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);
47 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags);
48 static void blk_mq_request_bypass_insert(struct request *rq,
50 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
51 struct list_head *list);
52 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
53 struct io_comp_batch *iob, unsigned int flags);
56 * Check if any of the ctx, dispatch list or elevator
57 * have pending work in this hardware queue.
59 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
61 return !list_empty_careful(&hctx->dispatch) ||
62 sbitmap_any_bit_set(&hctx->ctx_map) ||
63 blk_mq_sched_has_work(hctx);
67 * Mark this ctx as having pending work in this hardware queue
69 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
70 struct blk_mq_ctx *ctx)
72 const int bit = ctx->index_hw[hctx->type];
74 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
75 sbitmap_set_bit(&hctx->ctx_map, bit);
78 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
79 struct blk_mq_ctx *ctx)
81 const int bit = ctx->index_hw[hctx->type];
83 sbitmap_clear_bit(&hctx->ctx_map, bit);
87 struct block_device *part;
88 unsigned int inflight[2];
91 static bool blk_mq_check_inflight(struct request *rq, void *priv)
93 struct mq_inflight *mi = priv;
95 if (rq->part && blk_do_io_stat(rq) &&
96 (!mi->part->bd_partno || rq->part == mi->part) &&
97 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
98 mi->inflight[rq_data_dir(rq)]++;
103 unsigned int blk_mq_in_flight(struct request_queue *q,
104 struct block_device *part)
106 struct mq_inflight mi = { .part = part };
108 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
110 return mi.inflight[0] + mi.inflight[1];
113 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
114 unsigned int inflight[2])
116 struct mq_inflight mi = { .part = part };
118 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
119 inflight[0] = mi.inflight[0];
120 inflight[1] = mi.inflight[1];
123 void blk_freeze_queue_start(struct request_queue *q)
125 mutex_lock(&q->mq_freeze_lock);
126 if (++q->mq_freeze_depth == 1) {
127 percpu_ref_kill(&q->q_usage_counter);
128 mutex_unlock(&q->mq_freeze_lock);
130 blk_mq_run_hw_queues(q, false);
132 mutex_unlock(&q->mq_freeze_lock);
135 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
137 void blk_mq_freeze_queue_wait(struct request_queue *q)
139 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
141 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
143 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
144 unsigned long timeout)
146 return wait_event_timeout(q->mq_freeze_wq,
147 percpu_ref_is_zero(&q->q_usage_counter),
150 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
153 * Guarantee no request is in use, so we can change any data structure of
154 * the queue afterward.
156 void blk_freeze_queue(struct request_queue *q)
159 * In the !blk_mq case we are only calling this to kill the
160 * q_usage_counter, otherwise this increases the freeze depth
161 * and waits for it to return to zero. For this reason there is
162 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
163 * exported to drivers as the only user for unfreeze is blk_mq.
165 blk_freeze_queue_start(q);
166 blk_mq_freeze_queue_wait(q);
169 void blk_mq_freeze_queue(struct request_queue *q)
172 * ...just an alias to keep freeze and unfreeze actions balanced
173 * in the blk_mq_* namespace
177 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
179 void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
181 mutex_lock(&q->mq_freeze_lock);
183 q->q_usage_counter.data->force_atomic = true;
184 q->mq_freeze_depth--;
185 WARN_ON_ONCE(q->mq_freeze_depth < 0);
186 if (!q->mq_freeze_depth) {
187 percpu_ref_resurrect(&q->q_usage_counter);
188 wake_up_all(&q->mq_freeze_wq);
190 mutex_unlock(&q->mq_freeze_lock);
193 void blk_mq_unfreeze_queue(struct request_queue *q)
195 __blk_mq_unfreeze_queue(q, false);
197 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
200 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
201 * mpt3sas driver such that this function can be removed.
203 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
207 spin_lock_irqsave(&q->queue_lock, flags);
208 if (!q->quiesce_depth++)
209 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
210 spin_unlock_irqrestore(&q->queue_lock, flags);
212 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
215 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
216 * @set: tag_set to wait on
218 * Note: it is driver's responsibility for making sure that quiesce has
219 * been started on or more of the request_queues of the tag_set. This
220 * function only waits for the quiesce on those request_queues that had
221 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
223 void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
225 if (set->flags & BLK_MQ_F_BLOCKING)
226 synchronize_srcu(set->srcu);
230 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
233 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
236 * Note: this function does not prevent that the struct request end_io()
237 * callback function is invoked. Once this function is returned, we make
238 * sure no dispatch can happen until the queue is unquiesced via
239 * blk_mq_unquiesce_queue().
241 void blk_mq_quiesce_queue(struct request_queue *q)
243 blk_mq_quiesce_queue_nowait(q);
244 /* nothing to wait for non-mq queues */
246 blk_mq_wait_quiesce_done(q->tag_set);
248 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
251 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
254 * This function recovers queue into the state before quiescing
255 * which is done by blk_mq_quiesce_queue.
257 void blk_mq_unquiesce_queue(struct request_queue *q)
260 bool run_queue = false;
262 spin_lock_irqsave(&q->queue_lock, flags);
263 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
265 } else if (!--q->quiesce_depth) {
266 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
269 spin_unlock_irqrestore(&q->queue_lock, flags);
271 /* dispatch requests which are inserted during quiescing */
273 blk_mq_run_hw_queues(q, true);
275 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
277 void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
279 struct request_queue *q;
281 mutex_lock(&set->tag_list_lock);
282 list_for_each_entry(q, &set->tag_list, tag_set_list) {
283 if (!blk_queue_skip_tagset_quiesce(q))
284 blk_mq_quiesce_queue_nowait(q);
286 blk_mq_wait_quiesce_done(set);
287 mutex_unlock(&set->tag_list_lock);
289 EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
291 void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
293 struct request_queue *q;
295 mutex_lock(&set->tag_list_lock);
296 list_for_each_entry(q, &set->tag_list, tag_set_list) {
297 if (!blk_queue_skip_tagset_quiesce(q))
298 blk_mq_unquiesce_queue(q);
300 mutex_unlock(&set->tag_list_lock);
302 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
304 void blk_mq_wake_waiters(struct request_queue *q)
306 struct blk_mq_hw_ctx *hctx;
309 queue_for_each_hw_ctx(q, hctx, i)
310 if (blk_mq_hw_queue_mapped(hctx))
311 blk_mq_tag_wakeup_all(hctx->tags, true);
314 void blk_rq_init(struct request_queue *q, struct request *rq)
316 memset(rq, 0, sizeof(*rq));
318 INIT_LIST_HEAD(&rq->queuelist);
320 rq->__sector = (sector_t) -1;
321 INIT_HLIST_NODE(&rq->hash);
322 RB_CLEAR_NODE(&rq->rb_node);
323 rq->tag = BLK_MQ_NO_TAG;
324 rq->internal_tag = BLK_MQ_NO_TAG;
325 rq->start_time_ns = ktime_get_ns();
327 blk_crypto_rq_set_defaults(rq);
329 EXPORT_SYMBOL(blk_rq_init);
331 /* Set start and alloc time when the allocated request is actually used */
332 static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns)
334 if (blk_mq_need_time_stamp(rq))
335 rq->start_time_ns = ktime_get_ns();
337 rq->start_time_ns = 0;
339 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
340 if (blk_queue_rq_alloc_time(rq->q))
341 rq->alloc_time_ns = alloc_time_ns ?: rq->start_time_ns;
343 rq->alloc_time_ns = 0;
347 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
348 struct blk_mq_tags *tags, unsigned int tag)
350 struct blk_mq_ctx *ctx = data->ctx;
351 struct blk_mq_hw_ctx *hctx = data->hctx;
352 struct request_queue *q = data->q;
353 struct request *rq = tags->static_rqs[tag];
358 rq->cmd_flags = data->cmd_flags;
360 if (data->flags & BLK_MQ_REQ_PM)
361 data->rq_flags |= RQF_PM;
362 if (blk_queue_io_stat(q))
363 data->rq_flags |= RQF_IO_STAT;
364 rq->rq_flags = data->rq_flags;
366 if (data->rq_flags & RQF_SCHED_TAGS) {
367 rq->tag = BLK_MQ_NO_TAG;
368 rq->internal_tag = tag;
371 rq->internal_tag = BLK_MQ_NO_TAG;
376 rq->io_start_time_ns = 0;
377 rq->stats_sectors = 0;
378 rq->nr_phys_segments = 0;
379 #if defined(CONFIG_BLK_DEV_INTEGRITY)
380 rq->nr_integrity_segments = 0;
383 rq->end_io_data = NULL;
385 blk_crypto_rq_set_defaults(rq);
386 INIT_LIST_HEAD(&rq->queuelist);
387 /* tag was already set */
388 WRITE_ONCE(rq->deadline, 0);
391 if (rq->rq_flags & RQF_USE_SCHED) {
392 struct elevator_queue *e = data->q->elevator;
394 INIT_HLIST_NODE(&rq->hash);
395 RB_CLEAR_NODE(&rq->rb_node);
397 if (e->type->ops.prepare_request)
398 e->type->ops.prepare_request(rq);
404 static inline struct request *
405 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data)
407 unsigned int tag, tag_offset;
408 struct blk_mq_tags *tags;
410 unsigned long tag_mask;
413 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
414 if (unlikely(!tag_mask))
417 tags = blk_mq_tags_from_data(data);
418 for (i = 0; tag_mask; i++) {
419 if (!(tag_mask & (1UL << i)))
421 tag = tag_offset + i;
422 prefetch(tags->static_rqs[tag]);
423 tag_mask &= ~(1UL << i);
424 rq = blk_mq_rq_ctx_init(data, tags, tag);
425 rq_list_add(data->cached_rq, rq);
428 /* caller already holds a reference, add for remainder */
429 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
432 return rq_list_pop(data->cached_rq);
435 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
437 struct request_queue *q = data->q;
438 u64 alloc_time_ns = 0;
442 /* alloc_time includes depth and tag waits */
443 if (blk_queue_rq_alloc_time(q))
444 alloc_time_ns = ktime_get_ns();
446 if (data->cmd_flags & REQ_NOWAIT)
447 data->flags |= BLK_MQ_REQ_NOWAIT;
451 * All requests use scheduler tags when an I/O scheduler is
452 * enabled for the queue.
454 data->rq_flags |= RQF_SCHED_TAGS;
457 * Flush/passthrough requests are special and go directly to the
460 if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&
461 !blk_op_is_passthrough(data->cmd_flags)) {
462 struct elevator_mq_ops *ops = &q->elevator->type->ops;
464 WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
466 data->rq_flags |= RQF_USE_SCHED;
467 if (ops->limit_depth)
468 ops->limit_depth(data->cmd_flags, data);
473 data->ctx = blk_mq_get_ctx(q);
474 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
475 if (!(data->rq_flags & RQF_SCHED_TAGS))
476 blk_mq_tag_busy(data->hctx);
478 if (data->flags & BLK_MQ_REQ_RESERVED)
479 data->rq_flags |= RQF_RESV;
482 * Try batched alloc if we want more than 1 tag.
484 if (data->nr_tags > 1) {
485 rq = __blk_mq_alloc_requests_batch(data);
487 blk_mq_rq_time_init(rq, alloc_time_ns);
494 * Waiting allocations only fail because of an inactive hctx. In that
495 * case just retry the hctx assignment and tag allocation as CPU hotplug
496 * should have migrated us to an online CPU by now.
498 tag = blk_mq_get_tag(data);
499 if (tag == BLK_MQ_NO_TAG) {
500 if (data->flags & BLK_MQ_REQ_NOWAIT)
503 * Give up the CPU and sleep for a random short time to
504 * ensure that thread using a realtime scheduling class
505 * are migrated off the CPU, and thus off the hctx that
512 rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag);
513 blk_mq_rq_time_init(rq, alloc_time_ns);
517 static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
518 struct blk_plug *plug,
520 blk_mq_req_flags_t flags)
522 struct blk_mq_alloc_data data = {
526 .nr_tags = plug->nr_ios,
527 .cached_rq = &plug->cached_rq,
531 if (blk_queue_enter(q, flags))
536 rq = __blk_mq_alloc_requests(&data);
542 static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
544 blk_mq_req_flags_t flags)
546 struct blk_plug *plug = current->plug;
552 if (rq_list_empty(plug->cached_rq)) {
553 if (plug->nr_ios == 1)
555 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
559 rq = rq_list_peek(&plug->cached_rq);
560 if (!rq || rq->q != q)
563 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
565 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
568 plug->cached_rq = rq_list_next(rq);
569 blk_mq_rq_time_init(rq, 0);
573 INIT_LIST_HEAD(&rq->queuelist);
577 struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
578 blk_mq_req_flags_t flags)
582 rq = blk_mq_alloc_cached_request(q, opf, flags);
584 struct blk_mq_alloc_data data = {
592 ret = blk_queue_enter(q, flags);
596 rq = __blk_mq_alloc_requests(&data);
601 rq->__sector = (sector_t) -1;
602 rq->bio = rq->biotail = NULL;
606 return ERR_PTR(-EWOULDBLOCK);
608 EXPORT_SYMBOL(blk_mq_alloc_request);
610 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
611 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
613 struct blk_mq_alloc_data data = {
619 u64 alloc_time_ns = 0;
625 /* alloc_time includes depth and tag waits */
626 if (blk_queue_rq_alloc_time(q))
627 alloc_time_ns = ktime_get_ns();
630 * If the tag allocator sleeps we could get an allocation for a
631 * different hardware context. No need to complicate the low level
632 * allocator for this for the rare use case of a command tied to
635 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
636 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
637 return ERR_PTR(-EINVAL);
639 if (hctx_idx >= q->nr_hw_queues)
640 return ERR_PTR(-EIO);
642 ret = blk_queue_enter(q, flags);
647 * Check if the hardware context is actually mapped to anything.
648 * If not tell the caller that it should skip this queue.
651 data.hctx = xa_load(&q->hctx_table, hctx_idx);
652 if (!blk_mq_hw_queue_mapped(data.hctx))
654 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
655 if (cpu >= nr_cpu_ids)
657 data.ctx = __blk_mq_get_ctx(q, cpu);
660 data.rq_flags |= RQF_SCHED_TAGS;
662 blk_mq_tag_busy(data.hctx);
664 if (flags & BLK_MQ_REQ_RESERVED)
665 data.rq_flags |= RQF_RESV;
668 tag = blk_mq_get_tag(&data);
669 if (tag == BLK_MQ_NO_TAG)
671 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag);
672 blk_mq_rq_time_init(rq, alloc_time_ns);
674 rq->__sector = (sector_t) -1;
675 rq->bio = rq->biotail = NULL;
682 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
684 static void blk_mq_finish_request(struct request *rq)
686 struct request_queue *q = rq->q;
688 if (rq->rq_flags & RQF_USE_SCHED) {
689 q->elevator->type->ops.finish_request(rq);
691 * For postflush request that may need to be
692 * completed twice, we should clear this flag
693 * to avoid double finish_request() on the rq.
695 rq->rq_flags &= ~RQF_USE_SCHED;
699 static void __blk_mq_free_request(struct request *rq)
701 struct request_queue *q = rq->q;
702 struct blk_mq_ctx *ctx = rq->mq_ctx;
703 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
704 const int sched_tag = rq->internal_tag;
706 blk_crypto_free_request(rq);
707 blk_pm_mark_last_busy(rq);
710 if (rq->rq_flags & RQF_MQ_INFLIGHT)
711 __blk_mq_dec_active_requests(hctx);
713 if (rq->tag != BLK_MQ_NO_TAG)
714 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
715 if (sched_tag != BLK_MQ_NO_TAG)
716 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
717 blk_mq_sched_restart(hctx);
721 void blk_mq_free_request(struct request *rq)
723 struct request_queue *q = rq->q;
725 blk_mq_finish_request(rq);
727 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
728 laptop_io_completion(q->disk->bdi);
732 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
733 if (req_ref_put_and_test(rq))
734 __blk_mq_free_request(rq);
736 EXPORT_SYMBOL_GPL(blk_mq_free_request);
738 void blk_mq_free_plug_rqs(struct blk_plug *plug)
742 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
743 blk_mq_free_request(rq);
746 void blk_dump_rq_flags(struct request *rq, char *msg)
748 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
749 rq->q->disk ? rq->q->disk->disk_name : "?",
750 (__force unsigned long long) rq->cmd_flags);
752 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
753 (unsigned long long)blk_rq_pos(rq),
754 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
755 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
756 rq->bio, rq->biotail, blk_rq_bytes(rq));
758 EXPORT_SYMBOL(blk_dump_rq_flags);
760 static void req_bio_endio(struct request *rq, struct bio *bio,
761 unsigned int nbytes, blk_status_t error)
763 if (unlikely(error)) {
764 bio->bi_status = error;
765 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
767 * Partial zone append completions cannot be supported as the
768 * BIO fragments may end up not being written sequentially.
770 if (bio->bi_iter.bi_size != nbytes)
771 bio->bi_status = BLK_STS_IOERR;
773 bio->bi_iter.bi_sector = rq->__sector;
776 bio_advance(bio, nbytes);
778 if (unlikely(rq->rq_flags & RQF_QUIET))
779 bio_set_flag(bio, BIO_QUIET);
780 /* don't actually finish bio if it's part of flush sequence */
781 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
785 static void blk_account_io_completion(struct request *req, unsigned int bytes)
787 if (req->part && blk_do_io_stat(req)) {
788 const int sgrp = op_stat_group(req_op(req));
791 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
796 static void blk_print_req_error(struct request *req, blk_status_t status)
798 printk_ratelimited(KERN_ERR
799 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
800 "phys_seg %u prio class %u\n",
801 blk_status_to_str(status),
802 req->q->disk ? req->q->disk->disk_name : "?",
803 blk_rq_pos(req), (__force u32)req_op(req),
804 blk_op_str(req_op(req)),
805 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
806 req->nr_phys_segments,
807 IOPRIO_PRIO_CLASS(req->ioprio));
811 * Fully end IO on a request. Does not support partial completions, or
814 static void blk_complete_request(struct request *req)
816 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
817 int total_bytes = blk_rq_bytes(req);
818 struct bio *bio = req->bio;
820 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
825 #ifdef CONFIG_BLK_DEV_INTEGRITY
826 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
827 req->q->integrity.profile->complete_fn(req, total_bytes);
831 * Upper layers may call blk_crypto_evict_key() anytime after the last
832 * bio_endio(). Therefore, the keyslot must be released before that.
834 blk_crypto_rq_put_keyslot(req);
836 blk_account_io_completion(req, total_bytes);
839 struct bio *next = bio->bi_next;
841 /* Completion has already been traced */
842 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
844 if (req_op(req) == REQ_OP_ZONE_APPEND)
845 bio->bi_iter.bi_sector = req->__sector;
853 * Reset counters so that the request stacking driver
854 * can find how many bytes remain in the request
864 * blk_update_request - Complete multiple bytes without completing the request
865 * @req: the request being processed
866 * @error: block status code
867 * @nr_bytes: number of bytes to complete for @req
870 * Ends I/O on a number of bytes attached to @req, but doesn't complete
871 * the request structure even if @req doesn't have leftover.
872 * If @req has leftover, sets it up for the next range of segments.
874 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
875 * %false return from this function.
878 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
879 * except in the consistency check at the end of this function.
882 * %false - this request doesn't have any more data
883 * %true - this request has more data
885 bool blk_update_request(struct request *req, blk_status_t error,
886 unsigned int nr_bytes)
890 trace_block_rq_complete(req, error, nr_bytes);
895 #ifdef CONFIG_BLK_DEV_INTEGRITY
896 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
898 req->q->integrity.profile->complete_fn(req, nr_bytes);
902 * Upper layers may call blk_crypto_evict_key() anytime after the last
903 * bio_endio(). Therefore, the keyslot must be released before that.
905 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
906 __blk_crypto_rq_put_keyslot(req);
908 if (unlikely(error && !blk_rq_is_passthrough(req) &&
909 !(req->rq_flags & RQF_QUIET)) &&
910 !test_bit(GD_DEAD, &req->q->disk->state)) {
911 blk_print_req_error(req, error);
912 trace_block_rq_error(req, error, nr_bytes);
915 blk_account_io_completion(req, nr_bytes);
919 struct bio *bio = req->bio;
920 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
922 if (bio_bytes == bio->bi_iter.bi_size)
923 req->bio = bio->bi_next;
925 /* Completion has already been traced */
926 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
927 req_bio_endio(req, bio, bio_bytes, error);
929 total_bytes += bio_bytes;
930 nr_bytes -= bio_bytes;
941 * Reset counters so that the request stacking driver
942 * can find how many bytes remain in the request
949 req->__data_len -= total_bytes;
951 /* update sector only for requests with clear definition of sector */
952 if (!blk_rq_is_passthrough(req))
953 req->__sector += total_bytes >> 9;
955 /* mixed attributes always follow the first bio */
956 if (req->rq_flags & RQF_MIXED_MERGE) {
957 req->cmd_flags &= ~REQ_FAILFAST_MASK;
958 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
961 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
963 * If total number of sectors is less than the first segment
964 * size, something has gone terribly wrong.
966 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
967 blk_dump_rq_flags(req, "request botched");
968 req->__data_len = blk_rq_cur_bytes(req);
971 /* recalculate the number of segments */
972 req->nr_phys_segments = blk_recalc_rq_segments(req);
977 EXPORT_SYMBOL_GPL(blk_update_request);
979 static inline void blk_account_io_done(struct request *req, u64 now)
981 trace_block_io_done(req);
984 * Account IO completion. flush_rq isn't accounted as a
985 * normal IO on queueing nor completion. Accounting the
986 * containing request is enough.
988 if (blk_do_io_stat(req) && req->part &&
989 !(req->rq_flags & RQF_FLUSH_SEQ)) {
990 const int sgrp = op_stat_group(req_op(req));
993 update_io_ticks(req->part, jiffies, true);
994 part_stat_inc(req->part, ios[sgrp]);
995 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
1000 static inline void blk_account_io_start(struct request *req)
1002 trace_block_io_start(req);
1004 if (blk_do_io_stat(req)) {
1006 * All non-passthrough requests are created from a bio with one
1007 * exception: when a flush command that is part of a flush sequence
1008 * generated by the state machine in blk-flush.c is cloned onto the
1009 * lower device by dm-multipath we can get here without a bio.
1012 req->part = req->bio->bi_bdev;
1014 req->part = req->q->disk->part0;
1017 update_io_ticks(req->part, jiffies, false);
1022 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1024 if (rq->rq_flags & RQF_STATS)
1025 blk_stat_add(rq, now);
1027 blk_mq_sched_completed_request(rq, now);
1028 blk_account_io_done(rq, now);
1031 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1033 if (blk_mq_need_time_stamp(rq))
1034 __blk_mq_end_request_acct(rq, ktime_get_ns());
1036 blk_mq_finish_request(rq);
1039 rq_qos_done(rq->q, rq);
1040 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1041 blk_mq_free_request(rq);
1043 blk_mq_free_request(rq);
1046 EXPORT_SYMBOL(__blk_mq_end_request);
1048 void blk_mq_end_request(struct request *rq, blk_status_t error)
1050 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1052 __blk_mq_end_request(rq, error);
1054 EXPORT_SYMBOL(blk_mq_end_request);
1056 #define TAG_COMP_BATCH 32
1058 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1059 int *tag_array, int nr_tags)
1061 struct request_queue *q = hctx->queue;
1064 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1065 * update hctx->nr_active in batch
1067 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1068 __blk_mq_sub_active_requests(hctx, nr_tags);
1070 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1071 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1074 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1076 int tags[TAG_COMP_BATCH], nr_tags = 0;
1077 struct blk_mq_hw_ctx *cur_hctx = NULL;
1082 now = ktime_get_ns();
1084 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1086 prefetch(rq->rq_next);
1088 blk_complete_request(rq);
1090 __blk_mq_end_request_acct(rq, now);
1092 blk_mq_finish_request(rq);
1094 rq_qos_done(rq->q, rq);
1097 * If end_io handler returns NONE, then it still has
1098 * ownership of the request.
1100 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1103 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1104 if (!req_ref_put_and_test(rq))
1107 blk_crypto_free_request(rq);
1108 blk_pm_mark_last_busy(rq);
1110 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1112 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1114 cur_hctx = rq->mq_hctx;
1116 tags[nr_tags++] = rq->tag;
1120 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1122 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1124 static void blk_complete_reqs(struct llist_head *list)
1126 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1127 struct request *rq, *next;
1129 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1130 rq->q->mq_ops->complete(rq);
1133 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1135 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1138 static int blk_softirq_cpu_dead(unsigned int cpu)
1140 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1144 static void __blk_mq_complete_request_remote(void *data)
1146 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1149 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1151 int cpu = raw_smp_processor_id();
1153 if (!IS_ENABLED(CONFIG_SMP) ||
1154 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1157 * With force threaded interrupts enabled, raising softirq from an SMP
1158 * function call will always result in waking the ksoftirqd thread.
1159 * This is probably worse than completing the request on a different
1162 if (force_irqthreads())
1165 /* same CPU or cache domain? Complete locally */
1166 if (cpu == rq->mq_ctx->cpu ||
1167 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1168 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1171 /* don't try to IPI to an offline CPU */
1172 return cpu_online(rq->mq_ctx->cpu);
1175 static void blk_mq_complete_send_ipi(struct request *rq)
1177 struct llist_head *list;
1180 cpu = rq->mq_ctx->cpu;
1181 list = &per_cpu(blk_cpu_done, cpu);
1182 if (llist_add(&rq->ipi_list, list)) {
1183 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1184 smp_call_function_single_async(cpu, &rq->csd);
1188 static void blk_mq_raise_softirq(struct request *rq)
1190 struct llist_head *list;
1193 list = this_cpu_ptr(&blk_cpu_done);
1194 if (llist_add(&rq->ipi_list, list))
1195 raise_softirq(BLOCK_SOFTIRQ);
1199 bool blk_mq_complete_request_remote(struct request *rq)
1201 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1204 * For request which hctx has only one ctx mapping,
1205 * or a polled request, always complete locally,
1206 * it's pointless to redirect the completion.
1208 if ((rq->mq_hctx->nr_ctx == 1 &&
1209 rq->mq_ctx->cpu == raw_smp_processor_id()) ||
1210 rq->cmd_flags & REQ_POLLED)
1213 if (blk_mq_complete_need_ipi(rq)) {
1214 blk_mq_complete_send_ipi(rq);
1218 if (rq->q->nr_hw_queues == 1) {
1219 blk_mq_raise_softirq(rq);
1224 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1227 * blk_mq_complete_request - end I/O on a request
1228 * @rq: the request being processed
1231 * Complete a request by scheduling the ->complete_rq operation.
1233 void blk_mq_complete_request(struct request *rq)
1235 if (!blk_mq_complete_request_remote(rq))
1236 rq->q->mq_ops->complete(rq);
1238 EXPORT_SYMBOL(blk_mq_complete_request);
1241 * blk_mq_start_request - Start processing a request
1242 * @rq: Pointer to request to be started
1244 * Function used by device drivers to notify the block layer that a request
1245 * is going to be processed now, so blk layer can do proper initializations
1246 * such as starting the timeout timer.
1248 void blk_mq_start_request(struct request *rq)
1250 struct request_queue *q = rq->q;
1252 trace_block_rq_issue(rq);
1254 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1255 rq->io_start_time_ns = ktime_get_ns();
1256 rq->stats_sectors = blk_rq_sectors(rq);
1257 rq->rq_flags |= RQF_STATS;
1258 rq_qos_issue(q, rq);
1261 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1264 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1266 #ifdef CONFIG_BLK_DEV_INTEGRITY
1267 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1268 q->integrity.profile->prepare_fn(rq);
1270 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1271 WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num);
1273 EXPORT_SYMBOL(blk_mq_start_request);
1276 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1277 * queues. This is important for md arrays to benefit from merging
1280 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1282 if (plug->multiple_queues)
1283 return BLK_MAX_REQUEST_COUNT * 2;
1284 return BLK_MAX_REQUEST_COUNT;
1287 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1289 struct request *last = rq_list_peek(&plug->mq_list);
1291 if (!plug->rq_count) {
1292 trace_block_plug(rq->q);
1293 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1294 (!blk_queue_nomerges(rq->q) &&
1295 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1296 blk_mq_flush_plug_list(plug, false);
1298 trace_block_plug(rq->q);
1301 if (!plug->multiple_queues && last && last->q != rq->q)
1302 plug->multiple_queues = true;
1304 * Any request allocated from sched tags can't be issued to
1305 * ->queue_rqs() directly
1307 if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))
1308 plug->has_elevator = true;
1310 rq_list_add(&plug->mq_list, rq);
1315 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1316 * @rq: request to insert
1317 * @at_head: insert request at head or tail of queue
1320 * Insert a fully prepared request at the back of the I/O scheduler queue
1321 * for execution. Don't wait for completion.
1324 * This function will invoke @done directly if the queue is dead.
1326 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1328 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1330 WARN_ON(irqs_disabled());
1331 WARN_ON(!blk_rq_is_passthrough(rq));
1333 blk_account_io_start(rq);
1336 * As plugging can be enabled for passthrough requests on a zoned
1337 * device, directly accessing the plug instead of using blk_mq_plug()
1338 * should not have any consequences.
1340 if (current->plug && !at_head) {
1341 blk_add_rq_to_plug(current->plug, rq);
1345 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1346 blk_mq_run_hw_queue(hctx, false);
1348 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1350 struct blk_rq_wait {
1351 struct completion done;
1355 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1357 struct blk_rq_wait *wait = rq->end_io_data;
1360 complete(&wait->done);
1361 return RQ_END_IO_NONE;
1364 bool blk_rq_is_poll(struct request *rq)
1368 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1372 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1374 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1377 blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0);
1379 } while (!completion_done(wait));
1383 * blk_execute_rq - insert a request into queue for execution
1384 * @rq: request to insert
1385 * @at_head: insert request at head or tail of queue
1388 * Insert a fully prepared request at the back of the I/O scheduler queue
1389 * for execution and wait for completion.
1390 * Return: The blk_status_t result provided to blk_mq_end_request().
1392 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1394 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1395 struct blk_rq_wait wait = {
1396 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1399 WARN_ON(irqs_disabled());
1400 WARN_ON(!blk_rq_is_passthrough(rq));
1402 rq->end_io_data = &wait;
1403 rq->end_io = blk_end_sync_rq;
1405 blk_account_io_start(rq);
1406 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1407 blk_mq_run_hw_queue(hctx, false);
1409 if (blk_rq_is_poll(rq)) {
1410 blk_rq_poll_completion(rq, &wait.done);
1413 * Prevent hang_check timer from firing at us during very long
1416 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1419 while (!wait_for_completion_io_timeout(&wait.done,
1420 hang_check * (HZ/2)))
1423 wait_for_completion_io(&wait.done);
1428 EXPORT_SYMBOL(blk_execute_rq);
1430 static void __blk_mq_requeue_request(struct request *rq)
1432 struct request_queue *q = rq->q;
1434 blk_mq_put_driver_tag(rq);
1436 trace_block_rq_requeue(rq);
1437 rq_qos_requeue(q, rq);
1439 if (blk_mq_request_started(rq)) {
1440 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1441 rq->rq_flags &= ~RQF_TIMED_OUT;
1445 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1447 struct request_queue *q = rq->q;
1448 unsigned long flags;
1450 __blk_mq_requeue_request(rq);
1452 /* this request will be re-inserted to io scheduler queue */
1453 blk_mq_sched_requeue_request(rq);
1455 spin_lock_irqsave(&q->requeue_lock, flags);
1456 list_add_tail(&rq->queuelist, &q->requeue_list);
1457 spin_unlock_irqrestore(&q->requeue_lock, flags);
1459 if (kick_requeue_list)
1460 blk_mq_kick_requeue_list(q);
1462 EXPORT_SYMBOL(blk_mq_requeue_request);
1464 static void blk_mq_requeue_work(struct work_struct *work)
1466 struct request_queue *q =
1467 container_of(work, struct request_queue, requeue_work.work);
1469 LIST_HEAD(flush_list);
1472 spin_lock_irq(&q->requeue_lock);
1473 list_splice_init(&q->requeue_list, &rq_list);
1474 list_splice_init(&q->flush_list, &flush_list);
1475 spin_unlock_irq(&q->requeue_lock);
1477 while (!list_empty(&rq_list)) {
1478 rq = list_entry(rq_list.next, struct request, queuelist);
1480 * If RQF_DONTPREP ist set, the request has been started by the
1481 * driver already and might have driver-specific data allocated
1482 * already. Insert it into the hctx dispatch list to avoid
1483 * block layer merges for the request.
1485 if (rq->rq_flags & RQF_DONTPREP) {
1486 list_del_init(&rq->queuelist);
1487 blk_mq_request_bypass_insert(rq, 0);
1489 list_del_init(&rq->queuelist);
1490 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1494 while (!list_empty(&flush_list)) {
1495 rq = list_entry(flush_list.next, struct request, queuelist);
1496 list_del_init(&rq->queuelist);
1497 blk_mq_insert_request(rq, 0);
1500 blk_mq_run_hw_queues(q, false);
1503 void blk_mq_kick_requeue_list(struct request_queue *q)
1505 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1507 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1509 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1510 unsigned long msecs)
1512 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1513 msecs_to_jiffies(msecs));
1515 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1517 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1520 * If we find a request that isn't idle we know the queue is busy
1521 * as it's checked in the iter.
1522 * Return false to stop the iteration.
1524 if (blk_mq_request_started(rq)) {
1534 bool blk_mq_queue_inflight(struct request_queue *q)
1538 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1541 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1543 static void blk_mq_rq_timed_out(struct request *req)
1545 req->rq_flags |= RQF_TIMED_OUT;
1546 if (req->q->mq_ops->timeout) {
1547 enum blk_eh_timer_return ret;
1549 ret = req->q->mq_ops->timeout(req);
1550 if (ret == BLK_EH_DONE)
1552 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1558 struct blk_expired_data {
1559 bool has_timedout_rq;
1561 unsigned long timeout_start;
1564 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1566 unsigned long deadline;
1568 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1570 if (rq->rq_flags & RQF_TIMED_OUT)
1573 deadline = READ_ONCE(rq->deadline);
1574 if (time_after_eq(expired->timeout_start, deadline))
1577 if (expired->next == 0)
1578 expired->next = deadline;
1579 else if (time_after(expired->next, deadline))
1580 expired->next = deadline;
1584 void blk_mq_put_rq_ref(struct request *rq)
1586 if (is_flush_rq(rq)) {
1587 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1588 blk_mq_free_request(rq);
1589 } else if (req_ref_put_and_test(rq)) {
1590 __blk_mq_free_request(rq);
1594 static bool blk_mq_check_expired(struct request *rq, void *priv)
1596 struct blk_expired_data *expired = priv;
1599 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1600 * be reallocated underneath the timeout handler's processing, then
1601 * the expire check is reliable. If the request is not expired, then
1602 * it was completed and reallocated as a new request after returning
1603 * from blk_mq_check_expired().
1605 if (blk_mq_req_expired(rq, expired)) {
1606 expired->has_timedout_rq = true;
1612 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1614 struct blk_expired_data *expired = priv;
1616 if (blk_mq_req_expired(rq, expired))
1617 blk_mq_rq_timed_out(rq);
1621 static void blk_mq_timeout_work(struct work_struct *work)
1623 struct request_queue *q =
1624 container_of(work, struct request_queue, timeout_work);
1625 struct blk_expired_data expired = {
1626 .timeout_start = jiffies,
1628 struct blk_mq_hw_ctx *hctx;
1631 /* A deadlock might occur if a request is stuck requiring a
1632 * timeout at the same time a queue freeze is waiting
1633 * completion, since the timeout code would not be able to
1634 * acquire the queue reference here.
1636 * That's why we don't use blk_queue_enter here; instead, we use
1637 * percpu_ref_tryget directly, because we need to be able to
1638 * obtain a reference even in the short window between the queue
1639 * starting to freeze, by dropping the first reference in
1640 * blk_freeze_queue_start, and the moment the last request is
1641 * consumed, marked by the instant q_usage_counter reaches
1644 if (!percpu_ref_tryget(&q->q_usage_counter))
1647 /* check if there is any timed-out request */
1648 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1649 if (expired.has_timedout_rq) {
1651 * Before walking tags, we must ensure any submit started
1652 * before the current time has finished. Since the submit
1653 * uses srcu or rcu, wait for a synchronization point to
1654 * ensure all running submits have finished
1656 blk_mq_wait_quiesce_done(q->tag_set);
1659 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1662 if (expired.next != 0) {
1663 mod_timer(&q->timeout, expired.next);
1666 * Request timeouts are handled as a forward rolling timer. If
1667 * we end up here it means that no requests are pending and
1668 * also that no request has been pending for a while. Mark
1669 * each hctx as idle.
1671 queue_for_each_hw_ctx(q, hctx, i) {
1672 /* the hctx may be unmapped, so check it here */
1673 if (blk_mq_hw_queue_mapped(hctx))
1674 blk_mq_tag_idle(hctx);
1680 struct flush_busy_ctx_data {
1681 struct blk_mq_hw_ctx *hctx;
1682 struct list_head *list;
1685 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1687 struct flush_busy_ctx_data *flush_data = data;
1688 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1689 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1690 enum hctx_type type = hctx->type;
1692 spin_lock(&ctx->lock);
1693 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1694 sbitmap_clear_bit(sb, bitnr);
1695 spin_unlock(&ctx->lock);
1700 * Process software queues that have been marked busy, splicing them
1701 * to the for-dispatch
1703 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1705 struct flush_busy_ctx_data data = {
1710 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1712 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1714 struct dispatch_rq_data {
1715 struct blk_mq_hw_ctx *hctx;
1719 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1722 struct dispatch_rq_data *dispatch_data = data;
1723 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1724 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1725 enum hctx_type type = hctx->type;
1727 spin_lock(&ctx->lock);
1728 if (!list_empty(&ctx->rq_lists[type])) {
1729 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1730 list_del_init(&dispatch_data->rq->queuelist);
1731 if (list_empty(&ctx->rq_lists[type]))
1732 sbitmap_clear_bit(sb, bitnr);
1734 spin_unlock(&ctx->lock);
1736 return !dispatch_data->rq;
1739 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1740 struct blk_mq_ctx *start)
1742 unsigned off = start ? start->index_hw[hctx->type] : 0;
1743 struct dispatch_rq_data data = {
1748 __sbitmap_for_each_set(&hctx->ctx_map, off,
1749 dispatch_rq_from_ctx, &data);
1754 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1756 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1757 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1760 blk_mq_tag_busy(rq->mq_hctx);
1762 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1763 bt = &rq->mq_hctx->tags->breserved_tags;
1766 if (!hctx_may_queue(rq->mq_hctx, bt))
1770 tag = __sbitmap_queue_get(bt);
1771 if (tag == BLK_MQ_NO_TAG)
1774 rq->tag = tag + tag_offset;
1778 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1780 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1783 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1784 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1785 rq->rq_flags |= RQF_MQ_INFLIGHT;
1786 __blk_mq_inc_active_requests(hctx);
1788 hctx->tags->rqs[rq->tag] = rq;
1792 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1793 int flags, void *key)
1795 struct blk_mq_hw_ctx *hctx;
1797 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1799 spin_lock(&hctx->dispatch_wait_lock);
1800 if (!list_empty(&wait->entry)) {
1801 struct sbitmap_queue *sbq;
1803 list_del_init(&wait->entry);
1804 sbq = &hctx->tags->bitmap_tags;
1805 atomic_dec(&sbq->ws_active);
1807 spin_unlock(&hctx->dispatch_wait_lock);
1809 blk_mq_run_hw_queue(hctx, true);
1814 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1815 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1816 * restart. For both cases, take care to check the condition again after
1817 * marking us as waiting.
1819 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1822 struct sbitmap_queue *sbq;
1823 struct wait_queue_head *wq;
1824 wait_queue_entry_t *wait;
1827 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1828 !(blk_mq_is_shared_tags(hctx->flags))) {
1829 blk_mq_sched_mark_restart_hctx(hctx);
1832 * It's possible that a tag was freed in the window between the
1833 * allocation failure and adding the hardware queue to the wait
1836 * Don't clear RESTART here, someone else could have set it.
1837 * At most this will cost an extra queue run.
1839 return blk_mq_get_driver_tag(rq);
1842 wait = &hctx->dispatch_wait;
1843 if (!list_empty_careful(&wait->entry))
1846 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1847 sbq = &hctx->tags->breserved_tags;
1849 sbq = &hctx->tags->bitmap_tags;
1850 wq = &bt_wait_ptr(sbq, hctx)->wait;
1852 spin_lock_irq(&wq->lock);
1853 spin_lock(&hctx->dispatch_wait_lock);
1854 if (!list_empty(&wait->entry)) {
1855 spin_unlock(&hctx->dispatch_wait_lock);
1856 spin_unlock_irq(&wq->lock);
1860 atomic_inc(&sbq->ws_active);
1861 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1862 __add_wait_queue(wq, wait);
1865 * It's possible that a tag was freed in the window between the
1866 * allocation failure and adding the hardware queue to the wait
1869 ret = blk_mq_get_driver_tag(rq);
1871 spin_unlock(&hctx->dispatch_wait_lock);
1872 spin_unlock_irq(&wq->lock);
1877 * We got a tag, remove ourselves from the wait queue to ensure
1878 * someone else gets the wakeup.
1880 list_del_init(&wait->entry);
1881 atomic_dec(&sbq->ws_active);
1882 spin_unlock(&hctx->dispatch_wait_lock);
1883 spin_unlock_irq(&wq->lock);
1888 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1889 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1891 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1892 * - EWMA is one simple way to compute running average value
1893 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1894 * - take 4 as factor for avoiding to get too small(0) result, and this
1895 * factor doesn't matter because EWMA decreases exponentially
1897 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1901 ewma = hctx->dispatch_busy;
1906 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1908 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1909 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1911 hctx->dispatch_busy = ewma;
1914 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1916 static void blk_mq_handle_dev_resource(struct request *rq,
1917 struct list_head *list)
1919 list_add(&rq->queuelist, list);
1920 __blk_mq_requeue_request(rq);
1923 static void blk_mq_handle_zone_resource(struct request *rq,
1924 struct list_head *zone_list)
1927 * If we end up here it is because we cannot dispatch a request to a
1928 * specific zone due to LLD level zone-write locking or other zone
1929 * related resource not being available. In this case, set the request
1930 * aside in zone_list for retrying it later.
1932 list_add(&rq->queuelist, zone_list);
1933 __blk_mq_requeue_request(rq);
1936 enum prep_dispatch {
1938 PREP_DISPATCH_NO_TAG,
1939 PREP_DISPATCH_NO_BUDGET,
1942 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1945 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1946 int budget_token = -1;
1949 budget_token = blk_mq_get_dispatch_budget(rq->q);
1950 if (budget_token < 0) {
1951 blk_mq_put_driver_tag(rq);
1952 return PREP_DISPATCH_NO_BUDGET;
1954 blk_mq_set_rq_budget_token(rq, budget_token);
1957 if (!blk_mq_get_driver_tag(rq)) {
1959 * The initial allocation attempt failed, so we need to
1960 * rerun the hardware queue when a tag is freed. The
1961 * waitqueue takes care of that. If the queue is run
1962 * before we add this entry back on the dispatch list,
1963 * we'll re-run it below.
1965 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1967 * All budgets not got from this function will be put
1968 * together during handling partial dispatch
1971 blk_mq_put_dispatch_budget(rq->q, budget_token);
1972 return PREP_DISPATCH_NO_TAG;
1976 return PREP_DISPATCH_OK;
1979 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1980 static void blk_mq_release_budgets(struct request_queue *q,
1981 struct list_head *list)
1985 list_for_each_entry(rq, list, queuelist) {
1986 int budget_token = blk_mq_get_rq_budget_token(rq);
1988 if (budget_token >= 0)
1989 blk_mq_put_dispatch_budget(q, budget_token);
1994 * blk_mq_commit_rqs will notify driver using bd->last that there is no
1995 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
1997 * Attention, we should explicitly call this in unusual cases:
1998 * 1) did not queue everything initially scheduled to queue
1999 * 2) the last attempt to queue a request failed
2001 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
2004 if (hctx->queue->mq_ops->commit_rqs && queued) {
2005 trace_block_unplug(hctx->queue, queued, !from_schedule);
2006 hctx->queue->mq_ops->commit_rqs(hctx);
2011 * Returns true if we did some work AND can potentially do more.
2013 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
2014 unsigned int nr_budgets)
2016 enum prep_dispatch prep;
2017 struct request_queue *q = hctx->queue;
2020 blk_status_t ret = BLK_STS_OK;
2021 LIST_HEAD(zone_list);
2022 bool needs_resource = false;
2024 if (list_empty(list))
2028 * Now process all the entries, sending them to the driver.
2032 struct blk_mq_queue_data bd;
2034 rq = list_first_entry(list, struct request, queuelist);
2036 WARN_ON_ONCE(hctx != rq->mq_hctx);
2037 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2038 if (prep != PREP_DISPATCH_OK)
2041 list_del_init(&rq->queuelist);
2044 bd.last = list_empty(list);
2047 * once the request is queued to lld, no need to cover the
2052 ret = q->mq_ops->queue_rq(hctx, &bd);
2057 case BLK_STS_RESOURCE:
2058 needs_resource = true;
2060 case BLK_STS_DEV_RESOURCE:
2061 blk_mq_handle_dev_resource(rq, list);
2063 case BLK_STS_ZONE_RESOURCE:
2065 * Move the request to zone_list and keep going through
2066 * the dispatch list to find more requests the drive can
2069 blk_mq_handle_zone_resource(rq, &zone_list);
2070 needs_resource = true;
2073 blk_mq_end_request(rq, ret);
2075 } while (!list_empty(list));
2077 if (!list_empty(&zone_list))
2078 list_splice_tail_init(&zone_list, list);
2080 /* If we didn't flush the entire list, we could have told the driver
2081 * there was more coming, but that turned out to be a lie.
2083 if (!list_empty(list) || ret != BLK_STS_OK)
2084 blk_mq_commit_rqs(hctx, queued, false);
2087 * Any items that need requeuing? Stuff them into hctx->dispatch,
2088 * that is where we will continue on next queue run.
2090 if (!list_empty(list)) {
2092 /* For non-shared tags, the RESTART check will suffice */
2093 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2094 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2095 blk_mq_is_shared_tags(hctx->flags));
2098 blk_mq_release_budgets(q, list);
2100 spin_lock(&hctx->lock);
2101 list_splice_tail_init(list, &hctx->dispatch);
2102 spin_unlock(&hctx->lock);
2105 * Order adding requests to hctx->dispatch and checking
2106 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2107 * in blk_mq_sched_restart(). Avoid restart code path to
2108 * miss the new added requests to hctx->dispatch, meantime
2109 * SCHED_RESTART is observed here.
2114 * If SCHED_RESTART was set by the caller of this function and
2115 * it is no longer set that means that it was cleared by another
2116 * thread and hence that a queue rerun is needed.
2118 * If 'no_tag' is set, that means that we failed getting
2119 * a driver tag with an I/O scheduler attached. If our dispatch
2120 * waitqueue is no longer active, ensure that we run the queue
2121 * AFTER adding our entries back to the list.
2123 * If no I/O scheduler has been configured it is possible that
2124 * the hardware queue got stopped and restarted before requests
2125 * were pushed back onto the dispatch list. Rerun the queue to
2126 * avoid starvation. Notes:
2127 * - blk_mq_run_hw_queue() checks whether or not a queue has
2128 * been stopped before rerunning a queue.
2129 * - Some but not all block drivers stop a queue before
2130 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2133 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2134 * bit is set, run queue after a delay to avoid IO stalls
2135 * that could otherwise occur if the queue is idle. We'll do
2136 * similar if we couldn't get budget or couldn't lock a zone
2137 * and SCHED_RESTART is set.
2139 needs_restart = blk_mq_sched_needs_restart(hctx);
2140 if (prep == PREP_DISPATCH_NO_BUDGET)
2141 needs_resource = true;
2142 if (!needs_restart ||
2143 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2144 blk_mq_run_hw_queue(hctx, true);
2145 else if (needs_resource)
2146 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2148 blk_mq_update_dispatch_busy(hctx, true);
2152 blk_mq_update_dispatch_busy(hctx, false);
2156 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2158 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2160 if (cpu >= nr_cpu_ids)
2161 cpu = cpumask_first(hctx->cpumask);
2166 * It'd be great if the workqueue API had a way to pass
2167 * in a mask and had some smarts for more clever placement.
2168 * For now we just round-robin here, switching for every
2169 * BLK_MQ_CPU_WORK_BATCH queued items.
2171 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2174 int next_cpu = hctx->next_cpu;
2176 if (hctx->queue->nr_hw_queues == 1)
2177 return WORK_CPU_UNBOUND;
2179 if (--hctx->next_cpu_batch <= 0) {
2181 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2183 if (next_cpu >= nr_cpu_ids)
2184 next_cpu = blk_mq_first_mapped_cpu(hctx);
2185 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2189 * Do unbound schedule if we can't find a online CPU for this hctx,
2190 * and it should only happen in the path of handling CPU DEAD.
2192 if (!cpu_online(next_cpu)) {
2199 * Make sure to re-select CPU next time once after CPUs
2200 * in hctx->cpumask become online again.
2202 hctx->next_cpu = next_cpu;
2203 hctx->next_cpu_batch = 1;
2204 return WORK_CPU_UNBOUND;
2207 hctx->next_cpu = next_cpu;
2212 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2213 * @hctx: Pointer to the hardware queue to run.
2214 * @msecs: Milliseconds of delay to wait before running the queue.
2216 * Run a hardware queue asynchronously with a delay of @msecs.
2218 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2220 if (unlikely(blk_mq_hctx_stopped(hctx)))
2222 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2223 msecs_to_jiffies(msecs));
2225 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2228 * blk_mq_run_hw_queue - Start to run a hardware queue.
2229 * @hctx: Pointer to the hardware queue to run.
2230 * @async: If we want to run the queue asynchronously.
2232 * Check if the request queue is not in a quiesced state and if there are
2233 * pending requests to be sent. If this is true, run the queue to send requests
2236 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2241 * We can't run the queue inline with interrupts disabled.
2243 WARN_ON_ONCE(!async && in_interrupt());
2246 * When queue is quiesced, we may be switching io scheduler, or
2247 * updating nr_hw_queues, or other things, and we can't run queue
2248 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2250 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2253 __blk_mq_run_dispatch_ops(hctx->queue, false,
2254 need_run = !blk_queue_quiesced(hctx->queue) &&
2255 blk_mq_hctx_has_pending(hctx));
2260 if (async || (hctx->flags & BLK_MQ_F_BLOCKING) ||
2261 !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2262 blk_mq_delay_run_hw_queue(hctx, 0);
2266 blk_mq_run_dispatch_ops(hctx->queue,
2267 blk_mq_sched_dispatch_requests(hctx));
2269 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2272 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2275 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2277 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2279 * If the IO scheduler does not respect hardware queues when
2280 * dispatching, we just don't bother with multiple HW queues and
2281 * dispatch from hctx for the current CPU since running multiple queues
2282 * just causes lock contention inside the scheduler and pointless cache
2285 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2287 if (!blk_mq_hctx_stopped(hctx))
2293 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2294 * @q: Pointer to the request queue to run.
2295 * @async: If we want to run the queue asynchronously.
2297 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2299 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2303 if (blk_queue_sq_sched(q))
2304 sq_hctx = blk_mq_get_sq_hctx(q);
2305 queue_for_each_hw_ctx(q, hctx, i) {
2306 if (blk_mq_hctx_stopped(hctx))
2309 * Dispatch from this hctx either if there's no hctx preferred
2310 * by IO scheduler or if it has requests that bypass the
2313 if (!sq_hctx || sq_hctx == hctx ||
2314 !list_empty_careful(&hctx->dispatch))
2315 blk_mq_run_hw_queue(hctx, async);
2318 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2321 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2322 * @q: Pointer to the request queue to run.
2323 * @msecs: Milliseconds of delay to wait before running the queues.
2325 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2327 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2331 if (blk_queue_sq_sched(q))
2332 sq_hctx = blk_mq_get_sq_hctx(q);
2333 queue_for_each_hw_ctx(q, hctx, i) {
2334 if (blk_mq_hctx_stopped(hctx))
2337 * If there is already a run_work pending, leave the
2338 * pending delay untouched. Otherwise, a hctx can stall
2339 * if another hctx is re-delaying the other's work
2340 * before the work executes.
2342 if (delayed_work_pending(&hctx->run_work))
2345 * Dispatch from this hctx either if there's no hctx preferred
2346 * by IO scheduler or if it has requests that bypass the
2349 if (!sq_hctx || sq_hctx == hctx ||
2350 !list_empty_careful(&hctx->dispatch))
2351 blk_mq_delay_run_hw_queue(hctx, msecs);
2354 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2357 * This function is often used for pausing .queue_rq() by driver when
2358 * there isn't enough resource or some conditions aren't satisfied, and
2359 * BLK_STS_RESOURCE is usually returned.
2361 * We do not guarantee that dispatch can be drained or blocked
2362 * after blk_mq_stop_hw_queue() returns. Please use
2363 * blk_mq_quiesce_queue() for that requirement.
2365 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2367 cancel_delayed_work(&hctx->run_work);
2369 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2371 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2374 * This function is often used for pausing .queue_rq() by driver when
2375 * there isn't enough resource or some conditions aren't satisfied, and
2376 * BLK_STS_RESOURCE is usually returned.
2378 * We do not guarantee that dispatch can be drained or blocked
2379 * after blk_mq_stop_hw_queues() returns. Please use
2380 * blk_mq_quiesce_queue() for that requirement.
2382 void blk_mq_stop_hw_queues(struct request_queue *q)
2384 struct blk_mq_hw_ctx *hctx;
2387 queue_for_each_hw_ctx(q, hctx, i)
2388 blk_mq_stop_hw_queue(hctx);
2390 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2392 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2394 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2396 blk_mq_run_hw_queue(hctx, false);
2398 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2400 void blk_mq_start_hw_queues(struct request_queue *q)
2402 struct blk_mq_hw_ctx *hctx;
2405 queue_for_each_hw_ctx(q, hctx, i)
2406 blk_mq_start_hw_queue(hctx);
2408 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2410 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2412 if (!blk_mq_hctx_stopped(hctx))
2415 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2416 blk_mq_run_hw_queue(hctx, async);
2418 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2420 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2422 struct blk_mq_hw_ctx *hctx;
2425 queue_for_each_hw_ctx(q, hctx, i)
2426 blk_mq_start_stopped_hw_queue(hctx, async);
2428 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2430 static void blk_mq_run_work_fn(struct work_struct *work)
2432 struct blk_mq_hw_ctx *hctx =
2433 container_of(work, struct blk_mq_hw_ctx, run_work.work);
2435 blk_mq_run_dispatch_ops(hctx->queue,
2436 blk_mq_sched_dispatch_requests(hctx));
2440 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2441 * @rq: Pointer to request to be inserted.
2442 * @flags: BLK_MQ_INSERT_*
2444 * Should only be used carefully, when the caller knows we want to
2445 * bypass a potential IO scheduler on the target device.
2447 static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2449 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2451 spin_lock(&hctx->lock);
2452 if (flags & BLK_MQ_INSERT_AT_HEAD)
2453 list_add(&rq->queuelist, &hctx->dispatch);
2455 list_add_tail(&rq->queuelist, &hctx->dispatch);
2456 spin_unlock(&hctx->lock);
2459 static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2460 struct blk_mq_ctx *ctx, struct list_head *list,
2461 bool run_queue_async)
2464 enum hctx_type type = hctx->type;
2467 * Try to issue requests directly if the hw queue isn't busy to save an
2468 * extra enqueue & dequeue to the sw queue.
2470 if (!hctx->dispatch_busy && !run_queue_async) {
2471 blk_mq_run_dispatch_ops(hctx->queue,
2472 blk_mq_try_issue_list_directly(hctx, list));
2473 if (list_empty(list))
2478 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2481 list_for_each_entry(rq, list, queuelist) {
2482 BUG_ON(rq->mq_ctx != ctx);
2483 trace_block_rq_insert(rq);
2486 spin_lock(&ctx->lock);
2487 list_splice_tail_init(list, &ctx->rq_lists[type]);
2488 blk_mq_hctx_mark_pending(hctx, ctx);
2489 spin_unlock(&ctx->lock);
2491 blk_mq_run_hw_queue(hctx, run_queue_async);
2494 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2496 struct request_queue *q = rq->q;
2497 struct blk_mq_ctx *ctx = rq->mq_ctx;
2498 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2500 if (blk_rq_is_passthrough(rq)) {
2502 * Passthrough request have to be added to hctx->dispatch
2503 * directly. The device may be in a situation where it can't
2504 * handle FS request, and always returns BLK_STS_RESOURCE for
2505 * them, which gets them added to hctx->dispatch.
2507 * If a passthrough request is required to unblock the queues,
2508 * and it is added to the scheduler queue, there is no chance to
2509 * dispatch it given we prioritize requests in hctx->dispatch.
2511 blk_mq_request_bypass_insert(rq, flags);
2512 } else if (req_op(rq) == REQ_OP_FLUSH) {
2514 * Firstly normal IO request is inserted to scheduler queue or
2515 * sw queue, meantime we add flush request to dispatch queue(
2516 * hctx->dispatch) directly and there is at most one in-flight
2517 * flush request for each hw queue, so it doesn't matter to add
2518 * flush request to tail or front of the dispatch queue.
2520 * Secondly in case of NCQ, flush request belongs to non-NCQ
2521 * command, and queueing it will fail when there is any
2522 * in-flight normal IO request(NCQ command). When adding flush
2523 * rq to the front of hctx->dispatch, it is easier to introduce
2524 * extra time to flush rq's latency because of S_SCHED_RESTART
2525 * compared with adding to the tail of dispatch queue, then
2526 * chance of flush merge is increased, and less flush requests
2527 * will be issued to controller. It is observed that ~10% time
2528 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2529 * drive when adding flush rq to the front of hctx->dispatch.
2531 * Simply queue flush rq to the front of hctx->dispatch so that
2532 * intensive flush workloads can benefit in case of NCQ HW.
2534 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2535 } else if (q->elevator) {
2538 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2540 list_add(&rq->queuelist, &list);
2541 q->elevator->type->ops.insert_requests(hctx, &list, flags);
2543 trace_block_rq_insert(rq);
2545 spin_lock(&ctx->lock);
2546 if (flags & BLK_MQ_INSERT_AT_HEAD)
2547 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2549 list_add_tail(&rq->queuelist,
2550 &ctx->rq_lists[hctx->type]);
2551 blk_mq_hctx_mark_pending(hctx, ctx);
2552 spin_unlock(&ctx->lock);
2556 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2557 unsigned int nr_segs)
2561 if (bio->bi_opf & REQ_RAHEAD)
2562 rq->cmd_flags |= REQ_FAILFAST_MASK;
2564 rq->__sector = bio->bi_iter.bi_sector;
2565 blk_rq_bio_prep(rq, bio, nr_segs);
2567 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2568 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2571 blk_account_io_start(rq);
2574 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2575 struct request *rq, bool last)
2577 struct request_queue *q = rq->q;
2578 struct blk_mq_queue_data bd = {
2585 * For OK queue, we are done. For error, caller may kill it.
2586 * Any other error (busy), just add it to our list as we
2587 * previously would have done.
2589 ret = q->mq_ops->queue_rq(hctx, &bd);
2592 blk_mq_update_dispatch_busy(hctx, false);
2594 case BLK_STS_RESOURCE:
2595 case BLK_STS_DEV_RESOURCE:
2596 blk_mq_update_dispatch_busy(hctx, true);
2597 __blk_mq_requeue_request(rq);
2600 blk_mq_update_dispatch_busy(hctx, false);
2607 static bool blk_mq_get_budget_and_tag(struct request *rq)
2611 budget_token = blk_mq_get_dispatch_budget(rq->q);
2612 if (budget_token < 0)
2614 blk_mq_set_rq_budget_token(rq, budget_token);
2615 if (!blk_mq_get_driver_tag(rq)) {
2616 blk_mq_put_dispatch_budget(rq->q, budget_token);
2623 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2624 * @hctx: Pointer of the associated hardware queue.
2625 * @rq: Pointer to request to be sent.
2627 * If the device has enough resources to accept a new request now, send the
2628 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2629 * we can try send it another time in the future. Requests inserted at this
2630 * queue have higher priority.
2632 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2637 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2638 blk_mq_insert_request(rq, 0);
2642 if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
2643 blk_mq_insert_request(rq, 0);
2644 blk_mq_run_hw_queue(hctx, false);
2648 ret = __blk_mq_issue_directly(hctx, rq, true);
2652 case BLK_STS_RESOURCE:
2653 case BLK_STS_DEV_RESOURCE:
2654 blk_mq_request_bypass_insert(rq, 0);
2655 blk_mq_run_hw_queue(hctx, false);
2658 blk_mq_end_request(rq, ret);
2663 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2665 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2667 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2668 blk_mq_insert_request(rq, 0);
2672 if (!blk_mq_get_budget_and_tag(rq))
2673 return BLK_STS_RESOURCE;
2674 return __blk_mq_issue_directly(hctx, rq, last);
2677 static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2679 struct blk_mq_hw_ctx *hctx = NULL;
2682 blk_status_t ret = BLK_STS_OK;
2684 while ((rq = rq_list_pop(&plug->mq_list))) {
2685 bool last = rq_list_empty(plug->mq_list);
2687 if (hctx != rq->mq_hctx) {
2689 blk_mq_commit_rqs(hctx, queued, false);
2695 ret = blk_mq_request_issue_directly(rq, last);
2700 case BLK_STS_RESOURCE:
2701 case BLK_STS_DEV_RESOURCE:
2702 blk_mq_request_bypass_insert(rq, 0);
2703 blk_mq_run_hw_queue(hctx, false);
2706 blk_mq_end_request(rq, ret);
2712 if (ret != BLK_STS_OK)
2713 blk_mq_commit_rqs(hctx, queued, false);
2716 static void __blk_mq_flush_plug_list(struct request_queue *q,
2717 struct blk_plug *plug)
2719 if (blk_queue_quiesced(q))
2721 q->mq_ops->queue_rqs(&plug->mq_list);
2724 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2726 struct blk_mq_hw_ctx *this_hctx = NULL;
2727 struct blk_mq_ctx *this_ctx = NULL;
2728 struct request *requeue_list = NULL;
2729 struct request **requeue_lastp = &requeue_list;
2730 unsigned int depth = 0;
2731 bool is_passthrough = false;
2735 struct request *rq = rq_list_pop(&plug->mq_list);
2738 this_hctx = rq->mq_hctx;
2739 this_ctx = rq->mq_ctx;
2740 is_passthrough = blk_rq_is_passthrough(rq);
2741 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2742 is_passthrough != blk_rq_is_passthrough(rq)) {
2743 rq_list_add_tail(&requeue_lastp, rq);
2746 list_add(&rq->queuelist, &list);
2748 } while (!rq_list_empty(plug->mq_list));
2750 plug->mq_list = requeue_list;
2751 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2753 percpu_ref_get(&this_hctx->queue->q_usage_counter);
2754 /* passthrough requests should never be issued to the I/O scheduler */
2755 if (is_passthrough) {
2756 spin_lock(&this_hctx->lock);
2757 list_splice_tail_init(&list, &this_hctx->dispatch);
2758 spin_unlock(&this_hctx->lock);
2759 blk_mq_run_hw_queue(this_hctx, from_sched);
2760 } else if (this_hctx->queue->elevator) {
2761 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2763 blk_mq_run_hw_queue(this_hctx, from_sched);
2765 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2767 percpu_ref_put(&this_hctx->queue->q_usage_counter);
2770 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2775 * We may have been called recursively midway through handling
2776 * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2777 * To avoid mq_list changing under our feet, clear rq_count early and
2778 * bail out specifically if rq_count is 0 rather than checking
2779 * whether the mq_list is empty.
2781 if (plug->rq_count == 0)
2785 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2786 struct request_queue *q;
2788 rq = rq_list_peek(&plug->mq_list);
2792 * Peek first request and see if we have a ->queue_rqs() hook.
2793 * If we do, we can dispatch the whole plug list in one go. We
2794 * already know at this point that all requests belong to the
2795 * same queue, caller must ensure that's the case.
2797 * Since we pass off the full list to the driver at this point,
2798 * we do not increment the active request count for the queue.
2799 * Bypass shared tags for now because of that.
2801 if (q->mq_ops->queue_rqs &&
2802 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2803 blk_mq_run_dispatch_ops(q,
2804 __blk_mq_flush_plug_list(q, plug));
2805 if (rq_list_empty(plug->mq_list))
2809 blk_mq_run_dispatch_ops(q,
2810 blk_mq_plug_issue_direct(plug));
2811 if (rq_list_empty(plug->mq_list))
2816 blk_mq_dispatch_plug_list(plug, from_schedule);
2817 } while (!rq_list_empty(plug->mq_list));
2820 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2821 struct list_head *list)
2824 blk_status_t ret = BLK_STS_OK;
2826 while (!list_empty(list)) {
2827 struct request *rq = list_first_entry(list, struct request,
2830 list_del_init(&rq->queuelist);
2831 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2836 case BLK_STS_RESOURCE:
2837 case BLK_STS_DEV_RESOURCE:
2838 blk_mq_request_bypass_insert(rq, 0);
2839 if (list_empty(list))
2840 blk_mq_run_hw_queue(hctx, false);
2843 blk_mq_end_request(rq, ret);
2849 if (ret != BLK_STS_OK)
2850 blk_mq_commit_rqs(hctx, queued, false);
2853 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2854 struct bio *bio, unsigned int nr_segs)
2856 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2857 if (blk_attempt_plug_merge(q, bio, nr_segs))
2859 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2865 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2866 struct blk_plug *plug,
2870 struct blk_mq_alloc_data data = {
2873 .cmd_flags = bio->bi_opf,
2877 if (unlikely(bio_queue_enter(bio)))
2880 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2883 rq_qos_throttle(q, bio);
2886 data.nr_tags = plug->nr_ios;
2888 data.cached_rq = &plug->cached_rq;
2891 rq = __blk_mq_alloc_requests(&data);
2894 rq_qos_cleanup(q, bio);
2895 if (bio->bi_opf & REQ_NOWAIT)
2896 bio_wouldblock_error(bio);
2902 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2903 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2906 enum hctx_type type, hctx_type;
2910 rq = rq_list_peek(&plug->cached_rq);
2911 if (!rq || rq->q != q)
2914 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2919 type = blk_mq_get_hctx_type((*bio)->bi_opf);
2920 hctx_type = rq->mq_hctx->type;
2921 if (type != hctx_type &&
2922 !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2924 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2928 * If any qos ->throttle() end up blocking, we will have flushed the
2929 * plug and hence killed the cached_rq list as well. Pop this entry
2930 * before we throttle.
2932 plug->cached_rq = rq_list_next(rq);
2933 rq_qos_throttle(q, *bio);
2935 blk_mq_rq_time_init(rq, 0);
2936 rq->cmd_flags = (*bio)->bi_opf;
2937 INIT_LIST_HEAD(&rq->queuelist);
2941 static void bio_set_ioprio(struct bio *bio)
2943 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2944 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2945 bio->bi_ioprio = get_current_ioprio();
2946 blkcg_set_ioprio(bio);
2950 * blk_mq_submit_bio - Create and send a request to block device.
2951 * @bio: Bio pointer.
2953 * Builds up a request structure from @q and @bio and send to the device. The
2954 * request may not be queued directly to hardware if:
2955 * * This request can be merged with another one
2956 * * We want to place request at plug queue for possible future merging
2957 * * There is an IO scheduler active at this queue
2959 * It will not queue the request if there is an error with the bio, or at the
2962 void blk_mq_submit_bio(struct bio *bio)
2964 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2965 struct blk_plug *plug = blk_mq_plug(bio);
2966 const int is_sync = op_is_sync(bio->bi_opf);
2967 struct blk_mq_hw_ctx *hctx;
2969 unsigned int nr_segs = 1;
2972 bio = blk_queue_bounce(bio, q);
2973 if (bio_may_exceed_limits(bio, &q->limits)) {
2974 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2979 if (!bio_integrity_prep(bio))
2982 bio_set_ioprio(bio);
2984 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2988 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2993 trace_block_getrq(bio);
2995 rq_qos_track(q, rq, bio);
2997 blk_mq_bio_to_request(rq, bio, nr_segs);
2999 ret = blk_crypto_rq_get_keyslot(rq);
3000 if (ret != BLK_STS_OK) {
3001 bio->bi_status = ret;
3003 blk_mq_free_request(rq);
3007 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
3011 blk_add_rq_to_plug(plug, rq);
3016 if ((rq->rq_flags & RQF_USE_SCHED) ||
3017 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
3018 blk_mq_insert_request(rq, 0);
3019 blk_mq_run_hw_queue(hctx, true);
3021 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
3025 #ifdef CONFIG_BLK_MQ_STACKING
3027 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3028 * @rq: the request being queued
3030 blk_status_t blk_insert_cloned_request(struct request *rq)
3032 struct request_queue *q = rq->q;
3033 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3034 unsigned int max_segments = blk_rq_get_max_segments(rq);
3037 if (blk_rq_sectors(rq) > max_sectors) {
3039 * SCSI device does not have a good way to return if
3040 * Write Same/Zero is actually supported. If a device rejects
3041 * a non-read/write command (discard, write same,etc.) the
3042 * low-level device driver will set the relevant queue limit to
3043 * 0 to prevent blk-lib from issuing more of the offending
3044 * operations. Commands queued prior to the queue limit being
3045 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3046 * errors being propagated to upper layers.
3048 if (max_sectors == 0)
3049 return BLK_STS_NOTSUPP;
3051 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3052 __func__, blk_rq_sectors(rq), max_sectors);
3053 return BLK_STS_IOERR;
3057 * The queue settings related to segment counting may differ from the
3060 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3061 if (rq->nr_phys_segments > max_segments) {
3062 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3063 __func__, rq->nr_phys_segments, max_segments);
3064 return BLK_STS_IOERR;
3067 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3068 return BLK_STS_IOERR;
3070 ret = blk_crypto_rq_get_keyslot(rq);
3071 if (ret != BLK_STS_OK)
3074 blk_account_io_start(rq);
3077 * Since we have a scheduler attached on the top device,
3078 * bypass a potential scheduler on the bottom device for
3081 blk_mq_run_dispatch_ops(q,
3082 ret = blk_mq_request_issue_directly(rq, true));
3084 blk_account_io_done(rq, ktime_get_ns());
3087 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3090 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3091 * @rq: the clone request to be cleaned up
3094 * Free all bios in @rq for a cloned request.
3096 void blk_rq_unprep_clone(struct request *rq)
3100 while ((bio = rq->bio) != NULL) {
3101 rq->bio = bio->bi_next;
3106 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3109 * blk_rq_prep_clone - Helper function to setup clone request
3110 * @rq: the request to be setup
3111 * @rq_src: original request to be cloned
3112 * @bs: bio_set that bios for clone are allocated from
3113 * @gfp_mask: memory allocation mask for bio
3114 * @bio_ctr: setup function to be called for each clone bio.
3115 * Returns %0 for success, non %0 for failure.
3116 * @data: private data to be passed to @bio_ctr
3119 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3120 * Also, pages which the original bios are pointing to are not copied
3121 * and the cloned bios just point same pages.
3122 * So cloned bios must be completed before original bios, which means
3123 * the caller must complete @rq before @rq_src.
3125 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3126 struct bio_set *bs, gfp_t gfp_mask,
3127 int (*bio_ctr)(struct bio *, struct bio *, void *),
3130 struct bio *bio, *bio_src;
3135 __rq_for_each_bio(bio_src, rq_src) {
3136 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3141 if (bio_ctr && bio_ctr(bio, bio_src, data))
3145 rq->biotail->bi_next = bio;
3148 rq->bio = rq->biotail = bio;
3153 /* Copy attributes of the original request to the clone request. */
3154 rq->__sector = blk_rq_pos(rq_src);
3155 rq->__data_len = blk_rq_bytes(rq_src);
3156 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3157 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3158 rq->special_vec = rq_src->special_vec;
3160 rq->nr_phys_segments = rq_src->nr_phys_segments;
3161 rq->ioprio = rq_src->ioprio;
3163 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3171 blk_rq_unprep_clone(rq);
3175 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3176 #endif /* CONFIG_BLK_MQ_STACKING */
3179 * Steal bios from a request and add them to a bio list.
3180 * The request must not have been partially completed before.
3182 void blk_steal_bios(struct bio_list *list, struct request *rq)
3186 list->tail->bi_next = rq->bio;
3188 list->head = rq->bio;
3189 list->tail = rq->biotail;
3197 EXPORT_SYMBOL_GPL(blk_steal_bios);
3199 static size_t order_to_size(unsigned int order)
3201 return (size_t)PAGE_SIZE << order;
3204 /* called before freeing request pool in @tags */
3205 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3206 struct blk_mq_tags *tags)
3209 unsigned long flags;
3212 * There is no need to clear mapping if driver tags is not initialized
3213 * or the mapping belongs to the driver tags.
3215 if (!drv_tags || drv_tags == tags)
3218 list_for_each_entry(page, &tags->page_list, lru) {
3219 unsigned long start = (unsigned long)page_address(page);
3220 unsigned long end = start + order_to_size(page->private);
3223 for (i = 0; i < drv_tags->nr_tags; i++) {
3224 struct request *rq = drv_tags->rqs[i];
3225 unsigned long rq_addr = (unsigned long)rq;
3227 if (rq_addr >= start && rq_addr < end) {
3228 WARN_ON_ONCE(req_ref_read(rq) != 0);
3229 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3235 * Wait until all pending iteration is done.
3237 * Request reference is cleared and it is guaranteed to be observed
3238 * after the ->lock is released.
3240 spin_lock_irqsave(&drv_tags->lock, flags);
3241 spin_unlock_irqrestore(&drv_tags->lock, flags);
3244 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3245 unsigned int hctx_idx)
3247 struct blk_mq_tags *drv_tags;
3250 if (list_empty(&tags->page_list))
3253 if (blk_mq_is_shared_tags(set->flags))
3254 drv_tags = set->shared_tags;
3256 drv_tags = set->tags[hctx_idx];
3258 if (tags->static_rqs && set->ops->exit_request) {
3261 for (i = 0; i < tags->nr_tags; i++) {
3262 struct request *rq = tags->static_rqs[i];
3266 set->ops->exit_request(set, rq, hctx_idx);
3267 tags->static_rqs[i] = NULL;
3271 blk_mq_clear_rq_mapping(drv_tags, tags);
3273 while (!list_empty(&tags->page_list)) {
3274 page = list_first_entry(&tags->page_list, struct page, lru);
3275 list_del_init(&page->lru);
3277 * Remove kmemleak object previously allocated in
3278 * blk_mq_alloc_rqs().
3280 kmemleak_free(page_address(page));
3281 __free_pages(page, page->private);
3285 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3289 kfree(tags->static_rqs);
3290 tags->static_rqs = NULL;
3292 blk_mq_free_tags(tags);
3295 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3296 unsigned int hctx_idx)
3300 for (i = 0; i < set->nr_maps; i++) {
3301 unsigned int start = set->map[i].queue_offset;
3302 unsigned int end = start + set->map[i].nr_queues;
3304 if (hctx_idx >= start && hctx_idx < end)
3308 if (i >= set->nr_maps)
3309 i = HCTX_TYPE_DEFAULT;
3314 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3315 unsigned int hctx_idx)
3317 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3319 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3322 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3323 unsigned int hctx_idx,
3324 unsigned int nr_tags,
3325 unsigned int reserved_tags)
3327 int node = blk_mq_get_hctx_node(set, hctx_idx);
3328 struct blk_mq_tags *tags;
3330 if (node == NUMA_NO_NODE)
3331 node = set->numa_node;
3333 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3334 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3338 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3339 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3344 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3345 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3347 if (!tags->static_rqs)
3355 blk_mq_free_tags(tags);
3359 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3360 unsigned int hctx_idx, int node)
3364 if (set->ops->init_request) {
3365 ret = set->ops->init_request(set, rq, hctx_idx, node);
3370 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3374 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3375 struct blk_mq_tags *tags,
3376 unsigned int hctx_idx, unsigned int depth)
3378 unsigned int i, j, entries_per_page, max_order = 4;
3379 int node = blk_mq_get_hctx_node(set, hctx_idx);
3380 size_t rq_size, left;
3382 if (node == NUMA_NO_NODE)
3383 node = set->numa_node;
3385 INIT_LIST_HEAD(&tags->page_list);
3388 * rq_size is the size of the request plus driver payload, rounded
3389 * to the cacheline size
3391 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3393 left = rq_size * depth;
3395 for (i = 0; i < depth; ) {
3396 int this_order = max_order;
3401 while (this_order && left < order_to_size(this_order - 1))
3405 page = alloc_pages_node(node,
3406 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3412 if (order_to_size(this_order) < rq_size)
3419 page->private = this_order;
3420 list_add_tail(&page->lru, &tags->page_list);
3422 p = page_address(page);
3424 * Allow kmemleak to scan these pages as they contain pointers
3425 * to additional allocations like via ops->init_request().
3427 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3428 entries_per_page = order_to_size(this_order) / rq_size;
3429 to_do = min(entries_per_page, depth - i);
3430 left -= to_do * rq_size;
3431 for (j = 0; j < to_do; j++) {
3432 struct request *rq = p;
3434 tags->static_rqs[i] = rq;
3435 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3436 tags->static_rqs[i] = NULL;
3447 blk_mq_free_rqs(set, tags, hctx_idx);
3451 struct rq_iter_data {
3452 struct blk_mq_hw_ctx *hctx;
3456 static bool blk_mq_has_request(struct request *rq, void *data)
3458 struct rq_iter_data *iter_data = data;
3460 if (rq->mq_hctx != iter_data->hctx)
3462 iter_data->has_rq = true;
3466 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3468 struct blk_mq_tags *tags = hctx->sched_tags ?
3469 hctx->sched_tags : hctx->tags;
3470 struct rq_iter_data data = {
3474 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3478 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3479 struct blk_mq_hw_ctx *hctx)
3481 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3483 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3488 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3490 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3491 struct blk_mq_hw_ctx, cpuhp_online);
3493 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3494 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3498 * Prevent new request from being allocated on the current hctx.
3500 * The smp_mb__after_atomic() Pairs with the implied barrier in
3501 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3502 * seen once we return from the tag allocator.
3504 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3505 smp_mb__after_atomic();
3508 * Try to grab a reference to the queue and wait for any outstanding
3509 * requests. If we could not grab a reference the queue has been
3510 * frozen and there are no requests.
3512 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3513 while (blk_mq_hctx_has_requests(hctx))
3515 percpu_ref_put(&hctx->queue->q_usage_counter);
3521 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3523 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3524 struct blk_mq_hw_ctx, cpuhp_online);
3526 if (cpumask_test_cpu(cpu, hctx->cpumask))
3527 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3532 * 'cpu' is going away. splice any existing rq_list entries from this
3533 * software queue to the hw queue dispatch list, and ensure that it
3536 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3538 struct blk_mq_hw_ctx *hctx;
3539 struct blk_mq_ctx *ctx;
3541 enum hctx_type type;
3543 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3544 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3547 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3550 spin_lock(&ctx->lock);
3551 if (!list_empty(&ctx->rq_lists[type])) {
3552 list_splice_init(&ctx->rq_lists[type], &tmp);
3553 blk_mq_hctx_clear_pending(hctx, ctx);
3555 spin_unlock(&ctx->lock);
3557 if (list_empty(&tmp))
3560 spin_lock(&hctx->lock);
3561 list_splice_tail_init(&tmp, &hctx->dispatch);
3562 spin_unlock(&hctx->lock);
3564 blk_mq_run_hw_queue(hctx, true);
3568 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3570 if (!(hctx->flags & BLK_MQ_F_STACKING))
3571 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3572 &hctx->cpuhp_online);
3573 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3578 * Before freeing hw queue, clearing the flush request reference in
3579 * tags->rqs[] for avoiding potential UAF.
3581 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3582 unsigned int queue_depth, struct request *flush_rq)
3585 unsigned long flags;
3587 /* The hw queue may not be mapped yet */
3591 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3593 for (i = 0; i < queue_depth; i++)
3594 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3597 * Wait until all pending iteration is done.
3599 * Request reference is cleared and it is guaranteed to be observed
3600 * after the ->lock is released.
3602 spin_lock_irqsave(&tags->lock, flags);
3603 spin_unlock_irqrestore(&tags->lock, flags);
3606 /* hctx->ctxs will be freed in queue's release handler */
3607 static void blk_mq_exit_hctx(struct request_queue *q,
3608 struct blk_mq_tag_set *set,
3609 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3611 struct request *flush_rq = hctx->fq->flush_rq;
3613 if (blk_mq_hw_queue_mapped(hctx))
3614 blk_mq_tag_idle(hctx);
3616 if (blk_queue_init_done(q))
3617 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3618 set->queue_depth, flush_rq);
3619 if (set->ops->exit_request)
3620 set->ops->exit_request(set, flush_rq, hctx_idx);
3622 if (set->ops->exit_hctx)
3623 set->ops->exit_hctx(hctx, hctx_idx);
3625 blk_mq_remove_cpuhp(hctx);
3627 xa_erase(&q->hctx_table, hctx_idx);
3629 spin_lock(&q->unused_hctx_lock);
3630 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3631 spin_unlock(&q->unused_hctx_lock);
3634 static void blk_mq_exit_hw_queues(struct request_queue *q,
3635 struct blk_mq_tag_set *set, int nr_queue)
3637 struct blk_mq_hw_ctx *hctx;
3640 queue_for_each_hw_ctx(q, hctx, i) {
3643 blk_mq_exit_hctx(q, set, hctx, i);
3647 static int blk_mq_init_hctx(struct request_queue *q,
3648 struct blk_mq_tag_set *set,
3649 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3651 hctx->queue_num = hctx_idx;
3653 if (!(hctx->flags & BLK_MQ_F_STACKING))
3654 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3655 &hctx->cpuhp_online);
3656 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3658 hctx->tags = set->tags[hctx_idx];
3660 if (set->ops->init_hctx &&
3661 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3662 goto unregister_cpu_notifier;
3664 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3668 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3674 if (set->ops->exit_request)
3675 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3677 if (set->ops->exit_hctx)
3678 set->ops->exit_hctx(hctx, hctx_idx);
3679 unregister_cpu_notifier:
3680 blk_mq_remove_cpuhp(hctx);
3684 static struct blk_mq_hw_ctx *
3685 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3688 struct blk_mq_hw_ctx *hctx;
3689 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3691 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3693 goto fail_alloc_hctx;
3695 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3698 atomic_set(&hctx->nr_active, 0);
3699 if (node == NUMA_NO_NODE)
3700 node = set->numa_node;
3701 hctx->numa_node = node;
3703 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3704 spin_lock_init(&hctx->lock);
3705 INIT_LIST_HEAD(&hctx->dispatch);
3707 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3709 INIT_LIST_HEAD(&hctx->hctx_list);
3712 * Allocate space for all possible cpus to avoid allocation at
3715 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3720 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3721 gfp, node, false, false))
3725 spin_lock_init(&hctx->dispatch_wait_lock);
3726 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3727 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3729 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3733 blk_mq_hctx_kobj_init(hctx);
3738 sbitmap_free(&hctx->ctx_map);
3742 free_cpumask_var(hctx->cpumask);
3749 static void blk_mq_init_cpu_queues(struct request_queue *q,
3750 unsigned int nr_hw_queues)
3752 struct blk_mq_tag_set *set = q->tag_set;
3755 for_each_possible_cpu(i) {
3756 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3757 struct blk_mq_hw_ctx *hctx;
3761 spin_lock_init(&__ctx->lock);
3762 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3763 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3768 * Set local node, IFF we have more than one hw queue. If
3769 * not, we remain on the home node of the device
3771 for (j = 0; j < set->nr_maps; j++) {
3772 hctx = blk_mq_map_queue_type(q, j, i);
3773 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3774 hctx->numa_node = cpu_to_node(i);
3779 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3780 unsigned int hctx_idx,
3783 struct blk_mq_tags *tags;
3786 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3790 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3792 blk_mq_free_rq_map(tags);
3799 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3802 if (blk_mq_is_shared_tags(set->flags)) {
3803 set->tags[hctx_idx] = set->shared_tags;
3808 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3811 return set->tags[hctx_idx];
3814 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3815 struct blk_mq_tags *tags,
3816 unsigned int hctx_idx)
3819 blk_mq_free_rqs(set, tags, hctx_idx);
3820 blk_mq_free_rq_map(tags);
3824 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3825 unsigned int hctx_idx)
3827 if (!blk_mq_is_shared_tags(set->flags))
3828 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3830 set->tags[hctx_idx] = NULL;
3833 static void blk_mq_map_swqueue(struct request_queue *q)
3835 unsigned int j, hctx_idx;
3837 struct blk_mq_hw_ctx *hctx;
3838 struct blk_mq_ctx *ctx;
3839 struct blk_mq_tag_set *set = q->tag_set;
3841 queue_for_each_hw_ctx(q, hctx, i) {
3842 cpumask_clear(hctx->cpumask);
3844 hctx->dispatch_from = NULL;
3848 * Map software to hardware queues.
3850 * If the cpu isn't present, the cpu is mapped to first hctx.
3852 for_each_possible_cpu(i) {
3854 ctx = per_cpu_ptr(q->queue_ctx, i);
3855 for (j = 0; j < set->nr_maps; j++) {
3856 if (!set->map[j].nr_queues) {
3857 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3858 HCTX_TYPE_DEFAULT, i);
3861 hctx_idx = set->map[j].mq_map[i];
3862 /* unmapped hw queue can be remapped after CPU topo changed */
3863 if (!set->tags[hctx_idx] &&
3864 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3866 * If tags initialization fail for some hctx,
3867 * that hctx won't be brought online. In this
3868 * case, remap the current ctx to hctx[0] which
3869 * is guaranteed to always have tags allocated
3871 set->map[j].mq_map[i] = 0;
3874 hctx = blk_mq_map_queue_type(q, j, i);
3875 ctx->hctxs[j] = hctx;
3877 * If the CPU is already set in the mask, then we've
3878 * mapped this one already. This can happen if
3879 * devices share queues across queue maps.
3881 if (cpumask_test_cpu(i, hctx->cpumask))
3884 cpumask_set_cpu(i, hctx->cpumask);
3886 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3887 hctx->ctxs[hctx->nr_ctx++] = ctx;
3890 * If the nr_ctx type overflows, we have exceeded the
3891 * amount of sw queues we can support.
3893 BUG_ON(!hctx->nr_ctx);
3896 for (; j < HCTX_MAX_TYPES; j++)
3897 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3898 HCTX_TYPE_DEFAULT, i);
3901 queue_for_each_hw_ctx(q, hctx, i) {
3903 * If no software queues are mapped to this hardware queue,
3904 * disable it and free the request entries.
3906 if (!hctx->nr_ctx) {
3907 /* Never unmap queue 0. We need it as a
3908 * fallback in case of a new remap fails
3912 __blk_mq_free_map_and_rqs(set, i);
3918 hctx->tags = set->tags[i];
3919 WARN_ON(!hctx->tags);
3922 * Set the map size to the number of mapped software queues.
3923 * This is more accurate and more efficient than looping
3924 * over all possibly mapped software queues.
3926 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3929 * Initialize batch roundrobin counts
3931 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3932 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3937 * Caller needs to ensure that we're either frozen/quiesced, or that
3938 * the queue isn't live yet.
3940 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3942 struct blk_mq_hw_ctx *hctx;
3945 queue_for_each_hw_ctx(q, hctx, i) {
3947 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3949 blk_mq_tag_idle(hctx);
3950 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3955 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3958 struct request_queue *q;
3960 lockdep_assert_held(&set->tag_list_lock);
3962 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3963 blk_mq_freeze_queue(q);
3964 queue_set_hctx_shared(q, shared);
3965 blk_mq_unfreeze_queue(q);
3969 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3971 struct blk_mq_tag_set *set = q->tag_set;
3973 mutex_lock(&set->tag_list_lock);
3974 list_del(&q->tag_set_list);
3975 if (list_is_singular(&set->tag_list)) {
3976 /* just transitioned to unshared */
3977 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3978 /* update existing queue */
3979 blk_mq_update_tag_set_shared(set, false);
3981 mutex_unlock(&set->tag_list_lock);
3982 INIT_LIST_HEAD(&q->tag_set_list);
3985 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3986 struct request_queue *q)
3988 mutex_lock(&set->tag_list_lock);
3991 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3993 if (!list_empty(&set->tag_list) &&
3994 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3995 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3996 /* update existing queue */
3997 blk_mq_update_tag_set_shared(set, true);
3999 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
4000 queue_set_hctx_shared(q, true);
4001 list_add_tail(&q->tag_set_list, &set->tag_list);
4003 mutex_unlock(&set->tag_list_lock);
4006 /* All allocations will be freed in release handler of q->mq_kobj */
4007 static int blk_mq_alloc_ctxs(struct request_queue *q)
4009 struct blk_mq_ctxs *ctxs;
4012 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
4016 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
4017 if (!ctxs->queue_ctx)
4020 for_each_possible_cpu(cpu) {
4021 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4025 q->mq_kobj = &ctxs->kobj;
4026 q->queue_ctx = ctxs->queue_ctx;
4035 * It is the actual release handler for mq, but we do it from
4036 * request queue's release handler for avoiding use-after-free
4037 * and headache because q->mq_kobj shouldn't have been introduced,
4038 * but we can't group ctx/kctx kobj without it.
4040 void blk_mq_release(struct request_queue *q)
4042 struct blk_mq_hw_ctx *hctx, *next;
4045 queue_for_each_hw_ctx(q, hctx, i)
4046 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4048 /* all hctx are in .unused_hctx_list now */
4049 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4050 list_del_init(&hctx->hctx_list);
4051 kobject_put(&hctx->kobj);
4054 xa_destroy(&q->hctx_table);
4057 * release .mq_kobj and sw queue's kobject now because
4058 * both share lifetime with request queue.
4060 blk_mq_sysfs_deinit(q);
4063 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4066 struct request_queue *q;
4069 q = blk_alloc_queue(set->numa_node);
4071 return ERR_PTR(-ENOMEM);
4072 q->queuedata = queuedata;
4073 ret = blk_mq_init_allocated_queue(set, q);
4076 return ERR_PTR(ret);
4081 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4083 return blk_mq_init_queue_data(set, NULL);
4085 EXPORT_SYMBOL(blk_mq_init_queue);
4088 * blk_mq_destroy_queue - shutdown a request queue
4089 * @q: request queue to shutdown
4091 * This shuts down a request queue allocated by blk_mq_init_queue(). All future
4092 * requests will be failed with -ENODEV. The caller is responsible for dropping
4093 * the reference from blk_mq_init_queue() by calling blk_put_queue().
4095 * Context: can sleep
4097 void blk_mq_destroy_queue(struct request_queue *q)
4099 WARN_ON_ONCE(!queue_is_mq(q));
4100 WARN_ON_ONCE(blk_queue_registered(q));
4104 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4105 blk_queue_start_drain(q);
4106 blk_mq_freeze_queue_wait(q);
4109 blk_mq_cancel_work_sync(q);
4110 blk_mq_exit_queue(q);
4112 EXPORT_SYMBOL(blk_mq_destroy_queue);
4114 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4115 struct lock_class_key *lkclass)
4117 struct request_queue *q;
4118 struct gendisk *disk;
4120 q = blk_mq_init_queue_data(set, queuedata);
4124 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4126 blk_mq_destroy_queue(q);
4128 return ERR_PTR(-ENOMEM);
4130 set_bit(GD_OWNS_QUEUE, &disk->state);
4133 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4135 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4136 struct lock_class_key *lkclass)
4138 struct gendisk *disk;
4140 if (!blk_get_queue(q))
4142 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4147 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4149 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4150 struct blk_mq_tag_set *set, struct request_queue *q,
4151 int hctx_idx, int node)
4153 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4155 /* reuse dead hctx first */
4156 spin_lock(&q->unused_hctx_lock);
4157 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4158 if (tmp->numa_node == node) {
4164 list_del_init(&hctx->hctx_list);
4165 spin_unlock(&q->unused_hctx_lock);
4168 hctx = blk_mq_alloc_hctx(q, set, node);
4172 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4178 kobject_put(&hctx->kobj);
4183 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4184 struct request_queue *q)
4186 struct blk_mq_hw_ctx *hctx;
4189 /* protect against switching io scheduler */
4190 mutex_lock(&q->sysfs_lock);
4191 for (i = 0; i < set->nr_hw_queues; i++) {
4193 int node = blk_mq_get_hctx_node(set, i);
4194 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4197 old_node = old_hctx->numa_node;
4198 blk_mq_exit_hctx(q, set, old_hctx, i);
4201 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4204 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4206 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4207 WARN_ON_ONCE(!hctx);
4211 * Increasing nr_hw_queues fails. Free the newly allocated
4212 * hctxs and keep the previous q->nr_hw_queues.
4214 if (i != set->nr_hw_queues) {
4215 j = q->nr_hw_queues;
4218 q->nr_hw_queues = set->nr_hw_queues;
4221 xa_for_each_start(&q->hctx_table, j, hctx, j)
4222 blk_mq_exit_hctx(q, set, hctx, j);
4223 mutex_unlock(&q->sysfs_lock);
4226 static void blk_mq_update_poll_flag(struct request_queue *q)
4228 struct blk_mq_tag_set *set = q->tag_set;
4230 if (set->nr_maps > HCTX_TYPE_POLL &&
4231 set->map[HCTX_TYPE_POLL].nr_queues)
4232 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4234 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4237 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4238 struct request_queue *q)
4240 /* mark the queue as mq asap */
4241 q->mq_ops = set->ops;
4243 if (blk_mq_alloc_ctxs(q))
4246 /* init q->mq_kobj and sw queues' kobjects */
4247 blk_mq_sysfs_init(q);
4249 INIT_LIST_HEAD(&q->unused_hctx_list);
4250 spin_lock_init(&q->unused_hctx_lock);
4252 xa_init(&q->hctx_table);
4254 blk_mq_realloc_hw_ctxs(set, q);
4255 if (!q->nr_hw_queues)
4258 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4259 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4263 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4264 blk_mq_update_poll_flag(q);
4266 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4267 INIT_LIST_HEAD(&q->flush_list);
4268 INIT_LIST_HEAD(&q->requeue_list);
4269 spin_lock_init(&q->requeue_lock);
4271 q->nr_requests = set->queue_depth;
4273 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4274 blk_mq_add_queue_tag_set(set, q);
4275 blk_mq_map_swqueue(q);
4284 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4286 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4287 void blk_mq_exit_queue(struct request_queue *q)
4289 struct blk_mq_tag_set *set = q->tag_set;
4291 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4292 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4293 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4294 blk_mq_del_queue_tag_set(q);
4297 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4301 if (blk_mq_is_shared_tags(set->flags)) {
4302 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4305 if (!set->shared_tags)
4309 for (i = 0; i < set->nr_hw_queues; i++) {
4310 if (!__blk_mq_alloc_map_and_rqs(set, i))
4319 __blk_mq_free_map_and_rqs(set, i);
4321 if (blk_mq_is_shared_tags(set->flags)) {
4322 blk_mq_free_map_and_rqs(set, set->shared_tags,
4323 BLK_MQ_NO_HCTX_IDX);
4330 * Allocate the request maps associated with this tag_set. Note that this
4331 * may reduce the depth asked for, if memory is tight. set->queue_depth
4332 * will be updated to reflect the allocated depth.
4334 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4339 depth = set->queue_depth;
4341 err = __blk_mq_alloc_rq_maps(set);
4345 set->queue_depth >>= 1;
4346 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4350 } while (set->queue_depth);
4352 if (!set->queue_depth || err) {
4353 pr_err("blk-mq: failed to allocate request map\n");
4357 if (depth != set->queue_depth)
4358 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4359 depth, set->queue_depth);
4364 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4367 * blk_mq_map_queues() and multiple .map_queues() implementations
4368 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4369 * number of hardware queues.
4371 if (set->nr_maps == 1)
4372 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4374 if (set->ops->map_queues && !is_kdump_kernel()) {
4378 * transport .map_queues is usually done in the following
4381 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4382 * mask = get_cpu_mask(queue)
4383 * for_each_cpu(cpu, mask)
4384 * set->map[x].mq_map[cpu] = queue;
4387 * When we need to remap, the table has to be cleared for
4388 * killing stale mapping since one CPU may not be mapped
4391 for (i = 0; i < set->nr_maps; i++)
4392 blk_mq_clear_mq_map(&set->map[i]);
4394 set->ops->map_queues(set);
4396 BUG_ON(set->nr_maps > 1);
4397 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4401 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4402 int new_nr_hw_queues)
4404 struct blk_mq_tags **new_tags;
4406 if (set->nr_hw_queues >= new_nr_hw_queues)
4409 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4410 GFP_KERNEL, set->numa_node);
4415 memcpy(new_tags, set->tags, set->nr_hw_queues *
4416 sizeof(*set->tags));
4418 set->tags = new_tags;
4420 set->nr_hw_queues = new_nr_hw_queues;
4425 * Alloc a tag set to be associated with one or more request queues.
4426 * May fail with EINVAL for various error conditions. May adjust the
4427 * requested depth down, if it's too large. In that case, the set
4428 * value will be stored in set->queue_depth.
4430 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4434 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4436 if (!set->nr_hw_queues)
4438 if (!set->queue_depth)
4440 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4443 if (!set->ops->queue_rq)
4446 if (!set->ops->get_budget ^ !set->ops->put_budget)
4449 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4450 pr_info("blk-mq: reduced tag depth to %u\n",
4452 set->queue_depth = BLK_MQ_MAX_DEPTH;
4457 else if (set->nr_maps > HCTX_MAX_TYPES)
4461 * If a crashdump is active, then we are potentially in a very
4462 * memory constrained environment. Limit us to 1 queue and
4463 * 64 tags to prevent using too much memory.
4465 if (is_kdump_kernel()) {
4466 set->nr_hw_queues = 1;
4468 set->queue_depth = min(64U, set->queue_depth);
4471 * There is no use for more h/w queues than cpus if we just have
4474 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4475 set->nr_hw_queues = nr_cpu_ids;
4477 if (set->flags & BLK_MQ_F_BLOCKING) {
4478 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4481 ret = init_srcu_struct(set->srcu);
4487 set->tags = kcalloc_node(set->nr_hw_queues,
4488 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4491 goto out_cleanup_srcu;
4493 for (i = 0; i < set->nr_maps; i++) {
4494 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4495 sizeof(set->map[i].mq_map[0]),
4496 GFP_KERNEL, set->numa_node);
4497 if (!set->map[i].mq_map)
4498 goto out_free_mq_map;
4499 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4502 blk_mq_update_queue_map(set);
4504 ret = blk_mq_alloc_set_map_and_rqs(set);
4506 goto out_free_mq_map;
4508 mutex_init(&set->tag_list_lock);
4509 INIT_LIST_HEAD(&set->tag_list);
4514 for (i = 0; i < set->nr_maps; i++) {
4515 kfree(set->map[i].mq_map);
4516 set->map[i].mq_map = NULL;
4521 if (set->flags & BLK_MQ_F_BLOCKING)
4522 cleanup_srcu_struct(set->srcu);
4524 if (set->flags & BLK_MQ_F_BLOCKING)
4528 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4530 /* allocate and initialize a tagset for a simple single-queue device */
4531 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4532 const struct blk_mq_ops *ops, unsigned int queue_depth,
4533 unsigned int set_flags)
4535 memset(set, 0, sizeof(*set));
4537 set->nr_hw_queues = 1;
4539 set->queue_depth = queue_depth;
4540 set->numa_node = NUMA_NO_NODE;
4541 set->flags = set_flags;
4542 return blk_mq_alloc_tag_set(set);
4544 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4546 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4550 for (i = 0; i < set->nr_hw_queues; i++)
4551 __blk_mq_free_map_and_rqs(set, i);
4553 if (blk_mq_is_shared_tags(set->flags)) {
4554 blk_mq_free_map_and_rqs(set, set->shared_tags,
4555 BLK_MQ_NO_HCTX_IDX);
4558 for (j = 0; j < set->nr_maps; j++) {
4559 kfree(set->map[j].mq_map);
4560 set->map[j].mq_map = NULL;
4565 if (set->flags & BLK_MQ_F_BLOCKING) {
4566 cleanup_srcu_struct(set->srcu);
4570 EXPORT_SYMBOL(blk_mq_free_tag_set);
4572 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4574 struct blk_mq_tag_set *set = q->tag_set;
4575 struct blk_mq_hw_ctx *hctx;
4582 if (q->nr_requests == nr)
4585 blk_mq_freeze_queue(q);
4586 blk_mq_quiesce_queue(q);
4589 queue_for_each_hw_ctx(q, hctx, i) {
4593 * If we're using an MQ scheduler, just update the scheduler
4594 * queue depth. This is similar to what the old code would do.
4596 if (hctx->sched_tags) {
4597 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4600 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4605 if (q->elevator && q->elevator->type->ops.depth_updated)
4606 q->elevator->type->ops.depth_updated(hctx);
4609 q->nr_requests = nr;
4610 if (blk_mq_is_shared_tags(set->flags)) {
4612 blk_mq_tag_update_sched_shared_tags(q);
4614 blk_mq_tag_resize_shared_tags(set, nr);
4618 blk_mq_unquiesce_queue(q);
4619 blk_mq_unfreeze_queue(q);
4625 * request_queue and elevator_type pair.
4626 * It is just used by __blk_mq_update_nr_hw_queues to cache
4627 * the elevator_type associated with a request_queue.
4629 struct blk_mq_qe_pair {
4630 struct list_head node;
4631 struct request_queue *q;
4632 struct elevator_type *type;
4636 * Cache the elevator_type in qe pair list and switch the
4637 * io scheduler to 'none'
4639 static bool blk_mq_elv_switch_none(struct list_head *head,
4640 struct request_queue *q)
4642 struct blk_mq_qe_pair *qe;
4644 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4648 /* q->elevator needs protection from ->sysfs_lock */
4649 mutex_lock(&q->sysfs_lock);
4651 /* the check has to be done with holding sysfs_lock */
4657 INIT_LIST_HEAD(&qe->node);
4659 qe->type = q->elevator->type;
4660 /* keep a reference to the elevator module as we'll switch back */
4661 __elevator_get(qe->type);
4662 list_add(&qe->node, head);
4663 elevator_disable(q);
4665 mutex_unlock(&q->sysfs_lock);
4670 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4671 struct request_queue *q)
4673 struct blk_mq_qe_pair *qe;
4675 list_for_each_entry(qe, head, node)
4682 static void blk_mq_elv_switch_back(struct list_head *head,
4683 struct request_queue *q)
4685 struct blk_mq_qe_pair *qe;
4686 struct elevator_type *t;
4688 qe = blk_lookup_qe_pair(head, q);
4692 list_del(&qe->node);
4695 mutex_lock(&q->sysfs_lock);
4696 elevator_switch(q, t);
4697 /* drop the reference acquired in blk_mq_elv_switch_none */
4699 mutex_unlock(&q->sysfs_lock);
4702 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4705 struct request_queue *q;
4707 int prev_nr_hw_queues;
4709 lockdep_assert_held(&set->tag_list_lock);
4711 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4712 nr_hw_queues = nr_cpu_ids;
4713 if (nr_hw_queues < 1)
4715 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4718 list_for_each_entry(q, &set->tag_list, tag_set_list)
4719 blk_mq_freeze_queue(q);
4721 * Switch IO scheduler to 'none', cleaning up the data associated
4722 * with the previous scheduler. We will switch back once we are done
4723 * updating the new sw to hw queue mappings.
4725 list_for_each_entry(q, &set->tag_list, tag_set_list)
4726 if (!blk_mq_elv_switch_none(&head, q))
4729 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4730 blk_mq_debugfs_unregister_hctxs(q);
4731 blk_mq_sysfs_unregister_hctxs(q);
4734 prev_nr_hw_queues = set->nr_hw_queues;
4735 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4739 blk_mq_update_queue_map(set);
4740 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4741 blk_mq_realloc_hw_ctxs(set, q);
4742 blk_mq_update_poll_flag(q);
4743 if (q->nr_hw_queues != set->nr_hw_queues) {
4744 int i = prev_nr_hw_queues;
4746 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4747 nr_hw_queues, prev_nr_hw_queues);
4748 for (; i < set->nr_hw_queues; i++)
4749 __blk_mq_free_map_and_rqs(set, i);
4751 set->nr_hw_queues = prev_nr_hw_queues;
4752 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4755 blk_mq_map_swqueue(q);
4759 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4760 blk_mq_sysfs_register_hctxs(q);
4761 blk_mq_debugfs_register_hctxs(q);
4765 list_for_each_entry(q, &set->tag_list, tag_set_list)
4766 blk_mq_elv_switch_back(&head, q);
4768 list_for_each_entry(q, &set->tag_list, tag_set_list)
4769 blk_mq_unfreeze_queue(q);
4772 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4774 mutex_lock(&set->tag_list_lock);
4775 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4776 mutex_unlock(&set->tag_list_lock);
4778 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4780 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
4781 struct io_comp_batch *iob, unsigned int flags)
4783 long state = get_current_state();
4787 ret = q->mq_ops->poll(hctx, iob);
4789 __set_current_state(TASK_RUNNING);
4793 if (signal_pending_state(state, current))
4794 __set_current_state(TASK_RUNNING);
4795 if (task_is_running(current))
4798 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4801 } while (!need_resched());
4803 __set_current_state(TASK_RUNNING);
4807 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
4808 struct io_comp_batch *iob, unsigned int flags)
4810 struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie);
4812 return blk_hctx_poll(q, hctx, iob, flags);
4815 int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
4816 unsigned int poll_flags)
4818 struct request_queue *q = rq->q;
4821 if (!blk_rq_is_poll(rq))
4823 if (!percpu_ref_tryget(&q->q_usage_counter))
4826 ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags);
4831 EXPORT_SYMBOL_GPL(blk_rq_poll);
4833 unsigned int blk_mq_rq_cpu(struct request *rq)
4835 return rq->mq_ctx->cpu;
4837 EXPORT_SYMBOL(blk_mq_rq_cpu);
4839 void blk_mq_cancel_work_sync(struct request_queue *q)
4841 struct blk_mq_hw_ctx *hctx;
4844 cancel_delayed_work_sync(&q->requeue_work);
4846 queue_for_each_hw_ctx(q, hctx, i)
4847 cancel_delayed_work_sync(&hctx->run_work);
4850 static int __init blk_mq_init(void)
4854 for_each_possible_cpu(i)
4855 init_llist_head(&per_cpu(blk_cpu_done, i));
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);