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 (rq->part == mi->part && blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
109 mi->inflight[rq_data_dir(rq)]++;
114 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
116 struct mq_inflight mi = { .part = part };
118 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
120 return mi.inflight[0] + mi.inflight[1];
123 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
124 unsigned int inflight[2])
126 struct mq_inflight mi = { .part = part };
128 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
129 inflight[0] = mi.inflight[0];
130 inflight[1] = mi.inflight[1];
133 void blk_freeze_queue_start(struct request_queue *q)
135 mutex_lock(&q->mq_freeze_lock);
136 if (++q->mq_freeze_depth == 1) {
137 percpu_ref_kill(&q->q_usage_counter);
138 mutex_unlock(&q->mq_freeze_lock);
140 blk_mq_run_hw_queues(q, false);
142 mutex_unlock(&q->mq_freeze_lock);
145 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
147 void blk_mq_freeze_queue_wait(struct request_queue *q)
149 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
151 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
153 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
154 unsigned long timeout)
156 return wait_event_timeout(q->mq_freeze_wq,
157 percpu_ref_is_zero(&q->q_usage_counter),
160 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
163 * Guarantee no request is in use, so we can change any data structure of
164 * the queue afterward.
166 void blk_freeze_queue(struct request_queue *q)
169 * In the !blk_mq case we are only calling this to kill the
170 * q_usage_counter, otherwise this increases the freeze depth
171 * and waits for it to return to zero. For this reason there is
172 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
173 * exported to drivers as the only user for unfreeze is blk_mq.
175 blk_freeze_queue_start(q);
176 blk_mq_freeze_queue_wait(q);
179 void blk_mq_freeze_queue(struct request_queue *q)
182 * ...just an alias to keep freeze and unfreeze actions balanced
183 * in the blk_mq_* namespace
187 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
189 void blk_mq_unfreeze_queue(struct request_queue *q)
191 mutex_lock(&q->mq_freeze_lock);
192 q->mq_freeze_depth--;
193 WARN_ON_ONCE(q->mq_freeze_depth < 0);
194 if (!q->mq_freeze_depth) {
195 percpu_ref_resurrect(&q->q_usage_counter);
196 wake_up_all(&q->mq_freeze_wq);
198 mutex_unlock(&q->mq_freeze_lock);
200 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
203 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
204 * mpt3sas driver such that this function can be removed.
206 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
208 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
210 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
213 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
216 * Note: this function does not prevent that the struct request end_io()
217 * callback function is invoked. Once this function is returned, we make
218 * sure no dispatch can happen until the queue is unquiesced via
219 * blk_mq_unquiesce_queue().
221 void blk_mq_quiesce_queue(struct request_queue *q)
223 struct blk_mq_hw_ctx *hctx;
227 blk_mq_quiesce_queue_nowait(q);
229 queue_for_each_hw_ctx(q, hctx, i) {
230 if (hctx->flags & BLK_MQ_F_BLOCKING)
231 synchronize_srcu(hctx->srcu);
238 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
241 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
244 * This function recovers queue into the state before quiescing
245 * which is done by blk_mq_quiesce_queue.
247 void blk_mq_unquiesce_queue(struct request_queue *q)
249 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
251 /* dispatch requests which are inserted during quiescing */
252 blk_mq_run_hw_queues(q, true);
254 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
256 void blk_mq_wake_waiters(struct request_queue *q)
258 struct blk_mq_hw_ctx *hctx;
261 queue_for_each_hw_ctx(q, hctx, i)
262 if (blk_mq_hw_queue_mapped(hctx))
263 blk_mq_tag_wakeup_all(hctx->tags, true);
267 * Only need start/end time stamping if we have iostat or
268 * blk stats enabled, or using an IO scheduler.
270 static inline bool blk_mq_need_time_stamp(struct request *rq)
272 return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
275 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
276 unsigned int tag, u64 alloc_time_ns)
278 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
279 struct request *rq = tags->static_rqs[tag];
281 if (data->q->elevator) {
282 rq->tag = BLK_MQ_NO_TAG;
283 rq->internal_tag = tag;
286 rq->internal_tag = BLK_MQ_NO_TAG;
289 /* csd/requeue_work/fifo_time is initialized before use */
291 rq->mq_ctx = data->ctx;
292 rq->mq_hctx = data->hctx;
294 rq->cmd_flags = data->cmd_flags;
295 if (data->flags & BLK_MQ_REQ_PM)
296 rq->rq_flags |= RQF_PM;
297 if (blk_queue_io_stat(data->q))
298 rq->rq_flags |= RQF_IO_STAT;
299 INIT_LIST_HEAD(&rq->queuelist);
300 INIT_HLIST_NODE(&rq->hash);
301 RB_CLEAR_NODE(&rq->rb_node);
304 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
305 rq->alloc_time_ns = alloc_time_ns;
307 if (blk_mq_need_time_stamp(rq))
308 rq->start_time_ns = ktime_get_ns();
310 rq->start_time_ns = 0;
311 rq->io_start_time_ns = 0;
312 rq->stats_sectors = 0;
313 rq->nr_phys_segments = 0;
314 #if defined(CONFIG_BLK_DEV_INTEGRITY)
315 rq->nr_integrity_segments = 0;
317 blk_crypto_rq_set_defaults(rq);
318 /* tag was already set */
319 WRITE_ONCE(rq->deadline, 0);
324 rq->end_io_data = NULL;
326 data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++;
327 refcount_set(&rq->ref, 1);
329 if (!op_is_flush(data->cmd_flags)) {
330 struct elevator_queue *e = data->q->elevator;
333 if (e && e->type->ops.prepare_request) {
334 if (e->type->icq_cache)
335 blk_mq_sched_assign_ioc(rq);
337 e->type->ops.prepare_request(rq);
338 rq->rq_flags |= RQF_ELVPRIV;
342 data->hctx->queued++;
346 static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data)
348 struct request_queue *q = data->q;
349 struct elevator_queue *e = q->elevator;
350 u64 alloc_time_ns = 0;
353 /* alloc_time includes depth and tag waits */
354 if (blk_queue_rq_alloc_time(q))
355 alloc_time_ns = ktime_get_ns();
357 if (data->cmd_flags & REQ_NOWAIT)
358 data->flags |= BLK_MQ_REQ_NOWAIT;
362 * Flush requests are special and go directly to the
363 * dispatch list. Don't include reserved tags in the
364 * limiting, as it isn't useful.
366 if (!op_is_flush(data->cmd_flags) &&
367 e->type->ops.limit_depth &&
368 !(data->flags & BLK_MQ_REQ_RESERVED))
369 e->type->ops.limit_depth(data->cmd_flags, data);
373 data->ctx = blk_mq_get_ctx(q);
374 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
376 blk_mq_tag_busy(data->hctx);
379 * Waiting allocations only fail because of an inactive hctx. In that
380 * case just retry the hctx assignment and tag allocation as CPU hotplug
381 * should have migrated us to an online CPU by now.
383 tag = blk_mq_get_tag(data);
384 if (tag == BLK_MQ_NO_TAG) {
385 if (data->flags & BLK_MQ_REQ_NOWAIT)
389 * Give up the CPU and sleep for a random short time to ensure
390 * that thread using a realtime scheduling class are migrated
391 * off the CPU, and thus off the hctx that is going away.
396 return blk_mq_rq_ctx_init(data, tag, alloc_time_ns);
399 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
400 blk_mq_req_flags_t flags)
402 struct blk_mq_alloc_data data = {
410 ret = blk_queue_enter(q, flags);
414 rq = __blk_mq_alloc_request(&data);
418 rq->__sector = (sector_t) -1;
419 rq->bio = rq->biotail = NULL;
423 return ERR_PTR(-EWOULDBLOCK);
425 EXPORT_SYMBOL(blk_mq_alloc_request);
427 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
428 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
430 struct blk_mq_alloc_data data = {
435 u64 alloc_time_ns = 0;
440 /* alloc_time includes depth and tag waits */
441 if (blk_queue_rq_alloc_time(q))
442 alloc_time_ns = ktime_get_ns();
445 * If the tag allocator sleeps we could get an allocation for a
446 * different hardware context. No need to complicate the low level
447 * allocator for this for the rare use case of a command tied to
450 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
451 return ERR_PTR(-EINVAL);
453 if (hctx_idx >= q->nr_hw_queues)
454 return ERR_PTR(-EIO);
456 ret = blk_queue_enter(q, flags);
461 * Check if the hardware context is actually mapped to anything.
462 * If not tell the caller that it should skip this queue.
465 data.hctx = q->queue_hw_ctx[hctx_idx];
466 if (!blk_mq_hw_queue_mapped(data.hctx))
468 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
469 data.ctx = __blk_mq_get_ctx(q, cpu);
472 blk_mq_tag_busy(data.hctx);
475 tag = blk_mq_get_tag(&data);
476 if (tag == BLK_MQ_NO_TAG)
478 return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns);
484 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
486 static void __blk_mq_free_request(struct request *rq)
488 struct request_queue *q = rq->q;
489 struct blk_mq_ctx *ctx = rq->mq_ctx;
490 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
491 const int sched_tag = rq->internal_tag;
493 blk_crypto_free_request(rq);
494 blk_pm_mark_last_busy(rq);
496 if (rq->tag != BLK_MQ_NO_TAG)
497 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
498 if (sched_tag != BLK_MQ_NO_TAG)
499 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
500 blk_mq_sched_restart(hctx);
504 void blk_mq_free_request(struct request *rq)
506 struct request_queue *q = rq->q;
507 struct elevator_queue *e = q->elevator;
508 struct blk_mq_ctx *ctx = rq->mq_ctx;
509 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
511 if (rq->rq_flags & RQF_ELVPRIV) {
512 if (e && e->type->ops.finish_request)
513 e->type->ops.finish_request(rq);
515 put_io_context(rq->elv.icq->ioc);
520 ctx->rq_completed[rq_is_sync(rq)]++;
521 if (rq->rq_flags & RQF_MQ_INFLIGHT)
522 __blk_mq_dec_active_requests(hctx);
524 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
525 laptop_io_completion(q->backing_dev_info);
529 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
530 if (refcount_dec_and_test(&rq->ref))
531 __blk_mq_free_request(rq);
533 EXPORT_SYMBOL_GPL(blk_mq_free_request);
535 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
539 if (blk_mq_need_time_stamp(rq))
540 now = ktime_get_ns();
542 if (rq->rq_flags & RQF_STATS) {
543 blk_mq_poll_stats_start(rq->q);
544 blk_stat_add(rq, now);
547 blk_mq_sched_completed_request(rq, now);
549 blk_account_io_done(rq, now);
552 rq_qos_done(rq->q, rq);
553 rq->end_io(rq, error);
555 blk_mq_free_request(rq);
558 EXPORT_SYMBOL(__blk_mq_end_request);
560 void blk_mq_end_request(struct request *rq, blk_status_t error)
562 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
564 __blk_mq_end_request(rq, error);
566 EXPORT_SYMBOL(blk_mq_end_request);
569 * Softirq action handler - move entries to local list and loop over them
570 * while passing them to the queue registered handler.
572 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
574 struct list_head *cpu_list, local_list;
577 cpu_list = this_cpu_ptr(&blk_cpu_done);
578 list_replace_init(cpu_list, &local_list);
581 while (!list_empty(&local_list)) {
584 rq = list_entry(local_list.next, struct request, ipi_list);
585 list_del_init(&rq->ipi_list);
586 rq->q->mq_ops->complete(rq);
590 static void blk_mq_trigger_softirq(struct request *rq)
592 struct list_head *list;
595 local_irq_save(flags);
596 list = this_cpu_ptr(&blk_cpu_done);
597 list_add_tail(&rq->ipi_list, list);
600 * If the list only contains our just added request, signal a raise of
601 * the softirq. If there are already entries there, someone already
602 * raised the irq but it hasn't run yet.
604 if (list->next == &rq->ipi_list)
605 raise_softirq_irqoff(BLOCK_SOFTIRQ);
606 local_irq_restore(flags);
609 static int blk_softirq_cpu_dead(unsigned int cpu)
612 * If a CPU goes away, splice its entries to the current CPU
613 * and trigger a run of the softirq
616 list_splice_init(&per_cpu(blk_cpu_done, cpu),
617 this_cpu_ptr(&blk_cpu_done));
618 raise_softirq_irqoff(BLOCK_SOFTIRQ);
625 static void __blk_mq_complete_request_remote(void *data)
627 struct request *rq = data;
630 * For most of single queue controllers, there is only one irq vector
631 * for handling I/O completion, and the only irq's affinity is set
632 * to all possible CPUs. On most of ARCHs, this affinity means the irq
633 * is handled on one specific CPU.
635 * So complete I/O requests in softirq context in case of single queue
636 * devices to avoid degrading I/O performance due to irqsoff latency.
638 if (rq->q->nr_hw_queues == 1)
639 blk_mq_trigger_softirq(rq);
641 rq->q->mq_ops->complete(rq);
644 static inline bool blk_mq_complete_need_ipi(struct request *rq)
646 int cpu = raw_smp_processor_id();
648 if (!IS_ENABLED(CONFIG_SMP) ||
649 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
652 /* same CPU or cache domain? Complete locally */
653 if (cpu == rq->mq_ctx->cpu ||
654 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
655 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
658 /* don't try to IPI to an offline CPU */
659 return cpu_online(rq->mq_ctx->cpu);
662 bool blk_mq_complete_request_remote(struct request *rq)
664 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
667 * For a polled request, always complete locallly, it's pointless
668 * to redirect the completion.
670 if (rq->cmd_flags & REQ_HIPRI)
673 if (blk_mq_complete_need_ipi(rq)) {
674 rq->csd.func = __blk_mq_complete_request_remote;
677 smp_call_function_single_async(rq->mq_ctx->cpu, &rq->csd);
679 if (rq->q->nr_hw_queues > 1)
681 blk_mq_trigger_softirq(rq);
686 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
689 * blk_mq_complete_request - end I/O on a request
690 * @rq: the request being processed
693 * Complete a request by scheduling the ->complete_rq operation.
695 void blk_mq_complete_request(struct request *rq)
697 if (!blk_mq_complete_request_remote(rq))
698 rq->q->mq_ops->complete(rq);
700 EXPORT_SYMBOL(blk_mq_complete_request);
702 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
703 __releases(hctx->srcu)
705 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
708 srcu_read_unlock(hctx->srcu, srcu_idx);
711 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
712 __acquires(hctx->srcu)
714 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
715 /* shut up gcc false positive */
719 *srcu_idx = srcu_read_lock(hctx->srcu);
723 * blk_mq_start_request - Start processing a request
724 * @rq: Pointer to request to be started
726 * Function used by device drivers to notify the block layer that a request
727 * is going to be processed now, so blk layer can do proper initializations
728 * such as starting the timeout timer.
730 void blk_mq_start_request(struct request *rq)
732 struct request_queue *q = rq->q;
734 trace_block_rq_issue(q, rq);
736 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
737 rq->io_start_time_ns = ktime_get_ns();
738 rq->stats_sectors = blk_rq_sectors(rq);
739 rq->rq_flags |= RQF_STATS;
743 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
746 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
748 #ifdef CONFIG_BLK_DEV_INTEGRITY
749 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
750 q->integrity.profile->prepare_fn(rq);
753 EXPORT_SYMBOL(blk_mq_start_request);
755 static void __blk_mq_requeue_request(struct request *rq)
757 struct request_queue *q = rq->q;
759 blk_mq_put_driver_tag(rq);
761 trace_block_rq_requeue(q, rq);
762 rq_qos_requeue(q, rq);
764 if (blk_mq_request_started(rq)) {
765 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
766 rq->rq_flags &= ~RQF_TIMED_OUT;
770 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
772 __blk_mq_requeue_request(rq);
774 /* this request will be re-inserted to io scheduler queue */
775 blk_mq_sched_requeue_request(rq);
777 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
779 EXPORT_SYMBOL(blk_mq_requeue_request);
781 static void blk_mq_requeue_work(struct work_struct *work)
783 struct request_queue *q =
784 container_of(work, struct request_queue, requeue_work.work);
786 struct request *rq, *next;
788 spin_lock_irq(&q->requeue_lock);
789 list_splice_init(&q->requeue_list, &rq_list);
790 spin_unlock_irq(&q->requeue_lock);
792 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
793 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
796 rq->rq_flags &= ~RQF_SOFTBARRIER;
797 list_del_init(&rq->queuelist);
799 * If RQF_DONTPREP, rq has contained some driver specific
800 * data, so insert it to hctx dispatch list to avoid any
803 if (rq->rq_flags & RQF_DONTPREP)
804 blk_mq_request_bypass_insert(rq, false, false);
806 blk_mq_sched_insert_request(rq, true, false, false);
809 while (!list_empty(&rq_list)) {
810 rq = list_entry(rq_list.next, struct request, queuelist);
811 list_del_init(&rq->queuelist);
812 blk_mq_sched_insert_request(rq, false, false, false);
815 blk_mq_run_hw_queues(q, false);
818 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
819 bool kick_requeue_list)
821 struct request_queue *q = rq->q;
825 * We abuse this flag that is otherwise used by the I/O scheduler to
826 * request head insertion from the workqueue.
828 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
830 spin_lock_irqsave(&q->requeue_lock, flags);
832 rq->rq_flags |= RQF_SOFTBARRIER;
833 list_add(&rq->queuelist, &q->requeue_list);
835 list_add_tail(&rq->queuelist, &q->requeue_list);
837 spin_unlock_irqrestore(&q->requeue_lock, flags);
839 if (kick_requeue_list)
840 blk_mq_kick_requeue_list(q);
843 void blk_mq_kick_requeue_list(struct request_queue *q)
845 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
847 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
849 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
852 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
853 msecs_to_jiffies(msecs));
855 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
857 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
859 if (tag < tags->nr_tags) {
860 prefetch(tags->rqs[tag]);
861 return tags->rqs[tag];
866 EXPORT_SYMBOL(blk_mq_tag_to_rq);
868 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
869 void *priv, bool reserved)
872 * If we find a request that isn't idle and the queue matches,
873 * we know the queue is busy. Return false to stop the iteration.
875 if (blk_mq_request_started(rq) && rq->q == hctx->queue) {
885 bool blk_mq_queue_inflight(struct request_queue *q)
889 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
892 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
894 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
896 req->rq_flags |= RQF_TIMED_OUT;
897 if (req->q->mq_ops->timeout) {
898 enum blk_eh_timer_return ret;
900 ret = req->q->mq_ops->timeout(req, reserved);
901 if (ret == BLK_EH_DONE)
903 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
909 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
911 unsigned long deadline;
913 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
915 if (rq->rq_flags & RQF_TIMED_OUT)
918 deadline = READ_ONCE(rq->deadline);
919 if (time_after_eq(jiffies, deadline))
924 else if (time_after(*next, deadline))
929 void blk_mq_put_rq_ref(struct request *rq)
933 else if (refcount_dec_and_test(&rq->ref))
934 __blk_mq_free_request(rq);
937 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
938 struct request *rq, void *priv, bool reserved)
940 unsigned long *next = priv;
943 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
944 * be reallocated underneath the timeout handler's processing, then
945 * the expire check is reliable. If the request is not expired, then
946 * it was completed and reallocated as a new request after returning
947 * from blk_mq_check_expired().
949 if (blk_mq_req_expired(rq, next))
950 blk_mq_rq_timed_out(rq, reserved);
954 static void blk_mq_timeout_work(struct work_struct *work)
956 struct request_queue *q =
957 container_of(work, struct request_queue, timeout_work);
958 unsigned long next = 0;
959 struct blk_mq_hw_ctx *hctx;
962 /* A deadlock might occur if a request is stuck requiring a
963 * timeout at the same time a queue freeze is waiting
964 * completion, since the timeout code would not be able to
965 * acquire the queue reference here.
967 * That's why we don't use blk_queue_enter here; instead, we use
968 * percpu_ref_tryget directly, because we need to be able to
969 * obtain a reference even in the short window between the queue
970 * starting to freeze, by dropping the first reference in
971 * blk_freeze_queue_start, and the moment the last request is
972 * consumed, marked by the instant q_usage_counter reaches
975 if (!percpu_ref_tryget(&q->q_usage_counter))
978 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
981 mod_timer(&q->timeout, next);
984 * Request timeouts are handled as a forward rolling timer. If
985 * we end up here it means that no requests are pending and
986 * also that no request has been pending for a while. Mark
989 queue_for_each_hw_ctx(q, hctx, i) {
990 /* the hctx may be unmapped, so check it here */
991 if (blk_mq_hw_queue_mapped(hctx))
992 blk_mq_tag_idle(hctx);
998 struct flush_busy_ctx_data {
999 struct blk_mq_hw_ctx *hctx;
1000 struct list_head *list;
1003 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1005 struct flush_busy_ctx_data *flush_data = data;
1006 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1007 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1008 enum hctx_type type = hctx->type;
1010 spin_lock(&ctx->lock);
1011 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1012 sbitmap_clear_bit(sb, bitnr);
1013 spin_unlock(&ctx->lock);
1018 * Process software queues that have been marked busy, splicing them
1019 * to the for-dispatch
1021 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1023 struct flush_busy_ctx_data data = {
1028 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1030 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1032 struct dispatch_rq_data {
1033 struct blk_mq_hw_ctx *hctx;
1037 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1040 struct dispatch_rq_data *dispatch_data = data;
1041 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1042 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1043 enum hctx_type type = hctx->type;
1045 spin_lock(&ctx->lock);
1046 if (!list_empty(&ctx->rq_lists[type])) {
1047 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1048 list_del_init(&dispatch_data->rq->queuelist);
1049 if (list_empty(&ctx->rq_lists[type]))
1050 sbitmap_clear_bit(sb, bitnr);
1052 spin_unlock(&ctx->lock);
1054 return !dispatch_data->rq;
1057 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1058 struct blk_mq_ctx *start)
1060 unsigned off = start ? start->index_hw[hctx->type] : 0;
1061 struct dispatch_rq_data data = {
1066 __sbitmap_for_each_set(&hctx->ctx_map, off,
1067 dispatch_rq_from_ctx, &data);
1072 static inline unsigned int queued_to_index(unsigned int queued)
1077 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1080 static bool __blk_mq_get_driver_tag(struct request *rq)
1082 struct sbitmap_queue *bt = rq->mq_hctx->tags->bitmap_tags;
1083 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1086 blk_mq_tag_busy(rq->mq_hctx);
1088 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1089 bt = rq->mq_hctx->tags->breserved_tags;
1092 if (!hctx_may_queue(rq->mq_hctx, bt))
1096 tag = __sbitmap_queue_get(bt);
1097 if (tag == BLK_MQ_NO_TAG)
1100 rq->tag = tag + tag_offset;
1104 static bool blk_mq_get_driver_tag(struct request *rq)
1106 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1108 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq))
1111 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1112 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1113 rq->rq_flags |= RQF_MQ_INFLIGHT;
1114 __blk_mq_inc_active_requests(hctx);
1116 hctx->tags->rqs[rq->tag] = rq;
1120 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1121 int flags, void *key)
1123 struct blk_mq_hw_ctx *hctx;
1125 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1127 spin_lock(&hctx->dispatch_wait_lock);
1128 if (!list_empty(&wait->entry)) {
1129 struct sbitmap_queue *sbq;
1131 list_del_init(&wait->entry);
1132 sbq = hctx->tags->bitmap_tags;
1133 atomic_dec(&sbq->ws_active);
1135 spin_unlock(&hctx->dispatch_wait_lock);
1137 blk_mq_run_hw_queue(hctx, true);
1142 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1143 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1144 * restart. For both cases, take care to check the condition again after
1145 * marking us as waiting.
1147 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1150 struct sbitmap_queue *sbq = hctx->tags->bitmap_tags;
1151 struct wait_queue_head *wq;
1152 wait_queue_entry_t *wait;
1155 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1156 blk_mq_sched_mark_restart_hctx(hctx);
1159 * It's possible that a tag was freed in the window between the
1160 * allocation failure and adding the hardware queue to the wait
1163 * Don't clear RESTART here, someone else could have set it.
1164 * At most this will cost an extra queue run.
1166 return blk_mq_get_driver_tag(rq);
1169 wait = &hctx->dispatch_wait;
1170 if (!list_empty_careful(&wait->entry))
1173 wq = &bt_wait_ptr(sbq, hctx)->wait;
1175 spin_lock_irq(&wq->lock);
1176 spin_lock(&hctx->dispatch_wait_lock);
1177 if (!list_empty(&wait->entry)) {
1178 spin_unlock(&hctx->dispatch_wait_lock);
1179 spin_unlock_irq(&wq->lock);
1183 atomic_inc(&sbq->ws_active);
1184 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1185 __add_wait_queue(wq, wait);
1188 * It's possible that a tag was freed in the window between the
1189 * allocation failure and adding the hardware queue to the wait
1192 ret = blk_mq_get_driver_tag(rq);
1194 spin_unlock(&hctx->dispatch_wait_lock);
1195 spin_unlock_irq(&wq->lock);
1200 * We got a tag, remove ourselves from the wait queue to ensure
1201 * someone else gets the wakeup.
1203 list_del_init(&wait->entry);
1204 atomic_dec(&sbq->ws_active);
1205 spin_unlock(&hctx->dispatch_wait_lock);
1206 spin_unlock_irq(&wq->lock);
1211 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1212 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1214 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1215 * - EWMA is one simple way to compute running average value
1216 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1217 * - take 4 as factor for avoiding to get too small(0) result, and this
1218 * factor doesn't matter because EWMA decreases exponentially
1220 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1224 ewma = hctx->dispatch_busy;
1229 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1231 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1232 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1234 hctx->dispatch_busy = ewma;
1237 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1239 static void blk_mq_handle_dev_resource(struct request *rq,
1240 struct list_head *list)
1242 struct request *next =
1243 list_first_entry_or_null(list, struct request, queuelist);
1246 * If an I/O scheduler has been configured and we got a driver tag for
1247 * the next request already, free it.
1250 blk_mq_put_driver_tag(next);
1252 list_add(&rq->queuelist, list);
1253 __blk_mq_requeue_request(rq);
1256 static void blk_mq_handle_zone_resource(struct request *rq,
1257 struct list_head *zone_list)
1260 * If we end up here it is because we cannot dispatch a request to a
1261 * specific zone due to LLD level zone-write locking or other zone
1262 * related resource not being available. In this case, set the request
1263 * aside in zone_list for retrying it later.
1265 list_add(&rq->queuelist, zone_list);
1266 __blk_mq_requeue_request(rq);
1269 enum prep_dispatch {
1271 PREP_DISPATCH_NO_TAG,
1272 PREP_DISPATCH_NO_BUDGET,
1275 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1278 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1280 if (need_budget && !blk_mq_get_dispatch_budget(rq->q)) {
1281 blk_mq_put_driver_tag(rq);
1282 return PREP_DISPATCH_NO_BUDGET;
1285 if (!blk_mq_get_driver_tag(rq)) {
1287 * The initial allocation attempt failed, so we need to
1288 * rerun the hardware queue when a tag is freed. The
1289 * waitqueue takes care of that. If the queue is run
1290 * before we add this entry back on the dispatch list,
1291 * we'll re-run it below.
1293 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1295 * All budgets not got from this function will be put
1296 * together during handling partial dispatch
1299 blk_mq_put_dispatch_budget(rq->q);
1300 return PREP_DISPATCH_NO_TAG;
1304 return PREP_DISPATCH_OK;
1307 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1308 static void blk_mq_release_budgets(struct request_queue *q,
1309 unsigned int nr_budgets)
1313 for (i = 0; i < nr_budgets; i++)
1314 blk_mq_put_dispatch_budget(q);
1318 * Returns true if we did some work AND can potentially do more.
1320 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1321 unsigned int nr_budgets)
1323 enum prep_dispatch prep;
1324 struct request_queue *q = hctx->queue;
1325 struct request *rq, *nxt;
1327 blk_status_t ret = BLK_STS_OK;
1328 LIST_HEAD(zone_list);
1329 bool needs_resource = false;
1331 if (list_empty(list))
1335 * Now process all the entries, sending them to the driver.
1337 errors = queued = 0;
1339 struct blk_mq_queue_data bd;
1341 rq = list_first_entry(list, struct request, queuelist);
1343 WARN_ON_ONCE(hctx != rq->mq_hctx);
1344 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1345 if (prep != PREP_DISPATCH_OK)
1348 list_del_init(&rq->queuelist);
1353 * Flag last if we have no more requests, or if we have more
1354 * but can't assign a driver tag to it.
1356 if (list_empty(list))
1359 nxt = list_first_entry(list, struct request, queuelist);
1360 bd.last = !blk_mq_get_driver_tag(nxt);
1364 * once the request is queued to lld, no need to cover the
1369 ret = q->mq_ops->queue_rq(hctx, &bd);
1374 case BLK_STS_RESOURCE:
1375 needs_resource = true;
1377 case BLK_STS_DEV_RESOURCE:
1378 blk_mq_handle_dev_resource(rq, list);
1380 case BLK_STS_ZONE_RESOURCE:
1382 * Move the request to zone_list and keep going through
1383 * the dispatch list to find more requests the drive can
1386 blk_mq_handle_zone_resource(rq, &zone_list);
1387 needs_resource = true;
1391 blk_mq_end_request(rq, BLK_STS_IOERR);
1393 } while (!list_empty(list));
1395 if (!list_empty(&zone_list))
1396 list_splice_tail_init(&zone_list, list);
1398 hctx->dispatched[queued_to_index(queued)]++;
1400 /* If we didn't flush the entire list, we could have told the driver
1401 * there was more coming, but that turned out to be a lie.
1403 if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued)
1404 q->mq_ops->commit_rqs(hctx);
1406 * Any items that need requeuing? Stuff them into hctx->dispatch,
1407 * that is where we will continue on next queue run.
1409 if (!list_empty(list)) {
1411 /* For non-shared tags, the RESTART check will suffice */
1412 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1413 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1415 blk_mq_release_budgets(q, nr_budgets);
1417 spin_lock(&hctx->lock);
1418 list_splice_tail_init(list, &hctx->dispatch);
1419 spin_unlock(&hctx->lock);
1422 * Order adding requests to hctx->dispatch and checking
1423 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1424 * in blk_mq_sched_restart(). Avoid restart code path to
1425 * miss the new added requests to hctx->dispatch, meantime
1426 * SCHED_RESTART is observed here.
1431 * If SCHED_RESTART was set by the caller of this function and
1432 * it is no longer set that means that it was cleared by another
1433 * thread and hence that a queue rerun is needed.
1435 * If 'no_tag' is set, that means that we failed getting
1436 * a driver tag with an I/O scheduler attached. If our dispatch
1437 * waitqueue is no longer active, ensure that we run the queue
1438 * AFTER adding our entries back to the list.
1440 * If no I/O scheduler has been configured it is possible that
1441 * the hardware queue got stopped and restarted before requests
1442 * were pushed back onto the dispatch list. Rerun the queue to
1443 * avoid starvation. Notes:
1444 * - blk_mq_run_hw_queue() checks whether or not a queue has
1445 * been stopped before rerunning a queue.
1446 * - Some but not all block drivers stop a queue before
1447 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1450 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1451 * bit is set, run queue after a delay to avoid IO stalls
1452 * that could otherwise occur if the queue is idle. We'll do
1453 * similar if we couldn't get budget or couldn't lock a zone
1454 * and SCHED_RESTART is set.
1456 needs_restart = blk_mq_sched_needs_restart(hctx);
1457 if (prep == PREP_DISPATCH_NO_BUDGET)
1458 needs_resource = true;
1459 if (!needs_restart ||
1460 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1461 blk_mq_run_hw_queue(hctx, true);
1462 else if (needs_restart && needs_resource)
1463 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1465 blk_mq_update_dispatch_busy(hctx, true);
1468 blk_mq_update_dispatch_busy(hctx, false);
1470 return (queued + errors) != 0;
1474 * __blk_mq_run_hw_queue - Run a hardware queue.
1475 * @hctx: Pointer to the hardware queue to run.
1477 * Send pending requests to the hardware.
1479 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1484 * We should be running this queue from one of the CPUs that
1487 * There are at least two related races now between setting
1488 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1489 * __blk_mq_run_hw_queue():
1491 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1492 * but later it becomes online, then this warning is harmless
1495 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1496 * but later it becomes offline, then the warning can't be
1497 * triggered, and we depend on blk-mq timeout handler to
1498 * handle dispatched requests to this hctx
1500 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1501 cpu_online(hctx->next_cpu)) {
1502 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1503 raw_smp_processor_id(),
1504 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1509 * We can't run the queue inline with ints disabled. Ensure that
1510 * we catch bad users of this early.
1512 WARN_ON_ONCE(in_interrupt());
1514 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1516 hctx_lock(hctx, &srcu_idx);
1517 blk_mq_sched_dispatch_requests(hctx);
1518 hctx_unlock(hctx, srcu_idx);
1521 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1523 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1525 if (cpu >= nr_cpu_ids)
1526 cpu = cpumask_first(hctx->cpumask);
1531 * It'd be great if the workqueue API had a way to pass
1532 * in a mask and had some smarts for more clever placement.
1533 * For now we just round-robin here, switching for every
1534 * BLK_MQ_CPU_WORK_BATCH queued items.
1536 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1539 int next_cpu = hctx->next_cpu;
1541 if (hctx->queue->nr_hw_queues == 1)
1542 return WORK_CPU_UNBOUND;
1544 if (--hctx->next_cpu_batch <= 0) {
1546 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1548 if (next_cpu >= nr_cpu_ids)
1549 next_cpu = blk_mq_first_mapped_cpu(hctx);
1550 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1554 * Do unbound schedule if we can't find a online CPU for this hctx,
1555 * and it should only happen in the path of handling CPU DEAD.
1557 if (!cpu_online(next_cpu)) {
1564 * Make sure to re-select CPU next time once after CPUs
1565 * in hctx->cpumask become online again.
1567 hctx->next_cpu = next_cpu;
1568 hctx->next_cpu_batch = 1;
1569 return WORK_CPU_UNBOUND;
1572 hctx->next_cpu = next_cpu;
1577 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1578 * @hctx: Pointer to the hardware queue to run.
1579 * @async: If we want to run the queue asynchronously.
1580 * @msecs: Microseconds of delay to wait before running the queue.
1582 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1583 * with a delay of @msecs.
1585 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1586 unsigned long msecs)
1588 if (unlikely(blk_mq_hctx_stopped(hctx)))
1591 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1592 int cpu = get_cpu();
1593 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1594 __blk_mq_run_hw_queue(hctx);
1602 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1603 msecs_to_jiffies(msecs));
1607 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1608 * @hctx: Pointer to the hardware queue to run.
1609 * @msecs: Microseconds of delay to wait before running the queue.
1611 * Run a hardware queue asynchronously with a delay of @msecs.
1613 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1615 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1617 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1620 * blk_mq_run_hw_queue - Start to run a hardware queue.
1621 * @hctx: Pointer to the hardware queue to run.
1622 * @async: If we want to run the queue asynchronously.
1624 * Check if the request queue is not in a quiesced state and if there are
1625 * pending requests to be sent. If this is true, run the queue to send requests
1628 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1634 * When queue is quiesced, we may be switching io scheduler, or
1635 * updating nr_hw_queues, or other things, and we can't run queue
1636 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1638 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1641 hctx_lock(hctx, &srcu_idx);
1642 need_run = !blk_queue_quiesced(hctx->queue) &&
1643 blk_mq_hctx_has_pending(hctx);
1644 hctx_unlock(hctx, srcu_idx);
1647 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1649 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1652 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
1653 * @q: Pointer to the request queue to run.
1654 * @async: If we want to run the queue asynchronously.
1656 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1658 struct blk_mq_hw_ctx *hctx;
1661 queue_for_each_hw_ctx(q, hctx, i) {
1662 if (blk_mq_hctx_stopped(hctx))
1665 blk_mq_run_hw_queue(hctx, async);
1668 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1671 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1672 * @q: Pointer to the request queue to run.
1673 * @msecs: Microseconds of delay to wait before running the queues.
1675 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1677 struct blk_mq_hw_ctx *hctx;
1680 queue_for_each_hw_ctx(q, hctx, i) {
1681 if (blk_mq_hctx_stopped(hctx))
1684 blk_mq_delay_run_hw_queue(hctx, msecs);
1687 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1690 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1691 * @q: request queue.
1693 * The caller is responsible for serializing this function against
1694 * blk_mq_{start,stop}_hw_queue().
1696 bool blk_mq_queue_stopped(struct request_queue *q)
1698 struct blk_mq_hw_ctx *hctx;
1701 queue_for_each_hw_ctx(q, hctx, i)
1702 if (blk_mq_hctx_stopped(hctx))
1707 EXPORT_SYMBOL(blk_mq_queue_stopped);
1710 * This function is often used for pausing .queue_rq() by driver when
1711 * there isn't enough resource or some conditions aren't satisfied, and
1712 * BLK_STS_RESOURCE is usually returned.
1714 * We do not guarantee that dispatch can be drained or blocked
1715 * after blk_mq_stop_hw_queue() returns. Please use
1716 * blk_mq_quiesce_queue() for that requirement.
1718 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1720 cancel_delayed_work(&hctx->run_work);
1722 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1724 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1727 * This function is often used for pausing .queue_rq() by driver when
1728 * there isn't enough resource or some conditions aren't satisfied, and
1729 * BLK_STS_RESOURCE is usually returned.
1731 * We do not guarantee that dispatch can be drained or blocked
1732 * after blk_mq_stop_hw_queues() returns. Please use
1733 * blk_mq_quiesce_queue() for that requirement.
1735 void blk_mq_stop_hw_queues(struct request_queue *q)
1737 struct blk_mq_hw_ctx *hctx;
1740 queue_for_each_hw_ctx(q, hctx, i)
1741 blk_mq_stop_hw_queue(hctx);
1743 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1745 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1747 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1749 blk_mq_run_hw_queue(hctx, false);
1751 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1753 void blk_mq_start_hw_queues(struct request_queue *q)
1755 struct blk_mq_hw_ctx *hctx;
1758 queue_for_each_hw_ctx(q, hctx, i)
1759 blk_mq_start_hw_queue(hctx);
1761 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1763 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1765 if (!blk_mq_hctx_stopped(hctx))
1768 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1769 blk_mq_run_hw_queue(hctx, async);
1771 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1773 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1775 struct blk_mq_hw_ctx *hctx;
1778 queue_for_each_hw_ctx(q, hctx, i)
1779 blk_mq_start_stopped_hw_queue(hctx, async);
1781 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1783 static void blk_mq_run_work_fn(struct work_struct *work)
1785 struct blk_mq_hw_ctx *hctx;
1787 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1790 * If we are stopped, don't run the queue.
1792 if (blk_mq_hctx_stopped(hctx))
1795 __blk_mq_run_hw_queue(hctx);
1798 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1802 struct blk_mq_ctx *ctx = rq->mq_ctx;
1803 enum hctx_type type = hctx->type;
1805 lockdep_assert_held(&ctx->lock);
1807 trace_block_rq_insert(hctx->queue, rq);
1810 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1812 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1815 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1818 struct blk_mq_ctx *ctx = rq->mq_ctx;
1820 lockdep_assert_held(&ctx->lock);
1822 __blk_mq_insert_req_list(hctx, rq, at_head);
1823 blk_mq_hctx_mark_pending(hctx, ctx);
1827 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1828 * @rq: Pointer to request to be inserted.
1829 * @at_head: true if the request should be inserted at the head of the list.
1830 * @run_queue: If we should run the hardware queue after inserting the request.
1832 * Should only be used carefully, when the caller knows we want to
1833 * bypass a potential IO scheduler on the target device.
1835 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1838 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1840 spin_lock(&hctx->lock);
1842 list_add(&rq->queuelist, &hctx->dispatch);
1844 list_add_tail(&rq->queuelist, &hctx->dispatch);
1845 spin_unlock(&hctx->lock);
1848 blk_mq_run_hw_queue(hctx, false);
1851 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1852 struct list_head *list)
1856 enum hctx_type type = hctx->type;
1859 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1862 list_for_each_entry(rq, list, queuelist) {
1863 BUG_ON(rq->mq_ctx != ctx);
1864 trace_block_rq_insert(hctx->queue, rq);
1867 spin_lock(&ctx->lock);
1868 list_splice_tail_init(list, &ctx->rq_lists[type]);
1869 blk_mq_hctx_mark_pending(hctx, ctx);
1870 spin_unlock(&ctx->lock);
1873 static int plug_rq_cmp(void *priv, const struct list_head *a,
1874 const struct list_head *b)
1876 struct request *rqa = container_of(a, struct request, queuelist);
1877 struct request *rqb = container_of(b, struct request, queuelist);
1879 if (rqa->mq_ctx != rqb->mq_ctx)
1880 return rqa->mq_ctx > rqb->mq_ctx;
1881 if (rqa->mq_hctx != rqb->mq_hctx)
1882 return rqa->mq_hctx > rqb->mq_hctx;
1884 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1887 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1891 if (list_empty(&plug->mq_list))
1893 list_splice_init(&plug->mq_list, &list);
1895 if (plug->rq_count > 2 && plug->multiple_queues)
1896 list_sort(NULL, &list, plug_rq_cmp);
1901 struct list_head rq_list;
1902 struct request *rq, *head_rq = list_entry_rq(list.next);
1903 struct list_head *pos = &head_rq->queuelist; /* skip first */
1904 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1905 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1906 unsigned int depth = 1;
1908 list_for_each_continue(pos, &list) {
1909 rq = list_entry_rq(pos);
1911 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1916 list_cut_before(&rq_list, &list, pos);
1917 trace_block_unplug(head_rq->q, depth, !from_schedule);
1918 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1920 } while(!list_empty(&list));
1923 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1924 unsigned int nr_segs)
1928 if (bio->bi_opf & REQ_RAHEAD)
1929 rq->cmd_flags |= REQ_FAILFAST_MASK;
1931 rq->__sector = bio->bi_iter.bi_sector;
1932 rq->write_hint = bio->bi_write_hint;
1933 blk_rq_bio_prep(rq, bio, nr_segs);
1935 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
1936 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1939 blk_account_io_start(rq);
1942 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1944 blk_qc_t *cookie, bool last)
1946 struct request_queue *q = rq->q;
1947 struct blk_mq_queue_data bd = {
1951 blk_qc_t new_cookie;
1954 new_cookie = request_to_qc_t(hctx, rq);
1957 * For OK queue, we are done. For error, caller may kill it.
1958 * Any other error (busy), just add it to our list as we
1959 * previously would have done.
1961 ret = q->mq_ops->queue_rq(hctx, &bd);
1964 blk_mq_update_dispatch_busy(hctx, false);
1965 *cookie = new_cookie;
1967 case BLK_STS_RESOURCE:
1968 case BLK_STS_DEV_RESOURCE:
1969 blk_mq_update_dispatch_busy(hctx, true);
1970 __blk_mq_requeue_request(rq);
1973 blk_mq_update_dispatch_busy(hctx, false);
1974 *cookie = BLK_QC_T_NONE;
1981 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1984 bool bypass_insert, bool last)
1986 struct request_queue *q = rq->q;
1987 bool run_queue = true;
1990 * RCU or SRCU read lock is needed before checking quiesced flag.
1992 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1993 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1994 * and avoid driver to try to dispatch again.
1996 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1998 bypass_insert = false;
2002 if (q->elevator && !bypass_insert)
2005 if (!blk_mq_get_dispatch_budget(q))
2008 if (!blk_mq_get_driver_tag(rq)) {
2009 blk_mq_put_dispatch_budget(q);
2013 return __blk_mq_issue_directly(hctx, rq, cookie, last);
2016 return BLK_STS_RESOURCE;
2018 blk_mq_sched_insert_request(rq, false, run_queue, false);
2024 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2025 * @hctx: Pointer of the associated hardware queue.
2026 * @rq: Pointer to request to be sent.
2027 * @cookie: Request queue cookie.
2029 * If the device has enough resources to accept a new request now, send the
2030 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2031 * we can try send it another time in the future. Requests inserted at this
2032 * queue have higher priority.
2034 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2035 struct request *rq, blk_qc_t *cookie)
2040 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2042 hctx_lock(hctx, &srcu_idx);
2044 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
2045 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2046 blk_mq_request_bypass_insert(rq, false, true);
2047 else if (ret != BLK_STS_OK)
2048 blk_mq_end_request(rq, ret);
2050 hctx_unlock(hctx, srcu_idx);
2053 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2057 blk_qc_t unused_cookie;
2058 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2060 hctx_lock(hctx, &srcu_idx);
2061 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
2062 hctx_unlock(hctx, srcu_idx);
2067 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2068 struct list_head *list)
2073 while (!list_empty(list)) {
2075 struct request *rq = list_first_entry(list, struct request,
2078 list_del_init(&rq->queuelist);
2079 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2080 if (ret != BLK_STS_OK) {
2081 if (ret == BLK_STS_RESOURCE ||
2082 ret == BLK_STS_DEV_RESOURCE) {
2083 blk_mq_request_bypass_insert(rq, false,
2087 blk_mq_end_request(rq, ret);
2094 * If we didn't flush the entire list, we could have told
2095 * the driver there was more coming, but that turned out to
2098 if ((!list_empty(list) || errors) &&
2099 hctx->queue->mq_ops->commit_rqs && queued)
2100 hctx->queue->mq_ops->commit_rqs(hctx);
2103 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2105 list_add_tail(&rq->queuelist, &plug->mq_list);
2107 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
2108 struct request *tmp;
2110 tmp = list_first_entry(&plug->mq_list, struct request,
2112 if (tmp->q != rq->q)
2113 plug->multiple_queues = true;
2118 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
2119 * queues. This is important for md arrays to benefit from merging
2122 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
2124 if (plug->multiple_queues)
2125 return BLK_MAX_REQUEST_COUNT * 2;
2126 return BLK_MAX_REQUEST_COUNT;
2130 * blk_mq_submit_bio - Create and send a request to block device.
2131 * @bio: Bio pointer.
2133 * Builds up a request structure from @q and @bio and send to the device. The
2134 * request may not be queued directly to hardware if:
2135 * * This request can be merged with another one
2136 * * We want to place request at plug queue for possible future merging
2137 * * There is an IO scheduler active at this queue
2139 * It will not queue the request if there is an error with the bio, or at the
2142 * Returns: Request queue cookie.
2144 blk_qc_t blk_mq_submit_bio(struct bio *bio)
2146 struct request_queue *q = bio->bi_disk->queue;
2147 const int is_sync = op_is_sync(bio->bi_opf);
2148 const int is_flush_fua = op_is_flush(bio->bi_opf);
2149 struct blk_mq_alloc_data data = {
2153 struct blk_plug *plug;
2154 struct request *same_queue_rq = NULL;
2155 unsigned int nr_segs;
2159 blk_queue_bounce(q, &bio);
2160 __blk_queue_split(&bio, &nr_segs);
2162 if (!bio_integrity_prep(bio))
2165 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2166 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2169 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2172 rq_qos_throttle(q, bio);
2174 data.cmd_flags = bio->bi_opf;
2175 rq = __blk_mq_alloc_request(&data);
2176 if (unlikely(!rq)) {
2177 rq_qos_cleanup(q, bio);
2178 if (bio->bi_opf & REQ_NOWAIT)
2179 bio_wouldblock_error(bio);
2183 trace_block_getrq(q, bio, bio->bi_opf);
2185 rq_qos_track(q, rq, bio);
2187 cookie = request_to_qc_t(data.hctx, rq);
2189 blk_mq_bio_to_request(rq, bio, nr_segs);
2191 ret = blk_crypto_init_request(rq);
2192 if (ret != BLK_STS_OK) {
2193 bio->bi_status = ret;
2195 blk_mq_free_request(rq);
2196 return BLK_QC_T_NONE;
2199 plug = blk_mq_plug(q, bio);
2200 if (unlikely(is_flush_fua)) {
2201 /* Bypass scheduler for flush requests */
2202 blk_insert_flush(rq);
2203 blk_mq_run_hw_queue(data.hctx, true);
2204 } else if (plug && (q->nr_hw_queues == 1 ||
2205 blk_mq_is_sbitmap_shared(rq->mq_hctx->flags) ||
2206 q->mq_ops->commit_rqs || !blk_queue_nonrot(q))) {
2208 * Use plugging if we have a ->commit_rqs() hook as well, as
2209 * we know the driver uses bd->last in a smart fashion.
2211 * Use normal plugging if this disk is slow HDD, as sequential
2212 * IO may benefit a lot from plug merging.
2214 unsigned int request_count = plug->rq_count;
2215 struct request *last = NULL;
2218 trace_block_plug(q);
2220 last = list_entry_rq(plug->mq_list.prev);
2222 if (request_count >= blk_plug_max_rq_count(plug) || (last &&
2223 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2224 blk_flush_plug_list(plug, false);
2225 trace_block_plug(q);
2228 blk_add_rq_to_plug(plug, rq);
2229 } else if (q->elevator) {
2230 /* Insert the request at the IO scheduler queue */
2231 blk_mq_sched_insert_request(rq, false, true, true);
2232 } else if (plug && !blk_queue_nomerges(q)) {
2234 * We do limited plugging. If the bio can be merged, do that.
2235 * Otherwise the existing request in the plug list will be
2236 * issued. So the plug list will have one request at most
2237 * The plug list might get flushed before this. If that happens,
2238 * the plug list is empty, and same_queue_rq is invalid.
2240 if (list_empty(&plug->mq_list))
2241 same_queue_rq = NULL;
2242 if (same_queue_rq) {
2243 list_del_init(&same_queue_rq->queuelist);
2246 blk_add_rq_to_plug(plug, rq);
2247 trace_block_plug(q);
2249 if (same_queue_rq) {
2250 data.hctx = same_queue_rq->mq_hctx;
2251 trace_block_unplug(q, 1, true);
2252 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2255 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2256 !data.hctx->dispatch_busy) {
2258 * There is no scheduler and we can try to send directly
2261 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2264 blk_mq_sched_insert_request(rq, false, true, true);
2270 return BLK_QC_T_NONE;
2273 static size_t order_to_size(unsigned int order)
2275 return (size_t)PAGE_SIZE << order;
2278 /* called before freeing request pool in @tags */
2279 static void blk_mq_clear_rq_mapping(struct blk_mq_tag_set *set,
2280 struct blk_mq_tags *tags, unsigned int hctx_idx)
2282 struct blk_mq_tags *drv_tags = set->tags[hctx_idx];
2284 unsigned long flags;
2286 list_for_each_entry(page, &tags->page_list, lru) {
2287 unsigned long start = (unsigned long)page_address(page);
2288 unsigned long end = start + order_to_size(page->private);
2291 for (i = 0; i < set->queue_depth; i++) {
2292 struct request *rq = drv_tags->rqs[i];
2293 unsigned long rq_addr = (unsigned long)rq;
2295 if (rq_addr >= start && rq_addr < end) {
2296 WARN_ON_ONCE(refcount_read(&rq->ref) != 0);
2297 cmpxchg(&drv_tags->rqs[i], rq, NULL);
2303 * Wait until all pending iteration is done.
2305 * Request reference is cleared and it is guaranteed to be observed
2306 * after the ->lock is released.
2308 spin_lock_irqsave(&drv_tags->lock, flags);
2309 spin_unlock_irqrestore(&drv_tags->lock, flags);
2312 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2313 unsigned int hctx_idx)
2317 if (tags->rqs && set->ops->exit_request) {
2320 for (i = 0; i < tags->nr_tags; i++) {
2321 struct request *rq = tags->static_rqs[i];
2325 set->ops->exit_request(set, rq, hctx_idx);
2326 tags->static_rqs[i] = NULL;
2330 blk_mq_clear_rq_mapping(set, tags, hctx_idx);
2332 while (!list_empty(&tags->page_list)) {
2333 page = list_first_entry(&tags->page_list, struct page, lru);
2334 list_del_init(&page->lru);
2336 * Remove kmemleak object previously allocated in
2337 * blk_mq_alloc_rqs().
2339 kmemleak_free(page_address(page));
2340 __free_pages(page, page->private);
2344 void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags)
2348 kfree(tags->static_rqs);
2349 tags->static_rqs = NULL;
2351 blk_mq_free_tags(tags, flags);
2354 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2355 unsigned int hctx_idx,
2356 unsigned int nr_tags,
2357 unsigned int reserved_tags,
2360 struct blk_mq_tags *tags;
2363 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2364 if (node == NUMA_NO_NODE)
2365 node = set->numa_node;
2367 tags = blk_mq_init_tags(nr_tags, reserved_tags, node, flags);
2371 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2372 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2375 blk_mq_free_tags(tags, flags);
2379 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2380 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2382 if (!tags->static_rqs) {
2384 blk_mq_free_tags(tags, flags);
2391 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2392 unsigned int hctx_idx, int node)
2396 if (set->ops->init_request) {
2397 ret = set->ops->init_request(set, rq, hctx_idx, node);
2402 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2406 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2407 unsigned int hctx_idx, unsigned int depth)
2409 unsigned int i, j, entries_per_page, max_order = 4;
2410 size_t rq_size, left;
2413 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2414 if (node == NUMA_NO_NODE)
2415 node = set->numa_node;
2417 INIT_LIST_HEAD(&tags->page_list);
2420 * rq_size is the size of the request plus driver payload, rounded
2421 * to the cacheline size
2423 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2425 left = rq_size * depth;
2427 for (i = 0; i < depth; ) {
2428 int this_order = max_order;
2433 while (this_order && left < order_to_size(this_order - 1))
2437 page = alloc_pages_node(node,
2438 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2444 if (order_to_size(this_order) < rq_size)
2451 page->private = this_order;
2452 list_add_tail(&page->lru, &tags->page_list);
2454 p = page_address(page);
2456 * Allow kmemleak to scan these pages as they contain pointers
2457 * to additional allocations like via ops->init_request().
2459 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2460 entries_per_page = order_to_size(this_order) / rq_size;
2461 to_do = min(entries_per_page, depth - i);
2462 left -= to_do * rq_size;
2463 for (j = 0; j < to_do; j++) {
2464 struct request *rq = p;
2466 tags->static_rqs[i] = rq;
2467 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2468 tags->static_rqs[i] = NULL;
2479 blk_mq_free_rqs(set, tags, hctx_idx);
2483 struct rq_iter_data {
2484 struct blk_mq_hw_ctx *hctx;
2488 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2490 struct rq_iter_data *iter_data = data;
2492 if (rq->mq_hctx != iter_data->hctx)
2494 iter_data->has_rq = true;
2498 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2500 struct blk_mq_tags *tags = hctx->sched_tags ?
2501 hctx->sched_tags : hctx->tags;
2502 struct rq_iter_data data = {
2506 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2510 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2511 struct blk_mq_hw_ctx *hctx)
2513 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2515 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2520 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2522 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2523 struct blk_mq_hw_ctx, cpuhp_online);
2525 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2526 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2530 * Prevent new request from being allocated on the current hctx.
2532 * The smp_mb__after_atomic() Pairs with the implied barrier in
2533 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2534 * seen once we return from the tag allocator.
2536 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2537 smp_mb__after_atomic();
2540 * Try to grab a reference to the queue and wait for any outstanding
2541 * requests. If we could not grab a reference the queue has been
2542 * frozen and there are no requests.
2544 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2545 while (blk_mq_hctx_has_requests(hctx))
2547 percpu_ref_put(&hctx->queue->q_usage_counter);
2553 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2555 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2556 struct blk_mq_hw_ctx, cpuhp_online);
2558 if (cpumask_test_cpu(cpu, hctx->cpumask))
2559 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2564 * 'cpu' is going away. splice any existing rq_list entries from this
2565 * software queue to the hw queue dispatch list, and ensure that it
2568 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2570 struct blk_mq_hw_ctx *hctx;
2571 struct blk_mq_ctx *ctx;
2573 enum hctx_type type;
2575 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2576 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2579 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2582 spin_lock(&ctx->lock);
2583 if (!list_empty(&ctx->rq_lists[type])) {
2584 list_splice_init(&ctx->rq_lists[type], &tmp);
2585 blk_mq_hctx_clear_pending(hctx, ctx);
2587 spin_unlock(&ctx->lock);
2589 if (list_empty(&tmp))
2592 spin_lock(&hctx->lock);
2593 list_splice_tail_init(&tmp, &hctx->dispatch);
2594 spin_unlock(&hctx->lock);
2596 blk_mq_run_hw_queue(hctx, true);
2600 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2602 if (!(hctx->flags & BLK_MQ_F_STACKING))
2603 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2604 &hctx->cpuhp_online);
2605 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2610 * Before freeing hw queue, clearing the flush request reference in
2611 * tags->rqs[] for avoiding potential UAF.
2613 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
2614 unsigned int queue_depth, struct request *flush_rq)
2617 unsigned long flags;
2619 /* The hw queue may not be mapped yet */
2623 WARN_ON_ONCE(refcount_read(&flush_rq->ref) != 0);
2625 for (i = 0; i < queue_depth; i++)
2626 cmpxchg(&tags->rqs[i], flush_rq, NULL);
2629 * Wait until all pending iteration is done.
2631 * Request reference is cleared and it is guaranteed to be observed
2632 * after the ->lock is released.
2634 spin_lock_irqsave(&tags->lock, flags);
2635 spin_unlock_irqrestore(&tags->lock, flags);
2638 /* hctx->ctxs will be freed in queue's release handler */
2639 static void blk_mq_exit_hctx(struct request_queue *q,
2640 struct blk_mq_tag_set *set,
2641 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2643 struct request *flush_rq = hctx->fq->flush_rq;
2645 if (blk_mq_hw_queue_mapped(hctx))
2646 blk_mq_tag_idle(hctx);
2648 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
2649 set->queue_depth, flush_rq);
2650 if (set->ops->exit_request)
2651 set->ops->exit_request(set, flush_rq, hctx_idx);
2653 if (set->ops->exit_hctx)
2654 set->ops->exit_hctx(hctx, hctx_idx);
2656 blk_mq_remove_cpuhp(hctx);
2658 spin_lock(&q->unused_hctx_lock);
2659 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2660 spin_unlock(&q->unused_hctx_lock);
2663 static void blk_mq_exit_hw_queues(struct request_queue *q,
2664 struct blk_mq_tag_set *set, int nr_queue)
2666 struct blk_mq_hw_ctx *hctx;
2669 queue_for_each_hw_ctx(q, hctx, i) {
2672 blk_mq_debugfs_unregister_hctx(hctx);
2673 blk_mq_exit_hctx(q, set, hctx, i);
2677 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2679 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2681 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2682 __alignof__(struct blk_mq_hw_ctx)) !=
2683 sizeof(struct blk_mq_hw_ctx));
2685 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2686 hw_ctx_size += sizeof(struct srcu_struct);
2691 static int blk_mq_init_hctx(struct request_queue *q,
2692 struct blk_mq_tag_set *set,
2693 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2695 hctx->queue_num = hctx_idx;
2697 if (!(hctx->flags & BLK_MQ_F_STACKING))
2698 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2699 &hctx->cpuhp_online);
2700 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2702 hctx->tags = set->tags[hctx_idx];
2704 if (set->ops->init_hctx &&
2705 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2706 goto unregister_cpu_notifier;
2708 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2714 if (set->ops->exit_hctx)
2715 set->ops->exit_hctx(hctx, hctx_idx);
2716 unregister_cpu_notifier:
2717 blk_mq_remove_cpuhp(hctx);
2721 static struct blk_mq_hw_ctx *
2722 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2725 struct blk_mq_hw_ctx *hctx;
2726 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2728 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2730 goto fail_alloc_hctx;
2732 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2735 atomic_set(&hctx->nr_active, 0);
2736 atomic_set(&hctx->elevator_queued, 0);
2737 if (node == NUMA_NO_NODE)
2738 node = set->numa_node;
2739 hctx->numa_node = node;
2741 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2742 spin_lock_init(&hctx->lock);
2743 INIT_LIST_HEAD(&hctx->dispatch);
2745 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
2747 INIT_LIST_HEAD(&hctx->hctx_list);
2750 * Allocate space for all possible cpus to avoid allocation at
2753 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2758 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2763 spin_lock_init(&hctx->dispatch_wait_lock);
2764 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2765 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2767 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2771 if (hctx->flags & BLK_MQ_F_BLOCKING)
2772 init_srcu_struct(hctx->srcu);
2773 blk_mq_hctx_kobj_init(hctx);
2778 sbitmap_free(&hctx->ctx_map);
2782 free_cpumask_var(hctx->cpumask);
2789 static void blk_mq_init_cpu_queues(struct request_queue *q,
2790 unsigned int nr_hw_queues)
2792 struct blk_mq_tag_set *set = q->tag_set;
2795 for_each_possible_cpu(i) {
2796 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2797 struct blk_mq_hw_ctx *hctx;
2801 spin_lock_init(&__ctx->lock);
2802 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2803 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2808 * Set local node, IFF we have more than one hw queue. If
2809 * not, we remain on the home node of the device
2811 for (j = 0; j < set->nr_maps; j++) {
2812 hctx = blk_mq_map_queue_type(q, j, i);
2813 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2814 hctx->numa_node = cpu_to_node(i);
2819 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2822 unsigned int flags = set->flags;
2825 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2826 set->queue_depth, set->reserved_tags, flags);
2827 if (!set->tags[hctx_idx])
2830 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2835 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2836 set->tags[hctx_idx] = NULL;
2840 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2841 unsigned int hctx_idx)
2843 unsigned int flags = set->flags;
2845 if (set->tags && set->tags[hctx_idx]) {
2846 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2847 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2848 set->tags[hctx_idx] = NULL;
2852 static void blk_mq_map_swqueue(struct request_queue *q)
2854 unsigned int i, j, hctx_idx;
2855 struct blk_mq_hw_ctx *hctx;
2856 struct blk_mq_ctx *ctx;
2857 struct blk_mq_tag_set *set = q->tag_set;
2859 queue_for_each_hw_ctx(q, hctx, i) {
2860 cpumask_clear(hctx->cpumask);
2862 hctx->dispatch_from = NULL;
2866 * Map software to hardware queues.
2868 * If the cpu isn't present, the cpu is mapped to first hctx.
2870 for_each_possible_cpu(i) {
2872 ctx = per_cpu_ptr(q->queue_ctx, i);
2873 for (j = 0; j < set->nr_maps; j++) {
2874 if (!set->map[j].nr_queues) {
2875 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2876 HCTX_TYPE_DEFAULT, i);
2879 hctx_idx = set->map[j].mq_map[i];
2880 /* unmapped hw queue can be remapped after CPU topo changed */
2881 if (!set->tags[hctx_idx] &&
2882 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2884 * If tags initialization fail for some hctx,
2885 * that hctx won't be brought online. In this
2886 * case, remap the current ctx to hctx[0] which
2887 * is guaranteed to always have tags allocated
2889 set->map[j].mq_map[i] = 0;
2892 hctx = blk_mq_map_queue_type(q, j, i);
2893 ctx->hctxs[j] = hctx;
2895 * If the CPU is already set in the mask, then we've
2896 * mapped this one already. This can happen if
2897 * devices share queues across queue maps.
2899 if (cpumask_test_cpu(i, hctx->cpumask))
2902 cpumask_set_cpu(i, hctx->cpumask);
2904 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2905 hctx->ctxs[hctx->nr_ctx++] = ctx;
2908 * If the nr_ctx type overflows, we have exceeded the
2909 * amount of sw queues we can support.
2911 BUG_ON(!hctx->nr_ctx);
2914 for (; j < HCTX_MAX_TYPES; j++)
2915 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2916 HCTX_TYPE_DEFAULT, i);
2919 queue_for_each_hw_ctx(q, hctx, i) {
2921 * If no software queues are mapped to this hardware queue,
2922 * disable it and free the request entries.
2924 if (!hctx->nr_ctx) {
2925 /* Never unmap queue 0. We need it as a
2926 * fallback in case of a new remap fails
2929 if (i && set->tags[i])
2930 blk_mq_free_map_and_requests(set, i);
2936 hctx->tags = set->tags[i];
2937 WARN_ON(!hctx->tags);
2940 * Set the map size to the number of mapped software queues.
2941 * This is more accurate and more efficient than looping
2942 * over all possibly mapped software queues.
2944 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2947 * Initialize batch roundrobin counts
2949 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2950 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2955 * Caller needs to ensure that we're either frozen/quiesced, or that
2956 * the queue isn't live yet.
2958 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2960 struct blk_mq_hw_ctx *hctx;
2963 queue_for_each_hw_ctx(q, hctx, i) {
2965 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
2967 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
2971 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
2974 struct request_queue *q;
2976 lockdep_assert_held(&set->tag_list_lock);
2978 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2979 blk_mq_freeze_queue(q);
2980 queue_set_hctx_shared(q, shared);
2981 blk_mq_unfreeze_queue(q);
2985 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2987 struct blk_mq_tag_set *set = q->tag_set;
2989 mutex_lock(&set->tag_list_lock);
2990 list_del(&q->tag_set_list);
2991 if (list_is_singular(&set->tag_list)) {
2992 /* just transitioned to unshared */
2993 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
2994 /* update existing queue */
2995 blk_mq_update_tag_set_shared(set, false);
2997 mutex_unlock(&set->tag_list_lock);
2998 INIT_LIST_HEAD(&q->tag_set_list);
3001 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3002 struct request_queue *q)
3004 mutex_lock(&set->tag_list_lock);
3007 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3009 if (!list_empty(&set->tag_list) &&
3010 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3011 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3012 /* update existing queue */
3013 blk_mq_update_tag_set_shared(set, true);
3015 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3016 queue_set_hctx_shared(q, true);
3017 list_add_tail(&q->tag_set_list, &set->tag_list);
3019 mutex_unlock(&set->tag_list_lock);
3022 /* All allocations will be freed in release handler of q->mq_kobj */
3023 static int blk_mq_alloc_ctxs(struct request_queue *q)
3025 struct blk_mq_ctxs *ctxs;
3028 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3032 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3033 if (!ctxs->queue_ctx)
3036 for_each_possible_cpu(cpu) {
3037 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3041 q->mq_kobj = &ctxs->kobj;
3042 q->queue_ctx = ctxs->queue_ctx;
3051 * It is the actual release handler for mq, but we do it from
3052 * request queue's release handler for avoiding use-after-free
3053 * and headache because q->mq_kobj shouldn't have been introduced,
3054 * but we can't group ctx/kctx kobj without it.
3056 void blk_mq_release(struct request_queue *q)
3058 struct blk_mq_hw_ctx *hctx, *next;
3061 queue_for_each_hw_ctx(q, hctx, i)
3062 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3064 /* all hctx are in .unused_hctx_list now */
3065 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3066 list_del_init(&hctx->hctx_list);
3067 kobject_put(&hctx->kobj);
3070 kfree(q->queue_hw_ctx);
3073 * release .mq_kobj and sw queue's kobject now because
3074 * both share lifetime with request queue.
3076 blk_mq_sysfs_deinit(q);
3079 struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3082 struct request_queue *uninit_q, *q;
3084 uninit_q = blk_alloc_queue(set->numa_node);
3086 return ERR_PTR(-ENOMEM);
3087 uninit_q->queuedata = queuedata;
3090 * Initialize the queue without an elevator. device_add_disk() will do
3091 * the initialization.
3093 q = blk_mq_init_allocated_queue(set, uninit_q, false);
3095 blk_cleanup_queue(uninit_q);
3099 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data);
3101 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3103 return blk_mq_init_queue_data(set, NULL);
3105 EXPORT_SYMBOL(blk_mq_init_queue);
3108 * Helper for setting up a queue with mq ops, given queue depth, and
3109 * the passed in mq ops flags.
3111 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
3112 const struct blk_mq_ops *ops,
3113 unsigned int queue_depth,
3114 unsigned int set_flags)
3116 struct request_queue *q;
3119 memset(set, 0, sizeof(*set));
3121 set->nr_hw_queues = 1;
3123 set->queue_depth = queue_depth;
3124 set->numa_node = NUMA_NO_NODE;
3125 set->flags = set_flags;
3127 ret = blk_mq_alloc_tag_set(set);
3129 return ERR_PTR(ret);
3131 q = blk_mq_init_queue(set);
3133 blk_mq_free_tag_set(set);
3139 EXPORT_SYMBOL(blk_mq_init_sq_queue);
3141 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3142 struct blk_mq_tag_set *set, struct request_queue *q,
3143 int hctx_idx, int node)
3145 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3147 /* reuse dead hctx first */
3148 spin_lock(&q->unused_hctx_lock);
3149 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3150 if (tmp->numa_node == node) {
3156 list_del_init(&hctx->hctx_list);
3157 spin_unlock(&q->unused_hctx_lock);
3160 hctx = blk_mq_alloc_hctx(q, set, node);
3164 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3170 kobject_put(&hctx->kobj);
3175 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3176 struct request_queue *q)
3179 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3181 if (q->nr_hw_queues < set->nr_hw_queues) {
3182 struct blk_mq_hw_ctx **new_hctxs;
3184 new_hctxs = kcalloc_node(set->nr_hw_queues,
3185 sizeof(*new_hctxs), GFP_KERNEL,
3190 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3192 q->queue_hw_ctx = new_hctxs;
3197 /* protect against switching io scheduler */
3198 mutex_lock(&q->sysfs_lock);
3199 for (i = 0; i < set->nr_hw_queues; i++) {
3201 struct blk_mq_hw_ctx *hctx;
3203 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3205 * If the hw queue has been mapped to another numa node,
3206 * we need to realloc the hctx. If allocation fails, fallback
3207 * to use the previous one.
3209 if (hctxs[i] && (hctxs[i]->numa_node == node))
3212 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3215 blk_mq_exit_hctx(q, set, hctxs[i], i);
3219 pr_warn("Allocate new hctx on node %d fails,\
3220 fallback to previous one on node %d\n",
3221 node, hctxs[i]->numa_node);
3227 * Increasing nr_hw_queues fails. Free the newly allocated
3228 * hctxs and keep the previous q->nr_hw_queues.
3230 if (i != set->nr_hw_queues) {
3231 j = q->nr_hw_queues;
3235 end = q->nr_hw_queues;
3236 q->nr_hw_queues = set->nr_hw_queues;
3239 for (; j < end; j++) {
3240 struct blk_mq_hw_ctx *hctx = hctxs[j];
3244 blk_mq_free_map_and_requests(set, j);
3245 blk_mq_exit_hctx(q, set, hctx, j);
3249 mutex_unlock(&q->sysfs_lock);
3252 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3253 struct request_queue *q,
3256 /* mark the queue as mq asap */
3257 q->mq_ops = set->ops;
3259 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3260 blk_mq_poll_stats_bkt,
3261 BLK_MQ_POLL_STATS_BKTS, q);
3265 if (blk_mq_alloc_ctxs(q))
3268 /* init q->mq_kobj and sw queues' kobjects */
3269 blk_mq_sysfs_init(q);
3271 INIT_LIST_HEAD(&q->unused_hctx_list);
3272 spin_lock_init(&q->unused_hctx_lock);
3274 blk_mq_realloc_hw_ctxs(set, q);
3275 if (!q->nr_hw_queues)
3278 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3279 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3283 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3284 if (set->nr_maps > HCTX_TYPE_POLL &&
3285 set->map[HCTX_TYPE_POLL].nr_queues)
3286 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3288 q->sg_reserved_size = INT_MAX;
3290 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3291 INIT_LIST_HEAD(&q->requeue_list);
3292 spin_lock_init(&q->requeue_lock);
3294 q->nr_requests = set->queue_depth;
3297 * Default to classic polling
3299 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3301 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3302 blk_mq_add_queue_tag_set(set, q);
3303 blk_mq_map_swqueue(q);
3306 elevator_init_mq(q);
3311 kfree(q->queue_hw_ctx);
3312 q->nr_hw_queues = 0;
3313 blk_mq_sysfs_deinit(q);
3315 blk_stat_free_callback(q->poll_cb);
3319 return ERR_PTR(-ENOMEM);
3321 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3323 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3324 void blk_mq_exit_queue(struct request_queue *q)
3326 struct blk_mq_tag_set *set = q->tag_set;
3328 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
3329 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3330 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
3331 blk_mq_del_queue_tag_set(q);
3334 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3338 for (i = 0; i < set->nr_hw_queues; i++) {
3339 if (!__blk_mq_alloc_map_and_request(set, i))
3348 blk_mq_free_map_and_requests(set, i);
3354 * Allocate the request maps associated with this tag_set. Note that this
3355 * may reduce the depth asked for, if memory is tight. set->queue_depth
3356 * will be updated to reflect the allocated depth.
3358 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3363 depth = set->queue_depth;
3365 err = __blk_mq_alloc_rq_maps(set);
3369 set->queue_depth >>= 1;
3370 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3374 } while (set->queue_depth);
3376 if (!set->queue_depth || err) {
3377 pr_err("blk-mq: failed to allocate request map\n");
3381 if (depth != set->queue_depth)
3382 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3383 depth, set->queue_depth);
3388 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3391 * blk_mq_map_queues() and multiple .map_queues() implementations
3392 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3393 * number of hardware queues.
3395 if (set->nr_maps == 1)
3396 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3398 if (set->ops->map_queues && !is_kdump_kernel()) {
3402 * transport .map_queues is usually done in the following
3405 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3406 * mask = get_cpu_mask(queue)
3407 * for_each_cpu(cpu, mask)
3408 * set->map[x].mq_map[cpu] = queue;
3411 * When we need to remap, the table has to be cleared for
3412 * killing stale mapping since one CPU may not be mapped
3415 for (i = 0; i < set->nr_maps; i++)
3416 blk_mq_clear_mq_map(&set->map[i]);
3418 return set->ops->map_queues(set);
3420 BUG_ON(set->nr_maps > 1);
3421 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3425 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3426 int cur_nr_hw_queues, int new_nr_hw_queues)
3428 struct blk_mq_tags **new_tags;
3430 if (cur_nr_hw_queues >= new_nr_hw_queues)
3433 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3434 GFP_KERNEL, set->numa_node);
3439 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3440 sizeof(*set->tags));
3442 set->tags = new_tags;
3443 set->nr_hw_queues = new_nr_hw_queues;
3449 * Alloc a tag set to be associated with one or more request queues.
3450 * May fail with EINVAL for various error conditions. May adjust the
3451 * requested depth down, if it's too large. In that case, the set
3452 * value will be stored in set->queue_depth.
3454 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3458 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3460 if (!set->nr_hw_queues)
3462 if (!set->queue_depth)
3464 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3467 if (!set->ops->queue_rq)
3470 if (!set->ops->get_budget ^ !set->ops->put_budget)
3473 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3474 pr_info("blk-mq: reduced tag depth to %u\n",
3476 set->queue_depth = BLK_MQ_MAX_DEPTH;
3481 else if (set->nr_maps > HCTX_MAX_TYPES)
3485 * If a crashdump is active, then we are potentially in a very
3486 * memory constrained environment. Limit us to 1 queue and
3487 * 64 tags to prevent using too much memory.
3489 if (is_kdump_kernel()) {
3490 set->nr_hw_queues = 1;
3492 set->queue_depth = min(64U, set->queue_depth);
3495 * There is no use for more h/w queues than cpus if we just have
3498 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3499 set->nr_hw_queues = nr_cpu_ids;
3501 if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0)
3505 for (i = 0; i < set->nr_maps; i++) {
3506 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3507 sizeof(set->map[i].mq_map[0]),
3508 GFP_KERNEL, set->numa_node);
3509 if (!set->map[i].mq_map)
3510 goto out_free_mq_map;
3511 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3514 ret = blk_mq_update_queue_map(set);
3516 goto out_free_mq_map;
3518 ret = blk_mq_alloc_map_and_requests(set);
3520 goto out_free_mq_map;
3522 if (blk_mq_is_sbitmap_shared(set->flags)) {
3523 atomic_set(&set->active_queues_shared_sbitmap, 0);
3525 if (blk_mq_init_shared_sbitmap(set, set->flags)) {
3527 goto out_free_mq_rq_maps;
3531 mutex_init(&set->tag_list_lock);
3532 INIT_LIST_HEAD(&set->tag_list);
3536 out_free_mq_rq_maps:
3537 for (i = 0; i < set->nr_hw_queues; i++)
3538 blk_mq_free_map_and_requests(set, i);
3540 for (i = 0; i < set->nr_maps; i++) {
3541 kfree(set->map[i].mq_map);
3542 set->map[i].mq_map = NULL;
3548 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3550 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3554 for (i = 0; i < set->nr_hw_queues; i++)
3555 blk_mq_free_map_and_requests(set, i);
3557 if (blk_mq_is_sbitmap_shared(set->flags))
3558 blk_mq_exit_shared_sbitmap(set);
3560 for (j = 0; j < set->nr_maps; j++) {
3561 kfree(set->map[j].mq_map);
3562 set->map[j].mq_map = NULL;
3568 EXPORT_SYMBOL(blk_mq_free_tag_set);
3570 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3572 struct blk_mq_tag_set *set = q->tag_set;
3573 struct blk_mq_hw_ctx *hctx;
3579 if (q->nr_requests == nr)
3582 blk_mq_freeze_queue(q);
3583 blk_mq_quiesce_queue(q);
3586 queue_for_each_hw_ctx(q, hctx, i) {
3590 * If we're using an MQ scheduler, just update the scheduler
3591 * queue depth. This is similar to what the old code would do.
3593 if (!hctx->sched_tags) {
3594 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3596 if (!ret && blk_mq_is_sbitmap_shared(set->flags))
3597 blk_mq_tag_resize_shared_sbitmap(set, nr);
3599 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3604 if (q->elevator && q->elevator->type->ops.depth_updated)
3605 q->elevator->type->ops.depth_updated(hctx);
3609 q->nr_requests = nr;
3611 blk_mq_unquiesce_queue(q);
3612 blk_mq_unfreeze_queue(q);
3618 * request_queue and elevator_type pair.
3619 * It is just used by __blk_mq_update_nr_hw_queues to cache
3620 * the elevator_type associated with a request_queue.
3622 struct blk_mq_qe_pair {
3623 struct list_head node;
3624 struct request_queue *q;
3625 struct elevator_type *type;
3629 * Cache the elevator_type in qe pair list and switch the
3630 * io scheduler to 'none'
3632 static bool blk_mq_elv_switch_none(struct list_head *head,
3633 struct request_queue *q)
3635 struct blk_mq_qe_pair *qe;
3640 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3644 INIT_LIST_HEAD(&qe->node);
3646 qe->type = q->elevator->type;
3647 list_add(&qe->node, head);
3649 mutex_lock(&q->sysfs_lock);
3651 * After elevator_switch_mq, the previous elevator_queue will be
3652 * released by elevator_release. The reference of the io scheduler
3653 * module get by elevator_get will also be put. So we need to get
3654 * a reference of the io scheduler module here to prevent it to be
3657 __module_get(qe->type->elevator_owner);
3658 elevator_switch_mq(q, NULL);
3659 mutex_unlock(&q->sysfs_lock);
3664 static void blk_mq_elv_switch_back(struct list_head *head,
3665 struct request_queue *q)
3667 struct blk_mq_qe_pair *qe;
3668 struct elevator_type *t = NULL;
3670 list_for_each_entry(qe, head, node)
3679 list_del(&qe->node);
3682 mutex_lock(&q->sysfs_lock);
3683 elevator_switch_mq(q, t);
3684 mutex_unlock(&q->sysfs_lock);
3687 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3690 struct request_queue *q;
3692 int prev_nr_hw_queues;
3694 lockdep_assert_held(&set->tag_list_lock);
3696 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3697 nr_hw_queues = nr_cpu_ids;
3698 if (nr_hw_queues < 1)
3700 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3703 list_for_each_entry(q, &set->tag_list, tag_set_list)
3704 blk_mq_freeze_queue(q);
3706 * Switch IO scheduler to 'none', cleaning up the data associated
3707 * with the previous scheduler. We will switch back once we are done
3708 * updating the new sw to hw queue mappings.
3710 list_for_each_entry(q, &set->tag_list, tag_set_list)
3711 if (!blk_mq_elv_switch_none(&head, q))
3714 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3715 blk_mq_debugfs_unregister_hctxs(q);
3716 blk_mq_sysfs_unregister(q);
3719 prev_nr_hw_queues = set->nr_hw_queues;
3720 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3724 set->nr_hw_queues = nr_hw_queues;
3726 blk_mq_update_queue_map(set);
3727 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3728 blk_mq_realloc_hw_ctxs(set, q);
3729 if (q->nr_hw_queues != set->nr_hw_queues) {
3730 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3731 nr_hw_queues, prev_nr_hw_queues);
3732 set->nr_hw_queues = prev_nr_hw_queues;
3733 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3736 blk_mq_map_swqueue(q);
3740 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3741 blk_mq_sysfs_register(q);
3742 blk_mq_debugfs_register_hctxs(q);
3746 list_for_each_entry(q, &set->tag_list, tag_set_list)
3747 blk_mq_elv_switch_back(&head, q);
3749 list_for_each_entry(q, &set->tag_list, tag_set_list)
3750 blk_mq_unfreeze_queue(q);
3753 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3755 mutex_lock(&set->tag_list_lock);
3756 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3757 mutex_unlock(&set->tag_list_lock);
3759 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3761 /* Enable polling stats and return whether they were already enabled. */
3762 static bool blk_poll_stats_enable(struct request_queue *q)
3764 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3765 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3767 blk_stat_add_callback(q, q->poll_cb);
3771 static void blk_mq_poll_stats_start(struct request_queue *q)
3774 * We don't arm the callback if polling stats are not enabled or the
3775 * callback is already active.
3777 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3778 blk_stat_is_active(q->poll_cb))
3781 blk_stat_activate_msecs(q->poll_cb, 100);
3784 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3786 struct request_queue *q = cb->data;
3789 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3790 if (cb->stat[bucket].nr_samples)
3791 q->poll_stat[bucket] = cb->stat[bucket];
3795 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3798 unsigned long ret = 0;
3802 * If stats collection isn't on, don't sleep but turn it on for
3805 if (!blk_poll_stats_enable(q))
3809 * As an optimistic guess, use half of the mean service time
3810 * for this type of request. We can (and should) make this smarter.
3811 * For instance, if the completion latencies are tight, we can
3812 * get closer than just half the mean. This is especially
3813 * important on devices where the completion latencies are longer
3814 * than ~10 usec. We do use the stats for the relevant IO size
3815 * if available which does lead to better estimates.
3817 bucket = blk_mq_poll_stats_bkt(rq);
3821 if (q->poll_stat[bucket].nr_samples)
3822 ret = (q->poll_stat[bucket].mean + 1) / 2;
3827 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3830 struct hrtimer_sleeper hs;
3831 enum hrtimer_mode mode;
3835 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3839 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3841 * 0: use half of prev avg
3842 * >0: use this specific value
3844 if (q->poll_nsec > 0)
3845 nsecs = q->poll_nsec;
3847 nsecs = blk_mq_poll_nsecs(q, rq);
3852 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3855 * This will be replaced with the stats tracking code, using
3856 * 'avg_completion_time / 2' as the pre-sleep target.
3860 mode = HRTIMER_MODE_REL;
3861 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3862 hrtimer_set_expires(&hs.timer, kt);
3865 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3867 set_current_state(TASK_UNINTERRUPTIBLE);
3868 hrtimer_sleeper_start_expires(&hs, mode);
3871 hrtimer_cancel(&hs.timer);
3872 mode = HRTIMER_MODE_ABS;
3873 } while (hs.task && !signal_pending(current));
3875 __set_current_state(TASK_RUNNING);
3876 destroy_hrtimer_on_stack(&hs.timer);
3880 static bool blk_mq_poll_hybrid(struct request_queue *q,
3881 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3885 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3888 if (!blk_qc_t_is_internal(cookie))
3889 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3891 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3893 * With scheduling, if the request has completed, we'll
3894 * get a NULL return here, as we clear the sched tag when
3895 * that happens. The request still remains valid, like always,
3896 * so we should be safe with just the NULL check.
3902 return blk_mq_poll_hybrid_sleep(q, rq);
3906 * blk_poll - poll for IO completions
3908 * @cookie: cookie passed back at IO submission time
3909 * @spin: whether to spin for completions
3912 * Poll for completions on the passed in queue. Returns number of
3913 * completed entries found. If @spin is true, then blk_poll will continue
3914 * looping until at least one completion is found, unless the task is
3915 * otherwise marked running (or we need to reschedule).
3917 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3919 struct blk_mq_hw_ctx *hctx;
3922 if (!blk_qc_t_valid(cookie) ||
3923 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3927 blk_flush_plug_list(current->plug, false);
3929 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3932 * If we sleep, have the caller restart the poll loop to reset
3933 * the state. Like for the other success return cases, the
3934 * caller is responsible for checking if the IO completed. If
3935 * the IO isn't complete, we'll get called again and will go
3936 * straight to the busy poll loop.
3938 if (blk_mq_poll_hybrid(q, hctx, cookie))
3941 hctx->poll_considered++;
3943 state = current->state;
3947 hctx->poll_invoked++;
3949 ret = q->mq_ops->poll(hctx);
3951 hctx->poll_success++;
3952 __set_current_state(TASK_RUNNING);
3956 if (signal_pending_state(state, current))
3957 __set_current_state(TASK_RUNNING);
3959 if (current->state == TASK_RUNNING)
3961 if (ret < 0 || !spin)
3964 } while (!need_resched());
3966 __set_current_state(TASK_RUNNING);
3969 EXPORT_SYMBOL_GPL(blk_poll);
3971 unsigned int blk_mq_rq_cpu(struct request *rq)
3973 return rq->mq_ctx->cpu;
3975 EXPORT_SYMBOL(blk_mq_rq_cpu);
3977 static int __init blk_mq_init(void)
3981 for_each_possible_cpu(i)
3982 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3983 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
3985 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
3986 "block/softirq:dead", NULL,
3987 blk_softirq_cpu_dead);
3988 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3989 blk_mq_hctx_notify_dead);
3990 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
3991 blk_mq_hctx_notify_online,
3992 blk_mq_hctx_notify_offline);
3995 subsys_initcall(blk_mq_init);