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 void blk_mq_poll_stats_start(struct request_queue *q);
45 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
47 static int blk_mq_poll_stats_bkt(const struct request *rq)
49 int ddir, sectors, bucket;
51 ddir = rq_data_dir(rq);
52 sectors = blk_rq_stats_sectors(rq);
54 bucket = ddir + 2 * ilog2(sectors);
58 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
59 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
65 * Check if any of the ctx, dispatch list or elevator
66 * have pending work in this hardware queue.
68 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
70 return !list_empty_careful(&hctx->dispatch) ||
71 sbitmap_any_bit_set(&hctx->ctx_map) ||
72 blk_mq_sched_has_work(hctx);
76 * Mark this ctx as having pending work in this hardware queue
78 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
79 struct blk_mq_ctx *ctx)
81 const int bit = ctx->index_hw[hctx->type];
83 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
84 sbitmap_set_bit(&hctx->ctx_map, bit);
87 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
88 struct blk_mq_ctx *ctx)
90 const int bit = ctx->index_hw[hctx->type];
92 sbitmap_clear_bit(&hctx->ctx_map, bit);
96 struct hd_struct *part;
97 unsigned int inflight[2];
100 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
101 struct request *rq, void *priv,
104 struct mq_inflight *mi = priv;
106 if (rq->part == mi->part)
107 mi->inflight[rq_data_dir(rq)]++;
112 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
114 struct mq_inflight mi = { .part = part };
116 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
118 return mi.inflight[0] + mi.inflight[1];
121 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
122 unsigned int inflight[2])
124 struct mq_inflight mi = { .part = part };
126 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
127 inflight[0] = mi.inflight[0];
128 inflight[1] = mi.inflight[1];
131 void blk_freeze_queue_start(struct request_queue *q)
133 mutex_lock(&q->mq_freeze_lock);
134 if (++q->mq_freeze_depth == 1) {
135 percpu_ref_kill(&q->q_usage_counter);
136 mutex_unlock(&q->mq_freeze_lock);
138 blk_mq_run_hw_queues(q, false);
140 mutex_unlock(&q->mq_freeze_lock);
143 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
145 void blk_mq_freeze_queue_wait(struct request_queue *q)
147 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
149 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
151 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
152 unsigned long timeout)
154 return wait_event_timeout(q->mq_freeze_wq,
155 percpu_ref_is_zero(&q->q_usage_counter),
158 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
161 * Guarantee no request is in use, so we can change any data structure of
162 * the queue afterward.
164 void blk_freeze_queue(struct request_queue *q)
167 * In the !blk_mq case we are only calling this to kill the
168 * q_usage_counter, otherwise this increases the freeze depth
169 * and waits for it to return to zero. For this reason there is
170 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
171 * exported to drivers as the only user for unfreeze is blk_mq.
173 blk_freeze_queue_start(q);
174 blk_mq_freeze_queue_wait(q);
177 void blk_mq_freeze_queue(struct request_queue *q)
180 * ...just an alias to keep freeze and unfreeze actions balanced
181 * in the blk_mq_* namespace
185 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
187 void blk_mq_unfreeze_queue(struct request_queue *q)
189 mutex_lock(&q->mq_freeze_lock);
190 q->mq_freeze_depth--;
191 WARN_ON_ONCE(q->mq_freeze_depth < 0);
192 if (!q->mq_freeze_depth) {
193 percpu_ref_resurrect(&q->q_usage_counter);
194 wake_up_all(&q->mq_freeze_wq);
196 mutex_unlock(&q->mq_freeze_lock);
198 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
201 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
202 * mpt3sas driver such that this function can be removed.
204 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
206 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
208 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
211 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
214 * Note: this function does not prevent that the struct request end_io()
215 * callback function is invoked. Once this function is returned, we make
216 * sure no dispatch can happen until the queue is unquiesced via
217 * blk_mq_unquiesce_queue().
219 void blk_mq_quiesce_queue(struct request_queue *q)
221 struct blk_mq_hw_ctx *hctx;
225 blk_mq_quiesce_queue_nowait(q);
227 queue_for_each_hw_ctx(q, hctx, i) {
228 if (hctx->flags & BLK_MQ_F_BLOCKING)
229 synchronize_srcu(hctx->srcu);
236 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
239 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
242 * This function recovers queue into the state before quiescing
243 * which is done by blk_mq_quiesce_queue.
245 void blk_mq_unquiesce_queue(struct request_queue *q)
247 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
249 /* dispatch requests which are inserted during quiescing */
250 blk_mq_run_hw_queues(q, true);
252 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
254 void blk_mq_wake_waiters(struct request_queue *q)
256 struct blk_mq_hw_ctx *hctx;
259 queue_for_each_hw_ctx(q, hctx, i)
260 if (blk_mq_hw_queue_mapped(hctx))
261 blk_mq_tag_wakeup_all(hctx->tags, true);
265 * Only need start/end time stamping if we have iostat or
266 * blk stats enabled, or using an IO scheduler.
268 static inline bool blk_mq_need_time_stamp(struct request *rq)
270 return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
273 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
274 unsigned int tag, unsigned int op, u64 alloc_time_ns)
276 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
277 struct request *rq = tags->static_rqs[tag];
278 req_flags_t rq_flags = 0;
280 if (data->flags & BLK_MQ_REQ_INTERNAL) {
282 rq->internal_tag = tag;
284 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
285 rq_flags = RQF_MQ_INFLIGHT;
286 atomic_inc(&data->hctx->nr_active);
289 rq->internal_tag = -1;
290 data->hctx->tags->rqs[rq->tag] = rq;
293 /* csd/requeue_work/fifo_time is initialized before use */
295 rq->mq_ctx = data->ctx;
296 rq->mq_hctx = data->hctx;
297 rq->rq_flags = rq_flags;
299 if (data->flags & BLK_MQ_REQ_PREEMPT)
300 rq->rq_flags |= RQF_PREEMPT;
301 if (blk_queue_io_stat(data->q))
302 rq->rq_flags |= RQF_IO_STAT;
303 INIT_LIST_HEAD(&rq->queuelist);
304 INIT_HLIST_NODE(&rq->hash);
305 RB_CLEAR_NODE(&rq->rb_node);
308 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
309 rq->alloc_time_ns = alloc_time_ns;
311 if (blk_mq_need_time_stamp(rq))
312 rq->start_time_ns = ktime_get_ns();
314 rq->start_time_ns = 0;
315 rq->io_start_time_ns = 0;
316 rq->stats_sectors = 0;
317 rq->nr_phys_segments = 0;
318 #if defined(CONFIG_BLK_DEV_INTEGRITY)
319 rq->nr_integrity_segments = 0;
321 blk_crypto_rq_set_defaults(rq);
322 /* tag was already set */
323 WRITE_ONCE(rq->deadline, 0);
328 rq->end_io_data = NULL;
330 data->ctx->rq_dispatched[op_is_sync(op)]++;
331 refcount_set(&rq->ref, 1);
335 static struct request *blk_mq_get_request(struct request_queue *q,
337 struct blk_mq_alloc_data *data)
339 struct elevator_queue *e = q->elevator;
342 bool clear_ctx_on_error = false;
343 u64 alloc_time_ns = 0;
345 blk_queue_enter_live(q);
347 /* alloc_time includes depth and tag waits */
348 if (blk_queue_rq_alloc_time(q))
349 alloc_time_ns = ktime_get_ns();
352 if (likely(!data->ctx)) {
353 data->ctx = blk_mq_get_ctx(q);
354 clear_ctx_on_error = true;
356 if (likely(!data->hctx))
357 data->hctx = blk_mq_map_queue(q, data->cmd_flags,
359 if (data->cmd_flags & REQ_NOWAIT)
360 data->flags |= BLK_MQ_REQ_NOWAIT;
363 data->flags |= BLK_MQ_REQ_INTERNAL;
366 * Flush requests are special and go directly to the
367 * dispatch list. Don't include reserved tags in the
368 * limiting, as it isn't useful.
370 if (!op_is_flush(data->cmd_flags) &&
371 e->type->ops.limit_depth &&
372 !(data->flags & BLK_MQ_REQ_RESERVED))
373 e->type->ops.limit_depth(data->cmd_flags, data);
375 blk_mq_tag_busy(data->hctx);
378 tag = blk_mq_get_tag(data);
379 if (tag == BLK_MQ_TAG_FAIL) {
380 if (clear_ctx_on_error)
386 rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags, alloc_time_ns);
387 if (!op_is_flush(data->cmd_flags)) {
389 if (e && e->type->ops.prepare_request) {
390 if (e->type->icq_cache)
391 blk_mq_sched_assign_ioc(rq);
393 e->type->ops.prepare_request(rq, bio);
394 rq->rq_flags |= RQF_ELVPRIV;
397 data->hctx->queued++;
401 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
402 blk_mq_req_flags_t flags)
404 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
408 ret = blk_queue_enter(q, flags);
412 rq = blk_mq_get_request(q, NULL, &alloc_data);
416 return ERR_PTR(-EWOULDBLOCK);
419 rq->__sector = (sector_t) -1;
420 rq->bio = rq->biotail = NULL;
423 EXPORT_SYMBOL(blk_mq_alloc_request);
425 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
426 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
428 struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
434 * If the tag allocator sleeps we could get an allocation for a
435 * different hardware context. No need to complicate the low level
436 * allocator for this for the rare use case of a command tied to
439 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
440 return ERR_PTR(-EINVAL);
442 if (hctx_idx >= q->nr_hw_queues)
443 return ERR_PTR(-EIO);
445 ret = blk_queue_enter(q, flags);
450 * Check if the hardware context is actually mapped to anything.
451 * If not tell the caller that it should skip this queue.
453 alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
454 if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
456 return ERR_PTR(-EXDEV);
458 cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
459 alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
461 rq = blk_mq_get_request(q, NULL, &alloc_data);
465 return ERR_PTR(-EWOULDBLOCK);
469 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
471 static void __blk_mq_free_request(struct request *rq)
473 struct request_queue *q = rq->q;
474 struct blk_mq_ctx *ctx = rq->mq_ctx;
475 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
476 const int sched_tag = rq->internal_tag;
478 blk_crypto_free_request(rq);
479 blk_pm_mark_last_busy(rq);
482 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
484 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
485 blk_mq_sched_restart(hctx);
489 void blk_mq_free_request(struct request *rq)
491 struct request_queue *q = rq->q;
492 struct elevator_queue *e = q->elevator;
493 struct blk_mq_ctx *ctx = rq->mq_ctx;
494 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
496 if (rq->rq_flags & RQF_ELVPRIV) {
497 if (e && e->type->ops.finish_request)
498 e->type->ops.finish_request(rq);
500 put_io_context(rq->elv.icq->ioc);
505 ctx->rq_completed[rq_is_sync(rq)]++;
506 if (rq->rq_flags & RQF_MQ_INFLIGHT)
507 atomic_dec(&hctx->nr_active);
509 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
510 laptop_io_completion(q->backing_dev_info);
514 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
515 if (refcount_dec_and_test(&rq->ref))
516 __blk_mq_free_request(rq);
518 EXPORT_SYMBOL_GPL(blk_mq_free_request);
520 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
524 if (blk_mq_need_time_stamp(rq))
525 now = ktime_get_ns();
527 if (rq->rq_flags & RQF_STATS) {
528 blk_mq_poll_stats_start(rq->q);
529 blk_stat_add(rq, now);
532 if (rq->internal_tag != -1)
533 blk_mq_sched_completed_request(rq, now);
535 blk_account_io_done(rq, now);
538 rq_qos_done(rq->q, rq);
539 rq->end_io(rq, error);
541 blk_mq_free_request(rq);
544 EXPORT_SYMBOL(__blk_mq_end_request);
546 void blk_mq_end_request(struct request *rq, blk_status_t error)
548 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
550 __blk_mq_end_request(rq, error);
552 EXPORT_SYMBOL(blk_mq_end_request);
554 static void __blk_mq_complete_request_remote(void *data)
556 struct request *rq = data;
557 struct request_queue *q = rq->q;
559 q->mq_ops->complete(rq);
562 static void __blk_mq_complete_request(struct request *rq)
564 struct blk_mq_ctx *ctx = rq->mq_ctx;
565 struct request_queue *q = rq->q;
569 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
571 * Most of single queue controllers, there is only one irq vector
572 * for handling IO completion, and the only irq's affinity is set
573 * as all possible CPUs. On most of ARCHs, this affinity means the
574 * irq is handled on one specific CPU.
576 * So complete IO reqeust in softirq context in case of single queue
577 * for not degrading IO performance by irqsoff latency.
579 if (q->nr_hw_queues == 1) {
580 __blk_complete_request(rq);
585 * For a polled request, always complete locallly, it's pointless
586 * to redirect the completion.
588 if ((rq->cmd_flags & REQ_HIPRI) ||
589 !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
590 q->mq_ops->complete(rq);
595 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
596 shared = cpus_share_cache(cpu, ctx->cpu);
598 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
599 rq->csd.func = __blk_mq_complete_request_remote;
602 smp_call_function_single_async(ctx->cpu, &rq->csd);
604 q->mq_ops->complete(rq);
609 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
610 __releases(hctx->srcu)
612 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
615 srcu_read_unlock(hctx->srcu, srcu_idx);
618 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
619 __acquires(hctx->srcu)
621 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
622 /* shut up gcc false positive */
626 *srcu_idx = srcu_read_lock(hctx->srcu);
630 * blk_mq_complete_request - end I/O on a request
631 * @rq: the request being processed
634 * Ends all I/O on a request. It does not handle partial completions.
635 * The actual completion happens out-of-order, through a IPI handler.
637 bool blk_mq_complete_request(struct request *rq)
639 if (unlikely(blk_should_fake_timeout(rq->q)))
641 __blk_mq_complete_request(rq);
644 EXPORT_SYMBOL(blk_mq_complete_request);
647 * blk_mq_start_request - Start processing a request
648 * @rq: Pointer to request to be started
650 * Function used by device drivers to notify the block layer that a request
651 * is going to be processed now, so blk layer can do proper initializations
652 * such as starting the timeout timer.
654 void blk_mq_start_request(struct request *rq)
656 struct request_queue *q = rq->q;
658 trace_block_rq_issue(q, rq);
660 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
661 rq->io_start_time_ns = ktime_get_ns();
662 rq->stats_sectors = blk_rq_sectors(rq);
663 rq->rq_flags |= RQF_STATS;
667 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
670 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
672 #ifdef CONFIG_BLK_DEV_INTEGRITY
673 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
674 q->integrity.profile->prepare_fn(rq);
677 EXPORT_SYMBOL(blk_mq_start_request);
679 static void __blk_mq_requeue_request(struct request *rq)
681 struct request_queue *q = rq->q;
683 blk_mq_put_driver_tag(rq);
685 trace_block_rq_requeue(q, rq);
686 rq_qos_requeue(q, rq);
688 if (blk_mq_request_started(rq)) {
689 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
690 rq->rq_flags &= ~RQF_TIMED_OUT;
694 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
696 __blk_mq_requeue_request(rq);
698 /* this request will be re-inserted to io scheduler queue */
699 blk_mq_sched_requeue_request(rq);
701 BUG_ON(!list_empty(&rq->queuelist));
702 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
704 EXPORT_SYMBOL(blk_mq_requeue_request);
706 static void blk_mq_requeue_work(struct work_struct *work)
708 struct request_queue *q =
709 container_of(work, struct request_queue, requeue_work.work);
711 struct request *rq, *next;
713 spin_lock_irq(&q->requeue_lock);
714 list_splice_init(&q->requeue_list, &rq_list);
715 spin_unlock_irq(&q->requeue_lock);
717 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
718 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
721 rq->rq_flags &= ~RQF_SOFTBARRIER;
722 list_del_init(&rq->queuelist);
724 * If RQF_DONTPREP, rq has contained some driver specific
725 * data, so insert it to hctx dispatch list to avoid any
728 if (rq->rq_flags & RQF_DONTPREP)
729 blk_mq_request_bypass_insert(rq, false, false);
731 blk_mq_sched_insert_request(rq, true, false, false);
734 while (!list_empty(&rq_list)) {
735 rq = list_entry(rq_list.next, struct request, queuelist);
736 list_del_init(&rq->queuelist);
737 blk_mq_sched_insert_request(rq, false, false, false);
740 blk_mq_run_hw_queues(q, false);
743 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
744 bool kick_requeue_list)
746 struct request_queue *q = rq->q;
750 * We abuse this flag that is otherwise used by the I/O scheduler to
751 * request head insertion from the workqueue.
753 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
755 spin_lock_irqsave(&q->requeue_lock, flags);
757 rq->rq_flags |= RQF_SOFTBARRIER;
758 list_add(&rq->queuelist, &q->requeue_list);
760 list_add_tail(&rq->queuelist, &q->requeue_list);
762 spin_unlock_irqrestore(&q->requeue_lock, flags);
764 if (kick_requeue_list)
765 blk_mq_kick_requeue_list(q);
768 void blk_mq_kick_requeue_list(struct request_queue *q)
770 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
772 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
774 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
777 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
778 msecs_to_jiffies(msecs));
780 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
782 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
784 if (tag < tags->nr_tags) {
785 prefetch(tags->rqs[tag]);
786 return tags->rqs[tag];
791 EXPORT_SYMBOL(blk_mq_tag_to_rq);
793 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
794 void *priv, bool reserved)
797 * If we find a request that is inflight and the queue matches,
798 * we know the queue is busy. Return false to stop the iteration.
800 if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
810 bool blk_mq_queue_inflight(struct request_queue *q)
814 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
817 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
819 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
821 req->rq_flags |= RQF_TIMED_OUT;
822 if (req->q->mq_ops->timeout) {
823 enum blk_eh_timer_return ret;
825 ret = req->q->mq_ops->timeout(req, reserved);
826 if (ret == BLK_EH_DONE)
828 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
834 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
836 unsigned long deadline;
838 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
840 if (rq->rq_flags & RQF_TIMED_OUT)
843 deadline = READ_ONCE(rq->deadline);
844 if (time_after_eq(jiffies, deadline))
849 else if (time_after(*next, deadline))
854 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
855 struct request *rq, void *priv, bool reserved)
857 unsigned long *next = priv;
860 * Just do a quick check if it is expired before locking the request in
861 * so we're not unnecessarilly synchronizing across CPUs.
863 if (!blk_mq_req_expired(rq, next))
867 * We have reason to believe the request may be expired. Take a
868 * reference on the request to lock this request lifetime into its
869 * currently allocated context to prevent it from being reallocated in
870 * the event the completion by-passes this timeout handler.
872 * If the reference was already released, then the driver beat the
873 * timeout handler to posting a natural completion.
875 if (!refcount_inc_not_zero(&rq->ref))
879 * The request is now locked and cannot be reallocated underneath the
880 * timeout handler's processing. Re-verify this exact request is truly
881 * expired; if it is not expired, then the request was completed and
882 * reallocated as a new request.
884 if (blk_mq_req_expired(rq, next))
885 blk_mq_rq_timed_out(rq, reserved);
887 if (is_flush_rq(rq, hctx))
889 else if (refcount_dec_and_test(&rq->ref))
890 __blk_mq_free_request(rq);
895 static void blk_mq_timeout_work(struct work_struct *work)
897 struct request_queue *q =
898 container_of(work, struct request_queue, timeout_work);
899 unsigned long next = 0;
900 struct blk_mq_hw_ctx *hctx;
903 /* A deadlock might occur if a request is stuck requiring a
904 * timeout at the same time a queue freeze is waiting
905 * completion, since the timeout code would not be able to
906 * acquire the queue reference here.
908 * That's why we don't use blk_queue_enter here; instead, we use
909 * percpu_ref_tryget directly, because we need to be able to
910 * obtain a reference even in the short window between the queue
911 * starting to freeze, by dropping the first reference in
912 * blk_freeze_queue_start, and the moment the last request is
913 * consumed, marked by the instant q_usage_counter reaches
916 if (!percpu_ref_tryget(&q->q_usage_counter))
919 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
922 mod_timer(&q->timeout, next);
925 * Request timeouts are handled as a forward rolling timer. If
926 * we end up here it means that no requests are pending and
927 * also that no request has been pending for a while. Mark
930 queue_for_each_hw_ctx(q, hctx, i) {
931 /* the hctx may be unmapped, so check it here */
932 if (blk_mq_hw_queue_mapped(hctx))
933 blk_mq_tag_idle(hctx);
939 struct flush_busy_ctx_data {
940 struct blk_mq_hw_ctx *hctx;
941 struct list_head *list;
944 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
946 struct flush_busy_ctx_data *flush_data = data;
947 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
948 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
949 enum hctx_type type = hctx->type;
951 spin_lock(&ctx->lock);
952 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
953 sbitmap_clear_bit(sb, bitnr);
954 spin_unlock(&ctx->lock);
959 * Process software queues that have been marked busy, splicing them
960 * to the for-dispatch
962 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
964 struct flush_busy_ctx_data data = {
969 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
971 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
973 struct dispatch_rq_data {
974 struct blk_mq_hw_ctx *hctx;
978 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
981 struct dispatch_rq_data *dispatch_data = data;
982 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
983 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
984 enum hctx_type type = hctx->type;
986 spin_lock(&ctx->lock);
987 if (!list_empty(&ctx->rq_lists[type])) {
988 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
989 list_del_init(&dispatch_data->rq->queuelist);
990 if (list_empty(&ctx->rq_lists[type]))
991 sbitmap_clear_bit(sb, bitnr);
993 spin_unlock(&ctx->lock);
995 return !dispatch_data->rq;
998 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
999 struct blk_mq_ctx *start)
1001 unsigned off = start ? start->index_hw[hctx->type] : 0;
1002 struct dispatch_rq_data data = {
1007 __sbitmap_for_each_set(&hctx->ctx_map, off,
1008 dispatch_rq_from_ctx, &data);
1013 static inline unsigned int queued_to_index(unsigned int queued)
1018 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1021 bool blk_mq_get_driver_tag(struct request *rq)
1023 struct blk_mq_alloc_data data = {
1025 .hctx = rq->mq_hctx,
1026 .flags = BLK_MQ_REQ_NOWAIT,
1027 .cmd_flags = rq->cmd_flags,
1034 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1035 data.flags |= BLK_MQ_REQ_RESERVED;
1037 shared = blk_mq_tag_busy(data.hctx);
1038 rq->tag = blk_mq_get_tag(&data);
1041 rq->rq_flags |= RQF_MQ_INFLIGHT;
1042 atomic_inc(&data.hctx->nr_active);
1044 data.hctx->tags->rqs[rq->tag] = rq;
1047 return rq->tag != -1;
1050 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1051 int flags, void *key)
1053 struct blk_mq_hw_ctx *hctx;
1055 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1057 spin_lock(&hctx->dispatch_wait_lock);
1058 if (!list_empty(&wait->entry)) {
1059 struct sbitmap_queue *sbq;
1061 list_del_init(&wait->entry);
1062 sbq = &hctx->tags->bitmap_tags;
1063 atomic_dec(&sbq->ws_active);
1065 spin_unlock(&hctx->dispatch_wait_lock);
1067 blk_mq_run_hw_queue(hctx, true);
1072 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1073 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1074 * restart. For both cases, take care to check the condition again after
1075 * marking us as waiting.
1077 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1080 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1081 struct wait_queue_head *wq;
1082 wait_queue_entry_t *wait;
1085 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1086 blk_mq_sched_mark_restart_hctx(hctx);
1089 * It's possible that a tag was freed in the window between the
1090 * allocation failure and adding the hardware queue to the wait
1093 * Don't clear RESTART here, someone else could have set it.
1094 * At most this will cost an extra queue run.
1096 return blk_mq_get_driver_tag(rq);
1099 wait = &hctx->dispatch_wait;
1100 if (!list_empty_careful(&wait->entry))
1103 wq = &bt_wait_ptr(sbq, hctx)->wait;
1105 spin_lock_irq(&wq->lock);
1106 spin_lock(&hctx->dispatch_wait_lock);
1107 if (!list_empty(&wait->entry)) {
1108 spin_unlock(&hctx->dispatch_wait_lock);
1109 spin_unlock_irq(&wq->lock);
1113 atomic_inc(&sbq->ws_active);
1114 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1115 __add_wait_queue(wq, wait);
1118 * It's possible that a tag was freed in the window between the
1119 * allocation failure and adding the hardware queue to the wait
1122 ret = blk_mq_get_driver_tag(rq);
1124 spin_unlock(&hctx->dispatch_wait_lock);
1125 spin_unlock_irq(&wq->lock);
1130 * We got a tag, remove ourselves from the wait queue to ensure
1131 * someone else gets the wakeup.
1133 list_del_init(&wait->entry);
1134 atomic_dec(&sbq->ws_active);
1135 spin_unlock(&hctx->dispatch_wait_lock);
1136 spin_unlock_irq(&wq->lock);
1141 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1142 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1144 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1145 * - EWMA is one simple way to compute running average value
1146 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1147 * - take 4 as factor for avoiding to get too small(0) result, and this
1148 * factor doesn't matter because EWMA decreases exponentially
1150 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1154 if (hctx->queue->elevator)
1157 ewma = hctx->dispatch_busy;
1162 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1164 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1165 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1167 hctx->dispatch_busy = ewma;
1170 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1172 static void blk_mq_handle_dev_resource(struct request *rq,
1173 struct list_head *list)
1175 struct request *next =
1176 list_first_entry_or_null(list, struct request, queuelist);
1179 * If an I/O scheduler has been configured and we got a driver tag for
1180 * the next request already, free it.
1183 blk_mq_put_driver_tag(next);
1185 list_add(&rq->queuelist, list);
1186 __blk_mq_requeue_request(rq);
1189 static void blk_mq_handle_zone_resource(struct request *rq,
1190 struct list_head *zone_list)
1193 * If we end up here it is because we cannot dispatch a request to a
1194 * specific zone due to LLD level zone-write locking or other zone
1195 * related resource not being available. In this case, set the request
1196 * aside in zone_list for retrying it later.
1198 list_add(&rq->queuelist, zone_list);
1199 __blk_mq_requeue_request(rq);
1203 * Returns true if we did some work AND can potentially do more.
1205 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1208 struct blk_mq_hw_ctx *hctx;
1209 struct request *rq, *nxt;
1210 bool no_tag = false;
1212 blk_status_t ret = BLK_STS_OK;
1213 bool no_budget_avail = false;
1214 LIST_HEAD(zone_list);
1216 if (list_empty(list))
1219 WARN_ON(!list_is_singular(list) && got_budget);
1222 * Now process all the entries, sending them to the driver.
1224 errors = queued = 0;
1226 struct blk_mq_queue_data bd;
1228 rq = list_first_entry(list, struct request, queuelist);
1231 if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) {
1232 blk_mq_put_driver_tag(rq);
1233 no_budget_avail = true;
1237 if (!blk_mq_get_driver_tag(rq)) {
1239 * The initial allocation attempt failed, so we need to
1240 * rerun the hardware queue when a tag is freed. The
1241 * waitqueue takes care of that. If the queue is run
1242 * before we add this entry back on the dispatch list,
1243 * we'll re-run it below.
1245 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1246 blk_mq_put_dispatch_budget(hctx);
1248 * For non-shared tags, the RESTART check
1251 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1257 list_del_init(&rq->queuelist);
1262 * Flag last if we have no more requests, or if we have more
1263 * but can't assign a driver tag to it.
1265 if (list_empty(list))
1268 nxt = list_first_entry(list, struct request, queuelist);
1269 bd.last = !blk_mq_get_driver_tag(nxt);
1272 ret = q->mq_ops->queue_rq(hctx, &bd);
1273 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1274 blk_mq_handle_dev_resource(rq, list);
1276 } else if (ret == BLK_STS_ZONE_RESOURCE) {
1278 * Move the request to zone_list and keep going through
1279 * the dispatch list to find more requests the drive can
1282 blk_mq_handle_zone_resource(rq, &zone_list);
1283 if (list_empty(list))
1288 if (unlikely(ret != BLK_STS_OK)) {
1290 blk_mq_end_request(rq, BLK_STS_IOERR);
1295 } while (!list_empty(list));
1297 if (!list_empty(&zone_list))
1298 list_splice_tail_init(&zone_list, list);
1300 hctx->dispatched[queued_to_index(queued)]++;
1303 * Any items that need requeuing? Stuff them into hctx->dispatch,
1304 * that is where we will continue on next queue run.
1306 if (!list_empty(list)) {
1310 * If we didn't flush the entire list, we could have told
1311 * the driver there was more coming, but that turned out to
1314 if (q->mq_ops->commit_rqs && queued)
1315 q->mq_ops->commit_rqs(hctx);
1317 spin_lock(&hctx->lock);
1318 list_splice_tail_init(list, &hctx->dispatch);
1319 spin_unlock(&hctx->lock);
1322 * If SCHED_RESTART was set by the caller of this function and
1323 * it is no longer set that means that it was cleared by another
1324 * thread and hence that a queue rerun is needed.
1326 * If 'no_tag' is set, that means that we failed getting
1327 * a driver tag with an I/O scheduler attached. If our dispatch
1328 * waitqueue is no longer active, ensure that we run the queue
1329 * AFTER adding our entries back to the list.
1331 * If no I/O scheduler has been configured it is possible that
1332 * the hardware queue got stopped and restarted before requests
1333 * were pushed back onto the dispatch list. Rerun the queue to
1334 * avoid starvation. Notes:
1335 * - blk_mq_run_hw_queue() checks whether or not a queue has
1336 * been stopped before rerunning a queue.
1337 * - Some but not all block drivers stop a queue before
1338 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1341 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1342 * bit is set, run queue after a delay to avoid IO stalls
1343 * that could otherwise occur if the queue is idle. We'll do
1344 * similar if we couldn't get budget and SCHED_RESTART is set.
1346 needs_restart = blk_mq_sched_needs_restart(hctx);
1347 if (!needs_restart ||
1348 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1349 blk_mq_run_hw_queue(hctx, true);
1350 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1352 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1354 blk_mq_update_dispatch_busy(hctx, true);
1357 blk_mq_update_dispatch_busy(hctx, false);
1360 * If the host/device is unable to accept more work, inform the
1363 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1366 return (queued + errors) != 0;
1370 * __blk_mq_run_hw_queue - Run a hardware queue.
1371 * @hctx: Pointer to the hardware queue to run.
1373 * Send pending requests to the hardware.
1375 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1380 * We should be running this queue from one of the CPUs that
1383 * There are at least two related races now between setting
1384 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1385 * __blk_mq_run_hw_queue():
1387 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1388 * but later it becomes online, then this warning is harmless
1391 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1392 * but later it becomes offline, then the warning can't be
1393 * triggered, and we depend on blk-mq timeout handler to
1394 * handle dispatched requests to this hctx
1396 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1397 cpu_online(hctx->next_cpu)) {
1398 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1399 raw_smp_processor_id(),
1400 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1405 * We can't run the queue inline with ints disabled. Ensure that
1406 * we catch bad users of this early.
1408 WARN_ON_ONCE(in_interrupt());
1410 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1412 hctx_lock(hctx, &srcu_idx);
1413 blk_mq_sched_dispatch_requests(hctx);
1414 hctx_unlock(hctx, srcu_idx);
1417 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1419 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1421 if (cpu >= nr_cpu_ids)
1422 cpu = cpumask_first(hctx->cpumask);
1427 * It'd be great if the workqueue API had a way to pass
1428 * in a mask and had some smarts for more clever placement.
1429 * For now we just round-robin here, switching for every
1430 * BLK_MQ_CPU_WORK_BATCH queued items.
1432 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1435 int next_cpu = hctx->next_cpu;
1437 if (hctx->queue->nr_hw_queues == 1)
1438 return WORK_CPU_UNBOUND;
1440 if (--hctx->next_cpu_batch <= 0) {
1442 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1444 if (next_cpu >= nr_cpu_ids)
1445 next_cpu = blk_mq_first_mapped_cpu(hctx);
1446 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1450 * Do unbound schedule if we can't find a online CPU for this hctx,
1451 * and it should only happen in the path of handling CPU DEAD.
1453 if (!cpu_online(next_cpu)) {
1460 * Make sure to re-select CPU next time once after CPUs
1461 * in hctx->cpumask become online again.
1463 hctx->next_cpu = next_cpu;
1464 hctx->next_cpu_batch = 1;
1465 return WORK_CPU_UNBOUND;
1468 hctx->next_cpu = next_cpu;
1473 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1474 * @hctx: Pointer to the hardware queue to run.
1475 * @async: If we want to run the queue asynchronously.
1476 * @msecs: Microseconds of delay to wait before running the queue.
1478 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1479 * with a delay of @msecs.
1481 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1482 unsigned long msecs)
1484 if (unlikely(blk_mq_hctx_stopped(hctx)))
1487 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1488 int cpu = get_cpu();
1489 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1490 __blk_mq_run_hw_queue(hctx);
1498 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1499 msecs_to_jiffies(msecs));
1503 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1504 * @hctx: Pointer to the hardware queue to run.
1505 * @msecs: Microseconds of delay to wait before running the queue.
1507 * Run a hardware queue asynchronously with a delay of @msecs.
1509 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1511 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1513 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1516 * blk_mq_run_hw_queue - Start to run a hardware queue.
1517 * @hctx: Pointer to the hardware queue to run.
1518 * @async: If we want to run the queue asynchronously.
1520 * Check if the request queue is not in a quiesced state and if there are
1521 * pending requests to be sent. If this is true, run the queue to send requests
1524 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1530 * When queue is quiesced, we may be switching io scheduler, or
1531 * updating nr_hw_queues, or other things, and we can't run queue
1532 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1534 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1537 hctx_lock(hctx, &srcu_idx);
1538 need_run = !blk_queue_quiesced(hctx->queue) &&
1539 blk_mq_hctx_has_pending(hctx);
1540 hctx_unlock(hctx, srcu_idx);
1543 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1545 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1548 * blk_mq_run_hw_queue - Run all hardware queues in a request queue.
1549 * @q: Pointer to the request queue to run.
1550 * @async: If we want to run the queue asynchronously.
1552 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1554 struct blk_mq_hw_ctx *hctx;
1557 queue_for_each_hw_ctx(q, hctx, i) {
1558 if (blk_mq_hctx_stopped(hctx))
1561 blk_mq_run_hw_queue(hctx, async);
1564 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1567 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1568 * @q: Pointer to the request queue to run.
1569 * @msecs: Microseconds of delay to wait before running the queues.
1571 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1573 struct blk_mq_hw_ctx *hctx;
1576 queue_for_each_hw_ctx(q, hctx, i) {
1577 if (blk_mq_hctx_stopped(hctx))
1580 blk_mq_delay_run_hw_queue(hctx, msecs);
1583 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1586 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1587 * @q: request queue.
1589 * The caller is responsible for serializing this function against
1590 * blk_mq_{start,stop}_hw_queue().
1592 bool blk_mq_queue_stopped(struct request_queue *q)
1594 struct blk_mq_hw_ctx *hctx;
1597 queue_for_each_hw_ctx(q, hctx, i)
1598 if (blk_mq_hctx_stopped(hctx))
1603 EXPORT_SYMBOL(blk_mq_queue_stopped);
1606 * This function is often used for pausing .queue_rq() by driver when
1607 * there isn't enough resource or some conditions aren't satisfied, and
1608 * BLK_STS_RESOURCE is usually returned.
1610 * We do not guarantee that dispatch can be drained or blocked
1611 * after blk_mq_stop_hw_queue() returns. Please use
1612 * blk_mq_quiesce_queue() for that requirement.
1614 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1616 cancel_delayed_work(&hctx->run_work);
1618 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1620 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1623 * This function is often used for pausing .queue_rq() by driver when
1624 * there isn't enough resource or some conditions aren't satisfied, and
1625 * BLK_STS_RESOURCE is usually returned.
1627 * We do not guarantee that dispatch can be drained or blocked
1628 * after blk_mq_stop_hw_queues() returns. Please use
1629 * blk_mq_quiesce_queue() for that requirement.
1631 void blk_mq_stop_hw_queues(struct request_queue *q)
1633 struct blk_mq_hw_ctx *hctx;
1636 queue_for_each_hw_ctx(q, hctx, i)
1637 blk_mq_stop_hw_queue(hctx);
1639 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1641 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1643 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1645 blk_mq_run_hw_queue(hctx, false);
1647 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1649 void blk_mq_start_hw_queues(struct request_queue *q)
1651 struct blk_mq_hw_ctx *hctx;
1654 queue_for_each_hw_ctx(q, hctx, i)
1655 blk_mq_start_hw_queue(hctx);
1657 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1659 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1661 if (!blk_mq_hctx_stopped(hctx))
1664 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1665 blk_mq_run_hw_queue(hctx, async);
1667 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1669 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1671 struct blk_mq_hw_ctx *hctx;
1674 queue_for_each_hw_ctx(q, hctx, i)
1675 blk_mq_start_stopped_hw_queue(hctx, async);
1677 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1679 static void blk_mq_run_work_fn(struct work_struct *work)
1681 struct blk_mq_hw_ctx *hctx;
1683 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1686 * If we are stopped, don't run the queue.
1688 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1691 __blk_mq_run_hw_queue(hctx);
1694 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1698 struct blk_mq_ctx *ctx = rq->mq_ctx;
1699 enum hctx_type type = hctx->type;
1701 lockdep_assert_held(&ctx->lock);
1703 trace_block_rq_insert(hctx->queue, rq);
1706 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1708 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1711 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1714 struct blk_mq_ctx *ctx = rq->mq_ctx;
1716 lockdep_assert_held(&ctx->lock);
1718 __blk_mq_insert_req_list(hctx, rq, at_head);
1719 blk_mq_hctx_mark_pending(hctx, ctx);
1723 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1724 * @rq: Pointer to request to be inserted.
1725 * @run_queue: If we should run the hardware queue after inserting the request.
1727 * Should only be used carefully, when the caller knows we want to
1728 * bypass a potential IO scheduler on the target device.
1730 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1733 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1735 spin_lock(&hctx->lock);
1737 list_add(&rq->queuelist, &hctx->dispatch);
1739 list_add_tail(&rq->queuelist, &hctx->dispatch);
1740 spin_unlock(&hctx->lock);
1743 blk_mq_run_hw_queue(hctx, false);
1746 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1747 struct list_head *list)
1751 enum hctx_type type = hctx->type;
1754 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1757 list_for_each_entry(rq, list, queuelist) {
1758 BUG_ON(rq->mq_ctx != ctx);
1759 trace_block_rq_insert(hctx->queue, rq);
1762 spin_lock(&ctx->lock);
1763 list_splice_tail_init(list, &ctx->rq_lists[type]);
1764 blk_mq_hctx_mark_pending(hctx, ctx);
1765 spin_unlock(&ctx->lock);
1768 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1770 struct request *rqa = container_of(a, struct request, queuelist);
1771 struct request *rqb = container_of(b, struct request, queuelist);
1773 if (rqa->mq_ctx != rqb->mq_ctx)
1774 return rqa->mq_ctx > rqb->mq_ctx;
1775 if (rqa->mq_hctx != rqb->mq_hctx)
1776 return rqa->mq_hctx > rqb->mq_hctx;
1778 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1781 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1785 if (list_empty(&plug->mq_list))
1787 list_splice_init(&plug->mq_list, &list);
1789 if (plug->rq_count > 2 && plug->multiple_queues)
1790 list_sort(NULL, &list, plug_rq_cmp);
1795 struct list_head rq_list;
1796 struct request *rq, *head_rq = list_entry_rq(list.next);
1797 struct list_head *pos = &head_rq->queuelist; /* skip first */
1798 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1799 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1800 unsigned int depth = 1;
1802 list_for_each_continue(pos, &list) {
1803 rq = list_entry_rq(pos);
1805 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1810 list_cut_before(&rq_list, &list, pos);
1811 trace_block_unplug(head_rq->q, depth, !from_schedule);
1812 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1814 } while(!list_empty(&list));
1817 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1818 unsigned int nr_segs)
1820 if (bio->bi_opf & REQ_RAHEAD)
1821 rq->cmd_flags |= REQ_FAILFAST_MASK;
1823 rq->__sector = bio->bi_iter.bi_sector;
1824 rq->write_hint = bio->bi_write_hint;
1825 blk_rq_bio_prep(rq, bio, nr_segs);
1826 blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1828 blk_account_io_start(rq, true);
1831 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1833 blk_qc_t *cookie, bool last)
1835 struct request_queue *q = rq->q;
1836 struct blk_mq_queue_data bd = {
1840 blk_qc_t new_cookie;
1843 new_cookie = request_to_qc_t(hctx, rq);
1846 * For OK queue, we are done. For error, caller may kill it.
1847 * Any other error (busy), just add it to our list as we
1848 * previously would have done.
1850 ret = q->mq_ops->queue_rq(hctx, &bd);
1853 blk_mq_update_dispatch_busy(hctx, false);
1854 *cookie = new_cookie;
1856 case BLK_STS_RESOURCE:
1857 case BLK_STS_DEV_RESOURCE:
1858 blk_mq_update_dispatch_busy(hctx, true);
1859 __blk_mq_requeue_request(rq);
1862 blk_mq_update_dispatch_busy(hctx, false);
1863 *cookie = BLK_QC_T_NONE;
1870 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1873 bool bypass_insert, bool last)
1875 struct request_queue *q = rq->q;
1876 bool run_queue = true;
1879 * RCU or SRCU read lock is needed before checking quiesced flag.
1881 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1882 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1883 * and avoid driver to try to dispatch again.
1885 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1887 bypass_insert = false;
1891 if (q->elevator && !bypass_insert)
1894 if (!blk_mq_get_dispatch_budget(hctx))
1897 if (!blk_mq_get_driver_tag(rq)) {
1898 blk_mq_put_dispatch_budget(hctx);
1902 return __blk_mq_issue_directly(hctx, rq, cookie, last);
1905 return BLK_STS_RESOURCE;
1907 blk_mq_request_bypass_insert(rq, false, run_queue);
1912 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
1913 * @hctx: Pointer of the associated hardware queue.
1914 * @rq: Pointer to request to be sent.
1915 * @cookie: Request queue cookie.
1917 * If the device has enough resources to accept a new request now, send the
1918 * request directly to device driver. Else, insert at hctx->dispatch queue, so
1919 * we can try send it another time in the future. Requests inserted at this
1920 * queue have higher priority.
1922 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1923 struct request *rq, blk_qc_t *cookie)
1928 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1930 hctx_lock(hctx, &srcu_idx);
1932 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
1933 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1934 blk_mq_request_bypass_insert(rq, false, true);
1935 else if (ret != BLK_STS_OK)
1936 blk_mq_end_request(rq, ret);
1938 hctx_unlock(hctx, srcu_idx);
1941 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
1945 blk_qc_t unused_cookie;
1946 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1948 hctx_lock(hctx, &srcu_idx);
1949 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
1950 hctx_unlock(hctx, srcu_idx);
1955 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1956 struct list_head *list)
1960 while (!list_empty(list)) {
1962 struct request *rq = list_first_entry(list, struct request,
1965 list_del_init(&rq->queuelist);
1966 ret = blk_mq_request_issue_directly(rq, list_empty(list));
1967 if (ret != BLK_STS_OK) {
1968 if (ret == BLK_STS_RESOURCE ||
1969 ret == BLK_STS_DEV_RESOURCE) {
1970 blk_mq_request_bypass_insert(rq, false,
1974 blk_mq_end_request(rq, ret);
1980 * If we didn't flush the entire list, we could have told
1981 * the driver there was more coming, but that turned out to
1984 if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs && queued)
1985 hctx->queue->mq_ops->commit_rqs(hctx);
1988 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1990 list_add_tail(&rq->queuelist, &plug->mq_list);
1992 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1993 struct request *tmp;
1995 tmp = list_first_entry(&plug->mq_list, struct request,
1997 if (tmp->q != rq->q)
1998 plug->multiple_queues = true;
2003 * blk_mq_make_request - Create and send a request to block device.
2004 * @q: Request queue pointer.
2005 * @bio: Bio pointer.
2007 * Builds up a request structure from @q and @bio and send to the device. The
2008 * request may not be queued directly to hardware if:
2009 * * This request can be merged with another one
2010 * * We want to place request at plug queue for possible future merging
2011 * * There is an IO scheduler active at this queue
2013 * It will not queue the request if there is an error with the bio, or at the
2016 * Returns: Request queue cookie.
2018 blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
2020 const int is_sync = op_is_sync(bio->bi_opf);
2021 const int is_flush_fua = op_is_flush(bio->bi_opf);
2022 struct blk_mq_alloc_data data = { .flags = 0};
2024 struct blk_plug *plug;
2025 struct request *same_queue_rq = NULL;
2026 unsigned int nr_segs;
2030 blk_queue_bounce(q, &bio);
2031 __blk_queue_split(q, &bio, &nr_segs);
2033 if (!bio_integrity_prep(bio))
2034 return BLK_QC_T_NONE;
2036 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2037 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2038 return BLK_QC_T_NONE;
2040 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2041 return BLK_QC_T_NONE;
2043 rq_qos_throttle(q, bio);
2045 data.cmd_flags = bio->bi_opf;
2046 rq = blk_mq_get_request(q, bio, &data);
2047 if (unlikely(!rq)) {
2048 rq_qos_cleanup(q, bio);
2049 if (bio->bi_opf & REQ_NOWAIT)
2050 bio_wouldblock_error(bio);
2051 return BLK_QC_T_NONE;
2054 trace_block_getrq(q, bio, bio->bi_opf);
2056 rq_qos_track(q, rq, bio);
2058 cookie = request_to_qc_t(data.hctx, rq);
2060 blk_mq_bio_to_request(rq, bio, nr_segs);
2062 ret = blk_crypto_init_request(rq);
2063 if (ret != BLK_STS_OK) {
2064 bio->bi_status = ret;
2066 blk_mq_free_request(rq);
2067 return BLK_QC_T_NONE;
2070 plug = blk_mq_plug(q, bio);
2071 if (unlikely(is_flush_fua)) {
2072 /* Bypass scheduler for flush requests */
2073 blk_insert_flush(rq);
2074 blk_mq_run_hw_queue(data.hctx, true);
2075 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2076 !blk_queue_nonrot(q))) {
2078 * Use plugging if we have a ->commit_rqs() hook as well, as
2079 * we know the driver uses bd->last in a smart fashion.
2081 * Use normal plugging if this disk is slow HDD, as sequential
2082 * IO may benefit a lot from plug merging.
2084 unsigned int request_count = plug->rq_count;
2085 struct request *last = NULL;
2088 trace_block_plug(q);
2090 last = list_entry_rq(plug->mq_list.prev);
2092 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2093 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2094 blk_flush_plug_list(plug, false);
2095 trace_block_plug(q);
2098 blk_add_rq_to_plug(plug, rq);
2099 } else if (q->elevator) {
2100 /* Insert the request at the IO scheduler queue */
2101 blk_mq_sched_insert_request(rq, false, true, true);
2102 } else if (plug && !blk_queue_nomerges(q)) {
2104 * We do limited plugging. If the bio can be merged, do that.
2105 * Otherwise the existing request in the plug list will be
2106 * issued. So the plug list will have one request at most
2107 * The plug list might get flushed before this. If that happens,
2108 * the plug list is empty, and same_queue_rq is invalid.
2110 if (list_empty(&plug->mq_list))
2111 same_queue_rq = NULL;
2112 if (same_queue_rq) {
2113 list_del_init(&same_queue_rq->queuelist);
2116 blk_add_rq_to_plug(plug, rq);
2117 trace_block_plug(q);
2119 if (same_queue_rq) {
2120 data.hctx = same_queue_rq->mq_hctx;
2121 trace_block_unplug(q, 1, true);
2122 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2125 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2126 !data.hctx->dispatch_busy) {
2128 * There is no scheduler and we can try to send directly
2131 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2134 blk_mq_sched_insert_request(rq, false, true, true);
2139 EXPORT_SYMBOL_GPL(blk_mq_make_request); /* only for request based dm */
2141 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2142 unsigned int hctx_idx)
2146 if (tags->rqs && set->ops->exit_request) {
2149 for (i = 0; i < tags->nr_tags; i++) {
2150 struct request *rq = tags->static_rqs[i];
2154 set->ops->exit_request(set, rq, hctx_idx);
2155 tags->static_rqs[i] = NULL;
2159 while (!list_empty(&tags->page_list)) {
2160 page = list_first_entry(&tags->page_list, struct page, lru);
2161 list_del_init(&page->lru);
2163 * Remove kmemleak object previously allocated in
2164 * blk_mq_alloc_rqs().
2166 kmemleak_free(page_address(page));
2167 __free_pages(page, page->private);
2171 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2175 kfree(tags->static_rqs);
2176 tags->static_rqs = NULL;
2178 blk_mq_free_tags(tags);
2181 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2182 unsigned int hctx_idx,
2183 unsigned int nr_tags,
2184 unsigned int reserved_tags)
2186 struct blk_mq_tags *tags;
2189 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2190 if (node == NUMA_NO_NODE)
2191 node = set->numa_node;
2193 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2194 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2198 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2199 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2202 blk_mq_free_tags(tags);
2206 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2207 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2209 if (!tags->static_rqs) {
2211 blk_mq_free_tags(tags);
2218 static size_t order_to_size(unsigned int order)
2220 return (size_t)PAGE_SIZE << order;
2223 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2224 unsigned int hctx_idx, int node)
2228 if (set->ops->init_request) {
2229 ret = set->ops->init_request(set, rq, hctx_idx, node);
2234 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2238 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2239 unsigned int hctx_idx, unsigned int depth)
2241 unsigned int i, j, entries_per_page, max_order = 4;
2242 size_t rq_size, left;
2245 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2246 if (node == NUMA_NO_NODE)
2247 node = set->numa_node;
2249 INIT_LIST_HEAD(&tags->page_list);
2252 * rq_size is the size of the request plus driver payload, rounded
2253 * to the cacheline size
2255 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2257 left = rq_size * depth;
2259 for (i = 0; i < depth; ) {
2260 int this_order = max_order;
2265 while (this_order && left < order_to_size(this_order - 1))
2269 page = alloc_pages_node(node,
2270 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2276 if (order_to_size(this_order) < rq_size)
2283 page->private = this_order;
2284 list_add_tail(&page->lru, &tags->page_list);
2286 p = page_address(page);
2288 * Allow kmemleak to scan these pages as they contain pointers
2289 * to additional allocations like via ops->init_request().
2291 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2292 entries_per_page = order_to_size(this_order) / rq_size;
2293 to_do = min(entries_per_page, depth - i);
2294 left -= to_do * rq_size;
2295 for (j = 0; j < to_do; j++) {
2296 struct request *rq = p;
2298 tags->static_rqs[i] = rq;
2299 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2300 tags->static_rqs[i] = NULL;
2311 blk_mq_free_rqs(set, tags, hctx_idx);
2316 * 'cpu' is going away. splice any existing rq_list entries from this
2317 * software queue to the hw queue dispatch list, and ensure that it
2320 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2322 struct blk_mq_hw_ctx *hctx;
2323 struct blk_mq_ctx *ctx;
2325 enum hctx_type type;
2327 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2328 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2331 spin_lock(&ctx->lock);
2332 if (!list_empty(&ctx->rq_lists[type])) {
2333 list_splice_init(&ctx->rq_lists[type], &tmp);
2334 blk_mq_hctx_clear_pending(hctx, ctx);
2336 spin_unlock(&ctx->lock);
2338 if (list_empty(&tmp))
2341 spin_lock(&hctx->lock);
2342 list_splice_tail_init(&tmp, &hctx->dispatch);
2343 spin_unlock(&hctx->lock);
2345 blk_mq_run_hw_queue(hctx, true);
2349 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2351 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2355 /* hctx->ctxs will be freed in queue's release handler */
2356 static void blk_mq_exit_hctx(struct request_queue *q,
2357 struct blk_mq_tag_set *set,
2358 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2360 if (blk_mq_hw_queue_mapped(hctx))
2361 blk_mq_tag_idle(hctx);
2363 if (set->ops->exit_request)
2364 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2366 if (set->ops->exit_hctx)
2367 set->ops->exit_hctx(hctx, hctx_idx);
2369 blk_mq_remove_cpuhp(hctx);
2371 spin_lock(&q->unused_hctx_lock);
2372 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2373 spin_unlock(&q->unused_hctx_lock);
2376 static void blk_mq_exit_hw_queues(struct request_queue *q,
2377 struct blk_mq_tag_set *set, int nr_queue)
2379 struct blk_mq_hw_ctx *hctx;
2382 queue_for_each_hw_ctx(q, hctx, i) {
2385 blk_mq_debugfs_unregister_hctx(hctx);
2386 blk_mq_exit_hctx(q, set, hctx, i);
2390 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2392 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2394 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2395 __alignof__(struct blk_mq_hw_ctx)) !=
2396 sizeof(struct blk_mq_hw_ctx));
2398 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2399 hw_ctx_size += sizeof(struct srcu_struct);
2404 static int blk_mq_init_hctx(struct request_queue *q,
2405 struct blk_mq_tag_set *set,
2406 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2408 hctx->queue_num = hctx_idx;
2410 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2412 hctx->tags = set->tags[hctx_idx];
2414 if (set->ops->init_hctx &&
2415 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2416 goto unregister_cpu_notifier;
2418 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2424 if (set->ops->exit_hctx)
2425 set->ops->exit_hctx(hctx, hctx_idx);
2426 unregister_cpu_notifier:
2427 blk_mq_remove_cpuhp(hctx);
2431 static struct blk_mq_hw_ctx *
2432 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2435 struct blk_mq_hw_ctx *hctx;
2436 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2438 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2440 goto fail_alloc_hctx;
2442 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2445 atomic_set(&hctx->nr_active, 0);
2446 if (node == NUMA_NO_NODE)
2447 node = set->numa_node;
2448 hctx->numa_node = node;
2450 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2451 spin_lock_init(&hctx->lock);
2452 INIT_LIST_HEAD(&hctx->dispatch);
2454 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2456 INIT_LIST_HEAD(&hctx->hctx_list);
2459 * Allocate space for all possible cpus to avoid allocation at
2462 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2467 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2472 spin_lock_init(&hctx->dispatch_wait_lock);
2473 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2474 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2476 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2480 if (hctx->flags & BLK_MQ_F_BLOCKING)
2481 init_srcu_struct(hctx->srcu);
2482 blk_mq_hctx_kobj_init(hctx);
2487 sbitmap_free(&hctx->ctx_map);
2491 free_cpumask_var(hctx->cpumask);
2498 static void blk_mq_init_cpu_queues(struct request_queue *q,
2499 unsigned int nr_hw_queues)
2501 struct blk_mq_tag_set *set = q->tag_set;
2504 for_each_possible_cpu(i) {
2505 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2506 struct blk_mq_hw_ctx *hctx;
2510 spin_lock_init(&__ctx->lock);
2511 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2512 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2517 * Set local node, IFF we have more than one hw queue. If
2518 * not, we remain on the home node of the device
2520 for (j = 0; j < set->nr_maps; j++) {
2521 hctx = blk_mq_map_queue_type(q, j, i);
2522 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2523 hctx->numa_node = local_memory_node(cpu_to_node(i));
2528 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2533 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2534 set->queue_depth, set->reserved_tags);
2535 if (!set->tags[hctx_idx])
2538 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2543 blk_mq_free_rq_map(set->tags[hctx_idx]);
2544 set->tags[hctx_idx] = NULL;
2548 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2549 unsigned int hctx_idx)
2551 if (set->tags && set->tags[hctx_idx]) {
2552 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2553 blk_mq_free_rq_map(set->tags[hctx_idx]);
2554 set->tags[hctx_idx] = NULL;
2558 static void blk_mq_map_swqueue(struct request_queue *q)
2560 unsigned int i, j, hctx_idx;
2561 struct blk_mq_hw_ctx *hctx;
2562 struct blk_mq_ctx *ctx;
2563 struct blk_mq_tag_set *set = q->tag_set;
2565 queue_for_each_hw_ctx(q, hctx, i) {
2566 cpumask_clear(hctx->cpumask);
2568 hctx->dispatch_from = NULL;
2572 * Map software to hardware queues.
2574 * If the cpu isn't present, the cpu is mapped to first hctx.
2576 for_each_possible_cpu(i) {
2578 ctx = per_cpu_ptr(q->queue_ctx, i);
2579 for (j = 0; j < set->nr_maps; j++) {
2580 if (!set->map[j].nr_queues) {
2581 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2582 HCTX_TYPE_DEFAULT, i);
2585 hctx_idx = set->map[j].mq_map[i];
2586 /* unmapped hw queue can be remapped after CPU topo changed */
2587 if (!set->tags[hctx_idx] &&
2588 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2590 * If tags initialization fail for some hctx,
2591 * that hctx won't be brought online. In this
2592 * case, remap the current ctx to hctx[0] which
2593 * is guaranteed to always have tags allocated
2595 set->map[j].mq_map[i] = 0;
2598 hctx = blk_mq_map_queue_type(q, j, i);
2599 ctx->hctxs[j] = hctx;
2601 * If the CPU is already set in the mask, then we've
2602 * mapped this one already. This can happen if
2603 * devices share queues across queue maps.
2605 if (cpumask_test_cpu(i, hctx->cpumask))
2608 cpumask_set_cpu(i, hctx->cpumask);
2610 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2611 hctx->ctxs[hctx->nr_ctx++] = ctx;
2614 * If the nr_ctx type overflows, we have exceeded the
2615 * amount of sw queues we can support.
2617 BUG_ON(!hctx->nr_ctx);
2620 for (; j < HCTX_MAX_TYPES; j++)
2621 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2622 HCTX_TYPE_DEFAULT, i);
2625 queue_for_each_hw_ctx(q, hctx, i) {
2627 * If no software queues are mapped to this hardware queue,
2628 * disable it and free the request entries.
2630 if (!hctx->nr_ctx) {
2631 /* Never unmap queue 0. We need it as a
2632 * fallback in case of a new remap fails
2635 if (i && set->tags[i])
2636 blk_mq_free_map_and_requests(set, i);
2642 hctx->tags = set->tags[i];
2643 WARN_ON(!hctx->tags);
2646 * Set the map size to the number of mapped software queues.
2647 * This is more accurate and more efficient than looping
2648 * over all possibly mapped software queues.
2650 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2653 * Initialize batch roundrobin counts
2655 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2656 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2661 * Caller needs to ensure that we're either frozen/quiesced, or that
2662 * the queue isn't live yet.
2664 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2666 struct blk_mq_hw_ctx *hctx;
2669 queue_for_each_hw_ctx(q, hctx, i) {
2671 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2673 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2677 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2680 struct request_queue *q;
2682 lockdep_assert_held(&set->tag_list_lock);
2684 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2685 blk_mq_freeze_queue(q);
2686 queue_set_hctx_shared(q, shared);
2687 blk_mq_unfreeze_queue(q);
2691 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2693 struct blk_mq_tag_set *set = q->tag_set;
2695 mutex_lock(&set->tag_list_lock);
2696 list_del_rcu(&q->tag_set_list);
2697 if (list_is_singular(&set->tag_list)) {
2698 /* just transitioned to unshared */
2699 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2700 /* update existing queue */
2701 blk_mq_update_tag_set_depth(set, false);
2703 mutex_unlock(&set->tag_list_lock);
2704 INIT_LIST_HEAD(&q->tag_set_list);
2707 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2708 struct request_queue *q)
2710 mutex_lock(&set->tag_list_lock);
2713 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2715 if (!list_empty(&set->tag_list) &&
2716 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2717 set->flags |= BLK_MQ_F_TAG_SHARED;
2718 /* update existing queue */
2719 blk_mq_update_tag_set_depth(set, true);
2721 if (set->flags & BLK_MQ_F_TAG_SHARED)
2722 queue_set_hctx_shared(q, true);
2723 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2725 mutex_unlock(&set->tag_list_lock);
2728 /* All allocations will be freed in release handler of q->mq_kobj */
2729 static int blk_mq_alloc_ctxs(struct request_queue *q)
2731 struct blk_mq_ctxs *ctxs;
2734 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2738 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2739 if (!ctxs->queue_ctx)
2742 for_each_possible_cpu(cpu) {
2743 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2747 q->mq_kobj = &ctxs->kobj;
2748 q->queue_ctx = ctxs->queue_ctx;
2757 * It is the actual release handler for mq, but we do it from
2758 * request queue's release handler for avoiding use-after-free
2759 * and headache because q->mq_kobj shouldn't have been introduced,
2760 * but we can't group ctx/kctx kobj without it.
2762 void blk_mq_release(struct request_queue *q)
2764 struct blk_mq_hw_ctx *hctx, *next;
2767 queue_for_each_hw_ctx(q, hctx, i)
2768 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2770 /* all hctx are in .unused_hctx_list now */
2771 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2772 list_del_init(&hctx->hctx_list);
2773 kobject_put(&hctx->kobj);
2776 kfree(q->queue_hw_ctx);
2779 * release .mq_kobj and sw queue's kobject now because
2780 * both share lifetime with request queue.
2782 blk_mq_sysfs_deinit(q);
2785 struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
2788 struct request_queue *uninit_q, *q;
2790 uninit_q = __blk_alloc_queue(set->numa_node);
2792 return ERR_PTR(-ENOMEM);
2793 uninit_q->queuedata = queuedata;
2796 * Initialize the queue without an elevator. device_add_disk() will do
2797 * the initialization.
2799 q = blk_mq_init_allocated_queue(set, uninit_q, false);
2801 blk_cleanup_queue(uninit_q);
2805 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data);
2807 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2809 return blk_mq_init_queue_data(set, NULL);
2811 EXPORT_SYMBOL(blk_mq_init_queue);
2814 * Helper for setting up a queue with mq ops, given queue depth, and
2815 * the passed in mq ops flags.
2817 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2818 const struct blk_mq_ops *ops,
2819 unsigned int queue_depth,
2820 unsigned int set_flags)
2822 struct request_queue *q;
2825 memset(set, 0, sizeof(*set));
2827 set->nr_hw_queues = 1;
2829 set->queue_depth = queue_depth;
2830 set->numa_node = NUMA_NO_NODE;
2831 set->flags = set_flags;
2833 ret = blk_mq_alloc_tag_set(set);
2835 return ERR_PTR(ret);
2837 q = blk_mq_init_queue(set);
2839 blk_mq_free_tag_set(set);
2845 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2847 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2848 struct blk_mq_tag_set *set, struct request_queue *q,
2849 int hctx_idx, int node)
2851 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2853 /* reuse dead hctx first */
2854 spin_lock(&q->unused_hctx_lock);
2855 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2856 if (tmp->numa_node == node) {
2862 list_del_init(&hctx->hctx_list);
2863 spin_unlock(&q->unused_hctx_lock);
2866 hctx = blk_mq_alloc_hctx(q, set, node);
2870 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2876 kobject_put(&hctx->kobj);
2881 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2882 struct request_queue *q)
2885 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2887 if (q->nr_hw_queues < set->nr_hw_queues) {
2888 struct blk_mq_hw_ctx **new_hctxs;
2890 new_hctxs = kcalloc_node(set->nr_hw_queues,
2891 sizeof(*new_hctxs), GFP_KERNEL,
2896 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
2898 q->queue_hw_ctx = new_hctxs;
2903 /* protect against switching io scheduler */
2904 mutex_lock(&q->sysfs_lock);
2905 for (i = 0; i < set->nr_hw_queues; i++) {
2907 struct blk_mq_hw_ctx *hctx;
2909 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2911 * If the hw queue has been mapped to another numa node,
2912 * we need to realloc the hctx. If allocation fails, fallback
2913 * to use the previous one.
2915 if (hctxs[i] && (hctxs[i]->numa_node == node))
2918 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2921 blk_mq_exit_hctx(q, set, hctxs[i], i);
2925 pr_warn("Allocate new hctx on node %d fails,\
2926 fallback to previous one on node %d\n",
2927 node, hctxs[i]->numa_node);
2933 * Increasing nr_hw_queues fails. Free the newly allocated
2934 * hctxs and keep the previous q->nr_hw_queues.
2936 if (i != set->nr_hw_queues) {
2937 j = q->nr_hw_queues;
2941 end = q->nr_hw_queues;
2942 q->nr_hw_queues = set->nr_hw_queues;
2945 for (; j < end; j++) {
2946 struct blk_mq_hw_ctx *hctx = hctxs[j];
2950 blk_mq_free_map_and_requests(set, j);
2951 blk_mq_exit_hctx(q, set, hctx, j);
2955 mutex_unlock(&q->sysfs_lock);
2958 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2959 struct request_queue *q,
2962 /* mark the queue as mq asap */
2963 q->mq_ops = set->ops;
2965 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2966 blk_mq_poll_stats_bkt,
2967 BLK_MQ_POLL_STATS_BKTS, q);
2971 if (blk_mq_alloc_ctxs(q))
2974 /* init q->mq_kobj and sw queues' kobjects */
2975 blk_mq_sysfs_init(q);
2977 INIT_LIST_HEAD(&q->unused_hctx_list);
2978 spin_lock_init(&q->unused_hctx_lock);
2980 blk_mq_realloc_hw_ctxs(set, q);
2981 if (!q->nr_hw_queues)
2984 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2985 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2989 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2990 if (set->nr_maps > HCTX_TYPE_POLL &&
2991 set->map[HCTX_TYPE_POLL].nr_queues)
2992 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2994 q->sg_reserved_size = INT_MAX;
2996 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2997 INIT_LIST_HEAD(&q->requeue_list);
2998 spin_lock_init(&q->requeue_lock);
3000 q->nr_requests = set->queue_depth;
3003 * Default to classic polling
3005 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3007 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3008 blk_mq_add_queue_tag_set(set, q);
3009 blk_mq_map_swqueue(q);
3012 elevator_init_mq(q);
3017 kfree(q->queue_hw_ctx);
3018 q->nr_hw_queues = 0;
3019 blk_mq_sysfs_deinit(q);
3021 blk_stat_free_callback(q->poll_cb);
3025 return ERR_PTR(-ENOMEM);
3027 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3029 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3030 void blk_mq_exit_queue(struct request_queue *q)
3032 struct blk_mq_tag_set *set = q->tag_set;
3034 blk_mq_del_queue_tag_set(q);
3035 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3038 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3042 for (i = 0; i < set->nr_hw_queues; i++)
3043 if (!__blk_mq_alloc_map_and_request(set, i))
3050 blk_mq_free_map_and_requests(set, i);
3056 * Allocate the request maps associated with this tag_set. Note that this
3057 * may reduce the depth asked for, if memory is tight. set->queue_depth
3058 * will be updated to reflect the allocated depth.
3060 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3065 depth = set->queue_depth;
3067 err = __blk_mq_alloc_rq_maps(set);
3071 set->queue_depth >>= 1;
3072 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3076 } while (set->queue_depth);
3078 if (!set->queue_depth || err) {
3079 pr_err("blk-mq: failed to allocate request map\n");
3083 if (depth != set->queue_depth)
3084 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3085 depth, set->queue_depth);
3090 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3093 * blk_mq_map_queues() and multiple .map_queues() implementations
3094 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3095 * number of hardware queues.
3097 if (set->nr_maps == 1)
3098 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3100 if (set->ops->map_queues && !is_kdump_kernel()) {
3104 * transport .map_queues is usually done in the following
3107 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3108 * mask = get_cpu_mask(queue)
3109 * for_each_cpu(cpu, mask)
3110 * set->map[x].mq_map[cpu] = queue;
3113 * When we need to remap, the table has to be cleared for
3114 * killing stale mapping since one CPU may not be mapped
3117 for (i = 0; i < set->nr_maps; i++)
3118 blk_mq_clear_mq_map(&set->map[i]);
3120 return set->ops->map_queues(set);
3122 BUG_ON(set->nr_maps > 1);
3123 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3127 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3128 int cur_nr_hw_queues, int new_nr_hw_queues)
3130 struct blk_mq_tags **new_tags;
3132 if (cur_nr_hw_queues >= new_nr_hw_queues)
3135 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3136 GFP_KERNEL, set->numa_node);
3141 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3142 sizeof(*set->tags));
3144 set->tags = new_tags;
3145 set->nr_hw_queues = new_nr_hw_queues;
3151 * Alloc a tag set to be associated with one or more request queues.
3152 * May fail with EINVAL for various error conditions. May adjust the
3153 * requested depth down, if it's too large. In that case, the set
3154 * value will be stored in set->queue_depth.
3156 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3160 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3162 if (!set->nr_hw_queues)
3164 if (!set->queue_depth)
3166 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3169 if (!set->ops->queue_rq)
3172 if (!set->ops->get_budget ^ !set->ops->put_budget)
3175 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3176 pr_info("blk-mq: reduced tag depth to %u\n",
3178 set->queue_depth = BLK_MQ_MAX_DEPTH;
3183 else if (set->nr_maps > HCTX_MAX_TYPES)
3187 * If a crashdump is active, then we are potentially in a very
3188 * memory constrained environment. Limit us to 1 queue and
3189 * 64 tags to prevent using too much memory.
3191 if (is_kdump_kernel()) {
3192 set->nr_hw_queues = 1;
3194 set->queue_depth = min(64U, set->queue_depth);
3197 * There is no use for more h/w queues than cpus if we just have
3200 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3201 set->nr_hw_queues = nr_cpu_ids;
3203 if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0)
3207 for (i = 0; i < set->nr_maps; i++) {
3208 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3209 sizeof(set->map[i].mq_map[0]),
3210 GFP_KERNEL, set->numa_node);
3211 if (!set->map[i].mq_map)
3212 goto out_free_mq_map;
3213 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3216 ret = blk_mq_update_queue_map(set);
3218 goto out_free_mq_map;
3220 ret = blk_mq_alloc_map_and_requests(set);
3222 goto out_free_mq_map;
3224 mutex_init(&set->tag_list_lock);
3225 INIT_LIST_HEAD(&set->tag_list);
3230 for (i = 0; i < set->nr_maps; i++) {
3231 kfree(set->map[i].mq_map);
3232 set->map[i].mq_map = NULL;
3238 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3240 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3244 for (i = 0; i < set->nr_hw_queues; i++)
3245 blk_mq_free_map_and_requests(set, i);
3247 for (j = 0; j < set->nr_maps; j++) {
3248 kfree(set->map[j].mq_map);
3249 set->map[j].mq_map = NULL;
3255 EXPORT_SYMBOL(blk_mq_free_tag_set);
3257 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3259 struct blk_mq_tag_set *set = q->tag_set;
3260 struct blk_mq_hw_ctx *hctx;
3266 if (q->nr_requests == nr)
3269 blk_mq_freeze_queue(q);
3270 blk_mq_quiesce_queue(q);
3273 queue_for_each_hw_ctx(q, hctx, i) {
3277 * If we're using an MQ scheduler, just update the scheduler
3278 * queue depth. This is similar to what the old code would do.
3280 if (!hctx->sched_tags) {
3281 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3284 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3289 if (q->elevator && q->elevator->type->ops.depth_updated)
3290 q->elevator->type->ops.depth_updated(hctx);
3294 q->nr_requests = nr;
3296 blk_mq_unquiesce_queue(q);
3297 blk_mq_unfreeze_queue(q);
3303 * request_queue and elevator_type pair.
3304 * It is just used by __blk_mq_update_nr_hw_queues to cache
3305 * the elevator_type associated with a request_queue.
3307 struct blk_mq_qe_pair {
3308 struct list_head node;
3309 struct request_queue *q;
3310 struct elevator_type *type;
3314 * Cache the elevator_type in qe pair list and switch the
3315 * io scheduler to 'none'
3317 static bool blk_mq_elv_switch_none(struct list_head *head,
3318 struct request_queue *q)
3320 struct blk_mq_qe_pair *qe;
3325 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3329 INIT_LIST_HEAD(&qe->node);
3331 qe->type = q->elevator->type;
3332 list_add(&qe->node, head);
3334 mutex_lock(&q->sysfs_lock);
3336 * After elevator_switch_mq, the previous elevator_queue will be
3337 * released by elevator_release. The reference of the io scheduler
3338 * module get by elevator_get will also be put. So we need to get
3339 * a reference of the io scheduler module here to prevent it to be
3342 __module_get(qe->type->elevator_owner);
3343 elevator_switch_mq(q, NULL);
3344 mutex_unlock(&q->sysfs_lock);
3349 static void blk_mq_elv_switch_back(struct list_head *head,
3350 struct request_queue *q)
3352 struct blk_mq_qe_pair *qe;
3353 struct elevator_type *t = NULL;
3355 list_for_each_entry(qe, head, node)
3364 list_del(&qe->node);
3367 mutex_lock(&q->sysfs_lock);
3368 elevator_switch_mq(q, t);
3369 mutex_unlock(&q->sysfs_lock);
3372 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3375 struct request_queue *q;
3377 int prev_nr_hw_queues;
3379 lockdep_assert_held(&set->tag_list_lock);
3381 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3382 nr_hw_queues = nr_cpu_ids;
3383 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3386 list_for_each_entry(q, &set->tag_list, tag_set_list)
3387 blk_mq_freeze_queue(q);
3389 * Switch IO scheduler to 'none', cleaning up the data associated
3390 * with the previous scheduler. We will switch back once we are done
3391 * updating the new sw to hw queue mappings.
3393 list_for_each_entry(q, &set->tag_list, tag_set_list)
3394 if (!blk_mq_elv_switch_none(&head, q))
3397 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3398 blk_mq_debugfs_unregister_hctxs(q);
3399 blk_mq_sysfs_unregister(q);
3402 prev_nr_hw_queues = set->nr_hw_queues;
3403 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3407 set->nr_hw_queues = nr_hw_queues;
3409 blk_mq_update_queue_map(set);
3410 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3411 blk_mq_realloc_hw_ctxs(set, q);
3412 if (q->nr_hw_queues != set->nr_hw_queues) {
3413 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3414 nr_hw_queues, prev_nr_hw_queues);
3415 set->nr_hw_queues = prev_nr_hw_queues;
3416 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3419 blk_mq_map_swqueue(q);
3423 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3424 blk_mq_sysfs_register(q);
3425 blk_mq_debugfs_register_hctxs(q);
3429 list_for_each_entry(q, &set->tag_list, tag_set_list)
3430 blk_mq_elv_switch_back(&head, q);
3432 list_for_each_entry(q, &set->tag_list, tag_set_list)
3433 blk_mq_unfreeze_queue(q);
3436 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3438 mutex_lock(&set->tag_list_lock);
3439 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3440 mutex_unlock(&set->tag_list_lock);
3442 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3444 /* Enable polling stats and return whether they were already enabled. */
3445 static bool blk_poll_stats_enable(struct request_queue *q)
3447 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3448 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3450 blk_stat_add_callback(q, q->poll_cb);
3454 static void blk_mq_poll_stats_start(struct request_queue *q)
3457 * We don't arm the callback if polling stats are not enabled or the
3458 * callback is already active.
3460 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3461 blk_stat_is_active(q->poll_cb))
3464 blk_stat_activate_msecs(q->poll_cb, 100);
3467 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3469 struct request_queue *q = cb->data;
3472 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3473 if (cb->stat[bucket].nr_samples)
3474 q->poll_stat[bucket] = cb->stat[bucket];
3478 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3481 unsigned long ret = 0;
3485 * If stats collection isn't on, don't sleep but turn it on for
3488 if (!blk_poll_stats_enable(q))
3492 * As an optimistic guess, use half of the mean service time
3493 * for this type of request. We can (and should) make this smarter.
3494 * For instance, if the completion latencies are tight, we can
3495 * get closer than just half the mean. This is especially
3496 * important on devices where the completion latencies are longer
3497 * than ~10 usec. We do use the stats for the relevant IO size
3498 * if available which does lead to better estimates.
3500 bucket = blk_mq_poll_stats_bkt(rq);
3504 if (q->poll_stat[bucket].nr_samples)
3505 ret = (q->poll_stat[bucket].mean + 1) / 2;
3510 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3513 struct hrtimer_sleeper hs;
3514 enum hrtimer_mode mode;
3518 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3522 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3524 * 0: use half of prev avg
3525 * >0: use this specific value
3527 if (q->poll_nsec > 0)
3528 nsecs = q->poll_nsec;
3530 nsecs = blk_mq_poll_nsecs(q, rq);
3535 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3538 * This will be replaced with the stats tracking code, using
3539 * 'avg_completion_time / 2' as the pre-sleep target.
3543 mode = HRTIMER_MODE_REL;
3544 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3545 hrtimer_set_expires(&hs.timer, kt);
3548 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3550 set_current_state(TASK_UNINTERRUPTIBLE);
3551 hrtimer_sleeper_start_expires(&hs, mode);
3554 hrtimer_cancel(&hs.timer);
3555 mode = HRTIMER_MODE_ABS;
3556 } while (hs.task && !signal_pending(current));
3558 __set_current_state(TASK_RUNNING);
3559 destroy_hrtimer_on_stack(&hs.timer);
3563 static bool blk_mq_poll_hybrid(struct request_queue *q,
3564 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3568 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3571 if (!blk_qc_t_is_internal(cookie))
3572 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3574 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3576 * With scheduling, if the request has completed, we'll
3577 * get a NULL return here, as we clear the sched tag when
3578 * that happens. The request still remains valid, like always,
3579 * so we should be safe with just the NULL check.
3585 return blk_mq_poll_hybrid_sleep(q, rq);
3589 * blk_poll - poll for IO completions
3591 * @cookie: cookie passed back at IO submission time
3592 * @spin: whether to spin for completions
3595 * Poll for completions on the passed in queue. Returns number of
3596 * completed entries found. If @spin is true, then blk_poll will continue
3597 * looping until at least one completion is found, unless the task is
3598 * otherwise marked running (or we need to reschedule).
3600 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3602 struct blk_mq_hw_ctx *hctx;
3605 if (!blk_qc_t_valid(cookie) ||
3606 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3610 blk_flush_plug_list(current->plug, false);
3612 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3615 * If we sleep, have the caller restart the poll loop to reset
3616 * the state. Like for the other success return cases, the
3617 * caller is responsible for checking if the IO completed. If
3618 * the IO isn't complete, we'll get called again and will go
3619 * straight to the busy poll loop.
3621 if (blk_mq_poll_hybrid(q, hctx, cookie))
3624 hctx->poll_considered++;
3626 state = current->state;
3630 hctx->poll_invoked++;
3632 ret = q->mq_ops->poll(hctx);
3634 hctx->poll_success++;
3635 __set_current_state(TASK_RUNNING);
3639 if (signal_pending_state(state, current))
3640 __set_current_state(TASK_RUNNING);
3642 if (current->state == TASK_RUNNING)
3644 if (ret < 0 || !spin)
3647 } while (!need_resched());
3649 __set_current_state(TASK_RUNNING);
3652 EXPORT_SYMBOL_GPL(blk_poll);
3654 unsigned int blk_mq_rq_cpu(struct request *rq)
3656 return rq->mq_ctx->cpu;
3658 EXPORT_SYMBOL(blk_mq_rq_cpu);
3660 static int __init blk_mq_init(void)
3662 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3663 blk_mq_hctx_notify_dead);
3666 subsys_initcall(blk_mq_init);