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 list_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 hd_struct *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->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, struct hd_struct *part)
117 struct mq_inflight mi = { .part = part };
119 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
121 return mi.inflight[0] + mi.inflight[1];
124 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
125 unsigned int inflight[2])
127 struct mq_inflight mi = { .part = part };
129 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
130 inflight[0] = mi.inflight[0];
131 inflight[1] = mi.inflight[1];
134 void blk_freeze_queue_start(struct request_queue *q)
136 mutex_lock(&q->mq_freeze_lock);
137 if (++q->mq_freeze_depth == 1) {
138 percpu_ref_kill(&q->q_usage_counter);
139 mutex_unlock(&q->mq_freeze_lock);
141 blk_mq_run_hw_queues(q, false);
143 mutex_unlock(&q->mq_freeze_lock);
146 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
148 void blk_mq_freeze_queue_wait(struct request_queue *q)
150 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
152 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
154 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
155 unsigned long timeout)
157 return wait_event_timeout(q->mq_freeze_wq,
158 percpu_ref_is_zero(&q->q_usage_counter),
161 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
164 * Guarantee no request is in use, so we can change any data structure of
165 * the queue afterward.
167 void blk_freeze_queue(struct request_queue *q)
170 * In the !blk_mq case we are only calling this to kill the
171 * q_usage_counter, otherwise this increases the freeze depth
172 * and waits for it to return to zero. For this reason there is
173 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
174 * exported to drivers as the only user for unfreeze is blk_mq.
176 blk_freeze_queue_start(q);
177 blk_mq_freeze_queue_wait(q);
180 void blk_mq_freeze_queue(struct request_queue *q)
183 * ...just an alias to keep freeze and unfreeze actions balanced
184 * in the blk_mq_* namespace
188 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
190 void blk_mq_unfreeze_queue(struct request_queue *q)
192 mutex_lock(&q->mq_freeze_lock);
193 q->mq_freeze_depth--;
194 WARN_ON_ONCE(q->mq_freeze_depth < 0);
195 if (!q->mq_freeze_depth) {
196 percpu_ref_resurrect(&q->q_usage_counter);
197 wake_up_all(&q->mq_freeze_wq);
199 mutex_unlock(&q->mq_freeze_lock);
201 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
204 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
205 * mpt3sas driver such that this function can be removed.
207 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
209 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
211 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
214 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
217 * Note: this function does not prevent that the struct request end_io()
218 * callback function is invoked. Once this function is returned, we make
219 * sure no dispatch can happen until the queue is unquiesced via
220 * blk_mq_unquiesce_queue().
222 void blk_mq_quiesce_queue(struct request_queue *q)
224 struct blk_mq_hw_ctx *hctx;
228 blk_mq_quiesce_queue_nowait(q);
230 queue_for_each_hw_ctx(q, hctx, i) {
231 if (hctx->flags & BLK_MQ_F_BLOCKING)
232 synchronize_srcu(hctx->srcu);
239 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
242 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
245 * This function recovers queue into the state before quiescing
246 * which is done by blk_mq_quiesce_queue.
248 void blk_mq_unquiesce_queue(struct request_queue *q)
250 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
252 /* dispatch requests which are inserted during quiescing */
253 blk_mq_run_hw_queues(q, true);
255 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
257 void blk_mq_wake_waiters(struct request_queue *q)
259 struct blk_mq_hw_ctx *hctx;
262 queue_for_each_hw_ctx(q, hctx, i)
263 if (blk_mq_hw_queue_mapped(hctx))
264 blk_mq_tag_wakeup_all(hctx->tags, true);
268 * Only need start/end time stamping if we have iostat or
269 * blk stats enabled, or using an IO scheduler.
271 static inline bool blk_mq_need_time_stamp(struct request *rq)
273 return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
276 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
277 unsigned int tag, u64 alloc_time_ns)
279 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
280 struct request *rq = tags->static_rqs[tag];
282 if (data->q->elevator) {
283 rq->tag = BLK_MQ_NO_TAG;
284 rq->internal_tag = tag;
287 rq->internal_tag = BLK_MQ_NO_TAG;
290 /* csd/requeue_work/fifo_time is initialized before use */
292 rq->mq_ctx = data->ctx;
293 rq->mq_hctx = data->hctx;
295 rq->cmd_flags = data->cmd_flags;
296 if (data->flags & BLK_MQ_REQ_PM)
297 rq->rq_flags |= RQF_PM;
298 if (blk_queue_io_stat(data->q))
299 rq->rq_flags |= RQF_IO_STAT;
300 INIT_LIST_HEAD(&rq->queuelist);
301 INIT_HLIST_NODE(&rq->hash);
302 RB_CLEAR_NODE(&rq->rb_node);
305 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
306 rq->alloc_time_ns = alloc_time_ns;
308 if (blk_mq_need_time_stamp(rq))
309 rq->start_time_ns = ktime_get_ns();
311 rq->start_time_ns = 0;
312 rq->io_start_time_ns = 0;
313 rq->stats_sectors = 0;
314 rq->nr_phys_segments = 0;
315 #if defined(CONFIG_BLK_DEV_INTEGRITY)
316 rq->nr_integrity_segments = 0;
318 blk_crypto_rq_set_defaults(rq);
319 /* tag was already set */
320 WRITE_ONCE(rq->deadline, 0);
325 rq->end_io_data = NULL;
327 data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++;
328 refcount_set(&rq->ref, 1);
330 if (!op_is_flush(data->cmd_flags)) {
331 struct elevator_queue *e = data->q->elevator;
334 if (e && e->type->ops.prepare_request) {
335 if (e->type->icq_cache)
336 blk_mq_sched_assign_ioc(rq);
338 e->type->ops.prepare_request(rq);
339 rq->rq_flags |= RQF_ELVPRIV;
343 data->hctx->queued++;
347 static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data)
349 struct request_queue *q = data->q;
350 struct elevator_queue *e = q->elevator;
351 u64 alloc_time_ns = 0;
354 /* alloc_time includes depth and tag waits */
355 if (blk_queue_rq_alloc_time(q))
356 alloc_time_ns = ktime_get_ns();
358 if (data->cmd_flags & REQ_NOWAIT)
359 data->flags |= BLK_MQ_REQ_NOWAIT;
363 * Flush requests are special and go directly to the
364 * dispatch list. Don't include reserved tags in the
365 * limiting, as it isn't useful.
367 if (!op_is_flush(data->cmd_flags) &&
368 e->type->ops.limit_depth &&
369 !(data->flags & BLK_MQ_REQ_RESERVED))
370 e->type->ops.limit_depth(data->cmd_flags, data);
374 data->ctx = blk_mq_get_ctx(q);
375 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
377 blk_mq_tag_busy(data->hctx);
380 * Waiting allocations only fail because of an inactive hctx. In that
381 * case just retry the hctx assignment and tag allocation as CPU hotplug
382 * should have migrated us to an online CPU by now.
384 tag = blk_mq_get_tag(data);
385 if (tag == BLK_MQ_NO_TAG) {
386 if (data->flags & BLK_MQ_REQ_NOWAIT)
390 * Give up the CPU and sleep for a random short time to ensure
391 * that thread using a realtime scheduling class are migrated
392 * off the CPU, and thus off the hctx that is going away.
397 return blk_mq_rq_ctx_init(data, tag, alloc_time_ns);
400 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
401 blk_mq_req_flags_t flags)
403 struct blk_mq_alloc_data data = {
411 ret = blk_queue_enter(q, flags);
415 rq = __blk_mq_alloc_request(&data);
419 rq->__sector = (sector_t) -1;
420 rq->bio = rq->biotail = NULL;
424 return ERR_PTR(-EWOULDBLOCK);
426 EXPORT_SYMBOL(blk_mq_alloc_request);
428 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
429 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
431 struct blk_mq_alloc_data data = {
436 u64 alloc_time_ns = 0;
441 /* alloc_time includes depth and tag waits */
442 if (blk_queue_rq_alloc_time(q))
443 alloc_time_ns = ktime_get_ns();
446 * If the tag allocator sleeps we could get an allocation for a
447 * different hardware context. No need to complicate the low level
448 * allocator for this for the rare use case of a command tied to
451 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
452 WARN_ON_ONCE(!(flags & 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 if (cpu >= nr_cpu_ids)
473 data.ctx = __blk_mq_get_ctx(q, cpu);
476 blk_mq_tag_busy(data.hctx);
479 tag = blk_mq_get_tag(&data);
480 if (tag == BLK_MQ_NO_TAG)
482 return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns);
488 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
490 static void __blk_mq_free_request(struct request *rq)
492 struct request_queue *q = rq->q;
493 struct blk_mq_ctx *ctx = rq->mq_ctx;
494 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
495 const int sched_tag = rq->internal_tag;
497 blk_crypto_free_request(rq);
498 blk_pm_mark_last_busy(rq);
500 if (rq->tag != BLK_MQ_NO_TAG)
501 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
502 if (sched_tag != BLK_MQ_NO_TAG)
503 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
504 blk_mq_sched_restart(hctx);
508 void blk_mq_free_request(struct request *rq)
510 struct request_queue *q = rq->q;
511 struct elevator_queue *e = q->elevator;
512 struct blk_mq_ctx *ctx = rq->mq_ctx;
513 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
515 if (rq->rq_flags & RQF_ELVPRIV) {
516 if (e && e->type->ops.finish_request)
517 e->type->ops.finish_request(rq);
519 put_io_context(rq->elv.icq->ioc);
524 ctx->rq_completed[rq_is_sync(rq)]++;
525 if (rq->rq_flags & RQF_MQ_INFLIGHT)
526 __blk_mq_dec_active_requests(hctx);
528 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
529 laptop_io_completion(q->backing_dev_info);
533 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
534 if (refcount_dec_and_test(&rq->ref))
535 __blk_mq_free_request(rq);
537 EXPORT_SYMBOL_GPL(blk_mq_free_request);
539 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
543 if (blk_mq_need_time_stamp(rq))
544 now = ktime_get_ns();
546 if (rq->rq_flags & RQF_STATS) {
547 blk_mq_poll_stats_start(rq->q);
548 blk_stat_add(rq, now);
551 blk_mq_sched_completed_request(rq, now);
553 blk_account_io_done(rq, now);
556 rq_qos_done(rq->q, rq);
557 rq->end_io(rq, error);
559 blk_mq_free_request(rq);
562 EXPORT_SYMBOL(__blk_mq_end_request);
564 void blk_mq_end_request(struct request *rq, blk_status_t error)
566 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
568 __blk_mq_end_request(rq, error);
570 EXPORT_SYMBOL(blk_mq_end_request);
573 * Softirq action handler - move entries to local list and loop over them
574 * while passing them to the queue registered handler.
576 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
578 struct list_head *cpu_list, local_list;
581 cpu_list = this_cpu_ptr(&blk_cpu_done);
582 list_replace_init(cpu_list, &local_list);
585 while (!list_empty(&local_list)) {
588 rq = list_entry(local_list.next, struct request, ipi_list);
589 list_del_init(&rq->ipi_list);
590 rq->q->mq_ops->complete(rq);
594 static void blk_mq_trigger_softirq(struct request *rq)
596 struct list_head *list;
599 local_irq_save(flags);
600 list = this_cpu_ptr(&blk_cpu_done);
601 list_add_tail(&rq->ipi_list, list);
604 * If the list only contains our just added request, signal a raise of
605 * the softirq. If there are already entries there, someone already
606 * raised the irq but it hasn't run yet.
608 if (list->next == &rq->ipi_list)
609 raise_softirq_irqoff(BLOCK_SOFTIRQ);
610 local_irq_restore(flags);
613 static int blk_softirq_cpu_dead(unsigned int cpu)
616 * If a CPU goes away, splice its entries to the current CPU
617 * and trigger a run of the softirq
620 list_splice_init(&per_cpu(blk_cpu_done, cpu),
621 this_cpu_ptr(&blk_cpu_done));
622 raise_softirq_irqoff(BLOCK_SOFTIRQ);
629 static void __blk_mq_complete_request_remote(void *data)
631 struct request *rq = data;
634 * For most of single queue controllers, there is only one irq vector
635 * for handling I/O completion, and the only irq's affinity is set
636 * to all possible CPUs. On most of ARCHs, this affinity means the irq
637 * is handled on one specific CPU.
639 * So complete I/O requests in softirq context in case of single queue
640 * devices to avoid degrading I/O performance due to irqsoff latency.
642 if (rq->q->nr_hw_queues == 1)
643 blk_mq_trigger_softirq(rq);
645 rq->q->mq_ops->complete(rq);
648 static inline bool blk_mq_complete_need_ipi(struct request *rq)
650 int cpu = raw_smp_processor_id();
652 if (!IS_ENABLED(CONFIG_SMP) ||
653 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
656 /* same CPU or cache domain? Complete locally */
657 if (cpu == rq->mq_ctx->cpu ||
658 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
659 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
662 /* don't try to IPI to an offline CPU */
663 return cpu_online(rq->mq_ctx->cpu);
666 bool blk_mq_complete_request_remote(struct request *rq)
668 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
671 * For a polled request, always complete locallly, it's pointless
672 * to redirect the completion.
674 if (rq->cmd_flags & REQ_HIPRI)
677 if (blk_mq_complete_need_ipi(rq)) {
678 rq->csd.func = __blk_mq_complete_request_remote;
681 smp_call_function_single_async(rq->mq_ctx->cpu, &rq->csd);
683 if (rq->q->nr_hw_queues > 1)
685 blk_mq_trigger_softirq(rq);
690 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
693 * blk_mq_complete_request - end I/O on a request
694 * @rq: the request being processed
697 * Complete a request by scheduling the ->complete_rq operation.
699 void blk_mq_complete_request(struct request *rq)
701 if (!blk_mq_complete_request_remote(rq))
702 rq->q->mq_ops->complete(rq);
704 EXPORT_SYMBOL(blk_mq_complete_request);
706 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
707 __releases(hctx->srcu)
709 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
712 srcu_read_unlock(hctx->srcu, srcu_idx);
715 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
716 __acquires(hctx->srcu)
718 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
719 /* shut up gcc false positive */
723 *srcu_idx = srcu_read_lock(hctx->srcu);
727 * blk_mq_start_request - Start processing a request
728 * @rq: Pointer to request to be started
730 * Function used by device drivers to notify the block layer that a request
731 * is going to be processed now, so blk layer can do proper initializations
732 * such as starting the timeout timer.
734 void blk_mq_start_request(struct request *rq)
736 struct request_queue *q = rq->q;
738 trace_block_rq_issue(rq);
740 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
741 rq->io_start_time_ns = ktime_get_ns();
742 rq->stats_sectors = blk_rq_sectors(rq);
743 rq->rq_flags |= RQF_STATS;
747 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
750 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
752 #ifdef CONFIG_BLK_DEV_INTEGRITY
753 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
754 q->integrity.profile->prepare_fn(rq);
757 EXPORT_SYMBOL(blk_mq_start_request);
759 static void __blk_mq_requeue_request(struct request *rq)
761 struct request_queue *q = rq->q;
763 blk_mq_put_driver_tag(rq);
765 trace_block_rq_requeue(rq);
766 rq_qos_requeue(q, rq);
768 if (blk_mq_request_started(rq)) {
769 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
770 rq->rq_flags &= ~RQF_TIMED_OUT;
774 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
776 __blk_mq_requeue_request(rq);
778 /* this request will be re-inserted to io scheduler queue */
779 blk_mq_sched_requeue_request(rq);
781 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
783 EXPORT_SYMBOL(blk_mq_requeue_request);
785 static void blk_mq_requeue_work(struct work_struct *work)
787 struct request_queue *q =
788 container_of(work, struct request_queue, requeue_work.work);
790 struct request *rq, *next;
792 spin_lock_irq(&q->requeue_lock);
793 list_splice_init(&q->requeue_list, &rq_list);
794 spin_unlock_irq(&q->requeue_lock);
796 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
797 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
800 rq->rq_flags &= ~RQF_SOFTBARRIER;
801 list_del_init(&rq->queuelist);
803 * If RQF_DONTPREP, rq has contained some driver specific
804 * data, so insert it to hctx dispatch list to avoid any
807 if (rq->rq_flags & RQF_DONTPREP)
808 blk_mq_request_bypass_insert(rq, false, false);
810 blk_mq_sched_insert_request(rq, true, false, false);
813 while (!list_empty(&rq_list)) {
814 rq = list_entry(rq_list.next, struct request, queuelist);
815 list_del_init(&rq->queuelist);
816 blk_mq_sched_insert_request(rq, false, false, false);
819 blk_mq_run_hw_queues(q, false);
822 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
823 bool kick_requeue_list)
825 struct request_queue *q = rq->q;
829 * We abuse this flag that is otherwise used by the I/O scheduler to
830 * request head insertion from the workqueue.
832 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
834 spin_lock_irqsave(&q->requeue_lock, flags);
836 rq->rq_flags |= RQF_SOFTBARRIER;
837 list_add(&rq->queuelist, &q->requeue_list);
839 list_add_tail(&rq->queuelist, &q->requeue_list);
841 spin_unlock_irqrestore(&q->requeue_lock, flags);
843 if (kick_requeue_list)
844 blk_mq_kick_requeue_list(q);
847 void blk_mq_kick_requeue_list(struct request_queue *q)
849 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
851 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
853 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
856 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
857 msecs_to_jiffies(msecs));
859 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
861 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
863 if (tag < tags->nr_tags) {
864 prefetch(tags->rqs[tag]);
865 return tags->rqs[tag];
870 EXPORT_SYMBOL(blk_mq_tag_to_rq);
872 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
873 void *priv, bool reserved)
876 * If we find a request that isn't idle and the queue matches,
877 * we know the queue is busy. Return false to stop the iteration.
879 if (blk_mq_request_started(rq) && rq->q == hctx->queue) {
889 bool blk_mq_queue_inflight(struct request_queue *q)
893 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
896 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
898 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
900 req->rq_flags |= RQF_TIMED_OUT;
901 if (req->q->mq_ops->timeout) {
902 enum blk_eh_timer_return ret;
904 ret = req->q->mq_ops->timeout(req, reserved);
905 if (ret == BLK_EH_DONE)
907 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
913 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
915 unsigned long deadline;
917 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
919 if (rq->rq_flags & RQF_TIMED_OUT)
922 deadline = READ_ONCE(rq->deadline);
923 if (time_after_eq(jiffies, deadline))
928 else if (time_after(*next, deadline))
933 void blk_mq_put_rq_ref(struct request *rq)
937 else if (refcount_dec_and_test(&rq->ref))
938 __blk_mq_free_request(rq);
941 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
942 struct request *rq, void *priv, bool reserved)
944 unsigned long *next = priv;
947 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
948 * be reallocated underneath the timeout handler's processing, then
949 * the expire check is reliable. If the request is not expired, then
950 * it was completed and reallocated as a new request after returning
951 * from blk_mq_check_expired().
953 if (blk_mq_req_expired(rq, next))
954 blk_mq_rq_timed_out(rq, reserved);
958 static void blk_mq_timeout_work(struct work_struct *work)
960 struct request_queue *q =
961 container_of(work, struct request_queue, timeout_work);
962 unsigned long next = 0;
963 struct blk_mq_hw_ctx *hctx;
966 /* A deadlock might occur if a request is stuck requiring a
967 * timeout at the same time a queue freeze is waiting
968 * completion, since the timeout code would not be able to
969 * acquire the queue reference here.
971 * That's why we don't use blk_queue_enter here; instead, we use
972 * percpu_ref_tryget directly, because we need to be able to
973 * obtain a reference even in the short window between the queue
974 * starting to freeze, by dropping the first reference in
975 * blk_freeze_queue_start, and the moment the last request is
976 * consumed, marked by the instant q_usage_counter reaches
979 if (!percpu_ref_tryget(&q->q_usage_counter))
982 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
985 mod_timer(&q->timeout, next);
988 * Request timeouts are handled as a forward rolling timer. If
989 * we end up here it means that no requests are pending and
990 * also that no request has been pending for a while. Mark
993 queue_for_each_hw_ctx(q, hctx, i) {
994 /* the hctx may be unmapped, so check it here */
995 if (blk_mq_hw_queue_mapped(hctx))
996 blk_mq_tag_idle(hctx);
1002 struct flush_busy_ctx_data {
1003 struct blk_mq_hw_ctx *hctx;
1004 struct list_head *list;
1007 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1009 struct flush_busy_ctx_data *flush_data = data;
1010 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1011 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1012 enum hctx_type type = hctx->type;
1014 spin_lock(&ctx->lock);
1015 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1016 sbitmap_clear_bit(sb, bitnr);
1017 spin_unlock(&ctx->lock);
1022 * Process software queues that have been marked busy, splicing them
1023 * to the for-dispatch
1025 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1027 struct flush_busy_ctx_data data = {
1032 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1034 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1036 struct dispatch_rq_data {
1037 struct blk_mq_hw_ctx *hctx;
1041 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1044 struct dispatch_rq_data *dispatch_data = data;
1045 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1046 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1047 enum hctx_type type = hctx->type;
1049 spin_lock(&ctx->lock);
1050 if (!list_empty(&ctx->rq_lists[type])) {
1051 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1052 list_del_init(&dispatch_data->rq->queuelist);
1053 if (list_empty(&ctx->rq_lists[type]))
1054 sbitmap_clear_bit(sb, bitnr);
1056 spin_unlock(&ctx->lock);
1058 return !dispatch_data->rq;
1061 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1062 struct blk_mq_ctx *start)
1064 unsigned off = start ? start->index_hw[hctx->type] : 0;
1065 struct dispatch_rq_data data = {
1070 __sbitmap_for_each_set(&hctx->ctx_map, off,
1071 dispatch_rq_from_ctx, &data);
1076 static inline unsigned int queued_to_index(unsigned int queued)
1081 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1084 static bool __blk_mq_get_driver_tag(struct request *rq)
1086 struct sbitmap_queue *bt = rq->mq_hctx->tags->bitmap_tags;
1087 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1090 blk_mq_tag_busy(rq->mq_hctx);
1092 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1093 bt = rq->mq_hctx->tags->breserved_tags;
1096 if (!hctx_may_queue(rq->mq_hctx, bt))
1100 tag = __sbitmap_queue_get(bt);
1101 if (tag == BLK_MQ_NO_TAG)
1104 rq->tag = tag + tag_offset;
1108 static bool blk_mq_get_driver_tag(struct request *rq)
1110 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1112 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq))
1115 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1116 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1117 rq->rq_flags |= RQF_MQ_INFLIGHT;
1118 __blk_mq_inc_active_requests(hctx);
1120 hctx->tags->rqs[rq->tag] = rq;
1124 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1125 int flags, void *key)
1127 struct blk_mq_hw_ctx *hctx;
1129 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1131 spin_lock(&hctx->dispatch_wait_lock);
1132 if (!list_empty(&wait->entry)) {
1133 struct sbitmap_queue *sbq;
1135 list_del_init(&wait->entry);
1136 sbq = hctx->tags->bitmap_tags;
1137 atomic_dec(&sbq->ws_active);
1139 spin_unlock(&hctx->dispatch_wait_lock);
1141 blk_mq_run_hw_queue(hctx, true);
1146 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1147 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1148 * restart. For both cases, take care to check the condition again after
1149 * marking us as waiting.
1151 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1154 struct sbitmap_queue *sbq = hctx->tags->bitmap_tags;
1155 struct wait_queue_head *wq;
1156 wait_queue_entry_t *wait;
1159 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1160 blk_mq_sched_mark_restart_hctx(hctx);
1163 * It's possible that a tag was freed in the window between the
1164 * allocation failure and adding the hardware queue to the wait
1167 * Don't clear RESTART here, someone else could have set it.
1168 * At most this will cost an extra queue run.
1170 return blk_mq_get_driver_tag(rq);
1173 wait = &hctx->dispatch_wait;
1174 if (!list_empty_careful(&wait->entry))
1177 wq = &bt_wait_ptr(sbq, hctx)->wait;
1179 spin_lock_irq(&wq->lock);
1180 spin_lock(&hctx->dispatch_wait_lock);
1181 if (!list_empty(&wait->entry)) {
1182 spin_unlock(&hctx->dispatch_wait_lock);
1183 spin_unlock_irq(&wq->lock);
1187 atomic_inc(&sbq->ws_active);
1188 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1189 __add_wait_queue(wq, wait);
1192 * Add one explicit barrier since blk_mq_get_driver_tag() may
1193 * not imply barrier in case of failure.
1195 * Order adding us to wait queue and allocating driver tag.
1197 * The pair is the one implied in sbitmap_queue_wake_up() which
1198 * orders clearing sbitmap tag bits and waitqueue_active() in
1199 * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless
1201 * Otherwise, re-order of adding wait queue and getting driver tag
1202 * may cause __sbitmap_queue_wake_up() to wake up nothing because
1203 * the waitqueue_active() may not observe us in wait queue.
1208 * It's possible that a tag was freed in the window between the
1209 * allocation failure and adding the hardware queue to the wait
1212 ret = blk_mq_get_driver_tag(rq);
1214 spin_unlock(&hctx->dispatch_wait_lock);
1215 spin_unlock_irq(&wq->lock);
1220 * We got a tag, remove ourselves from the wait queue to ensure
1221 * someone else gets the wakeup.
1223 list_del_init(&wait->entry);
1224 atomic_dec(&sbq->ws_active);
1225 spin_unlock(&hctx->dispatch_wait_lock);
1226 spin_unlock_irq(&wq->lock);
1231 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1232 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1234 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1235 * - EWMA is one simple way to compute running average value
1236 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1237 * - take 4 as factor for avoiding to get too small(0) result, and this
1238 * factor doesn't matter because EWMA decreases exponentially
1240 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1244 ewma = hctx->dispatch_busy;
1249 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1251 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1252 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1254 hctx->dispatch_busy = ewma;
1257 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1259 static void blk_mq_handle_dev_resource(struct request *rq,
1260 struct list_head *list)
1262 struct request *next =
1263 list_first_entry_or_null(list, struct request, queuelist);
1266 * If an I/O scheduler has been configured and we got a driver tag for
1267 * the next request already, free it.
1270 blk_mq_put_driver_tag(next);
1272 list_add(&rq->queuelist, list);
1273 __blk_mq_requeue_request(rq);
1276 static void blk_mq_handle_zone_resource(struct request *rq,
1277 struct list_head *zone_list)
1280 * If we end up here it is because we cannot dispatch a request to a
1281 * specific zone due to LLD level zone-write locking or other zone
1282 * related resource not being available. In this case, set the request
1283 * aside in zone_list for retrying it later.
1285 list_add(&rq->queuelist, zone_list);
1286 __blk_mq_requeue_request(rq);
1289 enum prep_dispatch {
1291 PREP_DISPATCH_NO_TAG,
1292 PREP_DISPATCH_NO_BUDGET,
1295 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1298 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1300 if (need_budget && !blk_mq_get_dispatch_budget(rq->q)) {
1301 blk_mq_put_driver_tag(rq);
1302 return PREP_DISPATCH_NO_BUDGET;
1305 if (!blk_mq_get_driver_tag(rq)) {
1307 * The initial allocation attempt failed, so we need to
1308 * rerun the hardware queue when a tag is freed. The
1309 * waitqueue takes care of that. If the queue is run
1310 * before we add this entry back on the dispatch list,
1311 * we'll re-run it below.
1313 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1315 * All budgets not got from this function will be put
1316 * together during handling partial dispatch
1319 blk_mq_put_dispatch_budget(rq->q);
1320 return PREP_DISPATCH_NO_TAG;
1324 return PREP_DISPATCH_OK;
1327 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1328 static void blk_mq_release_budgets(struct request_queue *q,
1329 unsigned int nr_budgets)
1333 for (i = 0; i < nr_budgets; i++)
1334 blk_mq_put_dispatch_budget(q);
1338 * Returns true if we did some work AND can potentially do more.
1340 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1341 unsigned int nr_budgets)
1343 enum prep_dispatch prep;
1344 struct request_queue *q = hctx->queue;
1345 struct request *rq, *nxt;
1347 blk_status_t ret = BLK_STS_OK;
1348 LIST_HEAD(zone_list);
1349 bool needs_resource = false;
1351 if (list_empty(list))
1355 * Now process all the entries, sending them to the driver.
1357 errors = queued = 0;
1359 struct blk_mq_queue_data bd;
1361 rq = list_first_entry(list, struct request, queuelist);
1363 WARN_ON_ONCE(hctx != rq->mq_hctx);
1364 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1365 if (prep != PREP_DISPATCH_OK)
1368 list_del_init(&rq->queuelist);
1373 * Flag last if we have no more requests, or if we have more
1374 * but can't assign a driver tag to it.
1376 if (list_empty(list))
1379 nxt = list_first_entry(list, struct request, queuelist);
1380 bd.last = !blk_mq_get_driver_tag(nxt);
1384 * once the request is queued to lld, no need to cover the
1389 ret = q->mq_ops->queue_rq(hctx, &bd);
1394 case BLK_STS_RESOURCE:
1395 needs_resource = true;
1397 case BLK_STS_DEV_RESOURCE:
1398 blk_mq_handle_dev_resource(rq, list);
1400 case BLK_STS_ZONE_RESOURCE:
1402 * Move the request to zone_list and keep going through
1403 * the dispatch list to find more requests the drive can
1406 blk_mq_handle_zone_resource(rq, &zone_list);
1407 needs_resource = true;
1411 blk_mq_end_request(rq, BLK_STS_IOERR);
1413 } while (!list_empty(list));
1415 if (!list_empty(&zone_list))
1416 list_splice_tail_init(&zone_list, list);
1418 hctx->dispatched[queued_to_index(queued)]++;
1420 /* If we didn't flush the entire list, we could have told the driver
1421 * there was more coming, but that turned out to be a lie.
1423 if ((!list_empty(list) || errors || needs_resource ||
1424 ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued)
1425 q->mq_ops->commit_rqs(hctx);
1427 * Any items that need requeuing? Stuff them into hctx->dispatch,
1428 * that is where we will continue on next queue run.
1430 if (!list_empty(list)) {
1432 /* For non-shared tags, the RESTART check will suffice */
1433 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1434 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1436 blk_mq_release_budgets(q, nr_budgets);
1438 spin_lock(&hctx->lock);
1439 list_splice_tail_init(list, &hctx->dispatch);
1440 spin_unlock(&hctx->lock);
1443 * Order adding requests to hctx->dispatch and checking
1444 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1445 * in blk_mq_sched_restart(). Avoid restart code path to
1446 * miss the new added requests to hctx->dispatch, meantime
1447 * SCHED_RESTART is observed here.
1452 * If SCHED_RESTART was set by the caller of this function and
1453 * it is no longer set that means that it was cleared by another
1454 * thread and hence that a queue rerun is needed.
1456 * If 'no_tag' is set, that means that we failed getting
1457 * a driver tag with an I/O scheduler attached. If our dispatch
1458 * waitqueue is no longer active, ensure that we run the queue
1459 * AFTER adding our entries back to the list.
1461 * If no I/O scheduler has been configured it is possible that
1462 * the hardware queue got stopped and restarted before requests
1463 * were pushed back onto the dispatch list. Rerun the queue to
1464 * avoid starvation. Notes:
1465 * - blk_mq_run_hw_queue() checks whether or not a queue has
1466 * been stopped before rerunning a queue.
1467 * - Some but not all block drivers stop a queue before
1468 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1471 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1472 * bit is set, run queue after a delay to avoid IO stalls
1473 * that could otherwise occur if the queue is idle. We'll do
1474 * similar if we couldn't get budget or couldn't lock a zone
1475 * and SCHED_RESTART is set.
1477 needs_restart = blk_mq_sched_needs_restart(hctx);
1478 if (prep == PREP_DISPATCH_NO_BUDGET)
1479 needs_resource = true;
1480 if (!needs_restart ||
1481 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1482 blk_mq_run_hw_queue(hctx, true);
1483 else if (needs_restart && needs_resource)
1484 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1486 blk_mq_update_dispatch_busy(hctx, true);
1489 blk_mq_update_dispatch_busy(hctx, false);
1491 return (queued + errors) != 0;
1495 * __blk_mq_run_hw_queue - Run a hardware queue.
1496 * @hctx: Pointer to the hardware queue to run.
1498 * Send pending requests to the hardware.
1500 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1505 * We should be running this queue from one of the CPUs that
1508 * There are at least two related races now between setting
1509 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1510 * __blk_mq_run_hw_queue():
1512 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1513 * but later it becomes online, then this warning is harmless
1516 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1517 * but later it becomes offline, then the warning can't be
1518 * triggered, and we depend on blk-mq timeout handler to
1519 * handle dispatched requests to this hctx
1521 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1522 cpu_online(hctx->next_cpu)) {
1523 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1524 raw_smp_processor_id(),
1525 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1530 * We can't run the queue inline with ints disabled. Ensure that
1531 * we catch bad users of this early.
1533 WARN_ON_ONCE(in_interrupt());
1535 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1537 hctx_lock(hctx, &srcu_idx);
1538 blk_mq_sched_dispatch_requests(hctx);
1539 hctx_unlock(hctx, srcu_idx);
1542 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1544 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1546 if (cpu >= nr_cpu_ids)
1547 cpu = cpumask_first(hctx->cpumask);
1552 * It'd be great if the workqueue API had a way to pass
1553 * in a mask and had some smarts for more clever placement.
1554 * For now we just round-robin here, switching for every
1555 * BLK_MQ_CPU_WORK_BATCH queued items.
1557 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1560 int next_cpu = hctx->next_cpu;
1562 if (hctx->queue->nr_hw_queues == 1)
1563 return WORK_CPU_UNBOUND;
1565 if (--hctx->next_cpu_batch <= 0) {
1567 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1569 if (next_cpu >= nr_cpu_ids)
1570 next_cpu = blk_mq_first_mapped_cpu(hctx);
1571 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1575 * Do unbound schedule if we can't find a online CPU for this hctx,
1576 * and it should only happen in the path of handling CPU DEAD.
1578 if (!cpu_online(next_cpu)) {
1585 * Make sure to re-select CPU next time once after CPUs
1586 * in hctx->cpumask become online again.
1588 hctx->next_cpu = next_cpu;
1589 hctx->next_cpu_batch = 1;
1590 return WORK_CPU_UNBOUND;
1593 hctx->next_cpu = next_cpu;
1598 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1599 * @hctx: Pointer to the hardware queue to run.
1600 * @async: If we want to run the queue asynchronously.
1601 * @msecs: Microseconds of delay to wait before running the queue.
1603 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1604 * with a delay of @msecs.
1606 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1607 unsigned long msecs)
1609 if (unlikely(blk_mq_hctx_stopped(hctx)))
1612 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1613 int cpu = get_cpu();
1614 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1615 __blk_mq_run_hw_queue(hctx);
1623 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1624 msecs_to_jiffies(msecs));
1628 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1629 * @hctx: Pointer to the hardware queue to run.
1630 * @msecs: Microseconds of delay to wait before running the queue.
1632 * Run a hardware queue asynchronously with a delay of @msecs.
1634 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1636 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1638 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1641 * blk_mq_run_hw_queue - Start to run a hardware queue.
1642 * @hctx: Pointer to the hardware queue to run.
1643 * @async: If we want to run the queue asynchronously.
1645 * Check if the request queue is not in a quiesced state and if there are
1646 * pending requests to be sent. If this is true, run the queue to send requests
1649 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1655 * When queue is quiesced, we may be switching io scheduler, or
1656 * updating nr_hw_queues, or other things, and we can't run queue
1657 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1659 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1662 hctx_lock(hctx, &srcu_idx);
1663 need_run = !blk_queue_quiesced(hctx->queue) &&
1664 blk_mq_hctx_has_pending(hctx);
1665 hctx_unlock(hctx, srcu_idx);
1668 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1670 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1673 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
1674 * @q: Pointer to the request queue to run.
1675 * @async: If we want to run the queue asynchronously.
1677 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1679 struct blk_mq_hw_ctx *hctx;
1682 queue_for_each_hw_ctx(q, hctx, i) {
1683 if (blk_mq_hctx_stopped(hctx))
1686 blk_mq_run_hw_queue(hctx, async);
1689 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1692 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1693 * @q: Pointer to the request queue to run.
1694 * @msecs: Microseconds of delay to wait before running the queues.
1696 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1698 struct blk_mq_hw_ctx *hctx;
1701 queue_for_each_hw_ctx(q, hctx, i) {
1702 if (blk_mq_hctx_stopped(hctx))
1705 blk_mq_delay_run_hw_queue(hctx, msecs);
1708 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1711 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1712 * @q: request queue.
1714 * The caller is responsible for serializing this function against
1715 * blk_mq_{start,stop}_hw_queue().
1717 bool blk_mq_queue_stopped(struct request_queue *q)
1719 struct blk_mq_hw_ctx *hctx;
1722 queue_for_each_hw_ctx(q, hctx, i)
1723 if (blk_mq_hctx_stopped(hctx))
1728 EXPORT_SYMBOL(blk_mq_queue_stopped);
1731 * This function is often used for pausing .queue_rq() by driver when
1732 * there isn't enough resource or some conditions aren't satisfied, and
1733 * BLK_STS_RESOURCE is usually returned.
1735 * We do not guarantee that dispatch can be drained or blocked
1736 * after blk_mq_stop_hw_queue() returns. Please use
1737 * blk_mq_quiesce_queue() for that requirement.
1739 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1741 cancel_delayed_work(&hctx->run_work);
1743 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1745 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1748 * This function is often used for pausing .queue_rq() by driver when
1749 * there isn't enough resource or some conditions aren't satisfied, and
1750 * BLK_STS_RESOURCE is usually returned.
1752 * We do not guarantee that dispatch can be drained or blocked
1753 * after blk_mq_stop_hw_queues() returns. Please use
1754 * blk_mq_quiesce_queue() for that requirement.
1756 void blk_mq_stop_hw_queues(struct request_queue *q)
1758 struct blk_mq_hw_ctx *hctx;
1761 queue_for_each_hw_ctx(q, hctx, i)
1762 blk_mq_stop_hw_queue(hctx);
1764 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1766 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1768 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1770 blk_mq_run_hw_queue(hctx, false);
1772 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1774 void blk_mq_start_hw_queues(struct request_queue *q)
1776 struct blk_mq_hw_ctx *hctx;
1779 queue_for_each_hw_ctx(q, hctx, i)
1780 blk_mq_start_hw_queue(hctx);
1782 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1784 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1786 if (!blk_mq_hctx_stopped(hctx))
1789 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1790 blk_mq_run_hw_queue(hctx, async);
1792 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1794 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1796 struct blk_mq_hw_ctx *hctx;
1799 queue_for_each_hw_ctx(q, hctx, i)
1800 blk_mq_start_stopped_hw_queue(hctx, async);
1802 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1804 static void blk_mq_run_work_fn(struct work_struct *work)
1806 struct blk_mq_hw_ctx *hctx;
1808 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1811 * If we are stopped, don't run the queue.
1813 if (blk_mq_hctx_stopped(hctx))
1816 __blk_mq_run_hw_queue(hctx);
1819 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1823 struct blk_mq_ctx *ctx = rq->mq_ctx;
1824 enum hctx_type type = hctx->type;
1826 lockdep_assert_held(&ctx->lock);
1828 trace_block_rq_insert(rq);
1831 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1833 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1836 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1839 struct blk_mq_ctx *ctx = rq->mq_ctx;
1841 lockdep_assert_held(&ctx->lock);
1843 __blk_mq_insert_req_list(hctx, rq, at_head);
1844 blk_mq_hctx_mark_pending(hctx, ctx);
1848 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1849 * @rq: Pointer to request to be inserted.
1850 * @at_head: true if the request should be inserted at the head of the list.
1851 * @run_queue: If we should run the hardware queue after inserting the request.
1853 * Should only be used carefully, when the caller knows we want to
1854 * bypass a potential IO scheduler on the target device.
1856 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1859 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1861 spin_lock(&hctx->lock);
1863 list_add(&rq->queuelist, &hctx->dispatch);
1865 list_add_tail(&rq->queuelist, &hctx->dispatch);
1866 spin_unlock(&hctx->lock);
1869 blk_mq_run_hw_queue(hctx, false);
1872 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1873 struct list_head *list)
1877 enum hctx_type type = hctx->type;
1880 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1883 list_for_each_entry(rq, list, queuelist) {
1884 BUG_ON(rq->mq_ctx != ctx);
1885 trace_block_rq_insert(rq);
1888 spin_lock(&ctx->lock);
1889 list_splice_tail_init(list, &ctx->rq_lists[type]);
1890 blk_mq_hctx_mark_pending(hctx, ctx);
1891 spin_unlock(&ctx->lock);
1894 static int plug_rq_cmp(void *priv, const struct list_head *a,
1895 const struct list_head *b)
1897 struct request *rqa = container_of(a, struct request, queuelist);
1898 struct request *rqb = container_of(b, struct request, queuelist);
1900 if (rqa->mq_ctx != rqb->mq_ctx)
1901 return rqa->mq_ctx > rqb->mq_ctx;
1902 if (rqa->mq_hctx != rqb->mq_hctx)
1903 return rqa->mq_hctx > rqb->mq_hctx;
1905 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1908 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1912 if (list_empty(&plug->mq_list))
1914 list_splice_init(&plug->mq_list, &list);
1916 if (plug->rq_count > 2 && plug->multiple_queues)
1917 list_sort(NULL, &list, plug_rq_cmp);
1922 struct list_head rq_list;
1923 struct request *rq, *head_rq = list_entry_rq(list.next);
1924 struct list_head *pos = &head_rq->queuelist; /* skip first */
1925 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1926 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1927 unsigned int depth = 1;
1929 list_for_each_continue(pos, &list) {
1930 rq = list_entry_rq(pos);
1932 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1937 list_cut_before(&rq_list, &list, pos);
1938 trace_block_unplug(head_rq->q, depth, !from_schedule);
1939 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1941 } while(!list_empty(&list));
1944 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1945 unsigned int nr_segs)
1949 if (bio->bi_opf & REQ_RAHEAD)
1950 rq->cmd_flags |= REQ_FAILFAST_MASK;
1952 rq->__sector = bio->bi_iter.bi_sector;
1953 rq->write_hint = bio->bi_write_hint;
1954 blk_rq_bio_prep(rq, bio, nr_segs);
1956 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
1957 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1960 blk_account_io_start(rq);
1963 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1965 blk_qc_t *cookie, bool last)
1967 struct request_queue *q = rq->q;
1968 struct blk_mq_queue_data bd = {
1972 blk_qc_t new_cookie;
1975 new_cookie = request_to_qc_t(hctx, rq);
1978 * For OK queue, we are done. For error, caller may kill it.
1979 * Any other error (busy), just add it to our list as we
1980 * previously would have done.
1982 ret = q->mq_ops->queue_rq(hctx, &bd);
1985 blk_mq_update_dispatch_busy(hctx, false);
1986 *cookie = new_cookie;
1988 case BLK_STS_RESOURCE:
1989 case BLK_STS_DEV_RESOURCE:
1990 blk_mq_update_dispatch_busy(hctx, true);
1991 __blk_mq_requeue_request(rq);
1994 blk_mq_update_dispatch_busy(hctx, false);
1995 *cookie = BLK_QC_T_NONE;
2002 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2005 bool bypass_insert, bool last)
2007 struct request_queue *q = rq->q;
2008 bool run_queue = true;
2011 * RCU or SRCU read lock is needed before checking quiesced flag.
2013 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2014 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2015 * and avoid driver to try to dispatch again.
2017 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2019 bypass_insert = false;
2023 if (q->elevator && !bypass_insert)
2026 if (!blk_mq_get_dispatch_budget(q))
2029 if (!blk_mq_get_driver_tag(rq)) {
2030 blk_mq_put_dispatch_budget(q);
2034 return __blk_mq_issue_directly(hctx, rq, cookie, last);
2037 return BLK_STS_RESOURCE;
2039 blk_mq_sched_insert_request(rq, false, run_queue, false);
2045 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2046 * @hctx: Pointer of the associated hardware queue.
2047 * @rq: Pointer to request to be sent.
2048 * @cookie: Request queue cookie.
2050 * If the device has enough resources to accept a new request now, send the
2051 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2052 * we can try send it another time in the future. Requests inserted at this
2053 * queue have higher priority.
2055 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2056 struct request *rq, blk_qc_t *cookie)
2061 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2063 hctx_lock(hctx, &srcu_idx);
2065 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
2066 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2067 blk_mq_request_bypass_insert(rq, false, true);
2068 else if (ret != BLK_STS_OK)
2069 blk_mq_end_request(rq, ret);
2071 hctx_unlock(hctx, srcu_idx);
2074 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2078 blk_qc_t unused_cookie;
2079 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2081 hctx_lock(hctx, &srcu_idx);
2082 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
2083 hctx_unlock(hctx, srcu_idx);
2088 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2089 struct list_head *list)
2094 while (!list_empty(list)) {
2096 struct request *rq = list_first_entry(list, struct request,
2099 list_del_init(&rq->queuelist);
2100 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2101 if (ret != BLK_STS_OK) {
2103 if (ret == BLK_STS_RESOURCE ||
2104 ret == BLK_STS_DEV_RESOURCE) {
2105 blk_mq_request_bypass_insert(rq, false,
2109 blk_mq_end_request(rq, ret);
2115 * If we didn't flush the entire list, we could have told
2116 * the driver there was more coming, but that turned out to
2119 if ((!list_empty(list) || errors) &&
2120 hctx->queue->mq_ops->commit_rqs && queued)
2121 hctx->queue->mq_ops->commit_rqs(hctx);
2124 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2126 list_add_tail(&rq->queuelist, &plug->mq_list);
2128 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
2129 struct request *tmp;
2131 tmp = list_first_entry(&plug->mq_list, struct request,
2133 if (tmp->q != rq->q)
2134 plug->multiple_queues = true;
2139 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
2140 * queues. This is important for md arrays to benefit from merging
2143 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
2145 if (plug->multiple_queues)
2146 return BLK_MAX_REQUEST_COUNT * 2;
2147 return BLK_MAX_REQUEST_COUNT;
2151 * blk_mq_submit_bio - Create and send a request to block device.
2152 * @bio: Bio pointer.
2154 * Builds up a request structure from @q and @bio and send to the device. The
2155 * request may not be queued directly to hardware if:
2156 * * This request can be merged with another one
2157 * * We want to place request at plug queue for possible future merging
2158 * * There is an IO scheduler active at this queue
2160 * It will not queue the request if there is an error with the bio, or at the
2163 * Returns: Request queue cookie.
2165 blk_qc_t blk_mq_submit_bio(struct bio *bio)
2167 struct request_queue *q = bio->bi_disk->queue;
2168 const int is_sync = op_is_sync(bio->bi_opf);
2169 const int is_flush_fua = op_is_flush(bio->bi_opf);
2170 struct blk_mq_alloc_data data = {
2174 struct blk_plug *plug;
2175 struct request *same_queue_rq = NULL;
2176 unsigned int nr_segs;
2180 blk_queue_bounce(q, &bio);
2181 __blk_queue_split(&bio, &nr_segs);
2183 if (!bio_integrity_prep(bio))
2186 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2187 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2190 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2193 rq_qos_throttle(q, bio);
2195 data.cmd_flags = bio->bi_opf;
2196 rq = __blk_mq_alloc_request(&data);
2197 if (unlikely(!rq)) {
2198 rq_qos_cleanup(q, bio);
2199 if (bio->bi_opf & REQ_NOWAIT)
2200 bio_wouldblock_error(bio);
2204 trace_block_getrq(q, bio, bio->bi_opf);
2206 rq_qos_track(q, rq, bio);
2208 cookie = request_to_qc_t(data.hctx, rq);
2210 blk_mq_bio_to_request(rq, bio, nr_segs);
2212 ret = blk_crypto_rq_get_keyslot(rq);
2213 if (ret != BLK_STS_OK) {
2214 bio->bi_status = ret;
2216 blk_mq_free_request(rq);
2217 return BLK_QC_T_NONE;
2220 plug = blk_mq_plug(q, bio);
2221 if (unlikely(is_flush_fua)) {
2222 /* Bypass scheduler for flush requests */
2223 blk_insert_flush(rq);
2224 blk_mq_run_hw_queue(data.hctx, true);
2225 } else if (plug && (q->nr_hw_queues == 1 ||
2226 blk_mq_is_sbitmap_shared(rq->mq_hctx->flags) ||
2227 q->mq_ops->commit_rqs || !blk_queue_nonrot(q))) {
2229 * Use plugging if we have a ->commit_rqs() hook as well, as
2230 * we know the driver uses bd->last in a smart fashion.
2232 * Use normal plugging if this disk is slow HDD, as sequential
2233 * IO may benefit a lot from plug merging.
2235 unsigned int request_count = plug->rq_count;
2236 struct request *last = NULL;
2239 trace_block_plug(q);
2241 last = list_entry_rq(plug->mq_list.prev);
2243 if (request_count >= blk_plug_max_rq_count(plug) || (last &&
2244 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2245 blk_flush_plug_list(plug, false);
2246 trace_block_plug(q);
2249 blk_add_rq_to_plug(plug, rq);
2250 } else if (q->elevator) {
2251 /* Insert the request at the IO scheduler queue */
2252 blk_mq_sched_insert_request(rq, false, true, true);
2253 } else if (plug && !blk_queue_nomerges(q)) {
2255 * We do limited plugging. If the bio can be merged, do that.
2256 * Otherwise the existing request in the plug list will be
2257 * issued. So the plug list will have one request at most
2258 * The plug list might get flushed before this. If that happens,
2259 * the plug list is empty, and same_queue_rq is invalid.
2261 if (list_empty(&plug->mq_list))
2262 same_queue_rq = NULL;
2263 if (same_queue_rq) {
2264 list_del_init(&same_queue_rq->queuelist);
2267 blk_add_rq_to_plug(plug, rq);
2268 trace_block_plug(q);
2270 if (same_queue_rq) {
2271 data.hctx = same_queue_rq->mq_hctx;
2272 trace_block_unplug(q, 1, true);
2273 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2276 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2277 !data.hctx->dispatch_busy) {
2279 * There is no scheduler and we can try to send directly
2282 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2285 blk_mq_sched_insert_request(rq, false, true, true);
2291 return BLK_QC_T_NONE;
2294 static size_t order_to_size(unsigned int order)
2296 return (size_t)PAGE_SIZE << order;
2299 /* called before freeing request pool in @tags */
2300 static void blk_mq_clear_rq_mapping(struct blk_mq_tag_set *set,
2301 struct blk_mq_tags *tags, unsigned int hctx_idx)
2303 struct blk_mq_tags *drv_tags = set->tags[hctx_idx];
2305 unsigned long flags;
2307 list_for_each_entry(page, &tags->page_list, lru) {
2308 unsigned long start = (unsigned long)page_address(page);
2309 unsigned long end = start + order_to_size(page->private);
2312 for (i = 0; i < set->queue_depth; i++) {
2313 struct request *rq = drv_tags->rqs[i];
2314 unsigned long rq_addr = (unsigned long)rq;
2316 if (rq_addr >= start && rq_addr < end) {
2317 WARN_ON_ONCE(refcount_read(&rq->ref) != 0);
2318 cmpxchg(&drv_tags->rqs[i], rq, NULL);
2324 * Wait until all pending iteration is done.
2326 * Request reference is cleared and it is guaranteed to be observed
2327 * after the ->lock is released.
2329 spin_lock_irqsave(&drv_tags->lock, flags);
2330 spin_unlock_irqrestore(&drv_tags->lock, flags);
2333 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2334 unsigned int hctx_idx)
2338 if (tags->rqs && set->ops->exit_request) {
2341 for (i = 0; i < tags->nr_tags; i++) {
2342 struct request *rq = tags->static_rqs[i];
2346 set->ops->exit_request(set, rq, hctx_idx);
2347 tags->static_rqs[i] = NULL;
2351 blk_mq_clear_rq_mapping(set, tags, hctx_idx);
2353 while (!list_empty(&tags->page_list)) {
2354 page = list_first_entry(&tags->page_list, struct page, lru);
2355 list_del_init(&page->lru);
2357 * Remove kmemleak object previously allocated in
2358 * blk_mq_alloc_rqs().
2360 kmemleak_free(page_address(page));
2361 __free_pages(page, page->private);
2365 void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags)
2369 kfree(tags->static_rqs);
2370 tags->static_rqs = NULL;
2372 blk_mq_free_tags(tags, flags);
2375 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2376 unsigned int hctx_idx,
2377 unsigned int nr_tags,
2378 unsigned int reserved_tags,
2381 struct blk_mq_tags *tags;
2384 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2385 if (node == NUMA_NO_NODE)
2386 node = set->numa_node;
2388 tags = blk_mq_init_tags(nr_tags, reserved_tags, node, flags);
2392 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2393 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2396 blk_mq_free_tags(tags, flags);
2400 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2401 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2403 if (!tags->static_rqs) {
2405 blk_mq_free_tags(tags, flags);
2412 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2413 unsigned int hctx_idx, int node)
2417 if (set->ops->init_request) {
2418 ret = set->ops->init_request(set, rq, hctx_idx, node);
2423 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2427 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2428 unsigned int hctx_idx, unsigned int depth)
2430 unsigned int i, j, entries_per_page, max_order = 4;
2431 size_t rq_size, left;
2434 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2435 if (node == NUMA_NO_NODE)
2436 node = set->numa_node;
2438 INIT_LIST_HEAD(&tags->page_list);
2441 * rq_size is the size of the request plus driver payload, rounded
2442 * to the cacheline size
2444 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2446 left = rq_size * depth;
2448 for (i = 0; i < depth; ) {
2449 int this_order = max_order;
2454 while (this_order && left < order_to_size(this_order - 1))
2458 page = alloc_pages_node(node,
2459 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2465 if (order_to_size(this_order) < rq_size)
2472 page->private = this_order;
2473 list_add_tail(&page->lru, &tags->page_list);
2475 p = page_address(page);
2477 * Allow kmemleak to scan these pages as they contain pointers
2478 * to additional allocations like via ops->init_request().
2480 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2481 entries_per_page = order_to_size(this_order) / rq_size;
2482 to_do = min(entries_per_page, depth - i);
2483 left -= to_do * rq_size;
2484 for (j = 0; j < to_do; j++) {
2485 struct request *rq = p;
2487 tags->static_rqs[i] = rq;
2488 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2489 tags->static_rqs[i] = NULL;
2500 blk_mq_free_rqs(set, tags, hctx_idx);
2504 struct rq_iter_data {
2505 struct blk_mq_hw_ctx *hctx;
2509 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2511 struct rq_iter_data *iter_data = data;
2513 if (rq->mq_hctx != iter_data->hctx)
2515 iter_data->has_rq = true;
2519 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2521 struct blk_mq_tags *tags = hctx->sched_tags ?
2522 hctx->sched_tags : hctx->tags;
2523 struct rq_iter_data data = {
2527 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2531 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2532 struct blk_mq_hw_ctx *hctx)
2534 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2536 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2541 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2543 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2544 struct blk_mq_hw_ctx, cpuhp_online);
2546 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2547 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2551 * Prevent new request from being allocated on the current hctx.
2553 * The smp_mb__after_atomic() Pairs with the implied barrier in
2554 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2555 * seen once we return from the tag allocator.
2557 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2558 smp_mb__after_atomic();
2561 * Try to grab a reference to the queue and wait for any outstanding
2562 * requests. If we could not grab a reference the queue has been
2563 * frozen and there are no requests.
2565 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2566 while (blk_mq_hctx_has_requests(hctx))
2568 percpu_ref_put(&hctx->queue->q_usage_counter);
2574 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2576 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2577 struct blk_mq_hw_ctx, cpuhp_online);
2579 if (cpumask_test_cpu(cpu, hctx->cpumask))
2580 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2585 * 'cpu' is going away. splice any existing rq_list entries from this
2586 * software queue to the hw queue dispatch list, and ensure that it
2589 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2591 struct blk_mq_hw_ctx *hctx;
2592 struct blk_mq_ctx *ctx;
2594 enum hctx_type type;
2596 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2597 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2600 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2603 spin_lock(&ctx->lock);
2604 if (!list_empty(&ctx->rq_lists[type])) {
2605 list_splice_init(&ctx->rq_lists[type], &tmp);
2606 blk_mq_hctx_clear_pending(hctx, ctx);
2608 spin_unlock(&ctx->lock);
2610 if (list_empty(&tmp))
2613 spin_lock(&hctx->lock);
2614 list_splice_tail_init(&tmp, &hctx->dispatch);
2615 spin_unlock(&hctx->lock);
2617 blk_mq_run_hw_queue(hctx, true);
2621 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2623 if (!(hctx->flags & BLK_MQ_F_STACKING))
2624 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2625 &hctx->cpuhp_online);
2626 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2631 * Before freeing hw queue, clearing the flush request reference in
2632 * tags->rqs[] for avoiding potential UAF.
2634 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
2635 unsigned int queue_depth, struct request *flush_rq)
2638 unsigned long flags;
2640 /* The hw queue may not be mapped yet */
2644 WARN_ON_ONCE(refcount_read(&flush_rq->ref) != 0);
2646 for (i = 0; i < queue_depth; i++)
2647 cmpxchg(&tags->rqs[i], flush_rq, NULL);
2650 * Wait until all pending iteration is done.
2652 * Request reference is cleared and it is guaranteed to be observed
2653 * after the ->lock is released.
2655 spin_lock_irqsave(&tags->lock, flags);
2656 spin_unlock_irqrestore(&tags->lock, flags);
2659 /* hctx->ctxs will be freed in queue's release handler */
2660 static void blk_mq_exit_hctx(struct request_queue *q,
2661 struct blk_mq_tag_set *set,
2662 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2664 struct request *flush_rq = hctx->fq->flush_rq;
2666 if (blk_mq_hw_queue_mapped(hctx))
2667 blk_mq_tag_idle(hctx);
2669 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
2670 set->queue_depth, flush_rq);
2671 if (set->ops->exit_request)
2672 set->ops->exit_request(set, flush_rq, hctx_idx);
2674 if (set->ops->exit_hctx)
2675 set->ops->exit_hctx(hctx, hctx_idx);
2677 blk_mq_remove_cpuhp(hctx);
2679 spin_lock(&q->unused_hctx_lock);
2680 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2681 spin_unlock(&q->unused_hctx_lock);
2684 static void blk_mq_exit_hw_queues(struct request_queue *q,
2685 struct blk_mq_tag_set *set, int nr_queue)
2687 struct blk_mq_hw_ctx *hctx;
2690 queue_for_each_hw_ctx(q, hctx, i) {
2693 blk_mq_debugfs_unregister_hctx(hctx);
2694 blk_mq_exit_hctx(q, set, hctx, i);
2698 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2700 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2702 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2703 __alignof__(struct blk_mq_hw_ctx)) !=
2704 sizeof(struct blk_mq_hw_ctx));
2706 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2707 hw_ctx_size += sizeof(struct srcu_struct);
2712 static int blk_mq_init_hctx(struct request_queue *q,
2713 struct blk_mq_tag_set *set,
2714 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2716 hctx->queue_num = hctx_idx;
2718 if (!(hctx->flags & BLK_MQ_F_STACKING))
2719 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2720 &hctx->cpuhp_online);
2721 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2723 hctx->tags = set->tags[hctx_idx];
2725 if (set->ops->init_hctx &&
2726 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2727 goto unregister_cpu_notifier;
2729 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2735 if (set->ops->exit_hctx)
2736 set->ops->exit_hctx(hctx, hctx_idx);
2737 unregister_cpu_notifier:
2738 blk_mq_remove_cpuhp(hctx);
2742 static struct blk_mq_hw_ctx *
2743 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2746 struct blk_mq_hw_ctx *hctx;
2747 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2749 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2751 goto fail_alloc_hctx;
2753 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2756 atomic_set(&hctx->nr_active, 0);
2757 atomic_set(&hctx->elevator_queued, 0);
2758 if (node == NUMA_NO_NODE)
2759 node = set->numa_node;
2760 hctx->numa_node = node;
2762 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2763 spin_lock_init(&hctx->lock);
2764 INIT_LIST_HEAD(&hctx->dispatch);
2766 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
2768 INIT_LIST_HEAD(&hctx->hctx_list);
2771 * Allocate space for all possible cpus to avoid allocation at
2774 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2779 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2784 spin_lock_init(&hctx->dispatch_wait_lock);
2785 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2786 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2788 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2792 if (hctx->flags & BLK_MQ_F_BLOCKING)
2793 init_srcu_struct(hctx->srcu);
2794 blk_mq_hctx_kobj_init(hctx);
2799 sbitmap_free(&hctx->ctx_map);
2803 free_cpumask_var(hctx->cpumask);
2810 static void blk_mq_init_cpu_queues(struct request_queue *q,
2811 unsigned int nr_hw_queues)
2813 struct blk_mq_tag_set *set = q->tag_set;
2816 for_each_possible_cpu(i) {
2817 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2818 struct blk_mq_hw_ctx *hctx;
2822 spin_lock_init(&__ctx->lock);
2823 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2824 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2829 * Set local node, IFF we have more than one hw queue. If
2830 * not, we remain on the home node of the device
2832 for (j = 0; j < set->nr_maps; j++) {
2833 hctx = blk_mq_map_queue_type(q, j, i);
2834 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2835 hctx->numa_node = cpu_to_node(i);
2840 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2843 unsigned int flags = set->flags;
2846 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2847 set->queue_depth, set->reserved_tags, flags);
2848 if (!set->tags[hctx_idx])
2851 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2856 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2857 set->tags[hctx_idx] = NULL;
2861 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2862 unsigned int hctx_idx)
2864 unsigned int flags = set->flags;
2866 if (set->tags && set->tags[hctx_idx]) {
2867 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2868 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2869 set->tags[hctx_idx] = NULL;
2873 static void blk_mq_map_swqueue(struct request_queue *q)
2875 unsigned int i, j, hctx_idx;
2876 struct blk_mq_hw_ctx *hctx;
2877 struct blk_mq_ctx *ctx;
2878 struct blk_mq_tag_set *set = q->tag_set;
2880 queue_for_each_hw_ctx(q, hctx, i) {
2881 cpumask_clear(hctx->cpumask);
2883 hctx->dispatch_from = NULL;
2887 * Map software to hardware queues.
2889 * If the cpu isn't present, the cpu is mapped to first hctx.
2891 for_each_possible_cpu(i) {
2893 ctx = per_cpu_ptr(q->queue_ctx, i);
2894 for (j = 0; j < set->nr_maps; j++) {
2895 if (!set->map[j].nr_queues) {
2896 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2897 HCTX_TYPE_DEFAULT, i);
2900 hctx_idx = set->map[j].mq_map[i];
2901 /* unmapped hw queue can be remapped after CPU topo changed */
2902 if (!set->tags[hctx_idx] &&
2903 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2905 * If tags initialization fail for some hctx,
2906 * that hctx won't be brought online. In this
2907 * case, remap the current ctx to hctx[0] which
2908 * is guaranteed to always have tags allocated
2910 set->map[j].mq_map[i] = 0;
2913 hctx = blk_mq_map_queue_type(q, j, i);
2914 ctx->hctxs[j] = hctx;
2916 * If the CPU is already set in the mask, then we've
2917 * mapped this one already. This can happen if
2918 * devices share queues across queue maps.
2920 if (cpumask_test_cpu(i, hctx->cpumask))
2923 cpumask_set_cpu(i, hctx->cpumask);
2925 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2926 hctx->ctxs[hctx->nr_ctx++] = ctx;
2929 * If the nr_ctx type overflows, we have exceeded the
2930 * amount of sw queues we can support.
2932 BUG_ON(!hctx->nr_ctx);
2935 for (; j < HCTX_MAX_TYPES; j++)
2936 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2937 HCTX_TYPE_DEFAULT, i);
2940 queue_for_each_hw_ctx(q, hctx, i) {
2942 * If no software queues are mapped to this hardware queue,
2943 * disable it and free the request entries.
2945 if (!hctx->nr_ctx) {
2946 /* Never unmap queue 0. We need it as a
2947 * fallback in case of a new remap fails
2950 if (i && set->tags[i])
2951 blk_mq_free_map_and_requests(set, i);
2957 hctx->tags = set->tags[i];
2958 WARN_ON(!hctx->tags);
2961 * Set the map size to the number of mapped software queues.
2962 * This is more accurate and more efficient than looping
2963 * over all possibly mapped software queues.
2965 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2968 * Initialize batch roundrobin counts
2970 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2971 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2976 * Caller needs to ensure that we're either frozen/quiesced, or that
2977 * the queue isn't live yet.
2979 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2981 struct blk_mq_hw_ctx *hctx;
2984 queue_for_each_hw_ctx(q, hctx, i) {
2986 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
2988 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
2992 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
2995 struct request_queue *q;
2997 lockdep_assert_held(&set->tag_list_lock);
2999 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3000 blk_mq_freeze_queue(q);
3001 queue_set_hctx_shared(q, shared);
3002 blk_mq_unfreeze_queue(q);
3006 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3008 struct blk_mq_tag_set *set = q->tag_set;
3010 mutex_lock(&set->tag_list_lock);
3011 list_del(&q->tag_set_list);
3012 if (list_is_singular(&set->tag_list)) {
3013 /* just transitioned to unshared */
3014 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3015 /* update existing queue */
3016 blk_mq_update_tag_set_shared(set, false);
3018 mutex_unlock(&set->tag_list_lock);
3019 INIT_LIST_HEAD(&q->tag_set_list);
3022 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3023 struct request_queue *q)
3025 mutex_lock(&set->tag_list_lock);
3028 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3030 if (!list_empty(&set->tag_list) &&
3031 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3032 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3033 /* update existing queue */
3034 blk_mq_update_tag_set_shared(set, true);
3036 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3037 queue_set_hctx_shared(q, true);
3038 list_add_tail(&q->tag_set_list, &set->tag_list);
3040 mutex_unlock(&set->tag_list_lock);
3043 /* All allocations will be freed in release handler of q->mq_kobj */
3044 static int blk_mq_alloc_ctxs(struct request_queue *q)
3046 struct blk_mq_ctxs *ctxs;
3049 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3053 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3054 if (!ctxs->queue_ctx)
3057 for_each_possible_cpu(cpu) {
3058 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3062 q->mq_kobj = &ctxs->kobj;
3063 q->queue_ctx = ctxs->queue_ctx;
3072 * It is the actual release handler for mq, but we do it from
3073 * request queue's release handler for avoiding use-after-free
3074 * and headache because q->mq_kobj shouldn't have been introduced,
3075 * but we can't group ctx/kctx kobj without it.
3077 void blk_mq_release(struct request_queue *q)
3079 struct blk_mq_hw_ctx *hctx, *next;
3082 queue_for_each_hw_ctx(q, hctx, i)
3083 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3085 /* all hctx are in .unused_hctx_list now */
3086 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3087 list_del_init(&hctx->hctx_list);
3088 kobject_put(&hctx->kobj);
3091 kfree(q->queue_hw_ctx);
3094 * release .mq_kobj and sw queue's kobject now because
3095 * both share lifetime with request queue.
3097 blk_mq_sysfs_deinit(q);
3100 struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3103 struct request_queue *uninit_q, *q;
3105 uninit_q = blk_alloc_queue(set->numa_node);
3107 return ERR_PTR(-ENOMEM);
3108 uninit_q->queuedata = queuedata;
3111 * Initialize the queue without an elevator. device_add_disk() will do
3112 * the initialization.
3114 q = blk_mq_init_allocated_queue(set, uninit_q, false);
3116 blk_cleanup_queue(uninit_q);
3120 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data);
3122 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3124 return blk_mq_init_queue_data(set, NULL);
3126 EXPORT_SYMBOL(blk_mq_init_queue);
3129 * Helper for setting up a queue with mq ops, given queue depth, and
3130 * the passed in mq ops flags.
3132 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
3133 const struct blk_mq_ops *ops,
3134 unsigned int queue_depth,
3135 unsigned int set_flags)
3137 struct request_queue *q;
3140 memset(set, 0, sizeof(*set));
3142 set->nr_hw_queues = 1;
3144 set->queue_depth = queue_depth;
3145 set->numa_node = NUMA_NO_NODE;
3146 set->flags = set_flags;
3148 ret = blk_mq_alloc_tag_set(set);
3150 return ERR_PTR(ret);
3152 q = blk_mq_init_queue(set);
3154 blk_mq_free_tag_set(set);
3160 EXPORT_SYMBOL(blk_mq_init_sq_queue);
3162 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3163 struct blk_mq_tag_set *set, struct request_queue *q,
3164 int hctx_idx, int node)
3166 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3168 /* reuse dead hctx first */
3169 spin_lock(&q->unused_hctx_lock);
3170 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3171 if (tmp->numa_node == node) {
3177 list_del_init(&hctx->hctx_list);
3178 spin_unlock(&q->unused_hctx_lock);
3181 hctx = blk_mq_alloc_hctx(q, set, node);
3185 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3191 kobject_put(&hctx->kobj);
3196 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3197 struct request_queue *q)
3200 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3202 if (q->nr_hw_queues < set->nr_hw_queues) {
3203 struct blk_mq_hw_ctx **new_hctxs;
3205 new_hctxs = kcalloc_node(set->nr_hw_queues,
3206 sizeof(*new_hctxs), GFP_KERNEL,
3211 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3213 q->queue_hw_ctx = new_hctxs;
3218 /* protect against switching io scheduler */
3219 mutex_lock(&q->sysfs_lock);
3220 for (i = 0; i < set->nr_hw_queues; i++) {
3222 struct blk_mq_hw_ctx *hctx;
3224 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3226 * If the hw queue has been mapped to another numa node,
3227 * we need to realloc the hctx. If allocation fails, fallback
3228 * to use the previous one.
3230 if (hctxs[i] && (hctxs[i]->numa_node == node))
3233 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3236 blk_mq_exit_hctx(q, set, hctxs[i], i);
3240 pr_warn("Allocate new hctx on node %d fails,\
3241 fallback to previous one on node %d\n",
3242 node, hctxs[i]->numa_node);
3248 * Increasing nr_hw_queues fails. Free the newly allocated
3249 * hctxs and keep the previous q->nr_hw_queues.
3251 if (i != set->nr_hw_queues) {
3252 j = q->nr_hw_queues;
3256 end = q->nr_hw_queues;
3257 q->nr_hw_queues = set->nr_hw_queues;
3260 for (; j < end; j++) {
3261 struct blk_mq_hw_ctx *hctx = hctxs[j];
3265 blk_mq_free_map_and_requests(set, j);
3266 blk_mq_exit_hctx(q, set, hctx, j);
3270 mutex_unlock(&q->sysfs_lock);
3273 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3274 struct request_queue *q,
3277 /* mark the queue as mq asap */
3278 q->mq_ops = set->ops;
3280 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3281 blk_mq_poll_stats_bkt,
3282 BLK_MQ_POLL_STATS_BKTS, q);
3286 if (blk_mq_alloc_ctxs(q))
3289 /* init q->mq_kobj and sw queues' kobjects */
3290 blk_mq_sysfs_init(q);
3292 INIT_LIST_HEAD(&q->unused_hctx_list);
3293 spin_lock_init(&q->unused_hctx_lock);
3295 blk_mq_realloc_hw_ctxs(set, q);
3296 if (!q->nr_hw_queues)
3299 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3300 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3304 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3305 if (set->nr_maps > HCTX_TYPE_POLL &&
3306 set->map[HCTX_TYPE_POLL].nr_queues)
3307 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3309 q->sg_reserved_size = INT_MAX;
3311 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3312 INIT_LIST_HEAD(&q->requeue_list);
3313 spin_lock_init(&q->requeue_lock);
3315 q->nr_requests = set->queue_depth;
3318 * Default to classic polling
3320 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3322 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3323 blk_mq_add_queue_tag_set(set, q);
3324 blk_mq_map_swqueue(q);
3327 elevator_init_mq(q);
3332 kfree(q->queue_hw_ctx);
3333 q->nr_hw_queues = 0;
3334 blk_mq_sysfs_deinit(q);
3336 blk_stat_free_callback(q->poll_cb);
3340 return ERR_PTR(-ENOMEM);
3342 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3344 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3345 void blk_mq_exit_queue(struct request_queue *q)
3347 struct blk_mq_tag_set *set = q->tag_set;
3349 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
3350 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3351 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
3352 blk_mq_del_queue_tag_set(q);
3355 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3359 for (i = 0; i < set->nr_hw_queues; i++) {
3360 if (!__blk_mq_alloc_map_and_request(set, i))
3369 blk_mq_free_map_and_requests(set, i);
3375 * Allocate the request maps associated with this tag_set. Note that this
3376 * may reduce the depth asked for, if memory is tight. set->queue_depth
3377 * will be updated to reflect the allocated depth.
3379 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3384 depth = set->queue_depth;
3386 err = __blk_mq_alloc_rq_maps(set);
3390 set->queue_depth >>= 1;
3391 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3395 } while (set->queue_depth);
3397 if (!set->queue_depth || err) {
3398 pr_err("blk-mq: failed to allocate request map\n");
3402 if (depth != set->queue_depth)
3403 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3404 depth, set->queue_depth);
3409 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3412 * blk_mq_map_queues() and multiple .map_queues() implementations
3413 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3414 * number of hardware queues.
3416 if (set->nr_maps == 1)
3417 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3419 if (set->ops->map_queues && !is_kdump_kernel()) {
3423 * transport .map_queues is usually done in the following
3426 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3427 * mask = get_cpu_mask(queue)
3428 * for_each_cpu(cpu, mask)
3429 * set->map[x].mq_map[cpu] = queue;
3432 * When we need to remap, the table has to be cleared for
3433 * killing stale mapping since one CPU may not be mapped
3436 for (i = 0; i < set->nr_maps; i++)
3437 blk_mq_clear_mq_map(&set->map[i]);
3439 return set->ops->map_queues(set);
3441 BUG_ON(set->nr_maps > 1);
3442 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3446 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3447 int cur_nr_hw_queues, int new_nr_hw_queues)
3449 struct blk_mq_tags **new_tags;
3451 if (cur_nr_hw_queues >= new_nr_hw_queues)
3454 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3455 GFP_KERNEL, set->numa_node);
3460 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3461 sizeof(*set->tags));
3463 set->tags = new_tags;
3464 set->nr_hw_queues = new_nr_hw_queues;
3470 * Alloc a tag set to be associated with one or more request queues.
3471 * May fail with EINVAL for various error conditions. May adjust the
3472 * requested depth down, if it's too large. In that case, the set
3473 * value will be stored in set->queue_depth.
3475 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3479 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3481 if (!set->nr_hw_queues)
3483 if (!set->queue_depth)
3485 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3488 if (!set->ops->queue_rq)
3491 if (!set->ops->get_budget ^ !set->ops->put_budget)
3494 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3495 pr_info("blk-mq: reduced tag depth to %u\n",
3497 set->queue_depth = BLK_MQ_MAX_DEPTH;
3502 else if (set->nr_maps > HCTX_MAX_TYPES)
3506 * If a crashdump is active, then we are potentially in a very
3507 * memory constrained environment. Limit us to 1 queue and
3508 * 64 tags to prevent using too much memory.
3510 if (is_kdump_kernel()) {
3511 set->nr_hw_queues = 1;
3513 set->queue_depth = min(64U, set->queue_depth);
3516 * There is no use for more h/w queues than cpus if we just have
3519 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3520 set->nr_hw_queues = nr_cpu_ids;
3522 if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0)
3526 for (i = 0; i < set->nr_maps; i++) {
3527 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3528 sizeof(set->map[i].mq_map[0]),
3529 GFP_KERNEL, set->numa_node);
3530 if (!set->map[i].mq_map)
3531 goto out_free_mq_map;
3532 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3535 ret = blk_mq_update_queue_map(set);
3537 goto out_free_mq_map;
3539 ret = blk_mq_alloc_map_and_requests(set);
3541 goto out_free_mq_map;
3543 if (blk_mq_is_sbitmap_shared(set->flags)) {
3544 atomic_set(&set->active_queues_shared_sbitmap, 0);
3546 if (blk_mq_init_shared_sbitmap(set, set->flags)) {
3548 goto out_free_mq_rq_maps;
3552 mutex_init(&set->tag_list_lock);
3553 INIT_LIST_HEAD(&set->tag_list);
3557 out_free_mq_rq_maps:
3558 for (i = 0; i < set->nr_hw_queues; i++)
3559 blk_mq_free_map_and_requests(set, i);
3561 for (i = 0; i < set->nr_maps; i++) {
3562 kfree(set->map[i].mq_map);
3563 set->map[i].mq_map = NULL;
3569 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3571 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3575 for (i = 0; i < set->nr_hw_queues; i++)
3576 blk_mq_free_map_and_requests(set, i);
3578 if (blk_mq_is_sbitmap_shared(set->flags))
3579 blk_mq_exit_shared_sbitmap(set);
3581 for (j = 0; j < set->nr_maps; j++) {
3582 kfree(set->map[j].mq_map);
3583 set->map[j].mq_map = NULL;
3589 EXPORT_SYMBOL(blk_mq_free_tag_set);
3591 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3593 struct blk_mq_tag_set *set = q->tag_set;
3594 struct blk_mq_hw_ctx *hctx;
3600 if (q->nr_requests == nr)
3603 blk_mq_freeze_queue(q);
3604 blk_mq_quiesce_queue(q);
3607 queue_for_each_hw_ctx(q, hctx, i) {
3611 * If we're using an MQ scheduler, just update the scheduler
3612 * queue depth. This is similar to what the old code would do.
3614 if (!hctx->sched_tags) {
3615 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3617 if (!ret && blk_mq_is_sbitmap_shared(set->flags))
3618 blk_mq_tag_resize_shared_sbitmap(set, nr);
3620 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3625 if (q->elevator && q->elevator->type->ops.depth_updated)
3626 q->elevator->type->ops.depth_updated(hctx);
3630 q->nr_requests = nr;
3632 blk_mq_unquiesce_queue(q);
3633 blk_mq_unfreeze_queue(q);
3639 * request_queue and elevator_type pair.
3640 * It is just used by __blk_mq_update_nr_hw_queues to cache
3641 * the elevator_type associated with a request_queue.
3643 struct blk_mq_qe_pair {
3644 struct list_head node;
3645 struct request_queue *q;
3646 struct elevator_type *type;
3650 * Cache the elevator_type in qe pair list and switch the
3651 * io scheduler to 'none'
3653 static bool blk_mq_elv_switch_none(struct list_head *head,
3654 struct request_queue *q)
3656 struct blk_mq_qe_pair *qe;
3661 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3665 INIT_LIST_HEAD(&qe->node);
3667 qe->type = q->elevator->type;
3668 list_add(&qe->node, head);
3670 mutex_lock(&q->sysfs_lock);
3672 * After elevator_switch_mq, the previous elevator_queue will be
3673 * released by elevator_release. The reference of the io scheduler
3674 * module get by elevator_get will also be put. So we need to get
3675 * a reference of the io scheduler module here to prevent it to be
3678 __module_get(qe->type->elevator_owner);
3679 elevator_switch_mq(q, NULL);
3680 mutex_unlock(&q->sysfs_lock);
3685 static void blk_mq_elv_switch_back(struct list_head *head,
3686 struct request_queue *q)
3688 struct blk_mq_qe_pair *qe;
3689 struct elevator_type *t = NULL;
3691 list_for_each_entry(qe, head, node)
3700 list_del(&qe->node);
3703 mutex_lock(&q->sysfs_lock);
3704 elevator_switch_mq(q, t);
3705 mutex_unlock(&q->sysfs_lock);
3708 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3711 struct request_queue *q;
3713 int prev_nr_hw_queues;
3715 lockdep_assert_held(&set->tag_list_lock);
3717 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3718 nr_hw_queues = nr_cpu_ids;
3719 if (nr_hw_queues < 1)
3721 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3724 list_for_each_entry(q, &set->tag_list, tag_set_list)
3725 blk_mq_freeze_queue(q);
3727 * Switch IO scheduler to 'none', cleaning up the data associated
3728 * with the previous scheduler. We will switch back once we are done
3729 * updating the new sw to hw queue mappings.
3731 list_for_each_entry(q, &set->tag_list, tag_set_list)
3732 if (!blk_mq_elv_switch_none(&head, q))
3735 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3736 blk_mq_debugfs_unregister_hctxs(q);
3737 blk_mq_sysfs_unregister(q);
3740 prev_nr_hw_queues = set->nr_hw_queues;
3741 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3745 set->nr_hw_queues = nr_hw_queues;
3747 blk_mq_update_queue_map(set);
3748 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3749 blk_mq_realloc_hw_ctxs(set, q);
3750 if (q->nr_hw_queues != set->nr_hw_queues) {
3751 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3752 nr_hw_queues, prev_nr_hw_queues);
3753 set->nr_hw_queues = prev_nr_hw_queues;
3754 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3757 blk_mq_map_swqueue(q);
3761 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3762 blk_mq_sysfs_register(q);
3763 blk_mq_debugfs_register_hctxs(q);
3767 list_for_each_entry(q, &set->tag_list, tag_set_list)
3768 blk_mq_elv_switch_back(&head, q);
3770 list_for_each_entry(q, &set->tag_list, tag_set_list)
3771 blk_mq_unfreeze_queue(q);
3774 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3776 mutex_lock(&set->tag_list_lock);
3777 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3778 mutex_unlock(&set->tag_list_lock);
3780 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3782 /* Enable polling stats and return whether they were already enabled. */
3783 static bool blk_poll_stats_enable(struct request_queue *q)
3785 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3786 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3788 blk_stat_add_callback(q, q->poll_cb);
3792 static void blk_mq_poll_stats_start(struct request_queue *q)
3795 * We don't arm the callback if polling stats are not enabled or the
3796 * callback is already active.
3798 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3799 blk_stat_is_active(q->poll_cb))
3802 blk_stat_activate_msecs(q->poll_cb, 100);
3805 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3807 struct request_queue *q = cb->data;
3810 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3811 if (cb->stat[bucket].nr_samples)
3812 q->poll_stat[bucket] = cb->stat[bucket];
3816 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3819 unsigned long ret = 0;
3823 * If stats collection isn't on, don't sleep but turn it on for
3826 if (!blk_poll_stats_enable(q))
3830 * As an optimistic guess, use half of the mean service time
3831 * for this type of request. We can (and should) make this smarter.
3832 * For instance, if the completion latencies are tight, we can
3833 * get closer than just half the mean. This is especially
3834 * important on devices where the completion latencies are longer
3835 * than ~10 usec. We do use the stats for the relevant IO size
3836 * if available which does lead to better estimates.
3838 bucket = blk_mq_poll_stats_bkt(rq);
3842 if (q->poll_stat[bucket].nr_samples)
3843 ret = (q->poll_stat[bucket].mean + 1) / 2;
3848 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3851 struct hrtimer_sleeper hs;
3852 enum hrtimer_mode mode;
3856 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3860 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3862 * 0: use half of prev avg
3863 * >0: use this specific value
3865 if (q->poll_nsec > 0)
3866 nsecs = q->poll_nsec;
3868 nsecs = blk_mq_poll_nsecs(q, rq);
3873 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3876 * This will be replaced with the stats tracking code, using
3877 * 'avg_completion_time / 2' as the pre-sleep target.
3881 mode = HRTIMER_MODE_REL;
3882 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3883 hrtimer_set_expires(&hs.timer, kt);
3886 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3888 set_current_state(TASK_UNINTERRUPTIBLE);
3889 hrtimer_sleeper_start_expires(&hs, mode);
3892 hrtimer_cancel(&hs.timer);
3893 mode = HRTIMER_MODE_ABS;
3894 } while (hs.task && !signal_pending(current));
3896 __set_current_state(TASK_RUNNING);
3897 destroy_hrtimer_on_stack(&hs.timer);
3901 static bool blk_mq_poll_hybrid(struct request_queue *q,
3902 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3906 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3909 if (!blk_qc_t_is_internal(cookie))
3910 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3912 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3914 * With scheduling, if the request has completed, we'll
3915 * get a NULL return here, as we clear the sched tag when
3916 * that happens. The request still remains valid, like always,
3917 * so we should be safe with just the NULL check.
3923 return blk_mq_poll_hybrid_sleep(q, rq);
3927 * blk_poll - poll for IO completions
3929 * @cookie: cookie passed back at IO submission time
3930 * @spin: whether to spin for completions
3933 * Poll for completions on the passed in queue. Returns number of
3934 * completed entries found. If @spin is true, then blk_poll will continue
3935 * looping until at least one completion is found, unless the task is
3936 * otherwise marked running (or we need to reschedule).
3938 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3940 struct blk_mq_hw_ctx *hctx;
3943 if (!blk_qc_t_valid(cookie) ||
3944 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3948 blk_flush_plug_list(current->plug, false);
3950 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3953 * If we sleep, have the caller restart the poll loop to reset
3954 * the state. Like for the other success return cases, the
3955 * caller is responsible for checking if the IO completed. If
3956 * the IO isn't complete, we'll get called again and will go
3957 * straight to the busy poll loop.
3959 if (blk_mq_poll_hybrid(q, hctx, cookie))
3962 hctx->poll_considered++;
3964 state = current->state;
3968 hctx->poll_invoked++;
3970 ret = q->mq_ops->poll(hctx);
3972 hctx->poll_success++;
3973 __set_current_state(TASK_RUNNING);
3977 if (signal_pending_state(state, current))
3978 __set_current_state(TASK_RUNNING);
3980 if (current->state == TASK_RUNNING)
3982 if (ret < 0 || !spin)
3985 } while (!need_resched());
3987 __set_current_state(TASK_RUNNING);
3990 EXPORT_SYMBOL_GPL(blk_poll);
3992 unsigned int blk_mq_rq_cpu(struct request *rq)
3994 return rq->mq_ctx->cpu;
3996 EXPORT_SYMBOL(blk_mq_rq_cpu);
3998 static int __init blk_mq_init(void)
4002 for_each_possible_cpu(i)
4003 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
4004 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4006 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4007 "block/softirq:dead", NULL,
4008 blk_softirq_cpu_dead);
4009 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4010 blk_mq_hctx_notify_dead);
4011 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4012 blk_mq_hctx_notify_online,
4013 blk_mq_hctx_notify_offline);
4016 subsys_initcall(blk_mq_init);