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_PREEMPT)
296 rq->rq_flags |= RQF_PREEMPT;
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 BUG_ON(!list_empty(&rq->queuelist));
778 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
780 EXPORT_SYMBOL(blk_mq_requeue_request);
782 static void blk_mq_requeue_work(struct work_struct *work)
784 struct request_queue *q =
785 container_of(work, struct request_queue, requeue_work.work);
787 struct request *rq, *next;
789 spin_lock_irq(&q->requeue_lock);
790 list_splice_init(&q->requeue_list, &rq_list);
791 spin_unlock_irq(&q->requeue_lock);
793 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
794 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
797 rq->rq_flags &= ~RQF_SOFTBARRIER;
798 list_del_init(&rq->queuelist);
800 * If RQF_DONTPREP, rq has contained some driver specific
801 * data, so insert it to hctx dispatch list to avoid any
804 if (rq->rq_flags & RQF_DONTPREP)
805 blk_mq_request_bypass_insert(rq, false, false);
807 blk_mq_sched_insert_request(rq, true, false, false);
810 while (!list_empty(&rq_list)) {
811 rq = list_entry(rq_list.next, struct request, queuelist);
812 list_del_init(&rq->queuelist);
813 blk_mq_sched_insert_request(rq, false, false, false);
816 blk_mq_run_hw_queues(q, false);
819 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
820 bool kick_requeue_list)
822 struct request_queue *q = rq->q;
826 * We abuse this flag that is otherwise used by the I/O scheduler to
827 * request head insertion from the workqueue.
829 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
831 spin_lock_irqsave(&q->requeue_lock, flags);
833 rq->rq_flags |= RQF_SOFTBARRIER;
834 list_add(&rq->queuelist, &q->requeue_list);
836 list_add_tail(&rq->queuelist, &q->requeue_list);
838 spin_unlock_irqrestore(&q->requeue_lock, flags);
840 if (kick_requeue_list)
841 blk_mq_kick_requeue_list(q);
844 void blk_mq_kick_requeue_list(struct request_queue *q)
846 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
848 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
850 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
853 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
854 msecs_to_jiffies(msecs));
856 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
858 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
860 if (tag < tags->nr_tags) {
861 prefetch(tags->rqs[tag]);
862 return tags->rqs[tag];
867 EXPORT_SYMBOL(blk_mq_tag_to_rq);
869 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
870 void *priv, bool reserved)
873 * If we find a request that isn't idle and the queue matches,
874 * we know the queue is busy. Return false to stop the iteration.
876 if (blk_mq_request_started(rq) && rq->q == hctx->queue) {
886 bool blk_mq_queue_inflight(struct request_queue *q)
890 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
893 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
895 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
897 req->rq_flags |= RQF_TIMED_OUT;
898 if (req->q->mq_ops->timeout) {
899 enum blk_eh_timer_return ret;
901 ret = req->q->mq_ops->timeout(req, reserved);
902 if (ret == BLK_EH_DONE)
904 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
910 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
912 unsigned long deadline;
914 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
916 if (rq->rq_flags & RQF_TIMED_OUT)
919 deadline = READ_ONCE(rq->deadline);
920 if (time_after_eq(jiffies, deadline))
925 else if (time_after(*next, deadline))
930 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
931 struct request *rq, void *priv, bool reserved)
933 unsigned long *next = priv;
936 * Just do a quick check if it is expired before locking the request in
937 * so we're not unnecessarilly synchronizing across CPUs.
939 if (!blk_mq_req_expired(rq, next))
943 * We have reason to believe the request may be expired. Take a
944 * reference on the request to lock this request lifetime into its
945 * currently allocated context to prevent it from being reallocated in
946 * the event the completion by-passes this timeout handler.
948 * If the reference was already released, then the driver beat the
949 * timeout handler to posting a natural completion.
951 if (!refcount_inc_not_zero(&rq->ref))
955 * The request is now locked and cannot be reallocated underneath the
956 * timeout handler's processing. Re-verify this exact request is truly
957 * expired; if it is not expired, then the request was completed and
958 * reallocated as a new request.
960 if (blk_mq_req_expired(rq, next))
961 blk_mq_rq_timed_out(rq, reserved);
963 if (is_flush_rq(rq, hctx))
965 else if (refcount_dec_and_test(&rq->ref))
966 __blk_mq_free_request(rq);
971 static void blk_mq_timeout_work(struct work_struct *work)
973 struct request_queue *q =
974 container_of(work, struct request_queue, timeout_work);
975 unsigned long next = 0;
976 struct blk_mq_hw_ctx *hctx;
979 /* A deadlock might occur if a request is stuck requiring a
980 * timeout at the same time a queue freeze is waiting
981 * completion, since the timeout code would not be able to
982 * acquire the queue reference here.
984 * That's why we don't use blk_queue_enter here; instead, we use
985 * percpu_ref_tryget directly, because we need to be able to
986 * obtain a reference even in the short window between the queue
987 * starting to freeze, by dropping the first reference in
988 * blk_freeze_queue_start, and the moment the last request is
989 * consumed, marked by the instant q_usage_counter reaches
992 if (!percpu_ref_tryget(&q->q_usage_counter))
995 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
998 mod_timer(&q->timeout, next);
1001 * Request timeouts are handled as a forward rolling timer. If
1002 * we end up here it means that no requests are pending and
1003 * also that no request has been pending for a while. Mark
1004 * each hctx as idle.
1006 queue_for_each_hw_ctx(q, hctx, i) {
1007 /* the hctx may be unmapped, so check it here */
1008 if (blk_mq_hw_queue_mapped(hctx))
1009 blk_mq_tag_idle(hctx);
1015 struct flush_busy_ctx_data {
1016 struct blk_mq_hw_ctx *hctx;
1017 struct list_head *list;
1020 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1022 struct flush_busy_ctx_data *flush_data = data;
1023 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1024 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1025 enum hctx_type type = hctx->type;
1027 spin_lock(&ctx->lock);
1028 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1029 sbitmap_clear_bit(sb, bitnr);
1030 spin_unlock(&ctx->lock);
1035 * Process software queues that have been marked busy, splicing them
1036 * to the for-dispatch
1038 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1040 struct flush_busy_ctx_data data = {
1045 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1047 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1049 struct dispatch_rq_data {
1050 struct blk_mq_hw_ctx *hctx;
1054 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1057 struct dispatch_rq_data *dispatch_data = data;
1058 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1059 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1060 enum hctx_type type = hctx->type;
1062 spin_lock(&ctx->lock);
1063 if (!list_empty(&ctx->rq_lists[type])) {
1064 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1065 list_del_init(&dispatch_data->rq->queuelist);
1066 if (list_empty(&ctx->rq_lists[type]))
1067 sbitmap_clear_bit(sb, bitnr);
1069 spin_unlock(&ctx->lock);
1071 return !dispatch_data->rq;
1074 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1075 struct blk_mq_ctx *start)
1077 unsigned off = start ? start->index_hw[hctx->type] : 0;
1078 struct dispatch_rq_data data = {
1083 __sbitmap_for_each_set(&hctx->ctx_map, off,
1084 dispatch_rq_from_ctx, &data);
1089 static inline unsigned int queued_to_index(unsigned int queued)
1094 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1097 static bool __blk_mq_get_driver_tag(struct request *rq)
1099 struct sbitmap_queue *bt = rq->mq_hctx->tags->bitmap_tags;
1100 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1103 blk_mq_tag_busy(rq->mq_hctx);
1105 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1106 bt = rq->mq_hctx->tags->breserved_tags;
1109 if (!hctx_may_queue(rq->mq_hctx, bt))
1113 tag = __sbitmap_queue_get(bt);
1114 if (tag == BLK_MQ_NO_TAG)
1117 rq->tag = tag + tag_offset;
1121 static bool blk_mq_get_driver_tag(struct request *rq)
1123 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1125 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq))
1128 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1129 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1130 rq->rq_flags |= RQF_MQ_INFLIGHT;
1131 __blk_mq_inc_active_requests(hctx);
1133 hctx->tags->rqs[rq->tag] = rq;
1137 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1138 int flags, void *key)
1140 struct blk_mq_hw_ctx *hctx;
1142 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1144 spin_lock(&hctx->dispatch_wait_lock);
1145 if (!list_empty(&wait->entry)) {
1146 struct sbitmap_queue *sbq;
1148 list_del_init(&wait->entry);
1149 sbq = hctx->tags->bitmap_tags;
1150 atomic_dec(&sbq->ws_active);
1152 spin_unlock(&hctx->dispatch_wait_lock);
1154 blk_mq_run_hw_queue(hctx, true);
1159 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1160 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1161 * restart. For both cases, take care to check the condition again after
1162 * marking us as waiting.
1164 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1167 struct sbitmap_queue *sbq = hctx->tags->bitmap_tags;
1168 struct wait_queue_head *wq;
1169 wait_queue_entry_t *wait;
1172 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1173 blk_mq_sched_mark_restart_hctx(hctx);
1176 * It's possible that a tag was freed in the window between the
1177 * allocation failure and adding the hardware queue to the wait
1180 * Don't clear RESTART here, someone else could have set it.
1181 * At most this will cost an extra queue run.
1183 return blk_mq_get_driver_tag(rq);
1186 wait = &hctx->dispatch_wait;
1187 if (!list_empty_careful(&wait->entry))
1190 wq = &bt_wait_ptr(sbq, hctx)->wait;
1192 spin_lock_irq(&wq->lock);
1193 spin_lock(&hctx->dispatch_wait_lock);
1194 if (!list_empty(&wait->entry)) {
1195 spin_unlock(&hctx->dispatch_wait_lock);
1196 spin_unlock_irq(&wq->lock);
1200 atomic_inc(&sbq->ws_active);
1201 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1202 __add_wait_queue(wq, wait);
1205 * It's possible that a tag was freed in the window between the
1206 * allocation failure and adding the hardware queue to the wait
1209 ret = blk_mq_get_driver_tag(rq);
1211 spin_unlock(&hctx->dispatch_wait_lock);
1212 spin_unlock_irq(&wq->lock);
1217 * We got a tag, remove ourselves from the wait queue to ensure
1218 * someone else gets the wakeup.
1220 list_del_init(&wait->entry);
1221 atomic_dec(&sbq->ws_active);
1222 spin_unlock(&hctx->dispatch_wait_lock);
1223 spin_unlock_irq(&wq->lock);
1228 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1229 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1231 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1232 * - EWMA is one simple way to compute running average value
1233 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1234 * - take 4 as factor for avoiding to get too small(0) result, and this
1235 * factor doesn't matter because EWMA decreases exponentially
1237 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1241 if (hctx->queue->elevator)
1244 ewma = hctx->dispatch_busy;
1249 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1251 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1252 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1254 hctx->dispatch_busy = ewma;
1257 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1259 static void blk_mq_handle_dev_resource(struct request *rq,
1260 struct list_head *list)
1262 struct request *next =
1263 list_first_entry_or_null(list, struct request, queuelist);
1266 * If an I/O scheduler has been configured and we got a driver tag for
1267 * the next request already, free it.
1270 blk_mq_put_driver_tag(next);
1272 list_add(&rq->queuelist, list);
1273 __blk_mq_requeue_request(rq);
1276 static void blk_mq_handle_zone_resource(struct request *rq,
1277 struct list_head *zone_list)
1280 * If we end up here it is because we cannot dispatch a request to a
1281 * specific zone due to LLD level zone-write locking or other zone
1282 * related resource not being available. In this case, set the request
1283 * aside in zone_list for retrying it later.
1285 list_add(&rq->queuelist, zone_list);
1286 __blk_mq_requeue_request(rq);
1289 enum prep_dispatch {
1291 PREP_DISPATCH_NO_TAG,
1292 PREP_DISPATCH_NO_BUDGET,
1295 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1298 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1300 if (need_budget && !blk_mq_get_dispatch_budget(rq->q)) {
1301 blk_mq_put_driver_tag(rq);
1302 return PREP_DISPATCH_NO_BUDGET;
1305 if (!blk_mq_get_driver_tag(rq)) {
1307 * The initial allocation attempt failed, so we need to
1308 * rerun the hardware queue when a tag is freed. The
1309 * waitqueue takes care of that. If the queue is run
1310 * before we add this entry back on the dispatch list,
1311 * we'll re-run it below.
1313 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1315 * All budgets not got from this function will be put
1316 * together during handling partial dispatch
1319 blk_mq_put_dispatch_budget(rq->q);
1320 return PREP_DISPATCH_NO_TAG;
1324 return PREP_DISPATCH_OK;
1327 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1328 static void blk_mq_release_budgets(struct request_queue *q,
1329 unsigned int nr_budgets)
1333 for (i = 0; i < nr_budgets; i++)
1334 blk_mq_put_dispatch_budget(q);
1338 * Returns true if we did some work AND can potentially do more.
1340 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1341 unsigned int nr_budgets)
1343 enum prep_dispatch prep;
1344 struct request_queue *q = hctx->queue;
1345 struct request *rq, *nxt;
1347 blk_status_t ret = BLK_STS_OK;
1348 LIST_HEAD(zone_list);
1350 if (list_empty(list))
1354 * Now process all the entries, sending them to the driver.
1356 errors = queued = 0;
1358 struct blk_mq_queue_data bd;
1360 rq = list_first_entry(list, struct request, queuelist);
1362 WARN_ON_ONCE(hctx != rq->mq_hctx);
1363 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1364 if (prep != PREP_DISPATCH_OK)
1367 list_del_init(&rq->queuelist);
1372 * Flag last if we have no more requests, or if we have more
1373 * but can't assign a driver tag to it.
1375 if (list_empty(list))
1378 nxt = list_first_entry(list, struct request, queuelist);
1379 bd.last = !blk_mq_get_driver_tag(nxt);
1383 * once the request is queued to lld, no need to cover the
1388 ret = q->mq_ops->queue_rq(hctx, &bd);
1393 case BLK_STS_RESOURCE:
1394 case BLK_STS_DEV_RESOURCE:
1395 blk_mq_handle_dev_resource(rq, list);
1397 case BLK_STS_ZONE_RESOURCE:
1399 * Move the request to zone_list and keep going through
1400 * the dispatch list to find more requests the drive can
1403 blk_mq_handle_zone_resource(rq, &zone_list);
1407 blk_mq_end_request(rq, BLK_STS_IOERR);
1409 } while (!list_empty(list));
1411 if (!list_empty(&zone_list))
1412 list_splice_tail_init(&zone_list, list);
1414 hctx->dispatched[queued_to_index(queued)]++;
1416 /* If we didn't flush the entire list, we could have told the driver
1417 * there was more coming, but that turned out to be a lie.
1419 if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued)
1420 q->mq_ops->commit_rqs(hctx);
1422 * Any items that need requeuing? Stuff them into hctx->dispatch,
1423 * that is where we will continue on next queue run.
1425 if (!list_empty(list)) {
1427 /* For non-shared tags, the RESTART check will suffice */
1428 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1429 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1430 bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET;
1432 blk_mq_release_budgets(q, nr_budgets);
1434 spin_lock(&hctx->lock);
1435 list_splice_tail_init(list, &hctx->dispatch);
1436 spin_unlock(&hctx->lock);
1439 * Order adding requests to hctx->dispatch and checking
1440 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1441 * in blk_mq_sched_restart(). Avoid restart code path to
1442 * miss the new added requests to hctx->dispatch, meantime
1443 * SCHED_RESTART is observed here.
1448 * If SCHED_RESTART was set by the caller of this function and
1449 * it is no longer set that means that it was cleared by another
1450 * thread and hence that a queue rerun is needed.
1452 * If 'no_tag' is set, that means that we failed getting
1453 * a driver tag with an I/O scheduler attached. If our dispatch
1454 * waitqueue is no longer active, ensure that we run the queue
1455 * AFTER adding our entries back to the list.
1457 * If no I/O scheduler has been configured it is possible that
1458 * the hardware queue got stopped and restarted before requests
1459 * were pushed back onto the dispatch list. Rerun the queue to
1460 * avoid starvation. Notes:
1461 * - blk_mq_run_hw_queue() checks whether or not a queue has
1462 * been stopped before rerunning a queue.
1463 * - Some but not all block drivers stop a queue before
1464 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1467 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1468 * bit is set, run queue after a delay to avoid IO stalls
1469 * that could otherwise occur if the queue is idle. We'll do
1470 * similar if we couldn't get budget and SCHED_RESTART is set.
1472 needs_restart = blk_mq_sched_needs_restart(hctx);
1473 if (!needs_restart ||
1474 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1475 blk_mq_run_hw_queue(hctx, true);
1476 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1478 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1480 blk_mq_update_dispatch_busy(hctx, true);
1483 blk_mq_update_dispatch_busy(hctx, false);
1485 return (queued + errors) != 0;
1489 * __blk_mq_run_hw_queue - Run a hardware queue.
1490 * @hctx: Pointer to the hardware queue to run.
1492 * Send pending requests to the hardware.
1494 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1499 * We should be running this queue from one of the CPUs that
1502 * There are at least two related races now between setting
1503 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1504 * __blk_mq_run_hw_queue():
1506 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1507 * but later it becomes online, then this warning is harmless
1510 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1511 * but later it becomes offline, then the warning can't be
1512 * triggered, and we depend on blk-mq timeout handler to
1513 * handle dispatched requests to this hctx
1515 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1516 cpu_online(hctx->next_cpu)) {
1517 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1518 raw_smp_processor_id(),
1519 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1524 * We can't run the queue inline with ints disabled. Ensure that
1525 * we catch bad users of this early.
1527 WARN_ON_ONCE(in_interrupt());
1529 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1531 hctx_lock(hctx, &srcu_idx);
1532 blk_mq_sched_dispatch_requests(hctx);
1533 hctx_unlock(hctx, srcu_idx);
1536 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1538 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1540 if (cpu >= nr_cpu_ids)
1541 cpu = cpumask_first(hctx->cpumask);
1546 * It'd be great if the workqueue API had a way to pass
1547 * in a mask and had some smarts for more clever placement.
1548 * For now we just round-robin here, switching for every
1549 * BLK_MQ_CPU_WORK_BATCH queued items.
1551 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1554 int next_cpu = hctx->next_cpu;
1556 if (hctx->queue->nr_hw_queues == 1)
1557 return WORK_CPU_UNBOUND;
1559 if (--hctx->next_cpu_batch <= 0) {
1561 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1563 if (next_cpu >= nr_cpu_ids)
1564 next_cpu = blk_mq_first_mapped_cpu(hctx);
1565 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1569 * Do unbound schedule if we can't find a online CPU for this hctx,
1570 * and it should only happen in the path of handling CPU DEAD.
1572 if (!cpu_online(next_cpu)) {
1579 * Make sure to re-select CPU next time once after CPUs
1580 * in hctx->cpumask become online again.
1582 hctx->next_cpu = next_cpu;
1583 hctx->next_cpu_batch = 1;
1584 return WORK_CPU_UNBOUND;
1587 hctx->next_cpu = next_cpu;
1592 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1593 * @hctx: Pointer to the hardware queue to run.
1594 * @async: If we want to run the queue asynchronously.
1595 * @msecs: Microseconds of delay to wait before running the queue.
1597 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1598 * with a delay of @msecs.
1600 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1601 unsigned long msecs)
1603 if (unlikely(blk_mq_hctx_stopped(hctx)))
1606 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1607 int cpu = get_cpu();
1608 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1609 __blk_mq_run_hw_queue(hctx);
1617 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1618 msecs_to_jiffies(msecs));
1622 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1623 * @hctx: Pointer to the hardware queue to run.
1624 * @msecs: Microseconds of delay to wait before running the queue.
1626 * Run a hardware queue asynchronously with a delay of @msecs.
1628 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1630 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1632 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1635 * blk_mq_run_hw_queue - Start to run a hardware queue.
1636 * @hctx: Pointer to the hardware queue to run.
1637 * @async: If we want to run the queue asynchronously.
1639 * Check if the request queue is not in a quiesced state and if there are
1640 * pending requests to be sent. If this is true, run the queue to send requests
1643 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1649 * When queue is quiesced, we may be switching io scheduler, or
1650 * updating nr_hw_queues, or other things, and we can't run queue
1651 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1653 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1656 hctx_lock(hctx, &srcu_idx);
1657 need_run = !blk_queue_quiesced(hctx->queue) &&
1658 blk_mq_hctx_has_pending(hctx);
1659 hctx_unlock(hctx, srcu_idx);
1662 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1664 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1667 * blk_mq_run_hw_queue - Run all hardware queues in a request queue.
1668 * @q: Pointer to the request queue to run.
1669 * @async: If we want to run the queue asynchronously.
1671 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1673 struct blk_mq_hw_ctx *hctx;
1676 queue_for_each_hw_ctx(q, hctx, i) {
1677 if (blk_mq_hctx_stopped(hctx))
1680 blk_mq_run_hw_queue(hctx, async);
1683 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1686 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1687 * @q: Pointer to the request queue to run.
1688 * @msecs: Microseconds of delay to wait before running the queues.
1690 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1692 struct blk_mq_hw_ctx *hctx;
1695 queue_for_each_hw_ctx(q, hctx, i) {
1696 if (blk_mq_hctx_stopped(hctx))
1699 blk_mq_delay_run_hw_queue(hctx, msecs);
1702 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1705 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1706 * @q: request queue.
1708 * The caller is responsible for serializing this function against
1709 * blk_mq_{start,stop}_hw_queue().
1711 bool blk_mq_queue_stopped(struct request_queue *q)
1713 struct blk_mq_hw_ctx *hctx;
1716 queue_for_each_hw_ctx(q, hctx, i)
1717 if (blk_mq_hctx_stopped(hctx))
1722 EXPORT_SYMBOL(blk_mq_queue_stopped);
1725 * This function is often used for pausing .queue_rq() by driver when
1726 * there isn't enough resource or some conditions aren't satisfied, and
1727 * BLK_STS_RESOURCE is usually returned.
1729 * We do not guarantee that dispatch can be drained or blocked
1730 * after blk_mq_stop_hw_queue() returns. Please use
1731 * blk_mq_quiesce_queue() for that requirement.
1733 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1735 cancel_delayed_work(&hctx->run_work);
1737 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1739 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1742 * This function is often used for pausing .queue_rq() by driver when
1743 * there isn't enough resource or some conditions aren't satisfied, and
1744 * BLK_STS_RESOURCE is usually returned.
1746 * We do not guarantee that dispatch can be drained or blocked
1747 * after blk_mq_stop_hw_queues() returns. Please use
1748 * blk_mq_quiesce_queue() for that requirement.
1750 void blk_mq_stop_hw_queues(struct request_queue *q)
1752 struct blk_mq_hw_ctx *hctx;
1755 queue_for_each_hw_ctx(q, hctx, i)
1756 blk_mq_stop_hw_queue(hctx);
1758 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1760 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1762 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1764 blk_mq_run_hw_queue(hctx, false);
1766 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1768 void blk_mq_start_hw_queues(struct request_queue *q)
1770 struct blk_mq_hw_ctx *hctx;
1773 queue_for_each_hw_ctx(q, hctx, i)
1774 blk_mq_start_hw_queue(hctx);
1776 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1778 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1780 if (!blk_mq_hctx_stopped(hctx))
1783 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1784 blk_mq_run_hw_queue(hctx, async);
1786 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1788 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1790 struct blk_mq_hw_ctx *hctx;
1793 queue_for_each_hw_ctx(q, hctx, i)
1794 blk_mq_start_stopped_hw_queue(hctx, async);
1796 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1798 static void blk_mq_run_work_fn(struct work_struct *work)
1800 struct blk_mq_hw_ctx *hctx;
1802 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1805 * If we are stopped, don't run the queue.
1807 if (blk_mq_hctx_stopped(hctx))
1810 __blk_mq_run_hw_queue(hctx);
1813 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1817 struct blk_mq_ctx *ctx = rq->mq_ctx;
1818 enum hctx_type type = hctx->type;
1820 lockdep_assert_held(&ctx->lock);
1822 trace_block_rq_insert(hctx->queue, rq);
1825 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1827 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1830 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1833 struct blk_mq_ctx *ctx = rq->mq_ctx;
1835 lockdep_assert_held(&ctx->lock);
1837 __blk_mq_insert_req_list(hctx, rq, at_head);
1838 blk_mq_hctx_mark_pending(hctx, ctx);
1842 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1843 * @rq: Pointer to request to be inserted.
1844 * @at_head: true if the request should be inserted at the head of the list.
1845 * @run_queue: If we should run the hardware queue after inserting the request.
1847 * Should only be used carefully, when the caller knows we want to
1848 * bypass a potential IO scheduler on the target device.
1850 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1853 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1855 spin_lock(&hctx->lock);
1857 list_add(&rq->queuelist, &hctx->dispatch);
1859 list_add_tail(&rq->queuelist, &hctx->dispatch);
1860 spin_unlock(&hctx->lock);
1863 blk_mq_run_hw_queue(hctx, false);
1866 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1867 struct list_head *list)
1871 enum hctx_type type = hctx->type;
1874 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1877 list_for_each_entry(rq, list, queuelist) {
1878 BUG_ON(rq->mq_ctx != ctx);
1879 trace_block_rq_insert(hctx->queue, rq);
1882 spin_lock(&ctx->lock);
1883 list_splice_tail_init(list, &ctx->rq_lists[type]);
1884 blk_mq_hctx_mark_pending(hctx, ctx);
1885 spin_unlock(&ctx->lock);
1888 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1890 struct request *rqa = container_of(a, struct request, queuelist);
1891 struct request *rqb = container_of(b, struct request, queuelist);
1893 if (rqa->mq_ctx != rqb->mq_ctx)
1894 return rqa->mq_ctx > rqb->mq_ctx;
1895 if (rqa->mq_hctx != rqb->mq_hctx)
1896 return rqa->mq_hctx > rqb->mq_hctx;
1898 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1901 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1905 if (list_empty(&plug->mq_list))
1907 list_splice_init(&plug->mq_list, &list);
1909 if (plug->rq_count > 2 && plug->multiple_queues)
1910 list_sort(NULL, &list, plug_rq_cmp);
1915 struct list_head rq_list;
1916 struct request *rq, *head_rq = list_entry_rq(list.next);
1917 struct list_head *pos = &head_rq->queuelist; /* skip first */
1918 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1919 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1920 unsigned int depth = 1;
1922 list_for_each_continue(pos, &list) {
1923 rq = list_entry_rq(pos);
1925 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1930 list_cut_before(&rq_list, &list, pos);
1931 trace_block_unplug(head_rq->q, depth, !from_schedule);
1932 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1934 } while(!list_empty(&list));
1937 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1938 unsigned int nr_segs)
1942 if (bio->bi_opf & REQ_RAHEAD)
1943 rq->cmd_flags |= REQ_FAILFAST_MASK;
1945 rq->__sector = bio->bi_iter.bi_sector;
1946 rq->write_hint = bio->bi_write_hint;
1947 blk_rq_bio_prep(rq, bio, nr_segs);
1949 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
1950 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1953 blk_account_io_start(rq);
1956 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1958 blk_qc_t *cookie, bool last)
1960 struct request_queue *q = rq->q;
1961 struct blk_mq_queue_data bd = {
1965 blk_qc_t new_cookie;
1968 new_cookie = request_to_qc_t(hctx, rq);
1971 * For OK queue, we are done. For error, caller may kill it.
1972 * Any other error (busy), just add it to our list as we
1973 * previously would have done.
1975 ret = q->mq_ops->queue_rq(hctx, &bd);
1978 blk_mq_update_dispatch_busy(hctx, false);
1979 *cookie = new_cookie;
1981 case BLK_STS_RESOURCE:
1982 case BLK_STS_DEV_RESOURCE:
1983 blk_mq_update_dispatch_busy(hctx, true);
1984 __blk_mq_requeue_request(rq);
1987 blk_mq_update_dispatch_busy(hctx, false);
1988 *cookie = BLK_QC_T_NONE;
1995 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1998 bool bypass_insert, bool last)
2000 struct request_queue *q = rq->q;
2001 bool run_queue = true;
2004 * RCU or SRCU read lock is needed before checking quiesced flag.
2006 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2007 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2008 * and avoid driver to try to dispatch again.
2010 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2012 bypass_insert = false;
2016 if (q->elevator && !bypass_insert)
2019 if (!blk_mq_get_dispatch_budget(q))
2022 if (!blk_mq_get_driver_tag(rq)) {
2023 blk_mq_put_dispatch_budget(q);
2027 return __blk_mq_issue_directly(hctx, rq, cookie, last);
2030 return BLK_STS_RESOURCE;
2032 blk_mq_sched_insert_request(rq, false, run_queue, false);
2038 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2039 * @hctx: Pointer of the associated hardware queue.
2040 * @rq: Pointer to request to be sent.
2041 * @cookie: Request queue cookie.
2043 * If the device has enough resources to accept a new request now, send the
2044 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2045 * we can try send it another time in the future. Requests inserted at this
2046 * queue have higher priority.
2048 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2049 struct request *rq, blk_qc_t *cookie)
2054 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2056 hctx_lock(hctx, &srcu_idx);
2058 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
2059 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2060 blk_mq_request_bypass_insert(rq, false, true);
2061 else if (ret != BLK_STS_OK)
2062 blk_mq_end_request(rq, ret);
2064 hctx_unlock(hctx, srcu_idx);
2067 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2071 blk_qc_t unused_cookie;
2072 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2074 hctx_lock(hctx, &srcu_idx);
2075 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
2076 hctx_unlock(hctx, srcu_idx);
2081 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2082 struct list_head *list)
2087 while (!list_empty(list)) {
2089 struct request *rq = list_first_entry(list, struct request,
2092 list_del_init(&rq->queuelist);
2093 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2094 if (ret != BLK_STS_OK) {
2095 if (ret == BLK_STS_RESOURCE ||
2096 ret == BLK_STS_DEV_RESOURCE) {
2097 blk_mq_request_bypass_insert(rq, false,
2101 blk_mq_end_request(rq, ret);
2108 * If we didn't flush the entire list, we could have told
2109 * the driver there was more coming, but that turned out to
2112 if ((!list_empty(list) || errors) &&
2113 hctx->queue->mq_ops->commit_rqs && queued)
2114 hctx->queue->mq_ops->commit_rqs(hctx);
2117 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2119 list_add_tail(&rq->queuelist, &plug->mq_list);
2121 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
2122 struct request *tmp;
2124 tmp = list_first_entry(&plug->mq_list, struct request,
2126 if (tmp->q != rq->q)
2127 plug->multiple_queues = true;
2132 * blk_mq_submit_bio - Create and send a request to block device.
2133 * @bio: Bio pointer.
2135 * Builds up a request structure from @q and @bio and send to the device. The
2136 * request may not be queued directly to hardware if:
2137 * * This request can be merged with another one
2138 * * We want to place request at plug queue for possible future merging
2139 * * There is an IO scheduler active at this queue
2141 * It will not queue the request if there is an error with the bio, or at the
2144 * Returns: Request queue cookie.
2146 blk_qc_t blk_mq_submit_bio(struct bio *bio)
2148 struct request_queue *q = bio->bi_disk->queue;
2149 const int is_sync = op_is_sync(bio->bi_opf);
2150 const int is_flush_fua = op_is_flush(bio->bi_opf);
2151 struct blk_mq_alloc_data data = {
2155 struct blk_plug *plug;
2156 struct request *same_queue_rq = NULL;
2157 unsigned int nr_segs;
2161 blk_queue_bounce(q, &bio);
2162 __blk_queue_split(&bio, &nr_segs);
2164 if (!bio_integrity_prep(bio))
2167 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2168 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2171 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2174 rq_qos_throttle(q, bio);
2176 data.cmd_flags = bio->bi_opf;
2177 rq = __blk_mq_alloc_request(&data);
2178 if (unlikely(!rq)) {
2179 rq_qos_cleanup(q, bio);
2180 if (bio->bi_opf & REQ_NOWAIT)
2181 bio_wouldblock_error(bio);
2185 trace_block_getrq(q, bio, bio->bi_opf);
2187 rq_qos_track(q, rq, bio);
2189 cookie = request_to_qc_t(data.hctx, rq);
2191 blk_mq_bio_to_request(rq, bio, nr_segs);
2193 ret = blk_crypto_init_request(rq);
2194 if (ret != BLK_STS_OK) {
2195 bio->bi_status = ret;
2197 blk_mq_free_request(rq);
2198 return BLK_QC_T_NONE;
2201 plug = blk_mq_plug(q, bio);
2202 if (unlikely(is_flush_fua)) {
2203 /* Bypass scheduler for flush requests */
2204 blk_insert_flush(rq);
2205 blk_mq_run_hw_queue(data.hctx, true);
2206 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2207 !blk_queue_nonrot(q))) {
2209 * Use plugging if we have a ->commit_rqs() hook as well, as
2210 * we know the driver uses bd->last in a smart fashion.
2212 * Use normal plugging if this disk is slow HDD, as sequential
2213 * IO may benefit a lot from plug merging.
2215 unsigned int request_count = plug->rq_count;
2216 struct request *last = NULL;
2219 trace_block_plug(q);
2221 last = list_entry_rq(plug->mq_list.prev);
2223 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2224 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2225 blk_flush_plug_list(plug, false);
2226 trace_block_plug(q);
2229 blk_add_rq_to_plug(plug, rq);
2230 } else if (q->elevator) {
2231 /* Insert the request at the IO scheduler queue */
2232 blk_mq_sched_insert_request(rq, false, true, true);
2233 } else if (plug && !blk_queue_nomerges(q)) {
2235 * We do limited plugging. If the bio can be merged, do that.
2236 * Otherwise the existing request in the plug list will be
2237 * issued. So the plug list will have one request at most
2238 * The plug list might get flushed before this. If that happens,
2239 * the plug list is empty, and same_queue_rq is invalid.
2241 if (list_empty(&plug->mq_list))
2242 same_queue_rq = NULL;
2243 if (same_queue_rq) {
2244 list_del_init(&same_queue_rq->queuelist);
2247 blk_add_rq_to_plug(plug, rq);
2248 trace_block_plug(q);
2250 if (same_queue_rq) {
2251 data.hctx = same_queue_rq->mq_hctx;
2252 trace_block_unplug(q, 1, true);
2253 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2256 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2257 !data.hctx->dispatch_busy) {
2259 * There is no scheduler and we can try to send directly
2262 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2265 blk_mq_sched_insert_request(rq, false, true, true);
2271 return BLK_QC_T_NONE;
2274 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2275 unsigned int hctx_idx)
2279 if (tags->rqs && set->ops->exit_request) {
2282 for (i = 0; i < tags->nr_tags; i++) {
2283 struct request *rq = tags->static_rqs[i];
2287 set->ops->exit_request(set, rq, hctx_idx);
2288 tags->static_rqs[i] = NULL;
2292 while (!list_empty(&tags->page_list)) {
2293 page = list_first_entry(&tags->page_list, struct page, lru);
2294 list_del_init(&page->lru);
2296 * Remove kmemleak object previously allocated in
2297 * blk_mq_alloc_rqs().
2299 kmemleak_free(page_address(page));
2300 __free_pages(page, page->private);
2304 void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags)
2308 kfree(tags->static_rqs);
2309 tags->static_rqs = NULL;
2311 blk_mq_free_tags(tags, flags);
2314 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2315 unsigned int hctx_idx,
2316 unsigned int nr_tags,
2317 unsigned int reserved_tags,
2320 struct blk_mq_tags *tags;
2323 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2324 if (node == NUMA_NO_NODE)
2325 node = set->numa_node;
2327 tags = blk_mq_init_tags(nr_tags, reserved_tags, node, flags);
2331 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2332 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2335 blk_mq_free_tags(tags, flags);
2339 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2340 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2342 if (!tags->static_rqs) {
2344 blk_mq_free_tags(tags, flags);
2351 static size_t order_to_size(unsigned int order)
2353 return (size_t)PAGE_SIZE << order;
2356 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2357 unsigned int hctx_idx, int node)
2361 if (set->ops->init_request) {
2362 ret = set->ops->init_request(set, rq, hctx_idx, node);
2367 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2371 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2372 unsigned int hctx_idx, unsigned int depth)
2374 unsigned int i, j, entries_per_page, max_order = 4;
2375 size_t rq_size, left;
2378 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2379 if (node == NUMA_NO_NODE)
2380 node = set->numa_node;
2382 INIT_LIST_HEAD(&tags->page_list);
2385 * rq_size is the size of the request plus driver payload, rounded
2386 * to the cacheline size
2388 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2390 left = rq_size * depth;
2392 for (i = 0; i < depth; ) {
2393 int this_order = max_order;
2398 while (this_order && left < order_to_size(this_order - 1))
2402 page = alloc_pages_node(node,
2403 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2409 if (order_to_size(this_order) < rq_size)
2416 page->private = this_order;
2417 list_add_tail(&page->lru, &tags->page_list);
2419 p = page_address(page);
2421 * Allow kmemleak to scan these pages as they contain pointers
2422 * to additional allocations like via ops->init_request().
2424 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2425 entries_per_page = order_to_size(this_order) / rq_size;
2426 to_do = min(entries_per_page, depth - i);
2427 left -= to_do * rq_size;
2428 for (j = 0; j < to_do; j++) {
2429 struct request *rq = p;
2431 tags->static_rqs[i] = rq;
2432 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2433 tags->static_rqs[i] = NULL;
2444 blk_mq_free_rqs(set, tags, hctx_idx);
2448 struct rq_iter_data {
2449 struct blk_mq_hw_ctx *hctx;
2453 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2455 struct rq_iter_data *iter_data = data;
2457 if (rq->mq_hctx != iter_data->hctx)
2459 iter_data->has_rq = true;
2463 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2465 struct blk_mq_tags *tags = hctx->sched_tags ?
2466 hctx->sched_tags : hctx->tags;
2467 struct rq_iter_data data = {
2471 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2475 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2476 struct blk_mq_hw_ctx *hctx)
2478 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2480 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2485 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2487 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2488 struct blk_mq_hw_ctx, cpuhp_online);
2490 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2491 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2495 * Prevent new request from being allocated on the current hctx.
2497 * The smp_mb__after_atomic() Pairs with the implied barrier in
2498 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2499 * seen once we return from the tag allocator.
2501 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2502 smp_mb__after_atomic();
2505 * Try to grab a reference to the queue and wait for any outstanding
2506 * requests. If we could not grab a reference the queue has been
2507 * frozen and there are no requests.
2509 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2510 while (blk_mq_hctx_has_requests(hctx))
2512 percpu_ref_put(&hctx->queue->q_usage_counter);
2518 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2520 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2521 struct blk_mq_hw_ctx, cpuhp_online);
2523 if (cpumask_test_cpu(cpu, hctx->cpumask))
2524 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2529 * 'cpu' is going away. splice any existing rq_list entries from this
2530 * software queue to the hw queue dispatch list, and ensure that it
2533 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2535 struct blk_mq_hw_ctx *hctx;
2536 struct blk_mq_ctx *ctx;
2538 enum hctx_type type;
2540 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2541 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2544 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2547 spin_lock(&ctx->lock);
2548 if (!list_empty(&ctx->rq_lists[type])) {
2549 list_splice_init(&ctx->rq_lists[type], &tmp);
2550 blk_mq_hctx_clear_pending(hctx, ctx);
2552 spin_unlock(&ctx->lock);
2554 if (list_empty(&tmp))
2557 spin_lock(&hctx->lock);
2558 list_splice_tail_init(&tmp, &hctx->dispatch);
2559 spin_unlock(&hctx->lock);
2561 blk_mq_run_hw_queue(hctx, true);
2565 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2567 if (!(hctx->flags & BLK_MQ_F_STACKING))
2568 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2569 &hctx->cpuhp_online);
2570 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2574 /* hctx->ctxs will be freed in queue's release handler */
2575 static void blk_mq_exit_hctx(struct request_queue *q,
2576 struct blk_mq_tag_set *set,
2577 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2579 if (blk_mq_hw_queue_mapped(hctx))
2580 blk_mq_tag_idle(hctx);
2582 if (set->ops->exit_request)
2583 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2585 if (set->ops->exit_hctx)
2586 set->ops->exit_hctx(hctx, hctx_idx);
2588 blk_mq_remove_cpuhp(hctx);
2590 spin_lock(&q->unused_hctx_lock);
2591 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2592 spin_unlock(&q->unused_hctx_lock);
2595 static void blk_mq_exit_hw_queues(struct request_queue *q,
2596 struct blk_mq_tag_set *set, int nr_queue)
2598 struct blk_mq_hw_ctx *hctx;
2601 queue_for_each_hw_ctx(q, hctx, i) {
2604 blk_mq_debugfs_unregister_hctx(hctx);
2605 blk_mq_exit_hctx(q, set, hctx, i);
2609 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2611 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2613 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2614 __alignof__(struct blk_mq_hw_ctx)) !=
2615 sizeof(struct blk_mq_hw_ctx));
2617 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2618 hw_ctx_size += sizeof(struct srcu_struct);
2623 static int blk_mq_init_hctx(struct request_queue *q,
2624 struct blk_mq_tag_set *set,
2625 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2627 hctx->queue_num = hctx_idx;
2629 if (!(hctx->flags & BLK_MQ_F_STACKING))
2630 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2631 &hctx->cpuhp_online);
2632 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2634 hctx->tags = set->tags[hctx_idx];
2636 if (set->ops->init_hctx &&
2637 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2638 goto unregister_cpu_notifier;
2640 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2646 if (set->ops->exit_hctx)
2647 set->ops->exit_hctx(hctx, hctx_idx);
2648 unregister_cpu_notifier:
2649 blk_mq_remove_cpuhp(hctx);
2653 static struct blk_mq_hw_ctx *
2654 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2657 struct blk_mq_hw_ctx *hctx;
2658 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2660 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2662 goto fail_alloc_hctx;
2664 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2667 atomic_set(&hctx->nr_active, 0);
2668 atomic_set(&hctx->elevator_queued, 0);
2669 if (node == NUMA_NO_NODE)
2670 node = set->numa_node;
2671 hctx->numa_node = node;
2673 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2674 spin_lock_init(&hctx->lock);
2675 INIT_LIST_HEAD(&hctx->dispatch);
2677 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
2679 INIT_LIST_HEAD(&hctx->hctx_list);
2682 * Allocate space for all possible cpus to avoid allocation at
2685 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2690 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2695 spin_lock_init(&hctx->dispatch_wait_lock);
2696 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2697 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2699 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2703 if (hctx->flags & BLK_MQ_F_BLOCKING)
2704 init_srcu_struct(hctx->srcu);
2705 blk_mq_hctx_kobj_init(hctx);
2710 sbitmap_free(&hctx->ctx_map);
2714 free_cpumask_var(hctx->cpumask);
2721 static void blk_mq_init_cpu_queues(struct request_queue *q,
2722 unsigned int nr_hw_queues)
2724 struct blk_mq_tag_set *set = q->tag_set;
2727 for_each_possible_cpu(i) {
2728 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2729 struct blk_mq_hw_ctx *hctx;
2733 spin_lock_init(&__ctx->lock);
2734 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2735 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2740 * Set local node, IFF we have more than one hw queue. If
2741 * not, we remain on the home node of the device
2743 for (j = 0; j < set->nr_maps; j++) {
2744 hctx = blk_mq_map_queue_type(q, j, i);
2745 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2746 hctx->numa_node = local_memory_node(cpu_to_node(i));
2751 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2754 unsigned int flags = set->flags;
2757 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2758 set->queue_depth, set->reserved_tags, flags);
2759 if (!set->tags[hctx_idx])
2762 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2767 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2768 set->tags[hctx_idx] = NULL;
2772 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2773 unsigned int hctx_idx)
2775 unsigned int flags = set->flags;
2777 if (set->tags && set->tags[hctx_idx]) {
2778 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2779 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2780 set->tags[hctx_idx] = NULL;
2784 static void blk_mq_map_swqueue(struct request_queue *q)
2786 unsigned int i, j, hctx_idx;
2787 struct blk_mq_hw_ctx *hctx;
2788 struct blk_mq_ctx *ctx;
2789 struct blk_mq_tag_set *set = q->tag_set;
2791 queue_for_each_hw_ctx(q, hctx, i) {
2792 cpumask_clear(hctx->cpumask);
2794 hctx->dispatch_from = NULL;
2798 * Map software to hardware queues.
2800 * If the cpu isn't present, the cpu is mapped to first hctx.
2802 for_each_possible_cpu(i) {
2804 ctx = per_cpu_ptr(q->queue_ctx, i);
2805 for (j = 0; j < set->nr_maps; j++) {
2806 if (!set->map[j].nr_queues) {
2807 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2808 HCTX_TYPE_DEFAULT, i);
2811 hctx_idx = set->map[j].mq_map[i];
2812 /* unmapped hw queue can be remapped after CPU topo changed */
2813 if (!set->tags[hctx_idx] &&
2814 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2816 * If tags initialization fail for some hctx,
2817 * that hctx won't be brought online. In this
2818 * case, remap the current ctx to hctx[0] which
2819 * is guaranteed to always have tags allocated
2821 set->map[j].mq_map[i] = 0;
2824 hctx = blk_mq_map_queue_type(q, j, i);
2825 ctx->hctxs[j] = hctx;
2827 * If the CPU is already set in the mask, then we've
2828 * mapped this one already. This can happen if
2829 * devices share queues across queue maps.
2831 if (cpumask_test_cpu(i, hctx->cpumask))
2834 cpumask_set_cpu(i, hctx->cpumask);
2836 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2837 hctx->ctxs[hctx->nr_ctx++] = ctx;
2840 * If the nr_ctx type overflows, we have exceeded the
2841 * amount of sw queues we can support.
2843 BUG_ON(!hctx->nr_ctx);
2846 for (; j < HCTX_MAX_TYPES; j++)
2847 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2848 HCTX_TYPE_DEFAULT, i);
2851 queue_for_each_hw_ctx(q, hctx, i) {
2853 * If no software queues are mapped to this hardware queue,
2854 * disable it and free the request entries.
2856 if (!hctx->nr_ctx) {
2857 /* Never unmap queue 0. We need it as a
2858 * fallback in case of a new remap fails
2861 if (i && set->tags[i])
2862 blk_mq_free_map_and_requests(set, i);
2868 hctx->tags = set->tags[i];
2869 WARN_ON(!hctx->tags);
2872 * Set the map size to the number of mapped software queues.
2873 * This is more accurate and more efficient than looping
2874 * over all possibly mapped software queues.
2876 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2879 * Initialize batch roundrobin counts
2881 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2882 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2887 * Caller needs to ensure that we're either frozen/quiesced, or that
2888 * the queue isn't live yet.
2890 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2892 struct blk_mq_hw_ctx *hctx;
2895 queue_for_each_hw_ctx(q, hctx, i) {
2897 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
2899 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
2903 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
2906 struct request_queue *q;
2908 lockdep_assert_held(&set->tag_list_lock);
2910 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2911 blk_mq_freeze_queue(q);
2912 queue_set_hctx_shared(q, shared);
2913 blk_mq_unfreeze_queue(q);
2917 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2919 struct blk_mq_tag_set *set = q->tag_set;
2921 mutex_lock(&set->tag_list_lock);
2922 list_del(&q->tag_set_list);
2923 if (list_is_singular(&set->tag_list)) {
2924 /* just transitioned to unshared */
2925 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
2926 /* update existing queue */
2927 blk_mq_update_tag_set_shared(set, false);
2929 mutex_unlock(&set->tag_list_lock);
2930 INIT_LIST_HEAD(&q->tag_set_list);
2933 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2934 struct request_queue *q)
2936 mutex_lock(&set->tag_list_lock);
2939 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2941 if (!list_empty(&set->tag_list) &&
2942 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2943 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
2944 /* update existing queue */
2945 blk_mq_update_tag_set_shared(set, true);
2947 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
2948 queue_set_hctx_shared(q, true);
2949 list_add_tail(&q->tag_set_list, &set->tag_list);
2951 mutex_unlock(&set->tag_list_lock);
2954 /* All allocations will be freed in release handler of q->mq_kobj */
2955 static int blk_mq_alloc_ctxs(struct request_queue *q)
2957 struct blk_mq_ctxs *ctxs;
2960 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2964 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2965 if (!ctxs->queue_ctx)
2968 for_each_possible_cpu(cpu) {
2969 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2973 q->mq_kobj = &ctxs->kobj;
2974 q->queue_ctx = ctxs->queue_ctx;
2983 * It is the actual release handler for mq, but we do it from
2984 * request queue's release handler for avoiding use-after-free
2985 * and headache because q->mq_kobj shouldn't have been introduced,
2986 * but we can't group ctx/kctx kobj without it.
2988 void blk_mq_release(struct request_queue *q)
2990 struct blk_mq_hw_ctx *hctx, *next;
2993 queue_for_each_hw_ctx(q, hctx, i)
2994 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2996 /* all hctx are in .unused_hctx_list now */
2997 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2998 list_del_init(&hctx->hctx_list);
2999 kobject_put(&hctx->kobj);
3002 kfree(q->queue_hw_ctx);
3005 * release .mq_kobj and sw queue's kobject now because
3006 * both share lifetime with request queue.
3008 blk_mq_sysfs_deinit(q);
3011 struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3014 struct request_queue *uninit_q, *q;
3016 uninit_q = blk_alloc_queue(set->numa_node);
3018 return ERR_PTR(-ENOMEM);
3019 uninit_q->queuedata = queuedata;
3022 * Initialize the queue without an elevator. device_add_disk() will do
3023 * the initialization.
3025 q = blk_mq_init_allocated_queue(set, uninit_q, false);
3027 blk_cleanup_queue(uninit_q);
3031 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data);
3033 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3035 return blk_mq_init_queue_data(set, NULL);
3037 EXPORT_SYMBOL(blk_mq_init_queue);
3040 * Helper for setting up a queue with mq ops, given queue depth, and
3041 * the passed in mq ops flags.
3043 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
3044 const struct blk_mq_ops *ops,
3045 unsigned int queue_depth,
3046 unsigned int set_flags)
3048 struct request_queue *q;
3051 memset(set, 0, sizeof(*set));
3053 set->nr_hw_queues = 1;
3055 set->queue_depth = queue_depth;
3056 set->numa_node = NUMA_NO_NODE;
3057 set->flags = set_flags;
3059 ret = blk_mq_alloc_tag_set(set);
3061 return ERR_PTR(ret);
3063 q = blk_mq_init_queue(set);
3065 blk_mq_free_tag_set(set);
3071 EXPORT_SYMBOL(blk_mq_init_sq_queue);
3073 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3074 struct blk_mq_tag_set *set, struct request_queue *q,
3075 int hctx_idx, int node)
3077 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3079 /* reuse dead hctx first */
3080 spin_lock(&q->unused_hctx_lock);
3081 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3082 if (tmp->numa_node == node) {
3088 list_del_init(&hctx->hctx_list);
3089 spin_unlock(&q->unused_hctx_lock);
3092 hctx = blk_mq_alloc_hctx(q, set, node);
3096 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3102 kobject_put(&hctx->kobj);
3107 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3108 struct request_queue *q)
3111 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3113 if (q->nr_hw_queues < set->nr_hw_queues) {
3114 struct blk_mq_hw_ctx **new_hctxs;
3116 new_hctxs = kcalloc_node(set->nr_hw_queues,
3117 sizeof(*new_hctxs), GFP_KERNEL,
3122 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3124 q->queue_hw_ctx = new_hctxs;
3129 /* protect against switching io scheduler */
3130 mutex_lock(&q->sysfs_lock);
3131 for (i = 0; i < set->nr_hw_queues; i++) {
3133 struct blk_mq_hw_ctx *hctx;
3135 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3137 * If the hw queue has been mapped to another numa node,
3138 * we need to realloc the hctx. If allocation fails, fallback
3139 * to use the previous one.
3141 if (hctxs[i] && (hctxs[i]->numa_node == node))
3144 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3147 blk_mq_exit_hctx(q, set, hctxs[i], i);
3151 pr_warn("Allocate new hctx on node %d fails,\
3152 fallback to previous one on node %d\n",
3153 node, hctxs[i]->numa_node);
3159 * Increasing nr_hw_queues fails. Free the newly allocated
3160 * hctxs and keep the previous q->nr_hw_queues.
3162 if (i != set->nr_hw_queues) {
3163 j = q->nr_hw_queues;
3167 end = q->nr_hw_queues;
3168 q->nr_hw_queues = set->nr_hw_queues;
3171 for (; j < end; j++) {
3172 struct blk_mq_hw_ctx *hctx = hctxs[j];
3176 blk_mq_free_map_and_requests(set, j);
3177 blk_mq_exit_hctx(q, set, hctx, j);
3181 mutex_unlock(&q->sysfs_lock);
3184 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3185 struct request_queue *q,
3188 /* mark the queue as mq asap */
3189 q->mq_ops = set->ops;
3191 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3192 blk_mq_poll_stats_bkt,
3193 BLK_MQ_POLL_STATS_BKTS, q);
3197 if (blk_mq_alloc_ctxs(q))
3200 /* init q->mq_kobj and sw queues' kobjects */
3201 blk_mq_sysfs_init(q);
3203 INIT_LIST_HEAD(&q->unused_hctx_list);
3204 spin_lock_init(&q->unused_hctx_lock);
3206 blk_mq_realloc_hw_ctxs(set, q);
3207 if (!q->nr_hw_queues)
3210 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3211 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3215 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3216 if (set->nr_maps > HCTX_TYPE_POLL &&
3217 set->map[HCTX_TYPE_POLL].nr_queues)
3218 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3220 q->sg_reserved_size = INT_MAX;
3222 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3223 INIT_LIST_HEAD(&q->requeue_list);
3224 spin_lock_init(&q->requeue_lock);
3226 q->nr_requests = set->queue_depth;
3229 * Default to classic polling
3231 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3233 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3234 blk_mq_add_queue_tag_set(set, q);
3235 blk_mq_map_swqueue(q);
3238 elevator_init_mq(q);
3243 kfree(q->queue_hw_ctx);
3244 q->nr_hw_queues = 0;
3245 blk_mq_sysfs_deinit(q);
3247 blk_stat_free_callback(q->poll_cb);
3251 return ERR_PTR(-ENOMEM);
3253 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3255 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3256 void blk_mq_exit_queue(struct request_queue *q)
3258 struct blk_mq_tag_set *set = q->tag_set;
3260 blk_mq_del_queue_tag_set(q);
3261 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3264 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3268 for (i = 0; i < set->nr_hw_queues; i++) {
3269 if (!__blk_mq_alloc_map_and_request(set, i))
3278 blk_mq_free_map_and_requests(set, i);
3284 * Allocate the request maps associated with this tag_set. Note that this
3285 * may reduce the depth asked for, if memory is tight. set->queue_depth
3286 * will be updated to reflect the allocated depth.
3288 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3293 depth = set->queue_depth;
3295 err = __blk_mq_alloc_rq_maps(set);
3299 set->queue_depth >>= 1;
3300 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3304 } while (set->queue_depth);
3306 if (!set->queue_depth || err) {
3307 pr_err("blk-mq: failed to allocate request map\n");
3311 if (depth != set->queue_depth)
3312 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3313 depth, set->queue_depth);
3318 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3321 * blk_mq_map_queues() and multiple .map_queues() implementations
3322 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3323 * number of hardware queues.
3325 if (set->nr_maps == 1)
3326 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3328 if (set->ops->map_queues && !is_kdump_kernel()) {
3332 * transport .map_queues is usually done in the following
3335 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3336 * mask = get_cpu_mask(queue)
3337 * for_each_cpu(cpu, mask)
3338 * set->map[x].mq_map[cpu] = queue;
3341 * When we need to remap, the table has to be cleared for
3342 * killing stale mapping since one CPU may not be mapped
3345 for (i = 0; i < set->nr_maps; i++)
3346 blk_mq_clear_mq_map(&set->map[i]);
3348 return set->ops->map_queues(set);
3350 BUG_ON(set->nr_maps > 1);
3351 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3355 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3356 int cur_nr_hw_queues, int new_nr_hw_queues)
3358 struct blk_mq_tags **new_tags;
3360 if (cur_nr_hw_queues >= new_nr_hw_queues)
3363 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3364 GFP_KERNEL, set->numa_node);
3369 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3370 sizeof(*set->tags));
3372 set->tags = new_tags;
3373 set->nr_hw_queues = new_nr_hw_queues;
3379 * Alloc a tag set to be associated with one or more request queues.
3380 * May fail with EINVAL for various error conditions. May adjust the
3381 * requested depth down, if it's too large. In that case, the set
3382 * value will be stored in set->queue_depth.
3384 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3388 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3390 if (!set->nr_hw_queues)
3392 if (!set->queue_depth)
3394 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3397 if (!set->ops->queue_rq)
3400 if (!set->ops->get_budget ^ !set->ops->put_budget)
3403 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3404 pr_info("blk-mq: reduced tag depth to %u\n",
3406 set->queue_depth = BLK_MQ_MAX_DEPTH;
3411 else if (set->nr_maps > HCTX_MAX_TYPES)
3415 * If a crashdump is active, then we are potentially in a very
3416 * memory constrained environment. Limit us to 1 queue and
3417 * 64 tags to prevent using too much memory.
3419 if (is_kdump_kernel()) {
3420 set->nr_hw_queues = 1;
3422 set->queue_depth = min(64U, set->queue_depth);
3425 * There is no use for more h/w queues than cpus if we just have
3428 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3429 set->nr_hw_queues = nr_cpu_ids;
3431 if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0)
3435 for (i = 0; i < set->nr_maps; i++) {
3436 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3437 sizeof(set->map[i].mq_map[0]),
3438 GFP_KERNEL, set->numa_node);
3439 if (!set->map[i].mq_map)
3440 goto out_free_mq_map;
3441 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3444 ret = blk_mq_update_queue_map(set);
3446 goto out_free_mq_map;
3448 ret = blk_mq_alloc_map_and_requests(set);
3450 goto out_free_mq_map;
3452 if (blk_mq_is_sbitmap_shared(set->flags)) {
3453 atomic_set(&set->active_queues_shared_sbitmap, 0);
3455 if (blk_mq_init_shared_sbitmap(set, set->flags)) {
3457 goto out_free_mq_rq_maps;
3461 mutex_init(&set->tag_list_lock);
3462 INIT_LIST_HEAD(&set->tag_list);
3466 out_free_mq_rq_maps:
3467 for (i = 0; i < set->nr_hw_queues; i++)
3468 blk_mq_free_map_and_requests(set, i);
3470 for (i = 0; i < set->nr_maps; i++) {
3471 kfree(set->map[i].mq_map);
3472 set->map[i].mq_map = NULL;
3478 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3480 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3484 for (i = 0; i < set->nr_hw_queues; i++)
3485 blk_mq_free_map_and_requests(set, i);
3487 if (blk_mq_is_sbitmap_shared(set->flags))
3488 blk_mq_exit_shared_sbitmap(set);
3490 for (j = 0; j < set->nr_maps; j++) {
3491 kfree(set->map[j].mq_map);
3492 set->map[j].mq_map = NULL;
3498 EXPORT_SYMBOL(blk_mq_free_tag_set);
3500 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3502 struct blk_mq_tag_set *set = q->tag_set;
3503 struct blk_mq_hw_ctx *hctx;
3509 if (q->nr_requests == nr)
3512 blk_mq_freeze_queue(q);
3513 blk_mq_quiesce_queue(q);
3516 queue_for_each_hw_ctx(q, hctx, i) {
3520 * If we're using an MQ scheduler, just update the scheduler
3521 * queue depth. This is similar to what the old code would do.
3523 if (!hctx->sched_tags) {
3524 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3526 if (!ret && blk_mq_is_sbitmap_shared(set->flags))
3527 blk_mq_tag_resize_shared_sbitmap(set, nr);
3529 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3534 if (q->elevator && q->elevator->type->ops.depth_updated)
3535 q->elevator->type->ops.depth_updated(hctx);
3539 q->nr_requests = nr;
3541 blk_mq_unquiesce_queue(q);
3542 blk_mq_unfreeze_queue(q);
3548 * request_queue and elevator_type pair.
3549 * It is just used by __blk_mq_update_nr_hw_queues to cache
3550 * the elevator_type associated with a request_queue.
3552 struct blk_mq_qe_pair {
3553 struct list_head node;
3554 struct request_queue *q;
3555 struct elevator_type *type;
3559 * Cache the elevator_type in qe pair list and switch the
3560 * io scheduler to 'none'
3562 static bool blk_mq_elv_switch_none(struct list_head *head,
3563 struct request_queue *q)
3565 struct blk_mq_qe_pair *qe;
3570 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3574 INIT_LIST_HEAD(&qe->node);
3576 qe->type = q->elevator->type;
3577 list_add(&qe->node, head);
3579 mutex_lock(&q->sysfs_lock);
3581 * After elevator_switch_mq, the previous elevator_queue will be
3582 * released by elevator_release. The reference of the io scheduler
3583 * module get by elevator_get will also be put. So we need to get
3584 * a reference of the io scheduler module here to prevent it to be
3587 __module_get(qe->type->elevator_owner);
3588 elevator_switch_mq(q, NULL);
3589 mutex_unlock(&q->sysfs_lock);
3594 static void blk_mq_elv_switch_back(struct list_head *head,
3595 struct request_queue *q)
3597 struct blk_mq_qe_pair *qe;
3598 struct elevator_type *t = NULL;
3600 list_for_each_entry(qe, head, node)
3609 list_del(&qe->node);
3612 mutex_lock(&q->sysfs_lock);
3613 elevator_switch_mq(q, t);
3614 mutex_unlock(&q->sysfs_lock);
3617 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3620 struct request_queue *q;
3622 int prev_nr_hw_queues;
3624 lockdep_assert_held(&set->tag_list_lock);
3626 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3627 nr_hw_queues = nr_cpu_ids;
3628 if (nr_hw_queues < 1)
3630 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3633 list_for_each_entry(q, &set->tag_list, tag_set_list)
3634 blk_mq_freeze_queue(q);
3636 * Switch IO scheduler to 'none', cleaning up the data associated
3637 * with the previous scheduler. We will switch back once we are done
3638 * updating the new sw to hw queue mappings.
3640 list_for_each_entry(q, &set->tag_list, tag_set_list)
3641 if (!blk_mq_elv_switch_none(&head, q))
3644 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3645 blk_mq_debugfs_unregister_hctxs(q);
3646 blk_mq_sysfs_unregister(q);
3649 prev_nr_hw_queues = set->nr_hw_queues;
3650 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3654 set->nr_hw_queues = nr_hw_queues;
3656 blk_mq_update_queue_map(set);
3657 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3658 blk_mq_realloc_hw_ctxs(set, q);
3659 if (q->nr_hw_queues != set->nr_hw_queues) {
3660 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3661 nr_hw_queues, prev_nr_hw_queues);
3662 set->nr_hw_queues = prev_nr_hw_queues;
3663 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3666 blk_mq_map_swqueue(q);
3670 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3671 blk_mq_sysfs_register(q);
3672 blk_mq_debugfs_register_hctxs(q);
3676 list_for_each_entry(q, &set->tag_list, tag_set_list)
3677 blk_mq_elv_switch_back(&head, q);
3679 list_for_each_entry(q, &set->tag_list, tag_set_list)
3680 blk_mq_unfreeze_queue(q);
3683 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3685 mutex_lock(&set->tag_list_lock);
3686 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3687 mutex_unlock(&set->tag_list_lock);
3689 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3691 /* Enable polling stats and return whether they were already enabled. */
3692 static bool blk_poll_stats_enable(struct request_queue *q)
3694 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3695 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3697 blk_stat_add_callback(q, q->poll_cb);
3701 static void blk_mq_poll_stats_start(struct request_queue *q)
3704 * We don't arm the callback if polling stats are not enabled or the
3705 * callback is already active.
3707 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3708 blk_stat_is_active(q->poll_cb))
3711 blk_stat_activate_msecs(q->poll_cb, 100);
3714 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3716 struct request_queue *q = cb->data;
3719 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3720 if (cb->stat[bucket].nr_samples)
3721 q->poll_stat[bucket] = cb->stat[bucket];
3725 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3728 unsigned long ret = 0;
3732 * If stats collection isn't on, don't sleep but turn it on for
3735 if (!blk_poll_stats_enable(q))
3739 * As an optimistic guess, use half of the mean service time
3740 * for this type of request. We can (and should) make this smarter.
3741 * For instance, if the completion latencies are tight, we can
3742 * get closer than just half the mean. This is especially
3743 * important on devices where the completion latencies are longer
3744 * than ~10 usec. We do use the stats for the relevant IO size
3745 * if available which does lead to better estimates.
3747 bucket = blk_mq_poll_stats_bkt(rq);
3751 if (q->poll_stat[bucket].nr_samples)
3752 ret = (q->poll_stat[bucket].mean + 1) / 2;
3757 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3760 struct hrtimer_sleeper hs;
3761 enum hrtimer_mode mode;
3765 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3769 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3771 * 0: use half of prev avg
3772 * >0: use this specific value
3774 if (q->poll_nsec > 0)
3775 nsecs = q->poll_nsec;
3777 nsecs = blk_mq_poll_nsecs(q, rq);
3782 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3785 * This will be replaced with the stats tracking code, using
3786 * 'avg_completion_time / 2' as the pre-sleep target.
3790 mode = HRTIMER_MODE_REL;
3791 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3792 hrtimer_set_expires(&hs.timer, kt);
3795 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3797 set_current_state(TASK_UNINTERRUPTIBLE);
3798 hrtimer_sleeper_start_expires(&hs, mode);
3801 hrtimer_cancel(&hs.timer);
3802 mode = HRTIMER_MODE_ABS;
3803 } while (hs.task && !signal_pending(current));
3805 __set_current_state(TASK_RUNNING);
3806 destroy_hrtimer_on_stack(&hs.timer);
3810 static bool blk_mq_poll_hybrid(struct request_queue *q,
3811 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3815 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3818 if (!blk_qc_t_is_internal(cookie))
3819 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3821 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3823 * With scheduling, if the request has completed, we'll
3824 * get a NULL return here, as we clear the sched tag when
3825 * that happens. The request still remains valid, like always,
3826 * so we should be safe with just the NULL check.
3832 return blk_mq_poll_hybrid_sleep(q, rq);
3836 * blk_poll - poll for IO completions
3838 * @cookie: cookie passed back at IO submission time
3839 * @spin: whether to spin for completions
3842 * Poll for completions on the passed in queue. Returns number of
3843 * completed entries found. If @spin is true, then blk_poll will continue
3844 * looping until at least one completion is found, unless the task is
3845 * otherwise marked running (or we need to reschedule).
3847 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3849 struct blk_mq_hw_ctx *hctx;
3852 if (!blk_qc_t_valid(cookie) ||
3853 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3857 blk_flush_plug_list(current->plug, false);
3859 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3862 * If we sleep, have the caller restart the poll loop to reset
3863 * the state. Like for the other success return cases, the
3864 * caller is responsible for checking if the IO completed. If
3865 * the IO isn't complete, we'll get called again and will go
3866 * straight to the busy poll loop.
3868 if (blk_mq_poll_hybrid(q, hctx, cookie))
3871 hctx->poll_considered++;
3873 state = current->state;
3877 hctx->poll_invoked++;
3879 ret = q->mq_ops->poll(hctx);
3881 hctx->poll_success++;
3882 __set_current_state(TASK_RUNNING);
3886 if (signal_pending_state(state, current))
3887 __set_current_state(TASK_RUNNING);
3889 if (current->state == TASK_RUNNING)
3891 if (ret < 0 || !spin)
3894 } while (!need_resched());
3896 __set_current_state(TASK_RUNNING);
3899 EXPORT_SYMBOL_GPL(blk_poll);
3901 unsigned int blk_mq_rq_cpu(struct request *rq)
3903 return rq->mq_ctx->cpu;
3905 EXPORT_SYMBOL(blk_mq_rq_cpu);
3907 static int __init blk_mq_init(void)
3911 for_each_possible_cpu(i)
3912 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3913 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
3915 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
3916 "block/softirq:dead", NULL,
3917 blk_softirq_cpu_dead);
3918 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3919 blk_mq_hctx_notify_dead);
3920 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
3921 blk_mq_hctx_notify_online,
3922 blk_mq_hctx_notify_offline);
3925 subsys_initcall(blk_mq_init);