1 // SPDX-License-Identifier: GPL-2.0
3 * Block multiqueue core code
5 * Copyright (C) 2013-2014 Jens Axboe
6 * Copyright (C) 2013-2014 Christoph Hellwig
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/backing-dev.h>
11 #include <linux/bio.h>
12 #include <linux/blkdev.h>
13 #include <linux/kmemleak.h>
15 #include <linux/init.h>
16 #include <linux/slab.h>
17 #include <linux/workqueue.h>
18 #include <linux/smp.h>
19 #include <linux/llist.h>
20 #include <linux/list_sort.h>
21 #include <linux/cpu.h>
22 #include <linux/cache.h>
23 #include <linux/sched/sysctl.h>
24 #include <linux/sched/topology.h>
25 #include <linux/sched/signal.h>
26 #include <linux/delay.h>
27 #include <linux/crash_dump.h>
28 #include <linux/prefetch.h>
29 #include <linux/blk-crypto.h>
31 #include <trace/events/block.h>
33 #include <linux/blk-mq.h>
34 #include <linux/t10-pi.h>
37 #include "blk-mq-debugfs.h"
38 #include "blk-mq-tag.h"
41 #include "blk-mq-sched.h"
42 #include "blk-rq-qos.h"
44 static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
46 static void blk_mq_poll_stats_start(struct request_queue *q);
47 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
49 static int blk_mq_poll_stats_bkt(const struct request *rq)
51 int ddir, sectors, bucket;
53 ddir = rq_data_dir(rq);
54 sectors = blk_rq_stats_sectors(rq);
56 bucket = ddir + 2 * ilog2(sectors);
60 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
61 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
67 * Check if any of the ctx, dispatch list or elevator
68 * have pending work in this hardware queue.
70 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
72 return !list_empty_careful(&hctx->dispatch) ||
73 sbitmap_any_bit_set(&hctx->ctx_map) ||
74 blk_mq_sched_has_work(hctx);
78 * Mark this ctx as having pending work in this hardware queue
80 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
81 struct blk_mq_ctx *ctx)
83 const int bit = ctx->index_hw[hctx->type];
85 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
86 sbitmap_set_bit(&hctx->ctx_map, bit);
89 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
90 struct blk_mq_ctx *ctx)
92 const int bit = ctx->index_hw[hctx->type];
94 sbitmap_clear_bit(&hctx->ctx_map, bit);
98 struct block_device *part;
99 unsigned int inflight[2];
102 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
103 struct request *rq, void *priv,
106 struct mq_inflight *mi = priv;
108 if ((!mi->part->bd_partno || rq->part == mi->part) &&
109 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
110 mi->inflight[rq_data_dir(rq)]++;
115 unsigned int blk_mq_in_flight(struct request_queue *q,
116 struct block_device *part)
118 struct mq_inflight mi = { .part = part };
120 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
122 return mi.inflight[0] + mi.inflight[1];
125 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
126 unsigned int inflight[2])
128 struct mq_inflight mi = { .part = part };
130 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
131 inflight[0] = mi.inflight[0];
132 inflight[1] = mi.inflight[1];
135 void blk_freeze_queue_start(struct request_queue *q)
137 mutex_lock(&q->mq_freeze_lock);
138 if (++q->mq_freeze_depth == 1) {
139 percpu_ref_kill(&q->q_usage_counter);
140 mutex_unlock(&q->mq_freeze_lock);
142 blk_mq_run_hw_queues(q, false);
144 mutex_unlock(&q->mq_freeze_lock);
147 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
149 void blk_mq_freeze_queue_wait(struct request_queue *q)
151 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
153 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
155 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
156 unsigned long timeout)
158 return wait_event_timeout(q->mq_freeze_wq,
159 percpu_ref_is_zero(&q->q_usage_counter),
162 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
165 * Guarantee no request is in use, so we can change any data structure of
166 * the queue afterward.
168 void blk_freeze_queue(struct request_queue *q)
171 * In the !blk_mq case we are only calling this to kill the
172 * q_usage_counter, otherwise this increases the freeze depth
173 * and waits for it to return to zero. For this reason there is
174 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
175 * exported to drivers as the only user for unfreeze is blk_mq.
177 blk_freeze_queue_start(q);
178 blk_mq_freeze_queue_wait(q);
181 void blk_mq_freeze_queue(struct request_queue *q)
184 * ...just an alias to keep freeze and unfreeze actions balanced
185 * in the blk_mq_* namespace
189 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
191 void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
193 mutex_lock(&q->mq_freeze_lock);
195 q->q_usage_counter.data->force_atomic = true;
196 q->mq_freeze_depth--;
197 WARN_ON_ONCE(q->mq_freeze_depth < 0);
198 if (!q->mq_freeze_depth) {
199 percpu_ref_resurrect(&q->q_usage_counter);
200 wake_up_all(&q->mq_freeze_wq);
202 mutex_unlock(&q->mq_freeze_lock);
205 void blk_mq_unfreeze_queue(struct request_queue *q)
207 __blk_mq_unfreeze_queue(q, false);
209 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
212 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
213 * mpt3sas driver such that this function can be removed.
215 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
217 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
219 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
222 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
225 * Note: this function does not prevent that the struct request end_io()
226 * callback function is invoked. Once this function is returned, we make
227 * sure no dispatch can happen until the queue is unquiesced via
228 * blk_mq_unquiesce_queue().
230 void blk_mq_quiesce_queue(struct request_queue *q)
232 struct blk_mq_hw_ctx *hctx;
236 blk_mq_quiesce_queue_nowait(q);
238 queue_for_each_hw_ctx(q, hctx, i) {
239 if (hctx->flags & BLK_MQ_F_BLOCKING)
240 synchronize_srcu(hctx->srcu);
247 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
250 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
253 * This function recovers queue into the state before quiescing
254 * which is done by blk_mq_quiesce_queue.
256 void blk_mq_unquiesce_queue(struct request_queue *q)
258 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
260 /* dispatch requests which are inserted during quiescing */
261 blk_mq_run_hw_queues(q, true);
263 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
265 void blk_mq_wake_waiters(struct request_queue *q)
267 struct blk_mq_hw_ctx *hctx;
270 queue_for_each_hw_ctx(q, hctx, i)
271 if (blk_mq_hw_queue_mapped(hctx))
272 blk_mq_tag_wakeup_all(hctx->tags, true);
276 * Only need start/end time stamping if we have iostat or
277 * blk stats enabled, or using an IO scheduler.
279 static inline bool blk_mq_need_time_stamp(struct request *rq)
281 return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
284 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
285 unsigned int tag, u64 alloc_time_ns)
287 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
288 struct request *rq = tags->static_rqs[tag];
290 if (data->q->elevator) {
291 rq->tag = BLK_MQ_NO_TAG;
292 rq->internal_tag = tag;
295 rq->internal_tag = BLK_MQ_NO_TAG;
298 /* csd/requeue_work/fifo_time is initialized before use */
300 rq->mq_ctx = data->ctx;
301 rq->mq_hctx = data->hctx;
303 rq->cmd_flags = data->cmd_flags;
304 if (data->flags & BLK_MQ_REQ_PM)
305 rq->rq_flags |= RQF_PM;
306 if (blk_queue_io_stat(data->q))
307 rq->rq_flags |= RQF_IO_STAT;
308 INIT_LIST_HEAD(&rq->queuelist);
309 INIT_HLIST_NODE(&rq->hash);
310 RB_CLEAR_NODE(&rq->rb_node);
313 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
314 rq->alloc_time_ns = alloc_time_ns;
316 if (blk_mq_need_time_stamp(rq))
317 rq->start_time_ns = ktime_get_ns();
319 rq->start_time_ns = 0;
320 rq->io_start_time_ns = 0;
321 rq->stats_sectors = 0;
322 rq->nr_phys_segments = 0;
323 #if defined(CONFIG_BLK_DEV_INTEGRITY)
324 rq->nr_integrity_segments = 0;
326 blk_crypto_rq_set_defaults(rq);
327 /* tag was already set */
328 WRITE_ONCE(rq->deadline, 0);
333 rq->end_io_data = NULL;
335 data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++;
336 refcount_set(&rq->ref, 1);
338 if (!op_is_flush(data->cmd_flags)) {
339 struct elevator_queue *e = data->q->elevator;
342 if (e && e->type->ops.prepare_request) {
343 if (e->type->icq_cache)
344 blk_mq_sched_assign_ioc(rq);
346 e->type->ops.prepare_request(rq);
347 rq->rq_flags |= RQF_ELVPRIV;
351 data->hctx->queued++;
355 static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data)
357 struct request_queue *q = data->q;
358 struct elevator_queue *e = q->elevator;
359 u64 alloc_time_ns = 0;
362 /* alloc_time includes depth and tag waits */
363 if (blk_queue_rq_alloc_time(q))
364 alloc_time_ns = ktime_get_ns();
366 if (data->cmd_flags & REQ_NOWAIT)
367 data->flags |= BLK_MQ_REQ_NOWAIT;
371 * Flush/passthrough requests are special and go directly to the
372 * dispatch list. Don't include reserved tags in the
373 * limiting, as it isn't useful.
375 if (!op_is_flush(data->cmd_flags) &&
376 !blk_op_is_passthrough(data->cmd_flags) &&
377 e->type->ops.limit_depth &&
378 !(data->flags & BLK_MQ_REQ_RESERVED))
379 e->type->ops.limit_depth(data->cmd_flags, data);
383 data->ctx = blk_mq_get_ctx(q);
384 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
386 blk_mq_tag_busy(data->hctx);
389 * Waiting allocations only fail because of an inactive hctx. In that
390 * case just retry the hctx assignment and tag allocation as CPU hotplug
391 * should have migrated us to an online CPU by now.
393 tag = blk_mq_get_tag(data);
394 if (tag == BLK_MQ_NO_TAG) {
395 if (data->flags & BLK_MQ_REQ_NOWAIT)
399 * Give up the CPU and sleep for a random short time to ensure
400 * that thread using a realtime scheduling class are migrated
401 * off the CPU, and thus off the hctx that is going away.
406 return blk_mq_rq_ctx_init(data, tag, alloc_time_ns);
409 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
410 blk_mq_req_flags_t flags)
412 struct blk_mq_alloc_data data = {
420 ret = blk_queue_enter(q, flags);
424 rq = __blk_mq_alloc_request(&data);
428 rq->__sector = (sector_t) -1;
429 rq->bio = rq->biotail = NULL;
433 return ERR_PTR(-EWOULDBLOCK);
435 EXPORT_SYMBOL(blk_mq_alloc_request);
437 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
438 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
440 struct blk_mq_alloc_data data = {
445 u64 alloc_time_ns = 0;
450 /* alloc_time includes depth and tag waits */
451 if (blk_queue_rq_alloc_time(q))
452 alloc_time_ns = ktime_get_ns();
455 * If the tag allocator sleeps we could get an allocation for a
456 * different hardware context. No need to complicate the low level
457 * allocator for this for the rare use case of a command tied to
460 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
461 return ERR_PTR(-EINVAL);
463 if (hctx_idx >= q->nr_hw_queues)
464 return ERR_PTR(-EIO);
466 ret = blk_queue_enter(q, flags);
471 * Check if the hardware context is actually mapped to anything.
472 * If not tell the caller that it should skip this queue.
475 data.hctx = q->queue_hw_ctx[hctx_idx];
476 if (!blk_mq_hw_queue_mapped(data.hctx))
478 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
479 data.ctx = __blk_mq_get_ctx(q, cpu);
482 blk_mq_tag_busy(data.hctx);
485 tag = blk_mq_get_tag(&data);
486 if (tag == BLK_MQ_NO_TAG)
488 return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns);
494 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
496 static void __blk_mq_free_request(struct request *rq)
498 struct request_queue *q = rq->q;
499 struct blk_mq_ctx *ctx = rq->mq_ctx;
500 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
501 const int sched_tag = rq->internal_tag;
503 blk_crypto_free_request(rq);
504 blk_pm_mark_last_busy(rq);
506 if (rq->tag != BLK_MQ_NO_TAG)
507 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
508 if (sched_tag != BLK_MQ_NO_TAG)
509 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
510 blk_mq_sched_restart(hctx);
514 void blk_mq_free_request(struct request *rq)
516 struct request_queue *q = rq->q;
517 struct elevator_queue *e = q->elevator;
518 struct blk_mq_ctx *ctx = rq->mq_ctx;
519 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
521 if (rq->rq_flags & RQF_ELVPRIV) {
522 if (e && e->type->ops.finish_request)
523 e->type->ops.finish_request(rq);
525 put_io_context(rq->elv.icq->ioc);
530 ctx->rq_completed[rq_is_sync(rq)]++;
531 if (rq->rq_flags & RQF_MQ_INFLIGHT)
532 __blk_mq_dec_active_requests(hctx);
534 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
535 laptop_io_completion(q->disk->bdi);
539 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
540 if (refcount_dec_and_test(&rq->ref))
541 __blk_mq_free_request(rq);
543 EXPORT_SYMBOL_GPL(blk_mq_free_request);
545 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
549 if (blk_mq_need_time_stamp(rq))
550 now = ktime_get_ns();
552 if (rq->rq_flags & RQF_STATS) {
553 blk_mq_poll_stats_start(rq->q);
554 blk_stat_add(rq, now);
557 blk_mq_sched_completed_request(rq, now);
559 blk_account_io_done(rq, now);
562 rq_qos_done(rq->q, rq);
563 rq->end_io(rq, error);
565 blk_mq_free_request(rq);
568 EXPORT_SYMBOL(__blk_mq_end_request);
570 void blk_mq_end_request(struct request *rq, blk_status_t error)
572 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
574 __blk_mq_end_request(rq, error);
576 EXPORT_SYMBOL(blk_mq_end_request);
578 static void blk_complete_reqs(struct llist_head *list)
580 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
581 struct request *rq, *next;
583 llist_for_each_entry_safe(rq, next, entry, ipi_list)
584 rq->q->mq_ops->complete(rq);
587 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
589 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
592 static int blk_softirq_cpu_dead(unsigned int cpu)
594 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
598 static void __blk_mq_complete_request_remote(void *data)
600 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
603 static inline bool blk_mq_complete_need_ipi(struct request *rq)
605 int cpu = raw_smp_processor_id();
607 if (!IS_ENABLED(CONFIG_SMP) ||
608 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
611 * With force threaded interrupts enabled, raising softirq from an SMP
612 * function call will always result in waking the ksoftirqd thread.
613 * This is probably worse than completing the request on a different
616 if (force_irqthreads())
619 /* same CPU or cache domain? Complete locally */
620 if (cpu == rq->mq_ctx->cpu ||
621 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
622 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
625 /* don't try to IPI to an offline CPU */
626 return cpu_online(rq->mq_ctx->cpu);
629 static void blk_mq_complete_send_ipi(struct request *rq)
631 struct llist_head *list;
634 cpu = rq->mq_ctx->cpu;
635 list = &per_cpu(blk_cpu_done, cpu);
636 if (llist_add(&rq->ipi_list, list)) {
637 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
638 smp_call_function_single_async(cpu, &rq->csd);
642 static void blk_mq_raise_softirq(struct request *rq)
644 struct llist_head *list;
647 list = this_cpu_ptr(&blk_cpu_done);
648 if (llist_add(&rq->ipi_list, list))
649 raise_softirq(BLOCK_SOFTIRQ);
653 bool blk_mq_complete_request_remote(struct request *rq)
655 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
658 * For a polled request, always complete locallly, it's pointless
659 * to redirect the completion.
661 if (rq->cmd_flags & REQ_HIPRI)
664 if (blk_mq_complete_need_ipi(rq)) {
665 blk_mq_complete_send_ipi(rq);
669 if (rq->q->nr_hw_queues == 1) {
670 blk_mq_raise_softirq(rq);
675 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
678 * blk_mq_complete_request - end I/O on a request
679 * @rq: the request being processed
682 * Complete a request by scheduling the ->complete_rq operation.
684 void blk_mq_complete_request(struct request *rq)
686 if (!blk_mq_complete_request_remote(rq))
687 rq->q->mq_ops->complete(rq);
689 EXPORT_SYMBOL(blk_mq_complete_request);
691 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
692 __releases(hctx->srcu)
694 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
697 srcu_read_unlock(hctx->srcu, srcu_idx);
700 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
701 __acquires(hctx->srcu)
703 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
704 /* shut up gcc false positive */
708 *srcu_idx = srcu_read_lock(hctx->srcu);
712 * blk_mq_start_request - Start processing a request
713 * @rq: Pointer to request to be started
715 * Function used by device drivers to notify the block layer that a request
716 * is going to be processed now, so blk layer can do proper initializations
717 * such as starting the timeout timer.
719 void blk_mq_start_request(struct request *rq)
721 struct request_queue *q = rq->q;
723 trace_block_rq_issue(rq);
725 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
726 rq->io_start_time_ns = ktime_get_ns();
727 rq->stats_sectors = blk_rq_sectors(rq);
728 rq->rq_flags |= RQF_STATS;
732 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
735 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
737 #ifdef CONFIG_BLK_DEV_INTEGRITY
738 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
739 q->integrity.profile->prepare_fn(rq);
742 EXPORT_SYMBOL(blk_mq_start_request);
744 static void __blk_mq_requeue_request(struct request *rq)
746 struct request_queue *q = rq->q;
748 blk_mq_put_driver_tag(rq);
750 trace_block_rq_requeue(rq);
751 rq_qos_requeue(q, rq);
753 if (blk_mq_request_started(rq)) {
754 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
755 rq->rq_flags &= ~RQF_TIMED_OUT;
759 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
761 __blk_mq_requeue_request(rq);
763 /* this request will be re-inserted to io scheduler queue */
764 blk_mq_sched_requeue_request(rq);
766 BUG_ON(!list_empty(&rq->queuelist));
767 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
769 EXPORT_SYMBOL(blk_mq_requeue_request);
771 static void blk_mq_requeue_work(struct work_struct *work)
773 struct request_queue *q =
774 container_of(work, struct request_queue, requeue_work.work);
776 struct request *rq, *next;
778 spin_lock_irq(&q->requeue_lock);
779 list_splice_init(&q->requeue_list, &rq_list);
780 spin_unlock_irq(&q->requeue_lock);
782 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
783 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
786 rq->rq_flags &= ~RQF_SOFTBARRIER;
787 list_del_init(&rq->queuelist);
789 * If RQF_DONTPREP, rq has contained some driver specific
790 * data, so insert it to hctx dispatch list to avoid any
793 if (rq->rq_flags & RQF_DONTPREP)
794 blk_mq_request_bypass_insert(rq, false, false);
796 blk_mq_sched_insert_request(rq, true, false, false);
799 while (!list_empty(&rq_list)) {
800 rq = list_entry(rq_list.next, struct request, queuelist);
801 list_del_init(&rq->queuelist);
802 blk_mq_sched_insert_request(rq, false, false, false);
805 blk_mq_run_hw_queues(q, false);
808 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
809 bool kick_requeue_list)
811 struct request_queue *q = rq->q;
815 * We abuse this flag that is otherwise used by the I/O scheduler to
816 * request head insertion from the workqueue.
818 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
820 spin_lock_irqsave(&q->requeue_lock, flags);
822 rq->rq_flags |= RQF_SOFTBARRIER;
823 list_add(&rq->queuelist, &q->requeue_list);
825 list_add_tail(&rq->queuelist, &q->requeue_list);
827 spin_unlock_irqrestore(&q->requeue_lock, flags);
829 if (kick_requeue_list)
830 blk_mq_kick_requeue_list(q);
833 void blk_mq_kick_requeue_list(struct request_queue *q)
835 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
837 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
839 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
842 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
843 msecs_to_jiffies(msecs));
845 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
847 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
849 if (tag < tags->nr_tags) {
850 prefetch(tags->rqs[tag]);
851 return tags->rqs[tag];
856 EXPORT_SYMBOL(blk_mq_tag_to_rq);
858 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
859 void *priv, bool reserved)
862 * If we find a request that isn't idle and the queue matches,
863 * we know the queue is busy. Return false to stop the iteration.
865 if (blk_mq_request_started(rq) && rq->q == hctx->queue) {
875 bool blk_mq_queue_inflight(struct request_queue *q)
879 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
882 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
884 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
886 req->rq_flags |= RQF_TIMED_OUT;
887 if (req->q->mq_ops->timeout) {
888 enum blk_eh_timer_return ret;
890 ret = req->q->mq_ops->timeout(req, reserved);
891 if (ret == BLK_EH_DONE)
893 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
899 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
901 unsigned long deadline;
903 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
905 if (rq->rq_flags & RQF_TIMED_OUT)
908 deadline = READ_ONCE(rq->deadline);
909 if (time_after_eq(jiffies, deadline))
914 else if (time_after(*next, deadline))
919 void blk_mq_put_rq_ref(struct request *rq)
923 else if (refcount_dec_and_test(&rq->ref))
924 __blk_mq_free_request(rq);
927 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
928 struct request *rq, void *priv, bool reserved)
930 unsigned long *next = priv;
933 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
934 * be reallocated underneath the timeout handler's processing, then
935 * the expire check is reliable. If the request is not expired, then
936 * it was completed and reallocated as a new request after returning
937 * from blk_mq_check_expired().
939 if (blk_mq_req_expired(rq, next))
940 blk_mq_rq_timed_out(rq, reserved);
944 static void blk_mq_timeout_work(struct work_struct *work)
946 struct request_queue *q =
947 container_of(work, struct request_queue, timeout_work);
948 unsigned long next = 0;
949 struct blk_mq_hw_ctx *hctx;
952 /* A deadlock might occur if a request is stuck requiring a
953 * timeout at the same time a queue freeze is waiting
954 * completion, since the timeout code would not be able to
955 * acquire the queue reference here.
957 * That's why we don't use blk_queue_enter here; instead, we use
958 * percpu_ref_tryget directly, because we need to be able to
959 * obtain a reference even in the short window between the queue
960 * starting to freeze, by dropping the first reference in
961 * blk_freeze_queue_start, and the moment the last request is
962 * consumed, marked by the instant q_usage_counter reaches
965 if (!percpu_ref_tryget(&q->q_usage_counter))
968 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
971 mod_timer(&q->timeout, next);
974 * Request timeouts are handled as a forward rolling timer. If
975 * we end up here it means that no requests are pending and
976 * also that no request has been pending for a while. Mark
979 queue_for_each_hw_ctx(q, hctx, i) {
980 /* the hctx may be unmapped, so check it here */
981 if (blk_mq_hw_queue_mapped(hctx))
982 blk_mq_tag_idle(hctx);
988 struct flush_busy_ctx_data {
989 struct blk_mq_hw_ctx *hctx;
990 struct list_head *list;
993 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
995 struct flush_busy_ctx_data *flush_data = data;
996 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
997 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
998 enum hctx_type type = hctx->type;
1000 spin_lock(&ctx->lock);
1001 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1002 sbitmap_clear_bit(sb, bitnr);
1003 spin_unlock(&ctx->lock);
1008 * Process software queues that have been marked busy, splicing them
1009 * to the for-dispatch
1011 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1013 struct flush_busy_ctx_data data = {
1018 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1020 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1022 struct dispatch_rq_data {
1023 struct blk_mq_hw_ctx *hctx;
1027 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1030 struct dispatch_rq_data *dispatch_data = data;
1031 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1032 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1033 enum hctx_type type = hctx->type;
1035 spin_lock(&ctx->lock);
1036 if (!list_empty(&ctx->rq_lists[type])) {
1037 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1038 list_del_init(&dispatch_data->rq->queuelist);
1039 if (list_empty(&ctx->rq_lists[type]))
1040 sbitmap_clear_bit(sb, bitnr);
1042 spin_unlock(&ctx->lock);
1044 return !dispatch_data->rq;
1047 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1048 struct blk_mq_ctx *start)
1050 unsigned off = start ? start->index_hw[hctx->type] : 0;
1051 struct dispatch_rq_data data = {
1056 __sbitmap_for_each_set(&hctx->ctx_map, off,
1057 dispatch_rq_from_ctx, &data);
1062 static inline unsigned int queued_to_index(unsigned int queued)
1067 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1070 static bool __blk_mq_get_driver_tag(struct request *rq)
1072 struct sbitmap_queue *bt = rq->mq_hctx->tags->bitmap_tags;
1073 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1076 blk_mq_tag_busy(rq->mq_hctx);
1078 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1079 bt = rq->mq_hctx->tags->breserved_tags;
1082 if (!hctx_may_queue(rq->mq_hctx, bt))
1086 tag = __sbitmap_queue_get(bt);
1087 if (tag == BLK_MQ_NO_TAG)
1090 rq->tag = tag + tag_offset;
1094 bool blk_mq_get_driver_tag(struct request *rq)
1096 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1098 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq))
1101 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1102 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1103 rq->rq_flags |= RQF_MQ_INFLIGHT;
1104 __blk_mq_inc_active_requests(hctx);
1106 hctx->tags->rqs[rq->tag] = rq;
1110 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1111 int flags, void *key)
1113 struct blk_mq_hw_ctx *hctx;
1115 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1117 spin_lock(&hctx->dispatch_wait_lock);
1118 if (!list_empty(&wait->entry)) {
1119 struct sbitmap_queue *sbq;
1121 list_del_init(&wait->entry);
1122 sbq = hctx->tags->bitmap_tags;
1123 atomic_dec(&sbq->ws_active);
1125 spin_unlock(&hctx->dispatch_wait_lock);
1127 blk_mq_run_hw_queue(hctx, true);
1132 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1133 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1134 * restart. For both cases, take care to check the condition again after
1135 * marking us as waiting.
1137 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1140 struct sbitmap_queue *sbq = hctx->tags->bitmap_tags;
1141 struct wait_queue_head *wq;
1142 wait_queue_entry_t *wait;
1145 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1146 blk_mq_sched_mark_restart_hctx(hctx);
1149 * It's possible that a tag was freed in the window between the
1150 * allocation failure and adding the hardware queue to the wait
1153 * Don't clear RESTART here, someone else could have set it.
1154 * At most this will cost an extra queue run.
1156 return blk_mq_get_driver_tag(rq);
1159 wait = &hctx->dispatch_wait;
1160 if (!list_empty_careful(&wait->entry))
1163 wq = &bt_wait_ptr(sbq, hctx)->wait;
1165 spin_lock_irq(&wq->lock);
1166 spin_lock(&hctx->dispatch_wait_lock);
1167 if (!list_empty(&wait->entry)) {
1168 spin_unlock(&hctx->dispatch_wait_lock);
1169 spin_unlock_irq(&wq->lock);
1173 atomic_inc(&sbq->ws_active);
1174 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1175 __add_wait_queue(wq, wait);
1178 * It's possible that a tag was freed in the window between the
1179 * allocation failure and adding the hardware queue to the wait
1182 ret = blk_mq_get_driver_tag(rq);
1184 spin_unlock(&hctx->dispatch_wait_lock);
1185 spin_unlock_irq(&wq->lock);
1190 * We got a tag, remove ourselves from the wait queue to ensure
1191 * someone else gets the wakeup.
1193 list_del_init(&wait->entry);
1194 atomic_dec(&sbq->ws_active);
1195 spin_unlock(&hctx->dispatch_wait_lock);
1196 spin_unlock_irq(&wq->lock);
1201 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1202 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1204 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1205 * - EWMA is one simple way to compute running average value
1206 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1207 * - take 4 as factor for avoiding to get too small(0) result, and this
1208 * factor doesn't matter because EWMA decreases exponentially
1210 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1214 ewma = hctx->dispatch_busy;
1219 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1221 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1222 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1224 hctx->dispatch_busy = ewma;
1227 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1229 static void blk_mq_handle_dev_resource(struct request *rq,
1230 struct list_head *list)
1232 struct request *next =
1233 list_first_entry_or_null(list, struct request, queuelist);
1236 * If an I/O scheduler has been configured and we got a driver tag for
1237 * the next request already, free it.
1240 blk_mq_put_driver_tag(next);
1242 list_add(&rq->queuelist, list);
1243 __blk_mq_requeue_request(rq);
1246 static void blk_mq_handle_zone_resource(struct request *rq,
1247 struct list_head *zone_list)
1250 * If we end up here it is because we cannot dispatch a request to a
1251 * specific zone due to LLD level zone-write locking or other zone
1252 * related resource not being available. In this case, set the request
1253 * aside in zone_list for retrying it later.
1255 list_add(&rq->queuelist, zone_list);
1256 __blk_mq_requeue_request(rq);
1259 enum prep_dispatch {
1261 PREP_DISPATCH_NO_TAG,
1262 PREP_DISPATCH_NO_BUDGET,
1265 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1268 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1269 int budget_token = -1;
1272 budget_token = blk_mq_get_dispatch_budget(rq->q);
1273 if (budget_token < 0) {
1274 blk_mq_put_driver_tag(rq);
1275 return PREP_DISPATCH_NO_BUDGET;
1277 blk_mq_set_rq_budget_token(rq, budget_token);
1280 if (!blk_mq_get_driver_tag(rq)) {
1282 * The initial allocation attempt failed, so we need to
1283 * rerun the hardware queue when a tag is freed. The
1284 * waitqueue takes care of that. If the queue is run
1285 * before we add this entry back on the dispatch list,
1286 * we'll re-run it below.
1288 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1290 * All budgets not got from this function will be put
1291 * together during handling partial dispatch
1294 blk_mq_put_dispatch_budget(rq->q, budget_token);
1295 return PREP_DISPATCH_NO_TAG;
1299 return PREP_DISPATCH_OK;
1302 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1303 static void blk_mq_release_budgets(struct request_queue *q,
1304 struct list_head *list)
1308 list_for_each_entry(rq, list, queuelist) {
1309 int budget_token = blk_mq_get_rq_budget_token(rq);
1311 if (budget_token >= 0)
1312 blk_mq_put_dispatch_budget(q, budget_token);
1317 * Returns true if we did some work AND can potentially do more.
1319 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1320 unsigned int nr_budgets)
1322 enum prep_dispatch prep;
1323 struct request_queue *q = hctx->queue;
1324 struct request *rq, *nxt;
1326 blk_status_t ret = BLK_STS_OK;
1327 LIST_HEAD(zone_list);
1328 bool needs_resource = false;
1330 if (list_empty(list))
1334 * Now process all the entries, sending them to the driver.
1336 errors = queued = 0;
1338 struct blk_mq_queue_data bd;
1340 rq = list_first_entry(list, struct request, queuelist);
1342 WARN_ON_ONCE(hctx != rq->mq_hctx);
1343 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1344 if (prep != PREP_DISPATCH_OK)
1347 list_del_init(&rq->queuelist);
1352 * Flag last if we have no more requests, or if we have more
1353 * but can't assign a driver tag to it.
1355 if (list_empty(list))
1358 nxt = list_first_entry(list, struct request, queuelist);
1359 bd.last = !blk_mq_get_driver_tag(nxt);
1363 * once the request is queued to lld, no need to cover the
1368 ret = q->mq_ops->queue_rq(hctx, &bd);
1373 case BLK_STS_RESOURCE:
1374 needs_resource = true;
1376 case BLK_STS_DEV_RESOURCE:
1377 blk_mq_handle_dev_resource(rq, list);
1379 case BLK_STS_ZONE_RESOURCE:
1381 * Move the request to zone_list and keep going through
1382 * the dispatch list to find more requests the drive can
1385 blk_mq_handle_zone_resource(rq, &zone_list);
1386 needs_resource = true;
1390 blk_mq_end_request(rq, ret);
1392 } while (!list_empty(list));
1394 if (!list_empty(&zone_list))
1395 list_splice_tail_init(&zone_list, list);
1397 hctx->dispatched[queued_to_index(queued)]++;
1399 /* If we didn't flush the entire list, we could have told the driver
1400 * there was more coming, but that turned out to be a lie.
1402 if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued)
1403 q->mq_ops->commit_rqs(hctx);
1405 * Any items that need requeuing? Stuff them into hctx->dispatch,
1406 * that is where we will continue on next queue run.
1408 if (!list_empty(list)) {
1410 /* For non-shared tags, the RESTART check will suffice */
1411 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1412 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1415 blk_mq_release_budgets(q, list);
1417 spin_lock(&hctx->lock);
1418 list_splice_tail_init(list, &hctx->dispatch);
1419 spin_unlock(&hctx->lock);
1422 * Order adding requests to hctx->dispatch and checking
1423 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1424 * in blk_mq_sched_restart(). Avoid restart code path to
1425 * miss the new added requests to hctx->dispatch, meantime
1426 * SCHED_RESTART is observed here.
1431 * If SCHED_RESTART was set by the caller of this function and
1432 * it is no longer set that means that it was cleared by another
1433 * thread and hence that a queue rerun is needed.
1435 * If 'no_tag' is set, that means that we failed getting
1436 * a driver tag with an I/O scheduler attached. If our dispatch
1437 * waitqueue is no longer active, ensure that we run the queue
1438 * AFTER adding our entries back to the list.
1440 * If no I/O scheduler has been configured it is possible that
1441 * the hardware queue got stopped and restarted before requests
1442 * were pushed back onto the dispatch list. Rerun the queue to
1443 * avoid starvation. Notes:
1444 * - blk_mq_run_hw_queue() checks whether or not a queue has
1445 * been stopped before rerunning a queue.
1446 * - Some but not all block drivers stop a queue before
1447 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1450 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1451 * bit is set, run queue after a delay to avoid IO stalls
1452 * that could otherwise occur if the queue is idle. We'll do
1453 * similar if we couldn't get budget or couldn't lock a zone
1454 * and SCHED_RESTART is set.
1456 needs_restart = blk_mq_sched_needs_restart(hctx);
1457 if (prep == PREP_DISPATCH_NO_BUDGET)
1458 needs_resource = true;
1459 if (!needs_restart ||
1460 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1461 blk_mq_run_hw_queue(hctx, true);
1462 else if (needs_restart && needs_resource)
1463 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1465 blk_mq_update_dispatch_busy(hctx, true);
1468 blk_mq_update_dispatch_busy(hctx, false);
1470 return (queued + errors) != 0;
1474 * __blk_mq_run_hw_queue - Run a hardware queue.
1475 * @hctx: Pointer to the hardware queue to run.
1477 * Send pending requests to the hardware.
1479 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1484 * We can't run the queue inline with ints disabled. Ensure that
1485 * we catch bad users of this early.
1487 WARN_ON_ONCE(in_interrupt());
1489 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1491 hctx_lock(hctx, &srcu_idx);
1492 blk_mq_sched_dispatch_requests(hctx);
1493 hctx_unlock(hctx, srcu_idx);
1496 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1498 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1500 if (cpu >= nr_cpu_ids)
1501 cpu = cpumask_first(hctx->cpumask);
1506 * It'd be great if the workqueue API had a way to pass
1507 * in a mask and had some smarts for more clever placement.
1508 * For now we just round-robin here, switching for every
1509 * BLK_MQ_CPU_WORK_BATCH queued items.
1511 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1514 int next_cpu = hctx->next_cpu;
1516 if (hctx->queue->nr_hw_queues == 1)
1517 return WORK_CPU_UNBOUND;
1519 if (--hctx->next_cpu_batch <= 0) {
1521 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1523 if (next_cpu >= nr_cpu_ids)
1524 next_cpu = blk_mq_first_mapped_cpu(hctx);
1525 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1529 * Do unbound schedule if we can't find a online CPU for this hctx,
1530 * and it should only happen in the path of handling CPU DEAD.
1532 if (!cpu_online(next_cpu)) {
1539 * Make sure to re-select CPU next time once after CPUs
1540 * in hctx->cpumask become online again.
1542 hctx->next_cpu = next_cpu;
1543 hctx->next_cpu_batch = 1;
1544 return WORK_CPU_UNBOUND;
1547 hctx->next_cpu = next_cpu;
1552 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1553 * @hctx: Pointer to the hardware queue to run.
1554 * @async: If we want to run the queue asynchronously.
1555 * @msecs: Milliseconds of delay to wait before running the queue.
1557 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1558 * with a delay of @msecs.
1560 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1561 unsigned long msecs)
1563 if (unlikely(blk_mq_hctx_stopped(hctx)))
1566 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1567 int cpu = get_cpu();
1568 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1569 __blk_mq_run_hw_queue(hctx);
1577 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1578 msecs_to_jiffies(msecs));
1582 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1583 * @hctx: Pointer to the hardware queue to run.
1584 * @msecs: Milliseconds of delay to wait before running the queue.
1586 * Run a hardware queue asynchronously with a delay of @msecs.
1588 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1590 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1592 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1595 * blk_mq_run_hw_queue - Start to run a hardware queue.
1596 * @hctx: Pointer to the hardware queue to run.
1597 * @async: If we want to run the queue asynchronously.
1599 * Check if the request queue is not in a quiesced state and if there are
1600 * pending requests to be sent. If this is true, run the queue to send requests
1603 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1609 * When queue is quiesced, we may be switching io scheduler, or
1610 * updating nr_hw_queues, or other things, and we can't run queue
1611 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1613 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1616 hctx_lock(hctx, &srcu_idx);
1617 need_run = !blk_queue_quiesced(hctx->queue) &&
1618 blk_mq_hctx_has_pending(hctx);
1619 hctx_unlock(hctx, srcu_idx);
1622 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1624 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1627 * Is the request queue handled by an IO scheduler that does not respect
1628 * hardware queues when dispatching?
1630 static bool blk_mq_has_sqsched(struct request_queue *q)
1632 struct elevator_queue *e = q->elevator;
1634 if (e && e->type->ops.dispatch_request &&
1635 !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE))
1641 * Return prefered queue to dispatch from (if any) for non-mq aware IO
1644 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
1646 struct blk_mq_hw_ctx *hctx;
1649 * If the IO scheduler does not respect hardware queues when
1650 * dispatching, we just don't bother with multiple HW queues and
1651 * dispatch from hctx for the current CPU since running multiple queues
1652 * just causes lock contention inside the scheduler and pointless cache
1655 hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT,
1656 raw_smp_processor_id());
1657 if (!blk_mq_hctx_stopped(hctx))
1663 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
1664 * @q: Pointer to the request queue to run.
1665 * @async: If we want to run the queue asynchronously.
1667 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1669 struct blk_mq_hw_ctx *hctx, *sq_hctx;
1673 if (blk_mq_has_sqsched(q))
1674 sq_hctx = blk_mq_get_sq_hctx(q);
1675 queue_for_each_hw_ctx(q, hctx, i) {
1676 if (blk_mq_hctx_stopped(hctx))
1679 * Dispatch from this hctx either if there's no hctx preferred
1680 * by IO scheduler or if it has requests that bypass the
1683 if (!sq_hctx || sq_hctx == hctx ||
1684 !list_empty_careful(&hctx->dispatch))
1685 blk_mq_run_hw_queue(hctx, async);
1688 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1691 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1692 * @q: Pointer to the request queue to run.
1693 * @msecs: Milliseconds of delay to wait before running the queues.
1695 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1697 struct blk_mq_hw_ctx *hctx, *sq_hctx;
1701 if (blk_mq_has_sqsched(q))
1702 sq_hctx = blk_mq_get_sq_hctx(q);
1703 queue_for_each_hw_ctx(q, hctx, i) {
1704 if (blk_mq_hctx_stopped(hctx))
1707 * Dispatch from this hctx either if there's no hctx preferred
1708 * by IO scheduler or if it has requests that bypass the
1711 if (!sq_hctx || sq_hctx == hctx ||
1712 !list_empty_careful(&hctx->dispatch))
1713 blk_mq_delay_run_hw_queue(hctx, msecs);
1716 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1719 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1720 * @q: request queue.
1722 * The caller is responsible for serializing this function against
1723 * blk_mq_{start,stop}_hw_queue().
1725 bool blk_mq_queue_stopped(struct request_queue *q)
1727 struct blk_mq_hw_ctx *hctx;
1730 queue_for_each_hw_ctx(q, hctx, i)
1731 if (blk_mq_hctx_stopped(hctx))
1736 EXPORT_SYMBOL(blk_mq_queue_stopped);
1739 * This function is often used for pausing .queue_rq() by driver when
1740 * there isn't enough resource or some conditions aren't satisfied, and
1741 * BLK_STS_RESOURCE is usually returned.
1743 * We do not guarantee that dispatch can be drained or blocked
1744 * after blk_mq_stop_hw_queue() returns. Please use
1745 * blk_mq_quiesce_queue() for that requirement.
1747 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1749 cancel_delayed_work(&hctx->run_work);
1751 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1753 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1756 * This function is often used for pausing .queue_rq() by driver when
1757 * there isn't enough resource or some conditions aren't satisfied, and
1758 * BLK_STS_RESOURCE is usually returned.
1760 * We do not guarantee that dispatch can be drained or blocked
1761 * after blk_mq_stop_hw_queues() returns. Please use
1762 * blk_mq_quiesce_queue() for that requirement.
1764 void blk_mq_stop_hw_queues(struct request_queue *q)
1766 struct blk_mq_hw_ctx *hctx;
1769 queue_for_each_hw_ctx(q, hctx, i)
1770 blk_mq_stop_hw_queue(hctx);
1772 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1774 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1776 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1778 blk_mq_run_hw_queue(hctx, false);
1780 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1782 void blk_mq_start_hw_queues(struct request_queue *q)
1784 struct blk_mq_hw_ctx *hctx;
1787 queue_for_each_hw_ctx(q, hctx, i)
1788 blk_mq_start_hw_queue(hctx);
1790 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1792 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1794 if (!blk_mq_hctx_stopped(hctx))
1797 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1798 blk_mq_run_hw_queue(hctx, async);
1800 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1802 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1804 struct blk_mq_hw_ctx *hctx;
1807 queue_for_each_hw_ctx(q, hctx, i)
1808 blk_mq_start_stopped_hw_queue(hctx, async);
1810 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1812 static void blk_mq_run_work_fn(struct work_struct *work)
1814 struct blk_mq_hw_ctx *hctx;
1816 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1819 * If we are stopped, don't run the queue.
1821 if (blk_mq_hctx_stopped(hctx))
1824 __blk_mq_run_hw_queue(hctx);
1827 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1831 struct blk_mq_ctx *ctx = rq->mq_ctx;
1832 enum hctx_type type = hctx->type;
1834 lockdep_assert_held(&ctx->lock);
1836 trace_block_rq_insert(rq);
1839 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1841 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1844 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1847 struct blk_mq_ctx *ctx = rq->mq_ctx;
1849 lockdep_assert_held(&ctx->lock);
1851 __blk_mq_insert_req_list(hctx, rq, at_head);
1852 blk_mq_hctx_mark_pending(hctx, ctx);
1856 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1857 * @rq: Pointer to request to be inserted.
1858 * @at_head: true if the request should be inserted at the head of the list.
1859 * @run_queue: If we should run the hardware queue after inserting the request.
1861 * Should only be used carefully, when the caller knows we want to
1862 * bypass a potential IO scheduler on the target device.
1864 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1867 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1869 spin_lock(&hctx->lock);
1871 list_add(&rq->queuelist, &hctx->dispatch);
1873 list_add_tail(&rq->queuelist, &hctx->dispatch);
1874 spin_unlock(&hctx->lock);
1877 blk_mq_run_hw_queue(hctx, false);
1880 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1881 struct list_head *list)
1885 enum hctx_type type = hctx->type;
1888 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1891 list_for_each_entry(rq, list, queuelist) {
1892 BUG_ON(rq->mq_ctx != ctx);
1893 trace_block_rq_insert(rq);
1896 spin_lock(&ctx->lock);
1897 list_splice_tail_init(list, &ctx->rq_lists[type]);
1898 blk_mq_hctx_mark_pending(hctx, ctx);
1899 spin_unlock(&ctx->lock);
1902 static int plug_rq_cmp(void *priv, const struct list_head *a,
1903 const struct list_head *b)
1905 struct request *rqa = container_of(a, struct request, queuelist);
1906 struct request *rqb = container_of(b, struct request, queuelist);
1908 if (rqa->mq_ctx != rqb->mq_ctx)
1909 return rqa->mq_ctx > rqb->mq_ctx;
1910 if (rqa->mq_hctx != rqb->mq_hctx)
1911 return rqa->mq_hctx > rqb->mq_hctx;
1913 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1916 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1920 if (list_empty(&plug->mq_list))
1922 list_splice_init(&plug->mq_list, &list);
1924 if (plug->rq_count > 2 && plug->multiple_queues)
1925 list_sort(NULL, &list, plug_rq_cmp);
1930 struct list_head rq_list;
1931 struct request *rq, *head_rq = list_entry_rq(list.next);
1932 struct list_head *pos = &head_rq->queuelist; /* skip first */
1933 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1934 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1935 unsigned int depth = 1;
1937 list_for_each_continue(pos, &list) {
1938 rq = list_entry_rq(pos);
1940 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1945 list_cut_before(&rq_list, &list, pos);
1946 trace_block_unplug(head_rq->q, depth, !from_schedule);
1947 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1949 } while(!list_empty(&list));
1952 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1953 unsigned int nr_segs)
1957 if (bio->bi_opf & REQ_RAHEAD)
1958 rq->cmd_flags |= REQ_FAILFAST_MASK;
1960 rq->__sector = bio->bi_iter.bi_sector;
1961 rq->write_hint = bio->bi_write_hint;
1962 blk_rq_bio_prep(rq, bio, nr_segs);
1964 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
1965 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1968 blk_account_io_start(rq);
1971 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1973 blk_qc_t *cookie, bool last)
1975 struct request_queue *q = rq->q;
1976 struct blk_mq_queue_data bd = {
1980 blk_qc_t new_cookie;
1983 new_cookie = request_to_qc_t(hctx, rq);
1986 * For OK queue, we are done. For error, caller may kill it.
1987 * Any other error (busy), just add it to our list as we
1988 * previously would have done.
1990 ret = q->mq_ops->queue_rq(hctx, &bd);
1993 blk_mq_update_dispatch_busy(hctx, false);
1994 *cookie = new_cookie;
1996 case BLK_STS_RESOURCE:
1997 case BLK_STS_DEV_RESOURCE:
1998 blk_mq_update_dispatch_busy(hctx, true);
1999 __blk_mq_requeue_request(rq);
2002 blk_mq_update_dispatch_busy(hctx, false);
2003 *cookie = BLK_QC_T_NONE;
2010 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2013 bool bypass_insert, bool last)
2015 struct request_queue *q = rq->q;
2016 bool run_queue = true;
2020 * RCU or SRCU read lock is needed before checking quiesced flag.
2022 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2023 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2024 * and avoid driver to try to dispatch again.
2026 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2028 bypass_insert = false;
2032 if (q->elevator && !bypass_insert)
2035 budget_token = blk_mq_get_dispatch_budget(q);
2036 if (budget_token < 0)
2039 blk_mq_set_rq_budget_token(rq, budget_token);
2041 if (!blk_mq_get_driver_tag(rq)) {
2042 blk_mq_put_dispatch_budget(q, budget_token);
2046 return __blk_mq_issue_directly(hctx, rq, cookie, last);
2049 return BLK_STS_RESOURCE;
2051 blk_mq_sched_insert_request(rq, false, run_queue, false);
2057 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2058 * @hctx: Pointer of the associated hardware queue.
2059 * @rq: Pointer to request to be sent.
2060 * @cookie: Request queue cookie.
2062 * If the device has enough resources to accept a new request now, send the
2063 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2064 * we can try send it another time in the future. Requests inserted at this
2065 * queue have higher priority.
2067 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2068 struct request *rq, blk_qc_t *cookie)
2073 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2075 hctx_lock(hctx, &srcu_idx);
2077 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
2078 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2079 blk_mq_request_bypass_insert(rq, false, true);
2080 else if (ret != BLK_STS_OK)
2081 blk_mq_end_request(rq, ret);
2083 hctx_unlock(hctx, srcu_idx);
2086 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2090 blk_qc_t unused_cookie;
2091 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2093 hctx_lock(hctx, &srcu_idx);
2094 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
2095 hctx_unlock(hctx, srcu_idx);
2100 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2101 struct list_head *list)
2106 while (!list_empty(list)) {
2108 struct request *rq = list_first_entry(list, struct request,
2111 list_del_init(&rq->queuelist);
2112 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2113 if (ret != BLK_STS_OK) {
2114 if (ret == BLK_STS_RESOURCE ||
2115 ret == BLK_STS_DEV_RESOURCE) {
2116 blk_mq_request_bypass_insert(rq, false,
2120 blk_mq_end_request(rq, ret);
2127 * If we didn't flush the entire list, we could have told
2128 * the driver there was more coming, but that turned out to
2131 if ((!list_empty(list) || errors) &&
2132 hctx->queue->mq_ops->commit_rqs && queued)
2133 hctx->queue->mq_ops->commit_rqs(hctx);
2136 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2138 list_add_tail(&rq->queuelist, &plug->mq_list);
2140 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
2141 struct request *tmp;
2143 tmp = list_first_entry(&plug->mq_list, struct request,
2145 if (tmp->q != rq->q)
2146 plug->multiple_queues = true;
2151 * Allow 4x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
2152 * queues. This is important for md arrays to benefit from merging
2155 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
2157 if (plug->multiple_queues)
2158 return BLK_MAX_REQUEST_COUNT * 4;
2159 return BLK_MAX_REQUEST_COUNT;
2163 * blk_mq_submit_bio - Create and send a request to block device.
2164 * @bio: Bio pointer.
2166 * Builds up a request structure from @q and @bio and send to the device. The
2167 * request may not be queued directly to hardware if:
2168 * * This request can be merged with another one
2169 * * We want to place request at plug queue for possible future merging
2170 * * There is an IO scheduler active at this queue
2172 * It will not queue the request if there is an error with the bio, or at the
2175 * Returns: Request queue cookie.
2177 blk_qc_t blk_mq_submit_bio(struct bio *bio)
2179 struct request_queue *q = bio->bi_bdev->bd_disk->queue;
2180 const int is_sync = op_is_sync(bio->bi_opf);
2181 const int is_flush_fua = op_is_flush(bio->bi_opf);
2182 struct blk_mq_alloc_data data = {
2186 struct blk_plug *plug;
2187 struct request *same_queue_rq = NULL;
2188 unsigned int nr_segs;
2193 blk_queue_bounce(q, &bio);
2194 __blk_queue_split(&bio, &nr_segs);
2196 if (!bio_integrity_prep(bio))
2199 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2200 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2203 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2206 rq_qos_throttle(q, bio);
2208 hipri = bio->bi_opf & REQ_HIPRI;
2210 data.cmd_flags = bio->bi_opf;
2211 rq = __blk_mq_alloc_request(&data);
2212 if (unlikely(!rq)) {
2213 rq_qos_cleanup(q, bio);
2214 if (bio->bi_opf & REQ_NOWAIT)
2215 bio_wouldblock_error(bio);
2219 trace_block_getrq(bio);
2221 rq_qos_track(q, rq, bio);
2223 cookie = request_to_qc_t(data.hctx, rq);
2225 blk_mq_bio_to_request(rq, bio, nr_segs);
2227 ret = blk_crypto_init_request(rq);
2228 if (ret != BLK_STS_OK) {
2229 bio->bi_status = ret;
2231 blk_mq_free_request(rq);
2232 return BLK_QC_T_NONE;
2235 plug = blk_mq_plug(q, bio);
2236 if (unlikely(is_flush_fua)) {
2237 /* Bypass scheduler for flush requests */
2238 blk_insert_flush(rq);
2239 blk_mq_run_hw_queue(data.hctx, true);
2240 } else if (plug && (q->nr_hw_queues == 1 ||
2241 blk_mq_is_sbitmap_shared(rq->mq_hctx->flags) ||
2242 q->mq_ops->commit_rqs || !blk_queue_nonrot(q))) {
2244 * Use plugging if we have a ->commit_rqs() hook as well, as
2245 * we know the driver uses bd->last in a smart fashion.
2247 * Use normal plugging if this disk is slow HDD, as sequential
2248 * IO may benefit a lot from plug merging.
2250 unsigned int request_count = plug->rq_count;
2251 struct request *last = NULL;
2254 trace_block_plug(q);
2256 last = list_entry_rq(plug->mq_list.prev);
2258 if (request_count >= blk_plug_max_rq_count(plug) || (last &&
2259 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2260 blk_flush_plug_list(plug, false);
2261 trace_block_plug(q);
2264 blk_add_rq_to_plug(plug, rq);
2265 } else if (q->elevator) {
2266 /* Insert the request at the IO scheduler queue */
2267 blk_mq_sched_insert_request(rq, false, true, true);
2268 } else if (plug && !blk_queue_nomerges(q)) {
2270 * We do limited plugging. If the bio can be merged, do that.
2271 * Otherwise the existing request in the plug list will be
2272 * issued. So the plug list will have one request at most
2273 * The plug list might get flushed before this. If that happens,
2274 * the plug list is empty, and same_queue_rq is invalid.
2276 if (list_empty(&plug->mq_list))
2277 same_queue_rq = NULL;
2278 if (same_queue_rq) {
2279 list_del_init(&same_queue_rq->queuelist);
2282 blk_add_rq_to_plug(plug, rq);
2283 trace_block_plug(q);
2285 if (same_queue_rq) {
2286 data.hctx = same_queue_rq->mq_hctx;
2287 trace_block_unplug(q, 1, true);
2288 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2291 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2292 !data.hctx->dispatch_busy) {
2294 * There is no scheduler and we can try to send directly
2297 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2300 blk_mq_sched_insert_request(rq, false, true, true);
2304 return BLK_QC_T_NONE;
2308 return BLK_QC_T_NONE;
2311 static size_t order_to_size(unsigned int order)
2313 return (size_t)PAGE_SIZE << order;
2316 /* called before freeing request pool in @tags */
2317 static void blk_mq_clear_rq_mapping(struct blk_mq_tag_set *set,
2318 struct blk_mq_tags *tags, unsigned int hctx_idx)
2320 struct blk_mq_tags *drv_tags = set->tags[hctx_idx];
2322 unsigned long flags;
2324 list_for_each_entry(page, &tags->page_list, lru) {
2325 unsigned long start = (unsigned long)page_address(page);
2326 unsigned long end = start + order_to_size(page->private);
2329 for (i = 0; i < set->queue_depth; i++) {
2330 struct request *rq = drv_tags->rqs[i];
2331 unsigned long rq_addr = (unsigned long)rq;
2333 if (rq_addr >= start && rq_addr < end) {
2334 WARN_ON_ONCE(refcount_read(&rq->ref) != 0);
2335 cmpxchg(&drv_tags->rqs[i], rq, NULL);
2341 * Wait until all pending iteration is done.
2343 * Request reference is cleared and it is guaranteed to be observed
2344 * after the ->lock is released.
2346 spin_lock_irqsave(&drv_tags->lock, flags);
2347 spin_unlock_irqrestore(&drv_tags->lock, flags);
2350 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2351 unsigned int hctx_idx)
2355 if (tags->rqs && set->ops->exit_request) {
2358 for (i = 0; i < tags->nr_tags; i++) {
2359 struct request *rq = tags->static_rqs[i];
2363 set->ops->exit_request(set, rq, hctx_idx);
2364 tags->static_rqs[i] = NULL;
2368 blk_mq_clear_rq_mapping(set, tags, hctx_idx);
2370 while (!list_empty(&tags->page_list)) {
2371 page = list_first_entry(&tags->page_list, struct page, lru);
2372 list_del_init(&page->lru);
2374 * Remove kmemleak object previously allocated in
2375 * blk_mq_alloc_rqs().
2377 kmemleak_free(page_address(page));
2378 __free_pages(page, page->private);
2382 void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags)
2386 kfree(tags->static_rqs);
2387 tags->static_rqs = NULL;
2389 blk_mq_free_tags(tags, flags);
2392 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2393 unsigned int hctx_idx,
2394 unsigned int nr_tags,
2395 unsigned int reserved_tags,
2398 struct blk_mq_tags *tags;
2401 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2402 if (node == NUMA_NO_NODE)
2403 node = set->numa_node;
2405 tags = blk_mq_init_tags(nr_tags, reserved_tags, node, flags);
2409 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2410 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2413 blk_mq_free_tags(tags, flags);
2417 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2418 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2420 if (!tags->static_rqs) {
2422 blk_mq_free_tags(tags, flags);
2429 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2430 unsigned int hctx_idx, int node)
2434 if (set->ops->init_request) {
2435 ret = set->ops->init_request(set, rq, hctx_idx, node);
2440 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2444 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2445 unsigned int hctx_idx, unsigned int depth)
2447 unsigned int i, j, entries_per_page, max_order = 4;
2448 size_t rq_size, left;
2451 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2452 if (node == NUMA_NO_NODE)
2453 node = set->numa_node;
2455 INIT_LIST_HEAD(&tags->page_list);
2458 * rq_size is the size of the request plus driver payload, rounded
2459 * to the cacheline size
2461 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2463 left = rq_size * depth;
2465 for (i = 0; i < depth; ) {
2466 int this_order = max_order;
2471 while (this_order && left < order_to_size(this_order - 1))
2475 page = alloc_pages_node(node,
2476 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2482 if (order_to_size(this_order) < rq_size)
2489 page->private = this_order;
2490 list_add_tail(&page->lru, &tags->page_list);
2492 p = page_address(page);
2494 * Allow kmemleak to scan these pages as they contain pointers
2495 * to additional allocations like via ops->init_request().
2497 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2498 entries_per_page = order_to_size(this_order) / rq_size;
2499 to_do = min(entries_per_page, depth - i);
2500 left -= to_do * rq_size;
2501 for (j = 0; j < to_do; j++) {
2502 struct request *rq = p;
2504 tags->static_rqs[i] = rq;
2505 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2506 tags->static_rqs[i] = NULL;
2517 blk_mq_free_rqs(set, tags, hctx_idx);
2521 struct rq_iter_data {
2522 struct blk_mq_hw_ctx *hctx;
2526 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2528 struct rq_iter_data *iter_data = data;
2530 if (rq->mq_hctx != iter_data->hctx)
2532 iter_data->has_rq = true;
2536 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2538 struct blk_mq_tags *tags = hctx->sched_tags ?
2539 hctx->sched_tags : hctx->tags;
2540 struct rq_iter_data data = {
2544 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2548 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2549 struct blk_mq_hw_ctx *hctx)
2551 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2553 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2558 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2560 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2561 struct blk_mq_hw_ctx, cpuhp_online);
2563 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2564 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2568 * Prevent new request from being allocated on the current hctx.
2570 * The smp_mb__after_atomic() Pairs with the implied barrier in
2571 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2572 * seen once we return from the tag allocator.
2574 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2575 smp_mb__after_atomic();
2578 * Try to grab a reference to the queue and wait for any outstanding
2579 * requests. If we could not grab a reference the queue has been
2580 * frozen and there are no requests.
2582 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2583 while (blk_mq_hctx_has_requests(hctx))
2585 percpu_ref_put(&hctx->queue->q_usage_counter);
2591 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2593 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2594 struct blk_mq_hw_ctx, cpuhp_online);
2596 if (cpumask_test_cpu(cpu, hctx->cpumask))
2597 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2602 * 'cpu' is going away. splice any existing rq_list entries from this
2603 * software queue to the hw queue dispatch list, and ensure that it
2606 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2608 struct blk_mq_hw_ctx *hctx;
2609 struct blk_mq_ctx *ctx;
2611 enum hctx_type type;
2613 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2614 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2617 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2620 spin_lock(&ctx->lock);
2621 if (!list_empty(&ctx->rq_lists[type])) {
2622 list_splice_init(&ctx->rq_lists[type], &tmp);
2623 blk_mq_hctx_clear_pending(hctx, ctx);
2625 spin_unlock(&ctx->lock);
2627 if (list_empty(&tmp))
2630 spin_lock(&hctx->lock);
2631 list_splice_tail_init(&tmp, &hctx->dispatch);
2632 spin_unlock(&hctx->lock);
2634 blk_mq_run_hw_queue(hctx, true);
2638 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2640 if (!(hctx->flags & BLK_MQ_F_STACKING))
2641 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2642 &hctx->cpuhp_online);
2643 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2648 * Before freeing hw queue, clearing the flush request reference in
2649 * tags->rqs[] for avoiding potential UAF.
2651 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
2652 unsigned int queue_depth, struct request *flush_rq)
2655 unsigned long flags;
2657 /* The hw queue may not be mapped yet */
2661 WARN_ON_ONCE(refcount_read(&flush_rq->ref) != 0);
2663 for (i = 0; i < queue_depth; i++)
2664 cmpxchg(&tags->rqs[i], flush_rq, NULL);
2667 * Wait until all pending iteration is done.
2669 * Request reference is cleared and it is guaranteed to be observed
2670 * after the ->lock is released.
2672 spin_lock_irqsave(&tags->lock, flags);
2673 spin_unlock_irqrestore(&tags->lock, flags);
2676 /* hctx->ctxs will be freed in queue's release handler */
2677 static void blk_mq_exit_hctx(struct request_queue *q,
2678 struct blk_mq_tag_set *set,
2679 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2681 struct request *flush_rq = hctx->fq->flush_rq;
2683 if (blk_mq_hw_queue_mapped(hctx))
2684 blk_mq_tag_idle(hctx);
2686 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
2687 set->queue_depth, flush_rq);
2688 if (set->ops->exit_request)
2689 set->ops->exit_request(set, flush_rq, hctx_idx);
2691 if (set->ops->exit_hctx)
2692 set->ops->exit_hctx(hctx, hctx_idx);
2694 blk_mq_remove_cpuhp(hctx);
2696 spin_lock(&q->unused_hctx_lock);
2697 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2698 spin_unlock(&q->unused_hctx_lock);
2701 static void blk_mq_exit_hw_queues(struct request_queue *q,
2702 struct blk_mq_tag_set *set, int nr_queue)
2704 struct blk_mq_hw_ctx *hctx;
2707 queue_for_each_hw_ctx(q, hctx, i) {
2710 blk_mq_debugfs_unregister_hctx(hctx);
2711 blk_mq_exit_hctx(q, set, hctx, i);
2715 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2717 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2719 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2720 __alignof__(struct blk_mq_hw_ctx)) !=
2721 sizeof(struct blk_mq_hw_ctx));
2723 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2724 hw_ctx_size += sizeof(struct srcu_struct);
2729 static int blk_mq_init_hctx(struct request_queue *q,
2730 struct blk_mq_tag_set *set,
2731 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2733 hctx->queue_num = hctx_idx;
2735 if (!(hctx->flags & BLK_MQ_F_STACKING))
2736 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2737 &hctx->cpuhp_online);
2738 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2740 hctx->tags = set->tags[hctx_idx];
2742 if (set->ops->init_hctx &&
2743 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2744 goto unregister_cpu_notifier;
2746 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2752 if (set->ops->exit_hctx)
2753 set->ops->exit_hctx(hctx, hctx_idx);
2754 unregister_cpu_notifier:
2755 blk_mq_remove_cpuhp(hctx);
2759 static struct blk_mq_hw_ctx *
2760 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2763 struct blk_mq_hw_ctx *hctx;
2764 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2766 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2768 goto fail_alloc_hctx;
2770 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2773 atomic_set(&hctx->nr_active, 0);
2774 if (node == NUMA_NO_NODE)
2775 node = set->numa_node;
2776 hctx->numa_node = node;
2778 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2779 spin_lock_init(&hctx->lock);
2780 INIT_LIST_HEAD(&hctx->dispatch);
2782 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
2784 INIT_LIST_HEAD(&hctx->hctx_list);
2787 * Allocate space for all possible cpus to avoid allocation at
2790 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2795 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2796 gfp, node, false, false))
2800 spin_lock_init(&hctx->dispatch_wait_lock);
2801 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2802 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2804 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2808 if (hctx->flags & BLK_MQ_F_BLOCKING)
2809 init_srcu_struct(hctx->srcu);
2810 blk_mq_hctx_kobj_init(hctx);
2815 sbitmap_free(&hctx->ctx_map);
2819 free_cpumask_var(hctx->cpumask);
2826 static void blk_mq_init_cpu_queues(struct request_queue *q,
2827 unsigned int nr_hw_queues)
2829 struct blk_mq_tag_set *set = q->tag_set;
2832 for_each_possible_cpu(i) {
2833 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2834 struct blk_mq_hw_ctx *hctx;
2838 spin_lock_init(&__ctx->lock);
2839 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2840 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2845 * Set local node, IFF we have more than one hw queue. If
2846 * not, we remain on the home node of the device
2848 for (j = 0; j < set->nr_maps; j++) {
2849 hctx = blk_mq_map_queue_type(q, j, i);
2850 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2851 hctx->numa_node = cpu_to_node(i);
2856 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2859 unsigned int flags = set->flags;
2862 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2863 set->queue_depth, set->reserved_tags, flags);
2864 if (!set->tags[hctx_idx])
2867 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2872 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2873 set->tags[hctx_idx] = NULL;
2877 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2878 unsigned int hctx_idx)
2880 unsigned int flags = set->flags;
2882 if (set->tags && set->tags[hctx_idx]) {
2883 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2884 blk_mq_free_rq_map(set->tags[hctx_idx], flags);
2885 set->tags[hctx_idx] = NULL;
2889 static void blk_mq_map_swqueue(struct request_queue *q)
2891 unsigned int i, j, hctx_idx;
2892 struct blk_mq_hw_ctx *hctx;
2893 struct blk_mq_ctx *ctx;
2894 struct blk_mq_tag_set *set = q->tag_set;
2896 queue_for_each_hw_ctx(q, hctx, i) {
2897 cpumask_clear(hctx->cpumask);
2899 hctx->dispatch_from = NULL;
2903 * Map software to hardware queues.
2905 * If the cpu isn't present, the cpu is mapped to first hctx.
2907 for_each_possible_cpu(i) {
2909 ctx = per_cpu_ptr(q->queue_ctx, i);
2910 for (j = 0; j < set->nr_maps; j++) {
2911 if (!set->map[j].nr_queues) {
2912 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2913 HCTX_TYPE_DEFAULT, i);
2916 hctx_idx = set->map[j].mq_map[i];
2917 /* unmapped hw queue can be remapped after CPU topo changed */
2918 if (!set->tags[hctx_idx] &&
2919 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2921 * If tags initialization fail for some hctx,
2922 * that hctx won't be brought online. In this
2923 * case, remap the current ctx to hctx[0] which
2924 * is guaranteed to always have tags allocated
2926 set->map[j].mq_map[i] = 0;
2929 hctx = blk_mq_map_queue_type(q, j, i);
2930 ctx->hctxs[j] = hctx;
2932 * If the CPU is already set in the mask, then we've
2933 * mapped this one already. This can happen if
2934 * devices share queues across queue maps.
2936 if (cpumask_test_cpu(i, hctx->cpumask))
2939 cpumask_set_cpu(i, hctx->cpumask);
2941 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2942 hctx->ctxs[hctx->nr_ctx++] = ctx;
2945 * If the nr_ctx type overflows, we have exceeded the
2946 * amount of sw queues we can support.
2948 BUG_ON(!hctx->nr_ctx);
2951 for (; j < HCTX_MAX_TYPES; j++)
2952 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2953 HCTX_TYPE_DEFAULT, i);
2956 queue_for_each_hw_ctx(q, hctx, i) {
2958 * If no software queues are mapped to this hardware queue,
2959 * disable it and free the request entries.
2961 if (!hctx->nr_ctx) {
2962 /* Never unmap queue 0. We need it as a
2963 * fallback in case of a new remap fails
2966 if (i && set->tags[i])
2967 blk_mq_free_map_and_requests(set, i);
2973 hctx->tags = set->tags[i];
2974 WARN_ON(!hctx->tags);
2977 * Set the map size to the number of mapped software queues.
2978 * This is more accurate and more efficient than looping
2979 * over all possibly mapped software queues.
2981 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2984 * Initialize batch roundrobin counts
2986 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2987 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2992 * Caller needs to ensure that we're either frozen/quiesced, or that
2993 * the queue isn't live yet.
2995 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2997 struct blk_mq_hw_ctx *hctx;
3000 queue_for_each_hw_ctx(q, hctx, i) {
3002 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3004 blk_mq_tag_idle(hctx);
3005 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3010 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3013 struct request_queue *q;
3015 lockdep_assert_held(&set->tag_list_lock);
3017 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3018 blk_mq_freeze_queue(q);
3019 queue_set_hctx_shared(q, shared);
3020 blk_mq_unfreeze_queue(q);
3024 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3026 struct blk_mq_tag_set *set = q->tag_set;
3028 mutex_lock(&set->tag_list_lock);
3029 list_del(&q->tag_set_list);
3030 if (list_is_singular(&set->tag_list)) {
3031 /* just transitioned to unshared */
3032 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3033 /* update existing queue */
3034 blk_mq_update_tag_set_shared(set, false);
3036 mutex_unlock(&set->tag_list_lock);
3037 INIT_LIST_HEAD(&q->tag_set_list);
3040 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3041 struct request_queue *q)
3043 mutex_lock(&set->tag_list_lock);
3046 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3048 if (!list_empty(&set->tag_list) &&
3049 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3050 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3051 /* update existing queue */
3052 blk_mq_update_tag_set_shared(set, true);
3054 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3055 queue_set_hctx_shared(q, true);
3056 list_add_tail(&q->tag_set_list, &set->tag_list);
3058 mutex_unlock(&set->tag_list_lock);
3061 /* All allocations will be freed in release handler of q->mq_kobj */
3062 static int blk_mq_alloc_ctxs(struct request_queue *q)
3064 struct blk_mq_ctxs *ctxs;
3067 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3071 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3072 if (!ctxs->queue_ctx)
3075 for_each_possible_cpu(cpu) {
3076 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3080 q->mq_kobj = &ctxs->kobj;
3081 q->queue_ctx = ctxs->queue_ctx;
3090 * It is the actual release handler for mq, but we do it from
3091 * request queue's release handler for avoiding use-after-free
3092 * and headache because q->mq_kobj shouldn't have been introduced,
3093 * but we can't group ctx/kctx kobj without it.
3095 void blk_mq_release(struct request_queue *q)
3097 struct blk_mq_hw_ctx *hctx, *next;
3100 queue_for_each_hw_ctx(q, hctx, i)
3101 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3103 /* all hctx are in .unused_hctx_list now */
3104 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3105 list_del_init(&hctx->hctx_list);
3106 kobject_put(&hctx->kobj);
3109 kfree(q->queue_hw_ctx);
3112 * release .mq_kobj and sw queue's kobject now because
3113 * both share lifetime with request queue.
3115 blk_mq_sysfs_deinit(q);
3118 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3121 struct request_queue *q;
3124 q = blk_alloc_queue(set->numa_node);
3126 return ERR_PTR(-ENOMEM);
3127 q->queuedata = queuedata;
3128 ret = blk_mq_init_allocated_queue(set, q);
3130 blk_cleanup_queue(q);
3131 return ERR_PTR(ret);
3136 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3138 return blk_mq_init_queue_data(set, NULL);
3140 EXPORT_SYMBOL(blk_mq_init_queue);
3142 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
3143 struct lock_class_key *lkclass)
3145 struct request_queue *q;
3146 struct gendisk *disk;
3148 q = blk_mq_init_queue_data(set, queuedata);
3152 disk = __alloc_disk_node(q, set->numa_node, lkclass);
3154 blk_cleanup_queue(q);
3155 return ERR_PTR(-ENOMEM);
3159 EXPORT_SYMBOL(__blk_mq_alloc_disk);
3161 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3162 struct blk_mq_tag_set *set, struct request_queue *q,
3163 int hctx_idx, int node)
3165 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3167 /* reuse dead hctx first */
3168 spin_lock(&q->unused_hctx_lock);
3169 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3170 if (tmp->numa_node == node) {
3176 list_del_init(&hctx->hctx_list);
3177 spin_unlock(&q->unused_hctx_lock);
3180 hctx = blk_mq_alloc_hctx(q, set, node);
3184 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3190 kobject_put(&hctx->kobj);
3195 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3196 struct request_queue *q)
3199 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3201 if (q->nr_hw_queues < set->nr_hw_queues) {
3202 struct blk_mq_hw_ctx **new_hctxs;
3204 new_hctxs = kcalloc_node(set->nr_hw_queues,
3205 sizeof(*new_hctxs), GFP_KERNEL,
3210 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3212 q->queue_hw_ctx = new_hctxs;
3217 /* protect against switching io scheduler */
3218 mutex_lock(&q->sysfs_lock);
3219 for (i = 0; i < set->nr_hw_queues; i++) {
3221 struct blk_mq_hw_ctx *hctx;
3223 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3225 * If the hw queue has been mapped to another numa node,
3226 * we need to realloc the hctx. If allocation fails, fallback
3227 * to use the previous one.
3229 if (hctxs[i] && (hctxs[i]->numa_node == node))
3232 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3235 blk_mq_exit_hctx(q, set, hctxs[i], i);
3239 pr_warn("Allocate new hctx on node %d fails,\
3240 fallback to previous one on node %d\n",
3241 node, hctxs[i]->numa_node);
3247 * Increasing nr_hw_queues fails. Free the newly allocated
3248 * hctxs and keep the previous q->nr_hw_queues.
3250 if (i != set->nr_hw_queues) {
3251 j = q->nr_hw_queues;
3255 end = q->nr_hw_queues;
3256 q->nr_hw_queues = set->nr_hw_queues;
3259 for (; j < end; j++) {
3260 struct blk_mq_hw_ctx *hctx = hctxs[j];
3264 blk_mq_free_map_and_requests(set, j);
3265 blk_mq_exit_hctx(q, set, hctx, j);
3269 mutex_unlock(&q->sysfs_lock);
3272 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3273 struct request_queue *q)
3275 /* mark the queue as mq asap */
3276 q->mq_ops = set->ops;
3278 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3279 blk_mq_poll_stats_bkt,
3280 BLK_MQ_POLL_STATS_BKTS, q);
3284 if (blk_mq_alloc_ctxs(q))
3287 /* init q->mq_kobj and sw queues' kobjects */
3288 blk_mq_sysfs_init(q);
3290 INIT_LIST_HEAD(&q->unused_hctx_list);
3291 spin_lock_init(&q->unused_hctx_lock);
3293 blk_mq_realloc_hw_ctxs(set, q);
3294 if (!q->nr_hw_queues)
3297 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3298 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3302 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3303 if (set->nr_maps > HCTX_TYPE_POLL &&
3304 set->map[HCTX_TYPE_POLL].nr_queues)
3305 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3307 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3308 INIT_LIST_HEAD(&q->requeue_list);
3309 spin_lock_init(&q->requeue_lock);
3311 q->nr_requests = set->queue_depth;
3314 * Default to classic polling
3316 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3318 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3319 blk_mq_add_queue_tag_set(set, q);
3320 blk_mq_map_swqueue(q);
3324 kfree(q->queue_hw_ctx);
3325 q->nr_hw_queues = 0;
3326 blk_mq_sysfs_deinit(q);
3328 blk_stat_free_callback(q->poll_cb);
3334 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3336 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3337 void blk_mq_exit_queue(struct request_queue *q)
3339 struct blk_mq_tag_set *set = q->tag_set;
3341 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
3342 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3343 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
3344 blk_mq_del_queue_tag_set(q);
3347 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3351 for (i = 0; i < set->nr_hw_queues; i++) {
3352 if (!__blk_mq_alloc_map_and_request(set, i))
3361 blk_mq_free_map_and_requests(set, i);
3367 * Allocate the request maps associated with this tag_set. Note that this
3368 * may reduce the depth asked for, if memory is tight. set->queue_depth
3369 * will be updated to reflect the allocated depth.
3371 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3376 depth = set->queue_depth;
3378 err = __blk_mq_alloc_rq_maps(set);
3382 set->queue_depth >>= 1;
3383 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3387 } while (set->queue_depth);
3389 if (!set->queue_depth || err) {
3390 pr_err("blk-mq: failed to allocate request map\n");
3394 if (depth != set->queue_depth)
3395 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3396 depth, set->queue_depth);
3401 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3404 * blk_mq_map_queues() and multiple .map_queues() implementations
3405 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3406 * number of hardware queues.
3408 if (set->nr_maps == 1)
3409 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3411 if (set->ops->map_queues && !is_kdump_kernel()) {
3415 * transport .map_queues is usually done in the following
3418 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3419 * mask = get_cpu_mask(queue)
3420 * for_each_cpu(cpu, mask)
3421 * set->map[x].mq_map[cpu] = queue;
3424 * When we need to remap, the table has to be cleared for
3425 * killing stale mapping since one CPU may not be mapped
3428 for (i = 0; i < set->nr_maps; i++)
3429 blk_mq_clear_mq_map(&set->map[i]);
3431 return set->ops->map_queues(set);
3433 BUG_ON(set->nr_maps > 1);
3434 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3438 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3439 int cur_nr_hw_queues, int new_nr_hw_queues)
3441 struct blk_mq_tags **new_tags;
3443 if (cur_nr_hw_queues >= new_nr_hw_queues)
3446 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3447 GFP_KERNEL, set->numa_node);
3452 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3453 sizeof(*set->tags));
3455 set->tags = new_tags;
3456 set->nr_hw_queues = new_nr_hw_queues;
3461 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
3462 int new_nr_hw_queues)
3464 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
3468 * Alloc a tag set to be associated with one or more request queues.
3469 * May fail with EINVAL for various error conditions. May adjust the
3470 * requested depth down, if it's too large. In that case, the set
3471 * value will be stored in set->queue_depth.
3473 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3477 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3479 if (!set->nr_hw_queues)
3481 if (!set->queue_depth)
3483 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3486 if (!set->ops->queue_rq)
3489 if (!set->ops->get_budget ^ !set->ops->put_budget)
3492 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3493 pr_info("blk-mq: reduced tag depth to %u\n",
3495 set->queue_depth = BLK_MQ_MAX_DEPTH;
3500 else if (set->nr_maps > HCTX_MAX_TYPES)
3504 * If a crashdump is active, then we are potentially in a very
3505 * memory constrained environment. Limit us to 1 queue and
3506 * 64 tags to prevent using too much memory.
3508 if (is_kdump_kernel()) {
3509 set->nr_hw_queues = 1;
3511 set->queue_depth = min(64U, set->queue_depth);
3514 * There is no use for more h/w queues than cpus if we just have
3517 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3518 set->nr_hw_queues = nr_cpu_ids;
3520 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
3524 for (i = 0; i < set->nr_maps; i++) {
3525 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3526 sizeof(set->map[i].mq_map[0]),
3527 GFP_KERNEL, set->numa_node);
3528 if (!set->map[i].mq_map)
3529 goto out_free_mq_map;
3530 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3533 ret = blk_mq_update_queue_map(set);
3535 goto out_free_mq_map;
3537 ret = blk_mq_alloc_map_and_requests(set);
3539 goto out_free_mq_map;
3541 if (blk_mq_is_sbitmap_shared(set->flags)) {
3542 atomic_set(&set->active_queues_shared_sbitmap, 0);
3544 if (blk_mq_init_shared_sbitmap(set)) {
3546 goto out_free_mq_rq_maps;
3550 mutex_init(&set->tag_list_lock);
3551 INIT_LIST_HEAD(&set->tag_list);
3555 out_free_mq_rq_maps:
3556 for (i = 0; i < set->nr_hw_queues; i++)
3557 blk_mq_free_map_and_requests(set, i);
3559 for (i = 0; i < set->nr_maps; i++) {
3560 kfree(set->map[i].mq_map);
3561 set->map[i].mq_map = NULL;
3567 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3569 /* allocate and initialize a tagset for a simple single-queue device */
3570 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
3571 const struct blk_mq_ops *ops, unsigned int queue_depth,
3572 unsigned int set_flags)
3574 memset(set, 0, sizeof(*set));
3576 set->nr_hw_queues = 1;
3578 set->queue_depth = queue_depth;
3579 set->numa_node = NUMA_NO_NODE;
3580 set->flags = set_flags;
3581 return blk_mq_alloc_tag_set(set);
3583 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
3585 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3589 for (i = 0; i < set->nr_hw_queues; i++)
3590 blk_mq_free_map_and_requests(set, i);
3592 if (blk_mq_is_sbitmap_shared(set->flags))
3593 blk_mq_exit_shared_sbitmap(set);
3595 for (j = 0; j < set->nr_maps; j++) {
3596 kfree(set->map[j].mq_map);
3597 set->map[j].mq_map = NULL;
3603 EXPORT_SYMBOL(blk_mq_free_tag_set);
3605 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3607 struct blk_mq_tag_set *set = q->tag_set;
3608 struct blk_mq_hw_ctx *hctx;
3614 if (q->nr_requests == nr)
3617 blk_mq_freeze_queue(q);
3618 blk_mq_quiesce_queue(q);
3621 queue_for_each_hw_ctx(q, hctx, i) {
3625 * If we're using an MQ scheduler, just update the scheduler
3626 * queue depth. This is similar to what the old code would do.
3628 if (!hctx->sched_tags) {
3629 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3631 if (!ret && blk_mq_is_sbitmap_shared(set->flags))
3632 blk_mq_tag_resize_shared_sbitmap(set, nr);
3634 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3636 if (blk_mq_is_sbitmap_shared(set->flags)) {
3637 hctx->sched_tags->bitmap_tags =
3638 &q->sched_bitmap_tags;
3639 hctx->sched_tags->breserved_tags =
3640 &q->sched_breserved_tags;
3645 if (q->elevator && q->elevator->type->ops.depth_updated)
3646 q->elevator->type->ops.depth_updated(hctx);
3649 q->nr_requests = nr;
3650 if (q->elevator && blk_mq_is_sbitmap_shared(set->flags))
3651 sbitmap_queue_resize(&q->sched_bitmap_tags,
3652 nr - set->reserved_tags);
3655 blk_mq_unquiesce_queue(q);
3656 blk_mq_unfreeze_queue(q);
3662 * request_queue and elevator_type pair.
3663 * It is just used by __blk_mq_update_nr_hw_queues to cache
3664 * the elevator_type associated with a request_queue.
3666 struct blk_mq_qe_pair {
3667 struct list_head node;
3668 struct request_queue *q;
3669 struct elevator_type *type;
3673 * Cache the elevator_type in qe pair list and switch the
3674 * io scheduler to 'none'
3676 static bool blk_mq_elv_switch_none(struct list_head *head,
3677 struct request_queue *q)
3679 struct blk_mq_qe_pair *qe;
3684 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3688 INIT_LIST_HEAD(&qe->node);
3690 qe->type = q->elevator->type;
3691 list_add(&qe->node, head);
3693 mutex_lock(&q->sysfs_lock);
3695 * After elevator_switch_mq, the previous elevator_queue will be
3696 * released by elevator_release. The reference of the io scheduler
3697 * module get by elevator_get will also be put. So we need to get
3698 * a reference of the io scheduler module here to prevent it to be
3701 __module_get(qe->type->elevator_owner);
3702 elevator_switch_mq(q, NULL);
3703 mutex_unlock(&q->sysfs_lock);
3708 static void blk_mq_elv_switch_back(struct list_head *head,
3709 struct request_queue *q)
3711 struct blk_mq_qe_pair *qe;
3712 struct elevator_type *t = NULL;
3714 list_for_each_entry(qe, head, node)
3723 list_del(&qe->node);
3726 mutex_lock(&q->sysfs_lock);
3727 elevator_switch_mq(q, t);
3728 mutex_unlock(&q->sysfs_lock);
3731 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3734 struct request_queue *q;
3736 int prev_nr_hw_queues;
3738 lockdep_assert_held(&set->tag_list_lock);
3740 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3741 nr_hw_queues = nr_cpu_ids;
3742 if (nr_hw_queues < 1)
3744 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3747 list_for_each_entry(q, &set->tag_list, tag_set_list)
3748 blk_mq_freeze_queue(q);
3750 * Switch IO scheduler to 'none', cleaning up the data associated
3751 * with the previous scheduler. We will switch back once we are done
3752 * updating the new sw to hw queue mappings.
3754 list_for_each_entry(q, &set->tag_list, tag_set_list)
3755 if (!blk_mq_elv_switch_none(&head, q))
3758 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3759 blk_mq_debugfs_unregister_hctxs(q);
3760 blk_mq_sysfs_unregister(q);
3763 prev_nr_hw_queues = set->nr_hw_queues;
3764 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3768 set->nr_hw_queues = nr_hw_queues;
3770 blk_mq_update_queue_map(set);
3771 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3772 blk_mq_realloc_hw_ctxs(set, q);
3773 if (q->nr_hw_queues != set->nr_hw_queues) {
3774 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3775 nr_hw_queues, prev_nr_hw_queues);
3776 set->nr_hw_queues = prev_nr_hw_queues;
3777 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3780 blk_mq_map_swqueue(q);
3784 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3785 blk_mq_sysfs_register(q);
3786 blk_mq_debugfs_register_hctxs(q);
3790 list_for_each_entry(q, &set->tag_list, tag_set_list)
3791 blk_mq_elv_switch_back(&head, q);
3793 list_for_each_entry(q, &set->tag_list, tag_set_list)
3794 blk_mq_unfreeze_queue(q);
3797 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3799 mutex_lock(&set->tag_list_lock);
3800 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3801 mutex_unlock(&set->tag_list_lock);
3803 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3805 /* Enable polling stats and return whether they were already enabled. */
3806 static bool blk_poll_stats_enable(struct request_queue *q)
3808 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3809 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3811 blk_stat_add_callback(q, q->poll_cb);
3815 static void blk_mq_poll_stats_start(struct request_queue *q)
3818 * We don't arm the callback if polling stats are not enabled or the
3819 * callback is already active.
3821 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3822 blk_stat_is_active(q->poll_cb))
3825 blk_stat_activate_msecs(q->poll_cb, 100);
3828 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3830 struct request_queue *q = cb->data;
3833 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3834 if (cb->stat[bucket].nr_samples)
3835 q->poll_stat[bucket] = cb->stat[bucket];
3839 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3842 unsigned long ret = 0;
3846 * If stats collection isn't on, don't sleep but turn it on for
3849 if (!blk_poll_stats_enable(q))
3853 * As an optimistic guess, use half of the mean service time
3854 * for this type of request. We can (and should) make this smarter.
3855 * For instance, if the completion latencies are tight, we can
3856 * get closer than just half the mean. This is especially
3857 * important on devices where the completion latencies are longer
3858 * than ~10 usec. We do use the stats for the relevant IO size
3859 * if available which does lead to better estimates.
3861 bucket = blk_mq_poll_stats_bkt(rq);
3865 if (q->poll_stat[bucket].nr_samples)
3866 ret = (q->poll_stat[bucket].mean + 1) / 2;
3871 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3874 struct hrtimer_sleeper hs;
3875 enum hrtimer_mode mode;
3879 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3883 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3885 * 0: use half of prev avg
3886 * >0: use this specific value
3888 if (q->poll_nsec > 0)
3889 nsecs = q->poll_nsec;
3891 nsecs = blk_mq_poll_nsecs(q, rq);
3896 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3899 * This will be replaced with the stats tracking code, using
3900 * 'avg_completion_time / 2' as the pre-sleep target.
3904 mode = HRTIMER_MODE_REL;
3905 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3906 hrtimer_set_expires(&hs.timer, kt);
3909 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3911 set_current_state(TASK_UNINTERRUPTIBLE);
3912 hrtimer_sleeper_start_expires(&hs, mode);
3915 hrtimer_cancel(&hs.timer);
3916 mode = HRTIMER_MODE_ABS;
3917 } while (hs.task && !signal_pending(current));
3919 __set_current_state(TASK_RUNNING);
3920 destroy_hrtimer_on_stack(&hs.timer);
3924 static bool blk_mq_poll_hybrid(struct request_queue *q,
3925 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3929 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3932 if (!blk_qc_t_is_internal(cookie))
3933 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3935 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3937 * With scheduling, if the request has completed, we'll
3938 * get a NULL return here, as we clear the sched tag when
3939 * that happens. The request still remains valid, like always,
3940 * so we should be safe with just the NULL check.
3946 return blk_mq_poll_hybrid_sleep(q, rq);
3950 * blk_poll - poll for IO completions
3952 * @cookie: cookie passed back at IO submission time
3953 * @spin: whether to spin for completions
3956 * Poll for completions on the passed in queue. Returns number of
3957 * completed entries found. If @spin is true, then blk_poll will continue
3958 * looping until at least one completion is found, unless the task is
3959 * otherwise marked running (or we need to reschedule).
3961 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3963 struct blk_mq_hw_ctx *hctx;
3966 if (!blk_qc_t_valid(cookie) ||
3967 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3971 blk_flush_plug_list(current->plug, false);
3973 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3976 * If we sleep, have the caller restart the poll loop to reset
3977 * the state. Like for the other success return cases, the
3978 * caller is responsible for checking if the IO completed. If
3979 * the IO isn't complete, we'll get called again and will go
3980 * straight to the busy poll loop. If specified not to spin,
3981 * we also should not sleep.
3983 if (spin && blk_mq_poll_hybrid(q, hctx, cookie))
3986 hctx->poll_considered++;
3988 state = get_current_state();
3992 hctx->poll_invoked++;
3994 ret = q->mq_ops->poll(hctx);
3996 hctx->poll_success++;
3997 __set_current_state(TASK_RUNNING);
4001 if (signal_pending_state(state, current))
4002 __set_current_state(TASK_RUNNING);
4004 if (task_is_running(current))
4006 if (ret < 0 || !spin)
4009 } while (!need_resched());
4011 __set_current_state(TASK_RUNNING);
4014 EXPORT_SYMBOL_GPL(blk_poll);
4016 unsigned int blk_mq_rq_cpu(struct request *rq)
4018 return rq->mq_ctx->cpu;
4020 EXPORT_SYMBOL(blk_mq_rq_cpu);
4022 static int __init blk_mq_init(void)
4026 for_each_possible_cpu(i)
4027 init_llist_head(&per_cpu(blk_cpu_done, i));
4028 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4030 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4031 "block/softirq:dead", NULL,
4032 blk_softirq_cpu_dead);
4033 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4034 blk_mq_hctx_notify_dead);
4035 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4036 blk_mq_hctx_notify_online,
4037 blk_mq_hctx_notify_offline);
4040 subsys_initcall(blk_mq_init);