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/blk-integrity.h>
14 #include <linux/kmemleak.h>
16 #include <linux/init.h>
17 #include <linux/slab.h>
18 #include <linux/workqueue.h>
19 #include <linux/smp.h>
20 #include <linux/interrupt.h>
21 #include <linux/llist.h>
22 #include <linux/cpu.h>
23 #include <linux/cache.h>
24 #include <linux/sched/sysctl.h>
25 #include <linux/sched/topology.h>
26 #include <linux/sched/signal.h>
27 #include <linux/delay.h>
28 #include <linux/crash_dump.h>
29 #include <linux/prefetch.h>
30 #include <linux/blk-crypto.h>
31 #include <linux/part_stat.h>
33 #include <trace/events/block.h>
35 #include <linux/t10-pi.h>
38 #include "blk-mq-debugfs.h"
41 #include "blk-mq-sched.h"
42 #include "blk-rq-qos.h"
43 #include "blk-ioprio.h"
45 static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
46 static DEFINE_PER_CPU(call_single_data_t, blk_cpu_csd);
48 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags);
49 static void blk_mq_request_bypass_insert(struct request *rq,
51 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
52 struct list_head *list);
53 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
54 struct io_comp_batch *iob, unsigned int flags);
57 * Check if any of the ctx, dispatch list or elevator
58 * have pending work in this hardware queue.
60 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
62 return !list_empty_careful(&hctx->dispatch) ||
63 sbitmap_any_bit_set(&hctx->ctx_map) ||
64 blk_mq_sched_has_work(hctx);
68 * Mark this ctx as having pending work in this hardware queue
70 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
71 struct blk_mq_ctx *ctx)
73 const int bit = ctx->index_hw[hctx->type];
75 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
76 sbitmap_set_bit(&hctx->ctx_map, bit);
79 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
80 struct blk_mq_ctx *ctx)
82 const int bit = ctx->index_hw[hctx->type];
84 sbitmap_clear_bit(&hctx->ctx_map, bit);
88 struct block_device *part;
89 unsigned int inflight[2];
92 static bool blk_mq_check_inflight(struct request *rq, void *priv)
94 struct mq_inflight *mi = priv;
96 if (rq->part && blk_do_io_stat(rq) &&
97 (!mi->part->bd_partno || rq->part == mi->part) &&
98 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
99 mi->inflight[rq_data_dir(rq)]++;
104 unsigned int blk_mq_in_flight(struct request_queue *q,
105 struct block_device *part)
107 struct mq_inflight mi = { .part = part };
109 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
111 return mi.inflight[0] + mi.inflight[1];
114 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
115 unsigned int inflight[2])
117 struct mq_inflight mi = { .part = part };
119 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
120 inflight[0] = mi.inflight[0];
121 inflight[1] = mi.inflight[1];
124 void blk_freeze_queue_start(struct request_queue *q)
126 mutex_lock(&q->mq_freeze_lock);
127 if (++q->mq_freeze_depth == 1) {
128 percpu_ref_kill(&q->q_usage_counter);
129 mutex_unlock(&q->mq_freeze_lock);
131 blk_mq_run_hw_queues(q, false);
133 mutex_unlock(&q->mq_freeze_lock);
136 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
138 void blk_mq_freeze_queue_wait(struct request_queue *q)
140 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
142 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
144 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
145 unsigned long timeout)
147 return wait_event_timeout(q->mq_freeze_wq,
148 percpu_ref_is_zero(&q->q_usage_counter),
151 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
154 * Guarantee no request is in use, so we can change any data structure of
155 * the queue afterward.
157 void blk_freeze_queue(struct request_queue *q)
160 * In the !blk_mq case we are only calling this to kill the
161 * q_usage_counter, otherwise this increases the freeze depth
162 * and waits for it to return to zero. For this reason there is
163 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
164 * exported to drivers as the only user for unfreeze is blk_mq.
166 blk_freeze_queue_start(q);
167 blk_mq_freeze_queue_wait(q);
170 void blk_mq_freeze_queue(struct request_queue *q)
173 * ...just an alias to keep freeze and unfreeze actions balanced
174 * in the blk_mq_* namespace
178 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
180 void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
182 mutex_lock(&q->mq_freeze_lock);
184 q->q_usage_counter.data->force_atomic = true;
185 q->mq_freeze_depth--;
186 WARN_ON_ONCE(q->mq_freeze_depth < 0);
187 if (!q->mq_freeze_depth) {
188 percpu_ref_resurrect(&q->q_usage_counter);
189 wake_up_all(&q->mq_freeze_wq);
191 mutex_unlock(&q->mq_freeze_lock);
194 void blk_mq_unfreeze_queue(struct request_queue *q)
196 __blk_mq_unfreeze_queue(q, false);
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)
208 spin_lock_irqsave(&q->queue_lock, flags);
209 if (!q->quiesce_depth++)
210 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
211 spin_unlock_irqrestore(&q->queue_lock, flags);
213 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
216 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
217 * @set: tag_set to wait on
219 * Note: it is driver's responsibility for making sure that quiesce has
220 * been started on or more of the request_queues of the tag_set. This
221 * function only waits for the quiesce on those request_queues that had
222 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
224 void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
226 if (set->flags & BLK_MQ_F_BLOCKING)
227 synchronize_srcu(set->srcu);
231 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
234 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
237 * Note: this function does not prevent that the struct request end_io()
238 * callback function is invoked. Once this function is returned, we make
239 * sure no dispatch can happen until the queue is unquiesced via
240 * blk_mq_unquiesce_queue().
242 void blk_mq_quiesce_queue(struct request_queue *q)
244 blk_mq_quiesce_queue_nowait(q);
245 /* nothing to wait for non-mq queues */
247 blk_mq_wait_quiesce_done(q->tag_set);
249 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
252 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
255 * This function recovers queue into the state before quiescing
256 * which is done by blk_mq_quiesce_queue.
258 void blk_mq_unquiesce_queue(struct request_queue *q)
261 bool run_queue = false;
263 spin_lock_irqsave(&q->queue_lock, flags);
264 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
266 } else if (!--q->quiesce_depth) {
267 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
270 spin_unlock_irqrestore(&q->queue_lock, flags);
272 /* dispatch requests which are inserted during quiescing */
274 blk_mq_run_hw_queues(q, true);
276 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
278 void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
280 struct request_queue *q;
282 mutex_lock(&set->tag_list_lock);
283 list_for_each_entry(q, &set->tag_list, tag_set_list) {
284 if (!blk_queue_skip_tagset_quiesce(q))
285 blk_mq_quiesce_queue_nowait(q);
287 blk_mq_wait_quiesce_done(set);
288 mutex_unlock(&set->tag_list_lock);
290 EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
292 void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
294 struct request_queue *q;
296 mutex_lock(&set->tag_list_lock);
297 list_for_each_entry(q, &set->tag_list, tag_set_list) {
298 if (!blk_queue_skip_tagset_quiesce(q))
299 blk_mq_unquiesce_queue(q);
301 mutex_unlock(&set->tag_list_lock);
303 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
305 void blk_mq_wake_waiters(struct request_queue *q)
307 struct blk_mq_hw_ctx *hctx;
310 queue_for_each_hw_ctx(q, hctx, i)
311 if (blk_mq_hw_queue_mapped(hctx))
312 blk_mq_tag_wakeup_all(hctx->tags, true);
315 void blk_rq_init(struct request_queue *q, struct request *rq)
317 memset(rq, 0, sizeof(*rq));
319 INIT_LIST_HEAD(&rq->queuelist);
321 rq->__sector = (sector_t) -1;
322 INIT_HLIST_NODE(&rq->hash);
323 RB_CLEAR_NODE(&rq->rb_node);
324 rq->tag = BLK_MQ_NO_TAG;
325 rq->internal_tag = BLK_MQ_NO_TAG;
326 rq->start_time_ns = ktime_get_ns();
328 blk_crypto_rq_set_defaults(rq);
330 EXPORT_SYMBOL(blk_rq_init);
332 /* Set start and alloc time when the allocated request is actually used */
333 static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns)
335 if (blk_mq_need_time_stamp(rq))
336 rq->start_time_ns = ktime_get_ns();
338 rq->start_time_ns = 0;
340 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
341 if (blk_queue_rq_alloc_time(rq->q))
342 rq->alloc_time_ns = alloc_time_ns ?: rq->start_time_ns;
344 rq->alloc_time_ns = 0;
348 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
349 struct blk_mq_tags *tags, unsigned int tag)
351 struct blk_mq_ctx *ctx = data->ctx;
352 struct blk_mq_hw_ctx *hctx = data->hctx;
353 struct request_queue *q = data->q;
354 struct request *rq = tags->static_rqs[tag];
359 rq->cmd_flags = data->cmd_flags;
361 if (data->flags & BLK_MQ_REQ_PM)
362 data->rq_flags |= RQF_PM;
363 if (blk_queue_io_stat(q))
364 data->rq_flags |= RQF_IO_STAT;
365 rq->rq_flags = data->rq_flags;
367 if (data->rq_flags & RQF_SCHED_TAGS) {
368 rq->tag = BLK_MQ_NO_TAG;
369 rq->internal_tag = tag;
372 rq->internal_tag = BLK_MQ_NO_TAG;
377 rq->io_start_time_ns = 0;
378 rq->stats_sectors = 0;
379 rq->nr_phys_segments = 0;
380 #if defined(CONFIG_BLK_DEV_INTEGRITY)
381 rq->nr_integrity_segments = 0;
384 rq->end_io_data = NULL;
386 blk_crypto_rq_set_defaults(rq);
387 INIT_LIST_HEAD(&rq->queuelist);
388 /* tag was already set */
389 WRITE_ONCE(rq->deadline, 0);
392 if (rq->rq_flags & RQF_USE_SCHED) {
393 struct elevator_queue *e = data->q->elevator;
395 INIT_HLIST_NODE(&rq->hash);
396 RB_CLEAR_NODE(&rq->rb_node);
398 if (e->type->ops.prepare_request)
399 e->type->ops.prepare_request(rq);
405 static inline struct request *
406 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data)
408 unsigned int tag, tag_offset;
409 struct blk_mq_tags *tags;
411 unsigned long tag_mask;
414 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
415 if (unlikely(!tag_mask))
418 tags = blk_mq_tags_from_data(data);
419 for (i = 0; tag_mask; i++) {
420 if (!(tag_mask & (1UL << i)))
422 tag = tag_offset + i;
423 prefetch(tags->static_rqs[tag]);
424 tag_mask &= ~(1UL << i);
425 rq = blk_mq_rq_ctx_init(data, tags, tag);
426 rq_list_add(data->cached_rq, rq);
429 /* caller already holds a reference, add for remainder */
430 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
433 return rq_list_pop(data->cached_rq);
436 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
438 struct request_queue *q = data->q;
439 u64 alloc_time_ns = 0;
443 /* alloc_time includes depth and tag waits */
444 if (blk_queue_rq_alloc_time(q))
445 alloc_time_ns = ktime_get_ns();
447 if (data->cmd_flags & REQ_NOWAIT)
448 data->flags |= BLK_MQ_REQ_NOWAIT;
452 * All requests use scheduler tags when an I/O scheduler is
453 * enabled for the queue.
455 data->rq_flags |= RQF_SCHED_TAGS;
458 * Flush/passthrough requests are special and go directly to the
461 if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&
462 !blk_op_is_passthrough(data->cmd_flags)) {
463 struct elevator_mq_ops *ops = &q->elevator->type->ops;
465 WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
467 data->rq_flags |= RQF_USE_SCHED;
468 if (ops->limit_depth)
469 ops->limit_depth(data->cmd_flags, data);
474 data->ctx = blk_mq_get_ctx(q);
475 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
476 if (!(data->rq_flags & RQF_SCHED_TAGS))
477 blk_mq_tag_busy(data->hctx);
479 if (data->flags & BLK_MQ_REQ_RESERVED)
480 data->rq_flags |= RQF_RESV;
483 * Try batched alloc if we want more than 1 tag.
485 if (data->nr_tags > 1) {
486 rq = __blk_mq_alloc_requests_batch(data);
488 blk_mq_rq_time_init(rq, alloc_time_ns);
495 * Waiting allocations only fail because of an inactive hctx. In that
496 * case just retry the hctx assignment and tag allocation as CPU hotplug
497 * should have migrated us to an online CPU by now.
499 tag = blk_mq_get_tag(data);
500 if (tag == BLK_MQ_NO_TAG) {
501 if (data->flags & BLK_MQ_REQ_NOWAIT)
504 * Give up the CPU and sleep for a random short time to
505 * ensure that thread using a realtime scheduling class
506 * are migrated off the CPU, and thus off the hctx that
513 rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag);
514 blk_mq_rq_time_init(rq, alloc_time_ns);
518 static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
519 struct blk_plug *plug,
521 blk_mq_req_flags_t flags)
523 struct blk_mq_alloc_data data = {
527 .nr_tags = plug->nr_ios,
528 .cached_rq = &plug->cached_rq,
532 if (blk_queue_enter(q, flags))
537 rq = __blk_mq_alloc_requests(&data);
543 static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
545 blk_mq_req_flags_t flags)
547 struct blk_plug *plug = current->plug;
553 if (rq_list_empty(plug->cached_rq)) {
554 if (plug->nr_ios == 1)
556 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
560 rq = rq_list_peek(&plug->cached_rq);
561 if (!rq || rq->q != q)
564 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
566 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
569 plug->cached_rq = rq_list_next(rq);
570 blk_mq_rq_time_init(rq, 0);
574 INIT_LIST_HEAD(&rq->queuelist);
578 struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
579 blk_mq_req_flags_t flags)
583 rq = blk_mq_alloc_cached_request(q, opf, flags);
585 struct blk_mq_alloc_data data = {
593 ret = blk_queue_enter(q, flags);
597 rq = __blk_mq_alloc_requests(&data);
602 rq->__sector = (sector_t) -1;
603 rq->bio = rq->biotail = NULL;
607 return ERR_PTR(-EWOULDBLOCK);
609 EXPORT_SYMBOL(blk_mq_alloc_request);
611 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
612 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
614 struct blk_mq_alloc_data data = {
620 u64 alloc_time_ns = 0;
626 /* alloc_time includes depth and tag waits */
627 if (blk_queue_rq_alloc_time(q))
628 alloc_time_ns = ktime_get_ns();
631 * If the tag allocator sleeps we could get an allocation for a
632 * different hardware context. No need to complicate the low level
633 * allocator for this for the rare use case of a command tied to
636 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
637 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
638 return ERR_PTR(-EINVAL);
640 if (hctx_idx >= q->nr_hw_queues)
641 return ERR_PTR(-EIO);
643 ret = blk_queue_enter(q, flags);
648 * Check if the hardware context is actually mapped to anything.
649 * If not tell the caller that it should skip this queue.
652 data.hctx = xa_load(&q->hctx_table, hctx_idx);
653 if (!blk_mq_hw_queue_mapped(data.hctx))
655 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
656 if (cpu >= nr_cpu_ids)
658 data.ctx = __blk_mq_get_ctx(q, cpu);
661 data.rq_flags |= RQF_SCHED_TAGS;
663 blk_mq_tag_busy(data.hctx);
665 if (flags & BLK_MQ_REQ_RESERVED)
666 data.rq_flags |= RQF_RESV;
669 tag = blk_mq_get_tag(&data);
670 if (tag == BLK_MQ_NO_TAG)
672 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag);
673 blk_mq_rq_time_init(rq, alloc_time_ns);
675 rq->__sector = (sector_t) -1;
676 rq->bio = rq->biotail = NULL;
683 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
685 static void blk_mq_finish_request(struct request *rq)
687 struct request_queue *q = rq->q;
689 if (rq->rq_flags & RQF_USE_SCHED) {
690 q->elevator->type->ops.finish_request(rq);
692 * For postflush request that may need to be
693 * completed twice, we should clear this flag
694 * to avoid double finish_request() on the rq.
696 rq->rq_flags &= ~RQF_USE_SCHED;
700 static void __blk_mq_free_request(struct request *rq)
702 struct request_queue *q = rq->q;
703 struct blk_mq_ctx *ctx = rq->mq_ctx;
704 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
705 const int sched_tag = rq->internal_tag;
707 blk_crypto_free_request(rq);
708 blk_pm_mark_last_busy(rq);
711 if (rq->rq_flags & RQF_MQ_INFLIGHT)
712 __blk_mq_dec_active_requests(hctx);
714 if (rq->tag != BLK_MQ_NO_TAG)
715 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
716 if (sched_tag != BLK_MQ_NO_TAG)
717 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
718 blk_mq_sched_restart(hctx);
722 void blk_mq_free_request(struct request *rq)
724 struct request_queue *q = rq->q;
726 blk_mq_finish_request(rq);
728 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
729 laptop_io_completion(q->disk->bdi);
733 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
734 if (req_ref_put_and_test(rq))
735 __blk_mq_free_request(rq);
737 EXPORT_SYMBOL_GPL(blk_mq_free_request);
739 void blk_mq_free_plug_rqs(struct blk_plug *plug)
743 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
744 blk_mq_free_request(rq);
747 void blk_dump_rq_flags(struct request *rq, char *msg)
749 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
750 rq->q->disk ? rq->q->disk->disk_name : "?",
751 (__force unsigned long long) rq->cmd_flags);
753 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
754 (unsigned long long)blk_rq_pos(rq),
755 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
756 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
757 rq->bio, rq->biotail, blk_rq_bytes(rq));
759 EXPORT_SYMBOL(blk_dump_rq_flags);
761 static void req_bio_endio(struct request *rq, struct bio *bio,
762 unsigned int nbytes, blk_status_t error)
764 if (unlikely(error)) {
765 bio->bi_status = error;
766 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
768 * Partial zone append completions cannot be supported as the
769 * BIO fragments may end up not being written sequentially.
771 if (bio->bi_iter.bi_size != nbytes)
772 bio->bi_status = BLK_STS_IOERR;
774 bio->bi_iter.bi_sector = rq->__sector;
777 bio_advance(bio, nbytes);
779 if (unlikely(rq->rq_flags & RQF_QUIET))
780 bio_set_flag(bio, BIO_QUIET);
781 /* don't actually finish bio if it's part of flush sequence */
782 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
786 static void blk_account_io_completion(struct request *req, unsigned int bytes)
788 if (req->part && blk_do_io_stat(req)) {
789 const int sgrp = op_stat_group(req_op(req));
792 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
797 static void blk_print_req_error(struct request *req, blk_status_t status)
799 printk_ratelimited(KERN_ERR
800 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
801 "phys_seg %u prio class %u\n",
802 blk_status_to_str(status),
803 req->q->disk ? req->q->disk->disk_name : "?",
804 blk_rq_pos(req), (__force u32)req_op(req),
805 blk_op_str(req_op(req)),
806 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
807 req->nr_phys_segments,
808 IOPRIO_PRIO_CLASS(req->ioprio));
812 * Fully end IO on a request. Does not support partial completions, or
815 static void blk_complete_request(struct request *req)
817 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
818 int total_bytes = blk_rq_bytes(req);
819 struct bio *bio = req->bio;
821 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
826 #ifdef CONFIG_BLK_DEV_INTEGRITY
827 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
828 req->q->integrity.profile->complete_fn(req, total_bytes);
832 * Upper layers may call blk_crypto_evict_key() anytime after the last
833 * bio_endio(). Therefore, the keyslot must be released before that.
835 blk_crypto_rq_put_keyslot(req);
837 blk_account_io_completion(req, total_bytes);
840 struct bio *next = bio->bi_next;
842 /* Completion has already been traced */
843 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
845 if (req_op(req) == REQ_OP_ZONE_APPEND)
846 bio->bi_iter.bi_sector = req->__sector;
854 * Reset counters so that the request stacking driver
855 * can find how many bytes remain in the request
865 * blk_update_request - Complete multiple bytes without completing the request
866 * @req: the request being processed
867 * @error: block status code
868 * @nr_bytes: number of bytes to complete for @req
871 * Ends I/O on a number of bytes attached to @req, but doesn't complete
872 * the request structure even if @req doesn't have leftover.
873 * If @req has leftover, sets it up for the next range of segments.
875 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
876 * %false return from this function.
879 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
880 * except in the consistency check at the end of this function.
883 * %false - this request doesn't have any more data
884 * %true - this request has more data
886 bool blk_update_request(struct request *req, blk_status_t error,
887 unsigned int nr_bytes)
891 trace_block_rq_complete(req, error, nr_bytes);
896 #ifdef CONFIG_BLK_DEV_INTEGRITY
897 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
899 req->q->integrity.profile->complete_fn(req, nr_bytes);
903 * Upper layers may call blk_crypto_evict_key() anytime after the last
904 * bio_endio(). Therefore, the keyslot must be released before that.
906 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
907 __blk_crypto_rq_put_keyslot(req);
909 if (unlikely(error && !blk_rq_is_passthrough(req) &&
910 !(req->rq_flags & RQF_QUIET)) &&
911 !test_bit(GD_DEAD, &req->q->disk->state)) {
912 blk_print_req_error(req, error);
913 trace_block_rq_error(req, error, nr_bytes);
916 blk_account_io_completion(req, nr_bytes);
920 struct bio *bio = req->bio;
921 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
923 if (bio_bytes == bio->bi_iter.bi_size)
924 req->bio = bio->bi_next;
926 /* Completion has already been traced */
927 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
928 req_bio_endio(req, bio, bio_bytes, error);
930 total_bytes += bio_bytes;
931 nr_bytes -= bio_bytes;
942 * Reset counters so that the request stacking driver
943 * can find how many bytes remain in the request
950 req->__data_len -= total_bytes;
952 /* update sector only for requests with clear definition of sector */
953 if (!blk_rq_is_passthrough(req))
954 req->__sector += total_bytes >> 9;
956 /* mixed attributes always follow the first bio */
957 if (req->rq_flags & RQF_MIXED_MERGE) {
958 req->cmd_flags &= ~REQ_FAILFAST_MASK;
959 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
962 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
964 * If total number of sectors is less than the first segment
965 * size, something has gone terribly wrong.
967 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
968 blk_dump_rq_flags(req, "request botched");
969 req->__data_len = blk_rq_cur_bytes(req);
972 /* recalculate the number of segments */
973 req->nr_phys_segments = blk_recalc_rq_segments(req);
978 EXPORT_SYMBOL_GPL(blk_update_request);
980 static inline void blk_account_io_done(struct request *req, u64 now)
982 trace_block_io_done(req);
985 * Account IO completion. flush_rq isn't accounted as a
986 * normal IO on queueing nor completion. Accounting the
987 * containing request is enough.
989 if (blk_do_io_stat(req) && req->part &&
990 !(req->rq_flags & RQF_FLUSH_SEQ)) {
991 const int sgrp = op_stat_group(req_op(req));
994 update_io_ticks(req->part, jiffies, true);
995 part_stat_inc(req->part, ios[sgrp]);
996 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
1001 static inline void blk_account_io_start(struct request *req)
1003 trace_block_io_start(req);
1005 if (blk_do_io_stat(req)) {
1007 * All non-passthrough requests are created from a bio with one
1008 * exception: when a flush command that is part of a flush sequence
1009 * generated by the state machine in blk-flush.c is cloned onto the
1010 * lower device by dm-multipath we can get here without a bio.
1013 req->part = req->bio->bi_bdev;
1015 req->part = req->q->disk->part0;
1018 update_io_ticks(req->part, jiffies, false);
1023 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1025 if (rq->rq_flags & RQF_STATS)
1026 blk_stat_add(rq, now);
1028 blk_mq_sched_completed_request(rq, now);
1029 blk_account_io_done(rq, now);
1032 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1034 if (blk_mq_need_time_stamp(rq))
1035 __blk_mq_end_request_acct(rq, ktime_get_ns());
1037 blk_mq_finish_request(rq);
1040 rq_qos_done(rq->q, rq);
1041 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1042 blk_mq_free_request(rq);
1044 blk_mq_free_request(rq);
1047 EXPORT_SYMBOL(__blk_mq_end_request);
1049 void blk_mq_end_request(struct request *rq, blk_status_t error)
1051 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1053 __blk_mq_end_request(rq, error);
1055 EXPORT_SYMBOL(blk_mq_end_request);
1057 #define TAG_COMP_BATCH 32
1059 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1060 int *tag_array, int nr_tags)
1062 struct request_queue *q = hctx->queue;
1065 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1066 * update hctx->nr_active in batch
1068 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1069 __blk_mq_sub_active_requests(hctx, nr_tags);
1071 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1072 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1075 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1077 int tags[TAG_COMP_BATCH], nr_tags = 0;
1078 struct blk_mq_hw_ctx *cur_hctx = NULL;
1083 now = ktime_get_ns();
1085 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1087 prefetch(rq->rq_next);
1089 blk_complete_request(rq);
1091 __blk_mq_end_request_acct(rq, now);
1093 blk_mq_finish_request(rq);
1095 rq_qos_done(rq->q, rq);
1098 * If end_io handler returns NONE, then it still has
1099 * ownership of the request.
1101 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1104 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1105 if (!req_ref_put_and_test(rq))
1108 blk_crypto_free_request(rq);
1109 blk_pm_mark_last_busy(rq);
1111 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1113 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1115 cur_hctx = rq->mq_hctx;
1117 tags[nr_tags++] = rq->tag;
1121 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1123 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1125 static void blk_complete_reqs(struct llist_head *list)
1127 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1128 struct request *rq, *next;
1130 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1131 rq->q->mq_ops->complete(rq);
1134 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1136 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1139 static int blk_softirq_cpu_dead(unsigned int cpu)
1141 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1145 static void __blk_mq_complete_request_remote(void *data)
1147 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1150 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1152 int cpu = raw_smp_processor_id();
1154 if (!IS_ENABLED(CONFIG_SMP) ||
1155 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1158 * With force threaded interrupts enabled, raising softirq from an SMP
1159 * function call will always result in waking the ksoftirqd thread.
1160 * This is probably worse than completing the request on a different
1163 if (force_irqthreads())
1166 /* same CPU or cache domain? Complete locally */
1167 if (cpu == rq->mq_ctx->cpu ||
1168 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1169 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1172 /* don't try to IPI to an offline CPU */
1173 return cpu_online(rq->mq_ctx->cpu);
1176 static void blk_mq_complete_send_ipi(struct request *rq)
1180 cpu = rq->mq_ctx->cpu;
1181 if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu)))
1182 smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu));
1185 static void blk_mq_raise_softirq(struct request *rq)
1187 struct llist_head *list;
1190 list = this_cpu_ptr(&blk_cpu_done);
1191 if (llist_add(&rq->ipi_list, list))
1192 raise_softirq(BLOCK_SOFTIRQ);
1196 bool blk_mq_complete_request_remote(struct request *rq)
1198 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1201 * For request which hctx has only one ctx mapping,
1202 * or a polled request, always complete locally,
1203 * it's pointless to redirect the completion.
1205 if ((rq->mq_hctx->nr_ctx == 1 &&
1206 rq->mq_ctx->cpu == raw_smp_processor_id()) ||
1207 rq->cmd_flags & REQ_POLLED)
1210 if (blk_mq_complete_need_ipi(rq)) {
1211 blk_mq_complete_send_ipi(rq);
1215 if (rq->q->nr_hw_queues == 1) {
1216 blk_mq_raise_softirq(rq);
1221 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1224 * blk_mq_complete_request - end I/O on a request
1225 * @rq: the request being processed
1228 * Complete a request by scheduling the ->complete_rq operation.
1230 void blk_mq_complete_request(struct request *rq)
1232 if (!blk_mq_complete_request_remote(rq))
1233 rq->q->mq_ops->complete(rq);
1235 EXPORT_SYMBOL(blk_mq_complete_request);
1238 * blk_mq_start_request - Start processing a request
1239 * @rq: Pointer to request to be started
1241 * Function used by device drivers to notify the block layer that a request
1242 * is going to be processed now, so blk layer can do proper initializations
1243 * such as starting the timeout timer.
1245 void blk_mq_start_request(struct request *rq)
1247 struct request_queue *q = rq->q;
1249 trace_block_rq_issue(rq);
1251 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1252 rq->io_start_time_ns = ktime_get_ns();
1253 rq->stats_sectors = blk_rq_sectors(rq);
1254 rq->rq_flags |= RQF_STATS;
1255 rq_qos_issue(q, rq);
1258 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1261 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1263 #ifdef CONFIG_BLK_DEV_INTEGRITY
1264 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1265 q->integrity.profile->prepare_fn(rq);
1267 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1268 WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num);
1270 EXPORT_SYMBOL(blk_mq_start_request);
1273 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1274 * queues. This is important for md arrays to benefit from merging
1277 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1279 if (plug->multiple_queues)
1280 return BLK_MAX_REQUEST_COUNT * 2;
1281 return BLK_MAX_REQUEST_COUNT;
1284 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1286 struct request *last = rq_list_peek(&plug->mq_list);
1288 if (!plug->rq_count) {
1289 trace_block_plug(rq->q);
1290 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1291 (!blk_queue_nomerges(rq->q) &&
1292 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1293 blk_mq_flush_plug_list(plug, false);
1295 trace_block_plug(rq->q);
1298 if (!plug->multiple_queues && last && last->q != rq->q)
1299 plug->multiple_queues = true;
1301 * Any request allocated from sched tags can't be issued to
1302 * ->queue_rqs() directly
1304 if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))
1305 plug->has_elevator = true;
1307 rq_list_add(&plug->mq_list, rq);
1312 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1313 * @rq: request to insert
1314 * @at_head: insert request at head or tail of queue
1317 * Insert a fully prepared request at the back of the I/O scheduler queue
1318 * for execution. Don't wait for completion.
1321 * This function will invoke @done directly if the queue is dead.
1323 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1325 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1327 WARN_ON(irqs_disabled());
1328 WARN_ON(!blk_rq_is_passthrough(rq));
1330 blk_account_io_start(rq);
1333 * As plugging can be enabled for passthrough requests on a zoned
1334 * device, directly accessing the plug instead of using blk_mq_plug()
1335 * should not have any consequences.
1337 if (current->plug && !at_head) {
1338 blk_add_rq_to_plug(current->plug, rq);
1342 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1343 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
1345 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1347 struct blk_rq_wait {
1348 struct completion done;
1352 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1354 struct blk_rq_wait *wait = rq->end_io_data;
1357 complete(&wait->done);
1358 return RQ_END_IO_NONE;
1361 bool blk_rq_is_poll(struct request *rq)
1365 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1369 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1371 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1374 blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0);
1376 } while (!completion_done(wait));
1380 * blk_execute_rq - insert a request into queue for execution
1381 * @rq: request to insert
1382 * @at_head: insert request at head or tail of queue
1385 * Insert a fully prepared request at the back of the I/O scheduler queue
1386 * for execution and wait for completion.
1387 * Return: The blk_status_t result provided to blk_mq_end_request().
1389 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1391 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1392 struct blk_rq_wait wait = {
1393 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1396 WARN_ON(irqs_disabled());
1397 WARN_ON(!blk_rq_is_passthrough(rq));
1399 rq->end_io_data = &wait;
1400 rq->end_io = blk_end_sync_rq;
1402 blk_account_io_start(rq);
1403 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1404 blk_mq_run_hw_queue(hctx, false);
1406 if (blk_rq_is_poll(rq)) {
1407 blk_rq_poll_completion(rq, &wait.done);
1410 * Prevent hang_check timer from firing at us during very long
1413 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1416 while (!wait_for_completion_io_timeout(&wait.done,
1417 hang_check * (HZ/2)))
1420 wait_for_completion_io(&wait.done);
1425 EXPORT_SYMBOL(blk_execute_rq);
1427 static void __blk_mq_requeue_request(struct request *rq)
1429 struct request_queue *q = rq->q;
1431 blk_mq_put_driver_tag(rq);
1433 trace_block_rq_requeue(rq);
1434 rq_qos_requeue(q, rq);
1436 if (blk_mq_request_started(rq)) {
1437 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1438 rq->rq_flags &= ~RQF_TIMED_OUT;
1442 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1444 struct request_queue *q = rq->q;
1445 unsigned long flags;
1447 __blk_mq_requeue_request(rq);
1449 /* this request will be re-inserted to io scheduler queue */
1450 blk_mq_sched_requeue_request(rq);
1452 spin_lock_irqsave(&q->requeue_lock, flags);
1453 list_add_tail(&rq->queuelist, &q->requeue_list);
1454 spin_unlock_irqrestore(&q->requeue_lock, flags);
1456 if (kick_requeue_list)
1457 blk_mq_kick_requeue_list(q);
1459 EXPORT_SYMBOL(blk_mq_requeue_request);
1461 static void blk_mq_requeue_work(struct work_struct *work)
1463 struct request_queue *q =
1464 container_of(work, struct request_queue, requeue_work.work);
1466 LIST_HEAD(flush_list);
1469 spin_lock_irq(&q->requeue_lock);
1470 list_splice_init(&q->requeue_list, &rq_list);
1471 list_splice_init(&q->flush_list, &flush_list);
1472 spin_unlock_irq(&q->requeue_lock);
1474 while (!list_empty(&rq_list)) {
1475 rq = list_entry(rq_list.next, struct request, queuelist);
1477 * If RQF_DONTPREP ist set, the request has been started by the
1478 * driver already and might have driver-specific data allocated
1479 * already. Insert it into the hctx dispatch list to avoid
1480 * block layer merges for the request.
1482 if (rq->rq_flags & RQF_DONTPREP) {
1483 list_del_init(&rq->queuelist);
1484 blk_mq_request_bypass_insert(rq, 0);
1486 list_del_init(&rq->queuelist);
1487 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1491 while (!list_empty(&flush_list)) {
1492 rq = list_entry(flush_list.next, struct request, queuelist);
1493 list_del_init(&rq->queuelist);
1494 blk_mq_insert_request(rq, 0);
1497 blk_mq_run_hw_queues(q, false);
1500 void blk_mq_kick_requeue_list(struct request_queue *q)
1502 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1504 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1506 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1507 unsigned long msecs)
1509 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1510 msecs_to_jiffies(msecs));
1512 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1514 static bool blk_is_flush_data_rq(struct request *rq)
1516 return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq);
1519 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1522 * If we find a request that isn't idle we know the queue is busy
1523 * as it's checked in the iter.
1524 * Return false to stop the iteration.
1526 * In case of queue quiesce, if one flush data request is completed,
1527 * don't count it as inflight given the flush sequence is suspended,
1528 * and the original flush data request is invisible to driver, just
1529 * like other pending requests because of quiesce
1531 if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) &&
1532 blk_is_flush_data_rq(rq) &&
1533 blk_mq_request_completed(rq))) {
1543 bool blk_mq_queue_inflight(struct request_queue *q)
1547 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1550 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1552 static void blk_mq_rq_timed_out(struct request *req)
1554 req->rq_flags |= RQF_TIMED_OUT;
1555 if (req->q->mq_ops->timeout) {
1556 enum blk_eh_timer_return ret;
1558 ret = req->q->mq_ops->timeout(req);
1559 if (ret == BLK_EH_DONE)
1561 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1567 struct blk_expired_data {
1568 bool has_timedout_rq;
1570 unsigned long timeout_start;
1573 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1575 unsigned long deadline;
1577 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1579 if (rq->rq_flags & RQF_TIMED_OUT)
1582 deadline = READ_ONCE(rq->deadline);
1583 if (time_after_eq(expired->timeout_start, deadline))
1586 if (expired->next == 0)
1587 expired->next = deadline;
1588 else if (time_after(expired->next, deadline))
1589 expired->next = deadline;
1593 void blk_mq_put_rq_ref(struct request *rq)
1595 if (is_flush_rq(rq)) {
1596 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1597 blk_mq_free_request(rq);
1598 } else if (req_ref_put_and_test(rq)) {
1599 __blk_mq_free_request(rq);
1603 static bool blk_mq_check_expired(struct request *rq, void *priv)
1605 struct blk_expired_data *expired = priv;
1608 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1609 * be reallocated underneath the timeout handler's processing, then
1610 * the expire check is reliable. If the request is not expired, then
1611 * it was completed and reallocated as a new request after returning
1612 * from blk_mq_check_expired().
1614 if (blk_mq_req_expired(rq, expired)) {
1615 expired->has_timedout_rq = true;
1621 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1623 struct blk_expired_data *expired = priv;
1625 if (blk_mq_req_expired(rq, expired))
1626 blk_mq_rq_timed_out(rq);
1630 static void blk_mq_timeout_work(struct work_struct *work)
1632 struct request_queue *q =
1633 container_of(work, struct request_queue, timeout_work);
1634 struct blk_expired_data expired = {
1635 .timeout_start = jiffies,
1637 struct blk_mq_hw_ctx *hctx;
1640 /* A deadlock might occur if a request is stuck requiring a
1641 * timeout at the same time a queue freeze is waiting
1642 * completion, since the timeout code would not be able to
1643 * acquire the queue reference here.
1645 * That's why we don't use blk_queue_enter here; instead, we use
1646 * percpu_ref_tryget directly, because we need to be able to
1647 * obtain a reference even in the short window between the queue
1648 * starting to freeze, by dropping the first reference in
1649 * blk_freeze_queue_start, and the moment the last request is
1650 * consumed, marked by the instant q_usage_counter reaches
1653 if (!percpu_ref_tryget(&q->q_usage_counter))
1656 /* check if there is any timed-out request */
1657 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1658 if (expired.has_timedout_rq) {
1660 * Before walking tags, we must ensure any submit started
1661 * before the current time has finished. Since the submit
1662 * uses srcu or rcu, wait for a synchronization point to
1663 * ensure all running submits have finished
1665 blk_mq_wait_quiesce_done(q->tag_set);
1668 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1671 if (expired.next != 0) {
1672 mod_timer(&q->timeout, expired.next);
1675 * Request timeouts are handled as a forward rolling timer. If
1676 * we end up here it means that no requests are pending and
1677 * also that no request has been pending for a while. Mark
1678 * each hctx as idle.
1680 queue_for_each_hw_ctx(q, hctx, i) {
1681 /* the hctx may be unmapped, so check it here */
1682 if (blk_mq_hw_queue_mapped(hctx))
1683 blk_mq_tag_idle(hctx);
1689 struct flush_busy_ctx_data {
1690 struct blk_mq_hw_ctx *hctx;
1691 struct list_head *list;
1694 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1696 struct flush_busy_ctx_data *flush_data = data;
1697 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1698 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1699 enum hctx_type type = hctx->type;
1701 spin_lock(&ctx->lock);
1702 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1703 sbitmap_clear_bit(sb, bitnr);
1704 spin_unlock(&ctx->lock);
1709 * Process software queues that have been marked busy, splicing them
1710 * to the for-dispatch
1712 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1714 struct flush_busy_ctx_data data = {
1719 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1721 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1723 struct dispatch_rq_data {
1724 struct blk_mq_hw_ctx *hctx;
1728 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1731 struct dispatch_rq_data *dispatch_data = data;
1732 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1733 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1734 enum hctx_type type = hctx->type;
1736 spin_lock(&ctx->lock);
1737 if (!list_empty(&ctx->rq_lists[type])) {
1738 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1739 list_del_init(&dispatch_data->rq->queuelist);
1740 if (list_empty(&ctx->rq_lists[type]))
1741 sbitmap_clear_bit(sb, bitnr);
1743 spin_unlock(&ctx->lock);
1745 return !dispatch_data->rq;
1748 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1749 struct blk_mq_ctx *start)
1751 unsigned off = start ? start->index_hw[hctx->type] : 0;
1752 struct dispatch_rq_data data = {
1757 __sbitmap_for_each_set(&hctx->ctx_map, off,
1758 dispatch_rq_from_ctx, &data);
1763 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1765 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1766 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1769 blk_mq_tag_busy(rq->mq_hctx);
1771 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1772 bt = &rq->mq_hctx->tags->breserved_tags;
1775 if (!hctx_may_queue(rq->mq_hctx, bt))
1779 tag = __sbitmap_queue_get(bt);
1780 if (tag == BLK_MQ_NO_TAG)
1783 rq->tag = tag + tag_offset;
1787 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1789 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1792 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1793 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1794 rq->rq_flags |= RQF_MQ_INFLIGHT;
1795 __blk_mq_inc_active_requests(hctx);
1797 hctx->tags->rqs[rq->tag] = rq;
1801 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1802 int flags, void *key)
1804 struct blk_mq_hw_ctx *hctx;
1806 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1808 spin_lock(&hctx->dispatch_wait_lock);
1809 if (!list_empty(&wait->entry)) {
1810 struct sbitmap_queue *sbq;
1812 list_del_init(&wait->entry);
1813 sbq = &hctx->tags->bitmap_tags;
1814 atomic_dec(&sbq->ws_active);
1816 spin_unlock(&hctx->dispatch_wait_lock);
1818 blk_mq_run_hw_queue(hctx, true);
1823 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1824 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1825 * restart. For both cases, take care to check the condition again after
1826 * marking us as waiting.
1828 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1831 struct sbitmap_queue *sbq;
1832 struct wait_queue_head *wq;
1833 wait_queue_entry_t *wait;
1836 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1837 !(blk_mq_is_shared_tags(hctx->flags))) {
1838 blk_mq_sched_mark_restart_hctx(hctx);
1841 * It's possible that a tag was freed in the window between the
1842 * allocation failure and adding the hardware queue to the wait
1845 * Don't clear RESTART here, someone else could have set it.
1846 * At most this will cost an extra queue run.
1848 return blk_mq_get_driver_tag(rq);
1851 wait = &hctx->dispatch_wait;
1852 if (!list_empty_careful(&wait->entry))
1855 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1856 sbq = &hctx->tags->breserved_tags;
1858 sbq = &hctx->tags->bitmap_tags;
1859 wq = &bt_wait_ptr(sbq, hctx)->wait;
1861 spin_lock_irq(&wq->lock);
1862 spin_lock(&hctx->dispatch_wait_lock);
1863 if (!list_empty(&wait->entry)) {
1864 spin_unlock(&hctx->dispatch_wait_lock);
1865 spin_unlock_irq(&wq->lock);
1869 atomic_inc(&sbq->ws_active);
1870 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1871 __add_wait_queue(wq, wait);
1874 * It's possible that a tag was freed in the window between the
1875 * allocation failure and adding the hardware queue to the wait
1878 ret = blk_mq_get_driver_tag(rq);
1880 spin_unlock(&hctx->dispatch_wait_lock);
1881 spin_unlock_irq(&wq->lock);
1886 * We got a tag, remove ourselves from the wait queue to ensure
1887 * someone else gets the wakeup.
1889 list_del_init(&wait->entry);
1890 atomic_dec(&sbq->ws_active);
1891 spin_unlock(&hctx->dispatch_wait_lock);
1892 spin_unlock_irq(&wq->lock);
1897 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1898 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1900 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1901 * - EWMA is one simple way to compute running average value
1902 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1903 * - take 4 as factor for avoiding to get too small(0) result, and this
1904 * factor doesn't matter because EWMA decreases exponentially
1906 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1910 ewma = hctx->dispatch_busy;
1915 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1917 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1918 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1920 hctx->dispatch_busy = ewma;
1923 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1925 static void blk_mq_handle_dev_resource(struct request *rq,
1926 struct list_head *list)
1928 list_add(&rq->queuelist, list);
1929 __blk_mq_requeue_request(rq);
1932 static void blk_mq_handle_zone_resource(struct request *rq,
1933 struct list_head *zone_list)
1936 * If we end up here it is because we cannot dispatch a request to a
1937 * specific zone due to LLD level zone-write locking or other zone
1938 * related resource not being available. In this case, set the request
1939 * aside in zone_list for retrying it later.
1941 list_add(&rq->queuelist, zone_list);
1942 __blk_mq_requeue_request(rq);
1945 enum prep_dispatch {
1947 PREP_DISPATCH_NO_TAG,
1948 PREP_DISPATCH_NO_BUDGET,
1951 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1954 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1955 int budget_token = -1;
1958 budget_token = blk_mq_get_dispatch_budget(rq->q);
1959 if (budget_token < 0) {
1960 blk_mq_put_driver_tag(rq);
1961 return PREP_DISPATCH_NO_BUDGET;
1963 blk_mq_set_rq_budget_token(rq, budget_token);
1966 if (!blk_mq_get_driver_tag(rq)) {
1968 * The initial allocation attempt failed, so we need to
1969 * rerun the hardware queue when a tag is freed. The
1970 * waitqueue takes care of that. If the queue is run
1971 * before we add this entry back on the dispatch list,
1972 * we'll re-run it below.
1974 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1976 * All budgets not got from this function will be put
1977 * together during handling partial dispatch
1980 blk_mq_put_dispatch_budget(rq->q, budget_token);
1981 return PREP_DISPATCH_NO_TAG;
1985 return PREP_DISPATCH_OK;
1988 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1989 static void blk_mq_release_budgets(struct request_queue *q,
1990 struct list_head *list)
1994 list_for_each_entry(rq, list, queuelist) {
1995 int budget_token = blk_mq_get_rq_budget_token(rq);
1997 if (budget_token >= 0)
1998 blk_mq_put_dispatch_budget(q, budget_token);
2003 * blk_mq_commit_rqs will notify driver using bd->last that there is no
2004 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
2006 * Attention, we should explicitly call this in unusual cases:
2007 * 1) did not queue everything initially scheduled to queue
2008 * 2) the last attempt to queue a request failed
2010 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
2013 if (hctx->queue->mq_ops->commit_rqs && queued) {
2014 trace_block_unplug(hctx->queue, queued, !from_schedule);
2015 hctx->queue->mq_ops->commit_rqs(hctx);
2020 * Returns true if we did some work AND can potentially do more.
2022 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
2023 unsigned int nr_budgets)
2025 enum prep_dispatch prep;
2026 struct request_queue *q = hctx->queue;
2029 blk_status_t ret = BLK_STS_OK;
2030 LIST_HEAD(zone_list);
2031 bool needs_resource = false;
2033 if (list_empty(list))
2037 * Now process all the entries, sending them to the driver.
2041 struct blk_mq_queue_data bd;
2043 rq = list_first_entry(list, struct request, queuelist);
2045 WARN_ON_ONCE(hctx != rq->mq_hctx);
2046 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2047 if (prep != PREP_DISPATCH_OK)
2050 list_del_init(&rq->queuelist);
2053 bd.last = list_empty(list);
2056 * once the request is queued to lld, no need to cover the
2061 ret = q->mq_ops->queue_rq(hctx, &bd);
2066 case BLK_STS_RESOURCE:
2067 needs_resource = true;
2069 case BLK_STS_DEV_RESOURCE:
2070 blk_mq_handle_dev_resource(rq, list);
2072 case BLK_STS_ZONE_RESOURCE:
2074 * Move the request to zone_list and keep going through
2075 * the dispatch list to find more requests the drive can
2078 blk_mq_handle_zone_resource(rq, &zone_list);
2079 needs_resource = true;
2082 blk_mq_end_request(rq, ret);
2084 } while (!list_empty(list));
2086 if (!list_empty(&zone_list))
2087 list_splice_tail_init(&zone_list, list);
2089 /* If we didn't flush the entire list, we could have told the driver
2090 * there was more coming, but that turned out to be a lie.
2092 if (!list_empty(list) || ret != BLK_STS_OK)
2093 blk_mq_commit_rqs(hctx, queued, false);
2096 * Any items that need requeuing? Stuff them into hctx->dispatch,
2097 * that is where we will continue on next queue run.
2099 if (!list_empty(list)) {
2101 /* For non-shared tags, the RESTART check will suffice */
2102 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2103 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2104 blk_mq_is_shared_tags(hctx->flags));
2107 blk_mq_release_budgets(q, list);
2109 spin_lock(&hctx->lock);
2110 list_splice_tail_init(list, &hctx->dispatch);
2111 spin_unlock(&hctx->lock);
2114 * Order adding requests to hctx->dispatch and checking
2115 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2116 * in blk_mq_sched_restart(). Avoid restart code path to
2117 * miss the new added requests to hctx->dispatch, meantime
2118 * SCHED_RESTART is observed here.
2123 * If SCHED_RESTART was set by the caller of this function and
2124 * it is no longer set that means that it was cleared by another
2125 * thread and hence that a queue rerun is needed.
2127 * If 'no_tag' is set, that means that we failed getting
2128 * a driver tag with an I/O scheduler attached. If our dispatch
2129 * waitqueue is no longer active, ensure that we run the queue
2130 * AFTER adding our entries back to the list.
2132 * If no I/O scheduler has been configured it is possible that
2133 * the hardware queue got stopped and restarted before requests
2134 * were pushed back onto the dispatch list. Rerun the queue to
2135 * avoid starvation. Notes:
2136 * - blk_mq_run_hw_queue() checks whether or not a queue has
2137 * been stopped before rerunning a queue.
2138 * - Some but not all block drivers stop a queue before
2139 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2142 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2143 * bit is set, run queue after a delay to avoid IO stalls
2144 * that could otherwise occur if the queue is idle. We'll do
2145 * similar if we couldn't get budget or couldn't lock a zone
2146 * and SCHED_RESTART is set.
2148 needs_restart = blk_mq_sched_needs_restart(hctx);
2149 if (prep == PREP_DISPATCH_NO_BUDGET)
2150 needs_resource = true;
2151 if (!needs_restart ||
2152 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2153 blk_mq_run_hw_queue(hctx, true);
2154 else if (needs_resource)
2155 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2157 blk_mq_update_dispatch_busy(hctx, true);
2161 blk_mq_update_dispatch_busy(hctx, false);
2165 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2167 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2169 if (cpu >= nr_cpu_ids)
2170 cpu = cpumask_first(hctx->cpumask);
2175 * It'd be great if the workqueue API had a way to pass
2176 * in a mask and had some smarts for more clever placement.
2177 * For now we just round-robin here, switching for every
2178 * BLK_MQ_CPU_WORK_BATCH queued items.
2180 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2183 int next_cpu = hctx->next_cpu;
2185 if (hctx->queue->nr_hw_queues == 1)
2186 return WORK_CPU_UNBOUND;
2188 if (--hctx->next_cpu_batch <= 0) {
2190 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2192 if (next_cpu >= nr_cpu_ids)
2193 next_cpu = blk_mq_first_mapped_cpu(hctx);
2194 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2198 * Do unbound schedule if we can't find a online CPU for this hctx,
2199 * and it should only happen in the path of handling CPU DEAD.
2201 if (!cpu_online(next_cpu)) {
2208 * Make sure to re-select CPU next time once after CPUs
2209 * in hctx->cpumask become online again.
2211 hctx->next_cpu = next_cpu;
2212 hctx->next_cpu_batch = 1;
2213 return WORK_CPU_UNBOUND;
2216 hctx->next_cpu = next_cpu;
2221 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2222 * @hctx: Pointer to the hardware queue to run.
2223 * @msecs: Milliseconds of delay to wait before running the queue.
2225 * Run a hardware queue asynchronously with a delay of @msecs.
2227 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2229 if (unlikely(blk_mq_hctx_stopped(hctx)))
2231 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2232 msecs_to_jiffies(msecs));
2234 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2237 * blk_mq_run_hw_queue - Start to run a hardware queue.
2238 * @hctx: Pointer to the hardware queue to run.
2239 * @async: If we want to run the queue asynchronously.
2241 * Check if the request queue is not in a quiesced state and if there are
2242 * pending requests to be sent. If this is true, run the queue to send requests
2245 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2250 * We can't run the queue inline with interrupts disabled.
2252 WARN_ON_ONCE(!async && in_interrupt());
2254 might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING);
2257 * When queue is quiesced, we may be switching io scheduler, or
2258 * updating nr_hw_queues, or other things, and we can't run queue
2259 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2261 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2264 __blk_mq_run_dispatch_ops(hctx->queue, false,
2265 need_run = !blk_queue_quiesced(hctx->queue) &&
2266 blk_mq_hctx_has_pending(hctx));
2271 if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2272 blk_mq_delay_run_hw_queue(hctx, 0);
2276 blk_mq_run_dispatch_ops(hctx->queue,
2277 blk_mq_sched_dispatch_requests(hctx));
2279 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2282 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2285 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2287 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2289 * If the IO scheduler does not respect hardware queues when
2290 * dispatching, we just don't bother with multiple HW queues and
2291 * dispatch from hctx for the current CPU since running multiple queues
2292 * just causes lock contention inside the scheduler and pointless cache
2295 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2297 if (!blk_mq_hctx_stopped(hctx))
2303 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2304 * @q: Pointer to the request queue to run.
2305 * @async: If we want to run the queue asynchronously.
2307 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2309 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2313 if (blk_queue_sq_sched(q))
2314 sq_hctx = blk_mq_get_sq_hctx(q);
2315 queue_for_each_hw_ctx(q, hctx, i) {
2316 if (blk_mq_hctx_stopped(hctx))
2319 * Dispatch from this hctx either if there's no hctx preferred
2320 * by IO scheduler or if it has requests that bypass the
2323 if (!sq_hctx || sq_hctx == hctx ||
2324 !list_empty_careful(&hctx->dispatch))
2325 blk_mq_run_hw_queue(hctx, async);
2328 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2331 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2332 * @q: Pointer to the request queue to run.
2333 * @msecs: Milliseconds of delay to wait before running the queues.
2335 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2337 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2341 if (blk_queue_sq_sched(q))
2342 sq_hctx = blk_mq_get_sq_hctx(q);
2343 queue_for_each_hw_ctx(q, hctx, i) {
2344 if (blk_mq_hctx_stopped(hctx))
2347 * If there is already a run_work pending, leave the
2348 * pending delay untouched. Otherwise, a hctx can stall
2349 * if another hctx is re-delaying the other's work
2350 * before the work executes.
2352 if (delayed_work_pending(&hctx->run_work))
2355 * Dispatch from this hctx either if there's no hctx preferred
2356 * by IO scheduler or if it has requests that bypass the
2359 if (!sq_hctx || sq_hctx == hctx ||
2360 !list_empty_careful(&hctx->dispatch))
2361 blk_mq_delay_run_hw_queue(hctx, msecs);
2364 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2367 * This function is often used for pausing .queue_rq() by driver when
2368 * there isn't enough resource or some conditions aren't satisfied, and
2369 * BLK_STS_RESOURCE is usually returned.
2371 * We do not guarantee that dispatch can be drained or blocked
2372 * after blk_mq_stop_hw_queue() returns. Please use
2373 * blk_mq_quiesce_queue() for that requirement.
2375 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2377 cancel_delayed_work(&hctx->run_work);
2379 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2381 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2384 * This function is often used for pausing .queue_rq() by driver when
2385 * there isn't enough resource or some conditions aren't satisfied, and
2386 * BLK_STS_RESOURCE is usually returned.
2388 * We do not guarantee that dispatch can be drained or blocked
2389 * after blk_mq_stop_hw_queues() returns. Please use
2390 * blk_mq_quiesce_queue() for that requirement.
2392 void blk_mq_stop_hw_queues(struct request_queue *q)
2394 struct blk_mq_hw_ctx *hctx;
2397 queue_for_each_hw_ctx(q, hctx, i)
2398 blk_mq_stop_hw_queue(hctx);
2400 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2402 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2404 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2406 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
2408 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2410 void blk_mq_start_hw_queues(struct request_queue *q)
2412 struct blk_mq_hw_ctx *hctx;
2415 queue_for_each_hw_ctx(q, hctx, i)
2416 blk_mq_start_hw_queue(hctx);
2418 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2420 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2422 if (!blk_mq_hctx_stopped(hctx))
2425 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2426 blk_mq_run_hw_queue(hctx, async);
2428 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2430 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2432 struct blk_mq_hw_ctx *hctx;
2435 queue_for_each_hw_ctx(q, hctx, i)
2436 blk_mq_start_stopped_hw_queue(hctx, async ||
2437 (hctx->flags & BLK_MQ_F_BLOCKING));
2439 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2441 static void blk_mq_run_work_fn(struct work_struct *work)
2443 struct blk_mq_hw_ctx *hctx =
2444 container_of(work, struct blk_mq_hw_ctx, run_work.work);
2446 blk_mq_run_dispatch_ops(hctx->queue,
2447 blk_mq_sched_dispatch_requests(hctx));
2451 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2452 * @rq: Pointer to request to be inserted.
2453 * @flags: BLK_MQ_INSERT_*
2455 * Should only be used carefully, when the caller knows we want to
2456 * bypass a potential IO scheduler on the target device.
2458 static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2460 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2462 spin_lock(&hctx->lock);
2463 if (flags & BLK_MQ_INSERT_AT_HEAD)
2464 list_add(&rq->queuelist, &hctx->dispatch);
2466 list_add_tail(&rq->queuelist, &hctx->dispatch);
2467 spin_unlock(&hctx->lock);
2470 static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2471 struct blk_mq_ctx *ctx, struct list_head *list,
2472 bool run_queue_async)
2475 enum hctx_type type = hctx->type;
2478 * Try to issue requests directly if the hw queue isn't busy to save an
2479 * extra enqueue & dequeue to the sw queue.
2481 if (!hctx->dispatch_busy && !run_queue_async) {
2482 blk_mq_run_dispatch_ops(hctx->queue,
2483 blk_mq_try_issue_list_directly(hctx, list));
2484 if (list_empty(list))
2489 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2492 list_for_each_entry(rq, list, queuelist) {
2493 BUG_ON(rq->mq_ctx != ctx);
2494 trace_block_rq_insert(rq);
2495 if (rq->cmd_flags & REQ_NOWAIT)
2496 run_queue_async = true;
2499 spin_lock(&ctx->lock);
2500 list_splice_tail_init(list, &ctx->rq_lists[type]);
2501 blk_mq_hctx_mark_pending(hctx, ctx);
2502 spin_unlock(&ctx->lock);
2504 blk_mq_run_hw_queue(hctx, run_queue_async);
2507 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2509 struct request_queue *q = rq->q;
2510 struct blk_mq_ctx *ctx = rq->mq_ctx;
2511 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2513 if (blk_rq_is_passthrough(rq)) {
2515 * Passthrough request have to be added to hctx->dispatch
2516 * directly. The device may be in a situation where it can't
2517 * handle FS request, and always returns BLK_STS_RESOURCE for
2518 * them, which gets them added to hctx->dispatch.
2520 * If a passthrough request is required to unblock the queues,
2521 * and it is added to the scheduler queue, there is no chance to
2522 * dispatch it given we prioritize requests in hctx->dispatch.
2524 blk_mq_request_bypass_insert(rq, flags);
2525 } else if (req_op(rq) == REQ_OP_FLUSH) {
2527 * Firstly normal IO request is inserted to scheduler queue or
2528 * sw queue, meantime we add flush request to dispatch queue(
2529 * hctx->dispatch) directly and there is at most one in-flight
2530 * flush request for each hw queue, so it doesn't matter to add
2531 * flush request to tail or front of the dispatch queue.
2533 * Secondly in case of NCQ, flush request belongs to non-NCQ
2534 * command, and queueing it will fail when there is any
2535 * in-flight normal IO request(NCQ command). When adding flush
2536 * rq to the front of hctx->dispatch, it is easier to introduce
2537 * extra time to flush rq's latency because of S_SCHED_RESTART
2538 * compared with adding to the tail of dispatch queue, then
2539 * chance of flush merge is increased, and less flush requests
2540 * will be issued to controller. It is observed that ~10% time
2541 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2542 * drive when adding flush rq to the front of hctx->dispatch.
2544 * Simply queue flush rq to the front of hctx->dispatch so that
2545 * intensive flush workloads can benefit in case of NCQ HW.
2547 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2548 } else if (q->elevator) {
2551 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2553 list_add(&rq->queuelist, &list);
2554 q->elevator->type->ops.insert_requests(hctx, &list, flags);
2556 trace_block_rq_insert(rq);
2558 spin_lock(&ctx->lock);
2559 if (flags & BLK_MQ_INSERT_AT_HEAD)
2560 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2562 list_add_tail(&rq->queuelist,
2563 &ctx->rq_lists[hctx->type]);
2564 blk_mq_hctx_mark_pending(hctx, ctx);
2565 spin_unlock(&ctx->lock);
2569 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2570 unsigned int nr_segs)
2574 if (bio->bi_opf & REQ_RAHEAD)
2575 rq->cmd_flags |= REQ_FAILFAST_MASK;
2577 rq->__sector = bio->bi_iter.bi_sector;
2578 blk_rq_bio_prep(rq, bio, nr_segs);
2580 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2581 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2584 blk_account_io_start(rq);
2587 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2588 struct request *rq, bool last)
2590 struct request_queue *q = rq->q;
2591 struct blk_mq_queue_data bd = {
2598 * For OK queue, we are done. For error, caller may kill it.
2599 * Any other error (busy), just add it to our list as we
2600 * previously would have done.
2602 ret = q->mq_ops->queue_rq(hctx, &bd);
2605 blk_mq_update_dispatch_busy(hctx, false);
2607 case BLK_STS_RESOURCE:
2608 case BLK_STS_DEV_RESOURCE:
2609 blk_mq_update_dispatch_busy(hctx, true);
2610 __blk_mq_requeue_request(rq);
2613 blk_mq_update_dispatch_busy(hctx, false);
2620 static bool blk_mq_get_budget_and_tag(struct request *rq)
2624 budget_token = blk_mq_get_dispatch_budget(rq->q);
2625 if (budget_token < 0)
2627 blk_mq_set_rq_budget_token(rq, budget_token);
2628 if (!blk_mq_get_driver_tag(rq)) {
2629 blk_mq_put_dispatch_budget(rq->q, budget_token);
2636 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2637 * @hctx: Pointer of the associated hardware queue.
2638 * @rq: Pointer to request to be sent.
2640 * If the device has enough resources to accept a new request now, send the
2641 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2642 * we can try send it another time in the future. Requests inserted at this
2643 * queue have higher priority.
2645 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2650 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2651 blk_mq_insert_request(rq, 0);
2655 if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
2656 blk_mq_insert_request(rq, 0);
2657 blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT);
2661 ret = __blk_mq_issue_directly(hctx, rq, true);
2665 case BLK_STS_RESOURCE:
2666 case BLK_STS_DEV_RESOURCE:
2667 blk_mq_request_bypass_insert(rq, 0);
2668 blk_mq_run_hw_queue(hctx, false);
2671 blk_mq_end_request(rq, ret);
2676 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2678 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2680 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2681 blk_mq_insert_request(rq, 0);
2685 if (!blk_mq_get_budget_and_tag(rq))
2686 return BLK_STS_RESOURCE;
2687 return __blk_mq_issue_directly(hctx, rq, last);
2690 static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2692 struct blk_mq_hw_ctx *hctx = NULL;
2695 blk_status_t ret = BLK_STS_OK;
2697 while ((rq = rq_list_pop(&plug->mq_list))) {
2698 bool last = rq_list_empty(plug->mq_list);
2700 if (hctx != rq->mq_hctx) {
2702 blk_mq_commit_rqs(hctx, queued, false);
2708 ret = blk_mq_request_issue_directly(rq, last);
2713 case BLK_STS_RESOURCE:
2714 case BLK_STS_DEV_RESOURCE:
2715 blk_mq_request_bypass_insert(rq, 0);
2716 blk_mq_run_hw_queue(hctx, false);
2719 blk_mq_end_request(rq, ret);
2725 if (ret != BLK_STS_OK)
2726 blk_mq_commit_rqs(hctx, queued, false);
2729 static void __blk_mq_flush_plug_list(struct request_queue *q,
2730 struct blk_plug *plug)
2732 if (blk_queue_quiesced(q))
2734 q->mq_ops->queue_rqs(&plug->mq_list);
2737 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2739 struct blk_mq_hw_ctx *this_hctx = NULL;
2740 struct blk_mq_ctx *this_ctx = NULL;
2741 struct request *requeue_list = NULL;
2742 struct request **requeue_lastp = &requeue_list;
2743 unsigned int depth = 0;
2744 bool is_passthrough = false;
2748 struct request *rq = rq_list_pop(&plug->mq_list);
2751 this_hctx = rq->mq_hctx;
2752 this_ctx = rq->mq_ctx;
2753 is_passthrough = blk_rq_is_passthrough(rq);
2754 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2755 is_passthrough != blk_rq_is_passthrough(rq)) {
2756 rq_list_add_tail(&requeue_lastp, rq);
2759 list_add(&rq->queuelist, &list);
2761 } while (!rq_list_empty(plug->mq_list));
2763 plug->mq_list = requeue_list;
2764 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2766 percpu_ref_get(&this_hctx->queue->q_usage_counter);
2767 /* passthrough requests should never be issued to the I/O scheduler */
2768 if (is_passthrough) {
2769 spin_lock(&this_hctx->lock);
2770 list_splice_tail_init(&list, &this_hctx->dispatch);
2771 spin_unlock(&this_hctx->lock);
2772 blk_mq_run_hw_queue(this_hctx, from_sched);
2773 } else if (this_hctx->queue->elevator) {
2774 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2776 blk_mq_run_hw_queue(this_hctx, from_sched);
2778 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2780 percpu_ref_put(&this_hctx->queue->q_usage_counter);
2783 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2788 * We may have been called recursively midway through handling
2789 * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2790 * To avoid mq_list changing under our feet, clear rq_count early and
2791 * bail out specifically if rq_count is 0 rather than checking
2792 * whether the mq_list is empty.
2794 if (plug->rq_count == 0)
2798 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2799 struct request_queue *q;
2801 rq = rq_list_peek(&plug->mq_list);
2805 * Peek first request and see if we have a ->queue_rqs() hook.
2806 * If we do, we can dispatch the whole plug list in one go. We
2807 * already know at this point that all requests belong to the
2808 * same queue, caller must ensure that's the case.
2810 * Since we pass off the full list to the driver at this point,
2811 * we do not increment the active request count for the queue.
2812 * Bypass shared tags for now because of that.
2814 if (q->mq_ops->queue_rqs &&
2815 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2816 blk_mq_run_dispatch_ops(q,
2817 __blk_mq_flush_plug_list(q, plug));
2818 if (rq_list_empty(plug->mq_list))
2822 blk_mq_run_dispatch_ops(q,
2823 blk_mq_plug_issue_direct(plug));
2824 if (rq_list_empty(plug->mq_list))
2829 blk_mq_dispatch_plug_list(plug, from_schedule);
2830 } while (!rq_list_empty(plug->mq_list));
2833 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2834 struct list_head *list)
2837 blk_status_t ret = BLK_STS_OK;
2839 while (!list_empty(list)) {
2840 struct request *rq = list_first_entry(list, struct request,
2843 list_del_init(&rq->queuelist);
2844 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2849 case BLK_STS_RESOURCE:
2850 case BLK_STS_DEV_RESOURCE:
2851 blk_mq_request_bypass_insert(rq, 0);
2852 if (list_empty(list))
2853 blk_mq_run_hw_queue(hctx, false);
2856 blk_mq_end_request(rq, ret);
2862 if (ret != BLK_STS_OK)
2863 blk_mq_commit_rqs(hctx, queued, false);
2866 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2867 struct bio *bio, unsigned int nr_segs)
2869 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2870 if (blk_attempt_plug_merge(q, bio, nr_segs))
2872 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2878 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2879 struct blk_plug *plug,
2883 struct blk_mq_alloc_data data = {
2886 .cmd_flags = bio->bi_opf,
2890 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2893 rq_qos_throttle(q, bio);
2896 data.nr_tags = plug->nr_ios;
2898 data.cached_rq = &plug->cached_rq;
2901 rq = __blk_mq_alloc_requests(&data);
2904 rq_qos_cleanup(q, bio);
2905 if (bio->bi_opf & REQ_NOWAIT)
2906 bio_wouldblock_error(bio);
2910 /* return true if this @rq can be used for @bio */
2911 static bool blk_mq_can_use_cached_rq(struct request *rq, struct blk_plug *plug,
2914 enum hctx_type type = blk_mq_get_hctx_type(bio->bi_opf);
2915 enum hctx_type hctx_type = rq->mq_hctx->type;
2917 WARN_ON_ONCE(rq_list_peek(&plug->cached_rq) != rq);
2919 if (type != hctx_type &&
2920 !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2922 if (op_is_flush(rq->cmd_flags) != op_is_flush(bio->bi_opf))
2926 * If any qos ->throttle() end up blocking, we will have flushed the
2927 * plug and hence killed the cached_rq list as well. Pop this entry
2928 * before we throttle.
2930 plug->cached_rq = rq_list_next(rq);
2931 rq_qos_throttle(rq->q, bio);
2933 blk_mq_rq_time_init(rq, 0);
2934 rq->cmd_flags = bio->bi_opf;
2935 INIT_LIST_HEAD(&rq->queuelist);
2939 static void bio_set_ioprio(struct bio *bio)
2941 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2942 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2943 bio->bi_ioprio = get_current_ioprio();
2944 blkcg_set_ioprio(bio);
2948 * blk_mq_submit_bio - Create and send a request to block device.
2949 * @bio: Bio pointer.
2951 * Builds up a request structure from @q and @bio and send to the device. The
2952 * request may not be queued directly to hardware if:
2953 * * This request can be merged with another one
2954 * * We want to place request at plug queue for possible future merging
2955 * * There is an IO scheduler active at this queue
2957 * It will not queue the request if there is an error with the bio, or at the
2960 void blk_mq_submit_bio(struct bio *bio)
2962 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2963 struct blk_plug *plug = blk_mq_plug(bio);
2964 const int is_sync = op_is_sync(bio->bi_opf);
2965 struct blk_mq_hw_ctx *hctx;
2966 struct request *rq = NULL;
2967 unsigned int nr_segs = 1;
2970 bio = blk_queue_bounce(bio, q);
2971 bio_set_ioprio(bio);
2974 rq = rq_list_peek(&plug->cached_rq);
2975 if (rq && rq->q != q)
2979 if (unlikely(bio_may_exceed_limits(bio, &q->limits))) {
2980 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2984 if (!bio_integrity_prep(bio))
2986 if (blk_mq_attempt_bio_merge(q, bio, nr_segs))
2988 if (blk_mq_can_use_cached_rq(rq, plug, bio))
2990 percpu_ref_get(&q->q_usage_counter);
2992 if (unlikely(bio_queue_enter(bio)))
2994 if (unlikely(bio_may_exceed_limits(bio, &q->limits))) {
2995 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2999 if (!bio_integrity_prep(bio))
3003 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
3004 if (unlikely(!rq)) {
3011 trace_block_getrq(bio);
3013 rq_qos_track(q, rq, bio);
3015 blk_mq_bio_to_request(rq, bio, nr_segs);
3017 ret = blk_crypto_rq_get_keyslot(rq);
3018 if (ret != BLK_STS_OK) {
3019 bio->bi_status = ret;
3021 blk_mq_free_request(rq);
3025 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
3029 blk_add_rq_to_plug(plug, rq);
3034 if ((rq->rq_flags & RQF_USE_SCHED) ||
3035 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
3036 blk_mq_insert_request(rq, 0);
3037 blk_mq_run_hw_queue(hctx, true);
3039 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
3043 #ifdef CONFIG_BLK_MQ_STACKING
3045 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3046 * @rq: the request being queued
3048 blk_status_t blk_insert_cloned_request(struct request *rq)
3050 struct request_queue *q = rq->q;
3051 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3052 unsigned int max_segments = blk_rq_get_max_segments(rq);
3055 if (blk_rq_sectors(rq) > max_sectors) {
3057 * SCSI device does not have a good way to return if
3058 * Write Same/Zero is actually supported. If a device rejects
3059 * a non-read/write command (discard, write same,etc.) the
3060 * low-level device driver will set the relevant queue limit to
3061 * 0 to prevent blk-lib from issuing more of the offending
3062 * operations. Commands queued prior to the queue limit being
3063 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3064 * errors being propagated to upper layers.
3066 if (max_sectors == 0)
3067 return BLK_STS_NOTSUPP;
3069 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3070 __func__, blk_rq_sectors(rq), max_sectors);
3071 return BLK_STS_IOERR;
3075 * The queue settings related to segment counting may differ from the
3078 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3079 if (rq->nr_phys_segments > max_segments) {
3080 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3081 __func__, rq->nr_phys_segments, max_segments);
3082 return BLK_STS_IOERR;
3085 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3086 return BLK_STS_IOERR;
3088 ret = blk_crypto_rq_get_keyslot(rq);
3089 if (ret != BLK_STS_OK)
3092 blk_account_io_start(rq);
3095 * Since we have a scheduler attached on the top device,
3096 * bypass a potential scheduler on the bottom device for
3099 blk_mq_run_dispatch_ops(q,
3100 ret = blk_mq_request_issue_directly(rq, true));
3102 blk_account_io_done(rq, ktime_get_ns());
3105 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3108 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3109 * @rq: the clone request to be cleaned up
3112 * Free all bios in @rq for a cloned request.
3114 void blk_rq_unprep_clone(struct request *rq)
3118 while ((bio = rq->bio) != NULL) {
3119 rq->bio = bio->bi_next;
3124 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3127 * blk_rq_prep_clone - Helper function to setup clone request
3128 * @rq: the request to be setup
3129 * @rq_src: original request to be cloned
3130 * @bs: bio_set that bios for clone are allocated from
3131 * @gfp_mask: memory allocation mask for bio
3132 * @bio_ctr: setup function to be called for each clone bio.
3133 * Returns %0 for success, non %0 for failure.
3134 * @data: private data to be passed to @bio_ctr
3137 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3138 * Also, pages which the original bios are pointing to are not copied
3139 * and the cloned bios just point same pages.
3140 * So cloned bios must be completed before original bios, which means
3141 * the caller must complete @rq before @rq_src.
3143 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3144 struct bio_set *bs, gfp_t gfp_mask,
3145 int (*bio_ctr)(struct bio *, struct bio *, void *),
3148 struct bio *bio, *bio_src;
3153 __rq_for_each_bio(bio_src, rq_src) {
3154 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3159 if (bio_ctr && bio_ctr(bio, bio_src, data))
3163 rq->biotail->bi_next = bio;
3166 rq->bio = rq->biotail = bio;
3171 /* Copy attributes of the original request to the clone request. */
3172 rq->__sector = blk_rq_pos(rq_src);
3173 rq->__data_len = blk_rq_bytes(rq_src);
3174 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3175 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3176 rq->special_vec = rq_src->special_vec;
3178 rq->nr_phys_segments = rq_src->nr_phys_segments;
3179 rq->ioprio = rq_src->ioprio;
3181 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3189 blk_rq_unprep_clone(rq);
3193 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3194 #endif /* CONFIG_BLK_MQ_STACKING */
3197 * Steal bios from a request and add them to a bio list.
3198 * The request must not have been partially completed before.
3200 void blk_steal_bios(struct bio_list *list, struct request *rq)
3204 list->tail->bi_next = rq->bio;
3206 list->head = rq->bio;
3207 list->tail = rq->biotail;
3215 EXPORT_SYMBOL_GPL(blk_steal_bios);
3217 static size_t order_to_size(unsigned int order)
3219 return (size_t)PAGE_SIZE << order;
3222 /* called before freeing request pool in @tags */
3223 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3224 struct blk_mq_tags *tags)
3227 unsigned long flags;
3230 * There is no need to clear mapping if driver tags is not initialized
3231 * or the mapping belongs to the driver tags.
3233 if (!drv_tags || drv_tags == tags)
3236 list_for_each_entry(page, &tags->page_list, lru) {
3237 unsigned long start = (unsigned long)page_address(page);
3238 unsigned long end = start + order_to_size(page->private);
3241 for (i = 0; i < drv_tags->nr_tags; i++) {
3242 struct request *rq = drv_tags->rqs[i];
3243 unsigned long rq_addr = (unsigned long)rq;
3245 if (rq_addr >= start && rq_addr < end) {
3246 WARN_ON_ONCE(req_ref_read(rq) != 0);
3247 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3253 * Wait until all pending iteration is done.
3255 * Request reference is cleared and it is guaranteed to be observed
3256 * after the ->lock is released.
3258 spin_lock_irqsave(&drv_tags->lock, flags);
3259 spin_unlock_irqrestore(&drv_tags->lock, flags);
3262 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3263 unsigned int hctx_idx)
3265 struct blk_mq_tags *drv_tags;
3268 if (list_empty(&tags->page_list))
3271 if (blk_mq_is_shared_tags(set->flags))
3272 drv_tags = set->shared_tags;
3274 drv_tags = set->tags[hctx_idx];
3276 if (tags->static_rqs && set->ops->exit_request) {
3279 for (i = 0; i < tags->nr_tags; i++) {
3280 struct request *rq = tags->static_rqs[i];
3284 set->ops->exit_request(set, rq, hctx_idx);
3285 tags->static_rqs[i] = NULL;
3289 blk_mq_clear_rq_mapping(drv_tags, tags);
3291 while (!list_empty(&tags->page_list)) {
3292 page = list_first_entry(&tags->page_list, struct page, lru);
3293 list_del_init(&page->lru);
3295 * Remove kmemleak object previously allocated in
3296 * blk_mq_alloc_rqs().
3298 kmemleak_free(page_address(page));
3299 __free_pages(page, page->private);
3303 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3307 kfree(tags->static_rqs);
3308 tags->static_rqs = NULL;
3310 blk_mq_free_tags(tags);
3313 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3314 unsigned int hctx_idx)
3318 for (i = 0; i < set->nr_maps; i++) {
3319 unsigned int start = set->map[i].queue_offset;
3320 unsigned int end = start + set->map[i].nr_queues;
3322 if (hctx_idx >= start && hctx_idx < end)
3326 if (i >= set->nr_maps)
3327 i = HCTX_TYPE_DEFAULT;
3332 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3333 unsigned int hctx_idx)
3335 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3337 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3340 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3341 unsigned int hctx_idx,
3342 unsigned int nr_tags,
3343 unsigned int reserved_tags)
3345 int node = blk_mq_get_hctx_node(set, hctx_idx);
3346 struct blk_mq_tags *tags;
3348 if (node == NUMA_NO_NODE)
3349 node = set->numa_node;
3351 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3352 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3356 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3357 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3362 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3363 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3365 if (!tags->static_rqs)
3373 blk_mq_free_tags(tags);
3377 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3378 unsigned int hctx_idx, int node)
3382 if (set->ops->init_request) {
3383 ret = set->ops->init_request(set, rq, hctx_idx, node);
3388 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3392 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3393 struct blk_mq_tags *tags,
3394 unsigned int hctx_idx, unsigned int depth)
3396 unsigned int i, j, entries_per_page, max_order = 4;
3397 int node = blk_mq_get_hctx_node(set, hctx_idx);
3398 size_t rq_size, left;
3400 if (node == NUMA_NO_NODE)
3401 node = set->numa_node;
3403 INIT_LIST_HEAD(&tags->page_list);
3406 * rq_size is the size of the request plus driver payload, rounded
3407 * to the cacheline size
3409 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3411 left = rq_size * depth;
3413 for (i = 0; i < depth; ) {
3414 int this_order = max_order;
3419 while (this_order && left < order_to_size(this_order - 1))
3423 page = alloc_pages_node(node,
3424 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3430 if (order_to_size(this_order) < rq_size)
3437 page->private = this_order;
3438 list_add_tail(&page->lru, &tags->page_list);
3440 p = page_address(page);
3442 * Allow kmemleak to scan these pages as they contain pointers
3443 * to additional allocations like via ops->init_request().
3445 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3446 entries_per_page = order_to_size(this_order) / rq_size;
3447 to_do = min(entries_per_page, depth - i);
3448 left -= to_do * rq_size;
3449 for (j = 0; j < to_do; j++) {
3450 struct request *rq = p;
3452 tags->static_rqs[i] = rq;
3453 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3454 tags->static_rqs[i] = NULL;
3465 blk_mq_free_rqs(set, tags, hctx_idx);
3469 struct rq_iter_data {
3470 struct blk_mq_hw_ctx *hctx;
3474 static bool blk_mq_has_request(struct request *rq, void *data)
3476 struct rq_iter_data *iter_data = data;
3478 if (rq->mq_hctx != iter_data->hctx)
3480 iter_data->has_rq = true;
3484 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3486 struct blk_mq_tags *tags = hctx->sched_tags ?
3487 hctx->sched_tags : hctx->tags;
3488 struct rq_iter_data data = {
3492 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3496 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3497 struct blk_mq_hw_ctx *hctx)
3499 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3501 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3506 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3508 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3509 struct blk_mq_hw_ctx, cpuhp_online);
3511 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3512 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3516 * Prevent new request from being allocated on the current hctx.
3518 * The smp_mb__after_atomic() Pairs with the implied barrier in
3519 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3520 * seen once we return from the tag allocator.
3522 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3523 smp_mb__after_atomic();
3526 * Try to grab a reference to the queue and wait for any outstanding
3527 * requests. If we could not grab a reference the queue has been
3528 * frozen and there are no requests.
3530 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3531 while (blk_mq_hctx_has_requests(hctx))
3533 percpu_ref_put(&hctx->queue->q_usage_counter);
3539 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3541 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3542 struct blk_mq_hw_ctx, cpuhp_online);
3544 if (cpumask_test_cpu(cpu, hctx->cpumask))
3545 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3550 * 'cpu' is going away. splice any existing rq_list entries from this
3551 * software queue to the hw queue dispatch list, and ensure that it
3554 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3556 struct blk_mq_hw_ctx *hctx;
3557 struct blk_mq_ctx *ctx;
3559 enum hctx_type type;
3561 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3562 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3565 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3568 spin_lock(&ctx->lock);
3569 if (!list_empty(&ctx->rq_lists[type])) {
3570 list_splice_init(&ctx->rq_lists[type], &tmp);
3571 blk_mq_hctx_clear_pending(hctx, ctx);
3573 spin_unlock(&ctx->lock);
3575 if (list_empty(&tmp))
3578 spin_lock(&hctx->lock);
3579 list_splice_tail_init(&tmp, &hctx->dispatch);
3580 spin_unlock(&hctx->lock);
3582 blk_mq_run_hw_queue(hctx, true);
3586 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3588 if (!(hctx->flags & BLK_MQ_F_STACKING))
3589 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3590 &hctx->cpuhp_online);
3591 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3596 * Before freeing hw queue, clearing the flush request reference in
3597 * tags->rqs[] for avoiding potential UAF.
3599 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3600 unsigned int queue_depth, struct request *flush_rq)
3603 unsigned long flags;
3605 /* The hw queue may not be mapped yet */
3609 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3611 for (i = 0; i < queue_depth; i++)
3612 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3615 * Wait until all pending iteration is done.
3617 * Request reference is cleared and it is guaranteed to be observed
3618 * after the ->lock is released.
3620 spin_lock_irqsave(&tags->lock, flags);
3621 spin_unlock_irqrestore(&tags->lock, flags);
3624 /* hctx->ctxs will be freed in queue's release handler */
3625 static void blk_mq_exit_hctx(struct request_queue *q,
3626 struct blk_mq_tag_set *set,
3627 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3629 struct request *flush_rq = hctx->fq->flush_rq;
3631 if (blk_mq_hw_queue_mapped(hctx))
3632 blk_mq_tag_idle(hctx);
3634 if (blk_queue_init_done(q))
3635 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3636 set->queue_depth, flush_rq);
3637 if (set->ops->exit_request)
3638 set->ops->exit_request(set, flush_rq, hctx_idx);
3640 if (set->ops->exit_hctx)
3641 set->ops->exit_hctx(hctx, hctx_idx);
3643 blk_mq_remove_cpuhp(hctx);
3645 xa_erase(&q->hctx_table, hctx_idx);
3647 spin_lock(&q->unused_hctx_lock);
3648 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3649 spin_unlock(&q->unused_hctx_lock);
3652 static void blk_mq_exit_hw_queues(struct request_queue *q,
3653 struct blk_mq_tag_set *set, int nr_queue)
3655 struct blk_mq_hw_ctx *hctx;
3658 queue_for_each_hw_ctx(q, hctx, i) {
3661 blk_mq_exit_hctx(q, set, hctx, i);
3665 static int blk_mq_init_hctx(struct request_queue *q,
3666 struct blk_mq_tag_set *set,
3667 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3669 hctx->queue_num = hctx_idx;
3671 if (!(hctx->flags & BLK_MQ_F_STACKING))
3672 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3673 &hctx->cpuhp_online);
3674 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3676 hctx->tags = set->tags[hctx_idx];
3678 if (set->ops->init_hctx &&
3679 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3680 goto unregister_cpu_notifier;
3682 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3686 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3692 if (set->ops->exit_request)
3693 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3695 if (set->ops->exit_hctx)
3696 set->ops->exit_hctx(hctx, hctx_idx);
3697 unregister_cpu_notifier:
3698 blk_mq_remove_cpuhp(hctx);
3702 static struct blk_mq_hw_ctx *
3703 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3706 struct blk_mq_hw_ctx *hctx;
3707 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3709 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3711 goto fail_alloc_hctx;
3713 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3716 atomic_set(&hctx->nr_active, 0);
3717 if (node == NUMA_NO_NODE)
3718 node = set->numa_node;
3719 hctx->numa_node = node;
3721 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3722 spin_lock_init(&hctx->lock);
3723 INIT_LIST_HEAD(&hctx->dispatch);
3725 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3727 INIT_LIST_HEAD(&hctx->hctx_list);
3730 * Allocate space for all possible cpus to avoid allocation at
3733 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3738 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3739 gfp, node, false, false))
3743 spin_lock_init(&hctx->dispatch_wait_lock);
3744 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3745 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3747 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3751 blk_mq_hctx_kobj_init(hctx);
3756 sbitmap_free(&hctx->ctx_map);
3760 free_cpumask_var(hctx->cpumask);
3767 static void blk_mq_init_cpu_queues(struct request_queue *q,
3768 unsigned int nr_hw_queues)
3770 struct blk_mq_tag_set *set = q->tag_set;
3773 for_each_possible_cpu(i) {
3774 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3775 struct blk_mq_hw_ctx *hctx;
3779 spin_lock_init(&__ctx->lock);
3780 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3781 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3786 * Set local node, IFF we have more than one hw queue. If
3787 * not, we remain on the home node of the device
3789 for (j = 0; j < set->nr_maps; j++) {
3790 hctx = blk_mq_map_queue_type(q, j, i);
3791 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3792 hctx->numa_node = cpu_to_node(i);
3797 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3798 unsigned int hctx_idx,
3801 struct blk_mq_tags *tags;
3804 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3808 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3810 blk_mq_free_rq_map(tags);
3817 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3820 if (blk_mq_is_shared_tags(set->flags)) {
3821 set->tags[hctx_idx] = set->shared_tags;
3826 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3829 return set->tags[hctx_idx];
3832 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3833 struct blk_mq_tags *tags,
3834 unsigned int hctx_idx)
3837 blk_mq_free_rqs(set, tags, hctx_idx);
3838 blk_mq_free_rq_map(tags);
3842 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3843 unsigned int hctx_idx)
3845 if (!blk_mq_is_shared_tags(set->flags))
3846 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3848 set->tags[hctx_idx] = NULL;
3851 static void blk_mq_map_swqueue(struct request_queue *q)
3853 unsigned int j, hctx_idx;
3855 struct blk_mq_hw_ctx *hctx;
3856 struct blk_mq_ctx *ctx;
3857 struct blk_mq_tag_set *set = q->tag_set;
3859 queue_for_each_hw_ctx(q, hctx, i) {
3860 cpumask_clear(hctx->cpumask);
3862 hctx->dispatch_from = NULL;
3866 * Map software to hardware queues.
3868 * If the cpu isn't present, the cpu is mapped to first hctx.
3870 for_each_possible_cpu(i) {
3872 ctx = per_cpu_ptr(q->queue_ctx, i);
3873 for (j = 0; j < set->nr_maps; j++) {
3874 if (!set->map[j].nr_queues) {
3875 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3876 HCTX_TYPE_DEFAULT, i);
3879 hctx_idx = set->map[j].mq_map[i];
3880 /* unmapped hw queue can be remapped after CPU topo changed */
3881 if (!set->tags[hctx_idx] &&
3882 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3884 * If tags initialization fail for some hctx,
3885 * that hctx won't be brought online. In this
3886 * case, remap the current ctx to hctx[0] which
3887 * is guaranteed to always have tags allocated
3889 set->map[j].mq_map[i] = 0;
3892 hctx = blk_mq_map_queue_type(q, j, i);
3893 ctx->hctxs[j] = hctx;
3895 * If the CPU is already set in the mask, then we've
3896 * mapped this one already. This can happen if
3897 * devices share queues across queue maps.
3899 if (cpumask_test_cpu(i, hctx->cpumask))
3902 cpumask_set_cpu(i, hctx->cpumask);
3904 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3905 hctx->ctxs[hctx->nr_ctx++] = ctx;
3908 * If the nr_ctx type overflows, we have exceeded the
3909 * amount of sw queues we can support.
3911 BUG_ON(!hctx->nr_ctx);
3914 for (; j < HCTX_MAX_TYPES; j++)
3915 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3916 HCTX_TYPE_DEFAULT, i);
3919 queue_for_each_hw_ctx(q, hctx, i) {
3921 * If no software queues are mapped to this hardware queue,
3922 * disable it and free the request entries.
3924 if (!hctx->nr_ctx) {
3925 /* Never unmap queue 0. We need it as a
3926 * fallback in case of a new remap fails
3930 __blk_mq_free_map_and_rqs(set, i);
3936 hctx->tags = set->tags[i];
3937 WARN_ON(!hctx->tags);
3940 * Set the map size to the number of mapped software queues.
3941 * This is more accurate and more efficient than looping
3942 * over all possibly mapped software queues.
3944 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3947 * Initialize batch roundrobin counts
3949 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3950 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3955 * Caller needs to ensure that we're either frozen/quiesced, or that
3956 * the queue isn't live yet.
3958 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3960 struct blk_mq_hw_ctx *hctx;
3963 queue_for_each_hw_ctx(q, hctx, i) {
3965 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3967 blk_mq_tag_idle(hctx);
3968 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3973 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3976 struct request_queue *q;
3978 lockdep_assert_held(&set->tag_list_lock);
3980 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3981 blk_mq_freeze_queue(q);
3982 queue_set_hctx_shared(q, shared);
3983 blk_mq_unfreeze_queue(q);
3987 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3989 struct blk_mq_tag_set *set = q->tag_set;
3991 mutex_lock(&set->tag_list_lock);
3992 list_del(&q->tag_set_list);
3993 if (list_is_singular(&set->tag_list)) {
3994 /* just transitioned to unshared */
3995 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3996 /* update existing queue */
3997 blk_mq_update_tag_set_shared(set, false);
3999 mutex_unlock(&set->tag_list_lock);
4000 INIT_LIST_HEAD(&q->tag_set_list);
4003 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
4004 struct request_queue *q)
4006 mutex_lock(&set->tag_list_lock);
4009 * Check to see if we're transitioning to shared (from 1 to 2 queues).
4011 if (!list_empty(&set->tag_list) &&
4012 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
4013 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
4014 /* update existing queue */
4015 blk_mq_update_tag_set_shared(set, true);
4017 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
4018 queue_set_hctx_shared(q, true);
4019 list_add_tail(&q->tag_set_list, &set->tag_list);
4021 mutex_unlock(&set->tag_list_lock);
4024 /* All allocations will be freed in release handler of q->mq_kobj */
4025 static int blk_mq_alloc_ctxs(struct request_queue *q)
4027 struct blk_mq_ctxs *ctxs;
4030 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
4034 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
4035 if (!ctxs->queue_ctx)
4038 for_each_possible_cpu(cpu) {
4039 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4043 q->mq_kobj = &ctxs->kobj;
4044 q->queue_ctx = ctxs->queue_ctx;
4053 * It is the actual release handler for mq, but we do it from
4054 * request queue's release handler for avoiding use-after-free
4055 * and headache because q->mq_kobj shouldn't have been introduced,
4056 * but we can't group ctx/kctx kobj without it.
4058 void blk_mq_release(struct request_queue *q)
4060 struct blk_mq_hw_ctx *hctx, *next;
4063 queue_for_each_hw_ctx(q, hctx, i)
4064 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4066 /* all hctx are in .unused_hctx_list now */
4067 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4068 list_del_init(&hctx->hctx_list);
4069 kobject_put(&hctx->kobj);
4072 xa_destroy(&q->hctx_table);
4075 * release .mq_kobj and sw queue's kobject now because
4076 * both share lifetime with request queue.
4078 blk_mq_sysfs_deinit(q);
4081 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4084 struct request_queue *q;
4087 q = blk_alloc_queue(set->numa_node);
4089 return ERR_PTR(-ENOMEM);
4090 q->queuedata = queuedata;
4091 ret = blk_mq_init_allocated_queue(set, q);
4094 return ERR_PTR(ret);
4099 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4101 return blk_mq_init_queue_data(set, NULL);
4103 EXPORT_SYMBOL(blk_mq_init_queue);
4106 * blk_mq_destroy_queue - shutdown a request queue
4107 * @q: request queue to shutdown
4109 * This shuts down a request queue allocated by blk_mq_init_queue(). All future
4110 * requests will be failed with -ENODEV. The caller is responsible for dropping
4111 * the reference from blk_mq_init_queue() by calling blk_put_queue().
4113 * Context: can sleep
4115 void blk_mq_destroy_queue(struct request_queue *q)
4117 WARN_ON_ONCE(!queue_is_mq(q));
4118 WARN_ON_ONCE(blk_queue_registered(q));
4122 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4123 blk_queue_start_drain(q);
4124 blk_mq_freeze_queue_wait(q);
4127 blk_mq_cancel_work_sync(q);
4128 blk_mq_exit_queue(q);
4130 EXPORT_SYMBOL(blk_mq_destroy_queue);
4132 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4133 struct lock_class_key *lkclass)
4135 struct request_queue *q;
4136 struct gendisk *disk;
4138 q = blk_mq_init_queue_data(set, queuedata);
4142 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4144 blk_mq_destroy_queue(q);
4146 return ERR_PTR(-ENOMEM);
4148 set_bit(GD_OWNS_QUEUE, &disk->state);
4151 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4153 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4154 struct lock_class_key *lkclass)
4156 struct gendisk *disk;
4158 if (!blk_get_queue(q))
4160 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4165 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4167 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4168 struct blk_mq_tag_set *set, struct request_queue *q,
4169 int hctx_idx, int node)
4171 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4173 /* reuse dead hctx first */
4174 spin_lock(&q->unused_hctx_lock);
4175 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4176 if (tmp->numa_node == node) {
4182 list_del_init(&hctx->hctx_list);
4183 spin_unlock(&q->unused_hctx_lock);
4186 hctx = blk_mq_alloc_hctx(q, set, node);
4190 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4196 kobject_put(&hctx->kobj);
4201 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4202 struct request_queue *q)
4204 struct blk_mq_hw_ctx *hctx;
4207 /* protect against switching io scheduler */
4208 mutex_lock(&q->sysfs_lock);
4209 for (i = 0; i < set->nr_hw_queues; i++) {
4211 int node = blk_mq_get_hctx_node(set, i);
4212 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4215 old_node = old_hctx->numa_node;
4216 blk_mq_exit_hctx(q, set, old_hctx, i);
4219 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4222 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4224 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4225 WARN_ON_ONCE(!hctx);
4229 * Increasing nr_hw_queues fails. Free the newly allocated
4230 * hctxs and keep the previous q->nr_hw_queues.
4232 if (i != set->nr_hw_queues) {
4233 j = q->nr_hw_queues;
4236 q->nr_hw_queues = set->nr_hw_queues;
4239 xa_for_each_start(&q->hctx_table, j, hctx, j)
4240 blk_mq_exit_hctx(q, set, hctx, j);
4241 mutex_unlock(&q->sysfs_lock);
4244 static void blk_mq_update_poll_flag(struct request_queue *q)
4246 struct blk_mq_tag_set *set = q->tag_set;
4248 if (set->nr_maps > HCTX_TYPE_POLL &&
4249 set->map[HCTX_TYPE_POLL].nr_queues)
4250 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4252 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4255 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4256 struct request_queue *q)
4258 /* mark the queue as mq asap */
4259 q->mq_ops = set->ops;
4261 if (blk_mq_alloc_ctxs(q))
4264 /* init q->mq_kobj and sw queues' kobjects */
4265 blk_mq_sysfs_init(q);
4267 INIT_LIST_HEAD(&q->unused_hctx_list);
4268 spin_lock_init(&q->unused_hctx_lock);
4270 xa_init(&q->hctx_table);
4272 blk_mq_realloc_hw_ctxs(set, q);
4273 if (!q->nr_hw_queues)
4276 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4277 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4281 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4282 blk_mq_update_poll_flag(q);
4284 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4285 INIT_LIST_HEAD(&q->flush_list);
4286 INIT_LIST_HEAD(&q->requeue_list);
4287 spin_lock_init(&q->requeue_lock);
4289 q->nr_requests = set->queue_depth;
4291 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4292 blk_mq_add_queue_tag_set(set, q);
4293 blk_mq_map_swqueue(q);
4302 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4304 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4305 void blk_mq_exit_queue(struct request_queue *q)
4307 struct blk_mq_tag_set *set = q->tag_set;
4309 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4310 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4311 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4312 blk_mq_del_queue_tag_set(q);
4315 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4319 if (blk_mq_is_shared_tags(set->flags)) {
4320 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4323 if (!set->shared_tags)
4327 for (i = 0; i < set->nr_hw_queues; i++) {
4328 if (!__blk_mq_alloc_map_and_rqs(set, i))
4337 __blk_mq_free_map_and_rqs(set, i);
4339 if (blk_mq_is_shared_tags(set->flags)) {
4340 blk_mq_free_map_and_rqs(set, set->shared_tags,
4341 BLK_MQ_NO_HCTX_IDX);
4348 * Allocate the request maps associated with this tag_set. Note that this
4349 * may reduce the depth asked for, if memory is tight. set->queue_depth
4350 * will be updated to reflect the allocated depth.
4352 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4357 depth = set->queue_depth;
4359 err = __blk_mq_alloc_rq_maps(set);
4363 set->queue_depth >>= 1;
4364 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4368 } while (set->queue_depth);
4370 if (!set->queue_depth || err) {
4371 pr_err("blk-mq: failed to allocate request map\n");
4375 if (depth != set->queue_depth)
4376 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4377 depth, set->queue_depth);
4382 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4385 * blk_mq_map_queues() and multiple .map_queues() implementations
4386 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4387 * number of hardware queues.
4389 if (set->nr_maps == 1)
4390 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4392 if (set->ops->map_queues && !is_kdump_kernel()) {
4396 * transport .map_queues is usually done in the following
4399 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4400 * mask = get_cpu_mask(queue)
4401 * for_each_cpu(cpu, mask)
4402 * set->map[x].mq_map[cpu] = queue;
4405 * When we need to remap, the table has to be cleared for
4406 * killing stale mapping since one CPU may not be mapped
4409 for (i = 0; i < set->nr_maps; i++)
4410 blk_mq_clear_mq_map(&set->map[i]);
4412 set->ops->map_queues(set);
4414 BUG_ON(set->nr_maps > 1);
4415 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4419 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4420 int new_nr_hw_queues)
4422 struct blk_mq_tags **new_tags;
4425 if (set->nr_hw_queues >= new_nr_hw_queues)
4428 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4429 GFP_KERNEL, set->numa_node);
4434 memcpy(new_tags, set->tags, set->nr_hw_queues *
4435 sizeof(*set->tags));
4437 set->tags = new_tags;
4439 for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) {
4440 if (!__blk_mq_alloc_map_and_rqs(set, i)) {
4441 while (--i >= set->nr_hw_queues)
4442 __blk_mq_free_map_and_rqs(set, i);
4449 set->nr_hw_queues = new_nr_hw_queues;
4454 * Alloc a tag set to be associated with one or more request queues.
4455 * May fail with EINVAL for various error conditions. May adjust the
4456 * requested depth down, if it's too large. In that case, the set
4457 * value will be stored in set->queue_depth.
4459 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4463 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4465 if (!set->nr_hw_queues)
4467 if (!set->queue_depth)
4469 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4472 if (!set->ops->queue_rq)
4475 if (!set->ops->get_budget ^ !set->ops->put_budget)
4478 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4479 pr_info("blk-mq: reduced tag depth to %u\n",
4481 set->queue_depth = BLK_MQ_MAX_DEPTH;
4486 else if (set->nr_maps > HCTX_MAX_TYPES)
4490 * If a crashdump is active, then we are potentially in a very
4491 * memory constrained environment. Limit us to 1 queue and
4492 * 64 tags to prevent using too much memory.
4494 if (is_kdump_kernel()) {
4495 set->nr_hw_queues = 1;
4497 set->queue_depth = min(64U, set->queue_depth);
4500 * There is no use for more h/w queues than cpus if we just have
4503 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4504 set->nr_hw_queues = nr_cpu_ids;
4506 if (set->flags & BLK_MQ_F_BLOCKING) {
4507 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4510 ret = init_srcu_struct(set->srcu);
4516 set->tags = kcalloc_node(set->nr_hw_queues,
4517 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4520 goto out_cleanup_srcu;
4522 for (i = 0; i < set->nr_maps; i++) {
4523 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4524 sizeof(set->map[i].mq_map[0]),
4525 GFP_KERNEL, set->numa_node);
4526 if (!set->map[i].mq_map)
4527 goto out_free_mq_map;
4528 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4531 blk_mq_update_queue_map(set);
4533 ret = blk_mq_alloc_set_map_and_rqs(set);
4535 goto out_free_mq_map;
4537 mutex_init(&set->tag_list_lock);
4538 INIT_LIST_HEAD(&set->tag_list);
4543 for (i = 0; i < set->nr_maps; i++) {
4544 kfree(set->map[i].mq_map);
4545 set->map[i].mq_map = NULL;
4550 if (set->flags & BLK_MQ_F_BLOCKING)
4551 cleanup_srcu_struct(set->srcu);
4553 if (set->flags & BLK_MQ_F_BLOCKING)
4557 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4559 /* allocate and initialize a tagset for a simple single-queue device */
4560 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4561 const struct blk_mq_ops *ops, unsigned int queue_depth,
4562 unsigned int set_flags)
4564 memset(set, 0, sizeof(*set));
4566 set->nr_hw_queues = 1;
4568 set->queue_depth = queue_depth;
4569 set->numa_node = NUMA_NO_NODE;
4570 set->flags = set_flags;
4571 return blk_mq_alloc_tag_set(set);
4573 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4575 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4579 for (i = 0; i < set->nr_hw_queues; i++)
4580 __blk_mq_free_map_and_rqs(set, i);
4582 if (blk_mq_is_shared_tags(set->flags)) {
4583 blk_mq_free_map_and_rqs(set, set->shared_tags,
4584 BLK_MQ_NO_HCTX_IDX);
4587 for (j = 0; j < set->nr_maps; j++) {
4588 kfree(set->map[j].mq_map);
4589 set->map[j].mq_map = NULL;
4594 if (set->flags & BLK_MQ_F_BLOCKING) {
4595 cleanup_srcu_struct(set->srcu);
4599 EXPORT_SYMBOL(blk_mq_free_tag_set);
4601 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4603 struct blk_mq_tag_set *set = q->tag_set;
4604 struct blk_mq_hw_ctx *hctx;
4611 if (q->nr_requests == nr)
4614 blk_mq_freeze_queue(q);
4615 blk_mq_quiesce_queue(q);
4618 queue_for_each_hw_ctx(q, hctx, i) {
4622 * If we're using an MQ scheduler, just update the scheduler
4623 * queue depth. This is similar to what the old code would do.
4625 if (hctx->sched_tags) {
4626 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4629 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4634 if (q->elevator && q->elevator->type->ops.depth_updated)
4635 q->elevator->type->ops.depth_updated(hctx);
4638 q->nr_requests = nr;
4639 if (blk_mq_is_shared_tags(set->flags)) {
4641 blk_mq_tag_update_sched_shared_tags(q);
4643 blk_mq_tag_resize_shared_tags(set, nr);
4647 blk_mq_unquiesce_queue(q);
4648 blk_mq_unfreeze_queue(q);
4654 * request_queue and elevator_type pair.
4655 * It is just used by __blk_mq_update_nr_hw_queues to cache
4656 * the elevator_type associated with a request_queue.
4658 struct blk_mq_qe_pair {
4659 struct list_head node;
4660 struct request_queue *q;
4661 struct elevator_type *type;
4665 * Cache the elevator_type in qe pair list and switch the
4666 * io scheduler to 'none'
4668 static bool blk_mq_elv_switch_none(struct list_head *head,
4669 struct request_queue *q)
4671 struct blk_mq_qe_pair *qe;
4673 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4677 /* q->elevator needs protection from ->sysfs_lock */
4678 mutex_lock(&q->sysfs_lock);
4680 /* the check has to be done with holding sysfs_lock */
4686 INIT_LIST_HEAD(&qe->node);
4688 qe->type = q->elevator->type;
4689 /* keep a reference to the elevator module as we'll switch back */
4690 __elevator_get(qe->type);
4691 list_add(&qe->node, head);
4692 elevator_disable(q);
4694 mutex_unlock(&q->sysfs_lock);
4699 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4700 struct request_queue *q)
4702 struct blk_mq_qe_pair *qe;
4704 list_for_each_entry(qe, head, node)
4711 static void blk_mq_elv_switch_back(struct list_head *head,
4712 struct request_queue *q)
4714 struct blk_mq_qe_pair *qe;
4715 struct elevator_type *t;
4717 qe = blk_lookup_qe_pair(head, q);
4721 list_del(&qe->node);
4724 mutex_lock(&q->sysfs_lock);
4725 elevator_switch(q, t);
4726 /* drop the reference acquired in blk_mq_elv_switch_none */
4728 mutex_unlock(&q->sysfs_lock);
4731 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4734 struct request_queue *q;
4736 int prev_nr_hw_queues = set->nr_hw_queues;
4739 lockdep_assert_held(&set->tag_list_lock);
4741 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4742 nr_hw_queues = nr_cpu_ids;
4743 if (nr_hw_queues < 1)
4745 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4748 list_for_each_entry(q, &set->tag_list, tag_set_list)
4749 blk_mq_freeze_queue(q);
4751 * Switch IO scheduler to 'none', cleaning up the data associated
4752 * with the previous scheduler. We will switch back once we are done
4753 * updating the new sw to hw queue mappings.
4755 list_for_each_entry(q, &set->tag_list, tag_set_list)
4756 if (!blk_mq_elv_switch_none(&head, q))
4759 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4760 blk_mq_debugfs_unregister_hctxs(q);
4761 blk_mq_sysfs_unregister_hctxs(q);
4764 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4768 blk_mq_update_queue_map(set);
4769 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4770 blk_mq_realloc_hw_ctxs(set, q);
4771 blk_mq_update_poll_flag(q);
4772 if (q->nr_hw_queues != set->nr_hw_queues) {
4773 int i = prev_nr_hw_queues;
4775 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4776 nr_hw_queues, prev_nr_hw_queues);
4777 for (; i < set->nr_hw_queues; i++)
4778 __blk_mq_free_map_and_rqs(set, i);
4780 set->nr_hw_queues = prev_nr_hw_queues;
4783 blk_mq_map_swqueue(q);
4787 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4788 blk_mq_sysfs_register_hctxs(q);
4789 blk_mq_debugfs_register_hctxs(q);
4793 list_for_each_entry(q, &set->tag_list, tag_set_list)
4794 blk_mq_elv_switch_back(&head, q);
4796 list_for_each_entry(q, &set->tag_list, tag_set_list)
4797 blk_mq_unfreeze_queue(q);
4799 /* Free the excess tags when nr_hw_queues shrink. */
4800 for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)
4801 __blk_mq_free_map_and_rqs(set, i);
4804 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4806 mutex_lock(&set->tag_list_lock);
4807 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4808 mutex_unlock(&set->tag_list_lock);
4810 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4812 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
4813 struct io_comp_batch *iob, unsigned int flags)
4815 long state = get_current_state();
4819 ret = q->mq_ops->poll(hctx, iob);
4821 __set_current_state(TASK_RUNNING);
4825 if (signal_pending_state(state, current))
4826 __set_current_state(TASK_RUNNING);
4827 if (task_is_running(current))
4830 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4833 } while (!need_resched());
4835 __set_current_state(TASK_RUNNING);
4839 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
4840 struct io_comp_batch *iob, unsigned int flags)
4842 struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie);
4844 return blk_hctx_poll(q, hctx, iob, flags);
4847 int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
4848 unsigned int poll_flags)
4850 struct request_queue *q = rq->q;
4853 if (!blk_rq_is_poll(rq))
4855 if (!percpu_ref_tryget(&q->q_usage_counter))
4858 ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags);
4863 EXPORT_SYMBOL_GPL(blk_rq_poll);
4865 unsigned int blk_mq_rq_cpu(struct request *rq)
4867 return rq->mq_ctx->cpu;
4869 EXPORT_SYMBOL(blk_mq_rq_cpu);
4871 void blk_mq_cancel_work_sync(struct request_queue *q)
4873 struct blk_mq_hw_ctx *hctx;
4876 cancel_delayed_work_sync(&q->requeue_work);
4878 queue_for_each_hw_ctx(q, hctx, i)
4879 cancel_delayed_work_sync(&hctx->run_work);
4882 static int __init blk_mq_init(void)
4886 for_each_possible_cpu(i)
4887 init_llist_head(&per_cpu(blk_cpu_done, i));
4888 for_each_possible_cpu(i)
4889 INIT_CSD(&per_cpu(blk_cpu_csd, i),
4890 __blk_mq_complete_request_remote, NULL);
4891 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4893 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4894 "block/softirq:dead", NULL,
4895 blk_softirq_cpu_dead);
4896 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4897 blk_mq_hctx_notify_dead);
4898 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4899 blk_mq_hctx_notify_online,
4900 blk_mq_hctx_notify_offline);
4903 subsys_initcall(blk_mq_init);