 for (i = 0; i < set->nr_hw_queues; i++) {
3297 if (!__blk_mq_alloc_map_and_request(set, i))
3306 blk_mq_free_map_and_requests(set, i);
3312 * Allocate the request maps associated with this tag_set. Note that this
3313 * may reduce the depth asked for, if memory is tight. set->queue_depth
3314 * will be updated to reflect the allocated depth.
3316 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3321 depth = set->queue_depth;
3323 err = __blk_mq_alloc_rq_maps(set);
3327 set->queue_depth >>= 1;
3328 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3332 } while (set->queue_depth);
3334 if (!set->queue_depth || err) {
3335 pr_err("blk-mq: failed to allocate request map\n");
3339 if (depth != set->queue_depth)
3340 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3341 depth, set->queue_depth);
3346 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3349 * blk_mq_map_queues() and multiple .map_queues() implementations
3350 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3351 * number of hardware queues.
3353 if (set->nr_maps == 1)
3354 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3356 if (set->ops->map_queues && !is_kdump_kernel()) {
3360 * transport .map_queues is usually done in the following
3363 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3364 * mask = get_cpu_mask(queue)
3365 * for_each_cpu(cpu, mask)
3366 * set->map[x].mq_map[cpu] = queue;
3369 * When we need to remap, the table has to be cleared for
3370 * killing stale mapping since one CPU may not be mapped
3373 for (i = 0; i < set->nr_maps; i++)
3374 blk_mq_clear_mq_map(&set->map[i]);
3376 return set->ops->map_queues(set);
3378 BUG_ON(set->nr_maps > 1);
3379 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3383 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3384 int cur_nr_hw_queues, int new_nr_hw_queues)
3386 struct blk_mq_tags **new_tags;
3388 if (cur_nr_hw_queues >= new_nr_hw_queues)
3391 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3392 GFP_KERNEL, set->numa_node);
3397 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3398 sizeof(*set->tags));
3400 set->tags = new_tags;
3401 set->nr_hw_queues = new_nr_hw_queues;
3406 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
3407 int new_nr_hw_queues)
3409 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
3413 * Alloc a tag set to be associated with one or more request queues.
3414 * May fail with EINVAL for various error conditions. May adjust the
3415 * requested depth down, if it's too large. In that case, the set
3416 * value will be stored in set->queue_depth.
3418 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3422 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3424 if (!set->nr_hw_queues)
3426 if (!set->queue_depth)
3428 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3431 if (!set->ops->queue_rq)
3434 if (!set->ops->get_budget ^ !set->ops->put_budget)
3437 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3438 pr_info("blk-mq: reduced tag depth to %u\n",
3440 set->queue_depth = BLK_MQ_MAX_DEPTH;
3445 else if (set->nr_maps > HCTX_MAX_TYPES)
3449 * If a crashdump is active, then we are potentially in a very
3450 * memory constrained environment. Limit us to 1 queue and
3451 * 64 tags to prevent using too much memory.
3453 if (is_kdump_kernel()) {
3454 set->nr_hw_queues = 1;
3456 set->queue_depth = min(64U, set->queue_depth);
3459 * There is no use for more h/w queues than cpus if we just have
3462 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3463 set->nr_hw_queues = nr_cpu_ids;
3465 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
3469 for (i = 0; i < set->nr_maps; i++) {
3470 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3471 sizeof(set->map[i].mq_map[0]),
3472 GFP_KERNEL, set->numa_node);
3473 if (!set->map[i].mq_map)
3474 goto out_free_mq_map;
3475 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3478 ret = blk_mq_update_queue_map(set);
3480 goto out_free_mq_map;
3482 ret = blk_mq_alloc_map_and_requests(set);
3484 goto out_free_mq_map;
3486 if (blk_mq_is_sbitmap_shared(set->flags)) {
3487 atomic_set(&set->active_queues_shared_sbitmap, 0);
3489 if (blk_mq_init_shared_sbitmap(set, set->flags)) {
3491 goto out_free_mq_rq_maps;
3495 mutex_init(&set->tag_list_lock);
3496 INIT_LIST_HEAD(&set->tag_list);
3500 out_free_mq_rq_maps:
3501 for (i = 0; i < set->nr_hw_queues; i++)
3502 blk_mq_free_map_and_requests(set, i);
3504 for (i = 0; i < set->nr_maps; i++) {
3505 kfree(set->map[i].mq_map);
3506 set->map[i].mq_map = NULL;
3512 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3514 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3518 for (i = 0; i < set->nr_hw_queues; i++)
3519 blk_mq_free_map_and_requests(set, i);
3521 if (blk_mq_is_sbitmap_shared(set->flags))
3522 blk_mq_exit_shared_sbitmap(set);
3524 for (j = 0; j < set->nr_maps; j++) {
3525 kfree(set->map[j].mq_map);
3526 set->map[j].mq_map = NULL;
3532 EXPORT_SYMBOL(blk_mq_free_tag_set);
3534 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3536 struct blk_mq_tag_set *set = q->tag_set;
3537 struct blk_mq_hw_ctx *hctx;
3543 if (q->nr_requests == nr)
3546 blk_mq_freeze_queue(q);
3547 blk_mq_quiesce_queue(q);
3550 queue_for_each_hw_ctx(q, hctx, i) {
3554 * If we're using an MQ scheduler, just update the scheduler
3555 * queue depth. This is similar to what the old code would do.
3557 if (!hctx->sched_tags) {
3558 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3560 if (!ret && blk_mq_is_sbitmap_shared(set->flags))
3561 blk_mq_tag_resize_shared_sbitmap(set, nr);
3563 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3568 if (q->elevator && q->elevator->type->ops.depth_updated)
3569 q->elevator->type->ops.depth_updated(hctx);
3573 q->nr_requests = nr;
3575 blk_mq_unquiesce_queue(q);
3576 blk_mq_unfreeze_queue(q);
3582 * request_queue and elevator_type pair.
3583 * It is just used by __blk_mq_update_nr_hw_queues to cache
3584 * the elevator_type associated with a request_queue.
3586 struct blk_mq_qe_pair {
3587 struct list_head node;
3588 struct request_queue *q;
3589 struct elevator_type *type;
3593 * Cache the elevator_type in qe pair list and switch the
3594 * io scheduler to 'none'
3596 static bool blk_mq_elv_switch_none(struct list_head *head,
3597 struct request_queue *q)
3599 struct blk_mq_qe_pair *qe;
3604 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3608 INIT_LIST_HEAD(&qe->node);
3610 qe->type = q->elevator->type;
3611 list_add(&qe->node, head);
3613 mutex_lock(&q->sysfs_lock);
3615 * After elevator_switch_mq, the previous elevator_queue will be
3616 * released by elevator_release. The reference of the io scheduler
3617 * module get by elevator_get will also be put. So we need to get
3618 * a reference of the io scheduler module here to prevent it to be
3621 __module_get(qe->type->elevator_owner);
3622 elevator_switch_mq(q, NULL);
3623 mutex_unlock(&q->sysfs_lock);
3628 static void blk_mq_elv_switch_back(struct list_head *head,
3629 struct request_queue *q)
3631 struct blk_mq_qe_pair *qe;
3632 struct elevator_type *t = NULL;
3634 list_for_each_entry(qe, head, node)
3643 list_del(&qe->node);
3646 mutex_lock(&q->sysfs_lock);
3647 elevator_switch_mq(q, t);
3648 mutex_unlock(&q->sysfs_lock);
3651 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3654 struct request_queue *q;
3656 int prev_nr_hw_queues;
3658 lockdep_assert_held(&set->tag_list_lock);
3660 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3661 nr_hw_queues = nr_cpu_ids;
3662 if (nr_hw_queues < 1)
3664 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3667 list_for_each_entry(q, &set->tag_list, tag_set_list)
3668 blk_mq_freeze_queue(q);
3670 * Switch IO scheduler to 'none', cleaning up the data associated
3671 * with the previous scheduler. We will switch back once we are done
3672 * updating the new sw to hw queue mappings.
3674 list_for_each_entry(q, &set->tag_list, tag_set_list)
3675 if (!blk_mq_elv_switch_none(&head, q))
3678 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3679 blk_mq_debugfs_unregister_hctxs(q);
3680 blk_mq_sysfs_unregister(q);
3683 prev_nr_hw_queues = set->nr_hw_queues;
3684 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3688 set->nr_hw_queues = nr_hw_queues;
3690 blk_mq_update_queue_map(set);
3691 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3692 blk_mq_realloc_hw_ctxs(set, q);
3693 if (q->nr_hw_queues != set->nr_hw_queues) {
3694 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3695 nr_hw_queues, prev_nr_hw_queues);
3696 set->nr_hw_queues = prev_nr_hw_queues;
3697 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3700 blk_mq_map_swqueue(q);
3704 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3705 blk_mq_sysfs_register(q);
3706 blk_mq_debugfs_register_hctxs(q);
3710 list_for_each_entry(q, &set->tag_list, tag_set_list)
3711 blk_mq_elv_switch_back(&head, q);
3713 list_for_each_entry(q, &set->tag_list, tag_set_list)
3714 blk_mq_unfreeze_queue(q);
3717 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3719 mutex_lock(&set->tag_list_lock);
3720 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3721 mutex_unlock(&set->tag_list_lock);
3723 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3725 /* Enable polling stats and return whether they were already enabled. */
3726 static bool blk_poll_stats_enable(struct request_queue *q)
3728 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3729 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3731 blk_stat_add_callback(q, q->poll_cb);
3735 static void blk_mq_poll_stats_start(struct request_queue *q)
3738 * We don't arm the callback if polling stats are not enabled or the
3739 * callback is already active.
3741 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3742 blk_stat_is_active(q->poll_cb))
3745 blk_stat_activate_msecs(q->poll_cb, 100);
3748 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3750 struct request_queue *q = cb->data;
3753 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3754 if (cb->stat[bucket].nr_samples)
3755 q->poll_stat[bucket] = cb->stat[bucket];
3759 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3762 unsigned long ret = 0;
3766 * If stats collection isn't on, don't sleep but turn it on for
3769 if (!blk_poll_stats_enable(q))
3773 * As an optimistic guess, use half of the mean service time
3774 * for this type of request. We can (and should) make this smarter.
3775 * For instance, if the completion latencies are tight, we can
3776 * get closer than just half the mean. This is especially
3777 * important on devices where the completion latencies are longer
3778 * than ~10 usec. We do use the stats for the relevant IO size
3779 * if available which does lead to better estimates.
3781 bucket = blk_mq_poll_stats_bkt(rq);
3785 if (q->poll_stat[bucket].nr_samples)
3786 ret = (q->poll_stat[bucket].mean + 1) / 2;
3791 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3794 struct hrtimer_sleeper hs;
3795 enum hrtimer_mode mode;
3799 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3803 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3805 * 0: use half of prev avg
3806 * >0: use this specific value
3808 if (q->poll_nsec > 0)
3809 nsecs = q->poll_nsec;
3811 nsecs = blk_mq_poll_nsecs(q, rq);
3816 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3819 * This will be replaced with the stats tracking code, using
3820 * 'avg_completion_time / 2' as the pre-sleep target.
3824 mode = HRTIMER_MODE_REL;
3825 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3826 hrtimer_set_expires(&hs.timer, kt);
3829 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3831 set_current_state(TASK_UNINTERRUPTIBLE);
3832 hrtimer_sleeper_start_expires(&hs, mode);
3835 hrtimer_cancel(&hs.timer);
3836 mode = HRTIMER_MODE_ABS;
3837 } while (hs.task && !signal_pending(current));
3839 __set_current_state(TASK_RUNNING);
3840 destroy_hrtimer_on_stack(&hs.timer);
3844 static bool blk_mq_poll_hybrid(struct request_queue *q,
3845 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3849 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3852 if (!blk_qc_t_is_internal(cookie))
3853 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3855 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3857 * With scheduling, if the request has completed, we'll
3858 * get a NULL return here, as we clear the sched tag when
3859 * that happens. The request still remains valid, like always,
3860 * so we should be safe with just the NULL check.
3866 return blk_mq_poll_hybrid_sleep(q, rq);
3870 * blk_poll - poll for IO completions
3872 * @cookie: cookie passed back at IO submission time
3873 * @spin: whether to spin for completions
3876 * Poll for completions on the passed in queue. Returns number of
3877 * completed entries found. If @spin is true, then blk_poll will continue
3878 * looping until at least one completion is found, unless the task is
3879 * otherwise marked running (or we need to reschedule).
3881 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3883 struct blk_mq_hw_ctx *hctx;
3886 if (!blk_qc_t_valid(cookie) ||
3887 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3891 blk_flush_plug_list(current->plug, false);
3893 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3896 * If we sleep, have the caller restart the poll loop to reset
3897 * the state. Like for the other success return cases, the
3898 * caller is responsible for checking if the IO completed. If
3899 * the IO isn't complete, we'll get called again and will go
3900 * straight to the busy poll loop. If specified not to spin,
3901 * we also should not sleep.
3903 if (spin && blk_mq_poll_hybrid(q, hctx, cookie))
3906 hctx->poll_considered++;
3908 state = current->state;
3912 hctx->poll_invoked++;
3914 ret = q->mq_ops->poll(hctx);
3916 hctx->poll_success++;
3917 __set_current_state(TASK_RUNNING);
3921 if (signal_pending_state(state, current))
3922 __set_current_state(TASK_RUNNING);
3924 if (current->state == TASK_RUNNING)
3926 if (ret < 0 || !spin)
3929 } while (!need_resched());
3931 __set_current_state(TASK_RUNNING);
3934 EXPORT_SYMBOL_GPL(blk_poll);
3936 unsigned int blk_mq_rq_cpu(struct request *rq)
3938 return rq->mq_ctx->cpu;
3940 EXPORT_SYMBOL(blk_mq_rq_cpu);
3942 static int __init blk_mq_init(void)
3946 for_each_possible_cpu(i)
3947 init_llist_head(&per_cpu(blk_cpu_done, i));
3948 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
3950 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
3951 "block/softirq:dead", NULL,
3952 blk_softirq_cpu_dead);
3953 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3954 blk_mq_hctx_notify_dead);
3955 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
3956 blk_mq_hctx_notify_online,
3957 blk_mq_hctx_notify_offline);
3960 subsys_initcall(blk_mq_init);