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 if (!(data->rq_flags & RQF_SCHED_TAGS))
430 blk_mq_add_active_requests(data->hctx, nr);
431 /* caller already holds a reference, add for remainder */
432 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
435 return rq_list_pop(data->cached_rq);
438 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
440 struct request_queue *q = data->q;
441 u64 alloc_time_ns = 0;
445 /* alloc_time includes depth and tag waits */
446 if (blk_queue_rq_alloc_time(q))
447 alloc_time_ns = ktime_get_ns();
449 if (data->cmd_flags & REQ_NOWAIT)
450 data->flags |= BLK_MQ_REQ_NOWAIT;
454 * All requests use scheduler tags when an I/O scheduler is
455 * enabled for the queue.
457 data->rq_flags |= RQF_SCHED_TAGS;
460 * Flush/passthrough requests are special and go directly to the
463 if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&
464 !blk_op_is_passthrough(data->cmd_flags)) {
465 struct elevator_mq_ops *ops = &q->elevator->type->ops;
467 WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
469 data->rq_flags |= RQF_USE_SCHED;
470 if (ops->limit_depth)
471 ops->limit_depth(data->cmd_flags, data);
476 data->ctx = blk_mq_get_ctx(q);
477 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
478 if (!(data->rq_flags & RQF_SCHED_TAGS))
479 blk_mq_tag_busy(data->hctx);
481 if (data->flags & BLK_MQ_REQ_RESERVED)
482 data->rq_flags |= RQF_RESV;
485 * Try batched alloc if we want more than 1 tag.
487 if (data->nr_tags > 1) {
488 rq = __blk_mq_alloc_requests_batch(data);
490 blk_mq_rq_time_init(rq, alloc_time_ns);
497 * Waiting allocations only fail because of an inactive hctx. In that
498 * case just retry the hctx assignment and tag allocation as CPU hotplug
499 * should have migrated us to an online CPU by now.
501 tag = blk_mq_get_tag(data);
502 if (tag == BLK_MQ_NO_TAG) {
503 if (data->flags & BLK_MQ_REQ_NOWAIT)
506 * Give up the CPU and sleep for a random short time to
507 * ensure that thread using a realtime scheduling class
508 * are migrated off the CPU, and thus off the hctx that
515 if (!(data->rq_flags & RQF_SCHED_TAGS))
516 blk_mq_inc_active_requests(data->hctx);
517 rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag);
518 blk_mq_rq_time_init(rq, alloc_time_ns);
522 static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
523 struct blk_plug *plug,
525 blk_mq_req_flags_t flags)
527 struct blk_mq_alloc_data data = {
531 .nr_tags = plug->nr_ios,
532 .cached_rq = &plug->cached_rq,
536 if (blk_queue_enter(q, flags))
541 rq = __blk_mq_alloc_requests(&data);
547 static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
549 blk_mq_req_flags_t flags)
551 struct blk_plug *plug = current->plug;
557 if (rq_list_empty(plug->cached_rq)) {
558 if (plug->nr_ios == 1)
560 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
564 rq = rq_list_peek(&plug->cached_rq);
565 if (!rq || rq->q != q)
568 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
570 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
573 plug->cached_rq = rq_list_next(rq);
574 blk_mq_rq_time_init(rq, 0);
578 INIT_LIST_HEAD(&rq->queuelist);
582 struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
583 blk_mq_req_flags_t flags)
587 rq = blk_mq_alloc_cached_request(q, opf, flags);
589 struct blk_mq_alloc_data data = {
597 ret = blk_queue_enter(q, flags);
601 rq = __blk_mq_alloc_requests(&data);
606 rq->__sector = (sector_t) -1;
607 rq->bio = rq->biotail = NULL;
611 return ERR_PTR(-EWOULDBLOCK);
613 EXPORT_SYMBOL(blk_mq_alloc_request);
615 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
616 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
618 struct blk_mq_alloc_data data = {
624 u64 alloc_time_ns = 0;
630 /* alloc_time includes depth and tag waits */
631 if (blk_queue_rq_alloc_time(q))
632 alloc_time_ns = ktime_get_ns();
635 * If the tag allocator sleeps we could get an allocation for a
636 * different hardware context. No need to complicate the low level
637 * allocator for this for the rare use case of a command tied to
640 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
641 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
642 return ERR_PTR(-EINVAL);
644 if (hctx_idx >= q->nr_hw_queues)
645 return ERR_PTR(-EIO);
647 ret = blk_queue_enter(q, flags);
652 * Check if the hardware context is actually mapped to anything.
653 * If not tell the caller that it should skip this queue.
656 data.hctx = xa_load(&q->hctx_table, hctx_idx);
657 if (!blk_mq_hw_queue_mapped(data.hctx))
659 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
660 if (cpu >= nr_cpu_ids)
662 data.ctx = __blk_mq_get_ctx(q, cpu);
665 data.rq_flags |= RQF_SCHED_TAGS;
667 blk_mq_tag_busy(data.hctx);
669 if (flags & BLK_MQ_REQ_RESERVED)
670 data.rq_flags |= RQF_RESV;
673 tag = blk_mq_get_tag(&data);
674 if (tag == BLK_MQ_NO_TAG)
676 if (!(data.rq_flags & RQF_SCHED_TAGS))
677 blk_mq_inc_active_requests(data.hctx);
678 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag);
679 blk_mq_rq_time_init(rq, alloc_time_ns);
681 rq->__sector = (sector_t) -1;
682 rq->bio = rq->biotail = NULL;
689 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
691 static void blk_mq_finish_request(struct request *rq)
693 struct request_queue *q = rq->q;
695 if (rq->rq_flags & RQF_USE_SCHED) {
696 q->elevator->type->ops.finish_request(rq);
698 * For postflush request that may need to be
699 * completed twice, we should clear this flag
700 * to avoid double finish_request() on the rq.
702 rq->rq_flags &= ~RQF_USE_SCHED;
706 static void __blk_mq_free_request(struct request *rq)
708 struct request_queue *q = rq->q;
709 struct blk_mq_ctx *ctx = rq->mq_ctx;
710 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
711 const int sched_tag = rq->internal_tag;
713 blk_crypto_free_request(rq);
714 blk_pm_mark_last_busy(rq);
717 if (rq->tag != BLK_MQ_NO_TAG) {
718 blk_mq_dec_active_requests(hctx);
719 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
721 if (sched_tag != BLK_MQ_NO_TAG)
722 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
723 blk_mq_sched_restart(hctx);
727 void blk_mq_free_request(struct request *rq)
729 struct request_queue *q = rq->q;
731 blk_mq_finish_request(rq);
733 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
734 laptop_io_completion(q->disk->bdi);
738 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
739 if (req_ref_put_and_test(rq))
740 __blk_mq_free_request(rq);
742 EXPORT_SYMBOL_GPL(blk_mq_free_request);
744 void blk_mq_free_plug_rqs(struct blk_plug *plug)
748 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
749 blk_mq_free_request(rq);
752 void blk_dump_rq_flags(struct request *rq, char *msg)
754 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
755 rq->q->disk ? rq->q->disk->disk_name : "?",
756 (__force unsigned long long) rq->cmd_flags);
758 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
759 (unsigned long long)blk_rq_pos(rq),
760 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
761 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
762 rq->bio, rq->biotail, blk_rq_bytes(rq));
764 EXPORT_SYMBOL(blk_dump_rq_flags);
766 static void req_bio_endio(struct request *rq, struct bio *bio,
767 unsigned int nbytes, blk_status_t error)
769 if (unlikely(error)) {
770 bio->bi_status = error;
771 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
773 * Partial zone append completions cannot be supported as the
774 * BIO fragments may end up not being written sequentially.
776 if (bio->bi_iter.bi_size != nbytes)
777 bio->bi_status = BLK_STS_IOERR;
779 bio->bi_iter.bi_sector = rq->__sector;
782 bio_advance(bio, nbytes);
784 if (unlikely(rq->rq_flags & RQF_QUIET))
785 bio_set_flag(bio, BIO_QUIET);
786 /* don't actually finish bio if it's part of flush sequence */
787 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
791 static void blk_account_io_completion(struct request *req, unsigned int bytes)
793 if (req->part && blk_do_io_stat(req)) {
794 const int sgrp = op_stat_group(req_op(req));
797 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
802 static void blk_print_req_error(struct request *req, blk_status_t status)
804 printk_ratelimited(KERN_ERR
805 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
806 "phys_seg %u prio class %u\n",
807 blk_status_to_str(status),
808 req->q->disk ? req->q->disk->disk_name : "?",
809 blk_rq_pos(req), (__force u32)req_op(req),
810 blk_op_str(req_op(req)),
811 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
812 req->nr_phys_segments,
813 IOPRIO_PRIO_CLASS(req->ioprio));
817 * Fully end IO on a request. Does not support partial completions, or
820 static void blk_complete_request(struct request *req)
822 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
823 int total_bytes = blk_rq_bytes(req);
824 struct bio *bio = req->bio;
826 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
831 #ifdef CONFIG_BLK_DEV_INTEGRITY
832 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
833 req->q->integrity.profile->complete_fn(req, total_bytes);
837 * Upper layers may call blk_crypto_evict_key() anytime after the last
838 * bio_endio(). Therefore, the keyslot must be released before that.
840 blk_crypto_rq_put_keyslot(req);
842 blk_account_io_completion(req, total_bytes);
845 struct bio *next = bio->bi_next;
847 /* Completion has already been traced */
848 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
850 if (req_op(req) == REQ_OP_ZONE_APPEND)
851 bio->bi_iter.bi_sector = req->__sector;
859 * Reset counters so that the request stacking driver
860 * can find how many bytes remain in the request
870 * blk_update_request - Complete multiple bytes without completing the request
871 * @req: the request being processed
872 * @error: block status code
873 * @nr_bytes: number of bytes to complete for @req
876 * Ends I/O on a number of bytes attached to @req, but doesn't complete
877 * the request structure even if @req doesn't have leftover.
878 * If @req has leftover, sets it up for the next range of segments.
880 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
881 * %false return from this function.
884 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
885 * except in the consistency check at the end of this function.
888 * %false - this request doesn't have any more data
889 * %true - this request has more data
891 bool blk_update_request(struct request *req, blk_status_t error,
892 unsigned int nr_bytes)
896 trace_block_rq_complete(req, error, nr_bytes);
901 #ifdef CONFIG_BLK_DEV_INTEGRITY
902 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
904 req->q->integrity.profile->complete_fn(req, nr_bytes);
908 * Upper layers may call blk_crypto_evict_key() anytime after the last
909 * bio_endio(). Therefore, the keyslot must be released before that.
911 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
912 __blk_crypto_rq_put_keyslot(req);
914 if (unlikely(error && !blk_rq_is_passthrough(req) &&
915 !(req->rq_flags & RQF_QUIET)) &&
916 !test_bit(GD_DEAD, &req->q->disk->state)) {
917 blk_print_req_error(req, error);
918 trace_block_rq_error(req, error, nr_bytes);
921 blk_account_io_completion(req, nr_bytes);
925 struct bio *bio = req->bio;
926 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
928 if (bio_bytes == bio->bi_iter.bi_size)
929 req->bio = bio->bi_next;
931 /* Completion has already been traced */
932 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
933 req_bio_endio(req, bio, bio_bytes, error);
935 total_bytes += bio_bytes;
936 nr_bytes -= bio_bytes;
947 * Reset counters so that the request stacking driver
948 * can find how many bytes remain in the request
955 req->__data_len -= total_bytes;
957 /* update sector only for requests with clear definition of sector */
958 if (!blk_rq_is_passthrough(req))
959 req->__sector += total_bytes >> 9;
961 /* mixed attributes always follow the first bio */
962 if (req->rq_flags & RQF_MIXED_MERGE) {
963 req->cmd_flags &= ~REQ_FAILFAST_MASK;
964 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
967 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
969 * If total number of sectors is less than the first segment
970 * size, something has gone terribly wrong.
972 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
973 blk_dump_rq_flags(req, "request botched");
974 req->__data_len = blk_rq_cur_bytes(req);
977 /* recalculate the number of segments */
978 req->nr_phys_segments = blk_recalc_rq_segments(req);
983 EXPORT_SYMBOL_GPL(blk_update_request);
985 static inline void blk_account_io_done(struct request *req, u64 now)
987 trace_block_io_done(req);
990 * Account IO completion. flush_rq isn't accounted as a
991 * normal IO on queueing nor completion. Accounting the
992 * containing request is enough.
994 if (blk_do_io_stat(req) && req->part &&
995 !(req->rq_flags & RQF_FLUSH_SEQ)) {
996 const int sgrp = op_stat_group(req_op(req));
999 update_io_ticks(req->part, jiffies, true);
1000 part_stat_inc(req->part, ios[sgrp]);
1001 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
1006 static inline void blk_account_io_start(struct request *req)
1008 trace_block_io_start(req);
1010 if (blk_do_io_stat(req)) {
1012 * All non-passthrough requests are created from a bio with one
1013 * exception: when a flush command that is part of a flush sequence
1014 * generated by the state machine in blk-flush.c is cloned onto the
1015 * lower device by dm-multipath we can get here without a bio.
1018 req->part = req->bio->bi_bdev;
1020 req->part = req->q->disk->part0;
1023 update_io_ticks(req->part, jiffies, false);
1028 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1030 if (rq->rq_flags & RQF_STATS)
1031 blk_stat_add(rq, now);
1033 blk_mq_sched_completed_request(rq, now);
1034 blk_account_io_done(rq, now);
1037 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1039 if (blk_mq_need_time_stamp(rq))
1040 __blk_mq_end_request_acct(rq, ktime_get_ns());
1042 blk_mq_finish_request(rq);
1045 rq_qos_done(rq->q, rq);
1046 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1047 blk_mq_free_request(rq);
1049 blk_mq_free_request(rq);
1052 EXPORT_SYMBOL(__blk_mq_end_request);
1054 void blk_mq_end_request(struct request *rq, blk_status_t error)
1056 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1058 __blk_mq_end_request(rq, error);
1060 EXPORT_SYMBOL(blk_mq_end_request);
1062 #define TAG_COMP_BATCH 32
1064 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1065 int *tag_array, int nr_tags)
1067 struct request_queue *q = hctx->queue;
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_mq_rq_inflight(struct request *rq, void *priv)
1517 * If we find a request that isn't idle we know the queue is busy
1518 * as it's checked in the iter.
1519 * Return false to stop the iteration.
1521 if (blk_mq_request_started(rq)) {
1531 bool blk_mq_queue_inflight(struct request_queue *q)
1535 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1538 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1540 static void blk_mq_rq_timed_out(struct request *req)
1542 req->rq_flags |= RQF_TIMED_OUT;
1543 if (req->q->mq_ops->timeout) {
1544 enum blk_eh_timer_return ret;
1546 ret = req->q->mq_ops->timeout(req);
1547 if (ret == BLK_EH_DONE)
1549 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1555 struct blk_expired_data {
1556 bool has_timedout_rq;
1558 unsigned long timeout_start;
1561 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1563 unsigned long deadline;
1565 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1567 if (rq->rq_flags & RQF_TIMED_OUT)
1570 deadline = READ_ONCE(rq->deadline);
1571 if (time_after_eq(expired->timeout_start, deadline))
1574 if (expired->next == 0)
1575 expired->next = deadline;
1576 else if (time_after(expired->next, deadline))
1577 expired->next = deadline;
1581 void blk_mq_put_rq_ref(struct request *rq)
1583 if (is_flush_rq(rq)) {
1584 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1585 blk_mq_free_request(rq);
1586 } else if (req_ref_put_and_test(rq)) {
1587 __blk_mq_free_request(rq);
1591 static bool blk_mq_check_expired(struct request *rq, void *priv)
1593 struct blk_expired_data *expired = priv;
1596 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1597 * be reallocated underneath the timeout handler's processing, then
1598 * the expire check is reliable. If the request is not expired, then
1599 * it was completed and reallocated as a new request after returning
1600 * from blk_mq_check_expired().
1602 if (blk_mq_req_expired(rq, expired)) {
1603 expired->has_timedout_rq = true;
1609 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1611 struct blk_expired_data *expired = priv;
1613 if (blk_mq_req_expired(rq, expired))
1614 blk_mq_rq_timed_out(rq);
1618 static void blk_mq_timeout_work(struct work_struct *work)
1620 struct request_queue *q =
1621 container_of(work, struct request_queue, timeout_work);
1622 struct blk_expired_data expired = {
1623 .timeout_start = jiffies,
1625 struct blk_mq_hw_ctx *hctx;
1628 /* A deadlock might occur if a request is stuck requiring a
1629 * timeout at the same time a queue freeze is waiting
1630 * completion, since the timeout code would not be able to
1631 * acquire the queue reference here.
1633 * That's why we don't use blk_queue_enter here; instead, we use
1634 * percpu_ref_tryget directly, because we need to be able to
1635 * obtain a reference even in the short window between the queue
1636 * starting to freeze, by dropping the first reference in
1637 * blk_freeze_queue_start, and the moment the last request is
1638 * consumed, marked by the instant q_usage_counter reaches
1641 if (!percpu_ref_tryget(&q->q_usage_counter))
1644 /* check if there is any timed-out request */
1645 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1646 if (expired.has_timedout_rq) {
1648 * Before walking tags, we must ensure any submit started
1649 * before the current time has finished. Since the submit
1650 * uses srcu or rcu, wait for a synchronization point to
1651 * ensure all running submits have finished
1653 blk_mq_wait_quiesce_done(q->tag_set);
1656 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1659 if (expired.next != 0) {
1660 mod_timer(&q->timeout, expired.next);
1663 * Request timeouts are handled as a forward rolling timer. If
1664 * we end up here it means that no requests are pending and
1665 * also that no request has been pending for a while. Mark
1666 * each hctx as idle.
1668 queue_for_each_hw_ctx(q, hctx, i) {
1669 /* the hctx may be unmapped, so check it here */
1670 if (blk_mq_hw_queue_mapped(hctx))
1671 blk_mq_tag_idle(hctx);
1677 struct flush_busy_ctx_data {
1678 struct blk_mq_hw_ctx *hctx;
1679 struct list_head *list;
1682 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1684 struct flush_busy_ctx_data *flush_data = data;
1685 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1686 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1687 enum hctx_type type = hctx->type;
1689 spin_lock(&ctx->lock);
1690 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1691 sbitmap_clear_bit(sb, bitnr);
1692 spin_unlock(&ctx->lock);
1697 * Process software queues that have been marked busy, splicing them
1698 * to the for-dispatch
1700 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1702 struct flush_busy_ctx_data data = {
1707 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1709 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1711 struct dispatch_rq_data {
1712 struct blk_mq_hw_ctx *hctx;
1716 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1719 struct dispatch_rq_data *dispatch_data = data;
1720 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1721 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1722 enum hctx_type type = hctx->type;
1724 spin_lock(&ctx->lock);
1725 if (!list_empty(&ctx->rq_lists[type])) {
1726 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1727 list_del_init(&dispatch_data->rq->queuelist);
1728 if (list_empty(&ctx->rq_lists[type]))
1729 sbitmap_clear_bit(sb, bitnr);
1731 spin_unlock(&ctx->lock);
1733 return !dispatch_data->rq;
1736 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1737 struct blk_mq_ctx *start)
1739 unsigned off = start ? start->index_hw[hctx->type] : 0;
1740 struct dispatch_rq_data data = {
1745 __sbitmap_for_each_set(&hctx->ctx_map, off,
1746 dispatch_rq_from_ctx, &data);
1751 bool __blk_mq_alloc_driver_tag(struct request *rq)
1753 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1754 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1757 blk_mq_tag_busy(rq->mq_hctx);
1759 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1760 bt = &rq->mq_hctx->tags->breserved_tags;
1763 if (!hctx_may_queue(rq->mq_hctx, bt))
1767 tag = __sbitmap_queue_get(bt);
1768 if (tag == BLK_MQ_NO_TAG)
1771 rq->tag = tag + tag_offset;
1772 blk_mq_inc_active_requests(rq->mq_hctx);
1776 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1777 int flags, void *key)
1779 struct blk_mq_hw_ctx *hctx;
1781 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1783 spin_lock(&hctx->dispatch_wait_lock);
1784 if (!list_empty(&wait->entry)) {
1785 struct sbitmap_queue *sbq;
1787 list_del_init(&wait->entry);
1788 sbq = &hctx->tags->bitmap_tags;
1789 atomic_dec(&sbq->ws_active);
1791 spin_unlock(&hctx->dispatch_wait_lock);
1793 blk_mq_run_hw_queue(hctx, true);
1798 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1799 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1800 * restart. For both cases, take care to check the condition again after
1801 * marking us as waiting.
1803 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1806 struct sbitmap_queue *sbq;
1807 struct wait_queue_head *wq;
1808 wait_queue_entry_t *wait;
1811 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1812 !(blk_mq_is_shared_tags(hctx->flags))) {
1813 blk_mq_sched_mark_restart_hctx(hctx);
1816 * It's possible that a tag was freed in the window between the
1817 * allocation failure and adding the hardware queue to the wait
1820 * Don't clear RESTART here, someone else could have set it.
1821 * At most this will cost an extra queue run.
1823 return blk_mq_get_driver_tag(rq);
1826 wait = &hctx->dispatch_wait;
1827 if (!list_empty_careful(&wait->entry))
1830 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1831 sbq = &hctx->tags->breserved_tags;
1833 sbq = &hctx->tags->bitmap_tags;
1834 wq = &bt_wait_ptr(sbq, hctx)->wait;
1836 spin_lock_irq(&wq->lock);
1837 spin_lock(&hctx->dispatch_wait_lock);
1838 if (!list_empty(&wait->entry)) {
1839 spin_unlock(&hctx->dispatch_wait_lock);
1840 spin_unlock_irq(&wq->lock);
1844 atomic_inc(&sbq->ws_active);
1845 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1846 __add_wait_queue(wq, wait);
1849 * It's possible that a tag was freed in the window between the
1850 * allocation failure and adding the hardware queue to the wait
1853 ret = blk_mq_get_driver_tag(rq);
1855 spin_unlock(&hctx->dispatch_wait_lock);
1856 spin_unlock_irq(&wq->lock);
1861 * We got a tag, remove ourselves from the wait queue to ensure
1862 * someone else gets the wakeup.
1864 list_del_init(&wait->entry);
1865 atomic_dec(&sbq->ws_active);
1866 spin_unlock(&hctx->dispatch_wait_lock);
1867 spin_unlock_irq(&wq->lock);
1872 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1873 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1875 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1876 * - EWMA is one simple way to compute running average value
1877 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1878 * - take 4 as factor for avoiding to get too small(0) result, and this
1879 * factor doesn't matter because EWMA decreases exponentially
1881 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1885 ewma = hctx->dispatch_busy;
1890 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1892 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1893 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1895 hctx->dispatch_busy = ewma;
1898 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1900 static void blk_mq_handle_dev_resource(struct request *rq,
1901 struct list_head *list)
1903 list_add(&rq->queuelist, list);
1904 __blk_mq_requeue_request(rq);
1907 static void blk_mq_handle_zone_resource(struct request *rq,
1908 struct list_head *zone_list)
1911 * If we end up here it is because we cannot dispatch a request to a
1912 * specific zone due to LLD level zone-write locking or other zone
1913 * related resource not being available. In this case, set the request
1914 * aside in zone_list for retrying it later.
1916 list_add(&rq->queuelist, zone_list);
1917 __blk_mq_requeue_request(rq);
1920 enum prep_dispatch {
1922 PREP_DISPATCH_NO_TAG,
1923 PREP_DISPATCH_NO_BUDGET,
1926 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1929 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1930 int budget_token = -1;
1933 budget_token = blk_mq_get_dispatch_budget(rq->q);
1934 if (budget_token < 0) {
1935 blk_mq_put_driver_tag(rq);
1936 return PREP_DISPATCH_NO_BUDGET;
1938 blk_mq_set_rq_budget_token(rq, budget_token);
1941 if (!blk_mq_get_driver_tag(rq)) {
1943 * The initial allocation attempt failed, so we need to
1944 * rerun the hardware queue when a tag is freed. The
1945 * waitqueue takes care of that. If the queue is run
1946 * before we add this entry back on the dispatch list,
1947 * we'll re-run it below.
1949 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1951 * All budgets not got from this function will be put
1952 * together during handling partial dispatch
1955 blk_mq_put_dispatch_budget(rq->q, budget_token);
1956 return PREP_DISPATCH_NO_TAG;
1960 return PREP_DISPATCH_OK;
1963 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1964 static void blk_mq_release_budgets(struct request_queue *q,
1965 struct list_head *list)
1969 list_for_each_entry(rq, list, queuelist) {
1970 int budget_token = blk_mq_get_rq_budget_token(rq);
1972 if (budget_token >= 0)
1973 blk_mq_put_dispatch_budget(q, budget_token);
1978 * blk_mq_commit_rqs will notify driver using bd->last that there is no
1979 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
1981 * Attention, we should explicitly call this in unusual cases:
1982 * 1) did not queue everything initially scheduled to queue
1983 * 2) the last attempt to queue a request failed
1985 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
1988 if (hctx->queue->mq_ops->commit_rqs && queued) {
1989 trace_block_unplug(hctx->queue, queued, !from_schedule);
1990 hctx->queue->mq_ops->commit_rqs(hctx);
1995 * Returns true if we did some work AND can potentially do more.
1997 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1998 unsigned int nr_budgets)
2000 enum prep_dispatch prep;
2001 struct request_queue *q = hctx->queue;
2004 blk_status_t ret = BLK_STS_OK;
2005 LIST_HEAD(zone_list);
2006 bool needs_resource = false;
2008 if (list_empty(list))
2012 * Now process all the entries, sending them to the driver.
2016 struct blk_mq_queue_data bd;
2018 rq = list_first_entry(list, struct request, queuelist);
2020 WARN_ON_ONCE(hctx != rq->mq_hctx);
2021 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2022 if (prep != PREP_DISPATCH_OK)
2025 list_del_init(&rq->queuelist);
2028 bd.last = list_empty(list);
2031 * once the request is queued to lld, no need to cover the
2036 ret = q->mq_ops->queue_rq(hctx, &bd);
2041 case BLK_STS_RESOURCE:
2042 needs_resource = true;
2044 case BLK_STS_DEV_RESOURCE:
2045 blk_mq_handle_dev_resource(rq, list);
2047 case BLK_STS_ZONE_RESOURCE:
2049 * Move the request to zone_list and keep going through
2050 * the dispatch list to find more requests the drive can
2053 blk_mq_handle_zone_resource(rq, &zone_list);
2054 needs_resource = true;
2057 blk_mq_end_request(rq, ret);
2059 } while (!list_empty(list));
2061 if (!list_empty(&zone_list))
2062 list_splice_tail_init(&zone_list, list);
2064 /* If we didn't flush the entire list, we could have told the driver
2065 * there was more coming, but that turned out to be a lie.
2067 if (!list_empty(list) || ret != BLK_STS_OK)
2068 blk_mq_commit_rqs(hctx, queued, false);
2071 * Any items that need requeuing? Stuff them into hctx->dispatch,
2072 * that is where we will continue on next queue run.
2074 if (!list_empty(list)) {
2076 /* For non-shared tags, the RESTART check will suffice */
2077 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2078 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2079 blk_mq_is_shared_tags(hctx->flags));
2082 blk_mq_release_budgets(q, list);
2084 spin_lock(&hctx->lock);
2085 list_splice_tail_init(list, &hctx->dispatch);
2086 spin_unlock(&hctx->lock);
2089 * Order adding requests to hctx->dispatch and checking
2090 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2091 * in blk_mq_sched_restart(). Avoid restart code path to
2092 * miss the new added requests to hctx->dispatch, meantime
2093 * SCHED_RESTART is observed here.
2098 * If SCHED_RESTART was set by the caller of this function and
2099 * it is no longer set that means that it was cleared by another
2100 * thread and hence that a queue rerun is needed.
2102 * If 'no_tag' is set, that means that we failed getting
2103 * a driver tag with an I/O scheduler attached. If our dispatch
2104 * waitqueue is no longer active, ensure that we run the queue
2105 * AFTER adding our entries back to the list.
2107 * If no I/O scheduler has been configured it is possible that
2108 * the hardware queue got stopped and restarted before requests
2109 * were pushed back onto the dispatch list. Rerun the queue to
2110 * avoid starvation. Notes:
2111 * - blk_mq_run_hw_queue() checks whether or not a queue has
2112 * been stopped before rerunning a queue.
2113 * - Some but not all block drivers stop a queue before
2114 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2117 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2118 * bit is set, run queue after a delay to avoid IO stalls
2119 * that could otherwise occur if the queue is idle. We'll do
2120 * similar if we couldn't get budget or couldn't lock a zone
2121 * and SCHED_RESTART is set.
2123 needs_restart = blk_mq_sched_needs_restart(hctx);
2124 if (prep == PREP_DISPATCH_NO_BUDGET)
2125 needs_resource = true;
2126 if (!needs_restart ||
2127 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2128 blk_mq_run_hw_queue(hctx, true);
2129 else if (needs_resource)
2130 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2132 blk_mq_update_dispatch_busy(hctx, true);
2136 blk_mq_update_dispatch_busy(hctx, false);
2140 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2142 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2144 if (cpu >= nr_cpu_ids)
2145 cpu = cpumask_first(hctx->cpumask);
2150 * It'd be great if the workqueue API had a way to pass
2151 * in a mask and had some smarts for more clever placement.
2152 * For now we just round-robin here, switching for every
2153 * BLK_MQ_CPU_WORK_BATCH queued items.
2155 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2158 int next_cpu = hctx->next_cpu;
2160 if (hctx->queue->nr_hw_queues == 1)
2161 return WORK_CPU_UNBOUND;
2163 if (--hctx->next_cpu_batch <= 0) {
2165 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2167 if (next_cpu >= nr_cpu_ids)
2168 next_cpu = blk_mq_first_mapped_cpu(hctx);
2169 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2173 * Do unbound schedule if we can't find a online CPU for this hctx,
2174 * and it should only happen in the path of handling CPU DEAD.
2176 if (!cpu_online(next_cpu)) {
2183 * Make sure to re-select CPU next time once after CPUs
2184 * in hctx->cpumask become online again.
2186 hctx->next_cpu = next_cpu;
2187 hctx->next_cpu_batch = 1;
2188 return WORK_CPU_UNBOUND;
2191 hctx->next_cpu = next_cpu;
2196 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2197 * @hctx: Pointer to the hardware queue to run.
2198 * @msecs: Milliseconds of delay to wait before running the queue.
2200 * Run a hardware queue asynchronously with a delay of @msecs.
2202 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2204 if (unlikely(blk_mq_hctx_stopped(hctx)))
2206 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2207 msecs_to_jiffies(msecs));
2209 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2212 * blk_mq_run_hw_queue - Start to run a hardware queue.
2213 * @hctx: Pointer to the hardware queue to run.
2214 * @async: If we want to run the queue asynchronously.
2216 * Check if the request queue is not in a quiesced state and if there are
2217 * pending requests to be sent. If this is true, run the queue to send requests
2220 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2225 * We can't run the queue inline with interrupts disabled.
2227 WARN_ON_ONCE(!async && in_interrupt());
2229 might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING);
2232 * When queue is quiesced, we may be switching io scheduler, or
2233 * updating nr_hw_queues, or other things, and we can't run queue
2234 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2236 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2239 __blk_mq_run_dispatch_ops(hctx->queue, false,
2240 need_run = !blk_queue_quiesced(hctx->queue) &&
2241 blk_mq_hctx_has_pending(hctx));
2246 if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2247 blk_mq_delay_run_hw_queue(hctx, 0);
2251 blk_mq_run_dispatch_ops(hctx->queue,
2252 blk_mq_sched_dispatch_requests(hctx));
2254 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2257 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2260 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2262 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2264 * If the IO scheduler does not respect hardware queues when
2265 * dispatching, we just don't bother with multiple HW queues and
2266 * dispatch from hctx for the current CPU since running multiple queues
2267 * just causes lock contention inside the scheduler and pointless cache
2270 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2272 if (!blk_mq_hctx_stopped(hctx))
2278 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2279 * @q: Pointer to the request queue to run.
2280 * @async: If we want to run the queue asynchronously.
2282 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2284 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2288 if (blk_queue_sq_sched(q))
2289 sq_hctx = blk_mq_get_sq_hctx(q);
2290 queue_for_each_hw_ctx(q, hctx, i) {
2291 if (blk_mq_hctx_stopped(hctx))
2294 * Dispatch from this hctx either if there's no hctx preferred
2295 * by IO scheduler or if it has requests that bypass the
2298 if (!sq_hctx || sq_hctx == hctx ||
2299 !list_empty_careful(&hctx->dispatch))
2300 blk_mq_run_hw_queue(hctx, async);
2303 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2306 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2307 * @q: Pointer to the request queue to run.
2308 * @msecs: Milliseconds of delay to wait before running the queues.
2310 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2312 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2316 if (blk_queue_sq_sched(q))
2317 sq_hctx = blk_mq_get_sq_hctx(q);
2318 queue_for_each_hw_ctx(q, hctx, i) {
2319 if (blk_mq_hctx_stopped(hctx))
2322 * If there is already a run_work pending, leave the
2323 * pending delay untouched. Otherwise, a hctx can stall
2324 * if another hctx is re-delaying the other's work
2325 * before the work executes.
2327 if (delayed_work_pending(&hctx->run_work))
2330 * Dispatch from this hctx either if there's no hctx preferred
2331 * by IO scheduler or if it has requests that bypass the
2334 if (!sq_hctx || sq_hctx == hctx ||
2335 !list_empty_careful(&hctx->dispatch))
2336 blk_mq_delay_run_hw_queue(hctx, msecs);
2339 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2342 * This function is often used for pausing .queue_rq() by driver when
2343 * there isn't enough resource or some conditions aren't satisfied, and
2344 * BLK_STS_RESOURCE is usually returned.
2346 * We do not guarantee that dispatch can be drained or blocked
2347 * after blk_mq_stop_hw_queue() returns. Please use
2348 * blk_mq_quiesce_queue() for that requirement.
2350 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2352 cancel_delayed_work(&hctx->run_work);
2354 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2356 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2359 * This function is often used for pausing .queue_rq() by driver when
2360 * there isn't enough resource or some conditions aren't satisfied, and
2361 * BLK_STS_RESOURCE is usually returned.
2363 * We do not guarantee that dispatch can be drained or blocked
2364 * after blk_mq_stop_hw_queues() returns. Please use
2365 * blk_mq_quiesce_queue() for that requirement.
2367 void blk_mq_stop_hw_queues(struct request_queue *q)
2369 struct blk_mq_hw_ctx *hctx;
2372 queue_for_each_hw_ctx(q, hctx, i)
2373 blk_mq_stop_hw_queue(hctx);
2375 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2377 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2379 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2381 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
2383 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2385 void blk_mq_start_hw_queues(struct request_queue *q)
2387 struct blk_mq_hw_ctx *hctx;
2390 queue_for_each_hw_ctx(q, hctx, i)
2391 blk_mq_start_hw_queue(hctx);
2393 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2395 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2397 if (!blk_mq_hctx_stopped(hctx))
2400 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2401 blk_mq_run_hw_queue(hctx, async);
2403 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2405 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2407 struct blk_mq_hw_ctx *hctx;
2410 queue_for_each_hw_ctx(q, hctx, i)
2411 blk_mq_start_stopped_hw_queue(hctx, async ||
2412 (hctx->flags & BLK_MQ_F_BLOCKING));
2414 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2416 static void blk_mq_run_work_fn(struct work_struct *work)
2418 struct blk_mq_hw_ctx *hctx =
2419 container_of(work, struct blk_mq_hw_ctx, run_work.work);
2421 blk_mq_run_dispatch_ops(hctx->queue,
2422 blk_mq_sched_dispatch_requests(hctx));
2426 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2427 * @rq: Pointer to request to be inserted.
2428 * @flags: BLK_MQ_INSERT_*
2430 * Should only be used carefully, when the caller knows we want to
2431 * bypass a potential IO scheduler on the target device.
2433 static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2435 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2437 spin_lock(&hctx->lock);
2438 if (flags & BLK_MQ_INSERT_AT_HEAD)
2439 list_add(&rq->queuelist, &hctx->dispatch);
2441 list_add_tail(&rq->queuelist, &hctx->dispatch);
2442 spin_unlock(&hctx->lock);
2445 static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2446 struct blk_mq_ctx *ctx, struct list_head *list,
2447 bool run_queue_async)
2450 enum hctx_type type = hctx->type;
2453 * Try to issue requests directly if the hw queue isn't busy to save an
2454 * extra enqueue & dequeue to the sw queue.
2456 if (!hctx->dispatch_busy && !run_queue_async) {
2457 blk_mq_run_dispatch_ops(hctx->queue,
2458 blk_mq_try_issue_list_directly(hctx, list));
2459 if (list_empty(list))
2464 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2467 list_for_each_entry(rq, list, queuelist) {
2468 BUG_ON(rq->mq_ctx != ctx);
2469 trace_block_rq_insert(rq);
2470 if (rq->cmd_flags & REQ_NOWAIT)
2471 run_queue_async = true;
2474 spin_lock(&ctx->lock);
2475 list_splice_tail_init(list, &ctx->rq_lists[type]);
2476 blk_mq_hctx_mark_pending(hctx, ctx);
2477 spin_unlock(&ctx->lock);
2479 blk_mq_run_hw_queue(hctx, run_queue_async);
2482 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2484 struct request_queue *q = rq->q;
2485 struct blk_mq_ctx *ctx = rq->mq_ctx;
2486 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2488 if (blk_rq_is_passthrough(rq)) {
2490 * Passthrough request have to be added to hctx->dispatch
2491 * directly. The device may be in a situation where it can't
2492 * handle FS request, and always returns BLK_STS_RESOURCE for
2493 * them, which gets them added to hctx->dispatch.
2495 * If a passthrough request is required to unblock the queues,
2496 * and it is added to the scheduler queue, there is no chance to
2497 * dispatch it given we prioritize requests in hctx->dispatch.
2499 blk_mq_request_bypass_insert(rq, flags);
2500 } else if (req_op(rq) == REQ_OP_FLUSH) {
2502 * Firstly normal IO request is inserted to scheduler queue or
2503 * sw queue, meantime we add flush request to dispatch queue(
2504 * hctx->dispatch) directly and there is at most one in-flight
2505 * flush request for each hw queue, so it doesn't matter to add
2506 * flush request to tail or front of the dispatch queue.
2508 * Secondly in case of NCQ, flush request belongs to non-NCQ
2509 * command, and queueing it will fail when there is any
2510 * in-flight normal IO request(NCQ command). When adding flush
2511 * rq to the front of hctx->dispatch, it is easier to introduce
2512 * extra time to flush rq's latency because of S_SCHED_RESTART
2513 * compared with adding to the tail of dispatch queue, then
2514 * chance of flush merge is increased, and less flush requests
2515 * will be issued to controller. It is observed that ~10% time
2516 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2517 * drive when adding flush rq to the front of hctx->dispatch.
2519 * Simply queue flush rq to the front of hctx->dispatch so that
2520 * intensive flush workloads can benefit in case of NCQ HW.
2522 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2523 } else if (q->elevator) {
2526 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2528 list_add(&rq->queuelist, &list);
2529 q->elevator->type->ops.insert_requests(hctx, &list, flags);
2531 trace_block_rq_insert(rq);
2533 spin_lock(&ctx->lock);
2534 if (flags & BLK_MQ_INSERT_AT_HEAD)
2535 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2537 list_add_tail(&rq->queuelist,
2538 &ctx->rq_lists[hctx->type]);
2539 blk_mq_hctx_mark_pending(hctx, ctx);
2540 spin_unlock(&ctx->lock);
2544 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2545 unsigned int nr_segs)
2549 if (bio->bi_opf & REQ_RAHEAD)
2550 rq->cmd_flags |= REQ_FAILFAST_MASK;
2552 rq->__sector = bio->bi_iter.bi_sector;
2553 blk_rq_bio_prep(rq, bio, nr_segs);
2555 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2556 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2559 blk_account_io_start(rq);
2562 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2563 struct request *rq, bool last)
2565 struct request_queue *q = rq->q;
2566 struct blk_mq_queue_data bd = {
2573 * For OK queue, we are done. For error, caller may kill it.
2574 * Any other error (busy), just add it to our list as we
2575 * previously would have done.
2577 ret = q->mq_ops->queue_rq(hctx, &bd);
2580 blk_mq_update_dispatch_busy(hctx, false);
2582 case BLK_STS_RESOURCE:
2583 case BLK_STS_DEV_RESOURCE:
2584 blk_mq_update_dispatch_busy(hctx, true);
2585 __blk_mq_requeue_request(rq);
2588 blk_mq_update_dispatch_busy(hctx, false);
2595 static bool blk_mq_get_budget_and_tag(struct request *rq)
2599 budget_token = blk_mq_get_dispatch_budget(rq->q);
2600 if (budget_token < 0)
2602 blk_mq_set_rq_budget_token(rq, budget_token);
2603 if (!blk_mq_get_driver_tag(rq)) {
2604 blk_mq_put_dispatch_budget(rq->q, budget_token);
2611 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2612 * @hctx: Pointer of the associated hardware queue.
2613 * @rq: Pointer to request to be sent.
2615 * If the device has enough resources to accept a new request now, send the
2616 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2617 * we can try send it another time in the future. Requests inserted at this
2618 * queue have higher priority.
2620 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2625 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2626 blk_mq_insert_request(rq, 0);
2630 if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
2631 blk_mq_insert_request(rq, 0);
2632 blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT);
2636 ret = __blk_mq_issue_directly(hctx, rq, true);
2640 case BLK_STS_RESOURCE:
2641 case BLK_STS_DEV_RESOURCE:
2642 blk_mq_request_bypass_insert(rq, 0);
2643 blk_mq_run_hw_queue(hctx, false);
2646 blk_mq_end_request(rq, ret);
2651 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2653 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2655 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2656 blk_mq_insert_request(rq, 0);
2660 if (!blk_mq_get_budget_and_tag(rq))
2661 return BLK_STS_RESOURCE;
2662 return __blk_mq_issue_directly(hctx, rq, last);
2665 static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2667 struct blk_mq_hw_ctx *hctx = NULL;
2670 blk_status_t ret = BLK_STS_OK;
2672 while ((rq = rq_list_pop(&plug->mq_list))) {
2673 bool last = rq_list_empty(plug->mq_list);
2675 if (hctx != rq->mq_hctx) {
2677 blk_mq_commit_rqs(hctx, queued, false);
2683 ret = blk_mq_request_issue_directly(rq, last);
2688 case BLK_STS_RESOURCE:
2689 case BLK_STS_DEV_RESOURCE:
2690 blk_mq_request_bypass_insert(rq, 0);
2691 blk_mq_run_hw_queue(hctx, false);
2694 blk_mq_end_request(rq, ret);
2700 if (ret != BLK_STS_OK)
2701 blk_mq_commit_rqs(hctx, queued, false);
2704 static void __blk_mq_flush_plug_list(struct request_queue *q,
2705 struct blk_plug *plug)
2707 if (blk_queue_quiesced(q))
2709 q->mq_ops->queue_rqs(&plug->mq_list);
2712 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2714 struct blk_mq_hw_ctx *this_hctx = NULL;
2715 struct blk_mq_ctx *this_ctx = NULL;
2716 struct request *requeue_list = NULL;
2717 struct request **requeue_lastp = &requeue_list;
2718 unsigned int depth = 0;
2719 bool is_passthrough = false;
2723 struct request *rq = rq_list_pop(&plug->mq_list);
2726 this_hctx = rq->mq_hctx;
2727 this_ctx = rq->mq_ctx;
2728 is_passthrough = blk_rq_is_passthrough(rq);
2729 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2730 is_passthrough != blk_rq_is_passthrough(rq)) {
2731 rq_list_add_tail(&requeue_lastp, rq);
2734 list_add(&rq->queuelist, &list);
2736 } while (!rq_list_empty(plug->mq_list));
2738 plug->mq_list = requeue_list;
2739 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2741 percpu_ref_get(&this_hctx->queue->q_usage_counter);
2742 /* passthrough requests should never be issued to the I/O scheduler */
2743 if (is_passthrough) {
2744 spin_lock(&this_hctx->lock);
2745 list_splice_tail_init(&list, &this_hctx->dispatch);
2746 spin_unlock(&this_hctx->lock);
2747 blk_mq_run_hw_queue(this_hctx, from_sched);
2748 } else if (this_hctx->queue->elevator) {
2749 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2751 blk_mq_run_hw_queue(this_hctx, from_sched);
2753 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2755 percpu_ref_put(&this_hctx->queue->q_usage_counter);
2758 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2763 * We may have been called recursively midway through handling
2764 * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2765 * To avoid mq_list changing under our feet, clear rq_count early and
2766 * bail out specifically if rq_count is 0 rather than checking
2767 * whether the mq_list is empty.
2769 if (plug->rq_count == 0)
2773 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2774 struct request_queue *q;
2776 rq = rq_list_peek(&plug->mq_list);
2780 * Peek first request and see if we have a ->queue_rqs() hook.
2781 * If we do, we can dispatch the whole plug list in one go. We
2782 * already know at this point that all requests belong to the
2783 * same queue, caller must ensure that's the case.
2785 if (q->mq_ops->queue_rqs) {
2786 blk_mq_run_dispatch_ops(q,
2787 __blk_mq_flush_plug_list(q, plug));
2788 if (rq_list_empty(plug->mq_list))
2792 blk_mq_run_dispatch_ops(q,
2793 blk_mq_plug_issue_direct(plug));
2794 if (rq_list_empty(plug->mq_list))
2799 blk_mq_dispatch_plug_list(plug, from_schedule);
2800 } while (!rq_list_empty(plug->mq_list));
2803 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2804 struct list_head *list)
2807 blk_status_t ret = BLK_STS_OK;
2809 while (!list_empty(list)) {
2810 struct request *rq = list_first_entry(list, struct request,
2813 list_del_init(&rq->queuelist);
2814 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2819 case BLK_STS_RESOURCE:
2820 case BLK_STS_DEV_RESOURCE:
2821 blk_mq_request_bypass_insert(rq, 0);
2822 if (list_empty(list))
2823 blk_mq_run_hw_queue(hctx, false);
2826 blk_mq_end_request(rq, ret);
2832 if (ret != BLK_STS_OK)
2833 blk_mq_commit_rqs(hctx, queued, false);
2836 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2837 struct bio *bio, unsigned int nr_segs)
2839 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2840 if (blk_attempt_plug_merge(q, bio, nr_segs))
2842 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2848 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2849 struct blk_plug *plug,
2853 struct blk_mq_alloc_data data = {
2856 .cmd_flags = bio->bi_opf,
2860 if (unlikely(bio_queue_enter(bio)))
2863 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2866 rq_qos_throttle(q, bio);
2869 data.nr_tags = plug->nr_ios;
2871 data.cached_rq = &plug->cached_rq;
2874 rq = __blk_mq_alloc_requests(&data);
2877 rq_qos_cleanup(q, bio);
2878 if (bio->bi_opf & REQ_NOWAIT)
2879 bio_wouldblock_error(bio);
2885 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2886 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2889 enum hctx_type type, hctx_type;
2893 rq = rq_list_peek(&plug->cached_rq);
2894 if (!rq || rq->q != q)
2897 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2902 type = blk_mq_get_hctx_type((*bio)->bi_opf);
2903 hctx_type = rq->mq_hctx->type;
2904 if (type != hctx_type &&
2905 !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2907 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2911 * If any qos ->throttle() end up blocking, we will have flushed the
2912 * plug and hence killed the cached_rq list as well. Pop this entry
2913 * before we throttle.
2915 plug->cached_rq = rq_list_next(rq);
2916 rq_qos_throttle(q, *bio);
2918 blk_mq_rq_time_init(rq, 0);
2919 rq->cmd_flags = (*bio)->bi_opf;
2920 INIT_LIST_HEAD(&rq->queuelist);
2924 static void bio_set_ioprio(struct bio *bio)
2926 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2927 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2928 bio->bi_ioprio = get_current_ioprio();
2929 blkcg_set_ioprio(bio);
2933 * blk_mq_submit_bio - Create and send a request to block device.
2934 * @bio: Bio pointer.
2936 * Builds up a request structure from @q and @bio and send to the device. The
2937 * request may not be queued directly to hardware if:
2938 * * This request can be merged with another one
2939 * * We want to place request at plug queue for possible future merging
2940 * * There is an IO scheduler active at this queue
2942 * It will not queue the request if there is an error with the bio, or at the
2945 void blk_mq_submit_bio(struct bio *bio)
2947 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2948 struct blk_plug *plug = blk_mq_plug(bio);
2949 const int is_sync = op_is_sync(bio->bi_opf);
2950 struct blk_mq_hw_ctx *hctx;
2952 unsigned int nr_segs = 1;
2955 bio = blk_queue_bounce(bio, q);
2956 if (bio_may_exceed_limits(bio, &q->limits)) {
2957 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2962 if (!bio_integrity_prep(bio))
2965 bio_set_ioprio(bio);
2967 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2971 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2976 trace_block_getrq(bio);
2978 rq_qos_track(q, rq, bio);
2980 blk_mq_bio_to_request(rq, bio, nr_segs);
2982 ret = blk_crypto_rq_get_keyslot(rq);
2983 if (ret != BLK_STS_OK) {
2984 bio->bi_status = ret;
2986 blk_mq_free_request(rq);
2990 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
2994 blk_add_rq_to_plug(plug, rq);
2999 if ((rq->rq_flags & RQF_USE_SCHED) ||
3000 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
3001 blk_mq_insert_request(rq, 0);
3002 blk_mq_run_hw_queue(hctx, true);
3004 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
3008 #ifdef CONFIG_BLK_MQ_STACKING
3010 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3011 * @rq: the request being queued
3013 blk_status_t blk_insert_cloned_request(struct request *rq)
3015 struct request_queue *q = rq->q;
3016 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3017 unsigned int max_segments = blk_rq_get_max_segments(rq);
3020 if (blk_rq_sectors(rq) > max_sectors) {
3022 * SCSI device does not have a good way to return if
3023 * Write Same/Zero is actually supported. If a device rejects
3024 * a non-read/write command (discard, write same,etc.) the
3025 * low-level device driver will set the relevant queue limit to
3026 * 0 to prevent blk-lib from issuing more of the offending
3027 * operations. Commands queued prior to the queue limit being
3028 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3029 * errors being propagated to upper layers.
3031 if (max_sectors == 0)
3032 return BLK_STS_NOTSUPP;
3034 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3035 __func__, blk_rq_sectors(rq), max_sectors);
3036 return BLK_STS_IOERR;
3040 * The queue settings related to segment counting may differ from the
3043 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3044 if (rq->nr_phys_segments > max_segments) {
3045 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3046 __func__, rq->nr_phys_segments, max_segments);
3047 return BLK_STS_IOERR;
3050 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3051 return BLK_STS_IOERR;
3053 ret = blk_crypto_rq_get_keyslot(rq);
3054 if (ret != BLK_STS_OK)
3057 blk_account_io_start(rq);
3060 * Since we have a scheduler attached on the top device,
3061 * bypass a potential scheduler on the bottom device for
3064 blk_mq_run_dispatch_ops(q,
3065 ret = blk_mq_request_issue_directly(rq, true));
3067 blk_account_io_done(rq, ktime_get_ns());
3070 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3073 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3074 * @rq: the clone request to be cleaned up
3077 * Free all bios in @rq for a cloned request.
3079 void blk_rq_unprep_clone(struct request *rq)
3083 while ((bio = rq->bio) != NULL) {
3084 rq->bio = bio->bi_next;
3089 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3092 * blk_rq_prep_clone - Helper function to setup clone request
3093 * @rq: the request to be setup
3094 * @rq_src: original request to be cloned
3095 * @bs: bio_set that bios for clone are allocated from
3096 * @gfp_mask: memory allocation mask for bio
3097 * @bio_ctr: setup function to be called for each clone bio.
3098 * Returns %0 for success, non %0 for failure.
3099 * @data: private data to be passed to @bio_ctr
3102 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3103 * Also, pages which the original bios are pointing to are not copied
3104 * and the cloned bios just point same pages.
3105 * So cloned bios must be completed before original bios, which means
3106 * the caller must complete @rq before @rq_src.
3108 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3109 struct bio_set *bs, gfp_t gfp_mask,
3110 int (*bio_ctr)(struct bio *, struct bio *, void *),
3113 struct bio *bio, *bio_src;
3118 __rq_for_each_bio(bio_src, rq_src) {
3119 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3124 if (bio_ctr && bio_ctr(bio, bio_src, data))
3128 rq->biotail->bi_next = bio;
3131 rq->bio = rq->biotail = bio;
3136 /* Copy attributes of the original request to the clone request. */
3137 rq->__sector = blk_rq_pos(rq_src);
3138 rq->__data_len = blk_rq_bytes(rq_src);
3139 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3140 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3141 rq->special_vec = rq_src->special_vec;
3143 rq->nr_phys_segments = rq_src->nr_phys_segments;
3144 rq->ioprio = rq_src->ioprio;
3146 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3154 blk_rq_unprep_clone(rq);
3158 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3159 #endif /* CONFIG_BLK_MQ_STACKING */
3162 * Steal bios from a request and add them to a bio list.
3163 * The request must not have been partially completed before.
3165 void blk_steal_bios(struct bio_list *list, struct request *rq)
3169 list->tail->bi_next = rq->bio;
3171 list->head = rq->bio;
3172 list->tail = rq->biotail;
3180 EXPORT_SYMBOL_GPL(blk_steal_bios);
3182 static size_t order_to_size(unsigned int order)
3184 return (size_t)PAGE_SIZE << order;
3187 /* called before freeing request pool in @tags */
3188 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3189 struct blk_mq_tags *tags)
3192 unsigned long flags;
3195 * There is no need to clear mapping if driver tags is not initialized
3196 * or the mapping belongs to the driver tags.
3198 if (!drv_tags || drv_tags == tags)
3201 list_for_each_entry(page, &tags->page_list, lru) {
3202 unsigned long start = (unsigned long)page_address(page);
3203 unsigned long end = start + order_to_size(page->private);
3206 for (i = 0; i < drv_tags->nr_tags; i++) {
3207 struct request *rq = drv_tags->rqs[i];
3208 unsigned long rq_addr = (unsigned long)rq;
3210 if (rq_addr >= start && rq_addr < end) {
3211 WARN_ON_ONCE(req_ref_read(rq) != 0);
3212 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3218 * Wait until all pending iteration is done.
3220 * Request reference is cleared and it is guaranteed to be observed
3221 * after the ->lock is released.
3223 spin_lock_irqsave(&drv_tags->lock, flags);
3224 spin_unlock_irqrestore(&drv_tags->lock, flags);
3227 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3228 unsigned int hctx_idx)
3230 struct blk_mq_tags *drv_tags;
3233 if (list_empty(&tags->page_list))
3236 if (blk_mq_is_shared_tags(set->flags))
3237 drv_tags = set->shared_tags;
3239 drv_tags = set->tags[hctx_idx];
3241 if (tags->static_rqs && set->ops->exit_request) {
3244 for (i = 0; i < tags->nr_tags; i++) {
3245 struct request *rq = tags->static_rqs[i];
3249 set->ops->exit_request(set, rq, hctx_idx);
3250 tags->static_rqs[i] = NULL;
3254 blk_mq_clear_rq_mapping(drv_tags, tags);
3256 while (!list_empty(&tags->page_list)) {
3257 page = list_first_entry(&tags->page_list, struct page, lru);
3258 list_del_init(&page->lru);
3260 * Remove kmemleak object previously allocated in
3261 * blk_mq_alloc_rqs().
3263 kmemleak_free(page_address(page));
3264 __free_pages(page, page->private);
3268 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3272 kfree(tags->static_rqs);
3273 tags->static_rqs = NULL;
3275 blk_mq_free_tags(tags);
3278 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3279 unsigned int hctx_idx)
3283 for (i = 0; i < set->nr_maps; i++) {
3284 unsigned int start = set->map[i].queue_offset;
3285 unsigned int end = start + set->map[i].nr_queues;
3287 if (hctx_idx >= start && hctx_idx < end)
3291 if (i >= set->nr_maps)
3292 i = HCTX_TYPE_DEFAULT;
3297 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3298 unsigned int hctx_idx)
3300 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3302 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3305 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3306 unsigned int hctx_idx,
3307 unsigned int nr_tags,
3308 unsigned int reserved_tags)
3310 int node = blk_mq_get_hctx_node(set, hctx_idx);
3311 struct blk_mq_tags *tags;
3313 if (node == NUMA_NO_NODE)
3314 node = set->numa_node;
3316 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3317 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3321 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3322 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3327 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3328 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3330 if (!tags->static_rqs)
3338 blk_mq_free_tags(tags);
3342 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3343 unsigned int hctx_idx, int node)
3347 if (set->ops->init_request) {
3348 ret = set->ops->init_request(set, rq, hctx_idx, node);
3353 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3357 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3358 struct blk_mq_tags *tags,
3359 unsigned int hctx_idx, unsigned int depth)
3361 unsigned int i, j, entries_per_page, max_order = 4;
3362 int node = blk_mq_get_hctx_node(set, hctx_idx);
3363 size_t rq_size, left;
3365 if (node == NUMA_NO_NODE)
3366 node = set->numa_node;
3368 INIT_LIST_HEAD(&tags->page_list);
3371 * rq_size is the size of the request plus driver payload, rounded
3372 * to the cacheline size
3374 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3376 left = rq_size * depth;
3378 for (i = 0; i < depth; ) {
3379 int this_order = max_order;
3384 while (this_order && left < order_to_size(this_order - 1))
3388 page = alloc_pages_node(node,
3389 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3395 if (order_to_size(this_order) < rq_size)
3402 page->private = this_order;
3403 list_add_tail(&page->lru, &tags->page_list);
3405 p = page_address(page);
3407 * Allow kmemleak to scan these pages as they contain pointers
3408 * to additional allocations like via ops->init_request().
3410 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3411 entries_per_page = order_to_size(this_order) / rq_size;
3412 to_do = min(entries_per_page, depth - i);
3413 left -= to_do * rq_size;
3414 for (j = 0; j < to_do; j++) {
3415 struct request *rq = p;
3417 tags->static_rqs[i] = rq;
3418 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3419 tags->static_rqs[i] = NULL;
3430 blk_mq_free_rqs(set, tags, hctx_idx);
3434 struct rq_iter_data {
3435 struct blk_mq_hw_ctx *hctx;
3439 static bool blk_mq_has_request(struct request *rq, void *data)
3441 struct rq_iter_data *iter_data = data;
3443 if (rq->mq_hctx != iter_data->hctx)
3445 iter_data->has_rq = true;
3449 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3451 struct blk_mq_tags *tags = hctx->sched_tags ?
3452 hctx->sched_tags : hctx->tags;
3453 struct rq_iter_data data = {
3457 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3461 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3462 struct blk_mq_hw_ctx *hctx)
3464 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3466 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3471 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3473 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3474 struct blk_mq_hw_ctx, cpuhp_online);
3476 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3477 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3481 * Prevent new request from being allocated on the current hctx.
3483 * The smp_mb__after_atomic() Pairs with the implied barrier in
3484 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3485 * seen once we return from the tag allocator.
3487 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3488 smp_mb__after_atomic();
3491 * Try to grab a reference to the queue and wait for any outstanding
3492 * requests. If we could not grab a reference the queue has been
3493 * frozen and there are no requests.
3495 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3496 while (blk_mq_hctx_has_requests(hctx))
3498 percpu_ref_put(&hctx->queue->q_usage_counter);
3504 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3506 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3507 struct blk_mq_hw_ctx, cpuhp_online);
3509 if (cpumask_test_cpu(cpu, hctx->cpumask))
3510 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3515 * 'cpu' is going away. splice any existing rq_list entries from this
3516 * software queue to the hw queue dispatch list, and ensure that it
3519 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3521 struct blk_mq_hw_ctx *hctx;
3522 struct blk_mq_ctx *ctx;
3524 enum hctx_type type;
3526 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3527 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3530 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3533 spin_lock(&ctx->lock);
3534 if (!list_empty(&ctx->rq_lists[type])) {
3535 list_splice_init(&ctx->rq_lists[type], &tmp);
3536 blk_mq_hctx_clear_pending(hctx, ctx);
3538 spin_unlock(&ctx->lock);
3540 if (list_empty(&tmp))
3543 spin_lock(&hctx->lock);
3544 list_splice_tail_init(&tmp, &hctx->dispatch);
3545 spin_unlock(&hctx->lock);
3547 blk_mq_run_hw_queue(hctx, true);
3551 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3553 if (!(hctx->flags & BLK_MQ_F_STACKING))
3554 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3555 &hctx->cpuhp_online);
3556 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3561 * Before freeing hw queue, clearing the flush request reference in
3562 * tags->rqs[] for avoiding potential UAF.
3564 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3565 unsigned int queue_depth, struct request *flush_rq)
3568 unsigned long flags;
3570 /* The hw queue may not be mapped yet */
3574 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3576 for (i = 0; i < queue_depth; i++)
3577 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3580 * Wait until all pending iteration is done.
3582 * Request reference is cleared and it is guaranteed to be observed
3583 * after the ->lock is released.
3585 spin_lock_irqsave(&tags->lock, flags);
3586 spin_unlock_irqrestore(&tags->lock, flags);
3589 /* hctx->ctxs will be freed in queue's release handler */
3590 static void blk_mq_exit_hctx(struct request_queue *q,
3591 struct blk_mq_tag_set *set,
3592 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3594 struct request *flush_rq = hctx->fq->flush_rq;
3596 if (blk_mq_hw_queue_mapped(hctx))
3597 blk_mq_tag_idle(hctx);
3599 if (blk_queue_init_done(q))
3600 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3601 set->queue_depth, flush_rq);
3602 if (set->ops->exit_request)
3603 set->ops->exit_request(set, flush_rq, hctx_idx);
3605 if (set->ops->exit_hctx)
3606 set->ops->exit_hctx(hctx, hctx_idx);
3608 blk_mq_remove_cpuhp(hctx);
3610 xa_erase(&q->hctx_table, hctx_idx);
3612 spin_lock(&q->unused_hctx_lock);
3613 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3614 spin_unlock(&q->unused_hctx_lock);
3617 static void blk_mq_exit_hw_queues(struct request_queue *q,
3618 struct blk_mq_tag_set *set, int nr_queue)
3620 struct blk_mq_hw_ctx *hctx;
3623 queue_for_each_hw_ctx(q, hctx, i) {
3626 blk_mq_exit_hctx(q, set, hctx, i);
3630 static int blk_mq_init_hctx(struct request_queue *q,
3631 struct blk_mq_tag_set *set,
3632 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3634 hctx->queue_num = hctx_idx;
3636 if (!(hctx->flags & BLK_MQ_F_STACKING))
3637 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3638 &hctx->cpuhp_online);
3639 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3641 hctx->tags = set->tags[hctx_idx];
3643 if (set->ops->init_hctx &&
3644 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3645 goto unregister_cpu_notifier;
3647 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3651 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3657 if (set->ops->exit_request)
3658 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3660 if (set->ops->exit_hctx)
3661 set->ops->exit_hctx(hctx, hctx_idx);
3662 unregister_cpu_notifier:
3663 blk_mq_remove_cpuhp(hctx);
3667 static struct blk_mq_hw_ctx *
3668 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3671 struct blk_mq_hw_ctx *hctx;
3672 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3674 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3676 goto fail_alloc_hctx;
3678 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3681 atomic_set(&hctx->nr_active, 0);
3682 if (node == NUMA_NO_NODE)
3683 node = set->numa_node;
3684 hctx->numa_node = node;
3686 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3687 spin_lock_init(&hctx->lock);
3688 INIT_LIST_HEAD(&hctx->dispatch);
3690 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3692 INIT_LIST_HEAD(&hctx->hctx_list);
3695 * Allocate space for all possible cpus to avoid allocation at
3698 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3703 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3704 gfp, node, false, false))
3708 spin_lock_init(&hctx->dispatch_wait_lock);
3709 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3710 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3712 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3716 blk_mq_hctx_kobj_init(hctx);
3721 sbitmap_free(&hctx->ctx_map);
3725 free_cpumask_var(hctx->cpumask);
3732 static void blk_mq_init_cpu_queues(struct request_queue *q,
3733 unsigned int nr_hw_queues)
3735 struct blk_mq_tag_set *set = q->tag_set;
3738 for_each_possible_cpu(i) {
3739 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3740 struct blk_mq_hw_ctx *hctx;
3744 spin_lock_init(&__ctx->lock);
3745 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3746 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3751 * Set local node, IFF we have more than one hw queue. If
3752 * not, we remain on the home node of the device
3754 for (j = 0; j < set->nr_maps; j++) {
3755 hctx = blk_mq_map_queue_type(q, j, i);
3756 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3757 hctx->numa_node = cpu_to_node(i);
3762 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3763 unsigned int hctx_idx,
3766 struct blk_mq_tags *tags;
3769 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3773 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3775 blk_mq_free_rq_map(tags);
3782 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3785 if (blk_mq_is_shared_tags(set->flags)) {
3786 set->tags[hctx_idx] = set->shared_tags;
3791 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3794 return set->tags[hctx_idx];
3797 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3798 struct blk_mq_tags *tags,
3799 unsigned int hctx_idx)
3802 blk_mq_free_rqs(set, tags, hctx_idx);
3803 blk_mq_free_rq_map(tags);
3807 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3808 unsigned int hctx_idx)
3810 if (!blk_mq_is_shared_tags(set->flags))
3811 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3813 set->tags[hctx_idx] = NULL;
3816 static void blk_mq_map_swqueue(struct request_queue *q)
3818 unsigned int j, hctx_idx;
3820 struct blk_mq_hw_ctx *hctx;
3821 struct blk_mq_ctx *ctx;
3822 struct blk_mq_tag_set *set = q->tag_set;
3824 queue_for_each_hw_ctx(q, hctx, i) {
3825 cpumask_clear(hctx->cpumask);
3827 hctx->dispatch_from = NULL;
3831 * Map software to hardware queues.
3833 * If the cpu isn't present, the cpu is mapped to first hctx.
3835 for_each_possible_cpu(i) {
3837 ctx = per_cpu_ptr(q->queue_ctx, i);
3838 for (j = 0; j < set->nr_maps; j++) {
3839 if (!set->map[j].nr_queues) {
3840 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3841 HCTX_TYPE_DEFAULT, i);
3844 hctx_idx = set->map[j].mq_map[i];
3845 /* unmapped hw queue can be remapped after CPU topo changed */
3846 if (!set->tags[hctx_idx] &&
3847 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3849 * If tags initialization fail for some hctx,
3850 * that hctx won't be brought online. In this
3851 * case, remap the current ctx to hctx[0] which
3852 * is guaranteed to always have tags allocated
3854 set->map[j].mq_map[i] = 0;
3857 hctx = blk_mq_map_queue_type(q, j, i);
3858 ctx->hctxs[j] = hctx;
3860 * If the CPU is already set in the mask, then we've
3861 * mapped this one already. This can happen if
3862 * devices share queues across queue maps.
3864 if (cpumask_test_cpu(i, hctx->cpumask))
3867 cpumask_set_cpu(i, hctx->cpumask);
3869 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3870 hctx->ctxs[hctx->nr_ctx++] = ctx;
3873 * If the nr_ctx type overflows, we have exceeded the
3874 * amount of sw queues we can support.
3876 BUG_ON(!hctx->nr_ctx);
3879 for (; j < HCTX_MAX_TYPES; j++)
3880 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3881 HCTX_TYPE_DEFAULT, i);
3884 queue_for_each_hw_ctx(q, hctx, i) {
3886 * If no software queues are mapped to this hardware queue,
3887 * disable it and free the request entries.
3889 if (!hctx->nr_ctx) {
3890 /* Never unmap queue 0. We need it as a
3891 * fallback in case of a new remap fails
3895 __blk_mq_free_map_and_rqs(set, i);
3901 hctx->tags = set->tags[i];
3902 WARN_ON(!hctx->tags);
3905 * Set the map size to the number of mapped software queues.
3906 * This is more accurate and more efficient than looping
3907 * over all possibly mapped software queues.
3909 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3912 * Initialize batch roundrobin counts
3914 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3915 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3920 * Caller needs to ensure that we're either frozen/quiesced, or that
3921 * the queue isn't live yet.
3923 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3925 struct blk_mq_hw_ctx *hctx;
3928 queue_for_each_hw_ctx(q, hctx, i) {
3930 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3932 blk_mq_tag_idle(hctx);
3933 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3938 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3941 struct request_queue *q;
3943 lockdep_assert_held(&set->tag_list_lock);
3945 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3946 blk_mq_freeze_queue(q);
3947 queue_set_hctx_shared(q, shared);
3948 blk_mq_unfreeze_queue(q);
3952 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3954 struct blk_mq_tag_set *set = q->tag_set;
3956 mutex_lock(&set->tag_list_lock);
3957 list_del(&q->tag_set_list);
3958 if (list_is_singular(&set->tag_list)) {
3959 /* just transitioned to unshared */
3960 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3961 /* update existing queue */
3962 blk_mq_update_tag_set_shared(set, false);
3964 mutex_unlock(&set->tag_list_lock);
3965 INIT_LIST_HEAD(&q->tag_set_list);
3968 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3969 struct request_queue *q)
3971 mutex_lock(&set->tag_list_lock);
3974 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3976 if (!list_empty(&set->tag_list) &&
3977 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3978 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3979 /* update existing queue */
3980 blk_mq_update_tag_set_shared(set, true);
3982 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3983 queue_set_hctx_shared(q, true);
3984 list_add_tail(&q->tag_set_list, &set->tag_list);
3986 mutex_unlock(&set->tag_list_lock);
3989 /* All allocations will be freed in release handler of q->mq_kobj */
3990 static int blk_mq_alloc_ctxs(struct request_queue *q)
3992 struct blk_mq_ctxs *ctxs;
3995 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3999 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
4000 if (!ctxs->queue_ctx)
4003 for_each_possible_cpu(cpu) {
4004 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4008 q->mq_kobj = &ctxs->kobj;
4009 q->queue_ctx = ctxs->queue_ctx;
4018 * It is the actual release handler for mq, but we do it from
4019 * request queue's release handler for avoiding use-after-free
4020 * and headache because q->mq_kobj shouldn't have been introduced,
4021 * but we can't group ctx/kctx kobj without it.
4023 void blk_mq_release(struct request_queue *q)
4025 struct blk_mq_hw_ctx *hctx, *next;
4028 queue_for_each_hw_ctx(q, hctx, i)
4029 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4031 /* all hctx are in .unused_hctx_list now */
4032 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4033 list_del_init(&hctx->hctx_list);
4034 kobject_put(&hctx->kobj);
4037 xa_destroy(&q->hctx_table);
4040 * release .mq_kobj and sw queue's kobject now because
4041 * both share lifetime with request queue.
4043 blk_mq_sysfs_deinit(q);
4046 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4049 struct request_queue *q;
4052 q = blk_alloc_queue(set->numa_node);
4054 return ERR_PTR(-ENOMEM);
4055 q->queuedata = queuedata;
4056 ret = blk_mq_init_allocated_queue(set, q);
4059 return ERR_PTR(ret);
4064 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4066 return blk_mq_init_queue_data(set, NULL);
4068 EXPORT_SYMBOL(blk_mq_init_queue);
4071 * blk_mq_destroy_queue - shutdown a request queue
4072 * @q: request queue to shutdown
4074 * This shuts down a request queue allocated by blk_mq_init_queue(). All future
4075 * requests will be failed with -ENODEV. The caller is responsible for dropping
4076 * the reference from blk_mq_init_queue() by calling blk_put_queue().
4078 * Context: can sleep
4080 void blk_mq_destroy_queue(struct request_queue *q)
4082 WARN_ON_ONCE(!queue_is_mq(q));
4083 WARN_ON_ONCE(blk_queue_registered(q));
4087 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4088 blk_queue_start_drain(q);
4089 blk_mq_freeze_queue_wait(q);
4092 blk_mq_cancel_work_sync(q);
4093 blk_mq_exit_queue(q);
4095 EXPORT_SYMBOL(blk_mq_destroy_queue);
4097 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4098 struct lock_class_key *lkclass)
4100 struct request_queue *q;
4101 struct gendisk *disk;
4103 q = blk_mq_init_queue_data(set, queuedata);
4107 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4109 blk_mq_destroy_queue(q);
4111 return ERR_PTR(-ENOMEM);
4113 set_bit(GD_OWNS_QUEUE, &disk->state);
4116 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4118 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4119 struct lock_class_key *lkclass)
4121 struct gendisk *disk;
4123 if (!blk_get_queue(q))
4125 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4130 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4132 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4133 struct blk_mq_tag_set *set, struct request_queue *q,
4134 int hctx_idx, int node)
4136 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4138 /* reuse dead hctx first */
4139 spin_lock(&q->unused_hctx_lock);
4140 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4141 if (tmp->numa_node == node) {
4147 list_del_init(&hctx->hctx_list);
4148 spin_unlock(&q->unused_hctx_lock);
4151 hctx = blk_mq_alloc_hctx(q, set, node);
4155 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4161 kobject_put(&hctx->kobj);
4166 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4167 struct request_queue *q)
4169 struct blk_mq_hw_ctx *hctx;
4172 /* protect against switching io scheduler */
4173 mutex_lock(&q->sysfs_lock);
4174 for (i = 0; i < set->nr_hw_queues; i++) {
4176 int node = blk_mq_get_hctx_node(set, i);
4177 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4180 old_node = old_hctx->numa_node;
4181 blk_mq_exit_hctx(q, set, old_hctx, i);
4184 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4187 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4189 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4190 WARN_ON_ONCE(!hctx);
4194 * Increasing nr_hw_queues fails. Free the newly allocated
4195 * hctxs and keep the previous q->nr_hw_queues.
4197 if (i != set->nr_hw_queues) {
4198 j = q->nr_hw_queues;
4201 q->nr_hw_queues = set->nr_hw_queues;
4204 xa_for_each_start(&q->hctx_table, j, hctx, j)
4205 blk_mq_exit_hctx(q, set, hctx, j);
4206 mutex_unlock(&q->sysfs_lock);
4209 static void blk_mq_update_poll_flag(struct request_queue *q)
4211 struct blk_mq_tag_set *set = q->tag_set;
4213 if (set->nr_maps > HCTX_TYPE_POLL &&
4214 set->map[HCTX_TYPE_POLL].nr_queues)
4215 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4217 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4220 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4221 struct request_queue *q)
4223 /* mark the queue as mq asap */
4224 q->mq_ops = set->ops;
4226 if (blk_mq_alloc_ctxs(q))
4229 /* init q->mq_kobj and sw queues' kobjects */
4230 blk_mq_sysfs_init(q);
4232 INIT_LIST_HEAD(&q->unused_hctx_list);
4233 spin_lock_init(&q->unused_hctx_lock);
4235 xa_init(&q->hctx_table);
4237 blk_mq_realloc_hw_ctxs(set, q);
4238 if (!q->nr_hw_queues)
4241 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4242 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4246 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4247 blk_mq_update_poll_flag(q);
4249 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4250 INIT_LIST_HEAD(&q->flush_list);
4251 INIT_LIST_HEAD(&q->requeue_list);
4252 spin_lock_init(&q->requeue_lock);
4254 q->nr_requests = set->queue_depth;
4256 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4257 blk_mq_add_queue_tag_set(set, q);
4258 blk_mq_map_swqueue(q);
4267 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4269 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4270 void blk_mq_exit_queue(struct request_queue *q)
4272 struct blk_mq_tag_set *set = q->tag_set;
4274 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4275 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4276 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4277 blk_mq_del_queue_tag_set(q);
4280 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4284 if (blk_mq_is_shared_tags(set->flags)) {
4285 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4288 if (!set->shared_tags)
4292 for (i = 0; i < set->nr_hw_queues; i++) {
4293 if (!__blk_mq_alloc_map_and_rqs(set, i))
4302 __blk_mq_free_map_and_rqs(set, i);
4304 if (blk_mq_is_shared_tags(set->flags)) {
4305 blk_mq_free_map_and_rqs(set, set->shared_tags,
4306 BLK_MQ_NO_HCTX_IDX);
4313 * Allocate the request maps associated with this tag_set. Note that this
4314 * may reduce the depth asked for, if memory is tight. set->queue_depth
4315 * will be updated to reflect the allocated depth.
4317 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4322 depth = set->queue_depth;
4324 err = __blk_mq_alloc_rq_maps(set);
4328 set->queue_depth >>= 1;
4329 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4333 } while (set->queue_depth);
4335 if (!set->queue_depth || err) {
4336 pr_err("blk-mq: failed to allocate request map\n");
4340 if (depth != set->queue_depth)
4341 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4342 depth, set->queue_depth);
4347 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4350 * blk_mq_map_queues() and multiple .map_queues() implementations
4351 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4352 * number of hardware queues.
4354 if (set->nr_maps == 1)
4355 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4357 if (set->ops->map_queues && !is_kdump_kernel()) {
4361 * transport .map_queues is usually done in the following
4364 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4365 * mask = get_cpu_mask(queue)
4366 * for_each_cpu(cpu, mask)
4367 * set->map[x].mq_map[cpu] = queue;
4370 * When we need to remap, the table has to be cleared for
4371 * killing stale mapping since one CPU may not be mapped
4374 for (i = 0; i < set->nr_maps; i++)
4375 blk_mq_clear_mq_map(&set->map[i]);
4377 set->ops->map_queues(set);
4379 BUG_ON(set->nr_maps > 1);
4380 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4384 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4385 int new_nr_hw_queues)
4387 struct blk_mq_tags **new_tags;
4390 if (set->nr_hw_queues >= new_nr_hw_queues)
4393 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4394 GFP_KERNEL, set->numa_node);
4399 memcpy(new_tags, set->tags, set->nr_hw_queues *
4400 sizeof(*set->tags));
4402 set->tags = new_tags;
4404 for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) {
4405 if (!__blk_mq_alloc_map_and_rqs(set, i)) {
4406 while (--i >= set->nr_hw_queues)
4407 __blk_mq_free_map_and_rqs(set, i);
4414 set->nr_hw_queues = new_nr_hw_queues;
4419 * Alloc a tag set to be associated with one or more request queues.
4420 * May fail with EINVAL for various error conditions. May adjust the
4421 * requested depth down, if it's too large. In that case, the set
4422 * value will be stored in set->queue_depth.
4424 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4428 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4430 if (!set->nr_hw_queues)
4432 if (!set->queue_depth)
4434 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4437 if (!set->ops->queue_rq)
4440 if (!set->ops->get_budget ^ !set->ops->put_budget)
4443 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4444 pr_info("blk-mq: reduced tag depth to %u\n",
4446 set->queue_depth = BLK_MQ_MAX_DEPTH;
4451 else if (set->nr_maps > HCTX_MAX_TYPES)
4455 * If a crashdump is active, then we are potentially in a very
4456 * memory constrained environment. Limit us to 1 queue and
4457 * 64 tags to prevent using too much memory.
4459 if (is_kdump_kernel()) {
4460 set->nr_hw_queues = 1;
4462 set->queue_depth = min(64U, set->queue_depth);
4465 * There is no use for more h/w queues than cpus if we just have
4468 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4469 set->nr_hw_queues = nr_cpu_ids;
4471 if (set->flags & BLK_MQ_F_BLOCKING) {
4472 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4475 ret = init_srcu_struct(set->srcu);
4481 set->tags = kcalloc_node(set->nr_hw_queues,
4482 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4485 goto out_cleanup_srcu;
4487 for (i = 0; i < set->nr_maps; i++) {
4488 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4489 sizeof(set->map[i].mq_map[0]),
4490 GFP_KERNEL, set->numa_node);
4491 if (!set->map[i].mq_map)
4492 goto out_free_mq_map;
4493 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4496 blk_mq_update_queue_map(set);
4498 ret = blk_mq_alloc_set_map_and_rqs(set);
4500 goto out_free_mq_map;
4502 mutex_init(&set->tag_list_lock);
4503 INIT_LIST_HEAD(&set->tag_list);
4508 for (i = 0; i < set->nr_maps; i++) {
4509 kfree(set->map[i].mq_map);
4510 set->map[i].mq_map = NULL;
4515 if (set->flags & BLK_MQ_F_BLOCKING)
4516 cleanup_srcu_struct(set->srcu);
4518 if (set->flags & BLK_MQ_F_BLOCKING)
4522 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4524 /* allocate and initialize a tagset for a simple single-queue device */
4525 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4526 const struct blk_mq_ops *ops, unsigned int queue_depth,
4527 unsigned int set_flags)
4529 memset(set, 0, sizeof(*set));
4531 set->nr_hw_queues = 1;
4533 set->queue_depth = queue_depth;
4534 set->numa_node = NUMA_NO_NODE;
4535 set->flags = set_flags;
4536 return blk_mq_alloc_tag_set(set);
4538 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4540 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4544 for (i = 0; i < set->nr_hw_queues; i++)
4545 __blk_mq_free_map_and_rqs(set, i);
4547 if (blk_mq_is_shared_tags(set->flags)) {
4548 blk_mq_free_map_and_rqs(set, set->shared_tags,
4549 BLK_MQ_NO_HCTX_IDX);
4552 for (j = 0; j < set->nr_maps; j++) {
4553 kfree(set->map[j].mq_map);
4554 set->map[j].mq_map = NULL;
4559 if (set->flags & BLK_MQ_F_BLOCKING) {
4560 cleanup_srcu_struct(set->srcu);
4564 EXPORT_SYMBOL(blk_mq_free_tag_set);
4566 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4568 struct blk_mq_tag_set *set = q->tag_set;
4569 struct blk_mq_hw_ctx *hctx;
4576 if (q->nr_requests == nr)
4579 blk_mq_freeze_queue(q);
4580 blk_mq_quiesce_queue(q);
4583 queue_for_each_hw_ctx(q, hctx, i) {
4587 * If we're using an MQ scheduler, just update the scheduler
4588 * queue depth. This is similar to what the old code would do.
4590 if (hctx->sched_tags) {
4591 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4594 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4599 if (q->elevator && q->elevator->type->ops.depth_updated)
4600 q->elevator->type->ops.depth_updated(hctx);
4603 q->nr_requests = nr;
4604 if (blk_mq_is_shared_tags(set->flags)) {
4606 blk_mq_tag_update_sched_shared_tags(q);
4608 blk_mq_tag_resize_shared_tags(set, nr);
4612 blk_mq_unquiesce_queue(q);
4613 blk_mq_unfreeze_queue(q);
4619 * request_queue and elevator_type pair.
4620 * It is just used by __blk_mq_update_nr_hw_queues to cache
4621 * the elevator_type associated with a request_queue.
4623 struct blk_mq_qe_pair {
4624 struct list_head node;
4625 struct request_queue *q;
4626 struct elevator_type *type;
4630 * Cache the elevator_type in qe pair list and switch the
4631 * io scheduler to 'none'
4633 static bool blk_mq_elv_switch_none(struct list_head *head,
4634 struct request_queue *q)
4636 struct blk_mq_qe_pair *qe;
4638 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4642 /* q->elevator needs protection from ->sysfs_lock */
4643 mutex_lock(&q->sysfs_lock);
4645 /* the check has to be done with holding sysfs_lock */
4651 INIT_LIST_HEAD(&qe->node);
4653 qe->type = q->elevator->type;
4654 /* keep a reference to the elevator module as we'll switch back */
4655 __elevator_get(qe->type);
4656 list_add(&qe->node, head);
4657 elevator_disable(q);
4659 mutex_unlock(&q->sysfs_lock);
4664 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4665 struct request_queue *q)
4667 struct blk_mq_qe_pair *qe;
4669 list_for_each_entry(qe, head, node)
4676 static void blk_mq_elv_switch_back(struct list_head *head,
4677 struct request_queue *q)
4679 struct blk_mq_qe_pair *qe;
4680 struct elevator_type *t;
4682 qe = blk_lookup_qe_pair(head, q);
4686 list_del(&qe->node);
4689 mutex_lock(&q->sysfs_lock);
4690 elevator_switch(q, t);
4691 /* drop the reference acquired in blk_mq_elv_switch_none */
4693 mutex_unlock(&q->sysfs_lock);
4696 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4699 struct request_queue *q;
4701 int prev_nr_hw_queues = set->nr_hw_queues;
4704 lockdep_assert_held(&set->tag_list_lock);
4706 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4707 nr_hw_queues = nr_cpu_ids;
4708 if (nr_hw_queues < 1)
4710 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4713 list_for_each_entry(q, &set->tag_list, tag_set_list)
4714 blk_mq_freeze_queue(q);
4716 * Switch IO scheduler to 'none', cleaning up the data associated
4717 * with the previous scheduler. We will switch back once we are done
4718 * updating the new sw to hw queue mappings.
4720 list_for_each_entry(q, &set->tag_list, tag_set_list)
4721 if (!blk_mq_elv_switch_none(&head, q))
4724 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4725 blk_mq_debugfs_unregister_hctxs(q);
4726 blk_mq_sysfs_unregister_hctxs(q);
4729 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4733 blk_mq_update_queue_map(set);
4734 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4735 blk_mq_realloc_hw_ctxs(set, q);
4736 blk_mq_update_poll_flag(q);
4737 if (q->nr_hw_queues != set->nr_hw_queues) {
4738 int i = prev_nr_hw_queues;
4740 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4741 nr_hw_queues, prev_nr_hw_queues);
4742 for (; i < set->nr_hw_queues; i++)
4743 __blk_mq_free_map_and_rqs(set, i);
4745 set->nr_hw_queues = prev_nr_hw_queues;
4748 blk_mq_map_swqueue(q);
4752 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4753 blk_mq_sysfs_register_hctxs(q);
4754 blk_mq_debugfs_register_hctxs(q);
4758 list_for_each_entry(q, &set->tag_list, tag_set_list)
4759 blk_mq_elv_switch_back(&head, q);
4761 list_for_each_entry(q, &set->tag_list, tag_set_list)
4762 blk_mq_unfreeze_queue(q);
4764 /* Free the excess tags when nr_hw_queues shrink. */
4765 for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)
4766 __blk_mq_free_map_and_rqs(set, i);
4769 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4771 mutex_lock(&set->tag_list_lock);
4772 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4773 mutex_unlock(&set->tag_list_lock);
4775 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4777 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
4778 struct io_comp_batch *iob, unsigned int flags)
4780 long state = get_current_state();
4784 ret = q->mq_ops->poll(hctx, iob);
4786 __set_current_state(TASK_RUNNING);
4790 if (signal_pending_state(state, current))
4791 __set_current_state(TASK_RUNNING);
4792 if (task_is_running(current))
4795 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4798 } while (!need_resched());
4800 __set_current_state(TASK_RUNNING);
4804 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
4805 struct io_comp_batch *iob, unsigned int flags)
4807 struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie);
4809 return blk_hctx_poll(q, hctx, iob, flags);
4812 int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
4813 unsigned int poll_flags)
4815 struct request_queue *q = rq->q;
4818 if (!blk_rq_is_poll(rq))
4820 if (!percpu_ref_tryget(&q->q_usage_counter))
4823 ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags);
4828 EXPORT_SYMBOL_GPL(blk_rq_poll);
4830 unsigned int blk_mq_rq_cpu(struct request *rq)
4832 return rq->mq_ctx->cpu;
4834 EXPORT_SYMBOL(blk_mq_rq_cpu);
4836 void blk_mq_cancel_work_sync(struct request_queue *q)
4838 struct blk_mq_hw_ctx *hctx;
4841 cancel_delayed_work_sync(&q->requeue_work);
4843 queue_for_each_hw_ctx(q, hctx, i)
4844 cancel_delayed_work_sync(&hctx->run_work);
4847 static int __init blk_mq_init(void)
4851 for_each_possible_cpu(i)
4852 init_llist_head(&per_cpu(blk_cpu_done, i));
4853 for_each_possible_cpu(i)
4854 INIT_CSD(&per_cpu(blk_cpu_csd, i),
4855 __blk_mq_complete_request_remote, NULL);
4856 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4858 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4859 "block/softirq:dead", NULL,
4860 blk_softirq_cpu_dead);
4861 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4862 blk_mq_hctx_notify_dead);
4863 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4864 blk_mq_hctx_notify_online,
4865 blk_mq_hctx_notify_offline);
4868 subsys_initcall(blk_mq_init);