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 block_device *part;
99 unsigned int inflight[2];
102 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
103 struct request *rq, void *priv,
106 struct mq_inflight *mi = priv;
108 if ((!mi->part->bd_partno || rq->part == mi->part) &&
109 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
110 mi->inflight[rq_data_dir(rq)]++;
115 unsigned int blk_mq_in_flight(struct request_queue *q,
116 struct block_device *part)
118 struct mq_inflight mi = { .part = part };
120 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
122 return mi.inflight[0] + mi.inflight[1];
125 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
126 unsigned int inflight[2])
128 struct mq_inflight mi = { .part = part };
130 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
131 inflight[0] = mi.inflight[0];
132 inflight[1] = mi.inflight[1];
135 void blk_freeze_queue_start(struct request_queue *q)
137 mutex_lock(&q->mq_freeze_lock);
138 if (++q->mq_freeze_depth == 1) {
139 percpu_ref_kill(&q->q_usage_counter);
140 mutex_unlock(&q->mq_freeze_lock);
142 blk_mq_run_hw_queues(q, false);
144 mutex_unlock(&q->mq_freeze_lock);
147 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
149 void blk_mq_freeze_queue_wait(struct request_queue *q)
151 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
153 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
155 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
156 unsigned long timeout)
158 return wait_event_timeout(q->mq_freeze_wq,
159 percpu_ref_is_zero(&q->q_usage_counter),
162 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
165 * Guarantee no request is in use, so we can change any data structure of
166 * the queue afterward.
168 void blk_freeze_queue(struct request_queue *q)
171 * In the !blk_mq case we are only calling this to kill the
172 * q_usage_counter, otherwise this increases the freeze depth
173 * and waits for it to return to zero. For this reason there is
174 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
175 * exported to drivers as the only user for unfreeze is blk_mq.
177 blk_freeze_queue_start(q);
178 blk_mq_freeze_queue_wait(q);
181 void blk_mq_freeze_queue(struct request_queue *q)
184 * ...just an alias to keep freeze and unfreeze actions balanced
185 * in the blk_mq_* namespace
189 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
191 void blk_mq_unfreeze_queue(struct request_queue *q)
193 mutex_lock(&q->mq_freeze_lock);
194 q->mq_freeze_depth--;
195 WARN_ON_ONCE(q->mq_freeze_depth < 0);
196 if (!q->mq_freeze_depth) {
197 percpu_ref_resurrect(&q->q_usage_counter);
198 wake_up_all(&q->mq_freeze_wq);
200 mutex_unlock(&q->mq_freeze_lock);
202 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
205 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
206 * mpt3sas driver such that this function can be removed.
208 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
210 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
212 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
215 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
218 * Note: this function does not prevent that the struct request end_io()
219 * callback function is invoked. Once this function is returned, we make
220 * sure no dispatch can happen until the queue is unquiesced via
221 * blk_mq_unquiesce_queue().
223 void blk_mq_quiesce_queue(struct request_queue *q)
225 struct blk_mq_hw_ctx *hctx;
229 blk_mq_quiesce_queue_nowait(q);
231 queue_for_each_hw_ctx(q, hctx, i) {
232 if (hctx->flags & BLK_MQ_F_BLOCKING)
233 synchronize_srcu(hctx->srcu);
240 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
243 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
246 * This function recovers queue into the state before quiescing
247 * which is done by blk_mq_quiesce_queue.
249 void blk_mq_unquiesce_queue(struct request_queue *q)
251 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
253 /* dispatch requests which are inserted during quiescing */
254 blk_mq_run_hw_queues(q, true);
256 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
258 void blk_mq_wake_waiters(struct request_queue *q)
260 struct blk_mq_hw_ctx *hctx;
263 queue_for_each_hw_ctx(q, hctx, i)
264 if (blk_mq_hw_queue_mapped(hctx))
265 blk_mq_tag_wakeup_all(hctx->tags, true);
269 * Only need start/end time stamping if we have iostat or
270 * blk stats enabled, or using an IO scheduler.
272 static inline bool blk_mq_need_time_stamp(struct request *rq)
274 return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
277 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
278 unsigned int tag, u64 alloc_time_ns)
280 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
281 struct request *rq = tags->static_rqs[tag];
283 if (data->q->elevator) {
284 rq->tag = BLK_MQ_NO_TAG;
285 rq->internal_tag = tag;
288 rq->internal_tag = BLK_MQ_NO_TAG;
291 /* csd/requeue_work/fifo_time is initialized before use */
293 rq->mq_ctx = data->ctx;
294 rq->mq_hctx = data->hctx;
296 rq->cmd_flags = data->cmd_flags;
297 if (data->flags & BLK_MQ_REQ_PM)
298 rq->rq_flags |= RQF_PM;
299 if (blk_queue_io_stat(data->q))
300 rq->rq_flags |= RQF_IO_STAT;
301 INIT_LIST_HEAD(&rq->queuelist);
302 INIT_HLIST_NODE(&rq->hash);
303 RB_CLEAR_NODE(&rq->rb_node);
306 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
307 rq->alloc_time_ns = alloc_time_ns;
309 if (blk_mq_need_time_stamp(rq))
310 rq->start_time_ns = ktime_get_ns();
312 rq->start_time_ns = 0;
313 rq->io_start_time_ns = 0;
314 rq->stats_sectors = 0;
315 rq->nr_phys_segments = 0;
316 #if defined(CONFIG_BLK_DEV_INTEGRITY)
317 rq->nr_integrity_segments = 0;
319 blk_crypto_rq_set_defaults(rq);
320 /* tag was already set */
321 WRITE_ONCE(rq->deadline, 0);
326 rq->end_io_data = NULL;
328 data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++;
329 refcount_set(&rq->ref, 1);
331 if (!op_is_flush(data->cmd_flags)) {
332 struct elevator_queue *e = data->q->elevator;
335 if (e && e->type->ops.prepare_request) {
336 if (e->type->icq_cache)
337 blk_mq_sched_assign_ioc(rq);
339 e->type->ops.prepare_request(rq);
340 rq->rq_flags |= RQF_ELVPRIV;
344 data->hctx->queued++;
348 static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data)
350 struct request_queue *q = data->q;
351 struct elevator_queue *e = q->elevator;
352 u64 alloc_time_ns = 0;
355 /* alloc_time includes depth and tag waits */
356 if (blk_queue_rq_alloc_time(q))
357 alloc_time_ns = ktime_get_ns();
359 if (data->cmd_flags & REQ_NOWAIT)
360 data->flags |= BLK_MQ_REQ_NOWAIT;
364 * Flush requests are special and go directly to the
365 * dispatch list. Don't include reserved tags in the
366 * limiting, as it isn't useful.
368 if (!op_is_flush(data->cmd_flags) &&
369 e->type->ops.limit_depth &&
370 !(data->flags & BLK_MQ_REQ_RESERVED))
371 e->type->ops.limit_depth(data->cmd_flags, data);
375 data->ctx = blk_mq_get_ctx(q);
376 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
378 blk_mq_tag_busy(data->hctx);
381 * Waiting allocations only fail because of an inactive hctx. In that
382 * case just retry the hctx assignment and tag allocation as CPU hotplug
383 * should have migrated us to an online CPU by now.
385 tag = blk_mq_get_tag(data);
386 if (tag == BLK_MQ_NO_TAG) {
387 if (data->flags & BLK_MQ_REQ_NOWAIT)
391 * Give up the CPU and sleep for a random short time to ensure
392 * that thread using a realtime scheduling class are migrated
393 * off the CPU, and thus off the hctx that is going away.
398 return blk_mq_rq_ctx_init(data, tag, alloc_time_ns);
401 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
402 blk_mq_req_flags_t flags)
404 struct blk_mq_alloc_data data = {
412 ret = blk_queue_enter(q, flags);
416 rq = __blk_mq_alloc_request(&data);
420 rq->__sector = (sector_t) -1;
421 rq->bio = rq->biotail = NULL;
425 return ERR_PTR(-EWOULDBLOCK);
427 EXPORT_SYMBOL(blk_mq_alloc_request);
429 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
430 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
432 struct blk_mq_alloc_data data = {
437 u64 alloc_time_ns = 0;
442 /* alloc_time includes depth and tag waits */
443 if (blk_queue_rq_alloc_time(q))
444 alloc_time_ns = ktime_get_ns();
447 * If the tag allocator sleeps we could get an allocation for a
448 * different hardware context. No need to complicate the low level
449 * allocator for this for the rare use case of a command tied to
452 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
453 return ERR_PTR(-EINVAL);
455 if (hctx_idx >= q->nr_hw_queues)
456 return ERR_PTR(-EIO);
458 ret = blk_queue_enter(q, flags);
463 * Check if the hardware context is actually mapped to anything.
464 * If not tell the caller that it should skip this queue.
467 data.hctx = q->queue_hw_ctx[hctx_idx];
468 if (!blk_mq_hw_queue_mapped(data.hctx))
470 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
471 data.ctx = __blk_mq_get_ctx(q, cpu);
474 blk_mq_tag_busy(data.hctx);
477 tag = blk_mq_get_tag(&data);
478 if (tag == BLK_MQ_NO_TAG)
480 return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns);
486 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
488 static void __blk_mq_free_request(struct request *rq)
490 struct request_queue *q = rq->q;
491 struct blk_mq_ctx *ctx = rq->mq_ctx;
492 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
493 const int sched_tag = rq->internal_tag;
495 blk_crypto_free_request(rq);
496 blk_pm_mark_last_busy(rq);
498 if (rq->tag != BLK_MQ_NO_TAG)
499 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
500 if (sched_tag != BLK_MQ_NO_TAG)
501 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
502 blk_mq_sched_restart(hctx);
506 void blk_mq_free_request(struct request *rq)
508 struct request_queue *q = rq->q;
509 struct elevator_queue *e = q->elevator;
510 struct blk_mq_ctx *ctx = rq->mq_ctx;
511 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
513 if (rq->rq_flags & RQF_ELVPRIV) {
514 if (e && e->type->ops.finish_request)
515 e->type->ops.finish_request(rq);
517 put_io_context(rq->elv.icq->ioc);
522 ctx->rq_completed[rq_is_sync(rq)]++;
523 if (rq->rq_flags & RQF_MQ_INFLIGHT)
524 __blk_mq_dec_active_requests(hctx);
526 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
527 laptop_io_completion(q->backing_dev_info);
531 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
532 if (refcount_dec_and_test(&rq->ref))
533 __blk_mq_free_request(rq);
535 EXPORT_SYMBOL_GPL(blk_mq_free_request);
537 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
541 if (blk_mq_need_time_stamp(rq))
542 now = ktime_get_ns();
544 if (rq->rq_flags & RQF_STATS) {
545 blk_mq_poll_stats_start(rq->q);
546 blk_stat_add(rq, now);
549 blk_mq_sched_completed_request(rq, now);
551 blk_account_io_done(rq, now);
554 rq_qos_done(rq->q, rq);
555 rq->end_io(rq, error);
557 blk_mq_free_request(rq);
560 EXPORT_SYMBOL(__blk_mq_end_request);
562 void blk_mq_end_request(struct request *rq, blk_status_t error)
564 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
566 __blk_mq_end_request(rq, error);
568 EXPORT_SYMBOL(blk_mq_end_request);
571 * Softirq action handler - move entries to local list and loop over them
572 * while passing them to the queue registered handler.
574 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
576 struct list_head *cpu_list, local_list;
579 cpu_list = this_cpu_ptr(&blk_cpu_done);
580 list_replace_init(cpu_list, &local_list);
583 while (!list_empty(&local_list)) {
586 rq = list_entry(local_list.next, struct request, ipi_list);
587 list_del_init(&rq->ipi_list);
588 rq->q->mq_ops->complete(rq);
592 static void blk_mq_trigger_softirq(struct request *rq)
594 struct list_head *list;
597 local_irq_save(flags);
598 list = this_cpu_ptr(&blk_cpu_done);
599 list_add_tail(&rq->ipi_list, list);
602 * If the list only contains our just added request, signal a raise of
603 * the softirq. If there are already entries there, someone already
604 * raised the irq but it hasn't run yet.
606 if (list->next == &rq->ipi_list)
607 raise_softirq_irqoff(BLOCK_SOFTIRQ);
608 local_irq_restore(flags);
611 static int blk_softirq_cpu_dead(unsigned int cpu)
614 * If a CPU goes away, splice its entries to the current CPU
615 * and trigger a run of the softirq
618 list_splice_init(&per_cpu(blk_cpu_done, cpu),
619 this_cpu_ptr(&blk_cpu_done));
620 raise_softirq_irqoff(BLOCK_SOFTIRQ);
627 static void __blk_mq_complete_request_remote(void *data)
629 struct request *rq = data;
632 * For most of single queue controllers, there is only one irq vector
633 * for handling I/O completion, and the only irq's affinity is set
634 * to all possible CPUs. On most of ARCHs, this affinity means the irq
635 * is handled on one specific CPU.
637 * So complete I/O requests in softirq context in case of single queue
638 * devices to avoid degrading I/O performance due to irqsoff latency.
640 if (rq->q->nr_hw_queues == 1)
641 blk_mq_trigger_softirq(rq);
643 rq->q->mq_ops->complete(rq);
646 static inline bool blk_mq_complete_need_ipi(struct request *rq)
648 int cpu = raw_smp_processor_id();
650 if (!IS_ENABLED(CONFIG_SMP) ||
651 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
654 * With force threaded interrupts enabled, raising softirq from an SMP
655 * function call will always result in waking the ksoftirqd thread.
656 * This is probably worse than completing the request on a different
659 if (force_irqthreads)
662 /* same CPU or cache domain? Complete locally */
663 if (cpu == rq->mq_ctx->cpu ||
664 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
665 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
668 /* don't try to IPI to an offline CPU */
669 return cpu_online(rq->mq_ctx->cpu);
672 bool blk_mq_complete_request_remote(struct request *rq)
674 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
677 * For a polled request, always complete locallly, it's pointless
678 * to redirect the completion.
680 if (rq->cmd_flags & REQ_HIPRI)
683 if (blk_mq_complete_need_ipi(rq)) {
684 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
685 smp_call_function_single_async(rq->mq_ctx->cpu, &rq->csd);
687 if (rq->q->nr_hw_queues > 1)
689 blk_mq_trigger_softirq(rq);
694 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
697 * blk_mq_complete_request - end I/O on a request
698 * @rq: the request being processed
701 * Complete a request by scheduling the ->complete_rq operation.
703 void blk_mq_complete_request(struct request *rq)
705 if (!blk_mq_complete_request_remote(rq))
706 rq->q->mq_ops->complete(rq);
708 EXPORT_SYMBOL(blk_mq_complete_request);
710 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
711 __releases(hctx->srcu)
713 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
716 srcu_read_unlock(hctx->srcu, srcu_idx);
719 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
720 __acquires(hctx->srcu)
722 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
723 /* shut up gcc false positive */
727 *srcu_idx = srcu_read_lock(hctx->srcu);
731 * blk_mq_start_request - Start processing a request
732 * @rq: Pointer to request to be started
734 * Function used by device drivers to notify the block layer that a request
735 * is going to be processed now, so blk layer can do proper initializations
736 * such as starting the timeout timer.
738 void blk_mq_start_request(struct request *rq)
740 struct request_queue *q = rq->q;
742 trace_block_rq_issue(rq);
744 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
745 rq->io_start_time_ns = ktime_get_ns();
746 rq->stats_sectors = blk_rq_sectors(rq);
747 rq->rq_flags |= RQF_STATS;
751 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
754 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
756 #ifdef CONFIG_BLK_DEV_INTEGRITY
757 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
758 q->integrity.profile->prepare_fn(rq);
761 EXPORT_SYMBOL(blk_mq_start_request);
763 static void __blk_mq_requeue_request(struct request *rq)
765 struct request_queue *q = rq->q;
767 blk_mq_put_driver_tag(rq);
769 trace_block_rq_requeue(rq);
770 rq_qos_requeue(q, rq);
772 if (blk_mq_request_started(rq)) {
773 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
774 rq->rq_flags &= ~RQF_TIMED_OUT;
778 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
780 __blk_mq_requeue_request(rq);
782 /* this request will be re-inserted to io scheduler queue */
783 blk_mq_sched_requeue_request(rq);
785 BUG_ON(!list_empty(&rq->queuelist));
786 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
788 EXPORT_SYMBOL(blk_mq_requeue_request);
790 static void blk_mq_requeue_work(struct work_struct *work)
792 struct request_queue *q =
793 container_of(work, struct request_queue, requeue_work.work);
795 struct request *rq, *next;
797 spin_lock_irq(&q->requeue_lock);
798 list_splice_init(&q->requeue_list, &rq_list);
799 spin_unlock_irq(&q->requeue_lock);
801 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
802 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
805 rq->rq_flags &= ~RQF_SOFTBARRIER;
806 list_del_init(&rq->queuelist);
808 * If RQF_DONTPREP, rq has contained some driver specific
809 * data, so insert it to hctx dispatch list to avoid any
812 if (rq->rq_flags & RQF_DONTPREP)
813 blk_mq_request_bypass_insert(rq, false, false);
815 blk_mq_sched_insert_request(rq, true, false, false);
818 while (!list_empty(&rq_list)) {
819 rq = list_entry(rq_list.next, struct request, queuelist);
820 list_del_init(&rq->queuelist);
821 blk_mq_sched_insert_request(rq, false, false, false);
824 blk_mq_run_hw_queues(q, false);
827 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
828 bool kick_requeue_list)
830 struct request_queue *q = rq->q;
834 * We abuse this flag that is otherwise used by the I/O scheduler to
835 * request head insertion from the workqueue.
837 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
839 spin_lock_irqsave(&q->requeue_lock, flags);
841 rq->rq_flags |= RQF_SOFTBARRIER;
842 list_add(&rq->queuelist, &q->requeue_list);
844 list_add_tail(&rq->queuelist, &q->requeue_list);
846 spin_unlock_irqrestore(&q->requeue_lock, flags);
848 if (kick_requeue_list)
849 blk_mq_kick_requeue_list(q);
852 void blk_mq_kick_requeue_list(struct request_queue *q)
854 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
856 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
858 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
861 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
862 msecs_to_jiffies(msecs));
864 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
866 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
868 if (tag < tags->nr_tags) {
869 prefetch(tags->rqs[tag]);
870 return tags->rqs[tag];
875 EXPORT_SYMBOL(blk_mq_tag_to_rq);
877 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
878 void *priv, bool reserved)
881 * If we find a request that isn't idle and the queue matches,
882 * we know the queue is busy. Return false to stop the iteration.
884 if (blk_mq_request_started(rq) && rq->q == hctx->queue) {
894 bool blk_mq_queue_inflight(struct request_queue *q)
898 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
901 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
903 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
905 req->rq_flags |= RQF_TIMED_OUT;
906 if (req->q->mq_ops->timeout) {
907 enum blk_eh_timer_return ret;
909 ret = req->q->mq_ops->timeout(req, reserved);
910 if (ret == BLK_EH_DONE)
912 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
918 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
920 unsigned long deadline;
922 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
924 if (rq->rq_flags & RQF_TIMED_OUT)
927 deadline = READ_ONCE(rq->deadline);
928 if (time_after_eq(jiffies, deadline))
933 else if (time_after(*next, deadline))
938 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
939 struct request *rq, void *priv, bool reserved)
941 unsigned long *next = priv;
944 * Just do a quick check if it is expired before locking the request in
945 * so we're not unnecessarilly synchronizing across CPUs.
947 if (!blk_mq_req_expired(rq, next))
951 * We have reason to believe the request may be expired. Take a
952 * reference on the request to lock this request lifetime into its
953 * currently allocated context to prevent it from being reallocated in
954 * the event the completion by-passes this timeout handler.
956 * If the reference was already released, then the driver beat the
957 * timeout handler to posting a natural completion.
959 if (!refcount_inc_not_zero(&rq->ref))
963 * The request is now locked and cannot be reallocated underneath the
964 * timeout handler's processing. Re-verify this exact request is truly
965 * expired; if it is not expired, then the request was completed and
966 * reallocated as a new request.
968 if (blk_mq_req_expired(rq, next))
969 blk_mq_rq_timed_out(rq, reserved);
971 if (is_flush_rq(rq, hctx))
973 else if (refcount_dec_and_test(&rq->ref))
974 __blk_mq_free_request(rq);
979 static void blk_mq_timeout_work(struct work_struct *work)
981 struct request_queue *q =
982 container_of(work, struct request_queue, timeout_work);
983 unsigned long next = 0;
984 struct blk_mq_hw_ctx *hctx;
987 /* A deadlock might occur if a request is stuck requiring a
988 * timeout at the same time a queue freeze is waiting
989 * completion, since the timeout code would not be able to
990 * acquire the queue reference here.
992 * That's why we don't use blk_queue_enter here; instead, we use
993 * percpu_ref_tryget directly, because we need to be able to
994 * obtain a reference even in the short window between the queue
995 * starting to freeze, by dropping the first reference in
996 * blk_freeze_queue_start, and the moment the last request is
997 * consumed, marked by the instant q_usage_counter reaches
1000 if (!percpu_ref_tryget(&q->q_usage_counter))
1003 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
1006 mod_timer(&q->timeout, next);
1009 * Request timeouts are handled as a forward rolling timer. If
1010 * we end up here it means that no requests are pending and
1011 * also that no request has been pending for a while. Mark
1012 * each hctx as idle.
1014 queue_for_each_hw_ctx(q, hctx, i) {
1015 /* the hctx may be unmapped, so check it here */
1016 if (blk_mq_hw_queue_mapped(hctx))
1017 blk_mq_tag_idle(hctx);
1023 struct flush_busy_ctx_data {
1024 struct blk_mq_hw_ctx *hctx;
1025 struct list_head *list;
1028 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1030 struct flush_busy_ctx_data *flush_data = data;
1031 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1032 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1033 enum hctx_type type = hctx->type;
1035 spin_lock(&ctx->lock);
1036 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1037 sbitmap_clear_bit(sb, bitnr);
1038 spin_unlock(&ctx->lock);
1043 * Process software queues that have been marked busy, splicing them
1044 * to the for-dispatch
1046 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1048 struct flush_busy_ctx_data data = {
1053 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1055 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1057 struct dispatch_rq_data {
1058 struct blk_mq_hw_ctx *hctx;
1062 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1065 struct dispatch_rq_data *dispatch_data = data;
1066 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1067 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1068 enum hctx_type type = hctx->type;
1070 spin_lock(&ctx->lock);
1071 if (!list_empty(&ctx->rq_lists[type])) {
1072 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1073 list_del_init(&dispatch_data->rq->queuelist);
1074 if (list_empty(&ctx->rq_lists[type]))
1075 sbitmap_clear_bit(sb, bitnr);
1077 spin_unlock(&ctx->lock);
1079 return !dispatch_data->rq;
1082 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1083 struct blk_mq_ctx *start)
1085 unsigned off = start ? start->index_hw[hctx->type] : 0;
1086 struct dispatch_rq_data data = {
1091 __sbitmap_for_each_set(&hctx->ctx_map, off,
1092 dispatch_rq_from_ctx, &data);
1097 static inline unsigned int queued_to_index(unsigned int queued)
1102 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1105 static bool __blk_mq_get_driver_tag(struct request *rq)
1107 struct sbitmap_queue *bt = rq->mq_hctx->tags->bitmap_tags;
1108 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1111 blk_mq_tag_busy(rq->mq_hctx);
1113 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1114 bt = rq->mq_hctx->tags->breserved_tags;
1117 if (!hctx_may_queue(rq->mq_hctx, bt))
1121 tag = __sbitmap_queue_get(bt);
1122 if (tag == BLK_MQ_NO_TAG)
1125 rq->tag = tag + tag_offset;
1129 static bool blk_mq_get_driver_tag(struct request *rq)
1131 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1133 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq))
1136 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1137 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1138 rq->rq_flags |= RQF_MQ_INFLIGHT;
1139 __blk_mq_inc_active_requests(hctx);
1141 hctx->tags->rqs[rq->tag] = rq;
1145 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1146 int flags, void *key)
1148 struct blk_mq_hw_ctx *hctx;
1150 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1152 spin_lock(&hctx->dispatch_wait_lock);
1153 if (!list_empty(&wait->entry)) {
1154 struct sbitmap_queue *sbq;
1156 list_del_init(&wait->entry);
1157 sbq = hctx->tags->bitmap_tags;
1158 atomic_dec(&sbq->ws_active);
1160 spin_unlock(&hctx->dispatch_wait_lock);
1162 blk_mq_run_hw_queue(hctx, true);
1167 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1168 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1169 * restart. For both cases, take care to check the condition again after
1170 * marking us as waiting.
1172 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1175 struct sbitmap_queue *sbq = hctx->tags->bitmap_tags;
1176 struct wait_queue_head *wq;
1177 wait_queue_entry_t *wait;
1180 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1181 blk_mq_sched_mark_restart_hctx(hctx);
1184 * It's possible that a tag was freed in the window between the
1185 * allocation failure and adding the hardware queue to the wait
1188 * Don't clear RESTART here, someone else could have set it.
1189 * At most this will cost an extra queue run.
1191 return blk_mq_get_driver_tag(rq);
1194 wait = &hctx->dispatch_wait;
1195 if (!list_empty_careful(&wait->entry))
1198 wq = &bt_wait_ptr(sbq, hctx)->wait;
1200 spin_lock_irq(&wq->lock);
1201 spin_lock(&hctx->dispatch_wait_lock);
1202 if (!list_empty(&wait->entry)) {
1203 spin_unlock(&hctx->dispatch_wait_lock);
1204 spin_unlock_irq(&wq->lock);
1208 atomic_inc(&sbq->ws_active);
1209 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1210 __add_wait_queue(wq, wait);
1213 * It's possible that a tag was freed in the window between the
1214 * allocation failure and adding the hardware queue to the wait
1217 ret = blk_mq_get_driver_tag(rq);
1219 spin_unlock(&hctx->dispatch_wait_lock);
1220 spin_unlock_irq(&wq->lock);
1225 * We got a tag, remove ourselves from the wait queue to ensure
1226 * someone else gets the wakeup.
1228 list_del_init(&wait->entry);
1229 atomic_dec(&sbq->ws_active);
1230 spin_unlock(&hctx->dispatch_wait_lock);
1231 spin_unlock_irq(&wq->lock);
1236 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1237 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1239 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1240 * - EWMA is one simple way to compute running average value
1241 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1242 * - take 4 as factor for avoiding to get too small(0) result, and this
1243 * factor doesn't matter because EWMA decreases exponentially
1245 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1249 if (hctx->queue->elevator)
1252 ewma = hctx->dispatch_busy;
1257 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1259 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1260 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1262 hctx->dispatch_busy = ewma;
1265 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1267 static void blk_mq_handle_dev_resource(struct request *rq,
1268 struct list_head *list)
1270 struct request *next =
1271 list_first_entry_or_null(list, struct request, queuelist);
1274 * If an I/O scheduler has been configured and we got a driver tag for
1275 * the next request already, free it.
1278 blk_mq_put_driver_tag(next);
1280 list_add(&rq->queuelist, list);
1281 __blk_mq_requeue_request(rq);
1284 static void blk_mq_handle_zone_resource(struct request *rq,
1285 struct list_head *zone_list)
1288 * If we end up here it is because we cannot dispatch a request to a
1289 * specific zone due to LLD level zone-write locking or other zone
1290 * related resource not being available. In this case, set the request
1291 * aside in zone_list for retrying it later.
1293 list_add(&rq->queuelist, zone_list);
1294 __blk_mq_requeue_request(rq);
1297 enum prep_dispatch {
1299 PREP_DISPATCH_NO_TAG,
1300 PREP_DISPATCH_NO_BUDGET,
1303 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1306 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1308 if (need_budget && !blk_mq_get_dispatch_budget(rq->q)) {
1309 blk_mq_put_driver_tag(rq);
1310 return PREP_DISPATCH_NO_BUDGET;
1313 if (!blk_mq_get_driver_tag(rq)) {
1315 * The initial allocation attempt failed, so we need to
1316 * rerun the hardware queue when a tag is freed. The
1317 * waitqueue takes care of that. If the queue is run
1318 * before we add this entry back on the dispatch list,
1319 * we'll re-run it below.
1321 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1323 * All budgets not got from this function will be put
1324 * together during handling partial dispatch
1327 blk_mq_put_dispatch_budget(rq->q);
1328 return PREP_DISPATCH_NO_TAG;
1332 return PREP_DISPATCH_OK;
1335 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1336 static void blk_mq_release_budgets(struct request_queue *q,
1337 unsigned int nr_budgets)
1341 for (i = 0; i < nr_budgets; i++)
1342 blk_mq_put_dispatch_budget(q);
1346 * Returns true if we did some work AND can potentially do more.
1348 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1349 unsigned int nr_budgets)
1351 enum prep_dispatch prep;
1352 struct request_queue *q = hctx->queue;
1353 struct request *rq, *nxt;
1355 blk_status_t ret = BLK_STS_OK;
1356 LIST_HEAD(zone_list);
1358 if (list_empty(list))
1362 * Now process all the entries, sending them to the driver.
1364 errors = queued = 0;
1366 struct blk_mq_queue_data bd;
1368 rq = list_first_entry(list, struct request, queuelist);
1370 WARN_ON_ONCE(hctx != rq->mq_hctx);
1371 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1372 if (prep != PREP_DISPATCH_OK)
1375 list_del_init(&rq->queuelist);
1380 * Flag last if we have no more requests, or if we have more
1381 * but can't assign a driver tag to it.
1383 if (list_empty(list))
1386 nxt = list_first_entry(list, struct request, queuelist);
1387 bd.last = !blk_mq_get_driver_tag(nxt);
1391 * once the request is queued to lld, no need to cover the
1396 ret = q->mq_ops->queue_rq(hctx, &bd);
1401 case BLK_STS_RESOURCE:
1402 case BLK_STS_DEV_RESOURCE:
1403 blk_mq_handle_dev_resource(rq, list);
1405 case BLK_STS_ZONE_RESOURCE:
1407 * Move the request to zone_list and keep going through
1408 * the dispatch list to find more requests the drive can
1411 blk_mq_handle_zone_resource(rq, &zone_list);
1415 blk_mq_end_request(rq, ret);
1417 } while (!list_empty(list));
1419 if (!list_empty(&zone_list))
1420 list_splice_tail_init(&zone_list, list);
1422 hctx->dispatched[queued_to_index(queued)]++;
1424 /* If we didn't flush the entire list, we could have told the driver
1425 * there was more coming, but that turned out to be a lie.
1427 if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued)
1428 q->mq_ops->commit_rqs(hctx);
1430 * Any items that need requeuing? Stuff them into hctx->dispatch,
1431 * that is where we will continue on next queue run.
1433 if (!list_empty(list)) {
1435 /* For non-shared tags, the RESTART check will suffice */
1436 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1437 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1438 bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET;
1440 blk_mq_release_budgets(q, nr_budgets);
1442 spin_lock(&hctx->lock);
1443 list_splice_tail_init(list, &hctx->dispatch);
1444 spin_unlock(&hctx->lock);
1447 * Order adding requests to hctx->dispatch and checking
1448 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1449 * in blk_mq_sched_restart(). Avoid restart code path to
1450 * miss the new added requests to hctx->dispatch, meantime
1451 * SCHED_RESTART is observed here.
1456 * If SCHED_RESTART was set by the caller of this function and
1457 * it is no longer set that means that it was cleared by another
1458 * thread and hence that a queue rerun is needed.
1460 * If 'no_tag' is set, that means that we failed getting
1461 * a driver tag with an I/O scheduler attached. If our dispatch
1462 * waitqueue is no longer active, ensure that we run the queue
1463 * AFTER adding our entries back to the list.
1465 * If no I/O scheduler has been configured it is possible that
1466 * the hardware queue got stopped and restarted before requests
1467 * were pushed back onto the dispatch list. Rerun the queue to
1468 * avoid starvation. Notes:
1469 * - blk_mq_run_hw_queue() checks whether or not a queue has
1470 * been stopped before rerunning a queue.
1471 * - Some but not all block drivers stop a queue before
1472 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1475 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1476 * bit is set, run queue after a delay to avoid IO stalls
1477 * that could otherwise occur if the queue is idle. We'll do
1478 * similar if we couldn't get budget and SCHED_RESTART is set.
1480 needs_restart = blk_mq_sched_needs_restart(hctx);
1481 if (!needs_restart ||
1482 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1483 blk_mq_run_hw_queue(hctx, true);
1484 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1486 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1488 blk_mq_update_dispatch_busy(hctx, true);
1491 blk_mq_update_dispatch_busy(hctx, false);
1493 return (queued + errors) != 0;
1497 * __blk_mq_run_hw_queue - Run a hardware queue.
1498 * @hctx: Pointer to the hardware queue to run.
1500 * Send pending requests to the hardware.
1502 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1507 * We can't run the queue inline with ints disabled. Ensure that
1508 * we catch bad users of this early.
1510 WARN_ON_ONCE(in_interrupt());
1512 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1514 hctx_lock(hctx, &srcu_idx);
1515 blk_mq_sched_dispatch_requests(hctx);
1516 hctx_unlock(hctx, srcu_idx);
1519 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1521 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1523 if (cpu >= nr_cpu_ids)
1524 cpu = cpumask_first(hctx->cpumask);
1529 * It'd be great if the workqueue API had a way to pass
1530 * in a mask and had some smarts for more clever placement.
1531 * For now we just round-robin here, switching for every
1532 * BLK_MQ_CPU_WORK_BATCH queued items.
1534 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1537 int next_cpu = hctx->next_cpu;
1539 if (hctx->queue->nr_hw_queues == 1)
1540 return WORK_CPU_UNBOUND;
1542 if (--hctx->next_cpu_batch <= 0) {
1544 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1546 if (next_cpu >= nr_cpu_ids)
1547 next_cpu = blk_mq_first_mapped_cpu(hctx);
1548 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1552 * Do unbound schedule if we can't find a online CPU for this hctx,
1553 * and it should only happen in the path of handling CPU DEAD.
1555 if (!cpu_online(next_cpu)) {
1562 * Make sure to re-select CPU next time once after CPUs
1563 * in hctx->cpumask become online again.
1565 hctx->next_cpu = next_cpu;
1566 hctx->next_cpu_batch = 1;
1567 return WORK_CPU_UNBOUND;
1570 hctx->next_cpu = next_cpu;
1575 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1576 * @hctx: Pointer to the hardware queue to run.
1577 * @async: If we want to run the queue asynchronously.
1578 * @msecs: Milliseconds of delay to wait before running the queue.
1580 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1581 * with a delay of @msecs.
1583 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1584 unsigned long msecs)
1586 if (unlikely(blk_mq_hctx_stopped(hctx)))
1589 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1590 int cpu = get_cpu();
1591 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1592 __blk_mq_run_hw_queue(hctx);
1600 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1601 msecs_to_jiffies(msecs));
1605 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1606 * @hctx: Pointer to the hardware queue to run.
1607 * @msecs: Milliseconds of delay to wait before running the queue.
1609 * Run a hardware queue asynchronously with a delay of @msecs.
1611 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1613 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1615 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1618 * blk_mq_run_hw_queue - Start to run a hardware queue.
1619 * @hctx: Pointer to the hardware queue to run.
1620 * @async: If we want to run the queue asynchronously.
1622 * Check if the request queue is not in a quiesced state and if there are
1623 * pending requests to be sent. If this is true, run the queue to send requests
1626 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1632 * When queue is quiesced, we may be switching io scheduler, or
1633 * updating nr_hw_queues, or other things, and we can't run queue
1634 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1636 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1639 hctx_lock(hctx, &srcu_idx);
1640 need_run = !blk_queue_quiesced(hctx->queue) &&
1641 blk_mq_hctx_has_pending(hctx);
1642 hctx_unlock(hctx, srcu_idx);
1645 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1647 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1650 * Is the request queue handled by an IO scheduler that does not respect
1651 * hardware queues when dispatching?
1653 static bool blk_mq_has_sqsched(struct request_queue *q)
1655 struct elevator_queue *e = q->elevator;
1657 if (e && e->type->ops.dispatch_request &&
1658 !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE))
1664 * Return prefered queue to dispatch from (if any) for non-mq aware IO
1667 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
1669 struct blk_mq_hw_ctx *hctx;
1672 * If the IO scheduler does not respect hardware queues when
1673 * dispatching, we just don't bother with multiple HW queues and
1674 * dispatch from hctx for the current CPU since running multiple queues
1675 * just causes lock contention inside the scheduler and pointless cache
1678 hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT,
1679 raw_smp_processor_id());
1680 if (!blk_mq_hctx_stopped(hctx))
1686 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
1687 * @q: Pointer to the request queue to run.
1688 * @async: If we want to run the queue asynchronously.
1690 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1692 struct blk_mq_hw_ctx *hctx, *sq_hctx;
1696 if (blk_mq_has_sqsched(q))
1697 sq_hctx = blk_mq_get_sq_hctx(q);
1698 queue_for_each_hw_ctx(q, hctx, i) {
1699 if (blk_mq_hctx_stopped(hctx))
1702 * Dispatch from this hctx either if there's no hctx preferred
1703 * by IO scheduler or if it has requests that bypass the
1706 if (!sq_hctx || sq_hctx == hctx ||
1707 !list_empty_careful(&hctx->dispatch))
1708 blk_mq_run_hw_queue(hctx, async);
1711 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1714 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1715 * @q: Pointer to the request queue to run.
1716 * @msecs: Milliseconds of delay to wait before running the queues.
1718 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1720 struct blk_mq_hw_ctx *hctx, *sq_hctx;
1724 if (blk_mq_has_sqsched(q))
1725 sq_hctx = blk_mq_get_sq_hctx(q);
1726 queue_for_each_hw_ctx(q, hctx, i) {
1727 if (blk_mq_hctx_stopped(hctx))
1730 * Dispatch from this hctx either if there's no hctx preferred
1731 * by IO scheduler or if it has requests that bypass the
1734 if (!sq_hctx || sq_hctx == hctx ||
1735 !list_empty_careful(&hctx->dispatch))
1736 blk_mq_delay_run_hw_queue(hctx, msecs);
1739 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1742 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1743 * @q: request queue.
1745 * The caller is responsible for serializing this function against
1746 * blk_mq_{start,stop}_hw_queue().
1748 bool blk_mq_queue_stopped(struct request_queue *q)
1750 struct blk_mq_hw_ctx *hctx;
1753 queue_for_each_hw_ctx(q, hctx, i)
1754 if (blk_mq_hctx_stopped(hctx))
1759 EXPORT_SYMBOL(blk_mq_queue_stopped);
1762 * This function is often used for pausing .queue_rq() by driver when
1763 * there isn't enough resource or some conditions aren't satisfied, and
1764 * BLK_STS_RESOURCE is usually returned.
1766 * We do not guarantee that dispatch can be drained or blocked
1767 * after blk_mq_stop_hw_queue() returns. Please use
1768 * blk_mq_quiesce_queue() for that requirement.
1770 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1772 cancel_delayed_work(&hctx->run_work);
1774 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1776 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1779 * This function is often used for pausing .queue_rq() by driver when
1780 * there isn't enough resource or some conditions aren't satisfied, and
1781 * BLK_STS_RESOURCE is usually returned.
1783 * We do not guarantee that dispatch can be drained or blocked
1784 * after blk_mq_stop_hw_queues() returns. Please use
1785 * blk_mq_quiesce_queue() for that requirement.
1787 void blk_mq_stop_hw_queues(struct request_queue *q)
1789 struct blk_mq_hw_ctx *hctx;
1792 queue_for_each_hw_ctx(q, hctx, i)
1793 blk_mq_stop_hw_queue(hctx);
1795 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1797 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1799 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1801 blk_mq_run_hw_queue(hctx, false);
1803 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1805 void blk_mq_start_hw_queues(struct request_queue *q)
1807 struct blk_mq_hw_ctx *hctx;
1810 queue_for_each_hw_ctx(q, hctx, i)
1811 blk_mq_start_hw_queue(hctx);
1813 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1815 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1817 if (!blk_mq_hctx_stopped(hctx))
1820 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1821 blk_mq_run_hw_queue(hctx, async);
1823 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1825 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1827 struct blk_mq_hw_ctx *hctx;
1830 queue_for_each_hw_ctx(q, hctx, i)
1831 blk_mq_start_stopped_hw_queue(hctx, async);
1833 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1835 static void blk_mq_run_work_fn(struct work_struct *work)
1837 struct blk_mq_hw_ctx *hctx;
1839 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1842 * If we are stopped, don't run the queue.
1844 if (blk_mq_hctx_stopped(hctx))
1847 __blk_mq_run_hw_queue(hctx);
1850 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1854 struct blk_mq_ctx *ctx = rq->mq_ctx;
1855 enum hctx_type type = hctx->type;
1857 lockdep_assert_held(&ctx->lock);
1859 trace_block_rq_insert(rq);
1862 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1864 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1867 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1870 struct blk_mq_ctx *ctx = rq->mq_ctx;
1872 lockdep_assert_held(&ctx->lock);
1874 __blk_mq_insert_req_list(hctx, rq, at_head);
1875 blk_mq_hctx_mark_pending(hctx, ctx);
1879 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1880 * @rq: Pointer to request to be inserted.
1881 * @at_head: true if the request should be inserted at the head of the list.
1882 * @run_queue: If we should run the hardware queue after inserting the request.
1884 * Should only be used carefully, when the caller knows we want to
1885 * bypass a potential IO scheduler on the target device.
1887 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1890 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1892 spin_lock(&hctx->lock);
1894 list_add(&rq->queuelist, &hctx->dispatch);
1896 list_add_tail(&rq->queuelist, &hctx->dispatch);
1897 spin_unlock(&hctx->lock);
1900 blk_mq_run_hw_queue(hctx, false);
1903 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1904 struct list_head *list)
1908 enum hctx_type type = hctx->type;
1911 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1914 list_for_each_entry(rq, list, queuelist) {
1915 BUG_ON(rq->mq_ctx != ctx);
1916 trace_block_rq_insert(rq);
1919 spin_lock(&ctx->lock);
1920 list_splice_tail_init(list, &ctx->rq_lists[type]);
1921 blk_mq_hctx_mark_pending(hctx, ctx);
1922 spin_unlock(&ctx->lock);
1925 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1927 struct request *rqa = container_of(a, struct request, queuelist);
1928 struct request *rqb = container_of(b, struct request, queuelist);
1930 if (rqa->mq_ctx != rqb->mq_ctx)
1931 return rqa->mq_ctx > rqb->mq_ctx;
1932 if (rqa->mq_hctx != rqb->mq_hctx)
1933 return rqa->mq_hctx > rqb->mq_hctx;
1935 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1938 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1942 if (list_empty(&plug->mq_list))
1944 list_splice_init(&plug->mq_list, &list);
1946 if (plug->rq_count > 2 && plug->multiple_queues)
1947 list_sort(NULL, &list, plug_rq_cmp);
1952 struct list_head rq_list;
1953 struct request *rq, *head_rq = list_entry_rq(list.next);
1954 struct list_head *pos = &head_rq->queuelist; /* skip first */
1955 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1956 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1957 unsigned int depth = 1;
1959 list_for_each_continue(pos, &list) {
1960 rq = list_entry_rq(pos);
1962 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1967 list_cut_before(&rq_list, &list, pos);
1968 trace_block_unplug(head_rq->q, depth, !from_schedule);
1969 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1971 } while(!list_empty(&list));
1974 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1975 unsigned int nr_segs)
1979 if (bio->bi_opf & REQ_RAHEAD)
1980 rq->cmd_flags |= REQ_FAILFAST_MASK;
1982 rq->__sector = bio->bi_iter.bi_sector;
1983 rq->write_hint = bio->bi_write_hint;
1984 blk_rq_bio_prep(rq, bio, nr_segs);
1986 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
1987 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1990 blk_account_io_start(rq);
1993 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1995 blk_qc_t *cookie, bool last)
1997 struct request_queue *q = rq->q;
1998 struct blk_mq_queue_data bd = {
2002 blk_qc_t new_cookie;
2005 new_cookie = request_to_qc_t(hctx, rq);
2008 * For OK queue, we are done. For error, caller may kill it.
2009 * Any other error (busy), just add it to our list as we
2010 * previously would have done.
2012 ret = q->mq_ops->queue_rq(hctx, &bd);
2015 blk_mq_update_dispatch_busy(hctx, false);
2016 *cookie = new_cookie;
2018 case BLK_STS_RESOURCE:
2019 case BLK_STS_DEV_RESOURCE:
2020 blk_mq_update_dispatch_busy(hctx, true);
2021 __blk_mq_requeue_request(rq);
2024 blk_mq_update_dispatch_busy(hctx, false);
2025 *cookie = BLK_QC_T_NONE;
2032 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2035 bool bypass_insert, bool last)
2037 struct request_queue *q = rq->q;
2038 bool run_queue = true;
2041 * RCU or SRCU read lock is needed before checking quiesced flag.
2043 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2044 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2045 * and avoid driver to try to dispatch again.
2047 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2049 bypass_insert = false;
2053 if (q->elevator && !bypass_insert)
2056 if (!blk_mq_get_dispatch_budget(q))
2059 if (!blk_mq_get_driver_tag(rq)) {
2060 blk_mq_put_dispatch_budget(q);
2064 return __blk_mq_issue_directly(hctx, rq, cookie, last);
2067 return BLK_STS_RESOURCE;
2069 blk_mq_sched_insert_request(rq, false, run_queue, false);
2075 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2076 * @hctx: Pointer of the associated hardware queue.
2077 * @rq: Pointer to request to be sent.
2078 * @cookie: Request queue cookie.
2080 * If the device has enough resources to accept a new request now, send the
2081 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2082 * we can try send it another time in the future. Requests inserted at this
2083 * queue have higher priority.
2085 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2086 struct request *rq, blk_qc_t *cookie)
2091 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2093 hctx_lock(hctx, &srcu_idx);
2095 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
2096 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2097 blk_mq_request_bypass_insert(rq, false, true);
2098 else if (ret != BLK_STS_OK)
2099 blk_mq_end_request(rq, ret);
2101 hctx_unlock(hctx, srcu_idx);
2104 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2108 blk_qc_t unused_cookie;
2109 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2111 hctx_lock(hctx, &srcu_idx);
2112 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
2113 hctx_unlock(hctx, srcu_idx);
2118 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2119 struct list_head *list)
2124 while (!list_empty(list)) {
2126 struct request *rq = list_first_entry(list, struct request,
2129 list_del_init(&rq->queuelist);
2130 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2131 if (ret != BLK_STS_OK) {
2132 if (ret == BLK_STS_RESOURCE ||
2133 ret == BLK_STS_DEV_RESOURCE) {
2134 blk_mq_request_bypass_insert(rq, false,
2138 blk_mq_end_request(rq, ret);
2145 * If we didn't flush the entire list, we could have told
2146 * the driver there was more coming, but that turned out to
2149 if ((!list_empty(list) || errors) &&
2150 hctx->queue->mq_ops->commit_rqs && queued)
2151 hctx->queue->mq_ops->commit_rqs(hctx);
2154 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2156 list_add_tail(&rq->queuelist, &plug->mq_list);
2158 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
2159 struct request *tmp;
2161 tmp = list_first_entry(&plug->mq_list, struct request,
2163 if (tmp->q != rq->q)
2164 plug->multiple_queues = true;
2169 * blk_mq_submit_bio - Create and send a request to block device.
2170 * @bio: Bio pointer.
2172 * Builds up a request structure from @q and @bio and send to the device. The
2173 * request may not be queued directly to hardware if:
2174 * * This request can be merged with another one
2175 * * We want to place request at plug queue for possible future merging
2176 * * There is an IO scheduler active at this queue
2178 * It will not queue the request if there is an error with the bio, or at the
2181 * Returns: Request queue cookie.
2183 blk_qc_t blk_mq_submit_bio(struct bio *bio)
2185 struct request_queue *q = bio->bi_bdev->bd_disk->queue;
2186 const int is_sync = op_is_sync(bio->bi_opf);
2187 const int is_flush_fua = op_is_flush(bio->bi_opf);
2188 struct blk_mq_alloc_data data = {
2192 struct blk_plug *plug;
2193 struct request *same_queue_rq = NULL;
2194 unsigned int nr_segs;
2199 blk_queue_bounce(q, &bio);
2200 __blk_queue_split(&bio, &nr_segs);
2202 if (!bio_integrity_prep(bio))
2205 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2206 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2209 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2212 rq_qos_throttle(q, bio);
2214 hipri = bio->bi_opf & REQ_HIPRI;
2216 data.cmd_flags = bio->bi_opf;
2217 rq = __blk_mq_alloc_request(&data);
2218 if (unlikely(!rq)) {
2219 rq_qos_cleanup(q, bio);
2220 if (bio->bi_opf & REQ_NOWAIT)
2221 bio_wouldblock_error(bio);
2225 trace_block_getrq(bio);
2227 rq_qos_track(q, rq, bio);
2229 cookie = request_to_qc_t(data.hctx, rq);
2231 blk_mq_bio_to_request(rq, bio, nr_segs);
2233 ret = blk_crypto_init_request(rq);
2234 if (ret != BLK_STS_OK) {
2235 bio->bi_status = ret;
2237 blk_mq_free_request(rq);
2238 return BLK_QC_T_NONE;
2241 plug = blk_mq_plug(q, bio);
2242 if (unlikely(is_flush_fua)) {
2243 /* Bypass scheduler for flush requests */
2244 blk_insert_flush(rq);
2245 blk_mq_run_hw_queue(data.hctx, true);
2246 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2247 !blk_queue_nonrot(q))) {
2249 * Use plugging if we have a ->commit_rqs() hook as well, as
2250 * we know the driver uses bd->last in a smart fashion.
2252 * Use normal plugging if this disk is slow HDD, as sequential
2253 * IO may benefit a lot from plug merging.
2255 unsigned int request_count = plug->rq_count;
2256 struct request *last = NULL;
2259 trace_block_plug(q);
2261 last = list_entry_rq(plug->mq_list.prev);
2263 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2264 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2265 blk_flush_plug_list(plug, false);
2266 trace_block_plug(q);
2269 blk_add_rq_to_plug(plug, rq);
2270 } else if (q->elevator) {
2271 /* Insert the request at the IO scheduler queue */
2272 blk_mq_sched_insert_request(rq, false, true, true);
2273 } else if (plug && !blk_queue_nomerges(q)) {
2275 * We do limited plugging. If the bio can be merged, do that.
2276 * Otherwise the existing request in the plug list will be
2277 * issued. So the plug list will have one request at most
2278 * The plug list might get flushed before this. If that happens,
2279 * the plug list is empty, and same_queue_rq is invalid.
2281 if (list_empty(&plug->mq_list))
2282 same_queue_rq = NULL;
2283 if (same_queue_rq) {
2284 list_del_init(&same_queue_rq->queuelist);
2287 blk_add_rq_to_plug(plug, rq);
2288 trace_block_plug(q);
2290 if (same_queue_rq) {
2291 data.hctx = same_queue_rq->mq_hctx;
2292 trace_block_unplug(q, 1, true);
2293 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2296 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2297 !data.hctx->dispatch_busy) {
2299 * There is no scheduler and we can try to send directly
2302 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2305 blk_mq_sched_insert_request(rq, false, true, true);
2309 return BLK_QC_T_NONE;
2313 return BLK_QC_T_NONE;
2316 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2317 unsigned int hctx_idx)
2321 if (tags->rqs && set->ops->exit_request) {
2324 for (i = 0; i < tags->nr_tags; i++) {
2325 struct request *rq = tags->static_rqs[i];
2329 set->ops->exit_request(set, rq, hctx_idx);
2330 tags->static_rqs[i] = NULL;
2334 while (!list_empty(&tags->page_list)) {
2335 page = list_first_entry(&tags->page_list, struct page, lru);
2336 list_del_init(&page->lru);
2338 * Remove kmemleak object previously allocated in
2339 * blk_mq_alloc_rqs().
2341 kmemleak_free(page_address(page));
2342 __free_pages(page, page->private);
2346 void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags)
2350 kfree(tags->static_rqs);
2351 tags->static_rqs = NULL;
2353 blk_mq_free_tags(tags, flags);
2356 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2357 unsigned int hctx_idx,
2358 unsigned int nr_tags,
2359 unsigned int reserved_tags,
2362 struct blk_mq_tags *tags;
2365 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2366 if (node == NUMA_NO_NODE)
2367 node = set->numa_node;
2369 tags = blk_mq_init_tags(nr_tags, reserved_tags, node, flags);
2373 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2374 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2377 blk_mq_free_tags(tags, flags);
2381 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2382 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2384 if (!tags->static_rqs) {
2386 blk_mq_free_tags(tags, flags);
2393 static size_t order_to_size(unsigned int order)
2395 return (size_t)PAGE_SIZE << order;
2398 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2399 unsigned int hctx_idx, int node)
2403 if (set->ops->init_request) {
2404 ret = set->ops->init_request(set, rq, hctx_idx, node);
2409 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2413 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2414 unsigned int hctx_idx, unsigned int depth)
2416 unsigned int i, j, entries_per_page, max_order = 4;
2417 size_t rq_size, left;
2420 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2421 if (node == NUMA_NO_NODE)
2422 node = set->numa_node;
2424 INIT_LIST_HEAD(&tags->page_list);
2427 * rq_size is the size of the request plus driver payload, rounded
2428 * to the cacheline size
2430 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2432 left = rq_size * depth;
2434 for (i = 0; i < depth; ) {
2435 int this_order = max_order;
2440 while (this_order && left < order_to_size(this_order - 1))
2444 page = alloc_pages_node(node,
2445 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2451 if (order_to_size(this_order) < rq_size)
2458 page->private = this_order;
2459 list_add_tail(&page->lru, &tags->page_list);
2461 p = page_address(page);
2463 * Allow kmemleak to scan these pages as they contain pointers
2464 * to additional allocations like via ops->init_request().
2466 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2467 entries_per_page = order_to_size(this_order) / rq_size;
2468 to_do = min(entries_per_page, depth - i);
2469 left -= to_do * rq_size;
2470 for (j = 0; j < to_do; j++) {
2471 struct request *rq = p;
2473 tags->static_rqs[i] = rq;
2474 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2475 tags->static_rqs[i] = NULL;
2486 blk_mq_free_rqs(set, tags, hctx_idx);
2490 struct rq_iter_data {
2491 struct blk_mq_hw_ctx *hctx;
2495 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2497 struct rq_iter_data *iter_data = data;
2499 if (rq->mq_hctx != iter_data->hctx)
2501 iter_data->has_rq = true;
2505 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2507 struct blk_mq_tags *tags = hctx->sched_tags ?
2508 hctx->sched_tags : hctx->tags;
2509 struct rq_iter_data data = {
2513 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2517 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2518 struct blk_mq_hw_ctx *hctx)
2520 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2522 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2527 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2529 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2530 struct blk_mq_hw_ctx, cpuhp_online);
2532 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2533 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2537 * Prevent new request from being allocated on the current hctx.
2539 * The smp_mb__after_atomic() Pairs with the implied barrier in
2540 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2541 * seen once we return from the tag allocator.
2543 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2544 smp_mb__after_atomic();
2547 * Try to grab a reference to the queue and wait for any outstanding
2548 * requests. If we could not grab a reference the queue has been
2549 * frozen and there are no requests.
2551 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2552 while (blk_mq_hctx_has_requests(hctx))
2554 percpu_ref_put(&hctx->queue->q_usage_counter);
2560 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2562 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2563 struct blk_mq_hw_ctx, cpuhp_online);
2565 if (cpumask_test_cpu(cpu, hctx->cpumask))
2566 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2571 * 'cpu' is going away. splice any existing rq_list entries from this
2572 * software queue to the hw queue dispatch list, and ensure that it
2575 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2577 struct blk_mq_hw_ctx *hctx;
2578 struct blk_mq_ctx *ctx;
2580 enum hctx_type type;
2582 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2583 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2586 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2589 spin_lock(&ctx->lock);
2590 if (!list_empty(&ctx->rq_lists[type])) {
2591 list_splice_init(&ctx->rq_lists[type], &tmp);
2592 blk_mq_hctx_clear_pending(hctx, ctx);
2594 spin_unlock(&ctx->lock);
2596 if (list_empty(&tmp))
2599 spin_lock(&hctx->lock);
2600 list_splice_tail_init(&tmp, &hctx->dispatch);
2601 spin_unlock(&hctx->lock);
2603 blk_mq_run_hw_queue(hctx, true);
2607 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2609 if (!(hctx->flags & BLK_MQ_F_STACKING))
2610 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2611 &hctx->cpuhp_online);
2612 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2616 /* hctx->ctxs will be freed in queue's release handler */
2617 static void blk_mq_exit_hctx(struct request_queue *q,
2618 struct blk_mq_tag_set *set,
2619 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2621 if (blk_mq_hw_queue_mapped(hctx))
2622 blk_mq_tag_idle(hctx);
2624 if (set->ops->exit_request)
2625 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2627 if (set->ops->exit_hctx)
2628 set->ops->exit_hctx(hctx, hctx_idx);
2630 blk_mq_remove_cpuhp(hctx);
2632 spin_lock(&q->unused_hctx_lock);
2633 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2634 spin_unlock(&q->unused_hctx_lock);
2637 static void blk_mq_exit_hw_queues(struct request_queue *q,
2638 struct blk_mq_tag_set *set, int nr_queue)
2640 struct blk_mq_hw_ctx *hctx;
2643 queue_for_each_hw_ctx(q, hctx, i) {
2646 blk_mq_debugfs_unregister_hctx(hctx);
2647 blk_mq_exit_hctx(q, set, hctx, i);
2651 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2653 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2655 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2656 __alignof__(struct blk_mq_hw_ctx)) !=
2657 sizeof(struct blk_mq_hw_ctx));
2659 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2660 hw_ctx_size += sizeof(struct srcu_struct);
2665 static int blk_mq_init_hctx(struct request_queue *q,
2666 struct blk_mq_tag_set *set,
2667 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2669 hctx->queue_num = hctx_idx;
2671 if (!(hctx->flags & BLK_MQ_F_STACKING))
2672 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2673 &hctx->cpuhp_online);
2674 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2676 hctx->tags = set->tags[hctx_idx];
2678 if (set->ops->init_hctx &&
2679 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2680 goto unregister_cpu_notifier;
2682 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2688 if (set->ops->exit_hctx)
2689 set->ops->exit_hctx(hctx, hctx_idx);
2690 unregister_cpu_notifier:
2691 blk_mq_remove_cpuhp(hctx);
2695 static struct blk_mq_hw_ctx *
2696 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2699 struct blk_mq_hw_ctx *hctx;
2700 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2702 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2704 goto fail_alloc_hctx;
2706 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2709 atomic_set(&hctx->nr_active, 0);
2710 if (node == NUMA_NO_NODE)
2711 node = set->numa_node;
2712 hctx->numa_node = node;
2714 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2715 spin_lock_init(&hctx->lock);
2716 INIT_LIST_HEAD(&hctx->dispatch);
2718 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
2720 INIT_LIST_HEAD(&hctx->hctx_list);
2723 * Allocate space for all possible cpus to avoid allocation at
2726 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2731 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2736 spin_lock_init(&hctx->dispatch_wait_lock);
2737 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2738 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2740 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2744 if (hctx->flags & BLK_MQ_F_BLOCKING)
2745 init_srcu_struct(hctx->srcu);
2746 blk_mq_hctx_kobj_init(hctx);
2751 sbitmap_free(&hctx->ctx_map);
2755 free_cpumask_var(hctx->cpumask);
2762 static void blk_mq_init_cpu_queues(struct request_queue *q,
2763 unsigned int nr_hw_queues)
2765 struct blk_mq_tag_set *set = q->tag_set;
2768 for_each_possible_cpu(i) {
2769 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2770 struct blk_mq_hw_ctx *hctx;
2774 spin_lock_init(&__ctx->lock);
2775 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2776 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2781 * Set local node, IFF we have more than one hw queue. If
2782 * not, we remain on the home node of the device
2784 for (j = 0; j < set->nr_maps; j++) {
2785 hctx = blk_mq_map_queue_type(q, j, i);
2786 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2787 hctx->numa_node = cpu_to_node(i);
2792 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2795 unsigned int flags = set->flags;
2798 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2799 set->queue_depth, set->reserved_tags, flags);
2800 if (!set->tags[hctx_idx])
2803 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2808 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2809 set->tags[hctx_idx] = NULL;
2813 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2814 unsigned int hctx_idx)
2816 unsigned int flags = set->flags;
2818 if (set->tags && set->tags[hctx_idx]) {
2819 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2820 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2821 set->tags[hctx_idx] = NULL;
2825 static void blk_mq_map_swqueue(struct request_queue *q)
2827 unsigned int i, j, hctx_idx;
2828 struct blk_mq_hw_ctx *hctx;
2829 struct blk_mq_ctx *ctx;
2830 struct blk_mq_tag_set *set = q->tag_set;
2832 queue_for_each_hw_ctx(q, hctx, i) {
2833 cpumask_clear(hctx->cpumask);
2835 hctx->dispatch_from = NULL;
2839 * Map software to hardware queues.
2841 * If the cpu isn't present, the cpu is mapped to first hctx.
2843 for_each_possible_cpu(i) {
2845 ctx = per_cpu_ptr(q->queue_ctx, i);
2846 for (j = 0; j < set->nr_maps; j++) {
2847 if (!set->map[j].nr_queues) {
2848 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2849 HCTX_TYPE_DEFAULT, i);
2852 hctx_idx = set->map[j].mq_map[i];
2853 /* unmapped hw queue can be remapped after CPU topo changed */
2854 if (!set->tags[hctx_idx] &&
2855 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2857 * If tags initialization fail for some hctx,
2858 * that hctx won't be brought online. In this
2859 * case, remap the current ctx to hctx[0] which
2860 * is guaranteed to always have tags allocated
2862 set->map[j].mq_map[i] = 0;
2865 hctx = blk_mq_map_queue_type(q, j, i);
2866 ctx->hctxs[j] = hctx;
2868 * If the CPU is already set in the mask, then we've
2869 * mapped this one already. This can happen if
2870 * devices share queues across queue maps.
2872 if (cpumask_test_cpu(i, hctx->cpumask))
2875 cpumask_set_cpu(i, hctx->cpumask);
2877 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2878 hctx->ctxs[hctx->nr_ctx++] = ctx;
2881 * If the nr_ctx type overflows, we have exceeded the
2882 * amount of sw queues we can support.
2884 BUG_ON(!hctx->nr_ctx);
2887 for (; j < HCTX_MAX_TYPES; j++)
2888 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2889 HCTX_TYPE_DEFAULT, i);
2892 queue_for_each_hw_ctx(q, hctx, i) {
2894 * If no software queues are mapped to this hardware queue,
2895 * disable it and free the request entries.
2897 if (!hctx->nr_ctx) {
2898 /* Never unmap queue 0. We need it as a
2899 * fallback in case of a new remap fails
2902 if (i && set->tags[i])
2903 blk_mq_free_map_and_requests(set, i);
2909 hctx->tags = set->tags[i];
2910 WARN_ON(!hctx->tags);
2913 * Set the map size to the number of mapped software queues.
2914 * This is more accurate and more efficient than looping
2915 * over all possibly mapped software queues.
2917 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2920 * Initialize batch roundrobin counts
2922 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2923 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2928 * Caller needs to ensure that we're either frozen/quiesced, or that
2929 * the queue isn't live yet.
2931 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2933 struct blk_mq_hw_ctx *hctx;
2936 queue_for_each_hw_ctx(q, hctx, i) {
2938 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
2940 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
2944 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
2947 struct request_queue *q;
2949 lockdep_assert_held(&set->tag_list_lock);
2951 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2952 blk_mq_freeze_queue(q);
2953 queue_set_hctx_shared(q, shared);
2954 blk_mq_unfreeze_queue(q);
2958 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2960 struct blk_mq_tag_set *set = q->tag_set;
2962 mutex_lock(&set->tag_list_lock);
2963 list_del(&q->tag_set_list);
2964 if (list_is_singular(&set->tag_list)) {
2965 /* just transitioned to unshared */
2966 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
2967 /* update existing queue */
2968 blk_mq_update_tag_set_shared(set, false);
2970 mutex_unlock(&set->tag_list_lock);
2971 INIT_LIST_HEAD(&q->tag_set_list);
2974 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2975 struct request_queue *q)
2977 mutex_lock(&set->tag_list_lock);
2980 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2982 if (!list_empty(&set->tag_list) &&
2983 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2984 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
2985 /* update existing queue */
2986 blk_mq_update_tag_set_shared(set, true);
2988 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
2989 queue_set_hctx_shared(q, true);
2990 list_add_tail(&q->tag_set_list, &set->tag_list);
2992 mutex_unlock(&set->tag_list_lock);
2995 /* All allocations will be freed in release handler of q->mq_kobj */
2996 static int blk_mq_alloc_ctxs(struct request_queue *q)
2998 struct blk_mq_ctxs *ctxs;
3001 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3005 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3006 if (!ctxs->queue_ctx)
3009 for_each_possible_cpu(cpu) {
3010 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3014 q->mq_kobj = &ctxs->kobj;
3015 q->queue_ctx = ctxs->queue_ctx;
3024 * It is the actual release handler for mq, but we do it from
3025 * request queue's release handler for avoiding use-after-free
3026 * and headache because q->mq_kobj shouldn't have been introduced,
3027 * but we can't group ctx/kctx kobj without it.
3029 void blk_mq_release(struct request_queue *q)
3031 struct blk_mq_hw_ctx *hctx, *next;
3034 queue_for_each_hw_ctx(q, hctx, i)
3035 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3037 /* all hctx are in .unused_hctx_list now */
3038 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3039 list_del_init(&hctx->hctx_list);
3040 kobject_put(&hctx->kobj);
3043 kfree(q->queue_hw_ctx);
3046 * release .mq_kobj and sw queue's kobject now because
3047 * both share lifetime with request queue.
3049 blk_mq_sysfs_deinit(q);
3052 struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3055 struct request_queue *uninit_q, *q;
3057 uninit_q = blk_alloc_queue(set->numa_node);
3059 return ERR_PTR(-ENOMEM);
3060 uninit_q->queuedata = queuedata;
3063 * Initialize the queue without an elevator. device_add_disk() will do
3064 * the initialization.
3066 q = blk_mq_init_allocated_queue(set, uninit_q, false);
3068 blk_cleanup_queue(uninit_q);
3072 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data);
3074 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3076 return blk_mq_init_queue_data(set, NULL);
3078 EXPORT_SYMBOL(blk_mq_init_queue);
3081 * Helper for setting up a queue with mq ops, given queue depth, and
3082 * the passed in mq ops flags.
3084 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
3085 const struct blk_mq_ops *ops,
3086 unsigned int queue_depth,
3087 unsigned int set_flags)
3089 struct request_queue *q;
3092 memset(set, 0, sizeof(*set));
3094 set->nr_hw_queues = 1;
3096 set->queue_depth = queue_depth;
3097 set->numa_node = NUMA_NO_NODE;
3098 set->flags = set_flags;
3100 ret = blk_mq_alloc_tag_set(set);
3102 return ERR_PTR(ret);
3104 q = blk_mq_init_queue(set);
3106 blk_mq_free_tag_set(set);
3112 EXPORT_SYMBOL(blk_mq_init_sq_queue);
3114 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3115 struct blk_mq_tag_set *set, struct request_queue *q,
3116 int hctx_idx, int node)
3118 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3120 /* reuse dead hctx first */
3121 spin_lock(&q->unused_hctx_lock);
3122 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3123 if (tmp->numa_node == node) {
3129 list_del_init(&hctx->hctx_list);
3130 spin_unlock(&q->unused_hctx_lock);
3133 hctx = blk_mq_alloc_hctx(q, set, node);
3137 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3143 kobject_put(&hctx->kobj);
3148 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3149 struct request_queue *q)
3152 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3154 if (q->nr_hw_queues < set->nr_hw_queues) {
3155 struct blk_mq_hw_ctx **new_hctxs;
3157 new_hctxs = kcalloc_node(set->nr_hw_queues,
3158 sizeof(*new_hctxs), GFP_KERNEL,
3163 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3165 q->queue_hw_ctx = new_hctxs;
3170 /* protect against switching io scheduler */
3171 mutex_lock(&q->sysfs_lock);
3172 for (i = 0; i < set->nr_hw_queues; i++) {
3174 struct blk_mq_hw_ctx *hctx;
3176 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3178 * If the hw queue has been mapped to another numa node,
3179 * we need to realloc the hctx. If allocation fails, fallback
3180 * to use the previous one.
3182 if (hctxs[i] && (hctxs[i]->numa_node == node))
3185 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3188 blk_mq_exit_hctx(q, set, hctxs[i], i);
3192 pr_warn("Allocate new hctx on node %d fails,\
3193 fallback to previous one on node %d\n",
3194 node, hctxs[i]->numa_node);
3200 * Increasing nr_hw_queues fails. Free the newly allocated
3201 * hctxs and keep the previous q->nr_hw_queues.
3203 if (i != set->nr_hw_queues) {
3204 j = q->nr_hw_queues;
3208 end = q->nr_hw_queues;
3209 q->nr_hw_queues = set->nr_hw_queues;
3212 for (; j < end; j++) {
3213 struct blk_mq_hw_ctx *hctx = hctxs[j];
3217 blk_mq_free_map_and_requests(set, j);
3218 blk_mq_exit_hctx(q, set, hctx, j);
3222 mutex_unlock(&q->sysfs_lock);
3225 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3226 struct request_queue *q,
3229 /* mark the queue as mq asap */
3230 q->mq_ops = set->ops;
3232 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3233 blk_mq_poll_stats_bkt,
3234 BLK_MQ_POLL_STATS_BKTS, q);
3238 if (blk_mq_alloc_ctxs(q))
3241 /* init q->mq_kobj and sw queues' kobjects */
3242 blk_mq_sysfs_init(q);
3244 INIT_LIST_HEAD(&q->unused_hctx_list);
3245 spin_lock_init(&q->unused_hctx_lock);
3247 blk_mq_realloc_hw_ctxs(set, q);
3248 if (!q->nr_hw_queues)
3251 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3252 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3256 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3257 if (set->nr_maps > HCTX_TYPE_POLL &&
3258 set->map[HCTX_TYPE_POLL].nr_queues)
3259 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3261 q->sg_reserved_size = INT_MAX;
3263 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3264 INIT_LIST_HEAD(&q->requeue_list);
3265 spin_lock_init(&q->requeue_lock);
3267 q->nr_requests = set->queue_depth;
3270 * Default to classic polling
3272 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3274 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3275 blk_mq_add_queue_tag_set(set, q);
3276 blk_mq_map_swqueue(q);
3279 elevator_init_mq(q);
3284 kfree(q->queue_hw_ctx);
3285 q->nr_hw_queues = 0;
3286 blk_mq_sysfs_deinit(q);
3288 blk_stat_free_callback(q->poll_cb);
3292 return ERR_PTR(-ENOMEM);
3294 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3296 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3297 void blk_mq_exit_queue(struct request_queue *q)
3299 struct blk_mq_tag_set *set = q->tag_set;
3301 blk_mq_del_queue_tag_set(q);
3302 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3305 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3309 for (i = 0; i < set->nr_hw_queues; i++) {
3310 if (!__blk_mq_alloc_map_and_request(set, i))
3319 blk_mq_free_map_and_requests(set, i);
3325 * Allocate the request maps associated with this tag_set. Note that this
3326 * may reduce the depth asked for, if memory is tight. set->queue_depth
3327 * will be updated to reflect the allocated depth.
3329 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3334 depth = set->queue_depth;
3336 err = __blk_mq_alloc_rq_maps(set);
3340 set->queue_depth >>= 1;
3341 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3345 } while (set->queue_depth);
3347 if (!set->queue_depth || err) {
3348 pr_err("blk-mq: failed to allocate request map\n");
3352 if (depth != set->queue_depth)
3353 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3354 depth, set->queue_depth);
3359 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3362 * blk_mq_map_queues() and multiple .map_queues() implementations
3363 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3364 * number of hardware queues.
3366 if (set->nr_maps == 1)
3367 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3369 if (set->ops->map_queues && !is_kdump_kernel()) {
3373 * transport .map_queues is usually done in the following
3376 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3377 * mask = get_cpu_mask(queue)
3378 * for_each_cpu(cpu, mask)
3379 * set->map[x].mq_map[cpu] = queue;
3382 * When we need to remap, the table has to be cleared for
3383 * killing stale mapping since one CPU may not be mapped
3386 for (i = 0; i < set->nr_maps; i++)
3387 blk_mq_clear_mq_map(&set->map[i]);
3389 return set->ops->map_queues(set);
3391 BUG_ON(set->nr_maps > 1);
3392 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3396 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3397 int cur_nr_hw_queues, int new_nr_hw_queues)
3399 struct blk_mq_tags **new_tags;
3401 if (cur_nr_hw_queues >= new_nr_hw_queues)
3404 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3405 GFP_KERNEL, set->numa_node);
3410 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3411 sizeof(*set->tags));
3413 set->tags = new_tags;
3414 set->nr_hw_queues = new_nr_hw_queues;
3419 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
3420 int new_nr_hw_queues)
3422 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
3426 * Alloc a tag set to be associated with one or more request queues.
3427 * May fail with EINVAL for various error conditions. May adjust the
3428 * requested depth down, if it's too large. In that case, the set
3429 * value will be stored in set->queue_depth.
3431 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3435 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3437 if (!set->nr_hw_queues)
3439 if (!set->queue_depth)
3441 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3444 if (!set->ops->queue_rq)
3447 if (!set->ops->get_budget ^ !set->ops->put_budget)
3450 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3451 pr_info("blk-mq: reduced tag depth to %u\n",
3453 set->queue_depth = BLK_MQ_MAX_DEPTH;
3458 else if (set->nr_maps > HCTX_MAX_TYPES)
3462 * If a crashdump is active, then we are potentially in a very
3463 * memory constrained environment. Limit us to 1 queue and
3464 * 64 tags to prevent using too much memory.
3466 if (is_kdump_kernel()) {
3467 set->nr_hw_queues = 1;
3469 set->queue_depth = min(64U, set->queue_depth);
3472 * There is no use for more h/w queues than cpus if we just have
3475 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3476 set->nr_hw_queues = nr_cpu_ids;
3478 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
3482 for (i = 0; i < set->nr_maps; i++) {
3483 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3484 sizeof(set->map[i].mq_map[0]),
3485 GFP_KERNEL, set->numa_node);
3486 if (!set->map[i].mq_map)
3487 goto out_free_mq_map;
3488 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3491 ret = blk_mq_update_queue_map(set);
3493 goto out_free_mq_map;
3495 ret = blk_mq_alloc_map_and_requests(set);
3497 goto out_free_mq_map;
3499 if (blk_mq_is_sbitmap_shared(set->flags)) {
3500 atomic_set(&set->active_queues_shared_sbitmap, 0);
3502 if (blk_mq_init_shared_sbitmap(set, set->flags)) {
3504 goto out_free_mq_rq_maps;
3508 mutex_init(&set->tag_list_lock);
3509 INIT_LIST_HEAD(&set->tag_list);
3513 out_free_mq_rq_maps:
3514 for (i = 0; i < set->nr_hw_queues; i++)
3515 blk_mq_free_map_and_requests(set, i);
3517 for (i = 0; i < set->nr_maps; i++) {
3518 kfree(set->map[i].mq_map);
3519 set->map[i].mq_map = NULL;
3525 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3527 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3531 for (i = 0; i < set->nr_hw_queues; i++)
3532 blk_mq_free_map_and_requests(set, i);
3534 if (blk_mq_is_sbitmap_shared(set->flags))
3535 blk_mq_exit_shared_sbitmap(set);
3537 for (j = 0; j < set->nr_maps; j++) {
3538 kfree(set->map[j].mq_map);
3539 set->map[j].mq_map = NULL;
3545 EXPORT_SYMBOL(blk_mq_free_tag_set);
3547 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3549 struct blk_mq_tag_set *set = q->tag_set;
3550 struct blk_mq_hw_ctx *hctx;
3556 if (q->nr_requests == nr)
3559 blk_mq_freeze_queue(q);
3560 blk_mq_quiesce_queue(q);
3563 queue_for_each_hw_ctx(q, hctx, i) {
3567 * If we're using an MQ scheduler, just update the scheduler
3568 * queue depth. This is similar to what the old code would do.
3570 if (!hctx->sched_tags) {
3571 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3573 if (!ret && blk_mq_is_sbitmap_shared(set->flags))
3574 blk_mq_tag_resize_shared_sbitmap(set, nr);
3576 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3581 if (q->elevator && q->elevator->type->ops.depth_updated)
3582 q->elevator->type->ops.depth_updated(hctx);
3586 q->nr_requests = nr;
3588 blk_mq_unquiesce_queue(q);
3589 blk_mq_unfreeze_queue(q);
3595 * request_queue and elevator_type pair.
3596 * It is just used by __blk_mq_update_nr_hw_queues to cache
3597 * the elevator_type associated with a request_queue.
3599 struct blk_mq_qe_pair {
3600 struct list_head node;
3601 struct request_queue *q;
3602 struct elevator_type *type;
3606 * Cache the elevator_type in qe pair list and switch the
3607 * io scheduler to 'none'
3609 static bool blk_mq_elv_switch_none(struct list_head *head,
3610 struct request_queue *q)
3612 struct blk_mq_qe_pair *qe;
3617 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3621 INIT_LIST_HEAD(&qe->node);
3623 qe->type = q->elevator->type;
3624 list_add(&qe->node, head);
3626 mutex_lock(&q->sysfs_lock);
3628 * After elevator_switch_mq, the previous elevator_queue will be
3629 * released by elevator_release. The reference of the io scheduler
3630 * module get by elevator_get will also be put. So we need to get
3631 * a reference of the io scheduler module here to prevent it to be
3634 __module_get(qe->type->elevator_owner);
3635 elevator_switch_mq(q, NULL);
3636 mutex_unlock(&q->sysfs_lock);
3641 static void blk_mq_elv_switch_back(struct list_head *head,
3642 struct request_queue *q)
3644 struct blk_mq_qe_pair *qe;
3645 struct elevator_type *t = NULL;
3647 list_for_each_entry(qe, head, node)
3656 list_del(&qe->node);
3659 mutex_lock(&q->sysfs_lock);
3660 elevator_switch_mq(q, t);
3661 mutex_unlock(&q->sysfs_lock);
3664 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3667 struct request_queue *q;
3669 int prev_nr_hw_queues;
3671 lockdep_assert_held(&set->tag_list_lock);
3673 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3674 nr_hw_queues = nr_cpu_ids;
3675 if (nr_hw_queues < 1)
3677 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3680 list_for_each_entry(q, &set->tag_list, tag_set_list)
3681 blk_mq_freeze_queue(q);
3683 * Switch IO scheduler to 'none', cleaning up the data associated
3684 * with the previous scheduler. We will switch back once we are done
3685 * updating the new sw to hw queue mappings.
3687 list_for_each_entry(q, &set->tag_list, tag_set_list)
3688 if (!blk_mq_elv_switch_none(&head, q))
3691 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3692 blk_mq_debugfs_unregister_hctxs(q);
3693 blk_mq_sysfs_unregister(q);
3696 prev_nr_hw_queues = set->nr_hw_queues;
3697 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3701 set->nr_hw_queues = nr_hw_queues;
3703 blk_mq_update_queue_map(set);
3704 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3705 blk_mq_realloc_hw_ctxs(set, q);
3706 if (q->nr_hw_queues != set->nr_hw_queues) {
3707 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3708 nr_hw_queues, prev_nr_hw_queues);
3709 set->nr_hw_queues = prev_nr_hw_queues;
3710 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3713 blk_mq_map_swqueue(q);
3717 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3718 blk_mq_sysfs_register(q);
3719 blk_mq_debugfs_register_hctxs(q);
3723 list_for_each_entry(q, &set->tag_list, tag_set_list)
3724 blk_mq_elv_switch_back(&head, q);
3726 list_for_each_entry(q, &set->tag_list, tag_set_list)
3727 blk_mq_unfreeze_queue(q);
3730 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3732 mutex_lock(&set->tag_list_lock);
3733 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3734 mutex_unlock(&set->tag_list_lock);
3736 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3738 /* Enable polling stats and return whether they were already enabled. */
3739 static bool blk_poll_stats_enable(struct request_queue *q)
3741 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3742 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3744 blk_stat_add_callback(q, q->poll_cb);
3748 static void blk_mq_poll_stats_start(struct request_queue *q)
3751 * We don't arm the callback if polling stats are not enabled or the
3752 * callback is already active.
3754 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3755 blk_stat_is_active(q->poll_cb))
3758 blk_stat_activate_msecs(q->poll_cb, 100);
3761 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3763 struct request_queue *q = cb->data;
3766 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3767 if (cb->stat[bucket].nr_samples)
3768 q->poll_stat[bucket] = cb->stat[bucket];
3772 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3775 unsigned long ret = 0;
3779 * If stats collection isn't on, don't sleep but turn it on for
3782 if (!blk_poll_stats_enable(q))
3786 * As an optimistic guess, use half of the mean service time
3787 * for this type of request. We can (and should) make this smarter.
3788 * For instance, if the completion latencies are tight, we can
3789 * get closer than just half the mean. This is especially
3790 * important on devices where the completion latencies are longer
3791 * than ~10 usec. We do use the stats for the relevant IO size
3792 * if available which does lead to better estimates.
3794 bucket = blk_mq_poll_stats_bkt(rq);
3798 if (q->poll_stat[bucket].nr_samples)
3799 ret = (q->poll_stat[bucket].mean + 1) / 2;
3804 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3807 struct hrtimer_sleeper hs;
3808 enum hrtimer_mode mode;
3812 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3816 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3818 * 0: use half of prev avg
3819 * >0: use this specific value
3821 if (q->poll_nsec > 0)
3822 nsecs = q->poll_nsec;
3824 nsecs = blk_mq_poll_nsecs(q, rq);
3829 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3832 * This will be replaced with the stats tracking code, using
3833 * 'avg_completion_time / 2' as the pre-sleep target.
3837 mode = HRTIMER_MODE_REL;
3838 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3839 hrtimer_set_expires(&hs.timer, kt);
3842 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3844 set_current_state(TASK_UNINTERRUPTIBLE);
3845 hrtimer_sleeper_start_expires(&hs, mode);
3848 hrtimer_cancel(&hs.timer);
3849 mode = HRTIMER_MODE_ABS;
3850 } while (hs.task && !signal_pending(current));
3852 __set_current_state(TASK_RUNNING);
3853 destroy_hrtimer_on_stack(&hs.timer);
3857 static bool blk_mq_poll_hybrid(struct request_queue *q,
3858 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3862 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3865 if (!blk_qc_t_is_internal(cookie))
3866 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3868 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3870 * With scheduling, if the request has completed, we'll
3871 * get a NULL return here, as we clear the sched tag when
3872 * that happens. The request still remains valid, like always,
3873 * so we should be safe with just the NULL check.
3879 return blk_mq_poll_hybrid_sleep(q, rq);
3883 * blk_poll - poll for IO completions
3885 * @cookie: cookie passed back at IO submission time
3886 * @spin: whether to spin for completions
3889 * Poll for completions on the passed in queue. Returns number of
3890 * completed entries found. If @spin is true, then blk_poll will continue
3891 * looping until at least one completion is found, unless the task is
3892 * otherwise marked running (or we need to reschedule).
3894 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3896 struct blk_mq_hw_ctx *hctx;
3899 if (!blk_qc_t_valid(cookie) ||
3900 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3904 blk_flush_plug_list(current->plug, false);
3906 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3909 * If we sleep, have the caller restart the poll loop to reset
3910 * the state. Like for the other success return cases, the
3911 * caller is responsible for checking if the IO completed. If
3912 * the IO isn't complete, we'll get called again and will go
3913 * straight to the busy poll loop. If specified not to spin,
3914 * we also should not sleep.
3916 if (spin && blk_mq_poll_hybrid(q, hctx, cookie))
3919 hctx->poll_considered++;
3921 state = current->state;
3925 hctx->poll_invoked++;
3927 ret = q->mq_ops->poll(hctx);
3929 hctx->poll_success++;
3930 __set_current_state(TASK_RUNNING);
3934 if (signal_pending_state(state, current))
3935 __set_current_state(TASK_RUNNING);
3937 if (current->state == TASK_RUNNING)
3939 if (ret < 0 || !spin)
3942 } while (!need_resched());
3944 __set_current_state(TASK_RUNNING);
3947 EXPORT_SYMBOL_GPL(blk_poll);
3949 unsigned int blk_mq_rq_cpu(struct request *rq)
3951 return rq->mq_ctx->cpu;
3953 EXPORT_SYMBOL(blk_mq_rq_cpu);
3955 static int __init blk_mq_init(void)
3959 for_each_possible_cpu(i)
3960 INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i));
3961 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
3963 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
3964 "block/softirq:dead", NULL,
3965 blk_softirq_cpu_dead);
3966 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3967 blk_mq_hctx_notify_dead);
3968 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
3969 blk_mq_hctx_notify_online,
3970 blk_mq_hctx_notify_offline);
3973 subsys_initcall(blk_mq_init);