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, 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) {
281 rq->tag = BLK_MQ_NO_TAG;
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 = BLK_MQ_NO_TAG;
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;
298 rq->cmd_flags = data->cmd_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(data->cmd_flags)]++;
331 refcount_set(&rq->ref, 1);
333 if (!op_is_flush(data->cmd_flags)) {
334 struct elevator_queue *e = data->q->elevator;
337 if (e && e->type->ops.prepare_request) {
338 if (e->type->icq_cache)
339 blk_mq_sched_assign_ioc(rq);
341 e->type->ops.prepare_request(rq);
342 rq->rq_flags |= RQF_ELVPRIV;
346 data->hctx->queued++;
350 static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data)
352 struct request_queue *q = data->q;
353 struct elevator_queue *e = q->elevator;
355 bool clear_ctx_on_error = false;
356 u64 alloc_time_ns = 0;
358 /* alloc_time includes depth and tag waits */
359 if (blk_queue_rq_alloc_time(q))
360 alloc_time_ns = ktime_get_ns();
362 if (likely(!data->ctx)) {
363 data->ctx = blk_mq_get_ctx(q);
364 clear_ctx_on_error = true;
366 if (likely(!data->hctx))
367 data->hctx = blk_mq_map_queue(q, data->cmd_flags,
369 if (data->cmd_flags & REQ_NOWAIT)
370 data->flags |= BLK_MQ_REQ_NOWAIT;
373 data->flags |= BLK_MQ_REQ_INTERNAL;
376 * Flush requests are special and go directly to the
377 * dispatch list. Don't include reserved tags in the
378 * limiting, as it isn't useful.
380 if (!op_is_flush(data->cmd_flags) &&
381 e->type->ops.limit_depth &&
382 !(data->flags & BLK_MQ_REQ_RESERVED))
383 e->type->ops.limit_depth(data->cmd_flags, data);
385 blk_mq_tag_busy(data->hctx);
388 tag = blk_mq_get_tag(data);
389 if (tag == BLK_MQ_NO_TAG) {
390 if (clear_ctx_on_error)
395 return blk_mq_rq_ctx_init(data, tag, alloc_time_ns);
398 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
399 blk_mq_req_flags_t flags)
401 struct blk_mq_alloc_data data = {
409 ret = blk_queue_enter(q, flags);
413 rq = __blk_mq_alloc_request(&data);
417 rq->__sector = (sector_t) -1;
418 rq->bio = rq->biotail = NULL;
422 return ERR_PTR(-EWOULDBLOCK);
424 EXPORT_SYMBOL(blk_mq_alloc_request);
426 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
427 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
429 struct blk_mq_alloc_data data = {
439 * If the tag allocator sleeps we could get an allocation for a
440 * different hardware context. No need to complicate the low level
441 * allocator for this for the rare use case of a command tied to
444 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
445 return ERR_PTR(-EINVAL);
447 if (hctx_idx >= q->nr_hw_queues)
448 return ERR_PTR(-EIO);
450 ret = blk_queue_enter(q, flags);
455 * Check if the hardware context is actually mapped to anything.
456 * If not tell the caller that it should skip this queue.
459 data.hctx = q->queue_hw_ctx[hctx_idx];
460 if (!blk_mq_hw_queue_mapped(data.hctx))
462 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
463 data.ctx = __blk_mq_get_ctx(q, cpu);
466 rq = __blk_mq_alloc_request(&data);
474 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
476 static void __blk_mq_free_request(struct request *rq)
478 struct request_queue *q = rq->q;
479 struct blk_mq_ctx *ctx = rq->mq_ctx;
480 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
481 const int sched_tag = rq->internal_tag;
483 blk_crypto_free_request(rq);
484 blk_pm_mark_last_busy(rq);
486 if (rq->tag != BLK_MQ_NO_TAG)
487 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
488 if (sched_tag != BLK_MQ_NO_TAG)
489 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
490 blk_mq_sched_restart(hctx);
494 void blk_mq_free_request(struct request *rq)
496 struct request_queue *q = rq->q;
497 struct elevator_queue *e = q->elevator;
498 struct blk_mq_ctx *ctx = rq->mq_ctx;
499 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
501 if (rq->rq_flags & RQF_ELVPRIV) {
502 if (e && e->type->ops.finish_request)
503 e->type->ops.finish_request(rq);
505 put_io_context(rq->elv.icq->ioc);
510 ctx->rq_completed[rq_is_sync(rq)]++;
511 if (rq->rq_flags & RQF_MQ_INFLIGHT)
512 atomic_dec(&hctx->nr_active);
514 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
515 laptop_io_completion(q->backing_dev_info);
519 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
520 if (refcount_dec_and_test(&rq->ref))
521 __blk_mq_free_request(rq);
523 EXPORT_SYMBOL_GPL(blk_mq_free_request);
525 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
529 if (blk_mq_need_time_stamp(rq))
530 now = ktime_get_ns();
532 if (rq->rq_flags & RQF_STATS) {
533 blk_mq_poll_stats_start(rq->q);
534 blk_stat_add(rq, now);
537 if (rq->internal_tag != BLK_MQ_NO_TAG)
538 blk_mq_sched_completed_request(rq, now);
540 blk_account_io_done(rq, now);
543 rq_qos_done(rq->q, rq);
544 rq->end_io(rq, error);
546 blk_mq_free_request(rq);
549 EXPORT_SYMBOL(__blk_mq_end_request);
551 void blk_mq_end_request(struct request *rq, blk_status_t error)
553 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
555 __blk_mq_end_request(rq, error);
557 EXPORT_SYMBOL(blk_mq_end_request);
559 static void __blk_mq_complete_request_remote(void *data)
561 struct request *rq = data;
562 struct request_queue *q = rq->q;
564 q->mq_ops->complete(rq);
568 * blk_mq_force_complete_rq() - Force complete the request, bypassing any error
569 * injection that could drop the completion.
570 * @rq: Request to be force completed
572 * Drivers should use blk_mq_complete_request() to complete requests in their
573 * normal IO path. For timeout error recovery, drivers may call this forced
574 * completion routine after they've reclaimed timed out requests to bypass
575 * potentially subsequent fake timeouts.
577 void blk_mq_force_complete_rq(struct request *rq)
579 struct blk_mq_ctx *ctx = rq->mq_ctx;
580 struct request_queue *q = rq->q;
584 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
586 * Most of single queue controllers, there is only one irq vector
587 * for handling IO completion, and the only irq's affinity is set
588 * as all possible CPUs. On most of ARCHs, this affinity means the
589 * irq is handled on one specific CPU.
591 * So complete IO reqeust in softirq context in case of single queue
592 * for not degrading IO performance by irqsoff latency.
594 if (q->nr_hw_queues == 1) {
595 __blk_complete_request(rq);
600 * For a polled request, always complete locallly, it's pointless
601 * to redirect the completion.
603 if ((rq->cmd_flags & REQ_HIPRI) ||
604 !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
605 q->mq_ops->complete(rq);
610 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
611 shared = cpus_share_cache(cpu, ctx->cpu);
613 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
614 rq->csd.func = __blk_mq_complete_request_remote;
617 smp_call_function_single_async(ctx->cpu, &rq->csd);
619 q->mq_ops->complete(rq);
623 EXPORT_SYMBOL_GPL(blk_mq_force_complete_rq);
625 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
626 __releases(hctx->srcu)
628 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
631 srcu_read_unlock(hctx->srcu, srcu_idx);
634 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
635 __acquires(hctx->srcu)
637 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
638 /* shut up gcc false positive */
642 *srcu_idx = srcu_read_lock(hctx->srcu);
646 * blk_mq_complete_request - end I/O on a request
647 * @rq: the request being processed
650 * Ends all I/O on a request. It does not handle partial completions.
651 * The actual completion happens out-of-order, through a IPI handler.
653 bool blk_mq_complete_request(struct request *rq)
655 if (unlikely(blk_should_fake_timeout(rq->q)))
657 blk_mq_force_complete_rq(rq);
660 EXPORT_SYMBOL(blk_mq_complete_request);
663 * blk_mq_start_request - Start processing a request
664 * @rq: Pointer to request to be started
666 * Function used by device drivers to notify the block layer that a request
667 * is going to be processed now, so blk layer can do proper initializations
668 * such as starting the timeout timer.
670 void blk_mq_start_request(struct request *rq)
672 struct request_queue *q = rq->q;
674 trace_block_rq_issue(q, rq);
676 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
677 rq->io_start_time_ns = ktime_get_ns();
678 rq->stats_sectors = blk_rq_sectors(rq);
679 rq->rq_flags |= RQF_STATS;
683 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
686 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
688 #ifdef CONFIG_BLK_DEV_INTEGRITY
689 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
690 q->integrity.profile->prepare_fn(rq);
693 EXPORT_SYMBOL(blk_mq_start_request);
695 static void __blk_mq_requeue_request(struct request *rq)
697 struct request_queue *q = rq->q;
699 blk_mq_put_driver_tag(rq);
701 trace_block_rq_requeue(q, rq);
702 rq_qos_requeue(q, rq);
704 if (blk_mq_request_started(rq)) {
705 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
706 rq->rq_flags &= ~RQF_TIMED_OUT;
710 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
712 __blk_mq_requeue_request(rq);
714 /* this request will be re-inserted to io scheduler queue */
715 blk_mq_sched_requeue_request(rq);
717 BUG_ON(!list_empty(&rq->queuelist));
718 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
720 EXPORT_SYMBOL(blk_mq_requeue_request);
722 static void blk_mq_requeue_work(struct work_struct *work)
724 struct request_queue *q =
725 container_of(work, struct request_queue, requeue_work.work);
727 struct request *rq, *next;
729 spin_lock_irq(&q->requeue_lock);
730 list_splice_init(&q->requeue_list, &rq_list);
731 spin_unlock_irq(&q->requeue_lock);
733 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
734 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
737 rq->rq_flags &= ~RQF_SOFTBARRIER;
738 list_del_init(&rq->queuelist);
740 * If RQF_DONTPREP, rq has contained some driver specific
741 * data, so insert it to hctx dispatch list to avoid any
744 if (rq->rq_flags & RQF_DONTPREP)
745 blk_mq_request_bypass_insert(rq, false, false);
747 blk_mq_sched_insert_request(rq, true, false, false);
750 while (!list_empty(&rq_list)) {
751 rq = list_entry(rq_list.next, struct request, queuelist);
752 list_del_init(&rq->queuelist);
753 blk_mq_sched_insert_request(rq, false, false, false);
756 blk_mq_run_hw_queues(q, false);
759 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
760 bool kick_requeue_list)
762 struct request_queue *q = rq->q;
766 * We abuse this flag that is otherwise used by the I/O scheduler to
767 * request head insertion from the workqueue.
769 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
771 spin_lock_irqsave(&q->requeue_lock, flags);
773 rq->rq_flags |= RQF_SOFTBARRIER;
774 list_add(&rq->queuelist, &q->requeue_list);
776 list_add_tail(&rq->queuelist, &q->requeue_list);
778 spin_unlock_irqrestore(&q->requeue_lock, flags);
780 if (kick_requeue_list)
781 blk_mq_kick_requeue_list(q);
784 void blk_mq_kick_requeue_list(struct request_queue *q)
786 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
788 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
790 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
793 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
794 msecs_to_jiffies(msecs));
796 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
798 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
800 if (tag < tags->nr_tags) {
801 prefetch(tags->rqs[tag]);
802 return tags->rqs[tag];
807 EXPORT_SYMBOL(blk_mq_tag_to_rq);
809 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
810 void *priv, bool reserved)
813 * If we find a request that is inflight and the queue matches,
814 * we know the queue is busy. Return false to stop the iteration.
816 if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
826 bool blk_mq_queue_inflight(struct request_queue *q)
830 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
833 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
835 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
837 req->rq_flags |= RQF_TIMED_OUT;
838 if (req->q->mq_ops->timeout) {
839 enum blk_eh_timer_return ret;
841 ret = req->q->mq_ops->timeout(req, reserved);
842 if (ret == BLK_EH_DONE)
844 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
850 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
852 unsigned long deadline;
854 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
856 if (rq->rq_flags & RQF_TIMED_OUT)
859 deadline = READ_ONCE(rq->deadline);
860 if (time_after_eq(jiffies, deadline))
865 else if (time_after(*next, deadline))
870 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
871 struct request *rq, void *priv, bool reserved)
873 unsigned long *next = priv;
876 * Just do a quick check if it is expired before locking the request in
877 * so we're not unnecessarilly synchronizing across CPUs.
879 if (!blk_mq_req_expired(rq, next))
883 * We have reason to believe the request may be expired. Take a
884 * reference on the request to lock this request lifetime into its
885 * currently allocated context to prevent it from being reallocated in
886 * the event the completion by-passes this timeout handler.
888 * If the reference was already released, then the driver beat the
889 * timeout handler to posting a natural completion.
891 if (!refcount_inc_not_zero(&rq->ref))
895 * The request is now locked and cannot be reallocated underneath the
896 * timeout handler's processing. Re-verify this exact request is truly
897 * expired; if it is not expired, then the request was completed and
898 * reallocated as a new request.
900 if (blk_mq_req_expired(rq, next))
901 blk_mq_rq_timed_out(rq, reserved);
903 if (is_flush_rq(rq, hctx))
905 else if (refcount_dec_and_test(&rq->ref))
906 __blk_mq_free_request(rq);
911 static void blk_mq_timeout_work(struct work_struct *work)
913 struct request_queue *q =
914 container_of(work, struct request_queue, timeout_work);
915 unsigned long next = 0;
916 struct blk_mq_hw_ctx *hctx;
919 /* A deadlock might occur if a request is stuck requiring a
920 * timeout at the same time a queue freeze is waiting
921 * completion, since the timeout code would not be able to
922 * acquire the queue reference here.
924 * That's why we don't use blk_queue_enter here; instead, we use
925 * percpu_ref_tryget directly, because we need to be able to
926 * obtain a reference even in the short window between the queue
927 * starting to freeze, by dropping the first reference in
928 * blk_freeze_queue_start, and the moment the last request is
929 * consumed, marked by the instant q_usage_counter reaches
932 if (!percpu_ref_tryget(&q->q_usage_counter))
935 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
938 mod_timer(&q->timeout, next);
941 * Request timeouts are handled as a forward rolling timer. If
942 * we end up here it means that no requests are pending and
943 * also that no request has been pending for a while. Mark
946 queue_for_each_hw_ctx(q, hctx, i) {
947 /* the hctx may be unmapped, so check it here */
948 if (blk_mq_hw_queue_mapped(hctx))
949 blk_mq_tag_idle(hctx);
955 struct flush_busy_ctx_data {
956 struct blk_mq_hw_ctx *hctx;
957 struct list_head *list;
960 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
962 struct flush_busy_ctx_data *flush_data = data;
963 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
964 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
965 enum hctx_type type = hctx->type;
967 spin_lock(&ctx->lock);
968 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
969 sbitmap_clear_bit(sb, bitnr);
970 spin_unlock(&ctx->lock);
975 * Process software queues that have been marked busy, splicing them
976 * to the for-dispatch
978 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
980 struct flush_busy_ctx_data data = {
985 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
987 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
989 struct dispatch_rq_data {
990 struct blk_mq_hw_ctx *hctx;
994 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
997 struct dispatch_rq_data *dispatch_data = data;
998 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
999 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1000 enum hctx_type type = hctx->type;
1002 spin_lock(&ctx->lock);
1003 if (!list_empty(&ctx->rq_lists[type])) {
1004 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1005 list_del_init(&dispatch_data->rq->queuelist);
1006 if (list_empty(&ctx->rq_lists[type]))
1007 sbitmap_clear_bit(sb, bitnr);
1009 spin_unlock(&ctx->lock);
1011 return !dispatch_data->rq;
1014 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1015 struct blk_mq_ctx *start)
1017 unsigned off = start ? start->index_hw[hctx->type] : 0;
1018 struct dispatch_rq_data data = {
1023 __sbitmap_for_each_set(&hctx->ctx_map, off,
1024 dispatch_rq_from_ctx, &data);
1029 static inline unsigned int queued_to_index(unsigned int queued)
1034 return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1037 bool blk_mq_get_driver_tag(struct request *rq)
1039 struct blk_mq_alloc_data data = {
1041 .hctx = rq->mq_hctx,
1042 .flags = BLK_MQ_REQ_NOWAIT,
1043 .cmd_flags = rq->cmd_flags,
1047 if (rq->tag != BLK_MQ_NO_TAG)
1050 if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1051 data.flags |= BLK_MQ_REQ_RESERVED;
1053 shared = blk_mq_tag_busy(data.hctx);
1054 rq->tag = blk_mq_get_tag(&data);
1057 rq->rq_flags |= RQF_MQ_INFLIGHT;
1058 atomic_inc(&data.hctx->nr_active);
1060 data.hctx->tags->rqs[rq->tag] = rq;
1063 return rq->tag != BLK_MQ_NO_TAG;
1066 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1067 int flags, void *key)
1069 struct blk_mq_hw_ctx *hctx;
1071 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1073 spin_lock(&hctx->dispatch_wait_lock);
1074 if (!list_empty(&wait->entry)) {
1075 struct sbitmap_queue *sbq;
1077 list_del_init(&wait->entry);
1078 sbq = &hctx->tags->bitmap_tags;
1079 atomic_dec(&sbq->ws_active);
1081 spin_unlock(&hctx->dispatch_wait_lock);
1083 blk_mq_run_hw_queue(hctx, true);
1088 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1089 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1090 * restart. For both cases, take care to check the condition again after
1091 * marking us as waiting.
1093 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1096 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1097 struct wait_queue_head *wq;
1098 wait_queue_entry_t *wait;
1101 if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1102 blk_mq_sched_mark_restart_hctx(hctx);
1105 * It's possible that a tag was freed in the window between the
1106 * allocation failure and adding the hardware queue to the wait
1109 * Don't clear RESTART here, someone else could have set it.
1110 * At most this will cost an extra queue run.
1112 return blk_mq_get_driver_tag(rq);
1115 wait = &hctx->dispatch_wait;
1116 if (!list_empty_careful(&wait->entry))
1119 wq = &bt_wait_ptr(sbq, hctx)->wait;
1121 spin_lock_irq(&wq->lock);
1122 spin_lock(&hctx->dispatch_wait_lock);
1123 if (!list_empty(&wait->entry)) {
1124 spin_unlock(&hctx->dispatch_wait_lock);
1125 spin_unlock_irq(&wq->lock);
1129 atomic_inc(&sbq->ws_active);
1130 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1131 __add_wait_queue(wq, wait);
1134 * It's possible that a tag was freed in the window between the
1135 * allocation failure and adding the hardware queue to the wait
1138 ret = blk_mq_get_driver_tag(rq);
1140 spin_unlock(&hctx->dispatch_wait_lock);
1141 spin_unlock_irq(&wq->lock);
1146 * We got a tag, remove ourselves from the wait queue to ensure
1147 * someone else gets the wakeup.
1149 list_del_init(&wait->entry);
1150 atomic_dec(&sbq->ws_active);
1151 spin_unlock(&hctx->dispatch_wait_lock);
1152 spin_unlock_irq(&wq->lock);
1157 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1158 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1160 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1161 * - EWMA is one simple way to compute running average value
1162 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1163 * - take 4 as factor for avoiding to get too small(0) result, and this
1164 * factor doesn't matter because EWMA decreases exponentially
1166 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1170 if (hctx->queue->elevator)
1173 ewma = hctx->dispatch_busy;
1178 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1180 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1181 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1183 hctx->dispatch_busy = ewma;
1186 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1188 static void blk_mq_handle_dev_resource(struct request *rq,
1189 struct list_head *list)
1191 struct request *next =
1192 list_first_entry_or_null(list, struct request, queuelist);
1195 * If an I/O scheduler has been configured and we got a driver tag for
1196 * the next request already, free it.
1199 blk_mq_put_driver_tag(next);
1201 list_add(&rq->queuelist, list);
1202 __blk_mq_requeue_request(rq);
1205 static void blk_mq_handle_zone_resource(struct request *rq,
1206 struct list_head *zone_list)
1209 * If we end up here it is because we cannot dispatch a request to a
1210 * specific zone due to LLD level zone-write locking or other zone
1211 * related resource not being available. In this case, set the request
1212 * aside in zone_list for retrying it later.
1214 list_add(&rq->queuelist, zone_list);
1215 __blk_mq_requeue_request(rq);
1219 * Returns true if we did some work AND can potentially do more.
1221 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1224 struct blk_mq_hw_ctx *hctx;
1225 struct request *rq, *nxt;
1226 bool no_tag = false;
1228 blk_status_t ret = BLK_STS_OK;
1229 bool no_budget_avail = false;
1230 LIST_HEAD(zone_list);
1232 if (list_empty(list))
1235 WARN_ON(!list_is_singular(list) && got_budget);
1238 * Now process all the entries, sending them to the driver.
1240 errors = queued = 0;
1242 struct blk_mq_queue_data bd;
1244 rq = list_first_entry(list, struct request, queuelist);
1247 if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) {
1248 blk_mq_put_driver_tag(rq);
1249 no_budget_avail = true;
1253 if (!blk_mq_get_driver_tag(rq)) {
1255 * The initial allocation attempt failed, so we need to
1256 * rerun the hardware queue when a tag is freed. The
1257 * waitqueue takes care of that. If the queue is run
1258 * before we add this entry back on the dispatch list,
1259 * we'll re-run it below.
1261 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1262 blk_mq_put_dispatch_budget(hctx);
1264 * For non-shared tags, the RESTART check
1267 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1273 list_del_init(&rq->queuelist);
1278 * Flag last if we have no more requests, or if we have more
1279 * but can't assign a driver tag to it.
1281 if (list_empty(list))
1284 nxt = list_first_entry(list, struct request, queuelist);
1285 bd.last = !blk_mq_get_driver_tag(nxt);
1288 ret = q->mq_ops->queue_rq(hctx, &bd);
1289 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1290 blk_mq_handle_dev_resource(rq, list);
1292 } else if (ret == BLK_STS_ZONE_RESOURCE) {
1294 * Move the request to zone_list and keep going through
1295 * the dispatch list to find more requests the drive can
1298 blk_mq_handle_zone_resource(rq, &zone_list);
1299 if (list_empty(list))
1304 if (unlikely(ret != BLK_STS_OK)) {
1306 blk_mq_end_request(rq, BLK_STS_IOERR);
1311 } while (!list_empty(list));
1313 if (!list_empty(&zone_list))
1314 list_splice_tail_init(&zone_list, list);
1316 hctx->dispatched[queued_to_index(queued)]++;
1319 * Any items that need requeuing? Stuff them into hctx->dispatch,
1320 * that is where we will continue on next queue run.
1322 if (!list_empty(list)) {
1326 * If we didn't flush the entire list, we could have told
1327 * the driver there was more coming, but that turned out to
1330 if (q->mq_ops->commit_rqs && queued)
1331 q->mq_ops->commit_rqs(hctx);
1333 spin_lock(&hctx->lock);
1334 list_splice_tail_init(list, &hctx->dispatch);
1335 spin_unlock(&hctx->lock);
1338 * If SCHED_RESTART was set by the caller of this function and
1339 * it is no longer set that means that it was cleared by another
1340 * thread and hence that a queue rerun is needed.
1342 * If 'no_tag' is set, that means that we failed getting
1343 * a driver tag with an I/O scheduler attached. If our dispatch
1344 * waitqueue is no longer active, ensure that we run the queue
1345 * AFTER adding our entries back to the list.
1347 * If no I/O scheduler has been configured it is possible that
1348 * the hardware queue got stopped and restarted before requests
1349 * were pushed back onto the dispatch list. Rerun the queue to
1350 * avoid starvation. Notes:
1351 * - blk_mq_run_hw_queue() checks whether or not a queue has
1352 * been stopped before rerunning a queue.
1353 * - Some but not all block drivers stop a queue before
1354 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1357 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1358 * bit is set, run queue after a delay to avoid IO stalls
1359 * that could otherwise occur if the queue is idle. We'll do
1360 * similar if we couldn't get budget and SCHED_RESTART is set.
1362 needs_restart = blk_mq_sched_needs_restart(hctx);
1363 if (!needs_restart ||
1364 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1365 blk_mq_run_hw_queue(hctx, true);
1366 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1368 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1370 blk_mq_update_dispatch_busy(hctx, true);
1373 blk_mq_update_dispatch_busy(hctx, false);
1376 * If the host/device is unable to accept more work, inform the
1379 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1382 return (queued + errors) != 0;
1386 * __blk_mq_run_hw_queue - Run a hardware queue.
1387 * @hctx: Pointer to the hardware queue to run.
1389 * Send pending requests to the hardware.
1391 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1396 * We should be running this queue from one of the CPUs that
1399 * There are at least two related races now between setting
1400 * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1401 * __blk_mq_run_hw_queue():
1403 * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1404 * but later it becomes online, then this warning is harmless
1407 * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1408 * but later it becomes offline, then the warning can't be
1409 * triggered, and we depend on blk-mq timeout handler to
1410 * handle dispatched requests to this hctx
1412 if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1413 cpu_online(hctx->next_cpu)) {
1414 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1415 raw_smp_processor_id(),
1416 cpumask_empty(hctx->cpumask) ? "inactive": "active");
1421 * We can't run the queue inline with ints disabled. Ensure that
1422 * we catch bad users of this early.
1424 WARN_ON_ONCE(in_interrupt());
1426 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1428 hctx_lock(hctx, &srcu_idx);
1429 blk_mq_sched_dispatch_requests(hctx);
1430 hctx_unlock(hctx, srcu_idx);
1433 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1435 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1437 if (cpu >= nr_cpu_ids)
1438 cpu = cpumask_first(hctx->cpumask);
1443 * It'd be great if the workqueue API had a way to pass
1444 * in a mask and had some smarts for more clever placement.
1445 * For now we just round-robin here, switching for every
1446 * BLK_MQ_CPU_WORK_BATCH queued items.
1448 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1451 int next_cpu = hctx->next_cpu;
1453 if (hctx->queue->nr_hw_queues == 1)
1454 return WORK_CPU_UNBOUND;
1456 if (--hctx->next_cpu_batch <= 0) {
1458 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1460 if (next_cpu >= nr_cpu_ids)
1461 next_cpu = blk_mq_first_mapped_cpu(hctx);
1462 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1466 * Do unbound schedule if we can't find a online CPU for this hctx,
1467 * and it should only happen in the path of handling CPU DEAD.
1469 if (!cpu_online(next_cpu)) {
1476 * Make sure to re-select CPU next time once after CPUs
1477 * in hctx->cpumask become online again.
1479 hctx->next_cpu = next_cpu;
1480 hctx->next_cpu_batch = 1;
1481 return WORK_CPU_UNBOUND;
1484 hctx->next_cpu = next_cpu;
1489 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1490 * @hctx: Pointer to the hardware queue to run.
1491 * @async: If we want to run the queue asynchronously.
1492 * @msecs: Microseconds of delay to wait before running the queue.
1494 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1495 * with a delay of @msecs.
1497 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1498 unsigned long msecs)
1500 if (unlikely(blk_mq_hctx_stopped(hctx)))
1503 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1504 int cpu = get_cpu();
1505 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1506 __blk_mq_run_hw_queue(hctx);
1514 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1515 msecs_to_jiffies(msecs));
1519 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1520 * @hctx: Pointer to the hardware queue to run.
1521 * @msecs: Microseconds of delay to wait before running the queue.
1523 * Run a hardware queue asynchronously with a delay of @msecs.
1525 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1527 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1529 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1532 * blk_mq_run_hw_queue - Start to run a hardware queue.
1533 * @hctx: Pointer to the hardware queue to run.
1534 * @async: If we want to run the queue asynchronously.
1536 * Check if the request queue is not in a quiesced state and if there are
1537 * pending requests to be sent. If this is true, run the queue to send requests
1540 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1546 * When queue is quiesced, we may be switching io scheduler, or
1547 * updating nr_hw_queues, or other things, and we can't run queue
1548 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1550 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1553 hctx_lock(hctx, &srcu_idx);
1554 need_run = !blk_queue_quiesced(hctx->queue) &&
1555 blk_mq_hctx_has_pending(hctx);
1556 hctx_unlock(hctx, srcu_idx);
1559 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1561 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1564 * blk_mq_run_hw_queue - Run all hardware queues in a request queue.
1565 * @q: Pointer to the request queue to run.
1566 * @async: If we want to run the queue asynchronously.
1568 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1570 struct blk_mq_hw_ctx *hctx;
1573 queue_for_each_hw_ctx(q, hctx, i) {
1574 if (blk_mq_hctx_stopped(hctx))
1577 blk_mq_run_hw_queue(hctx, async);
1580 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1583 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1584 * @q: Pointer to the request queue to run.
1585 * @msecs: Microseconds of delay to wait before running the queues.
1587 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1589 struct blk_mq_hw_ctx *hctx;
1592 queue_for_each_hw_ctx(q, hctx, i) {
1593 if (blk_mq_hctx_stopped(hctx))
1596 blk_mq_delay_run_hw_queue(hctx, msecs);
1599 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1602 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1603 * @q: request queue.
1605 * The caller is responsible for serializing this function against
1606 * blk_mq_{start,stop}_hw_queue().
1608 bool blk_mq_queue_stopped(struct request_queue *q)
1610 struct blk_mq_hw_ctx *hctx;
1613 queue_for_each_hw_ctx(q, hctx, i)
1614 if (blk_mq_hctx_stopped(hctx))
1619 EXPORT_SYMBOL(blk_mq_queue_stopped);
1622 * This function is often used for pausing .queue_rq() by driver when
1623 * there isn't enough resource or some conditions aren't satisfied, and
1624 * BLK_STS_RESOURCE is usually returned.
1626 * We do not guarantee that dispatch can be drained or blocked
1627 * after blk_mq_stop_hw_queue() returns. Please use
1628 * blk_mq_quiesce_queue() for that requirement.
1630 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1632 cancel_delayed_work(&hctx->run_work);
1634 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1636 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1639 * This function is often used for pausing .queue_rq() by driver when
1640 * there isn't enough resource or some conditions aren't satisfied, and
1641 * BLK_STS_RESOURCE is usually returned.
1643 * We do not guarantee that dispatch can be drained or blocked
1644 * after blk_mq_stop_hw_queues() returns. Please use
1645 * blk_mq_quiesce_queue() for that requirement.
1647 void blk_mq_stop_hw_queues(struct request_queue *q)
1649 struct blk_mq_hw_ctx *hctx;
1652 queue_for_each_hw_ctx(q, hctx, i)
1653 blk_mq_stop_hw_queue(hctx);
1655 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1657 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1659 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1661 blk_mq_run_hw_queue(hctx, false);
1663 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1665 void blk_mq_start_hw_queues(struct request_queue *q)
1667 struct blk_mq_hw_ctx *hctx;
1670 queue_for_each_hw_ctx(q, hctx, i)
1671 blk_mq_start_hw_queue(hctx);
1673 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1675 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1677 if (!blk_mq_hctx_stopped(hctx))
1680 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1681 blk_mq_run_hw_queue(hctx, async);
1683 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1685 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1687 struct blk_mq_hw_ctx *hctx;
1690 queue_for_each_hw_ctx(q, hctx, i)
1691 blk_mq_start_stopped_hw_queue(hctx, async);
1693 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1695 static void blk_mq_run_work_fn(struct work_struct *work)
1697 struct blk_mq_hw_ctx *hctx;
1699 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1702 * If we are stopped, don't run the queue.
1704 if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1707 __blk_mq_run_hw_queue(hctx);
1710 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1714 struct blk_mq_ctx *ctx = rq->mq_ctx;
1715 enum hctx_type type = hctx->type;
1717 lockdep_assert_held(&ctx->lock);
1719 trace_block_rq_insert(hctx->queue, rq);
1722 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1724 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1727 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1730 struct blk_mq_ctx *ctx = rq->mq_ctx;
1732 lockdep_assert_held(&ctx->lock);
1734 __blk_mq_insert_req_list(hctx, rq, at_head);
1735 blk_mq_hctx_mark_pending(hctx, ctx);
1739 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
1740 * @rq: Pointer to request to be inserted.
1741 * @run_queue: If we should run the hardware queue after inserting the request.
1743 * Should only be used carefully, when the caller knows we want to
1744 * bypass a potential IO scheduler on the target device.
1746 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1749 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1751 spin_lock(&hctx->lock);
1753 list_add(&rq->queuelist, &hctx->dispatch);
1755 list_add_tail(&rq->queuelist, &hctx->dispatch);
1756 spin_unlock(&hctx->lock);
1759 blk_mq_run_hw_queue(hctx, false);
1762 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1763 struct list_head *list)
1767 enum hctx_type type = hctx->type;
1770 * preemption doesn't flush plug list, so it's possible ctx->cpu is
1773 list_for_each_entry(rq, list, queuelist) {
1774 BUG_ON(rq->mq_ctx != ctx);
1775 trace_block_rq_insert(hctx->queue, rq);
1778 spin_lock(&ctx->lock);
1779 list_splice_tail_init(list, &ctx->rq_lists[type]);
1780 blk_mq_hctx_mark_pending(hctx, ctx);
1781 spin_unlock(&ctx->lock);
1784 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1786 struct request *rqa = container_of(a, struct request, queuelist);
1787 struct request *rqb = container_of(b, struct request, queuelist);
1789 if (rqa->mq_ctx != rqb->mq_ctx)
1790 return rqa->mq_ctx > rqb->mq_ctx;
1791 if (rqa->mq_hctx != rqb->mq_hctx)
1792 return rqa->mq_hctx > rqb->mq_hctx;
1794 return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1797 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1801 if (list_empty(&plug->mq_list))
1803 list_splice_init(&plug->mq_list, &list);
1805 if (plug->rq_count > 2 && plug->multiple_queues)
1806 list_sort(NULL, &list, plug_rq_cmp);
1811 struct list_head rq_list;
1812 struct request *rq, *head_rq = list_entry_rq(list.next);
1813 struct list_head *pos = &head_rq->queuelist; /* skip first */
1814 struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx;
1815 struct blk_mq_ctx *this_ctx = head_rq->mq_ctx;
1816 unsigned int depth = 1;
1818 list_for_each_continue(pos, &list) {
1819 rq = list_entry_rq(pos);
1821 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx)
1826 list_cut_before(&rq_list, &list, pos);
1827 trace_block_unplug(head_rq->q, depth, !from_schedule);
1828 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1830 } while(!list_empty(&list));
1833 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1834 unsigned int nr_segs)
1836 if (bio->bi_opf & REQ_RAHEAD)
1837 rq->cmd_flags |= REQ_FAILFAST_MASK;
1839 rq->__sector = bio->bi_iter.bi_sector;
1840 rq->write_hint = bio->bi_write_hint;
1841 blk_rq_bio_prep(rq, bio, nr_segs);
1842 blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
1844 blk_account_io_start(rq);
1847 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1849 blk_qc_t *cookie, bool last)
1851 struct request_queue *q = rq->q;
1852 struct blk_mq_queue_data bd = {
1856 blk_qc_t new_cookie;
1859 new_cookie = request_to_qc_t(hctx, rq);
1862 * For OK queue, we are done. For error, caller may kill it.
1863 * Any other error (busy), just add it to our list as we
1864 * previously would have done.
1866 ret = q->mq_ops->queue_rq(hctx, &bd);
1869 blk_mq_update_dispatch_busy(hctx, false);
1870 *cookie = new_cookie;
1872 case BLK_STS_RESOURCE:
1873 case BLK_STS_DEV_RESOURCE:
1874 blk_mq_update_dispatch_busy(hctx, true);
1875 __blk_mq_requeue_request(rq);
1878 blk_mq_update_dispatch_busy(hctx, false);
1879 *cookie = BLK_QC_T_NONE;
1886 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1889 bool bypass_insert, bool last)
1891 struct request_queue *q = rq->q;
1892 bool run_queue = true;
1895 * RCU or SRCU read lock is needed before checking quiesced flag.
1897 * When queue is stopped or quiesced, ignore 'bypass_insert' from
1898 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1899 * and avoid driver to try to dispatch again.
1901 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1903 bypass_insert = false;
1907 if (q->elevator && !bypass_insert)
1910 if (!blk_mq_get_dispatch_budget(hctx))
1913 if (!blk_mq_get_driver_tag(rq)) {
1914 blk_mq_put_dispatch_budget(hctx);
1918 return __blk_mq_issue_directly(hctx, rq, cookie, last);
1921 return BLK_STS_RESOURCE;
1923 blk_mq_request_bypass_insert(rq, false, run_queue);
1928 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
1929 * @hctx: Pointer of the associated hardware queue.
1930 * @rq: Pointer to request to be sent.
1931 * @cookie: Request queue cookie.
1933 * If the device has enough resources to accept a new request now, send the
1934 * request directly to device driver. Else, insert at hctx->dispatch queue, so
1935 * we can try send it another time in the future. Requests inserted at this
1936 * queue have higher priority.
1938 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1939 struct request *rq, blk_qc_t *cookie)
1944 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1946 hctx_lock(hctx, &srcu_idx);
1948 ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
1949 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1950 blk_mq_request_bypass_insert(rq, false, true);
1951 else if (ret != BLK_STS_OK)
1952 blk_mq_end_request(rq, ret);
1954 hctx_unlock(hctx, srcu_idx);
1957 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
1961 blk_qc_t unused_cookie;
1962 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1964 hctx_lock(hctx, &srcu_idx);
1965 ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
1966 hctx_unlock(hctx, srcu_idx);
1971 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1972 struct list_head *list)
1976 while (!list_empty(list)) {
1978 struct request *rq = list_first_entry(list, struct request,
1981 list_del_init(&rq->queuelist);
1982 ret = blk_mq_request_issue_directly(rq, list_empty(list));
1983 if (ret != BLK_STS_OK) {
1984 if (ret == BLK_STS_RESOURCE ||
1985 ret == BLK_STS_DEV_RESOURCE) {
1986 blk_mq_request_bypass_insert(rq, false,
1990 blk_mq_end_request(rq, ret);
1996 * If we didn't flush the entire list, we could have told
1997 * the driver there was more coming, but that turned out to
2000 if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs && queued)
2001 hctx->queue->mq_ops->commit_rqs(hctx);
2004 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2006 list_add_tail(&rq->queuelist, &plug->mq_list);
2008 if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
2009 struct request *tmp;
2011 tmp = list_first_entry(&plug->mq_list, struct request,
2013 if (tmp->q != rq->q)
2014 plug->multiple_queues = true;
2019 * blk_mq_make_request - Create and send a request to block device.
2020 * @q: Request queue pointer.
2021 * @bio: Bio pointer.
2023 * Builds up a request structure from @q and @bio and send to the device. The
2024 * request may not be queued directly to hardware if:
2025 * * This request can be merged with another one
2026 * * We want to place request at plug queue for possible future merging
2027 * * There is an IO scheduler active at this queue
2029 * It will not queue the request if there is an error with the bio, or at the
2032 * Returns: Request queue cookie.
2034 blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
2036 const int is_sync = op_is_sync(bio->bi_opf);
2037 const int is_flush_fua = op_is_flush(bio->bi_opf);
2038 struct blk_mq_alloc_data data = {
2042 struct blk_plug *plug;
2043 struct request *same_queue_rq = NULL;
2044 unsigned int nr_segs;
2048 blk_queue_bounce(q, &bio);
2049 __blk_queue_split(q, &bio, &nr_segs);
2051 if (!bio_integrity_prep(bio))
2054 if (!is_flush_fua && !blk_queue_nomerges(q) &&
2055 blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2058 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2061 rq_qos_throttle(q, bio);
2063 data.cmd_flags = bio->bi_opf;
2064 rq = __blk_mq_alloc_request(&data);
2065 if (unlikely(!rq)) {
2066 rq_qos_cleanup(q, bio);
2067 if (bio->bi_opf & REQ_NOWAIT)
2068 bio_wouldblock_error(bio);
2072 trace_block_getrq(q, bio, bio->bi_opf);
2074 rq_qos_track(q, rq, bio);
2076 cookie = request_to_qc_t(data.hctx, rq);
2078 blk_mq_bio_to_request(rq, bio, nr_segs);
2080 ret = blk_crypto_init_request(rq);
2081 if (ret != BLK_STS_OK) {
2082 bio->bi_status = ret;
2084 blk_mq_free_request(rq);
2085 return BLK_QC_T_NONE;
2088 plug = blk_mq_plug(q, bio);
2089 if (unlikely(is_flush_fua)) {
2090 /* Bypass scheduler for flush requests */
2091 blk_insert_flush(rq);
2092 blk_mq_run_hw_queue(data.hctx, true);
2093 } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2094 !blk_queue_nonrot(q))) {
2096 * Use plugging if we have a ->commit_rqs() hook as well, as
2097 * we know the driver uses bd->last in a smart fashion.
2099 * Use normal plugging if this disk is slow HDD, as sequential
2100 * IO may benefit a lot from plug merging.
2102 unsigned int request_count = plug->rq_count;
2103 struct request *last = NULL;
2106 trace_block_plug(q);
2108 last = list_entry_rq(plug->mq_list.prev);
2110 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2111 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2112 blk_flush_plug_list(plug, false);
2113 trace_block_plug(q);
2116 blk_add_rq_to_plug(plug, rq);
2117 } else if (q->elevator) {
2118 /* Insert the request at the IO scheduler queue */
2119 blk_mq_sched_insert_request(rq, false, true, true);
2120 } else if (plug && !blk_queue_nomerges(q)) {
2122 * We do limited plugging. If the bio can be merged, do that.
2123 * Otherwise the existing request in the plug list will be
2124 * issued. So the plug list will have one request at most
2125 * The plug list might get flushed before this. If that happens,
2126 * the plug list is empty, and same_queue_rq is invalid.
2128 if (list_empty(&plug->mq_list))
2129 same_queue_rq = NULL;
2130 if (same_queue_rq) {
2131 list_del_init(&same_queue_rq->queuelist);
2134 blk_add_rq_to_plug(plug, rq);
2135 trace_block_plug(q);
2137 if (same_queue_rq) {
2138 data.hctx = same_queue_rq->mq_hctx;
2139 trace_block_unplug(q, 1, true);
2140 blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2143 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2144 !data.hctx->dispatch_busy) {
2146 * There is no scheduler and we can try to send directly
2149 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2152 blk_mq_sched_insert_request(rq, false, true, true);
2158 return BLK_QC_T_NONE;
2160 EXPORT_SYMBOL_GPL(blk_mq_make_request); /* only for request based dm */
2162 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2163 unsigned int hctx_idx)
2167 if (tags->rqs && set->ops->exit_request) {
2170 for (i = 0; i < tags->nr_tags; i++) {
2171 struct request *rq = tags->static_rqs[i];
2175 set->ops->exit_request(set, rq, hctx_idx);
2176 tags->static_rqs[i] = NULL;
2180 while (!list_empty(&tags->page_list)) {
2181 page = list_first_entry(&tags->page_list, struct page, lru);
2182 list_del_init(&page->lru);
2184 * Remove kmemleak object previously allocated in
2185 * blk_mq_alloc_rqs().
2187 kmemleak_free(page_address(page));
2188 __free_pages(page, page->private);
2192 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2196 kfree(tags->static_rqs);
2197 tags->static_rqs = NULL;
2199 blk_mq_free_tags(tags);
2202 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2203 unsigned int hctx_idx,
2204 unsigned int nr_tags,
2205 unsigned int reserved_tags)
2207 struct blk_mq_tags *tags;
2210 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2211 if (node == NUMA_NO_NODE)
2212 node = set->numa_node;
2214 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2215 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2219 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2220 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2223 blk_mq_free_tags(tags);
2227 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2228 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2230 if (!tags->static_rqs) {
2232 blk_mq_free_tags(tags);
2239 static size_t order_to_size(unsigned int order)
2241 return (size_t)PAGE_SIZE << order;
2244 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2245 unsigned int hctx_idx, int node)
2249 if (set->ops->init_request) {
2250 ret = set->ops->init_request(set, rq, hctx_idx, node);
2255 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2259 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2260 unsigned int hctx_idx, unsigned int depth)
2262 unsigned int i, j, entries_per_page, max_order = 4;
2263 size_t rq_size, left;
2266 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2267 if (node == NUMA_NO_NODE)
2268 node = set->numa_node;
2270 INIT_LIST_HEAD(&tags->page_list);
2273 * rq_size is the size of the request plus driver payload, rounded
2274 * to the cacheline size
2276 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2278 left = rq_size * depth;
2280 for (i = 0; i < depth; ) {
2281 int this_order = max_order;
2286 while (this_order && left < order_to_size(this_order - 1))
2290 page = alloc_pages_node(node,
2291 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2297 if (order_to_size(this_order) < rq_size)
2304 page->private = this_order;
2305 list_add_tail(&page->lru, &tags->page_list);
2307 p = page_address(page);
2309 * Allow kmemleak to scan these pages as they contain pointers
2310 * to additional allocations like via ops->init_request().
2312 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2313 entries_per_page = order_to_size(this_order) / rq_size;
2314 to_do = min(entries_per_page, depth - i);
2315 left -= to_do * rq_size;
2316 for (j = 0; j < to_do; j++) {
2317 struct request *rq = p;
2319 tags->static_rqs[i] = rq;
2320 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2321 tags->static_rqs[i] = NULL;
2332 blk_mq_free_rqs(set, tags, hctx_idx);
2337 * 'cpu' is going away. splice any existing rq_list entries from this
2338 * software queue to the hw queue dispatch list, and ensure that it
2341 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2343 struct blk_mq_hw_ctx *hctx;
2344 struct blk_mq_ctx *ctx;
2346 enum hctx_type type;
2348 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2349 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2352 spin_lock(&ctx->lock);
2353 if (!list_empty(&ctx->rq_lists[type])) {
2354 list_splice_init(&ctx->rq_lists[type], &tmp);
2355 blk_mq_hctx_clear_pending(hctx, ctx);
2357 spin_unlock(&ctx->lock);
2359 if (list_empty(&tmp))
2362 spin_lock(&hctx->lock);
2363 list_splice_tail_init(&tmp, &hctx->dispatch);
2364 spin_unlock(&hctx->lock);
2366 blk_mq_run_hw_queue(hctx, true);
2370 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2372 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2376 /* hctx->ctxs will be freed in queue's release handler */
2377 static void blk_mq_exit_hctx(struct request_queue *q,
2378 struct blk_mq_tag_set *set,
2379 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2381 if (blk_mq_hw_queue_mapped(hctx))
2382 blk_mq_tag_idle(hctx);
2384 if (set->ops->exit_request)
2385 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2387 if (set->ops->exit_hctx)
2388 set->ops->exit_hctx(hctx, hctx_idx);
2390 blk_mq_remove_cpuhp(hctx);
2392 spin_lock(&q->unused_hctx_lock);
2393 list_add(&hctx->hctx_list, &q->unused_hctx_list);
2394 spin_unlock(&q->unused_hctx_lock);
2397 static void blk_mq_exit_hw_queues(struct request_queue *q,
2398 struct blk_mq_tag_set *set, int nr_queue)
2400 struct blk_mq_hw_ctx *hctx;
2403 queue_for_each_hw_ctx(q, hctx, i) {
2406 blk_mq_debugfs_unregister_hctx(hctx);
2407 blk_mq_exit_hctx(q, set, hctx, i);
2411 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2413 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2415 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2416 __alignof__(struct blk_mq_hw_ctx)) !=
2417 sizeof(struct blk_mq_hw_ctx));
2419 if (tag_set->flags & BLK_MQ_F_BLOCKING)
2420 hw_ctx_size += sizeof(struct srcu_struct);
2425 static int blk_mq_init_hctx(struct request_queue *q,
2426 struct blk_mq_tag_set *set,
2427 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2429 hctx->queue_num = hctx_idx;
2431 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2433 hctx->tags = set->tags[hctx_idx];
2435 if (set->ops->init_hctx &&
2436 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2437 goto unregister_cpu_notifier;
2439 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2445 if (set->ops->exit_hctx)
2446 set->ops->exit_hctx(hctx, hctx_idx);
2447 unregister_cpu_notifier:
2448 blk_mq_remove_cpuhp(hctx);
2452 static struct blk_mq_hw_ctx *
2453 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2456 struct blk_mq_hw_ctx *hctx;
2457 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2459 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2461 goto fail_alloc_hctx;
2463 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2466 atomic_set(&hctx->nr_active, 0);
2467 if (node == NUMA_NO_NODE)
2468 node = set->numa_node;
2469 hctx->numa_node = node;
2471 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2472 spin_lock_init(&hctx->lock);
2473 INIT_LIST_HEAD(&hctx->dispatch);
2475 hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2477 INIT_LIST_HEAD(&hctx->hctx_list);
2480 * Allocate space for all possible cpus to avoid allocation at
2483 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2488 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2493 spin_lock_init(&hctx->dispatch_wait_lock);
2494 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2495 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2497 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
2501 if (hctx->flags & BLK_MQ_F_BLOCKING)
2502 init_srcu_struct(hctx->srcu);
2503 blk_mq_hctx_kobj_init(hctx);
2508 sbitmap_free(&hctx->ctx_map);
2512 free_cpumask_var(hctx->cpumask);
2519 static void blk_mq_init_cpu_queues(struct request_queue *q,
2520 unsigned int nr_hw_queues)
2522 struct blk_mq_tag_set *set = q->tag_set;
2525 for_each_possible_cpu(i) {
2526 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2527 struct blk_mq_hw_ctx *hctx;
2531 spin_lock_init(&__ctx->lock);
2532 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2533 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2538 * Set local node, IFF we have more than one hw queue. If
2539 * not, we remain on the home node of the device
2541 for (j = 0; j < set->nr_maps; j++) {
2542 hctx = blk_mq_map_queue_type(q, j, i);
2543 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2544 hctx->numa_node = local_memory_node(cpu_to_node(i));
2549 static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set,
2554 set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2555 set->queue_depth, set->reserved_tags);
2556 if (!set->tags[hctx_idx])
2559 ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2564 blk_mq_free_rq_map(set->tags[hctx_idx]);
2565 set->tags[hctx_idx] = NULL;
2569 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2570 unsigned int hctx_idx)
2572 if (set->tags && set->tags[hctx_idx]) {
2573 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2574 blk_mq_free_rq_map(set->tags[hctx_idx]);
2575 set->tags[hctx_idx] = NULL;
2579 static void blk_mq_map_swqueue(struct request_queue *q)
2581 unsigned int i, j, hctx_idx;
2582 struct blk_mq_hw_ctx *hctx;
2583 struct blk_mq_ctx *ctx;
2584 struct blk_mq_tag_set *set = q->tag_set;
2586 queue_for_each_hw_ctx(q, hctx, i) {
2587 cpumask_clear(hctx->cpumask);
2589 hctx->dispatch_from = NULL;
2593 * Map software to hardware queues.
2595 * If the cpu isn't present, the cpu is mapped to first hctx.
2597 for_each_possible_cpu(i) {
2599 ctx = per_cpu_ptr(q->queue_ctx, i);
2600 for (j = 0; j < set->nr_maps; j++) {
2601 if (!set->map[j].nr_queues) {
2602 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2603 HCTX_TYPE_DEFAULT, i);
2606 hctx_idx = set->map[j].mq_map[i];
2607 /* unmapped hw queue can be remapped after CPU topo changed */
2608 if (!set->tags[hctx_idx] &&
2609 !__blk_mq_alloc_map_and_request(set, hctx_idx)) {
2611 * If tags initialization fail for some hctx,
2612 * that hctx won't be brought online. In this
2613 * case, remap the current ctx to hctx[0] which
2614 * is guaranteed to always have tags allocated
2616 set->map[j].mq_map[i] = 0;
2619 hctx = blk_mq_map_queue_type(q, j, i);
2620 ctx->hctxs[j] = hctx;
2622 * If the CPU is already set in the mask, then we've
2623 * mapped this one already. This can happen if
2624 * devices share queues across queue maps.
2626 if (cpumask_test_cpu(i, hctx->cpumask))
2629 cpumask_set_cpu(i, hctx->cpumask);
2631 ctx->index_hw[hctx->type] = hctx->nr_ctx;
2632 hctx->ctxs[hctx->nr_ctx++] = ctx;
2635 * If the nr_ctx type overflows, we have exceeded the
2636 * amount of sw queues we can support.
2638 BUG_ON(!hctx->nr_ctx);
2641 for (; j < HCTX_MAX_TYPES; j++)
2642 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2643 HCTX_TYPE_DEFAULT, i);
2646 queue_for_each_hw_ctx(q, hctx, i) {
2648 * If no software queues are mapped to this hardware queue,
2649 * disable it and free the request entries.
2651 if (!hctx->nr_ctx) {
2652 /* Never unmap queue 0. We need it as a
2653 * fallback in case of a new remap fails
2656 if (i && set->tags[i])
2657 blk_mq_free_map_and_requests(set, i);
2663 hctx->tags = set->tags[i];
2664 WARN_ON(!hctx->tags);
2667 * Set the map size to the number of mapped software queues.
2668 * This is more accurate and more efficient than looping
2669 * over all possibly mapped software queues.
2671 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2674 * Initialize batch roundrobin counts
2676 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2677 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2682 * Caller needs to ensure that we're either frozen/quiesced, or that
2683 * the queue isn't live yet.
2685 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2687 struct blk_mq_hw_ctx *hctx;
2690 queue_for_each_hw_ctx(q, hctx, i) {
2692 hctx->flags |= BLK_MQ_F_TAG_SHARED;
2694 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2698 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2701 struct request_queue *q;
2703 lockdep_assert_held(&set->tag_list_lock);
2705 list_for_each_entry(q, &set->tag_list, tag_set_list) {
2706 blk_mq_freeze_queue(q);
2707 queue_set_hctx_shared(q, shared);
2708 blk_mq_unfreeze_queue(q);
2712 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2714 struct blk_mq_tag_set *set = q->tag_set;
2716 mutex_lock(&set->tag_list_lock);
2717 list_del_rcu(&q->tag_set_list);
2718 if (list_is_singular(&set->tag_list)) {
2719 /* just transitioned to unshared */
2720 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2721 /* update existing queue */
2722 blk_mq_update_tag_set_depth(set, false);
2724 mutex_unlock(&set->tag_list_lock);
2725 INIT_LIST_HEAD(&q->tag_set_list);
2728 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2729 struct request_queue *q)
2731 mutex_lock(&set->tag_list_lock);
2734 * Check to see if we're transitioning to shared (from 1 to 2 queues).
2736 if (!list_empty(&set->tag_list) &&
2737 !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2738 set->flags |= BLK_MQ_F_TAG_SHARED;
2739 /* update existing queue */
2740 blk_mq_update_tag_set_depth(set, true);
2742 if (set->flags & BLK_MQ_F_TAG_SHARED)
2743 queue_set_hctx_shared(q, true);
2744 list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2746 mutex_unlock(&set->tag_list_lock);
2749 /* All allocations will be freed in release handler of q->mq_kobj */
2750 static int blk_mq_alloc_ctxs(struct request_queue *q)
2752 struct blk_mq_ctxs *ctxs;
2755 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2759 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2760 if (!ctxs->queue_ctx)
2763 for_each_possible_cpu(cpu) {
2764 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2768 q->mq_kobj = &ctxs->kobj;
2769 q->queue_ctx = ctxs->queue_ctx;
2778 * It is the actual release handler for mq, but we do it from
2779 * request queue's release handler for avoiding use-after-free
2780 * and headache because q->mq_kobj shouldn't have been introduced,
2781 * but we can't group ctx/kctx kobj without it.
2783 void blk_mq_release(struct request_queue *q)
2785 struct blk_mq_hw_ctx *hctx, *next;
2788 queue_for_each_hw_ctx(q, hctx, i)
2789 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2791 /* all hctx are in .unused_hctx_list now */
2792 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2793 list_del_init(&hctx->hctx_list);
2794 kobject_put(&hctx->kobj);
2797 kfree(q->queue_hw_ctx);
2800 * release .mq_kobj and sw queue's kobject now because
2801 * both share lifetime with request queue.
2803 blk_mq_sysfs_deinit(q);
2806 struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
2809 struct request_queue *uninit_q, *q;
2811 uninit_q = __blk_alloc_queue(set->numa_node);
2813 return ERR_PTR(-ENOMEM);
2814 uninit_q->queuedata = queuedata;
2817 * Initialize the queue without an elevator. device_add_disk() will do
2818 * the initialization.
2820 q = blk_mq_init_allocated_queue(set, uninit_q, false);
2822 blk_cleanup_queue(uninit_q);
2826 EXPORT_SYMBOL_GPL(blk_mq_init_queue_data);
2828 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2830 return blk_mq_init_queue_data(set, NULL);
2832 EXPORT_SYMBOL(blk_mq_init_queue);
2835 * Helper for setting up a queue with mq ops, given queue depth, and
2836 * the passed in mq ops flags.
2838 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2839 const struct blk_mq_ops *ops,
2840 unsigned int queue_depth,
2841 unsigned int set_flags)
2843 struct request_queue *q;
2846 memset(set, 0, sizeof(*set));
2848 set->nr_hw_queues = 1;
2850 set->queue_depth = queue_depth;
2851 set->numa_node = NUMA_NO_NODE;
2852 set->flags = set_flags;
2854 ret = blk_mq_alloc_tag_set(set);
2856 return ERR_PTR(ret);
2858 q = blk_mq_init_queue(set);
2860 blk_mq_free_tag_set(set);
2866 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2868 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2869 struct blk_mq_tag_set *set, struct request_queue *q,
2870 int hctx_idx, int node)
2872 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2874 /* reuse dead hctx first */
2875 spin_lock(&q->unused_hctx_lock);
2876 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2877 if (tmp->numa_node == node) {
2883 list_del_init(&hctx->hctx_list);
2884 spin_unlock(&q->unused_hctx_lock);
2887 hctx = blk_mq_alloc_hctx(q, set, node);
2891 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2897 kobject_put(&hctx->kobj);
2902 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2903 struct request_queue *q)
2906 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2908 if (q->nr_hw_queues < set->nr_hw_queues) {
2909 struct blk_mq_hw_ctx **new_hctxs;
2911 new_hctxs = kcalloc_node(set->nr_hw_queues,
2912 sizeof(*new_hctxs), GFP_KERNEL,
2917 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
2919 q->queue_hw_ctx = new_hctxs;
2924 /* protect against switching io scheduler */
2925 mutex_lock(&q->sysfs_lock);
2926 for (i = 0; i < set->nr_hw_queues; i++) {
2928 struct blk_mq_hw_ctx *hctx;
2930 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2932 * If the hw queue has been mapped to another numa node,
2933 * we need to realloc the hctx. If allocation fails, fallback
2934 * to use the previous one.
2936 if (hctxs[i] && (hctxs[i]->numa_node == node))
2939 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2942 blk_mq_exit_hctx(q, set, hctxs[i], i);
2946 pr_warn("Allocate new hctx on node %d fails,\
2947 fallback to previous one on node %d\n",
2948 node, hctxs[i]->numa_node);
2954 * Increasing nr_hw_queues fails. Free the newly allocated
2955 * hctxs and keep the previous q->nr_hw_queues.
2957 if (i != set->nr_hw_queues) {
2958 j = q->nr_hw_queues;
2962 end = q->nr_hw_queues;
2963 q->nr_hw_queues = set->nr_hw_queues;
2966 for (; j < end; j++) {
2967 struct blk_mq_hw_ctx *hctx = hctxs[j];
2971 blk_mq_free_map_and_requests(set, j);
2972 blk_mq_exit_hctx(q, set, hctx, j);
2976 mutex_unlock(&q->sysfs_lock);
2979 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2980 struct request_queue *q,
2983 /* mark the queue as mq asap */
2984 q->mq_ops = set->ops;
2986 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2987 blk_mq_poll_stats_bkt,
2988 BLK_MQ_POLL_STATS_BKTS, q);
2992 if (blk_mq_alloc_ctxs(q))
2995 /* init q->mq_kobj and sw queues' kobjects */
2996 blk_mq_sysfs_init(q);
2998 INIT_LIST_HEAD(&q->unused_hctx_list);
2999 spin_lock_init(&q->unused_hctx_lock);
3001 blk_mq_realloc_hw_ctxs(set, q);
3002 if (!q->nr_hw_queues)
3005 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3006 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3010 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3011 if (set->nr_maps > HCTX_TYPE_POLL &&
3012 set->map[HCTX_TYPE_POLL].nr_queues)
3013 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3015 q->sg_reserved_size = INT_MAX;
3017 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3018 INIT_LIST_HEAD(&q->requeue_list);
3019 spin_lock_init(&q->requeue_lock);
3021 q->nr_requests = set->queue_depth;
3024 * Default to classic polling
3026 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3028 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3029 blk_mq_add_queue_tag_set(set, q);
3030 blk_mq_map_swqueue(q);
3033 elevator_init_mq(q);
3038 kfree(q->queue_hw_ctx);
3039 q->nr_hw_queues = 0;
3040 blk_mq_sysfs_deinit(q);
3042 blk_stat_free_callback(q->poll_cb);
3046 return ERR_PTR(-ENOMEM);
3048 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3050 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3051 void blk_mq_exit_queue(struct request_queue *q)
3053 struct blk_mq_tag_set *set = q->tag_set;
3055 blk_mq_del_queue_tag_set(q);
3056 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3059 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3063 for (i = 0; i < set->nr_hw_queues; i++)
3064 if (!__blk_mq_alloc_map_and_request(set, i))
3071 blk_mq_free_map_and_requests(set, i);
3077 * Allocate the request maps associated with this tag_set. Note that this
3078 * may reduce the depth asked for, if memory is tight. set->queue_depth
3079 * will be updated to reflect the allocated depth.
3081 static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set)
3086 depth = set->queue_depth;
3088 err = __blk_mq_alloc_rq_maps(set);
3092 set->queue_depth >>= 1;
3093 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3097 } while (set->queue_depth);
3099 if (!set->queue_depth || err) {
3100 pr_err("blk-mq: failed to allocate request map\n");
3104 if (depth != set->queue_depth)
3105 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3106 depth, set->queue_depth);
3111 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3114 * blk_mq_map_queues() and multiple .map_queues() implementations
3115 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3116 * number of hardware queues.
3118 if (set->nr_maps == 1)
3119 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3121 if (set->ops->map_queues && !is_kdump_kernel()) {
3125 * transport .map_queues is usually done in the following
3128 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3129 * mask = get_cpu_mask(queue)
3130 * for_each_cpu(cpu, mask)
3131 * set->map[x].mq_map[cpu] = queue;
3134 * When we need to remap, the table has to be cleared for
3135 * killing stale mapping since one CPU may not be mapped
3138 for (i = 0; i < set->nr_maps; i++)
3139 blk_mq_clear_mq_map(&set->map[i]);
3141 return set->ops->map_queues(set);
3143 BUG_ON(set->nr_maps > 1);
3144 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3148 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3149 int cur_nr_hw_queues, int new_nr_hw_queues)
3151 struct blk_mq_tags **new_tags;
3153 if (cur_nr_hw_queues >= new_nr_hw_queues)
3156 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3157 GFP_KERNEL, set->numa_node);
3162 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3163 sizeof(*set->tags));
3165 set->tags = new_tags;
3166 set->nr_hw_queues = new_nr_hw_queues;
3172 * Alloc a tag set to be associated with one or more request queues.
3173 * May fail with EINVAL for various error conditions. May adjust the
3174 * requested depth down, if it's too large. In that case, the set
3175 * value will be stored in set->queue_depth.
3177 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3181 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3183 if (!set->nr_hw_queues)
3185 if (!set->queue_depth)
3187 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3190 if (!set->ops->queue_rq)
3193 if (!set->ops->get_budget ^ !set->ops->put_budget)
3196 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3197 pr_info("blk-mq: reduced tag depth to %u\n",
3199 set->queue_depth = BLK_MQ_MAX_DEPTH;
3204 else if (set->nr_maps > HCTX_MAX_TYPES)
3208 * If a crashdump is active, then we are potentially in a very
3209 * memory constrained environment. Limit us to 1 queue and
3210 * 64 tags to prevent using too much memory.
3212 if (is_kdump_kernel()) {
3213 set->nr_hw_queues = 1;
3215 set->queue_depth = min(64U, set->queue_depth);
3218 * There is no use for more h/w queues than cpus if we just have
3221 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3222 set->nr_hw_queues = nr_cpu_ids;
3224 if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0)
3228 for (i = 0; i < set->nr_maps; i++) {
3229 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3230 sizeof(set->map[i].mq_map[0]),
3231 GFP_KERNEL, set->numa_node);
3232 if (!set->map[i].mq_map)
3233 goto out_free_mq_map;
3234 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3237 ret = blk_mq_update_queue_map(set);
3239 goto out_free_mq_map;
3241 ret = blk_mq_alloc_map_and_requests(set);
3243 goto out_free_mq_map;
3245 mutex_init(&set->tag_list_lock);
3246 INIT_LIST_HEAD(&set->tag_list);
3251 for (i = 0; i < set->nr_maps; i++) {
3252 kfree(set->map[i].mq_map);
3253 set->map[i].mq_map = NULL;
3259 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3261 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3265 for (i = 0; i < set->nr_hw_queues; i++)
3266 blk_mq_free_map_and_requests(set, i);
3268 for (j = 0; j < set->nr_maps; j++) {
3269 kfree(set->map[j].mq_map);
3270 set->map[j].mq_map = NULL;
3276 EXPORT_SYMBOL(blk_mq_free_tag_set);
3278 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3280 struct blk_mq_tag_set *set = q->tag_set;
3281 struct blk_mq_hw_ctx *hctx;
3287 if (q->nr_requests == nr)
3290 blk_mq_freeze_queue(q);
3291 blk_mq_quiesce_queue(q);
3294 queue_for_each_hw_ctx(q, hctx, i) {
3298 * If we're using an MQ scheduler, just update the scheduler
3299 * queue depth. This is similar to what the old code would do.
3301 if (!hctx->sched_tags) {
3302 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3305 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3310 if (q->elevator && q->elevator->type->ops.depth_updated)
3311 q->elevator->type->ops.depth_updated(hctx);
3315 q->nr_requests = nr;
3317 blk_mq_unquiesce_queue(q);
3318 blk_mq_unfreeze_queue(q);
3324 * request_queue and elevator_type pair.
3325 * It is just used by __blk_mq_update_nr_hw_queues to cache
3326 * the elevator_type associated with a request_queue.
3328 struct blk_mq_qe_pair {
3329 struct list_head node;
3330 struct request_queue *q;
3331 struct elevator_type *type;
3335 * Cache the elevator_type in qe pair list and switch the
3336 * io scheduler to 'none'
3338 static bool blk_mq_elv_switch_none(struct list_head *head,
3339 struct request_queue *q)
3341 struct blk_mq_qe_pair *qe;
3346 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3350 INIT_LIST_HEAD(&qe->node);
3352 qe->type = q->elevator->type;
3353 list_add(&qe->node, head);
3355 mutex_lock(&q->sysfs_lock);
3357 * After elevator_switch_mq, the previous elevator_queue will be
3358 * released by elevator_release. The reference of the io scheduler
3359 * module get by elevator_get will also be put. So we need to get
3360 * a reference of the io scheduler module here to prevent it to be
3363 __module_get(qe->type->elevator_owner);
3364 elevator_switch_mq(q, NULL);
3365 mutex_unlock(&q->sysfs_lock);
3370 static void blk_mq_elv_switch_back(struct list_head *head,
3371 struct request_queue *q)
3373 struct blk_mq_qe_pair *qe;
3374 struct elevator_type *t = NULL;
3376 list_for_each_entry(qe, head, node)
3385 list_del(&qe->node);
3388 mutex_lock(&q->sysfs_lock);
3389 elevator_switch_mq(q, t);
3390 mutex_unlock(&q->sysfs_lock);
3393 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3396 struct request_queue *q;
3398 int prev_nr_hw_queues;
3400 lockdep_assert_held(&set->tag_list_lock);
3402 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3403 nr_hw_queues = nr_cpu_ids;
3404 if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3407 list_for_each_entry(q, &set->tag_list, tag_set_list)
3408 blk_mq_freeze_queue(q);
3410 * Switch IO scheduler to 'none', cleaning up the data associated
3411 * with the previous scheduler. We will switch back once we are done
3412 * updating the new sw to hw queue mappings.
3414 list_for_each_entry(q, &set->tag_list, tag_set_list)
3415 if (!blk_mq_elv_switch_none(&head, q))
3418 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3419 blk_mq_debugfs_unregister_hctxs(q);
3420 blk_mq_sysfs_unregister(q);
3423 prev_nr_hw_queues = set->nr_hw_queues;
3424 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3428 set->nr_hw_queues = nr_hw_queues;
3430 blk_mq_update_queue_map(set);
3431 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3432 blk_mq_realloc_hw_ctxs(set, q);
3433 if (q->nr_hw_queues != set->nr_hw_queues) {
3434 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3435 nr_hw_queues, prev_nr_hw_queues);
3436 set->nr_hw_queues = prev_nr_hw_queues;
3437 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3440 blk_mq_map_swqueue(q);
3444 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3445 blk_mq_sysfs_register(q);
3446 blk_mq_debugfs_register_hctxs(q);
3450 list_for_each_entry(q, &set->tag_list, tag_set_list)
3451 blk_mq_elv_switch_back(&head, q);
3453 list_for_each_entry(q, &set->tag_list, tag_set_list)
3454 blk_mq_unfreeze_queue(q);
3457 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3459 mutex_lock(&set->tag_list_lock);
3460 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3461 mutex_unlock(&set->tag_list_lock);
3463 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3465 /* Enable polling stats and return whether they were already enabled. */
3466 static bool blk_poll_stats_enable(struct request_queue *q)
3468 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3469 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3471 blk_stat_add_callback(q, q->poll_cb);
3475 static void blk_mq_poll_stats_start(struct request_queue *q)
3478 * We don't arm the callback if polling stats are not enabled or the
3479 * callback is already active.
3481 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3482 blk_stat_is_active(q->poll_cb))
3485 blk_stat_activate_msecs(q->poll_cb, 100);
3488 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3490 struct request_queue *q = cb->data;
3493 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3494 if (cb->stat[bucket].nr_samples)
3495 q->poll_stat[bucket] = cb->stat[bucket];
3499 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3502 unsigned long ret = 0;
3506 * If stats collection isn't on, don't sleep but turn it on for
3509 if (!blk_poll_stats_enable(q))
3513 * As an optimistic guess, use half of the mean service time
3514 * for this type of request. We can (and should) make this smarter.
3515 * For instance, if the completion latencies are tight, we can
3516 * get closer than just half the mean. This is especially
3517 * important on devices where the completion latencies are longer
3518 * than ~10 usec. We do use the stats for the relevant IO size
3519 * if available which does lead to better estimates.
3521 bucket = blk_mq_poll_stats_bkt(rq);
3525 if (q->poll_stat[bucket].nr_samples)
3526 ret = (q->poll_stat[bucket].mean + 1) / 2;
3531 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3534 struct hrtimer_sleeper hs;
3535 enum hrtimer_mode mode;
3539 if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3543 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3545 * 0: use half of prev avg
3546 * >0: use this specific value
3548 if (q->poll_nsec > 0)
3549 nsecs = q->poll_nsec;
3551 nsecs = blk_mq_poll_nsecs(q, rq);
3556 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3559 * This will be replaced with the stats tracking code, using
3560 * 'avg_completion_time / 2' as the pre-sleep target.
3564 mode = HRTIMER_MODE_REL;
3565 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3566 hrtimer_set_expires(&hs.timer, kt);
3569 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3571 set_current_state(TASK_UNINTERRUPTIBLE);
3572 hrtimer_sleeper_start_expires(&hs, mode);
3575 hrtimer_cancel(&hs.timer);
3576 mode = HRTIMER_MODE_ABS;
3577 } while (hs.task && !signal_pending(current));
3579 __set_current_state(TASK_RUNNING);
3580 destroy_hrtimer_on_stack(&hs.timer);
3584 static bool blk_mq_poll_hybrid(struct request_queue *q,
3585 struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3589 if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3592 if (!blk_qc_t_is_internal(cookie))
3593 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3595 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3597 * With scheduling, if the request has completed, we'll
3598 * get a NULL return here, as we clear the sched tag when
3599 * that happens. The request still remains valid, like always,
3600 * so we should be safe with just the NULL check.
3606 return blk_mq_poll_hybrid_sleep(q, rq);
3610 * blk_poll - poll for IO completions
3612 * @cookie: cookie passed back at IO submission time
3613 * @spin: whether to spin for completions
3616 * Poll for completions on the passed in queue. Returns number of
3617 * completed entries found. If @spin is true, then blk_poll will continue
3618 * looping until at least one completion is found, unless the task is
3619 * otherwise marked running (or we need to reschedule).
3621 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3623 struct blk_mq_hw_ctx *hctx;
3626 if (!blk_qc_t_valid(cookie) ||
3627 !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3631 blk_flush_plug_list(current->plug, false);
3633 hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3636 * If we sleep, have the caller restart the poll loop to reset
3637 * the state. Like for the other success return cases, the
3638 * caller is responsible for checking if the IO completed. If
3639 * the IO isn't complete, we'll get called again and will go
3640 * straight to the busy poll loop.
3642 if (blk_mq_poll_hybrid(q, hctx, cookie))
3645 hctx->poll_considered++;
3647 state = current->state;
3651 hctx->poll_invoked++;
3653 ret = q->mq_ops->poll(hctx);
3655 hctx->poll_success++;
3656 __set_current_state(TASK_RUNNING);
3660 if (signal_pending_state(state, current))
3661 __set_current_state(TASK_RUNNING);
3663 if (current->state == TASK_RUNNING)
3665 if (ret < 0 || !spin)
3668 } while (!need_resched());
3670 __set_current_state(TASK_RUNNING);
3673 EXPORT_SYMBOL_GPL(blk_poll);
3675 unsigned int blk_mq_rq_cpu(struct request *rq)
3677 return rq->mq_ctx->cpu;
3679 EXPORT_SYMBOL(blk_mq_rq_cpu);
3681 static int __init blk_mq_init(void)
3683 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3684 blk_mq_hctx_notify_dead);
3687 subsys_initcall(blk_mq_init);