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/topology.h>
25 #include <linux/sched/signal.h>
26 #include <linux/delay.h>
27 #include <linux/crash_dump.h>
28 #include <linux/prefetch.h>
29 #include <linux/blk-crypto.h>
30 #include <linux/part_stat.h>
32 #include <trace/events/block.h>
34 #include <linux/t10-pi.h>
37 #include "blk-mq-debugfs.h"
40 #include "blk-mq-sched.h"
41 #include "blk-rq-qos.h"
43 static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
44 static DEFINE_PER_CPU(call_single_data_t, blk_cpu_csd);
46 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags);
47 static void blk_mq_request_bypass_insert(struct request *rq,
49 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
50 struct list_head *list);
51 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
52 struct io_comp_batch *iob, unsigned int flags);
55 * Check if any of the ctx, dispatch list or elevator
56 * have pending work in this hardware queue.
58 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
60 return !list_empty_careful(&hctx->dispatch) ||
61 sbitmap_any_bit_set(&hctx->ctx_map) ||
62 blk_mq_sched_has_work(hctx);
66 * Mark this ctx as having pending work in this hardware queue
68 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
69 struct blk_mq_ctx *ctx)
71 const int bit = ctx->index_hw[hctx->type];
73 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
74 sbitmap_set_bit(&hctx->ctx_map, bit);
77 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
78 struct blk_mq_ctx *ctx)
80 const int bit = ctx->index_hw[hctx->type];
82 sbitmap_clear_bit(&hctx->ctx_map, bit);
86 struct block_device *part;
87 unsigned int inflight[2];
90 static bool blk_mq_check_inflight(struct request *rq, void *priv)
92 struct mq_inflight *mi = priv;
94 if (rq->part && blk_do_io_stat(rq) &&
95 (!mi->part->bd_partno || rq->part == mi->part) &&
96 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
97 mi->inflight[rq_data_dir(rq)]++;
102 unsigned int blk_mq_in_flight(struct request_queue *q,
103 struct block_device *part)
105 struct mq_inflight mi = { .part = part };
107 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
109 return mi.inflight[0] + mi.inflight[1];
112 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
113 unsigned int inflight[2])
115 struct mq_inflight mi = { .part = part };
117 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
118 inflight[0] = mi.inflight[0];
119 inflight[1] = mi.inflight[1];
122 void blk_freeze_queue_start(struct request_queue *q)
124 mutex_lock(&q->mq_freeze_lock);
125 if (++q->mq_freeze_depth == 1) {
126 percpu_ref_kill(&q->q_usage_counter);
127 mutex_unlock(&q->mq_freeze_lock);
129 blk_mq_run_hw_queues(q, false);
131 mutex_unlock(&q->mq_freeze_lock);
134 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
136 void blk_mq_freeze_queue_wait(struct request_queue *q)
138 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
140 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
142 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
143 unsigned long timeout)
145 return wait_event_timeout(q->mq_freeze_wq,
146 percpu_ref_is_zero(&q->q_usage_counter),
149 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
152 * Guarantee no request is in use, so we can change any data structure of
153 * the queue afterward.
155 void blk_freeze_queue(struct request_queue *q)
158 * In the !blk_mq case we are only calling this to kill the
159 * q_usage_counter, otherwise this increases the freeze depth
160 * and waits for it to return to zero. For this reason there is
161 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
162 * exported to drivers as the only user for unfreeze is blk_mq.
164 blk_freeze_queue_start(q);
165 blk_mq_freeze_queue_wait(q);
168 void blk_mq_freeze_queue(struct request_queue *q)
171 * ...just an alias to keep freeze and unfreeze actions balanced
172 * in the blk_mq_* namespace
176 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
178 void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
180 mutex_lock(&q->mq_freeze_lock);
182 q->q_usage_counter.data->force_atomic = true;
183 q->mq_freeze_depth--;
184 WARN_ON_ONCE(q->mq_freeze_depth < 0);
185 if (!q->mq_freeze_depth) {
186 percpu_ref_resurrect(&q->q_usage_counter);
187 wake_up_all(&q->mq_freeze_wq);
189 mutex_unlock(&q->mq_freeze_lock);
192 void blk_mq_unfreeze_queue(struct request_queue *q)
194 __blk_mq_unfreeze_queue(q, false);
196 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
199 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
200 * mpt3sas driver such that this function can be removed.
202 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
206 spin_lock_irqsave(&q->queue_lock, flags);
207 if (!q->quiesce_depth++)
208 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
209 spin_unlock_irqrestore(&q->queue_lock, flags);
211 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
214 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
215 * @set: tag_set to wait on
217 * Note: it is driver's responsibility for making sure that quiesce has
218 * been started on or more of the request_queues of the tag_set. This
219 * function only waits for the quiesce on those request_queues that had
220 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
222 void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
224 if (set->flags & BLK_MQ_F_BLOCKING)
225 synchronize_srcu(set->srcu);
229 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
232 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
235 * Note: this function does not prevent that the struct request end_io()
236 * callback function is invoked. Once this function is returned, we make
237 * sure no dispatch can happen until the queue is unquiesced via
238 * blk_mq_unquiesce_queue().
240 void blk_mq_quiesce_queue(struct request_queue *q)
242 blk_mq_quiesce_queue_nowait(q);
243 /* nothing to wait for non-mq queues */
245 blk_mq_wait_quiesce_done(q->tag_set);
247 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
250 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
253 * This function recovers queue into the state before quiescing
254 * which is done by blk_mq_quiesce_queue.
256 void blk_mq_unquiesce_queue(struct request_queue *q)
259 bool run_queue = false;
261 spin_lock_irqsave(&q->queue_lock, flags);
262 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
264 } else if (!--q->quiesce_depth) {
265 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
268 spin_unlock_irqrestore(&q->queue_lock, flags);
270 /* dispatch requests which are inserted during quiescing */
272 blk_mq_run_hw_queues(q, true);
274 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
276 void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
278 struct request_queue *q;
280 mutex_lock(&set->tag_list_lock);
281 list_for_each_entry(q, &set->tag_list, tag_set_list) {
282 if (!blk_queue_skip_tagset_quiesce(q))
283 blk_mq_quiesce_queue_nowait(q);
285 blk_mq_wait_quiesce_done(set);
286 mutex_unlock(&set->tag_list_lock);
288 EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
290 void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
292 struct request_queue *q;
294 mutex_lock(&set->tag_list_lock);
295 list_for_each_entry(q, &set->tag_list, tag_set_list) {
296 if (!blk_queue_skip_tagset_quiesce(q))
297 blk_mq_unquiesce_queue(q);
299 mutex_unlock(&set->tag_list_lock);
301 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
303 void blk_mq_wake_waiters(struct request_queue *q)
305 struct blk_mq_hw_ctx *hctx;
308 queue_for_each_hw_ctx(q, hctx, i)
309 if (blk_mq_hw_queue_mapped(hctx))
310 blk_mq_tag_wakeup_all(hctx->tags, true);
313 void blk_rq_init(struct request_queue *q, struct request *rq)
315 memset(rq, 0, sizeof(*rq));
317 INIT_LIST_HEAD(&rq->queuelist);
319 rq->__sector = (sector_t) -1;
320 INIT_HLIST_NODE(&rq->hash);
321 RB_CLEAR_NODE(&rq->rb_node);
322 rq->tag = BLK_MQ_NO_TAG;
323 rq->internal_tag = BLK_MQ_NO_TAG;
324 rq->start_time_ns = blk_time_get_ns();
326 blk_crypto_rq_set_defaults(rq);
328 EXPORT_SYMBOL(blk_rq_init);
330 /* Set start and alloc time when the allocated request is actually used */
331 static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns)
333 if (blk_mq_need_time_stamp(rq))
334 rq->start_time_ns = blk_time_get_ns();
336 rq->start_time_ns = 0;
338 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
339 if (blk_queue_rq_alloc_time(rq->q))
340 rq->alloc_time_ns = alloc_time_ns ?: rq->start_time_ns;
342 rq->alloc_time_ns = 0;
346 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
347 struct blk_mq_tags *tags, unsigned int tag)
349 struct blk_mq_ctx *ctx = data->ctx;
350 struct blk_mq_hw_ctx *hctx = data->hctx;
351 struct request_queue *q = data->q;
352 struct request *rq = tags->static_rqs[tag];
357 rq->cmd_flags = data->cmd_flags;
359 if (data->flags & BLK_MQ_REQ_PM)
360 data->rq_flags |= RQF_PM;
361 if (blk_queue_io_stat(q))
362 data->rq_flags |= RQF_IO_STAT;
363 rq->rq_flags = data->rq_flags;
365 if (data->rq_flags & RQF_SCHED_TAGS) {
366 rq->tag = BLK_MQ_NO_TAG;
367 rq->internal_tag = tag;
370 rq->internal_tag = BLK_MQ_NO_TAG;
375 rq->io_start_time_ns = 0;
376 rq->stats_sectors = 0;
377 rq->nr_phys_segments = 0;
378 #if defined(CONFIG_BLK_DEV_INTEGRITY)
379 rq->nr_integrity_segments = 0;
382 rq->end_io_data = NULL;
384 blk_crypto_rq_set_defaults(rq);
385 INIT_LIST_HEAD(&rq->queuelist);
386 /* tag was already set */
387 WRITE_ONCE(rq->deadline, 0);
390 if (rq->rq_flags & RQF_USE_SCHED) {
391 struct elevator_queue *e = data->q->elevator;
393 INIT_HLIST_NODE(&rq->hash);
394 RB_CLEAR_NODE(&rq->rb_node);
396 if (e->type->ops.prepare_request)
397 e->type->ops.prepare_request(rq);
403 static inline struct request *
404 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data)
406 unsigned int tag, tag_offset;
407 struct blk_mq_tags *tags;
409 unsigned long tag_mask;
412 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
413 if (unlikely(!tag_mask))
416 tags = blk_mq_tags_from_data(data);
417 for (i = 0; tag_mask; i++) {
418 if (!(tag_mask & (1UL << i)))
420 tag = tag_offset + i;
421 prefetch(tags->static_rqs[tag]);
422 tag_mask &= ~(1UL << i);
423 rq = blk_mq_rq_ctx_init(data, tags, tag);
424 rq_list_add(data->cached_rq, rq);
427 if (!(data->rq_flags & RQF_SCHED_TAGS))
428 blk_mq_add_active_requests(data->hctx, nr);
429 /* caller already holds a reference, add for remainder */
430 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
433 return rq_list_pop(data->cached_rq);
436 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
438 struct request_queue *q = data->q;
439 u64 alloc_time_ns = 0;
443 /* alloc_time includes depth and tag waits */
444 if (blk_queue_rq_alloc_time(q))
445 alloc_time_ns = blk_time_get_ns();
447 if (data->cmd_flags & REQ_NOWAIT)
448 data->flags |= BLK_MQ_REQ_NOWAIT;
452 * All requests use scheduler tags when an I/O scheduler is
453 * enabled for the queue.
455 data->rq_flags |= RQF_SCHED_TAGS;
458 * Flush/passthrough requests are special and go directly to the
461 if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&
462 !blk_op_is_passthrough(data->cmd_flags)) {
463 struct elevator_mq_ops *ops = &q->elevator->type->ops;
465 WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
467 data->rq_flags |= RQF_USE_SCHED;
468 if (ops->limit_depth)
469 ops->limit_depth(data->cmd_flags, data);
474 data->ctx = blk_mq_get_ctx(q);
475 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
476 if (!(data->rq_flags & RQF_SCHED_TAGS))
477 blk_mq_tag_busy(data->hctx);
479 if (data->flags & BLK_MQ_REQ_RESERVED)
480 data->rq_flags |= RQF_RESV;
483 * Try batched alloc if we want more than 1 tag.
485 if (data->nr_tags > 1) {
486 rq = __blk_mq_alloc_requests_batch(data);
488 blk_mq_rq_time_init(rq, alloc_time_ns);
495 * Waiting allocations only fail because of an inactive hctx. In that
496 * case just retry the hctx assignment and tag allocation as CPU hotplug
497 * should have migrated us to an online CPU by now.
499 tag = blk_mq_get_tag(data);
500 if (tag == BLK_MQ_NO_TAG) {
501 if (data->flags & BLK_MQ_REQ_NOWAIT)
504 * Give up the CPU and sleep for a random short time to
505 * ensure that thread using a realtime scheduling class
506 * are migrated off the CPU, and thus off the hctx that
513 if (!(data->rq_flags & RQF_SCHED_TAGS))
514 blk_mq_inc_active_requests(data->hctx);
515 rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag);
516 blk_mq_rq_time_init(rq, alloc_time_ns);
520 static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
521 struct blk_plug *plug,
523 blk_mq_req_flags_t flags)
525 struct blk_mq_alloc_data data = {
529 .nr_tags = plug->nr_ios,
530 .cached_rq = &plug->cached_rq,
534 if (blk_queue_enter(q, flags))
539 rq = __blk_mq_alloc_requests(&data);
545 static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
547 blk_mq_req_flags_t flags)
549 struct blk_plug *plug = current->plug;
555 if (rq_list_empty(plug->cached_rq)) {
556 if (plug->nr_ios == 1)
558 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
562 rq = rq_list_peek(&plug->cached_rq);
563 if (!rq || rq->q != q)
566 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
568 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
571 plug->cached_rq = rq_list_next(rq);
572 blk_mq_rq_time_init(rq, 0);
576 INIT_LIST_HEAD(&rq->queuelist);
580 struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
581 blk_mq_req_flags_t flags)
585 rq = blk_mq_alloc_cached_request(q, opf, flags);
587 struct blk_mq_alloc_data data = {
595 ret = blk_queue_enter(q, flags);
599 rq = __blk_mq_alloc_requests(&data);
604 rq->__sector = (sector_t) -1;
605 rq->bio = rq->biotail = NULL;
609 return ERR_PTR(-EWOULDBLOCK);
611 EXPORT_SYMBOL(blk_mq_alloc_request);
613 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
614 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
616 struct blk_mq_alloc_data data = {
622 u64 alloc_time_ns = 0;
628 /* alloc_time includes depth and tag waits */
629 if (blk_queue_rq_alloc_time(q))
630 alloc_time_ns = blk_time_get_ns();
633 * If the tag allocator sleeps we could get an allocation for a
634 * different hardware context. No need to complicate the low level
635 * allocator for this for the rare use case of a command tied to
638 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
639 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
640 return ERR_PTR(-EINVAL);
642 if (hctx_idx >= q->nr_hw_queues)
643 return ERR_PTR(-EIO);
645 ret = blk_queue_enter(q, flags);
650 * Check if the hardware context is actually mapped to anything.
651 * If not tell the caller that it should skip this queue.
654 data.hctx = xa_load(&q->hctx_table, hctx_idx);
655 if (!blk_mq_hw_queue_mapped(data.hctx))
657 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
658 if (cpu >= nr_cpu_ids)
660 data.ctx = __blk_mq_get_ctx(q, cpu);
663 data.rq_flags |= RQF_SCHED_TAGS;
665 blk_mq_tag_busy(data.hctx);
667 if (flags & BLK_MQ_REQ_RESERVED)
668 data.rq_flags |= RQF_RESV;
671 tag = blk_mq_get_tag(&data);
672 if (tag == BLK_MQ_NO_TAG)
674 if (!(data.rq_flags & RQF_SCHED_TAGS))
675 blk_mq_inc_active_requests(data.hctx);
676 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag);
677 blk_mq_rq_time_init(rq, alloc_time_ns);
679 rq->__sector = (sector_t) -1;
680 rq->bio = rq->biotail = NULL;
687 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
689 static void blk_mq_finish_request(struct request *rq)
691 struct request_queue *q = rq->q;
693 if (rq->rq_flags & RQF_USE_SCHED) {
694 q->elevator->type->ops.finish_request(rq);
696 * For postflush request that may need to be
697 * completed twice, we should clear this flag
698 * to avoid double finish_request() on the rq.
700 rq->rq_flags &= ~RQF_USE_SCHED;
704 static void __blk_mq_free_request(struct request *rq)
706 struct request_queue *q = rq->q;
707 struct blk_mq_ctx *ctx = rq->mq_ctx;
708 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
709 const int sched_tag = rq->internal_tag;
711 blk_crypto_free_request(rq);
712 blk_pm_mark_last_busy(rq);
715 if (rq->tag != BLK_MQ_NO_TAG) {
716 blk_mq_dec_active_requests(hctx);
717 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
719 if (sched_tag != BLK_MQ_NO_TAG)
720 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
721 blk_mq_sched_restart(hctx);
725 void blk_mq_free_request(struct request *rq)
727 struct request_queue *q = rq->q;
729 blk_mq_finish_request(rq);
731 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
732 laptop_io_completion(q->disk->bdi);
736 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
737 if (req_ref_put_and_test(rq))
738 __blk_mq_free_request(rq);
740 EXPORT_SYMBOL_GPL(blk_mq_free_request);
742 void blk_mq_free_plug_rqs(struct blk_plug *plug)
746 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
747 blk_mq_free_request(rq);
750 void blk_dump_rq_flags(struct request *rq, char *msg)
752 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
753 rq->q->disk ? rq->q->disk->disk_name : "?",
754 (__force unsigned long long) rq->cmd_flags);
756 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
757 (unsigned long long)blk_rq_pos(rq),
758 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
759 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
760 rq->bio, rq->biotail, blk_rq_bytes(rq));
762 EXPORT_SYMBOL(blk_dump_rq_flags);
764 static void req_bio_endio(struct request *rq, struct bio *bio,
765 unsigned int nbytes, blk_status_t error)
767 if (unlikely(error)) {
768 bio->bi_status = error;
769 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
771 * Partial zone append completions cannot be supported as the
772 * BIO fragments may end up not being written sequentially.
773 * For such case, force the completed nbytes to be equal to
774 * the BIO size so that bio_advance() sets the BIO remaining
775 * size to 0 and we end up calling bio_endio() before returning.
777 if (bio->bi_iter.bi_size != nbytes) {
778 bio->bi_status = BLK_STS_IOERR;
779 nbytes = bio->bi_iter.bi_size;
781 bio->bi_iter.bi_sector = rq->__sector;
785 bio_advance(bio, nbytes);
787 if (unlikely(rq->rq_flags & RQF_QUIET))
788 bio_set_flag(bio, BIO_QUIET);
789 /* don't actually finish bio if it's part of flush sequence */
790 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
794 static void blk_account_io_completion(struct request *req, unsigned int bytes)
796 if (req->part && blk_do_io_stat(req)) {
797 const int sgrp = op_stat_group(req_op(req));
800 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
805 static void blk_print_req_error(struct request *req, blk_status_t status)
807 printk_ratelimited(KERN_ERR
808 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
809 "phys_seg %u prio class %u\n",
810 blk_status_to_str(status),
811 req->q->disk ? req->q->disk->disk_name : "?",
812 blk_rq_pos(req), (__force u32)req_op(req),
813 blk_op_str(req_op(req)),
814 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
815 req->nr_phys_segments,
816 IOPRIO_PRIO_CLASS(req->ioprio));
820 * Fully end IO on a request. Does not support partial completions, or
823 static void blk_complete_request(struct request *req)
825 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
826 int total_bytes = blk_rq_bytes(req);
827 struct bio *bio = req->bio;
829 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
834 #ifdef CONFIG_BLK_DEV_INTEGRITY
835 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
836 req->q->integrity.profile->complete_fn(req, total_bytes);
840 * Upper layers may call blk_crypto_evict_key() anytime after the last
841 * bio_endio(). Therefore, the keyslot must be released before that.
843 blk_crypto_rq_put_keyslot(req);
845 blk_account_io_completion(req, total_bytes);
848 struct bio *next = bio->bi_next;
850 /* Completion has already been traced */
851 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
853 if (req_op(req) == REQ_OP_ZONE_APPEND)
854 bio->bi_iter.bi_sector = req->__sector;
862 * Reset counters so that the request stacking driver
863 * can find how many bytes remain in the request
873 * blk_update_request - Complete multiple bytes without completing the request
874 * @req: the request being processed
875 * @error: block status code
876 * @nr_bytes: number of bytes to complete for @req
879 * Ends I/O on a number of bytes attached to @req, but doesn't complete
880 * the request structure even if @req doesn't have leftover.
881 * If @req has leftover, sets it up for the next range of segments.
883 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
884 * %false return from this function.
887 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
888 * except in the consistency check at the end of this function.
891 * %false - this request doesn't have any more data
892 * %true - this request has more data
894 bool blk_update_request(struct request *req, blk_status_t error,
895 unsigned int nr_bytes)
899 trace_block_rq_complete(req, error, nr_bytes);
904 #ifdef CONFIG_BLK_DEV_INTEGRITY
905 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
907 req->q->integrity.profile->complete_fn(req, nr_bytes);
911 * Upper layers may call blk_crypto_evict_key() anytime after the last
912 * bio_endio(). Therefore, the keyslot must be released before that.
914 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
915 __blk_crypto_rq_put_keyslot(req);
917 if (unlikely(error && !blk_rq_is_passthrough(req) &&
918 !(req->rq_flags & RQF_QUIET)) &&
919 !test_bit(GD_DEAD, &req->q->disk->state)) {
920 blk_print_req_error(req, error);
921 trace_block_rq_error(req, error, nr_bytes);
924 blk_account_io_completion(req, nr_bytes);
928 struct bio *bio = req->bio;
929 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
931 if (bio_bytes == bio->bi_iter.bi_size)
932 req->bio = bio->bi_next;
934 /* Completion has already been traced */
935 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
936 req_bio_endio(req, bio, bio_bytes, error);
938 total_bytes += bio_bytes;
939 nr_bytes -= bio_bytes;
950 * Reset counters so that the request stacking driver
951 * can find how many bytes remain in the request
958 req->__data_len -= total_bytes;
960 /* update sector only for requests with clear definition of sector */
961 if (!blk_rq_is_passthrough(req))
962 req->__sector += total_bytes >> 9;
964 /* mixed attributes always follow the first bio */
965 if (req->rq_flags & RQF_MIXED_MERGE) {
966 req->cmd_flags &= ~REQ_FAILFAST_MASK;
967 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
970 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
972 * If total number of sectors is less than the first segment
973 * size, something has gone terribly wrong.
975 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
976 blk_dump_rq_flags(req, "request botched");
977 req->__data_len = blk_rq_cur_bytes(req);
980 /* recalculate the number of segments */
981 req->nr_phys_segments = blk_recalc_rq_segments(req);
986 EXPORT_SYMBOL_GPL(blk_update_request);
988 static inline void blk_account_io_done(struct request *req, u64 now)
990 trace_block_io_done(req);
993 * Account IO completion. flush_rq isn't accounted as a
994 * normal IO on queueing nor completion. Accounting the
995 * containing request is enough.
997 if (blk_do_io_stat(req) && req->part &&
998 !(req->rq_flags & RQF_FLUSH_SEQ)) {
999 const int sgrp = op_stat_group(req_op(req));
1002 update_io_ticks(req->part, jiffies, true);
1003 part_stat_inc(req->part, ios[sgrp]);
1004 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
1009 static inline void blk_account_io_start(struct request *req)
1011 trace_block_io_start(req);
1013 if (blk_do_io_stat(req)) {
1015 * All non-passthrough requests are created from a bio with one
1016 * exception: when a flush command that is part of a flush sequence
1017 * generated by the state machine in blk-flush.c is cloned onto the
1018 * lower device by dm-multipath we can get here without a bio.
1021 req->part = req->bio->bi_bdev;
1023 req->part = req->q->disk->part0;
1026 update_io_ticks(req->part, jiffies, false);
1031 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1033 if (rq->rq_flags & RQF_STATS)
1034 blk_stat_add(rq, now);
1036 blk_mq_sched_completed_request(rq, now);
1037 blk_account_io_done(rq, now);
1040 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1042 if (blk_mq_need_time_stamp(rq))
1043 __blk_mq_end_request_acct(rq, blk_time_get_ns());
1045 blk_mq_finish_request(rq);
1048 rq_qos_done(rq->q, rq);
1049 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1050 blk_mq_free_request(rq);
1052 blk_mq_free_request(rq);
1055 EXPORT_SYMBOL(__blk_mq_end_request);
1057 void blk_mq_end_request(struct request *rq, blk_status_t error)
1059 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1061 __blk_mq_end_request(rq, error);
1063 EXPORT_SYMBOL(blk_mq_end_request);
1065 #define TAG_COMP_BATCH 32
1067 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1068 int *tag_array, int nr_tags)
1070 struct request_queue *q = hctx->queue;
1072 blk_mq_sub_active_requests(hctx, nr_tags);
1074 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1075 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1078 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1080 int tags[TAG_COMP_BATCH], nr_tags = 0;
1081 struct blk_mq_hw_ctx *cur_hctx = NULL;
1086 now = blk_time_get_ns();
1088 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1090 prefetch(rq->rq_next);
1092 blk_complete_request(rq);
1094 __blk_mq_end_request_acct(rq, now);
1096 blk_mq_finish_request(rq);
1098 rq_qos_done(rq->q, rq);
1101 * If end_io handler returns NONE, then it still has
1102 * ownership of the request.
1104 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1107 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1108 if (!req_ref_put_and_test(rq))
1111 blk_crypto_free_request(rq);
1112 blk_pm_mark_last_busy(rq);
1114 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1116 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1118 cur_hctx = rq->mq_hctx;
1120 tags[nr_tags++] = rq->tag;
1124 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1126 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1128 static void blk_complete_reqs(struct llist_head *list)
1130 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1131 struct request *rq, *next;
1133 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1134 rq->q->mq_ops->complete(rq);
1137 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1139 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1142 static int blk_softirq_cpu_dead(unsigned int cpu)
1144 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1148 static void __blk_mq_complete_request_remote(void *data)
1150 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1153 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1155 int cpu = raw_smp_processor_id();
1157 if (!IS_ENABLED(CONFIG_SMP) ||
1158 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1161 * With force threaded interrupts enabled, raising softirq from an SMP
1162 * function call will always result in waking the ksoftirqd thread.
1163 * This is probably worse than completing the request on a different
1166 if (force_irqthreads())
1169 /* same CPU or cache domain and capacity? Complete locally */
1170 if (cpu == rq->mq_ctx->cpu ||
1171 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1172 cpus_share_cache(cpu, rq->mq_ctx->cpu) &&
1173 cpus_equal_capacity(cpu, rq->mq_ctx->cpu)))
1176 /* don't try to IPI to an offline CPU */
1177 return cpu_online(rq->mq_ctx->cpu);
1180 static void blk_mq_complete_send_ipi(struct request *rq)
1184 cpu = rq->mq_ctx->cpu;
1185 if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu)))
1186 smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu));
1189 static void blk_mq_raise_softirq(struct request *rq)
1191 struct llist_head *list;
1194 list = this_cpu_ptr(&blk_cpu_done);
1195 if (llist_add(&rq->ipi_list, list))
1196 raise_softirq(BLOCK_SOFTIRQ);
1200 bool blk_mq_complete_request_remote(struct request *rq)
1202 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1205 * For request which hctx has only one ctx mapping,
1206 * or a polled request, always complete locally,
1207 * it's pointless to redirect the completion.
1209 if ((rq->mq_hctx->nr_ctx == 1 &&
1210 rq->mq_ctx->cpu == raw_smp_processor_id()) ||
1211 rq->cmd_flags & REQ_POLLED)
1214 if (blk_mq_complete_need_ipi(rq)) {
1215 blk_mq_complete_send_ipi(rq);
1219 if (rq->q->nr_hw_queues == 1) {
1220 blk_mq_raise_softirq(rq);
1225 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1228 * blk_mq_complete_request - end I/O on a request
1229 * @rq: the request being processed
1232 * Complete a request by scheduling the ->complete_rq operation.
1234 void blk_mq_complete_request(struct request *rq)
1236 if (!blk_mq_complete_request_remote(rq))
1237 rq->q->mq_ops->complete(rq);
1239 EXPORT_SYMBOL(blk_mq_complete_request);
1242 * blk_mq_start_request - Start processing a request
1243 * @rq: Pointer to request to be started
1245 * Function used by device drivers to notify the block layer that a request
1246 * is going to be processed now, so blk layer can do proper initializations
1247 * such as starting the timeout timer.
1249 void blk_mq_start_request(struct request *rq)
1251 struct request_queue *q = rq->q;
1253 trace_block_rq_issue(rq);
1255 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) &&
1256 !blk_rq_is_passthrough(rq)) {
1257 rq->io_start_time_ns = blk_time_get_ns();
1258 rq->stats_sectors = blk_rq_sectors(rq);
1259 rq->rq_flags |= RQF_STATS;
1260 rq_qos_issue(q, rq);
1263 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1266 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1267 rq->mq_hctx->tags->rqs[rq->tag] = rq;
1269 #ifdef CONFIG_BLK_DEV_INTEGRITY
1270 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1271 q->integrity.profile->prepare_fn(rq);
1273 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1274 WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num);
1276 EXPORT_SYMBOL(blk_mq_start_request);
1279 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1280 * queues. This is important for md arrays to benefit from merging
1283 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1285 if (plug->multiple_queues)
1286 return BLK_MAX_REQUEST_COUNT * 2;
1287 return BLK_MAX_REQUEST_COUNT;
1290 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1292 struct request *last = rq_list_peek(&plug->mq_list);
1294 if (!plug->rq_count) {
1295 trace_block_plug(rq->q);
1296 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1297 (!blk_queue_nomerges(rq->q) &&
1298 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1299 blk_mq_flush_plug_list(plug, false);
1301 trace_block_plug(rq->q);
1304 if (!plug->multiple_queues && last && last->q != rq->q)
1305 plug->multiple_queues = true;
1307 * Any request allocated from sched tags can't be issued to
1308 * ->queue_rqs() directly
1310 if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))
1311 plug->has_elevator = true;
1313 rq_list_add(&plug->mq_list, rq);
1318 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1319 * @rq: request to insert
1320 * @at_head: insert request at head or tail of queue
1323 * Insert a fully prepared request at the back of the I/O scheduler queue
1324 * for execution. Don't wait for completion.
1327 * This function will invoke @done directly if the queue is dead.
1329 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1331 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1333 WARN_ON(irqs_disabled());
1334 WARN_ON(!blk_rq_is_passthrough(rq));
1336 blk_account_io_start(rq);
1339 * As plugging can be enabled for passthrough requests on a zoned
1340 * device, directly accessing the plug instead of using blk_mq_plug()
1341 * should not have any consequences.
1343 if (current->plug && !at_head) {
1344 blk_add_rq_to_plug(current->plug, rq);
1348 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1349 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
1351 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1353 struct blk_rq_wait {
1354 struct completion done;
1358 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1360 struct blk_rq_wait *wait = rq->end_io_data;
1363 complete(&wait->done);
1364 return RQ_END_IO_NONE;
1367 bool blk_rq_is_poll(struct request *rq)
1371 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1375 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1377 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1380 blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0);
1382 } while (!completion_done(wait));
1386 * blk_execute_rq - insert a request into queue for execution
1387 * @rq: request to insert
1388 * @at_head: insert request at head or tail of queue
1391 * Insert a fully prepared request at the back of the I/O scheduler queue
1392 * for execution and wait for completion.
1393 * Return: The blk_status_t result provided to blk_mq_end_request().
1395 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1397 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1398 struct blk_rq_wait wait = {
1399 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1402 WARN_ON(irqs_disabled());
1403 WARN_ON(!blk_rq_is_passthrough(rq));
1405 rq->end_io_data = &wait;
1406 rq->end_io = blk_end_sync_rq;
1408 blk_account_io_start(rq);
1409 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1410 blk_mq_run_hw_queue(hctx, false);
1412 if (blk_rq_is_poll(rq))
1413 blk_rq_poll_completion(rq, &wait.done);
1415 blk_wait_io(&wait.done);
1419 EXPORT_SYMBOL(blk_execute_rq);
1421 static void __blk_mq_requeue_request(struct request *rq)
1423 struct request_queue *q = rq->q;
1425 blk_mq_put_driver_tag(rq);
1427 trace_block_rq_requeue(rq);
1428 rq_qos_requeue(q, rq);
1430 if (blk_mq_request_started(rq)) {
1431 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1432 rq->rq_flags &= ~RQF_TIMED_OUT;
1436 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1438 struct request_queue *q = rq->q;
1439 unsigned long flags;
1441 __blk_mq_requeue_request(rq);
1443 /* this request will be re-inserted to io scheduler queue */
1444 blk_mq_sched_requeue_request(rq);
1446 spin_lock_irqsave(&q->requeue_lock, flags);
1447 list_add_tail(&rq->queuelist, &q->requeue_list);
1448 spin_unlock_irqrestore(&q->requeue_lock, flags);
1450 if (kick_requeue_list)
1451 blk_mq_kick_requeue_list(q);
1453 EXPORT_SYMBOL(blk_mq_requeue_request);
1455 static void blk_mq_requeue_work(struct work_struct *work)
1457 struct request_queue *q =
1458 container_of(work, struct request_queue, requeue_work.work);
1460 LIST_HEAD(flush_list);
1463 spin_lock_irq(&q->requeue_lock);
1464 list_splice_init(&q->requeue_list, &rq_list);
1465 list_splice_init(&q->flush_list, &flush_list);
1466 spin_unlock_irq(&q->requeue_lock);
1468 while (!list_empty(&rq_list)) {
1469 rq = list_entry(rq_list.next, struct request, queuelist);
1471 * If RQF_DONTPREP ist set, the request has been started by the
1472 * driver already and might have driver-specific data allocated
1473 * already. Insert it into the hctx dispatch list to avoid
1474 * block layer merges for the request.
1476 if (rq->rq_flags & RQF_DONTPREP) {
1477 list_del_init(&rq->queuelist);
1478 blk_mq_request_bypass_insert(rq, 0);
1480 list_del_init(&rq->queuelist);
1481 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1485 while (!list_empty(&flush_list)) {
1486 rq = list_entry(flush_list.next, struct request, queuelist);
1487 list_del_init(&rq->queuelist);
1488 blk_mq_insert_request(rq, 0);
1491 blk_mq_run_hw_queues(q, false);
1494 void blk_mq_kick_requeue_list(struct request_queue *q)
1496 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1498 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1500 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1501 unsigned long msecs)
1503 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1504 msecs_to_jiffies(msecs));
1506 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1508 static bool blk_is_flush_data_rq(struct request *rq)
1510 return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq);
1513 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1516 * If we find a request that isn't idle we know the queue is busy
1517 * as it's checked in the iter.
1518 * Return false to stop the iteration.
1520 * In case of queue quiesce, if one flush data request is completed,
1521 * don't count it as inflight given the flush sequence is suspended,
1522 * and the original flush data request is invisible to driver, just
1523 * like other pending requests because of quiesce
1525 if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) &&
1526 blk_is_flush_data_rq(rq) &&
1527 blk_mq_request_completed(rq))) {
1537 bool blk_mq_queue_inflight(struct request_queue *q)
1541 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1544 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1546 static void blk_mq_rq_timed_out(struct request *req)
1548 req->rq_flags |= RQF_TIMED_OUT;
1549 if (req->q->mq_ops->timeout) {
1550 enum blk_eh_timer_return ret;
1552 ret = req->q->mq_ops->timeout(req);
1553 if (ret == BLK_EH_DONE)
1555 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1561 struct blk_expired_data {
1562 bool has_timedout_rq;
1564 unsigned long timeout_start;
1567 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1569 unsigned long deadline;
1571 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1573 if (rq->rq_flags & RQF_TIMED_OUT)
1576 deadline = READ_ONCE(rq->deadline);
1577 if (time_after_eq(expired->timeout_start, deadline))
1580 if (expired->next == 0)
1581 expired->next = deadline;
1582 else if (time_after(expired->next, deadline))
1583 expired->next = deadline;
1587 void blk_mq_put_rq_ref(struct request *rq)
1589 if (is_flush_rq(rq)) {
1590 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1591 blk_mq_free_request(rq);
1592 } else if (req_ref_put_and_test(rq)) {
1593 __blk_mq_free_request(rq);
1597 static bool blk_mq_check_expired(struct request *rq, void *priv)
1599 struct blk_expired_data *expired = priv;
1602 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1603 * be reallocated underneath the timeout handler's processing, then
1604 * the expire check is reliable. If the request is not expired, then
1605 * it was completed and reallocated as a new request after returning
1606 * from blk_mq_check_expired().
1608 if (blk_mq_req_expired(rq, expired)) {
1609 expired->has_timedout_rq = true;
1615 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1617 struct blk_expired_data *expired = priv;
1619 if (blk_mq_req_expired(rq, expired))
1620 blk_mq_rq_timed_out(rq);
1624 static void blk_mq_timeout_work(struct work_struct *work)
1626 struct request_queue *q =
1627 container_of(work, struct request_queue, timeout_work);
1628 struct blk_expired_data expired = {
1629 .timeout_start = jiffies,
1631 struct blk_mq_hw_ctx *hctx;
1634 /* A deadlock might occur if a request is stuck requiring a
1635 * timeout at the same time a queue freeze is waiting
1636 * completion, since the timeout code would not be able to
1637 * acquire the queue reference here.
1639 * That's why we don't use blk_queue_enter here; instead, we use
1640 * percpu_ref_tryget directly, because we need to be able to
1641 * obtain a reference even in the short window between the queue
1642 * starting to freeze, by dropping the first reference in
1643 * blk_freeze_queue_start, and the moment the last request is
1644 * consumed, marked by the instant q_usage_counter reaches
1647 if (!percpu_ref_tryget(&q->q_usage_counter))
1650 /* check if there is any timed-out request */
1651 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1652 if (expired.has_timedout_rq) {
1654 * Before walking tags, we must ensure any submit started
1655 * before the current time has finished. Since the submit
1656 * uses srcu or rcu, wait for a synchronization point to
1657 * ensure all running submits have finished
1659 blk_mq_wait_quiesce_done(q->tag_set);
1662 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1665 if (expired.next != 0) {
1666 mod_timer(&q->timeout, expired.next);
1669 * Request timeouts are handled as a forward rolling timer. If
1670 * we end up here it means that no requests are pending and
1671 * also that no request has been pending for a while. Mark
1672 * each hctx as idle.
1674 queue_for_each_hw_ctx(q, hctx, i) {
1675 /* the hctx may be unmapped, so check it here */
1676 if (blk_mq_hw_queue_mapped(hctx))
1677 blk_mq_tag_idle(hctx);
1683 struct flush_busy_ctx_data {
1684 struct blk_mq_hw_ctx *hctx;
1685 struct list_head *list;
1688 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1690 struct flush_busy_ctx_data *flush_data = data;
1691 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1692 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1693 enum hctx_type type = hctx->type;
1695 spin_lock(&ctx->lock);
1696 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1697 sbitmap_clear_bit(sb, bitnr);
1698 spin_unlock(&ctx->lock);
1703 * Process software queues that have been marked busy, splicing them
1704 * to the for-dispatch
1706 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1708 struct flush_busy_ctx_data data = {
1713 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1715 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1717 struct dispatch_rq_data {
1718 struct blk_mq_hw_ctx *hctx;
1722 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1725 struct dispatch_rq_data *dispatch_data = data;
1726 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1727 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1728 enum hctx_type type = hctx->type;
1730 spin_lock(&ctx->lock);
1731 if (!list_empty(&ctx->rq_lists[type])) {
1732 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1733 list_del_init(&dispatch_data->rq->queuelist);
1734 if (list_empty(&ctx->rq_lists[type]))
1735 sbitmap_clear_bit(sb, bitnr);
1737 spin_unlock(&ctx->lock);
1739 return !dispatch_data->rq;
1742 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1743 struct blk_mq_ctx *start)
1745 unsigned off = start ? start->index_hw[hctx->type] : 0;
1746 struct dispatch_rq_data data = {
1751 __sbitmap_for_each_set(&hctx->ctx_map, off,
1752 dispatch_rq_from_ctx, &data);
1757 bool __blk_mq_alloc_driver_tag(struct request *rq)
1759 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1760 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1763 blk_mq_tag_busy(rq->mq_hctx);
1765 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1766 bt = &rq->mq_hctx->tags->breserved_tags;
1769 if (!hctx_may_queue(rq->mq_hctx, bt))
1773 tag = __sbitmap_queue_get(bt);
1774 if (tag == BLK_MQ_NO_TAG)
1777 rq->tag = tag + tag_offset;
1778 blk_mq_inc_active_requests(rq->mq_hctx);
1782 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1783 int flags, void *key)
1785 struct blk_mq_hw_ctx *hctx;
1787 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1789 spin_lock(&hctx->dispatch_wait_lock);
1790 if (!list_empty(&wait->entry)) {
1791 struct sbitmap_queue *sbq;
1793 list_del_init(&wait->entry);
1794 sbq = &hctx->tags->bitmap_tags;
1795 atomic_dec(&sbq->ws_active);
1797 spin_unlock(&hctx->dispatch_wait_lock);
1799 blk_mq_run_hw_queue(hctx, true);
1804 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1805 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1806 * restart. For both cases, take care to check the condition again after
1807 * marking us as waiting.
1809 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1812 struct sbitmap_queue *sbq;
1813 struct wait_queue_head *wq;
1814 wait_queue_entry_t *wait;
1817 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1818 !(blk_mq_is_shared_tags(hctx->flags))) {
1819 blk_mq_sched_mark_restart_hctx(hctx);
1822 * It's possible that a tag was freed in the window between the
1823 * allocation failure and adding the hardware queue to the wait
1826 * Don't clear RESTART here, someone else could have set it.
1827 * At most this will cost an extra queue run.
1829 return blk_mq_get_driver_tag(rq);
1832 wait = &hctx->dispatch_wait;
1833 if (!list_empty_careful(&wait->entry))
1836 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1837 sbq = &hctx->tags->breserved_tags;
1839 sbq = &hctx->tags->bitmap_tags;
1840 wq = &bt_wait_ptr(sbq, hctx)->wait;
1842 spin_lock_irq(&wq->lock);
1843 spin_lock(&hctx->dispatch_wait_lock);
1844 if (!list_empty(&wait->entry)) {
1845 spin_unlock(&hctx->dispatch_wait_lock);
1846 spin_unlock_irq(&wq->lock);
1850 atomic_inc(&sbq->ws_active);
1851 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1852 __add_wait_queue(wq, wait);
1855 * Add one explicit barrier since blk_mq_get_driver_tag() may
1856 * not imply barrier in case of failure.
1858 * Order adding us to wait queue and allocating driver tag.
1860 * The pair is the one implied in sbitmap_queue_wake_up() which
1861 * orders clearing sbitmap tag bits and waitqueue_active() in
1862 * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless
1864 * Otherwise, re-order of adding wait queue and getting driver tag
1865 * may cause __sbitmap_queue_wake_up() to wake up nothing because
1866 * the waitqueue_active() may not observe us in wait queue.
1871 * It's possible that a tag was freed in the window between the
1872 * allocation failure and adding the hardware queue to the wait
1875 ret = blk_mq_get_driver_tag(rq);
1877 spin_unlock(&hctx->dispatch_wait_lock);
1878 spin_unlock_irq(&wq->lock);
1883 * We got a tag, remove ourselves from the wait queue to ensure
1884 * someone else gets the wakeup.
1886 list_del_init(&wait->entry);
1887 atomic_dec(&sbq->ws_active);
1888 spin_unlock(&hctx->dispatch_wait_lock);
1889 spin_unlock_irq(&wq->lock);
1894 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1895 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1897 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1898 * - EWMA is one simple way to compute running average value
1899 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1900 * - take 4 as factor for avoiding to get too small(0) result, and this
1901 * factor doesn't matter because EWMA decreases exponentially
1903 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1907 ewma = hctx->dispatch_busy;
1912 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1914 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1915 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1917 hctx->dispatch_busy = ewma;
1920 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1922 static void blk_mq_handle_dev_resource(struct request *rq,
1923 struct list_head *list)
1925 list_add(&rq->queuelist, list);
1926 __blk_mq_requeue_request(rq);
1929 static void blk_mq_handle_zone_resource(struct request *rq,
1930 struct list_head *zone_list)
1933 * If we end up here it is because we cannot dispatch a request to a
1934 * specific zone due to LLD level zone-write locking or other zone
1935 * related resource not being available. In this case, set the request
1936 * aside in zone_list for retrying it later.
1938 list_add(&rq->queuelist, zone_list);
1939 __blk_mq_requeue_request(rq);
1942 enum prep_dispatch {
1944 PREP_DISPATCH_NO_TAG,
1945 PREP_DISPATCH_NO_BUDGET,
1948 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1951 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1952 int budget_token = -1;
1955 budget_token = blk_mq_get_dispatch_budget(rq->q);
1956 if (budget_token < 0) {
1957 blk_mq_put_driver_tag(rq);
1958 return PREP_DISPATCH_NO_BUDGET;
1960 blk_mq_set_rq_budget_token(rq, budget_token);
1963 if (!blk_mq_get_driver_tag(rq)) {
1965 * The initial allocation attempt failed, so we need to
1966 * rerun the hardware queue when a tag is freed. The
1967 * waitqueue takes care of that. If the queue is run
1968 * before we add this entry back on the dispatch list,
1969 * we'll re-run it below.
1971 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1973 * All budgets not got from this function will be put
1974 * together during handling partial dispatch
1977 blk_mq_put_dispatch_budget(rq->q, budget_token);
1978 return PREP_DISPATCH_NO_TAG;
1982 return PREP_DISPATCH_OK;
1985 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1986 static void blk_mq_release_budgets(struct request_queue *q,
1987 struct list_head *list)
1991 list_for_each_entry(rq, list, queuelist) {
1992 int budget_token = blk_mq_get_rq_budget_token(rq);
1994 if (budget_token >= 0)
1995 blk_mq_put_dispatch_budget(q, budget_token);
2000 * blk_mq_commit_rqs will notify driver using bd->last that there is no
2001 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
2003 * Attention, we should explicitly call this in unusual cases:
2004 * 1) did not queue everything initially scheduled to queue
2005 * 2) the last attempt to queue a request failed
2007 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
2010 if (hctx->queue->mq_ops->commit_rqs && queued) {
2011 trace_block_unplug(hctx->queue, queued, !from_schedule);
2012 hctx->queue->mq_ops->commit_rqs(hctx);
2017 * Returns true if we did some work AND can potentially do more.
2019 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
2020 unsigned int nr_budgets)
2022 enum prep_dispatch prep;
2023 struct request_queue *q = hctx->queue;
2026 blk_status_t ret = BLK_STS_OK;
2027 LIST_HEAD(zone_list);
2028 bool needs_resource = false;
2030 if (list_empty(list))
2034 * Now process all the entries, sending them to the driver.
2038 struct blk_mq_queue_data bd;
2040 rq = list_first_entry(list, struct request, queuelist);
2042 WARN_ON_ONCE(hctx != rq->mq_hctx);
2043 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2044 if (prep != PREP_DISPATCH_OK)
2047 list_del_init(&rq->queuelist);
2050 bd.last = list_empty(list);
2053 * once the request is queued to lld, no need to cover the
2058 ret = q->mq_ops->queue_rq(hctx, &bd);
2063 case BLK_STS_RESOURCE:
2064 needs_resource = true;
2066 case BLK_STS_DEV_RESOURCE:
2067 blk_mq_handle_dev_resource(rq, list);
2069 case BLK_STS_ZONE_RESOURCE:
2071 * Move the request to zone_list and keep going through
2072 * the dispatch list to find more requests the drive can
2075 blk_mq_handle_zone_resource(rq, &zone_list);
2076 needs_resource = true;
2079 blk_mq_end_request(rq, ret);
2081 } while (!list_empty(list));
2083 if (!list_empty(&zone_list))
2084 list_splice_tail_init(&zone_list, list);
2086 /* If we didn't flush the entire list, we could have told the driver
2087 * there was more coming, but that turned out to be a lie.
2089 if (!list_empty(list) || ret != BLK_STS_OK)
2090 blk_mq_commit_rqs(hctx, queued, false);
2093 * Any items that need requeuing? Stuff them into hctx->dispatch,
2094 * that is where we will continue on next queue run.
2096 if (!list_empty(list)) {
2098 /* For non-shared tags, the RESTART check will suffice */
2099 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2100 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2101 blk_mq_is_shared_tags(hctx->flags));
2104 blk_mq_release_budgets(q, list);
2106 spin_lock(&hctx->lock);
2107 list_splice_tail_init(list, &hctx->dispatch);
2108 spin_unlock(&hctx->lock);
2111 * Order adding requests to hctx->dispatch and checking
2112 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2113 * in blk_mq_sched_restart(). Avoid restart code path to
2114 * miss the new added requests to hctx->dispatch, meantime
2115 * SCHED_RESTART is observed here.
2120 * If SCHED_RESTART was set by the caller of this function and
2121 * it is no longer set that means that it was cleared by another
2122 * thread and hence that a queue rerun is needed.
2124 * If 'no_tag' is set, that means that we failed getting
2125 * a driver tag with an I/O scheduler attached. If our dispatch
2126 * waitqueue is no longer active, ensure that we run the queue
2127 * AFTER adding our entries back to the list.
2129 * If no I/O scheduler has been configured it is possible that
2130 * the hardware queue got stopped and restarted before requests
2131 * were pushed back onto the dispatch list. Rerun the queue to
2132 * avoid starvation. Notes:
2133 * - blk_mq_run_hw_queue() checks whether or not a queue has
2134 * been stopped before rerunning a queue.
2135 * - Some but not all block drivers stop a queue before
2136 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2139 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2140 * bit is set, run queue after a delay to avoid IO stalls
2141 * that could otherwise occur if the queue is idle. We'll do
2142 * similar if we couldn't get budget or couldn't lock a zone
2143 * and SCHED_RESTART is set.
2145 needs_restart = blk_mq_sched_needs_restart(hctx);
2146 if (prep == PREP_DISPATCH_NO_BUDGET)
2147 needs_resource = true;
2148 if (!needs_restart ||
2149 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2150 blk_mq_run_hw_queue(hctx, true);
2151 else if (needs_resource)
2152 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2154 blk_mq_update_dispatch_busy(hctx, true);
2158 blk_mq_update_dispatch_busy(hctx, false);
2162 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2164 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2166 if (cpu >= nr_cpu_ids)
2167 cpu = cpumask_first(hctx->cpumask);
2172 * It'd be great if the workqueue API had a way to pass
2173 * in a mask and had some smarts for more clever placement.
2174 * For now we just round-robin here, switching for every
2175 * BLK_MQ_CPU_WORK_BATCH queued items.
2177 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2180 int next_cpu = hctx->next_cpu;
2182 if (hctx->queue->nr_hw_queues == 1)
2183 return WORK_CPU_UNBOUND;
2185 if (--hctx->next_cpu_batch <= 0) {
2187 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2189 if (next_cpu >= nr_cpu_ids)
2190 next_cpu = blk_mq_first_mapped_cpu(hctx);
2191 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2195 * Do unbound schedule if we can't find a online CPU for this hctx,
2196 * and it should only happen in the path of handling CPU DEAD.
2198 if (!cpu_online(next_cpu)) {
2205 * Make sure to re-select CPU next time once after CPUs
2206 * in hctx->cpumask become online again.
2208 hctx->next_cpu = next_cpu;
2209 hctx->next_cpu_batch = 1;
2210 return WORK_CPU_UNBOUND;
2213 hctx->next_cpu = next_cpu;
2218 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2219 * @hctx: Pointer to the hardware queue to run.
2220 * @msecs: Milliseconds of delay to wait before running the queue.
2222 * Run a hardware queue asynchronously with a delay of @msecs.
2224 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2226 if (unlikely(blk_mq_hctx_stopped(hctx)))
2228 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2229 msecs_to_jiffies(msecs));
2231 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2234 * blk_mq_run_hw_queue - Start to run a hardware queue.
2235 * @hctx: Pointer to the hardware queue to run.
2236 * @async: If we want to run the queue asynchronously.
2238 * Check if the request queue is not in a quiesced state and if there are
2239 * pending requests to be sent. If this is true, run the queue to send requests
2242 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2247 * We can't run the queue inline with interrupts disabled.
2249 WARN_ON_ONCE(!async && in_interrupt());
2251 might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING);
2254 * When queue is quiesced, we may be switching io scheduler, or
2255 * updating nr_hw_queues, or other things, and we can't run queue
2256 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2258 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2261 __blk_mq_run_dispatch_ops(hctx->queue, false,
2262 need_run = !blk_queue_quiesced(hctx->queue) &&
2263 blk_mq_hctx_has_pending(hctx));
2268 if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2269 blk_mq_delay_run_hw_queue(hctx, 0);
2273 blk_mq_run_dispatch_ops(hctx->queue,
2274 blk_mq_sched_dispatch_requests(hctx));
2276 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2279 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2282 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2284 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2286 * If the IO scheduler does not respect hardware queues when
2287 * dispatching, we just don't bother with multiple HW queues and
2288 * dispatch from hctx for the current CPU since running multiple queues
2289 * just causes lock contention inside the scheduler and pointless cache
2292 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2294 if (!blk_mq_hctx_stopped(hctx))
2300 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2301 * @q: Pointer to the request queue to run.
2302 * @async: If we want to run the queue asynchronously.
2304 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2306 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2310 if (blk_queue_sq_sched(q))
2311 sq_hctx = blk_mq_get_sq_hctx(q);
2312 queue_for_each_hw_ctx(q, hctx, i) {
2313 if (blk_mq_hctx_stopped(hctx))
2316 * Dispatch from this hctx either if there's no hctx preferred
2317 * by IO scheduler or if it has requests that bypass the
2320 if (!sq_hctx || sq_hctx == hctx ||
2321 !list_empty_careful(&hctx->dispatch))
2322 blk_mq_run_hw_queue(hctx, async);
2325 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2328 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2329 * @q: Pointer to the request queue to run.
2330 * @msecs: Milliseconds of delay to wait before running the queues.
2332 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2334 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2338 if (blk_queue_sq_sched(q))
2339 sq_hctx = blk_mq_get_sq_hctx(q);
2340 queue_for_each_hw_ctx(q, hctx, i) {
2341 if (blk_mq_hctx_stopped(hctx))
2344 * If there is already a run_work pending, leave the
2345 * pending delay untouched. Otherwise, a hctx can stall
2346 * if another hctx is re-delaying the other's work
2347 * before the work executes.
2349 if (delayed_work_pending(&hctx->run_work))
2352 * Dispatch from this hctx either if there's no hctx preferred
2353 * by IO scheduler or if it has requests that bypass the
2356 if (!sq_hctx || sq_hctx == hctx ||
2357 !list_empty_careful(&hctx->dispatch))
2358 blk_mq_delay_run_hw_queue(hctx, msecs);
2361 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2364 * This function is often used for pausing .queue_rq() by driver when
2365 * there isn't enough resource or some conditions aren't satisfied, and
2366 * BLK_STS_RESOURCE is usually returned.
2368 * We do not guarantee that dispatch can be drained or blocked
2369 * after blk_mq_stop_hw_queue() returns. Please use
2370 * blk_mq_quiesce_queue() for that requirement.
2372 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2374 cancel_delayed_work(&hctx->run_work);
2376 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2378 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2381 * This function is often used for pausing .queue_rq() by driver when
2382 * there isn't enough resource or some conditions aren't satisfied, and
2383 * BLK_STS_RESOURCE is usually returned.
2385 * We do not guarantee that dispatch can be drained or blocked
2386 * after blk_mq_stop_hw_queues() returns. Please use
2387 * blk_mq_quiesce_queue() for that requirement.
2389 void blk_mq_stop_hw_queues(struct request_queue *q)
2391 struct blk_mq_hw_ctx *hctx;
2394 queue_for_each_hw_ctx(q, hctx, i)
2395 blk_mq_stop_hw_queue(hctx);
2397 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2399 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2401 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2403 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
2405 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2407 void blk_mq_start_hw_queues(struct request_queue *q)
2409 struct blk_mq_hw_ctx *hctx;
2412 queue_for_each_hw_ctx(q, hctx, i)
2413 blk_mq_start_hw_queue(hctx);
2415 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2417 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2419 if (!blk_mq_hctx_stopped(hctx))
2422 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2423 blk_mq_run_hw_queue(hctx, async);
2425 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2427 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2429 struct blk_mq_hw_ctx *hctx;
2432 queue_for_each_hw_ctx(q, hctx, i)
2433 blk_mq_start_stopped_hw_queue(hctx, async ||
2434 (hctx->flags & BLK_MQ_F_BLOCKING));
2436 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2438 static void blk_mq_run_work_fn(struct work_struct *work)
2440 struct blk_mq_hw_ctx *hctx =
2441 container_of(work, struct blk_mq_hw_ctx, run_work.work);
2443 blk_mq_run_dispatch_ops(hctx->queue,
2444 blk_mq_sched_dispatch_requests(hctx));
2448 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2449 * @rq: Pointer to request to be inserted.
2450 * @flags: BLK_MQ_INSERT_*
2452 * Should only be used carefully, when the caller knows we want to
2453 * bypass a potential IO scheduler on the target device.
2455 static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2457 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2459 spin_lock(&hctx->lock);
2460 if (flags & BLK_MQ_INSERT_AT_HEAD)
2461 list_add(&rq->queuelist, &hctx->dispatch);
2463 list_add_tail(&rq->queuelist, &hctx->dispatch);
2464 spin_unlock(&hctx->lock);
2467 static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2468 struct blk_mq_ctx *ctx, struct list_head *list,
2469 bool run_queue_async)
2472 enum hctx_type type = hctx->type;
2475 * Try to issue requests directly if the hw queue isn't busy to save an
2476 * extra enqueue & dequeue to the sw queue.
2478 if (!hctx->dispatch_busy && !run_queue_async) {
2479 blk_mq_run_dispatch_ops(hctx->queue,
2480 blk_mq_try_issue_list_directly(hctx, list));
2481 if (list_empty(list))
2486 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2489 list_for_each_entry(rq, list, queuelist) {
2490 BUG_ON(rq->mq_ctx != ctx);
2491 trace_block_rq_insert(rq);
2492 if (rq->cmd_flags & REQ_NOWAIT)
2493 run_queue_async = true;
2496 spin_lock(&ctx->lock);
2497 list_splice_tail_init(list, &ctx->rq_lists[type]);
2498 blk_mq_hctx_mark_pending(hctx, ctx);
2499 spin_unlock(&ctx->lock);
2501 blk_mq_run_hw_queue(hctx, run_queue_async);
2504 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2506 struct request_queue *q = rq->q;
2507 struct blk_mq_ctx *ctx = rq->mq_ctx;
2508 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2510 if (blk_rq_is_passthrough(rq)) {
2512 * Passthrough request have to be added to hctx->dispatch
2513 * directly. The device may be in a situation where it can't
2514 * handle FS request, and always returns BLK_STS_RESOURCE for
2515 * them, which gets them added to hctx->dispatch.
2517 * If a passthrough request is required to unblock the queues,
2518 * and it is added to the scheduler queue, there is no chance to
2519 * dispatch it given we prioritize requests in hctx->dispatch.
2521 blk_mq_request_bypass_insert(rq, flags);
2522 } else if (req_op(rq) == REQ_OP_FLUSH) {
2524 * Firstly normal IO request is inserted to scheduler queue or
2525 * sw queue, meantime we add flush request to dispatch queue(
2526 * hctx->dispatch) directly and there is at most one in-flight
2527 * flush request for each hw queue, so it doesn't matter to add
2528 * flush request to tail or front of the dispatch queue.
2530 * Secondly in case of NCQ, flush request belongs to non-NCQ
2531 * command, and queueing it will fail when there is any
2532 * in-flight normal IO request(NCQ command). When adding flush
2533 * rq to the front of hctx->dispatch, it is easier to introduce
2534 * extra time to flush rq's latency because of S_SCHED_RESTART
2535 * compared with adding to the tail of dispatch queue, then
2536 * chance of flush merge is increased, and less flush requests
2537 * will be issued to controller. It is observed that ~10% time
2538 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2539 * drive when adding flush rq to the front of hctx->dispatch.
2541 * Simply queue flush rq to the front of hctx->dispatch so that
2542 * intensive flush workloads can benefit in case of NCQ HW.
2544 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2545 } else if (q->elevator) {
2548 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2550 list_add(&rq->queuelist, &list);
2551 q->elevator->type->ops.insert_requests(hctx, &list, flags);
2553 trace_block_rq_insert(rq);
2555 spin_lock(&ctx->lock);
2556 if (flags & BLK_MQ_INSERT_AT_HEAD)
2557 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2559 list_add_tail(&rq->queuelist,
2560 &ctx->rq_lists[hctx->type]);
2561 blk_mq_hctx_mark_pending(hctx, ctx);
2562 spin_unlock(&ctx->lock);
2566 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2567 unsigned int nr_segs)
2571 if (bio->bi_opf & REQ_RAHEAD)
2572 rq->cmd_flags |= REQ_FAILFAST_MASK;
2574 rq->__sector = bio->bi_iter.bi_sector;
2575 rq->write_hint = bio->bi_write_hint;
2576 blk_rq_bio_prep(rq, bio, nr_segs);
2578 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2579 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2582 blk_account_io_start(rq);
2585 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2586 struct request *rq, bool last)
2588 struct request_queue *q = rq->q;
2589 struct blk_mq_queue_data bd = {
2596 * For OK queue, we are done. For error, caller may kill it.
2597 * Any other error (busy), just add it to our list as we
2598 * previously would have done.
2600 ret = q->mq_ops->queue_rq(hctx, &bd);
2603 blk_mq_update_dispatch_busy(hctx, false);
2605 case BLK_STS_RESOURCE:
2606 case BLK_STS_DEV_RESOURCE:
2607 blk_mq_update_dispatch_busy(hctx, true);
2608 __blk_mq_requeue_request(rq);
2611 blk_mq_update_dispatch_busy(hctx, false);
2618 static bool blk_mq_get_budget_and_tag(struct request *rq)
2622 budget_token = blk_mq_get_dispatch_budget(rq->q);
2623 if (budget_token < 0)
2625 blk_mq_set_rq_budget_token(rq, budget_token);
2626 if (!blk_mq_get_driver_tag(rq)) {
2627 blk_mq_put_dispatch_budget(rq->q, budget_token);
2634 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2635 * @hctx: Pointer of the associated hardware queue.
2636 * @rq: Pointer to request to be sent.
2638 * If the device has enough resources to accept a new request now, send the
2639 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2640 * we can try send it another time in the future. Requests inserted at this
2641 * queue have higher priority.
2643 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2648 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2649 blk_mq_insert_request(rq, 0);
2653 if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
2654 blk_mq_insert_request(rq, 0);
2655 blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT);
2659 ret = __blk_mq_issue_directly(hctx, rq, true);
2663 case BLK_STS_RESOURCE:
2664 case BLK_STS_DEV_RESOURCE:
2665 blk_mq_request_bypass_insert(rq, 0);
2666 blk_mq_run_hw_queue(hctx, false);
2669 blk_mq_end_request(rq, ret);
2674 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2676 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2678 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2679 blk_mq_insert_request(rq, 0);
2683 if (!blk_mq_get_budget_and_tag(rq))
2684 return BLK_STS_RESOURCE;
2685 return __blk_mq_issue_directly(hctx, rq, last);
2688 static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2690 struct blk_mq_hw_ctx *hctx = NULL;
2693 blk_status_t ret = BLK_STS_OK;
2695 while ((rq = rq_list_pop(&plug->mq_list))) {
2696 bool last = rq_list_empty(plug->mq_list);
2698 if (hctx != rq->mq_hctx) {
2700 blk_mq_commit_rqs(hctx, queued, false);
2706 ret = blk_mq_request_issue_directly(rq, last);
2711 case BLK_STS_RESOURCE:
2712 case BLK_STS_DEV_RESOURCE:
2713 blk_mq_request_bypass_insert(rq, 0);
2714 blk_mq_run_hw_queue(hctx, false);
2717 blk_mq_end_request(rq, ret);
2723 if (ret != BLK_STS_OK)
2724 blk_mq_commit_rqs(hctx, queued, false);
2727 static void __blk_mq_flush_plug_list(struct request_queue *q,
2728 struct blk_plug *plug)
2730 if (blk_queue_quiesced(q))
2732 q->mq_ops->queue_rqs(&plug->mq_list);
2735 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2737 struct blk_mq_hw_ctx *this_hctx = NULL;
2738 struct blk_mq_ctx *this_ctx = NULL;
2739 struct request *requeue_list = NULL;
2740 struct request **requeue_lastp = &requeue_list;
2741 unsigned int depth = 0;
2742 bool is_passthrough = false;
2746 struct request *rq = rq_list_pop(&plug->mq_list);
2749 this_hctx = rq->mq_hctx;
2750 this_ctx = rq->mq_ctx;
2751 is_passthrough = blk_rq_is_passthrough(rq);
2752 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2753 is_passthrough != blk_rq_is_passthrough(rq)) {
2754 rq_list_add_tail(&requeue_lastp, rq);
2757 list_add(&rq->queuelist, &list);
2759 } while (!rq_list_empty(plug->mq_list));
2761 plug->mq_list = requeue_list;
2762 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2764 percpu_ref_get(&this_hctx->queue->q_usage_counter);
2765 /* passthrough requests should never be issued to the I/O scheduler */
2766 if (is_passthrough) {
2767 spin_lock(&this_hctx->lock);
2768 list_splice_tail_init(&list, &this_hctx->dispatch);
2769 spin_unlock(&this_hctx->lock);
2770 blk_mq_run_hw_queue(this_hctx, from_sched);
2771 } else if (this_hctx->queue->elevator) {
2772 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2774 blk_mq_run_hw_queue(this_hctx, from_sched);
2776 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2778 percpu_ref_put(&this_hctx->queue->q_usage_counter);
2781 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2786 * We may have been called recursively midway through handling
2787 * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2788 * To avoid mq_list changing under our feet, clear rq_count early and
2789 * bail out specifically if rq_count is 0 rather than checking
2790 * whether the mq_list is empty.
2792 if (plug->rq_count == 0)
2796 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2797 struct request_queue *q;
2799 rq = rq_list_peek(&plug->mq_list);
2803 * Peek first request and see if we have a ->queue_rqs() hook.
2804 * If we do, we can dispatch the whole plug list in one go. We
2805 * already know at this point that all requests belong to the
2806 * same queue, caller must ensure that's the case.
2808 if (q->mq_ops->queue_rqs) {
2809 blk_mq_run_dispatch_ops(q,
2810 __blk_mq_flush_plug_list(q, plug));
2811 if (rq_list_empty(plug->mq_list))
2815 blk_mq_run_dispatch_ops(q,
2816 blk_mq_plug_issue_direct(plug));
2817 if (rq_list_empty(plug->mq_list))
2822 blk_mq_dispatch_plug_list(plug, from_schedule);
2823 } while (!rq_list_empty(plug->mq_list));
2826 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2827 struct list_head *list)
2830 blk_status_t ret = BLK_STS_OK;
2832 while (!list_empty(list)) {
2833 struct request *rq = list_first_entry(list, struct request,
2836 list_del_init(&rq->queuelist);
2837 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2842 case BLK_STS_RESOURCE:
2843 case BLK_STS_DEV_RESOURCE:
2844 blk_mq_request_bypass_insert(rq, 0);
2845 if (list_empty(list))
2846 blk_mq_run_hw_queue(hctx, false);
2849 blk_mq_end_request(rq, ret);
2855 if (ret != BLK_STS_OK)
2856 blk_mq_commit_rqs(hctx, queued, false);
2859 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2860 struct bio *bio, unsigned int nr_segs)
2862 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2863 if (blk_attempt_plug_merge(q, bio, nr_segs))
2865 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2871 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2872 struct blk_plug *plug,
2876 struct blk_mq_alloc_data data = {
2879 .cmd_flags = bio->bi_opf,
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);
2901 * Check if there is a suitable cached request and return it.
2903 static struct request *blk_mq_peek_cached_request(struct blk_plug *plug,
2904 struct request_queue *q, blk_opf_t opf)
2906 enum hctx_type type = blk_mq_get_hctx_type(opf);
2911 rq = rq_list_peek(&plug->cached_rq);
2912 if (!rq || rq->q != q)
2914 if (type != rq->mq_hctx->type &&
2915 (type != HCTX_TYPE_READ || rq->mq_hctx->type != HCTX_TYPE_DEFAULT))
2917 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
2922 static void blk_mq_use_cached_rq(struct request *rq, struct blk_plug *plug,
2925 WARN_ON_ONCE(rq_list_peek(&plug->cached_rq) != rq);
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(rq->q, bio);
2935 blk_mq_rq_time_init(rq, 0);
2936 rq->cmd_flags = bio->bi_opf;
2937 INIT_LIST_HEAD(&rq->queuelist);
2941 * blk_mq_submit_bio - Create and send a request to block device.
2942 * @bio: Bio pointer.
2944 * Builds up a request structure from @q and @bio and send to the device. The
2945 * request may not be queued directly to hardware if:
2946 * * This request can be merged with another one
2947 * * We want to place request at plug queue for possible future merging
2948 * * There is an IO scheduler active at this queue
2950 * It will not queue the request if there is an error with the bio, or at the
2953 void blk_mq_submit_bio(struct bio *bio)
2955 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2956 struct blk_plug *plug = blk_mq_plug(bio);
2957 const int is_sync = op_is_sync(bio->bi_opf);
2958 struct blk_mq_hw_ctx *hctx;
2959 unsigned int nr_segs = 1;
2963 bio = blk_queue_bounce(bio, q);
2966 * If the plug has a cached request for this queue, try use it.
2968 * The cached request already holds a q_usage_counter reference and we
2969 * don't have to acquire a new one if we use it.
2971 rq = blk_mq_peek_cached_request(plug, q, bio->bi_opf);
2973 if (unlikely(bio_queue_enter(bio)))
2977 if (unlikely(bio_may_exceed_limits(bio, &q->limits))) {
2978 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2982 if (!bio_integrity_prep(bio))
2985 if (blk_mq_attempt_bio_merge(q, bio, nr_segs))
2989 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2993 blk_mq_use_cached_rq(rq, plug, bio);
2996 trace_block_getrq(bio);
2998 rq_qos_track(q, rq, bio);
3000 blk_mq_bio_to_request(rq, bio, nr_segs);
3002 ret = blk_crypto_rq_get_keyslot(rq);
3003 if (ret != BLK_STS_OK) {
3004 bio->bi_status = ret;
3006 blk_mq_free_request(rq);
3010 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
3014 blk_add_rq_to_plug(plug, rq);
3019 if ((rq->rq_flags & RQF_USE_SCHED) ||
3020 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
3021 blk_mq_insert_request(rq, 0);
3022 blk_mq_run_hw_queue(hctx, true);
3024 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
3030 * Don't drop the queue reference if we were trying to use a cached
3031 * request and thus didn't acquire one.
3037 #ifdef CONFIG_BLK_MQ_STACKING
3039 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3040 * @rq: the request being queued
3042 blk_status_t blk_insert_cloned_request(struct request *rq)
3044 struct request_queue *q = rq->q;
3045 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3046 unsigned int max_segments = blk_rq_get_max_segments(rq);
3049 if (blk_rq_sectors(rq) > max_sectors) {
3051 * SCSI device does not have a good way to return if
3052 * Write Same/Zero is actually supported. If a device rejects
3053 * a non-read/write command (discard, write same,etc.) the
3054 * low-level device driver will set the relevant queue limit to
3055 * 0 to prevent blk-lib from issuing more of the offending
3056 * operations. Commands queued prior to the queue limit being
3057 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3058 * errors being propagated to upper layers.
3060 if (max_sectors == 0)
3061 return BLK_STS_NOTSUPP;
3063 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3064 __func__, blk_rq_sectors(rq), max_sectors);
3065 return BLK_STS_IOERR;
3069 * The queue settings related to segment counting may differ from the
3072 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3073 if (rq->nr_phys_segments > max_segments) {
3074 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3075 __func__, rq->nr_phys_segments, max_segments);
3076 return BLK_STS_IOERR;
3079 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3080 return BLK_STS_IOERR;
3082 ret = blk_crypto_rq_get_keyslot(rq);
3083 if (ret != BLK_STS_OK)
3086 blk_account_io_start(rq);
3089 * Since we have a scheduler attached on the top device,
3090 * bypass a potential scheduler on the bottom device for
3093 blk_mq_run_dispatch_ops(q,
3094 ret = blk_mq_request_issue_directly(rq, true));
3096 blk_account_io_done(rq, blk_time_get_ns());
3099 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3102 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3103 * @rq: the clone request to be cleaned up
3106 * Free all bios in @rq for a cloned request.
3108 void blk_rq_unprep_clone(struct request *rq)
3112 while ((bio = rq->bio) != NULL) {
3113 rq->bio = bio->bi_next;
3118 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3121 * blk_rq_prep_clone - Helper function to setup clone request
3122 * @rq: the request to be setup
3123 * @rq_src: original request to be cloned
3124 * @bs: bio_set that bios for clone are allocated from
3125 * @gfp_mask: memory allocation mask for bio
3126 * @bio_ctr: setup function to be called for each clone bio.
3127 * Returns %0 for success, non %0 for failure.
3128 * @data: private data to be passed to @bio_ctr
3131 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3132 * Also, pages which the original bios are pointing to are not copied
3133 * and the cloned bios just point same pages.
3134 * So cloned bios must be completed before original bios, which means
3135 * the caller must complete @rq before @rq_src.
3137 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3138 struct bio_set *bs, gfp_t gfp_mask,
3139 int (*bio_ctr)(struct bio *, struct bio *, void *),
3142 struct bio *bio, *bio_src;
3147 __rq_for_each_bio(bio_src, rq_src) {
3148 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3153 if (bio_ctr && bio_ctr(bio, bio_src, data))
3157 rq->biotail->bi_next = bio;
3160 rq->bio = rq->biotail = bio;
3165 /* Copy attributes of the original request to the clone request. */
3166 rq->__sector = blk_rq_pos(rq_src);
3167 rq->__data_len = blk_rq_bytes(rq_src);
3168 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3169 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3170 rq->special_vec = rq_src->special_vec;
3172 rq->nr_phys_segments = rq_src->nr_phys_segments;
3173 rq->ioprio = rq_src->ioprio;
3174 rq->write_hint = rq_src->write_hint;
3176 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3184 blk_rq_unprep_clone(rq);
3188 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3189 #endif /* CONFIG_BLK_MQ_STACKING */
3192 * Steal bios from a request and add them to a bio list.
3193 * The request must not have been partially completed before.
3195 void blk_steal_bios(struct bio_list *list, struct request *rq)
3199 list->tail->bi_next = rq->bio;
3201 list->head = rq->bio;
3202 list->tail = rq->biotail;
3210 EXPORT_SYMBOL_GPL(blk_steal_bios);
3212 static size_t order_to_size(unsigned int order)
3214 return (size_t)PAGE_SIZE << order;
3217 /* called before freeing request pool in @tags */
3218 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3219 struct blk_mq_tags *tags)
3222 unsigned long flags;
3225 * There is no need to clear mapping if driver tags is not initialized
3226 * or the mapping belongs to the driver tags.
3228 if (!drv_tags || drv_tags == tags)
3231 list_for_each_entry(page, &tags->page_list, lru) {
3232 unsigned long start = (unsigned long)page_address(page);
3233 unsigned long end = start + order_to_size(page->private);
3236 for (i = 0; i < drv_tags->nr_tags; i++) {
3237 struct request *rq = drv_tags->rqs[i];
3238 unsigned long rq_addr = (unsigned long)rq;
3240 if (rq_addr >= start && rq_addr < end) {
3241 WARN_ON_ONCE(req_ref_read(rq) != 0);
3242 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3248 * Wait until all pending iteration is done.
3250 * Request reference is cleared and it is guaranteed to be observed
3251 * after the ->lock is released.
3253 spin_lock_irqsave(&drv_tags->lock, flags);
3254 spin_unlock_irqrestore(&drv_tags->lock, flags);
3257 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3258 unsigned int hctx_idx)
3260 struct blk_mq_tags *drv_tags;
3263 if (list_empty(&tags->page_list))
3266 if (blk_mq_is_shared_tags(set->flags))
3267 drv_tags = set->shared_tags;
3269 drv_tags = set->tags[hctx_idx];
3271 if (tags->static_rqs && set->ops->exit_request) {
3274 for (i = 0; i < tags->nr_tags; i++) {
3275 struct request *rq = tags->static_rqs[i];
3279 set->ops->exit_request(set, rq, hctx_idx);
3280 tags->static_rqs[i] = NULL;
3284 blk_mq_clear_rq_mapping(drv_tags, tags);
3286 while (!list_empty(&tags->page_list)) {
3287 page = list_first_entry(&tags->page_list, struct page, lru);
3288 list_del_init(&page->lru);
3290 * Remove kmemleak object previously allocated in
3291 * blk_mq_alloc_rqs().
3293 kmemleak_free(page_address(page));
3294 __free_pages(page, page->private);
3298 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3302 kfree(tags->static_rqs);
3303 tags->static_rqs = NULL;
3305 blk_mq_free_tags(tags);
3308 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3309 unsigned int hctx_idx)
3313 for (i = 0; i < set->nr_maps; i++) {
3314 unsigned int start = set->map[i].queue_offset;
3315 unsigned int end = start + set->map[i].nr_queues;
3317 if (hctx_idx >= start && hctx_idx < end)
3321 if (i >= set->nr_maps)
3322 i = HCTX_TYPE_DEFAULT;
3327 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3328 unsigned int hctx_idx)
3330 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3332 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3335 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3336 unsigned int hctx_idx,
3337 unsigned int nr_tags,
3338 unsigned int reserved_tags)
3340 int node = blk_mq_get_hctx_node(set, hctx_idx);
3341 struct blk_mq_tags *tags;
3343 if (node == NUMA_NO_NODE)
3344 node = set->numa_node;
3346 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3347 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3351 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3352 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3357 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3358 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3360 if (!tags->static_rqs)
3368 blk_mq_free_tags(tags);
3372 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3373 unsigned int hctx_idx, int node)
3377 if (set->ops->init_request) {
3378 ret = set->ops->init_request(set, rq, hctx_idx, node);
3383 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3387 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3388 struct blk_mq_tags *tags,
3389 unsigned int hctx_idx, unsigned int depth)
3391 unsigned int i, j, entries_per_page, max_order = 4;
3392 int node = blk_mq_get_hctx_node(set, hctx_idx);
3393 size_t rq_size, left;
3395 if (node == NUMA_NO_NODE)
3396 node = set->numa_node;
3398 INIT_LIST_HEAD(&tags->page_list);
3401 * rq_size is the size of the request plus driver payload, rounded
3402 * to the cacheline size
3404 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3406 left = rq_size * depth;
3408 for (i = 0; i < depth; ) {
3409 int this_order = max_order;
3414 while (this_order && left < order_to_size(this_order - 1))
3418 page = alloc_pages_node(node,
3419 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3425 if (order_to_size(this_order) < rq_size)
3432 page->private = this_order;
3433 list_add_tail(&page->lru, &tags->page_list);
3435 p = page_address(page);
3437 * Allow kmemleak to scan these pages as they contain pointers
3438 * to additional allocations like via ops->init_request().
3440 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3441 entries_per_page = order_to_size(this_order) / rq_size;
3442 to_do = min(entries_per_page, depth - i);
3443 left -= to_do * rq_size;
3444 for (j = 0; j < to_do; j++) {
3445 struct request *rq = p;
3447 tags->static_rqs[i] = rq;
3448 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3449 tags->static_rqs[i] = NULL;
3460 blk_mq_free_rqs(set, tags, hctx_idx);
3464 struct rq_iter_data {
3465 struct blk_mq_hw_ctx *hctx;
3469 static bool blk_mq_has_request(struct request *rq, void *data)
3471 struct rq_iter_data *iter_data = data;
3473 if (rq->mq_hctx != iter_data->hctx)
3475 iter_data->has_rq = true;
3479 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3481 struct blk_mq_tags *tags = hctx->sched_tags ?
3482 hctx->sched_tags : hctx->tags;
3483 struct rq_iter_data data = {
3487 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3491 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3492 struct blk_mq_hw_ctx *hctx)
3494 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3496 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3501 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3503 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3504 struct blk_mq_hw_ctx, cpuhp_online);
3506 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3507 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3511 * Prevent new request from being allocated on the current hctx.
3513 * The smp_mb__after_atomic() Pairs with the implied barrier in
3514 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3515 * seen once we return from the tag allocator.
3517 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3518 smp_mb__after_atomic();
3521 * Try to grab a reference to the queue and wait for any outstanding
3522 * requests. If we could not grab a reference the queue has been
3523 * frozen and there are no requests.
3525 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3526 while (blk_mq_hctx_has_requests(hctx))
3528 percpu_ref_put(&hctx->queue->q_usage_counter);
3534 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3536 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3537 struct blk_mq_hw_ctx, cpuhp_online);
3539 if (cpumask_test_cpu(cpu, hctx->cpumask))
3540 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3545 * 'cpu' is going away. splice any existing rq_list entries from this
3546 * software queue to the hw queue dispatch list, and ensure that it
3549 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3551 struct blk_mq_hw_ctx *hctx;
3552 struct blk_mq_ctx *ctx;
3554 enum hctx_type type;
3556 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3557 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3560 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3563 spin_lock(&ctx->lock);
3564 if (!list_empty(&ctx->rq_lists[type])) {
3565 list_splice_init(&ctx->rq_lists[type], &tmp);
3566 blk_mq_hctx_clear_pending(hctx, ctx);
3568 spin_unlock(&ctx->lock);
3570 if (list_empty(&tmp))
3573 spin_lock(&hctx->lock);
3574 list_splice_tail_init(&tmp, &hctx->dispatch);
3575 spin_unlock(&hctx->lock);
3577 blk_mq_run_hw_queue(hctx, true);
3581 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3583 if (!(hctx->flags & BLK_MQ_F_STACKING))
3584 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3585 &hctx->cpuhp_online);
3586 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3591 * Before freeing hw queue, clearing the flush request reference in
3592 * tags->rqs[] for avoiding potential UAF.
3594 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3595 unsigned int queue_depth, struct request *flush_rq)
3598 unsigned long flags;
3600 /* The hw queue may not be mapped yet */
3604 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3606 for (i = 0; i < queue_depth; i++)
3607 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3610 * Wait until all pending iteration is done.
3612 * Request reference is cleared and it is guaranteed to be observed
3613 * after the ->lock is released.
3615 spin_lock_irqsave(&tags->lock, flags);
3616 spin_unlock_irqrestore(&tags->lock, flags);
3619 /* hctx->ctxs will be freed in queue's release handler */
3620 static void blk_mq_exit_hctx(struct request_queue *q,
3621 struct blk_mq_tag_set *set,
3622 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3624 struct request *flush_rq = hctx->fq->flush_rq;
3626 if (blk_mq_hw_queue_mapped(hctx))
3627 blk_mq_tag_idle(hctx);
3629 if (blk_queue_init_done(q))
3630 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3631 set->queue_depth, flush_rq);
3632 if (set->ops->exit_request)
3633 set->ops->exit_request(set, flush_rq, hctx_idx);
3635 if (set->ops->exit_hctx)
3636 set->ops->exit_hctx(hctx, hctx_idx);
3638 blk_mq_remove_cpuhp(hctx);
3640 xa_erase(&q->hctx_table, hctx_idx);
3642 spin_lock(&q->unused_hctx_lock);
3643 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3644 spin_unlock(&q->unused_hctx_lock);
3647 static void blk_mq_exit_hw_queues(struct request_queue *q,
3648 struct blk_mq_tag_set *set, int nr_queue)
3650 struct blk_mq_hw_ctx *hctx;
3653 queue_for_each_hw_ctx(q, hctx, i) {
3656 blk_mq_exit_hctx(q, set, hctx, i);
3660 static int blk_mq_init_hctx(struct request_queue *q,
3661 struct blk_mq_tag_set *set,
3662 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3664 hctx->queue_num = hctx_idx;
3666 if (!(hctx->flags & BLK_MQ_F_STACKING))
3667 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3668 &hctx->cpuhp_online);
3669 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3671 hctx->tags = set->tags[hctx_idx];
3673 if (set->ops->init_hctx &&
3674 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3675 goto unregister_cpu_notifier;
3677 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3681 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3687 if (set->ops->exit_request)
3688 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3690 if (set->ops->exit_hctx)
3691 set->ops->exit_hctx(hctx, hctx_idx);
3692 unregister_cpu_notifier:
3693 blk_mq_remove_cpuhp(hctx);
3697 static struct blk_mq_hw_ctx *
3698 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3701 struct blk_mq_hw_ctx *hctx;
3702 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3704 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3706 goto fail_alloc_hctx;
3708 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3711 atomic_set(&hctx->nr_active, 0);
3712 if (node == NUMA_NO_NODE)
3713 node = set->numa_node;
3714 hctx->numa_node = node;
3716 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3717 spin_lock_init(&hctx->lock);
3718 INIT_LIST_HEAD(&hctx->dispatch);
3720 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3722 INIT_LIST_HEAD(&hctx->hctx_list);
3725 * Allocate space for all possible cpus to avoid allocation at
3728 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3733 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3734 gfp, node, false, false))
3738 spin_lock_init(&hctx->dispatch_wait_lock);
3739 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3740 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3742 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3746 blk_mq_hctx_kobj_init(hctx);
3751 sbitmap_free(&hctx->ctx_map);
3755 free_cpumask_var(hctx->cpumask);
3762 static void blk_mq_init_cpu_queues(struct request_queue *q,
3763 unsigned int nr_hw_queues)
3765 struct blk_mq_tag_set *set = q->tag_set;
3768 for_each_possible_cpu(i) {
3769 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3770 struct blk_mq_hw_ctx *hctx;
3774 spin_lock_init(&__ctx->lock);
3775 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3776 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3781 * Set local node, IFF we have more than one hw queue. If
3782 * not, we remain on the home node of the device
3784 for (j = 0; j < set->nr_maps; j++) {
3785 hctx = blk_mq_map_queue_type(q, j, i);
3786 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3787 hctx->numa_node = cpu_to_node(i);
3792 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3793 unsigned int hctx_idx,
3796 struct blk_mq_tags *tags;
3799 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3803 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3805 blk_mq_free_rq_map(tags);
3812 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3815 if (blk_mq_is_shared_tags(set->flags)) {
3816 set->tags[hctx_idx] = set->shared_tags;
3821 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3824 return set->tags[hctx_idx];
3827 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3828 struct blk_mq_tags *tags,
3829 unsigned int hctx_idx)
3832 blk_mq_free_rqs(set, tags, hctx_idx);
3833 blk_mq_free_rq_map(tags);
3837 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3838 unsigned int hctx_idx)
3840 if (!blk_mq_is_shared_tags(set->flags))
3841 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3843 set->tags[hctx_idx] = NULL;
3846 static void blk_mq_map_swqueue(struct request_queue *q)
3848 unsigned int j, hctx_idx;
3850 struct blk_mq_hw_ctx *hctx;
3851 struct blk_mq_ctx *ctx;
3852 struct blk_mq_tag_set *set = q->tag_set;
3854 queue_for_each_hw_ctx(q, hctx, i) {
3855 cpumask_clear(hctx->cpumask);
3857 hctx->dispatch_from = NULL;
3861 * Map software to hardware queues.
3863 * If the cpu isn't present, the cpu is mapped to first hctx.
3865 for_each_possible_cpu(i) {
3867 ctx = per_cpu_ptr(q->queue_ctx, i);
3868 for (j = 0; j < set->nr_maps; j++) {
3869 if (!set->map[j].nr_queues) {
3870 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3871 HCTX_TYPE_DEFAULT, i);
3874 hctx_idx = set->map[j].mq_map[i];
3875 /* unmapped hw queue can be remapped after CPU topo changed */
3876 if (!set->tags[hctx_idx] &&
3877 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3879 * If tags initialization fail for some hctx,
3880 * that hctx won't be brought online. In this
3881 * case, remap the current ctx to hctx[0] which
3882 * is guaranteed to always have tags allocated
3884 set->map[j].mq_map[i] = 0;
3887 hctx = blk_mq_map_queue_type(q, j, i);
3888 ctx->hctxs[j] = hctx;
3890 * If the CPU is already set in the mask, then we've
3891 * mapped this one already. This can happen if
3892 * devices share queues across queue maps.
3894 if (cpumask_test_cpu(i, hctx->cpumask))
3897 cpumask_set_cpu(i, hctx->cpumask);
3899 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3900 hctx->ctxs[hctx->nr_ctx++] = ctx;
3903 * If the nr_ctx type overflows, we have exceeded the
3904 * amount of sw queues we can support.
3906 BUG_ON(!hctx->nr_ctx);
3909 for (; j < HCTX_MAX_TYPES; j++)
3910 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3911 HCTX_TYPE_DEFAULT, i);
3914 queue_for_each_hw_ctx(q, hctx, i) {
3916 * If no software queues are mapped to this hardware queue,
3917 * disable it and free the request entries.
3919 if (!hctx->nr_ctx) {
3920 /* Never unmap queue 0. We need it as a
3921 * fallback in case of a new remap fails
3925 __blk_mq_free_map_and_rqs(set, i);
3931 hctx->tags = set->tags[i];
3932 WARN_ON(!hctx->tags);
3935 * Set the map size to the number of mapped software queues.
3936 * This is more accurate and more efficient than looping
3937 * over all possibly mapped software queues.
3939 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3942 * Initialize batch roundrobin counts
3944 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3945 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3950 * Caller needs to ensure that we're either frozen/quiesced, or that
3951 * the queue isn't live yet.
3953 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3955 struct blk_mq_hw_ctx *hctx;
3958 queue_for_each_hw_ctx(q, hctx, i) {
3960 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3962 blk_mq_tag_idle(hctx);
3963 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3968 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3971 struct request_queue *q;
3973 lockdep_assert_held(&set->tag_list_lock);
3975 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3976 blk_mq_freeze_queue(q);
3977 queue_set_hctx_shared(q, shared);
3978 blk_mq_unfreeze_queue(q);
3982 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3984 struct blk_mq_tag_set *set = q->tag_set;
3986 mutex_lock(&set->tag_list_lock);
3987 list_del(&q->tag_set_list);
3988 if (list_is_singular(&set->tag_list)) {
3989 /* just transitioned to unshared */
3990 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3991 /* update existing queue */
3992 blk_mq_update_tag_set_shared(set, false);
3994 mutex_unlock(&set->tag_list_lock);
3995 INIT_LIST_HEAD(&q->tag_set_list);
3998 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3999 struct request_queue *q)
4001 mutex_lock(&set->tag_list_lock);
4004 * Check to see if we're transitioning to shared (from 1 to 2 queues).
4006 if (!list_empty(&set->tag_list) &&
4007 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
4008 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
4009 /* update existing queue */
4010 blk_mq_update_tag_set_shared(set, true);
4012 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
4013 queue_set_hctx_shared(q, true);
4014 list_add_tail(&q->tag_set_list, &set->tag_list);
4016 mutex_unlock(&set->tag_list_lock);
4019 /* All allocations will be freed in release handler of q->mq_kobj */
4020 static int blk_mq_alloc_ctxs(struct request_queue *q)
4022 struct blk_mq_ctxs *ctxs;
4025 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
4029 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
4030 if (!ctxs->queue_ctx)
4033 for_each_possible_cpu(cpu) {
4034 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4038 q->mq_kobj = &ctxs->kobj;
4039 q->queue_ctx = ctxs->queue_ctx;
4048 * It is the actual release handler for mq, but we do it from
4049 * request queue's release handler for avoiding use-after-free
4050 * and headache because q->mq_kobj shouldn't have been introduced,
4051 * but we can't group ctx/kctx kobj without it.
4053 void blk_mq_release(struct request_queue *q)
4055 struct blk_mq_hw_ctx *hctx, *next;
4058 queue_for_each_hw_ctx(q, hctx, i)
4059 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4061 /* all hctx are in .unused_hctx_list now */
4062 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4063 list_del_init(&hctx->hctx_list);
4064 kobject_put(&hctx->kobj);
4067 xa_destroy(&q->hctx_table);
4070 * release .mq_kobj and sw queue's kobject now because
4071 * both share lifetime with request queue.
4073 blk_mq_sysfs_deinit(q);
4076 struct request_queue *blk_mq_alloc_queue(struct blk_mq_tag_set *set,
4077 struct queue_limits *lim, void *queuedata)
4079 struct queue_limits default_lim = { };
4080 struct request_queue *q;
4083 q = blk_alloc_queue(lim ? lim : &default_lim, set->numa_node);
4086 q->queuedata = queuedata;
4087 ret = blk_mq_init_allocated_queue(set, q);
4090 return ERR_PTR(ret);
4094 EXPORT_SYMBOL(blk_mq_alloc_queue);
4097 * blk_mq_destroy_queue - shutdown a request queue
4098 * @q: request queue to shutdown
4100 * This shuts down a request queue allocated by blk_mq_alloc_queue(). All future
4101 * requests will be failed with -ENODEV. The caller is responsible for dropping
4102 * the reference from blk_mq_alloc_queue() by calling blk_put_queue().
4104 * Context: can sleep
4106 void blk_mq_destroy_queue(struct request_queue *q)
4108 WARN_ON_ONCE(!queue_is_mq(q));
4109 WARN_ON_ONCE(blk_queue_registered(q));
4113 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4114 blk_queue_start_drain(q);
4115 blk_mq_freeze_queue_wait(q);
4118 blk_mq_cancel_work_sync(q);
4119 blk_mq_exit_queue(q);
4121 EXPORT_SYMBOL(blk_mq_destroy_queue);
4123 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set,
4124 struct queue_limits *lim, void *queuedata,
4125 struct lock_class_key *lkclass)
4127 struct request_queue *q;
4128 struct gendisk *disk;
4130 q = blk_mq_alloc_queue(set, lim, queuedata);
4134 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4136 blk_mq_destroy_queue(q);
4138 return ERR_PTR(-ENOMEM);
4140 set_bit(GD_OWNS_QUEUE, &disk->state);
4143 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4145 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4146 struct lock_class_key *lkclass)
4148 struct gendisk *disk;
4150 if (!blk_get_queue(q))
4152 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4157 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4159 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4160 struct blk_mq_tag_set *set, struct request_queue *q,
4161 int hctx_idx, int node)
4163 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4165 /* reuse dead hctx first */
4166 spin_lock(&q->unused_hctx_lock);
4167 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4168 if (tmp->numa_node == node) {
4174 list_del_init(&hctx->hctx_list);
4175 spin_unlock(&q->unused_hctx_lock);
4178 hctx = blk_mq_alloc_hctx(q, set, node);
4182 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4188 kobject_put(&hctx->kobj);
4193 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4194 struct request_queue *q)
4196 struct blk_mq_hw_ctx *hctx;
4199 /* protect against switching io scheduler */
4200 mutex_lock(&q->sysfs_lock);
4201 for (i = 0; i < set->nr_hw_queues; i++) {
4203 int node = blk_mq_get_hctx_node(set, i);
4204 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4207 old_node = old_hctx->numa_node;
4208 blk_mq_exit_hctx(q, set, old_hctx, i);
4211 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4214 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4216 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4217 WARN_ON_ONCE(!hctx);
4221 * Increasing nr_hw_queues fails. Free the newly allocated
4222 * hctxs and keep the previous q->nr_hw_queues.
4224 if (i != set->nr_hw_queues) {
4225 j = q->nr_hw_queues;
4228 q->nr_hw_queues = set->nr_hw_queues;
4231 xa_for_each_start(&q->hctx_table, j, hctx, j)
4232 blk_mq_exit_hctx(q, set, hctx, j);
4233 mutex_unlock(&q->sysfs_lock);
4236 static void blk_mq_update_poll_flag(struct request_queue *q)
4238 struct blk_mq_tag_set *set = q->tag_set;
4240 if (set->nr_maps > HCTX_TYPE_POLL &&
4241 set->map[HCTX_TYPE_POLL].nr_queues)
4242 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4244 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4247 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4248 struct request_queue *q)
4250 /* mark the queue as mq asap */
4251 q->mq_ops = set->ops;
4253 if (blk_mq_alloc_ctxs(q))
4256 /* init q->mq_kobj and sw queues' kobjects */
4257 blk_mq_sysfs_init(q);
4259 INIT_LIST_HEAD(&q->unused_hctx_list);
4260 spin_lock_init(&q->unused_hctx_lock);
4262 xa_init(&q->hctx_table);
4264 blk_mq_realloc_hw_ctxs(set, q);
4265 if (!q->nr_hw_queues)
4268 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4269 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4273 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4274 blk_mq_update_poll_flag(q);
4276 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4277 INIT_LIST_HEAD(&q->flush_list);
4278 INIT_LIST_HEAD(&q->requeue_list);
4279 spin_lock_init(&q->requeue_lock);
4281 q->nr_requests = set->queue_depth;
4283 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4284 blk_mq_add_queue_tag_set(set, q);
4285 blk_mq_map_swqueue(q);
4294 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4296 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4297 void blk_mq_exit_queue(struct request_queue *q)
4299 struct blk_mq_tag_set *set = q->tag_set;
4301 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4302 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4303 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4304 blk_mq_del_queue_tag_set(q);
4307 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4311 if (blk_mq_is_shared_tags(set->flags)) {
4312 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4315 if (!set->shared_tags)
4319 for (i = 0; i < set->nr_hw_queues; i++) {
4320 if (!__blk_mq_alloc_map_and_rqs(set, i))
4329 __blk_mq_free_map_and_rqs(set, i);
4331 if (blk_mq_is_shared_tags(set->flags)) {
4332 blk_mq_free_map_and_rqs(set, set->shared_tags,
4333 BLK_MQ_NO_HCTX_IDX);
4340 * Allocate the request maps associated with this tag_set. Note that this
4341 * may reduce the depth asked for, if memory is tight. set->queue_depth
4342 * will be updated to reflect the allocated depth.
4344 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4349 depth = set->queue_depth;
4351 err = __blk_mq_alloc_rq_maps(set);
4355 set->queue_depth >>= 1;
4356 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4360 } while (set->queue_depth);
4362 if (!set->queue_depth || err) {
4363 pr_err("blk-mq: failed to allocate request map\n");
4367 if (depth != set->queue_depth)
4368 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4369 depth, set->queue_depth);
4374 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4377 * blk_mq_map_queues() and multiple .map_queues() implementations
4378 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4379 * number of hardware queues.
4381 if (set->nr_maps == 1)
4382 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4384 if (set->ops->map_queues) {
4388 * transport .map_queues is usually done in the following
4391 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4392 * mask = get_cpu_mask(queue)
4393 * for_each_cpu(cpu, mask)
4394 * set->map[x].mq_map[cpu] = queue;
4397 * When we need to remap, the table has to be cleared for
4398 * killing stale mapping since one CPU may not be mapped
4401 for (i = 0; i < set->nr_maps; i++)
4402 blk_mq_clear_mq_map(&set->map[i]);
4404 set->ops->map_queues(set);
4406 BUG_ON(set->nr_maps > 1);
4407 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4411 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4412 int new_nr_hw_queues)
4414 struct blk_mq_tags **new_tags;
4417 if (set->nr_hw_queues >= new_nr_hw_queues)
4420 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4421 GFP_KERNEL, set->numa_node);
4426 memcpy(new_tags, set->tags, set->nr_hw_queues *
4427 sizeof(*set->tags));
4429 set->tags = new_tags;
4431 for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) {
4432 if (!__blk_mq_alloc_map_and_rqs(set, i)) {
4433 while (--i >= set->nr_hw_queues)
4434 __blk_mq_free_map_and_rqs(set, i);
4441 set->nr_hw_queues = new_nr_hw_queues;
4446 * Alloc a tag set to be associated with one or more request queues.
4447 * May fail with EINVAL for various error conditions. May adjust the
4448 * requested depth down, if it's too large. In that case, the set
4449 * value will be stored in set->queue_depth.
4451 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4455 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4457 if (!set->nr_hw_queues)
4459 if (!set->queue_depth)
4461 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4464 if (!set->ops->queue_rq)
4467 if (!set->ops->get_budget ^ !set->ops->put_budget)
4470 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4471 pr_info("blk-mq: reduced tag depth to %u\n",
4473 set->queue_depth = BLK_MQ_MAX_DEPTH;
4478 else if (set->nr_maps > HCTX_MAX_TYPES)
4482 * If a crashdump is active, then we are potentially in a very
4483 * memory constrained environment. Limit us to 64 tags to prevent
4484 * using too much memory.
4486 if (is_kdump_kernel())
4487 set->queue_depth = min(64U, set->queue_depth);
4490 * There is no use for more h/w queues than cpus if we just have
4493 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4494 set->nr_hw_queues = nr_cpu_ids;
4496 if (set->flags & BLK_MQ_F_BLOCKING) {
4497 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4500 ret = init_srcu_struct(set->srcu);
4506 set->tags = kcalloc_node(set->nr_hw_queues,
4507 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4510 goto out_cleanup_srcu;
4512 for (i = 0; i < set->nr_maps; i++) {
4513 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4514 sizeof(set->map[i].mq_map[0]),
4515 GFP_KERNEL, set->numa_node);
4516 if (!set->map[i].mq_map)
4517 goto out_free_mq_map;
4518 set->map[i].nr_queues = set->nr_hw_queues;
4521 blk_mq_update_queue_map(set);
4523 ret = blk_mq_alloc_set_map_and_rqs(set);
4525 goto out_free_mq_map;
4527 mutex_init(&set->tag_list_lock);
4528 INIT_LIST_HEAD(&set->tag_list);
4533 for (i = 0; i < set->nr_maps; i++) {
4534 kfree(set->map[i].mq_map);
4535 set->map[i].mq_map = NULL;
4540 if (set->flags & BLK_MQ_F_BLOCKING)
4541 cleanup_srcu_struct(set->srcu);
4543 if (set->flags & BLK_MQ_F_BLOCKING)
4547 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4549 /* allocate and initialize a tagset for a simple single-queue device */
4550 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4551 const struct blk_mq_ops *ops, unsigned int queue_depth,
4552 unsigned int set_flags)
4554 memset(set, 0, sizeof(*set));
4556 set->nr_hw_queues = 1;
4558 set->queue_depth = queue_depth;
4559 set->numa_node = NUMA_NO_NODE;
4560 set->flags = set_flags;
4561 return blk_mq_alloc_tag_set(set);
4563 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4565 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4569 for (i = 0; i < set->nr_hw_queues; i++)
4570 __blk_mq_free_map_and_rqs(set, i);
4572 if (blk_mq_is_shared_tags(set->flags)) {
4573 blk_mq_free_map_and_rqs(set, set->shared_tags,
4574 BLK_MQ_NO_HCTX_IDX);
4577 for (j = 0; j < set->nr_maps; j++) {
4578 kfree(set->map[j].mq_map);
4579 set->map[j].mq_map = NULL;
4584 if (set->flags & BLK_MQ_F_BLOCKING) {
4585 cleanup_srcu_struct(set->srcu);
4589 EXPORT_SYMBOL(blk_mq_free_tag_set);
4591 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4593 struct blk_mq_tag_set *set = q->tag_set;
4594 struct blk_mq_hw_ctx *hctx;
4601 if (q->nr_requests == nr)
4604 blk_mq_freeze_queue(q);
4605 blk_mq_quiesce_queue(q);
4608 queue_for_each_hw_ctx(q, hctx, i) {
4612 * If we're using an MQ scheduler, just update the scheduler
4613 * queue depth. This is similar to what the old code would do.
4615 if (hctx->sched_tags) {
4616 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4619 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4624 if (q->elevator && q->elevator->type->ops.depth_updated)
4625 q->elevator->type->ops.depth_updated(hctx);
4628 q->nr_requests = nr;
4629 if (blk_mq_is_shared_tags(set->flags)) {
4631 blk_mq_tag_update_sched_shared_tags(q);
4633 blk_mq_tag_resize_shared_tags(set, nr);
4637 blk_mq_unquiesce_queue(q);
4638 blk_mq_unfreeze_queue(q);
4644 * request_queue and elevator_type pair.
4645 * It is just used by __blk_mq_update_nr_hw_queues to cache
4646 * the elevator_type associated with a request_queue.
4648 struct blk_mq_qe_pair {
4649 struct list_head node;
4650 struct request_queue *q;
4651 struct elevator_type *type;
4655 * Cache the elevator_type in qe pair list and switch the
4656 * io scheduler to 'none'
4658 static bool blk_mq_elv_switch_none(struct list_head *head,
4659 struct request_queue *q)
4661 struct blk_mq_qe_pair *qe;
4663 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4667 /* q->elevator needs protection from ->sysfs_lock */
4668 mutex_lock(&q->sysfs_lock);
4670 /* the check has to be done with holding sysfs_lock */
4676 INIT_LIST_HEAD(&qe->node);
4678 qe->type = q->elevator->type;
4679 /* keep a reference to the elevator module as we'll switch back */
4680 __elevator_get(qe->type);
4681 list_add(&qe->node, head);
4682 elevator_disable(q);
4684 mutex_unlock(&q->sysfs_lock);
4689 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4690 struct request_queue *q)
4692 struct blk_mq_qe_pair *qe;
4694 list_for_each_entry(qe, head, node)
4701 static void blk_mq_elv_switch_back(struct list_head *head,
4702 struct request_queue *q)
4704 struct blk_mq_qe_pair *qe;
4705 struct elevator_type *t;
4707 qe = blk_lookup_qe_pair(head, q);
4711 list_del(&qe->node);
4714 mutex_lock(&q->sysfs_lock);
4715 elevator_switch(q, t);
4716 /* drop the reference acquired in blk_mq_elv_switch_none */
4718 mutex_unlock(&q->sysfs_lock);
4721 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4724 struct request_queue *q;
4726 int prev_nr_hw_queues = set->nr_hw_queues;
4729 lockdep_assert_held(&set->tag_list_lock);
4731 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4732 nr_hw_queues = nr_cpu_ids;
4733 if (nr_hw_queues < 1)
4735 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4738 list_for_each_entry(q, &set->tag_list, tag_set_list)
4739 blk_mq_freeze_queue(q);
4741 * Switch IO scheduler to 'none', cleaning up the data associated
4742 * with the previous scheduler. We will switch back once we are done
4743 * updating the new sw to hw queue mappings.
4745 list_for_each_entry(q, &set->tag_list, tag_set_list)
4746 if (!blk_mq_elv_switch_none(&head, q))
4749 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4750 blk_mq_debugfs_unregister_hctxs(q);
4751 blk_mq_sysfs_unregister_hctxs(q);
4754 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4758 blk_mq_update_queue_map(set);
4759 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4760 blk_mq_realloc_hw_ctxs(set, q);
4761 blk_mq_update_poll_flag(q);
4762 if (q->nr_hw_queues != set->nr_hw_queues) {
4763 int i = prev_nr_hw_queues;
4765 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4766 nr_hw_queues, prev_nr_hw_queues);
4767 for (; i < set->nr_hw_queues; i++)
4768 __blk_mq_free_map_and_rqs(set, i);
4770 set->nr_hw_queues = prev_nr_hw_queues;
4773 blk_mq_map_swqueue(q);
4777 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4778 blk_mq_sysfs_register_hctxs(q);
4779 blk_mq_debugfs_register_hctxs(q);
4783 list_for_each_entry(q, &set->tag_list, tag_set_list)
4784 blk_mq_elv_switch_back(&head, q);
4786 list_for_each_entry(q, &set->tag_list, tag_set_list)
4787 blk_mq_unfreeze_queue(q);
4789 /* Free the excess tags when nr_hw_queues shrink. */
4790 for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)
4791 __blk_mq_free_map_and_rqs(set, i);
4794 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4796 mutex_lock(&set->tag_list_lock);
4797 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4798 mutex_unlock(&set->tag_list_lock);
4800 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4802 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
4803 struct io_comp_batch *iob, unsigned int flags)
4805 long state = get_current_state();
4809 ret = q->mq_ops->poll(hctx, iob);
4811 __set_current_state(TASK_RUNNING);
4815 if (signal_pending_state(state, current))
4816 __set_current_state(TASK_RUNNING);
4817 if (task_is_running(current))
4820 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4823 } while (!need_resched());
4825 __set_current_state(TASK_RUNNING);
4829 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
4830 struct io_comp_batch *iob, unsigned int flags)
4832 struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie);
4834 return blk_hctx_poll(q, hctx, iob, flags);
4837 int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
4838 unsigned int poll_flags)
4840 struct request_queue *q = rq->q;
4843 if (!blk_rq_is_poll(rq))
4845 if (!percpu_ref_tryget(&q->q_usage_counter))
4848 ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags);
4853 EXPORT_SYMBOL_GPL(blk_rq_poll);
4855 unsigned int blk_mq_rq_cpu(struct request *rq)
4857 return rq->mq_ctx->cpu;
4859 EXPORT_SYMBOL(blk_mq_rq_cpu);
4861 void blk_mq_cancel_work_sync(struct request_queue *q)
4863 struct blk_mq_hw_ctx *hctx;
4866 cancel_delayed_work_sync(&q->requeue_work);
4868 queue_for_each_hw_ctx(q, hctx, i)
4869 cancel_delayed_work_sync(&hctx->run_work);
4872 static int __init blk_mq_init(void)
4876 for_each_possible_cpu(i)
4877 init_llist_head(&per_cpu(blk_cpu_done, i));
4878 for_each_possible_cpu(i)
4879 INIT_CSD(&per_cpu(blk_cpu_csd, i),
4880 __blk_mq_complete_request_remote, NULL);
4881 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4883 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4884 "block/softirq:dead", NULL,
4885 blk_softirq_cpu_dead);
4886 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4887 blk_mq_hctx_notify_dead);
4888 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4889 blk_mq_hctx_notify_online,
4890 blk_mq_hctx_notify_offline);
4893 subsys_initcall(blk_mq_init);