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? 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)))
1175 /* don't try to IPI to an offline CPU */
1176 return cpu_online(rq->mq_ctx->cpu);
1179 static void blk_mq_complete_send_ipi(struct request *rq)
1183 cpu = rq->mq_ctx->cpu;
1184 if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu)))
1185 smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu));
1188 static void blk_mq_raise_softirq(struct request *rq)
1190 struct llist_head *list;
1193 list = this_cpu_ptr(&blk_cpu_done);
1194 if (llist_add(&rq->ipi_list, list))
1195 raise_softirq(BLOCK_SOFTIRQ);
1199 bool blk_mq_complete_request_remote(struct request *rq)
1201 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1204 * For request which hctx has only one ctx mapping,
1205 * or a polled request, always complete locally,
1206 * it's pointless to redirect the completion.
1208 if ((rq->mq_hctx->nr_ctx == 1 &&
1209 rq->mq_ctx->cpu == raw_smp_processor_id()) ||
1210 rq->cmd_flags & REQ_POLLED)
1213 if (blk_mq_complete_need_ipi(rq)) {
1214 blk_mq_complete_send_ipi(rq);
1218 if (rq->q->nr_hw_queues == 1) {
1219 blk_mq_raise_softirq(rq);
1224 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1227 * blk_mq_complete_request - end I/O on a request
1228 * @rq: the request being processed
1231 * Complete a request by scheduling the ->complete_rq operation.
1233 void blk_mq_complete_request(struct request *rq)
1235 if (!blk_mq_complete_request_remote(rq))
1236 rq->q->mq_ops->complete(rq);
1238 EXPORT_SYMBOL(blk_mq_complete_request);
1241 * blk_mq_start_request - Start processing a request
1242 * @rq: Pointer to request to be started
1244 * Function used by device drivers to notify the block layer that a request
1245 * is going to be processed now, so blk layer can do proper initializations
1246 * such as starting the timeout timer.
1248 void blk_mq_start_request(struct request *rq)
1250 struct request_queue *q = rq->q;
1252 trace_block_rq_issue(rq);
1254 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) &&
1255 !blk_rq_is_passthrough(rq)) {
1256 rq->io_start_time_ns = blk_time_get_ns();
1257 rq->stats_sectors = blk_rq_sectors(rq);
1258 rq->rq_flags |= RQF_STATS;
1259 rq_qos_issue(q, rq);
1262 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1265 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1266 rq->mq_hctx->tags->rqs[rq->tag] = rq;
1268 #ifdef CONFIG_BLK_DEV_INTEGRITY
1269 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1270 q->integrity.profile->prepare_fn(rq);
1272 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1273 WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num);
1275 EXPORT_SYMBOL(blk_mq_start_request);
1278 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1279 * queues. This is important for md arrays to benefit from merging
1282 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1284 if (plug->multiple_queues)
1285 return BLK_MAX_REQUEST_COUNT * 2;
1286 return BLK_MAX_REQUEST_COUNT;
1289 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1291 struct request *last = rq_list_peek(&plug->mq_list);
1293 if (!plug->rq_count) {
1294 trace_block_plug(rq->q);
1295 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1296 (!blk_queue_nomerges(rq->q) &&
1297 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1298 blk_mq_flush_plug_list(plug, false);
1300 trace_block_plug(rq->q);
1303 if (!plug->multiple_queues && last && last->q != rq->q)
1304 plug->multiple_queues = true;
1306 * Any request allocated from sched tags can't be issued to
1307 * ->queue_rqs() directly
1309 if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))
1310 plug->has_elevator = true;
1312 rq_list_add(&plug->mq_list, rq);
1317 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1318 * @rq: request to insert
1319 * @at_head: insert request at head or tail of queue
1322 * Insert a fully prepared request at the back of the I/O scheduler queue
1323 * for execution. Don't wait for completion.
1326 * This function will invoke @done directly if the queue is dead.
1328 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1330 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1332 WARN_ON(irqs_disabled());
1333 WARN_ON(!blk_rq_is_passthrough(rq));
1335 blk_account_io_start(rq);
1338 * As plugging can be enabled for passthrough requests on a zoned
1339 * device, directly accessing the plug instead of using blk_mq_plug()
1340 * should not have any consequences.
1342 if (current->plug && !at_head) {
1343 blk_add_rq_to_plug(current->plug, rq);
1347 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1348 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
1350 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1352 struct blk_rq_wait {
1353 struct completion done;
1357 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1359 struct blk_rq_wait *wait = rq->end_io_data;
1362 complete(&wait->done);
1363 return RQ_END_IO_NONE;
1366 bool blk_rq_is_poll(struct request *rq)
1370 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1374 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1376 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1379 blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0);
1381 } while (!completion_done(wait));
1385 * blk_execute_rq - insert a request into queue for execution
1386 * @rq: request to insert
1387 * @at_head: insert request at head or tail of queue
1390 * Insert a fully prepared request at the back of the I/O scheduler queue
1391 * for execution and wait for completion.
1392 * Return: The blk_status_t result provided to blk_mq_end_request().
1394 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1396 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1397 struct blk_rq_wait wait = {
1398 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1401 WARN_ON(irqs_disabled());
1402 WARN_ON(!blk_rq_is_passthrough(rq));
1404 rq->end_io_data = &wait;
1405 rq->end_io = blk_end_sync_rq;
1407 blk_account_io_start(rq);
1408 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1409 blk_mq_run_hw_queue(hctx, false);
1411 if (blk_rq_is_poll(rq))
1412 blk_rq_poll_completion(rq, &wait.done);
1414 blk_wait_io(&wait.done);
1418 EXPORT_SYMBOL(blk_execute_rq);
1420 static void __blk_mq_requeue_request(struct request *rq)
1422 struct request_queue *q = rq->q;
1424 blk_mq_put_driver_tag(rq);
1426 trace_block_rq_requeue(rq);
1427 rq_qos_requeue(q, rq);
1429 if (blk_mq_request_started(rq)) {
1430 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1431 rq->rq_flags &= ~RQF_TIMED_OUT;
1435 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1437 struct request_queue *q = rq->q;
1438 unsigned long flags;
1440 __blk_mq_requeue_request(rq);
1442 /* this request will be re-inserted to io scheduler queue */
1443 blk_mq_sched_requeue_request(rq);
1445 spin_lock_irqsave(&q->requeue_lock, flags);
1446 list_add_tail(&rq->queuelist, &q->requeue_list);
1447 spin_unlock_irqrestore(&q->requeue_lock, flags);
1449 if (kick_requeue_list)
1450 blk_mq_kick_requeue_list(q);
1452 EXPORT_SYMBOL(blk_mq_requeue_request);
1454 static void blk_mq_requeue_work(struct work_struct *work)
1456 struct request_queue *q =
1457 container_of(work, struct request_queue, requeue_work.work);
1459 LIST_HEAD(flush_list);
1462 spin_lock_irq(&q->requeue_lock);
1463 list_splice_init(&q->requeue_list, &rq_list);
1464 list_splice_init(&q->flush_list, &flush_list);
1465 spin_unlock_irq(&q->requeue_lock);
1467 while (!list_empty(&rq_list)) {
1468 rq = list_entry(rq_list.next, struct request, queuelist);
1470 * If RQF_DONTPREP ist set, the request has been started by the
1471 * driver already and might have driver-specific data allocated
1472 * already. Insert it into the hctx dispatch list to avoid
1473 * block layer merges for the request.
1475 if (rq->rq_flags & RQF_DONTPREP) {
1476 list_del_init(&rq->queuelist);
1477 blk_mq_request_bypass_insert(rq, 0);
1479 list_del_init(&rq->queuelist);
1480 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1484 while (!list_empty(&flush_list)) {
1485 rq = list_entry(flush_list.next, struct request, queuelist);
1486 list_del_init(&rq->queuelist);
1487 blk_mq_insert_request(rq, 0);
1490 blk_mq_run_hw_queues(q, false);
1493 void blk_mq_kick_requeue_list(struct request_queue *q)
1495 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1497 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1499 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1500 unsigned long msecs)
1502 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1503 msecs_to_jiffies(msecs));
1505 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1507 static bool blk_is_flush_data_rq(struct request *rq)
1509 return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq);
1512 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1515 * If we find a request that isn't idle we know the queue is busy
1516 * as it's checked in the iter.
1517 * Return false to stop the iteration.
1519 * In case of queue quiesce, if one flush data request is completed,
1520 * don't count it as inflight given the flush sequence is suspended,
1521 * and the original flush data request is invisible to driver, just
1522 * like other pending requests because of quiesce
1524 if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) &&
1525 blk_is_flush_data_rq(rq) &&
1526 blk_mq_request_completed(rq))) {
1536 bool blk_mq_queue_inflight(struct request_queue *q)
1540 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1543 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1545 static void blk_mq_rq_timed_out(struct request *req)
1547 req->rq_flags |= RQF_TIMED_OUT;
1548 if (req->q->mq_ops->timeout) {
1549 enum blk_eh_timer_return ret;
1551 ret = req->q->mq_ops->timeout(req);
1552 if (ret == BLK_EH_DONE)
1554 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1560 struct blk_expired_data {
1561 bool has_timedout_rq;
1563 unsigned long timeout_start;
1566 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1568 unsigned long deadline;
1570 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1572 if (rq->rq_flags & RQF_TIMED_OUT)
1575 deadline = READ_ONCE(rq->deadline);
1576 if (time_after_eq(expired->timeout_start, deadline))
1579 if (expired->next == 0)
1580 expired->next = deadline;
1581 else if (time_after(expired->next, deadline))
1582 expired->next = deadline;
1586 void blk_mq_put_rq_ref(struct request *rq)
1588 if (is_flush_rq(rq)) {
1589 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1590 blk_mq_free_request(rq);
1591 } else if (req_ref_put_and_test(rq)) {
1592 __blk_mq_free_request(rq);
1596 static bool blk_mq_check_expired(struct request *rq, void *priv)
1598 struct blk_expired_data *expired = priv;
1601 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1602 * be reallocated underneath the timeout handler's processing, then
1603 * the expire check is reliable. If the request is not expired, then
1604 * it was completed and reallocated as a new request after returning
1605 * from blk_mq_check_expired().
1607 if (blk_mq_req_expired(rq, expired)) {
1608 expired->has_timedout_rq = true;
1614 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1616 struct blk_expired_data *expired = priv;
1618 if (blk_mq_req_expired(rq, expired))
1619 blk_mq_rq_timed_out(rq);
1623 static void blk_mq_timeout_work(struct work_struct *work)
1625 struct request_queue *q =
1626 container_of(work, struct request_queue, timeout_work);
1627 struct blk_expired_data expired = {
1628 .timeout_start = jiffies,
1630 struct blk_mq_hw_ctx *hctx;
1633 /* A deadlock might occur if a request is stuck requiring a
1634 * timeout at the same time a queue freeze is waiting
1635 * completion, since the timeout code would not be able to
1636 * acquire the queue reference here.
1638 * That's why we don't use blk_queue_enter here; instead, we use
1639 * percpu_ref_tryget directly, because we need to be able to
1640 * obtain a reference even in the short window between the queue
1641 * starting to freeze, by dropping the first reference in
1642 * blk_freeze_queue_start, and the moment the last request is
1643 * consumed, marked by the instant q_usage_counter reaches
1646 if (!percpu_ref_tryget(&q->q_usage_counter))
1649 /* check if there is any timed-out request */
1650 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1651 if (expired.has_timedout_rq) {
1653 * Before walking tags, we must ensure any submit started
1654 * before the current time has finished. Since the submit
1655 * uses srcu or rcu, wait for a synchronization point to
1656 * ensure all running submits have finished
1658 blk_mq_wait_quiesce_done(q->tag_set);
1661 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1664 if (expired.next != 0) {
1665 mod_timer(&q->timeout, expired.next);
1668 * Request timeouts are handled as a forward rolling timer. If
1669 * we end up here it means that no requests are pending and
1670 * also that no request has been pending for a while. Mark
1671 * each hctx as idle.
1673 queue_for_each_hw_ctx(q, hctx, i) {
1674 /* the hctx may be unmapped, so check it here */
1675 if (blk_mq_hw_queue_mapped(hctx))
1676 blk_mq_tag_idle(hctx);
1682 struct flush_busy_ctx_data {
1683 struct blk_mq_hw_ctx *hctx;
1684 struct list_head *list;
1687 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1689 struct flush_busy_ctx_data *flush_data = data;
1690 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1691 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1692 enum hctx_type type = hctx->type;
1694 spin_lock(&ctx->lock);
1695 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1696 sbitmap_clear_bit(sb, bitnr);
1697 spin_unlock(&ctx->lock);
1702 * Process software queues that have been marked busy, splicing them
1703 * to the for-dispatch
1705 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1707 struct flush_busy_ctx_data data = {
1712 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1714 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1716 struct dispatch_rq_data {
1717 struct blk_mq_hw_ctx *hctx;
1721 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1724 struct dispatch_rq_data *dispatch_data = data;
1725 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1726 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1727 enum hctx_type type = hctx->type;
1729 spin_lock(&ctx->lock);
1730 if (!list_empty(&ctx->rq_lists[type])) {
1731 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1732 list_del_init(&dispatch_data->rq->queuelist);
1733 if (list_empty(&ctx->rq_lists[type]))
1734 sbitmap_clear_bit(sb, bitnr);
1736 spin_unlock(&ctx->lock);
1738 return !dispatch_data->rq;
1741 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1742 struct blk_mq_ctx *start)
1744 unsigned off = start ? start->index_hw[hctx->type] : 0;
1745 struct dispatch_rq_data data = {
1750 __sbitmap_for_each_set(&hctx->ctx_map, off,
1751 dispatch_rq_from_ctx, &data);
1756 bool __blk_mq_alloc_driver_tag(struct request *rq)
1758 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1759 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1762 blk_mq_tag_busy(rq->mq_hctx);
1764 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1765 bt = &rq->mq_hctx->tags->breserved_tags;
1768 if (!hctx_may_queue(rq->mq_hctx, bt))
1772 tag = __sbitmap_queue_get(bt);
1773 if (tag == BLK_MQ_NO_TAG)
1776 rq->tag = tag + tag_offset;
1777 blk_mq_inc_active_requests(rq->mq_hctx);
1781 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1782 int flags, void *key)
1784 struct blk_mq_hw_ctx *hctx;
1786 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1788 spin_lock(&hctx->dispatch_wait_lock);
1789 if (!list_empty(&wait->entry)) {
1790 struct sbitmap_queue *sbq;
1792 list_del_init(&wait->entry);
1793 sbq = &hctx->tags->bitmap_tags;
1794 atomic_dec(&sbq->ws_active);
1796 spin_unlock(&hctx->dispatch_wait_lock);
1798 blk_mq_run_hw_queue(hctx, true);
1803 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1804 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1805 * restart. For both cases, take care to check the condition again after
1806 * marking us as waiting.
1808 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1811 struct sbitmap_queue *sbq;
1812 struct wait_queue_head *wq;
1813 wait_queue_entry_t *wait;
1816 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1817 !(blk_mq_is_shared_tags(hctx->flags))) {
1818 blk_mq_sched_mark_restart_hctx(hctx);
1821 * It's possible that a tag was freed in the window between the
1822 * allocation failure and adding the hardware queue to the wait
1825 * Don't clear RESTART here, someone else could have set it.
1826 * At most this will cost an extra queue run.
1828 return blk_mq_get_driver_tag(rq);
1831 wait = &hctx->dispatch_wait;
1832 if (!list_empty_careful(&wait->entry))
1835 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1836 sbq = &hctx->tags->breserved_tags;
1838 sbq = &hctx->tags->bitmap_tags;
1839 wq = &bt_wait_ptr(sbq, hctx)->wait;
1841 spin_lock_irq(&wq->lock);
1842 spin_lock(&hctx->dispatch_wait_lock);
1843 if (!list_empty(&wait->entry)) {
1844 spin_unlock(&hctx->dispatch_wait_lock);
1845 spin_unlock_irq(&wq->lock);
1849 atomic_inc(&sbq->ws_active);
1850 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1851 __add_wait_queue(wq, wait);
1854 * Add one explicit barrier since blk_mq_get_driver_tag() may
1855 * not imply barrier in case of failure.
1857 * Order adding us to wait queue and allocating driver tag.
1859 * The pair is the one implied in sbitmap_queue_wake_up() which
1860 * orders clearing sbitmap tag bits and waitqueue_active() in
1861 * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless
1863 * Otherwise, re-order of adding wait queue and getting driver tag
1864 * may cause __sbitmap_queue_wake_up() to wake up nothing because
1865 * the waitqueue_active() may not observe us in wait queue.
1870 * It's possible that a tag was freed in the window between the
1871 * allocation failure and adding the hardware queue to the wait
1874 ret = blk_mq_get_driver_tag(rq);
1876 spin_unlock(&hctx->dispatch_wait_lock);
1877 spin_unlock_irq(&wq->lock);
1882 * We got a tag, remove ourselves from the wait queue to ensure
1883 * someone else gets the wakeup.
1885 list_del_init(&wait->entry);
1886 atomic_dec(&sbq->ws_active);
1887 spin_unlock(&hctx->dispatch_wait_lock);
1888 spin_unlock_irq(&wq->lock);
1893 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1894 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1896 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1897 * - EWMA is one simple way to compute running average value
1898 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1899 * - take 4 as factor for avoiding to get too small(0) result, and this
1900 * factor doesn't matter because EWMA decreases exponentially
1902 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1906 ewma = hctx->dispatch_busy;
1911 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1913 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1914 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1916 hctx->dispatch_busy = ewma;
1919 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1921 static void blk_mq_handle_dev_resource(struct request *rq,
1922 struct list_head *list)
1924 list_add(&rq->queuelist, list);
1925 __blk_mq_requeue_request(rq);
1928 static void blk_mq_handle_zone_resource(struct request *rq,
1929 struct list_head *zone_list)
1932 * If we end up here it is because we cannot dispatch a request to a
1933 * specific zone due to LLD level zone-write locking or other zone
1934 * related resource not being available. In this case, set the request
1935 * aside in zone_list for retrying it later.
1937 list_add(&rq->queuelist, zone_list);
1938 __blk_mq_requeue_request(rq);
1941 enum prep_dispatch {
1943 PREP_DISPATCH_NO_TAG,
1944 PREP_DISPATCH_NO_BUDGET,
1947 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1950 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1951 int budget_token = -1;
1954 budget_token = blk_mq_get_dispatch_budget(rq->q);
1955 if (budget_token < 0) {
1956 blk_mq_put_driver_tag(rq);
1957 return PREP_DISPATCH_NO_BUDGET;
1959 blk_mq_set_rq_budget_token(rq, budget_token);
1962 if (!blk_mq_get_driver_tag(rq)) {
1964 * The initial allocation attempt failed, so we need to
1965 * rerun the hardware queue when a tag is freed. The
1966 * waitqueue takes care of that. If the queue is run
1967 * before we add this entry back on the dispatch list,
1968 * we'll re-run it below.
1970 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1972 * All budgets not got from this function will be put
1973 * together during handling partial dispatch
1976 blk_mq_put_dispatch_budget(rq->q, budget_token);
1977 return PREP_DISPATCH_NO_TAG;
1981 return PREP_DISPATCH_OK;
1984 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1985 static void blk_mq_release_budgets(struct request_queue *q,
1986 struct list_head *list)
1990 list_for_each_entry(rq, list, queuelist) {
1991 int budget_token = blk_mq_get_rq_budget_token(rq);
1993 if (budget_token >= 0)
1994 blk_mq_put_dispatch_budget(q, budget_token);
1999 * blk_mq_commit_rqs will notify driver using bd->last that there is no
2000 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
2002 * Attention, we should explicitly call this in unusual cases:
2003 * 1) did not queue everything initially scheduled to queue
2004 * 2) the last attempt to queue a request failed
2006 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
2009 if (hctx->queue->mq_ops->commit_rqs && queued) {
2010 trace_block_unplug(hctx->queue, queued, !from_schedule);
2011 hctx->queue->mq_ops->commit_rqs(hctx);
2016 * Returns true if we did some work AND can potentially do more.
2018 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
2019 unsigned int nr_budgets)
2021 enum prep_dispatch prep;
2022 struct request_queue *q = hctx->queue;
2025 blk_status_t ret = BLK_STS_OK;
2026 LIST_HEAD(zone_list);
2027 bool needs_resource = false;
2029 if (list_empty(list))
2033 * Now process all the entries, sending them to the driver.
2037 struct blk_mq_queue_data bd;
2039 rq = list_first_entry(list, struct request, queuelist);
2041 WARN_ON_ONCE(hctx != rq->mq_hctx);
2042 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2043 if (prep != PREP_DISPATCH_OK)
2046 list_del_init(&rq->queuelist);
2049 bd.last = list_empty(list);
2052 * once the request is queued to lld, no need to cover the
2057 ret = q->mq_ops->queue_rq(hctx, &bd);
2062 case BLK_STS_RESOURCE:
2063 needs_resource = true;
2065 case BLK_STS_DEV_RESOURCE:
2066 blk_mq_handle_dev_resource(rq, list);
2068 case BLK_STS_ZONE_RESOURCE:
2070 * Move the request to zone_list and keep going through
2071 * the dispatch list to find more requests the drive can
2074 blk_mq_handle_zone_resource(rq, &zone_list);
2075 needs_resource = true;
2078 blk_mq_end_request(rq, ret);
2080 } while (!list_empty(list));
2082 if (!list_empty(&zone_list))
2083 list_splice_tail_init(&zone_list, list);
2085 /* If we didn't flush the entire list, we could have told the driver
2086 * there was more coming, but that turned out to be a lie.
2088 if (!list_empty(list) || ret != BLK_STS_OK)
2089 blk_mq_commit_rqs(hctx, queued, false);
2092 * Any items that need requeuing? Stuff them into hctx->dispatch,
2093 * that is where we will continue on next queue run.
2095 if (!list_empty(list)) {
2097 /* For non-shared tags, the RESTART check will suffice */
2098 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2099 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2100 blk_mq_is_shared_tags(hctx->flags));
2103 blk_mq_release_budgets(q, list);
2105 spin_lock(&hctx->lock);
2106 list_splice_tail_init(list, &hctx->dispatch);
2107 spin_unlock(&hctx->lock);
2110 * Order adding requests to hctx->dispatch and checking
2111 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2112 * in blk_mq_sched_restart(). Avoid restart code path to
2113 * miss the new added requests to hctx->dispatch, meantime
2114 * SCHED_RESTART is observed here.
2119 * If SCHED_RESTART was set by the caller of this function and
2120 * it is no longer set that means that it was cleared by another
2121 * thread and hence that a queue rerun is needed.
2123 * If 'no_tag' is set, that means that we failed getting
2124 * a driver tag with an I/O scheduler attached. If our dispatch
2125 * waitqueue is no longer active, ensure that we run the queue
2126 * AFTER adding our entries back to the list.
2128 * If no I/O scheduler has been configured it is possible that
2129 * the hardware queue got stopped and restarted before requests
2130 * were pushed back onto the dispatch list. Rerun the queue to
2131 * avoid starvation. Notes:
2132 * - blk_mq_run_hw_queue() checks whether or not a queue has
2133 * been stopped before rerunning a queue.
2134 * - Some but not all block drivers stop a queue before
2135 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2138 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2139 * bit is set, run queue after a delay to avoid IO stalls
2140 * that could otherwise occur if the queue is idle. We'll do
2141 * similar if we couldn't get budget or couldn't lock a zone
2142 * and SCHED_RESTART is set.
2144 needs_restart = blk_mq_sched_needs_restart(hctx);
2145 if (prep == PREP_DISPATCH_NO_BUDGET)
2146 needs_resource = true;
2147 if (!needs_restart ||
2148 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2149 blk_mq_run_hw_queue(hctx, true);
2150 else if (needs_resource)
2151 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2153 blk_mq_update_dispatch_busy(hctx, true);
2157 blk_mq_update_dispatch_busy(hctx, false);
2161 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2163 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2165 if (cpu >= nr_cpu_ids)
2166 cpu = cpumask_first(hctx->cpumask);
2171 * It'd be great if the workqueue API had a way to pass
2172 * in a mask and had some smarts for more clever placement.
2173 * For now we just round-robin here, switching for every
2174 * BLK_MQ_CPU_WORK_BATCH queued items.
2176 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2179 int next_cpu = hctx->next_cpu;
2181 if (hctx->queue->nr_hw_queues == 1)
2182 return WORK_CPU_UNBOUND;
2184 if (--hctx->next_cpu_batch <= 0) {
2186 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2188 if (next_cpu >= nr_cpu_ids)
2189 next_cpu = blk_mq_first_mapped_cpu(hctx);
2190 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2194 * Do unbound schedule if we can't find a online CPU for this hctx,
2195 * and it should only happen in the path of handling CPU DEAD.
2197 if (!cpu_online(next_cpu)) {
2204 * Make sure to re-select CPU next time once after CPUs
2205 * in hctx->cpumask become online again.
2207 hctx->next_cpu = next_cpu;
2208 hctx->next_cpu_batch = 1;
2209 return WORK_CPU_UNBOUND;
2212 hctx->next_cpu = next_cpu;
2217 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2218 * @hctx: Pointer to the hardware queue to run.
2219 * @msecs: Milliseconds of delay to wait before running the queue.
2221 * Run a hardware queue asynchronously with a delay of @msecs.
2223 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2225 if (unlikely(blk_mq_hctx_stopped(hctx)))
2227 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2228 msecs_to_jiffies(msecs));
2230 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2233 * blk_mq_run_hw_queue - Start to run a hardware queue.
2234 * @hctx: Pointer to the hardware queue to run.
2235 * @async: If we want to run the queue asynchronously.
2237 * Check if the request queue is not in a quiesced state and if there are
2238 * pending requests to be sent. If this is true, run the queue to send requests
2241 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2246 * We can't run the queue inline with interrupts disabled.
2248 WARN_ON_ONCE(!async && in_interrupt());
2250 might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING);
2253 * When queue is quiesced, we may be switching io scheduler, or
2254 * updating nr_hw_queues, or other things, and we can't run queue
2255 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2257 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2260 __blk_mq_run_dispatch_ops(hctx->queue, false,
2261 need_run = !blk_queue_quiesced(hctx->queue) &&
2262 blk_mq_hctx_has_pending(hctx));
2267 if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2268 blk_mq_delay_run_hw_queue(hctx, 0);
2272 blk_mq_run_dispatch_ops(hctx->queue,
2273 blk_mq_sched_dispatch_requests(hctx));
2275 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2278 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2281 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2283 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2285 * If the IO scheduler does not respect hardware queues when
2286 * dispatching, we just don't bother with multiple HW queues and
2287 * dispatch from hctx for the current CPU since running multiple queues
2288 * just causes lock contention inside the scheduler and pointless cache
2291 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2293 if (!blk_mq_hctx_stopped(hctx))
2299 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2300 * @q: Pointer to the request queue to run.
2301 * @async: If we want to run the queue asynchronously.
2303 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2305 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2309 if (blk_queue_sq_sched(q))
2310 sq_hctx = blk_mq_get_sq_hctx(q);
2311 queue_for_each_hw_ctx(q, hctx, i) {
2312 if (blk_mq_hctx_stopped(hctx))
2315 * Dispatch from this hctx either if there's no hctx preferred
2316 * by IO scheduler or if it has requests that bypass the
2319 if (!sq_hctx || sq_hctx == hctx ||
2320 !list_empty_careful(&hctx->dispatch))
2321 blk_mq_run_hw_queue(hctx, async);
2324 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2327 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2328 * @q: Pointer to the request queue to run.
2329 * @msecs: Milliseconds of delay to wait before running the queues.
2331 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2333 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2337 if (blk_queue_sq_sched(q))
2338 sq_hctx = blk_mq_get_sq_hctx(q);
2339 queue_for_each_hw_ctx(q, hctx, i) {
2340 if (blk_mq_hctx_stopped(hctx))
2343 * If there is already a run_work pending, leave the
2344 * pending delay untouched. Otherwise, a hctx can stall
2345 * if another hctx is re-delaying the other's work
2346 * before the work executes.
2348 if (delayed_work_pending(&hctx->run_work))
2351 * Dispatch from this hctx either if there's no hctx preferred
2352 * by IO scheduler or if it has requests that bypass the
2355 if (!sq_hctx || sq_hctx == hctx ||
2356 !list_empty_careful(&hctx->dispatch))
2357 blk_mq_delay_run_hw_queue(hctx, msecs);
2360 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2363 * This function is often used for pausing .queue_rq() by driver when
2364 * there isn't enough resource or some conditions aren't satisfied, and
2365 * BLK_STS_RESOURCE is usually returned.
2367 * We do not guarantee that dispatch can be drained or blocked
2368 * after blk_mq_stop_hw_queue() returns. Please use
2369 * blk_mq_quiesce_queue() for that requirement.
2371 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2373 cancel_delayed_work(&hctx->run_work);
2375 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2377 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2380 * This function is often used for pausing .queue_rq() by driver when
2381 * there isn't enough resource or some conditions aren't satisfied, and
2382 * BLK_STS_RESOURCE is usually returned.
2384 * We do not guarantee that dispatch can be drained or blocked
2385 * after blk_mq_stop_hw_queues() returns. Please use
2386 * blk_mq_quiesce_queue() for that requirement.
2388 void blk_mq_stop_hw_queues(struct request_queue *q)
2390 struct blk_mq_hw_ctx *hctx;
2393 queue_for_each_hw_ctx(q, hctx, i)
2394 blk_mq_stop_hw_queue(hctx);
2396 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2398 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2400 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2402 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
2404 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2406 void blk_mq_start_hw_queues(struct request_queue *q)
2408 struct blk_mq_hw_ctx *hctx;
2411 queue_for_each_hw_ctx(q, hctx, i)
2412 blk_mq_start_hw_queue(hctx);
2414 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2416 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2418 if (!blk_mq_hctx_stopped(hctx))
2421 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2422 blk_mq_run_hw_queue(hctx, async);
2424 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2426 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2428 struct blk_mq_hw_ctx *hctx;
2431 queue_for_each_hw_ctx(q, hctx, i)
2432 blk_mq_start_stopped_hw_queue(hctx, async ||
2433 (hctx->flags & BLK_MQ_F_BLOCKING));
2435 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2437 static void blk_mq_run_work_fn(struct work_struct *work)
2439 struct blk_mq_hw_ctx *hctx =
2440 container_of(work, struct blk_mq_hw_ctx, run_work.work);
2442 blk_mq_run_dispatch_ops(hctx->queue,
2443 blk_mq_sched_dispatch_requests(hctx));
2447 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2448 * @rq: Pointer to request to be inserted.
2449 * @flags: BLK_MQ_INSERT_*
2451 * Should only be used carefully, when the caller knows we want to
2452 * bypass a potential IO scheduler on the target device.
2454 static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2456 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2458 spin_lock(&hctx->lock);
2459 if (flags & BLK_MQ_INSERT_AT_HEAD)
2460 list_add(&rq->queuelist, &hctx->dispatch);
2462 list_add_tail(&rq->queuelist, &hctx->dispatch);
2463 spin_unlock(&hctx->lock);
2466 static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2467 struct blk_mq_ctx *ctx, struct list_head *list,
2468 bool run_queue_async)
2471 enum hctx_type type = hctx->type;
2474 * Try to issue requests directly if the hw queue isn't busy to save an
2475 * extra enqueue & dequeue to the sw queue.
2477 if (!hctx->dispatch_busy && !run_queue_async) {
2478 blk_mq_run_dispatch_ops(hctx->queue,
2479 blk_mq_try_issue_list_directly(hctx, list));
2480 if (list_empty(list))
2485 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2488 list_for_each_entry(rq, list, queuelist) {
2489 BUG_ON(rq->mq_ctx != ctx);
2490 trace_block_rq_insert(rq);
2491 if (rq->cmd_flags & REQ_NOWAIT)
2492 run_queue_async = true;
2495 spin_lock(&ctx->lock);
2496 list_splice_tail_init(list, &ctx->rq_lists[type]);
2497 blk_mq_hctx_mark_pending(hctx, ctx);
2498 spin_unlock(&ctx->lock);
2500 blk_mq_run_hw_queue(hctx, run_queue_async);
2503 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2505 struct request_queue *q = rq->q;
2506 struct blk_mq_ctx *ctx = rq->mq_ctx;
2507 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2509 if (blk_rq_is_passthrough(rq)) {
2511 * Passthrough request have to be added to hctx->dispatch
2512 * directly. The device may be in a situation where it can't
2513 * handle FS request, and always returns BLK_STS_RESOURCE for
2514 * them, which gets them added to hctx->dispatch.
2516 * If a passthrough request is required to unblock the queues,
2517 * and it is added to the scheduler queue, there is no chance to
2518 * dispatch it given we prioritize requests in hctx->dispatch.
2520 blk_mq_request_bypass_insert(rq, flags);
2521 } else if (req_op(rq) == REQ_OP_FLUSH) {
2523 * Firstly normal IO request is inserted to scheduler queue or
2524 * sw queue, meantime we add flush request to dispatch queue(
2525 * hctx->dispatch) directly and there is at most one in-flight
2526 * flush request for each hw queue, so it doesn't matter to add
2527 * flush request to tail or front of the dispatch queue.
2529 * Secondly in case of NCQ, flush request belongs to non-NCQ
2530 * command, and queueing it will fail when there is any
2531 * in-flight normal IO request(NCQ command). When adding flush
2532 * rq to the front of hctx->dispatch, it is easier to introduce
2533 * extra time to flush rq's latency because of S_SCHED_RESTART
2534 * compared with adding to the tail of dispatch queue, then
2535 * chance of flush merge is increased, and less flush requests
2536 * will be issued to controller. It is observed that ~10% time
2537 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2538 * drive when adding flush rq to the front of hctx->dispatch.
2540 * Simply queue flush rq to the front of hctx->dispatch so that
2541 * intensive flush workloads can benefit in case of NCQ HW.
2543 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2544 } else if (q->elevator) {
2547 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2549 list_add(&rq->queuelist, &list);
2550 q->elevator->type->ops.insert_requests(hctx, &list, flags);
2552 trace_block_rq_insert(rq);
2554 spin_lock(&ctx->lock);
2555 if (flags & BLK_MQ_INSERT_AT_HEAD)
2556 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2558 list_add_tail(&rq->queuelist,
2559 &ctx->rq_lists[hctx->type]);
2560 blk_mq_hctx_mark_pending(hctx, ctx);
2561 spin_unlock(&ctx->lock);
2565 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2566 unsigned int nr_segs)
2570 if (bio->bi_opf & REQ_RAHEAD)
2571 rq->cmd_flags |= REQ_FAILFAST_MASK;
2573 rq->__sector = bio->bi_iter.bi_sector;
2574 blk_rq_bio_prep(rq, bio, nr_segs);
2576 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2577 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2580 blk_account_io_start(rq);
2583 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2584 struct request *rq, bool last)
2586 struct request_queue *q = rq->q;
2587 struct blk_mq_queue_data bd = {
2594 * For OK queue, we are done. For error, caller may kill it.
2595 * Any other error (busy), just add it to our list as we
2596 * previously would have done.
2598 ret = q->mq_ops->queue_rq(hctx, &bd);
2601 blk_mq_update_dispatch_busy(hctx, false);
2603 case BLK_STS_RESOURCE:
2604 case BLK_STS_DEV_RESOURCE:
2605 blk_mq_update_dispatch_busy(hctx, true);
2606 __blk_mq_requeue_request(rq);
2609 blk_mq_update_dispatch_busy(hctx, false);
2616 static bool blk_mq_get_budget_and_tag(struct request *rq)
2620 budget_token = blk_mq_get_dispatch_budget(rq->q);
2621 if (budget_token < 0)
2623 blk_mq_set_rq_budget_token(rq, budget_token);
2624 if (!blk_mq_get_driver_tag(rq)) {
2625 blk_mq_put_dispatch_budget(rq->q, budget_token);
2632 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2633 * @hctx: Pointer of the associated hardware queue.
2634 * @rq: Pointer to request to be sent.
2636 * If the device has enough resources to accept a new request now, send the
2637 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2638 * we can try send it another time in the future. Requests inserted at this
2639 * queue have higher priority.
2641 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2646 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2647 blk_mq_insert_request(rq, 0);
2651 if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
2652 blk_mq_insert_request(rq, 0);
2653 blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT);
2657 ret = __blk_mq_issue_directly(hctx, rq, true);
2661 case BLK_STS_RESOURCE:
2662 case BLK_STS_DEV_RESOURCE:
2663 blk_mq_request_bypass_insert(rq, 0);
2664 blk_mq_run_hw_queue(hctx, false);
2667 blk_mq_end_request(rq, ret);
2672 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2674 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2676 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2677 blk_mq_insert_request(rq, 0);
2681 if (!blk_mq_get_budget_and_tag(rq))
2682 return BLK_STS_RESOURCE;
2683 return __blk_mq_issue_directly(hctx, rq, last);
2686 static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2688 struct blk_mq_hw_ctx *hctx = NULL;
2691 blk_status_t ret = BLK_STS_OK;
2693 while ((rq = rq_list_pop(&plug->mq_list))) {
2694 bool last = rq_list_empty(plug->mq_list);
2696 if (hctx != rq->mq_hctx) {
2698 blk_mq_commit_rqs(hctx, queued, false);
2704 ret = blk_mq_request_issue_directly(rq, last);
2709 case BLK_STS_RESOURCE:
2710 case BLK_STS_DEV_RESOURCE:
2711 blk_mq_request_bypass_insert(rq, 0);
2712 blk_mq_run_hw_queue(hctx, false);
2715 blk_mq_end_request(rq, ret);
2721 if (ret != BLK_STS_OK)
2722 blk_mq_commit_rqs(hctx, queued, false);
2725 static void __blk_mq_flush_plug_list(struct request_queue *q,
2726 struct blk_plug *plug)
2728 if (blk_queue_quiesced(q))
2730 q->mq_ops->queue_rqs(&plug->mq_list);
2733 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2735 struct blk_mq_hw_ctx *this_hctx = NULL;
2736 struct blk_mq_ctx *this_ctx = NULL;
2737 struct request *requeue_list = NULL;
2738 struct request **requeue_lastp = &requeue_list;
2739 unsigned int depth = 0;
2740 bool is_passthrough = false;
2744 struct request *rq = rq_list_pop(&plug->mq_list);
2747 this_hctx = rq->mq_hctx;
2748 this_ctx = rq->mq_ctx;
2749 is_passthrough = blk_rq_is_passthrough(rq);
2750 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2751 is_passthrough != blk_rq_is_passthrough(rq)) {
2752 rq_list_add_tail(&requeue_lastp, rq);
2755 list_add(&rq->queuelist, &list);
2757 } while (!rq_list_empty(plug->mq_list));
2759 plug->mq_list = requeue_list;
2760 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2762 percpu_ref_get(&this_hctx->queue->q_usage_counter);
2763 /* passthrough requests should never be issued to the I/O scheduler */
2764 if (is_passthrough) {
2765 spin_lock(&this_hctx->lock);
2766 list_splice_tail_init(&list, &this_hctx->dispatch);
2767 spin_unlock(&this_hctx->lock);
2768 blk_mq_run_hw_queue(this_hctx, from_sched);
2769 } else if (this_hctx->queue->elevator) {
2770 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2772 blk_mq_run_hw_queue(this_hctx, from_sched);
2774 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2776 percpu_ref_put(&this_hctx->queue->q_usage_counter);
2779 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2784 * We may have been called recursively midway through handling
2785 * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2786 * To avoid mq_list changing under our feet, clear rq_count early and
2787 * bail out specifically if rq_count is 0 rather than checking
2788 * whether the mq_list is empty.
2790 if (plug->rq_count == 0)
2794 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2795 struct request_queue *q;
2797 rq = rq_list_peek(&plug->mq_list);
2801 * Peek first request and see if we have a ->queue_rqs() hook.
2802 * If we do, we can dispatch the whole plug list in one go. We
2803 * already know at this point that all requests belong to the
2804 * same queue, caller must ensure that's the case.
2806 if (q->mq_ops->queue_rqs) {
2807 blk_mq_run_dispatch_ops(q,
2808 __blk_mq_flush_plug_list(q, plug));
2809 if (rq_list_empty(plug->mq_list))
2813 blk_mq_run_dispatch_ops(q,
2814 blk_mq_plug_issue_direct(plug));
2815 if (rq_list_empty(plug->mq_list))
2820 blk_mq_dispatch_plug_list(plug, from_schedule);
2821 } while (!rq_list_empty(plug->mq_list));
2824 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2825 struct list_head *list)
2828 blk_status_t ret = BLK_STS_OK;
2830 while (!list_empty(list)) {
2831 struct request *rq = list_first_entry(list, struct request,
2834 list_del_init(&rq->queuelist);
2835 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2840 case BLK_STS_RESOURCE:
2841 case BLK_STS_DEV_RESOURCE:
2842 blk_mq_request_bypass_insert(rq, 0);
2843 if (list_empty(list))
2844 blk_mq_run_hw_queue(hctx, false);
2847 blk_mq_end_request(rq, ret);
2853 if (ret != BLK_STS_OK)
2854 blk_mq_commit_rqs(hctx, queued, false);
2857 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2858 struct bio *bio, unsigned int nr_segs)
2860 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2861 if (blk_attempt_plug_merge(q, bio, nr_segs))
2863 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2869 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2870 struct blk_plug *plug,
2874 struct blk_mq_alloc_data data = {
2877 .cmd_flags = bio->bi_opf,
2881 rq_qos_throttle(q, bio);
2884 data.nr_tags = plug->nr_ios;
2886 data.cached_rq = &plug->cached_rq;
2889 rq = __blk_mq_alloc_requests(&data);
2892 rq_qos_cleanup(q, bio);
2893 if (bio->bi_opf & REQ_NOWAIT)
2894 bio_wouldblock_error(bio);
2899 * Check if there is a suitable cached request and return it.
2901 static struct request *blk_mq_peek_cached_request(struct blk_plug *plug,
2902 struct request_queue *q, blk_opf_t opf)
2904 enum hctx_type type = blk_mq_get_hctx_type(opf);
2909 rq = rq_list_peek(&plug->cached_rq);
2910 if (!rq || rq->q != q)
2912 if (type != rq->mq_hctx->type &&
2913 (type != HCTX_TYPE_READ || rq->mq_hctx->type != HCTX_TYPE_DEFAULT))
2915 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
2920 static void blk_mq_use_cached_rq(struct request *rq, struct blk_plug *plug,
2923 WARN_ON_ONCE(rq_list_peek(&plug->cached_rq) != rq);
2926 * If any qos ->throttle() end up blocking, we will have flushed the
2927 * plug and hence killed the cached_rq list as well. Pop this entry
2928 * before we throttle.
2930 plug->cached_rq = rq_list_next(rq);
2931 rq_qos_throttle(rq->q, bio);
2933 blk_mq_rq_time_init(rq, 0);
2934 rq->cmd_flags = bio->bi_opf;
2935 INIT_LIST_HEAD(&rq->queuelist);
2939 * blk_mq_submit_bio - Create and send a request to block device.
2940 * @bio: Bio pointer.
2942 * Builds up a request structure from @q and @bio and send to the device. The
2943 * request may not be queued directly to hardware if:
2944 * * This request can be merged with another one
2945 * * We want to place request at plug queue for possible future merging
2946 * * There is an IO scheduler active at this queue
2948 * It will not queue the request if there is an error with the bio, or at the
2951 void blk_mq_submit_bio(struct bio *bio)
2953 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2954 struct blk_plug *plug = blk_mq_plug(bio);
2955 const int is_sync = op_is_sync(bio->bi_opf);
2956 struct blk_mq_hw_ctx *hctx;
2957 unsigned int nr_segs = 1;
2961 bio = blk_queue_bounce(bio, q);
2964 * If the plug has a cached request for this queue, try use it.
2966 * The cached request already holds a q_usage_counter reference and we
2967 * don't have to acquire a new one if we use it.
2969 rq = blk_mq_peek_cached_request(plug, q, bio->bi_opf);
2971 if (unlikely(bio_queue_enter(bio)))
2975 if (unlikely(bio_may_exceed_limits(bio, &q->limits))) {
2976 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2980 if (!bio_integrity_prep(bio))
2983 if (blk_mq_attempt_bio_merge(q, bio, nr_segs))
2987 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2991 blk_mq_use_cached_rq(rq, plug, bio);
2994 trace_block_getrq(bio);
2996 rq_qos_track(q, rq, bio);
2998 blk_mq_bio_to_request(rq, bio, nr_segs);
3000 ret = blk_crypto_rq_get_keyslot(rq);
3001 if (ret != BLK_STS_OK) {
3002 bio->bi_status = ret;
3004 blk_mq_free_request(rq);
3008 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
3012 blk_add_rq_to_plug(plug, rq);
3017 if ((rq->rq_flags & RQF_USE_SCHED) ||
3018 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
3019 blk_mq_insert_request(rq, 0);
3020 blk_mq_run_hw_queue(hctx, true);
3022 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
3028 * Don't drop the queue reference if we were trying to use a cached
3029 * request and thus didn't acquire one.
3035 #ifdef CONFIG_BLK_MQ_STACKING
3037 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3038 * @rq: the request being queued
3040 blk_status_t blk_insert_cloned_request(struct request *rq)
3042 struct request_queue *q = rq->q;
3043 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3044 unsigned int max_segments = blk_rq_get_max_segments(rq);
3047 if (blk_rq_sectors(rq) > max_sectors) {
3049 * SCSI device does not have a good way to return if
3050 * Write Same/Zero is actually supported. If a device rejects
3051 * a non-read/write command (discard, write same,etc.) the
3052 * low-level device driver will set the relevant queue limit to
3053 * 0 to prevent blk-lib from issuing more of the offending
3054 * operations. Commands queued prior to the queue limit being
3055 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3056 * errors being propagated to upper layers.
3058 if (max_sectors == 0)
3059 return BLK_STS_NOTSUPP;
3061 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3062 __func__, blk_rq_sectors(rq), max_sectors);
3063 return BLK_STS_IOERR;
3067 * The queue settings related to segment counting may differ from the
3070 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3071 if (rq->nr_phys_segments > max_segments) {
3072 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3073 __func__, rq->nr_phys_segments, max_segments);
3074 return BLK_STS_IOERR;
3077 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3078 return BLK_STS_IOERR;
3080 ret = blk_crypto_rq_get_keyslot(rq);
3081 if (ret != BLK_STS_OK)
3084 blk_account_io_start(rq);
3087 * Since we have a scheduler attached on the top device,
3088 * bypass a potential scheduler on the bottom device for
3091 blk_mq_run_dispatch_ops(q,
3092 ret = blk_mq_request_issue_directly(rq, true));
3094 blk_account_io_done(rq, blk_time_get_ns());
3097 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3100 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3101 * @rq: the clone request to be cleaned up
3104 * Free all bios in @rq for a cloned request.
3106 void blk_rq_unprep_clone(struct request *rq)
3110 while ((bio = rq->bio) != NULL) {
3111 rq->bio = bio->bi_next;
3116 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3119 * blk_rq_prep_clone - Helper function to setup clone request
3120 * @rq: the request to be setup
3121 * @rq_src: original request to be cloned
3122 * @bs: bio_set that bios for clone are allocated from
3123 * @gfp_mask: memory allocation mask for bio
3124 * @bio_ctr: setup function to be called for each clone bio.
3125 * Returns %0 for success, non %0 for failure.
3126 * @data: private data to be passed to @bio_ctr
3129 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3130 * Also, pages which the original bios are pointing to are not copied
3131 * and the cloned bios just point same pages.
3132 * So cloned bios must be completed before original bios, which means
3133 * the caller must complete @rq before @rq_src.
3135 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3136 struct bio_set *bs, gfp_t gfp_mask,
3137 int (*bio_ctr)(struct bio *, struct bio *, void *),
3140 struct bio *bio, *bio_src;
3145 __rq_for_each_bio(bio_src, rq_src) {
3146 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3151 if (bio_ctr && bio_ctr(bio, bio_src, data))
3155 rq->biotail->bi_next = bio;
3158 rq->bio = rq->biotail = bio;
3163 /* Copy attributes of the original request to the clone request. */
3164 rq->__sector = blk_rq_pos(rq_src);
3165 rq->__data_len = blk_rq_bytes(rq_src);
3166 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3167 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3168 rq->special_vec = rq_src->special_vec;
3170 rq->nr_phys_segments = rq_src->nr_phys_segments;
3171 rq->ioprio = rq_src->ioprio;
3173 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3181 blk_rq_unprep_clone(rq);
3185 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3186 #endif /* CONFIG_BLK_MQ_STACKING */
3189 * Steal bios from a request and add them to a bio list.
3190 * The request must not have been partially completed before.
3192 void blk_steal_bios(struct bio_list *list, struct request *rq)
3196 list->tail->bi_next = rq->bio;
3198 list->head = rq->bio;
3199 list->tail = rq->biotail;
3207 EXPORT_SYMBOL_GPL(blk_steal_bios);
3209 static size_t order_to_size(unsigned int order)
3211 return (size_t)PAGE_SIZE << order;
3214 /* called before freeing request pool in @tags */
3215 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3216 struct blk_mq_tags *tags)
3219 unsigned long flags;
3222 * There is no need to clear mapping if driver tags is not initialized
3223 * or the mapping belongs to the driver tags.
3225 if (!drv_tags || drv_tags == tags)
3228 list_for_each_entry(page, &tags->page_list, lru) {
3229 unsigned long start = (unsigned long)page_address(page);
3230 unsigned long end = start + order_to_size(page->private);
3233 for (i = 0; i < drv_tags->nr_tags; i++) {
3234 struct request *rq = drv_tags->rqs[i];
3235 unsigned long rq_addr = (unsigned long)rq;
3237 if (rq_addr >= start && rq_addr < end) {
3238 WARN_ON_ONCE(req_ref_read(rq) != 0);
3239 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3245 * Wait until all pending iteration is done.
3247 * Request reference is cleared and it is guaranteed to be observed
3248 * after the ->lock is released.
3250 spin_lock_irqsave(&drv_tags->lock, flags);
3251 spin_unlock_irqrestore(&drv_tags->lock, flags);
3254 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3255 unsigned int hctx_idx)
3257 struct blk_mq_tags *drv_tags;
3260 if (list_empty(&tags->page_list))
3263 if (blk_mq_is_shared_tags(set->flags))
3264 drv_tags = set->shared_tags;
3266 drv_tags = set->tags[hctx_idx];
3268 if (tags->static_rqs && set->ops->exit_request) {
3271 for (i = 0; i < tags->nr_tags; i++) {
3272 struct request *rq = tags->static_rqs[i];
3276 set->ops->exit_request(set, rq, hctx_idx);
3277 tags->static_rqs[i] = NULL;
3281 blk_mq_clear_rq_mapping(drv_tags, tags);
3283 while (!list_empty(&tags->page_list)) {
3284 page = list_first_entry(&tags->page_list, struct page, lru);
3285 list_del_init(&page->lru);
3287 * Remove kmemleak object previously allocated in
3288 * blk_mq_alloc_rqs().
3290 kmemleak_free(page_address(page));
3291 __free_pages(page, page->private);
3295 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3299 kfree(tags->static_rqs);
3300 tags->static_rqs = NULL;
3302 blk_mq_free_tags(tags);
3305 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3306 unsigned int hctx_idx)
3310 for (i = 0; i < set->nr_maps; i++) {
3311 unsigned int start = set->map[i].queue_offset;
3312 unsigned int end = start + set->map[i].nr_queues;
3314 if (hctx_idx >= start && hctx_idx < end)
3318 if (i >= set->nr_maps)
3319 i = HCTX_TYPE_DEFAULT;
3324 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3325 unsigned int hctx_idx)
3327 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3329 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3332 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3333 unsigned int hctx_idx,
3334 unsigned int nr_tags,
3335 unsigned int reserved_tags)
3337 int node = blk_mq_get_hctx_node(set, hctx_idx);
3338 struct blk_mq_tags *tags;
3340 if (node == NUMA_NO_NODE)
3341 node = set->numa_node;
3343 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3344 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3348 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3349 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3354 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3355 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3357 if (!tags->static_rqs)
3365 blk_mq_free_tags(tags);
3369 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3370 unsigned int hctx_idx, int node)
3374 if (set->ops->init_request) {
3375 ret = set->ops->init_request(set, rq, hctx_idx, node);
3380 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3384 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3385 struct blk_mq_tags *tags,
3386 unsigned int hctx_idx, unsigned int depth)
3388 unsigned int i, j, entries_per_page, max_order = 4;
3389 int node = blk_mq_get_hctx_node(set, hctx_idx);
3390 size_t rq_size, left;
3392 if (node == NUMA_NO_NODE)
3393 node = set->numa_node;
3395 INIT_LIST_HEAD(&tags->page_list);
3398 * rq_size is the size of the request plus driver payload, rounded
3399 * to the cacheline size
3401 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3403 left = rq_size * depth;
3405 for (i = 0; i < depth; ) {
3406 int this_order = max_order;
3411 while (this_order && left < order_to_size(this_order - 1))
3415 page = alloc_pages_node(node,
3416 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3422 if (order_to_size(this_order) < rq_size)
3429 page->private = this_order;
3430 list_add_tail(&page->lru, &tags->page_list);
3432 p = page_address(page);
3434 * Allow kmemleak to scan these pages as they contain pointers
3435 * to additional allocations like via ops->init_request().
3437 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3438 entries_per_page = order_to_size(this_order) / rq_size;
3439 to_do = min(entries_per_page, depth - i);
3440 left -= to_do * rq_size;
3441 for (j = 0; j < to_do; j++) {
3442 struct request *rq = p;
3444 tags->static_rqs[i] = rq;
3445 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3446 tags->static_rqs[i] = NULL;
3457 blk_mq_free_rqs(set, tags, hctx_idx);
3461 struct rq_iter_data {
3462 struct blk_mq_hw_ctx *hctx;
3466 static bool blk_mq_has_request(struct request *rq, void *data)
3468 struct rq_iter_data *iter_data = data;
3470 if (rq->mq_hctx != iter_data->hctx)
3472 iter_data->has_rq = true;
3476 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3478 struct blk_mq_tags *tags = hctx->sched_tags ?
3479 hctx->sched_tags : hctx->tags;
3480 struct rq_iter_data data = {
3484 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3488 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3489 struct blk_mq_hw_ctx *hctx)
3491 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3493 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3498 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3500 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3501 struct blk_mq_hw_ctx, cpuhp_online);
3503 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3504 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3508 * Prevent new request from being allocated on the current hctx.
3510 * The smp_mb__after_atomic() Pairs with the implied barrier in
3511 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3512 * seen once we return from the tag allocator.
3514 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3515 smp_mb__after_atomic();
3518 * Try to grab a reference to the queue and wait for any outstanding
3519 * requests. If we could not grab a reference the queue has been
3520 * frozen and there are no requests.
3522 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3523 while (blk_mq_hctx_has_requests(hctx))
3525 percpu_ref_put(&hctx->queue->q_usage_counter);
3531 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3533 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3534 struct blk_mq_hw_ctx, cpuhp_online);
3536 if (cpumask_test_cpu(cpu, hctx->cpumask))
3537 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3542 * 'cpu' is going away. splice any existing rq_list entries from this
3543 * software queue to the hw queue dispatch list, and ensure that it
3546 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3548 struct blk_mq_hw_ctx *hctx;
3549 struct blk_mq_ctx *ctx;
3551 enum hctx_type type;
3553 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3554 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3557 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3560 spin_lock(&ctx->lock);
3561 if (!list_empty(&ctx->rq_lists[type])) {
3562 list_splice_init(&ctx->rq_lists[type], &tmp);
3563 blk_mq_hctx_clear_pending(hctx, ctx);
3565 spin_unlock(&ctx->lock);
3567 if (list_empty(&tmp))
3570 spin_lock(&hctx->lock);
3571 list_splice_tail_init(&tmp, &hctx->dispatch);
3572 spin_unlock(&hctx->lock);
3574 blk_mq_run_hw_queue(hctx, true);
3578 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3580 if (!(hctx->flags & BLK_MQ_F_STACKING))
3581 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3582 &hctx->cpuhp_online);
3583 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3588 * Before freeing hw queue, clearing the flush request reference in
3589 * tags->rqs[] for avoiding potential UAF.
3591 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3592 unsigned int queue_depth, struct request *flush_rq)
3595 unsigned long flags;
3597 /* The hw queue may not be mapped yet */
3601 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3603 for (i = 0; i < queue_depth; i++)
3604 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3607 * Wait until all pending iteration is done.
3609 * Request reference is cleared and it is guaranteed to be observed
3610 * after the ->lock is released.
3612 spin_lock_irqsave(&tags->lock, flags);
3613 spin_unlock_irqrestore(&tags->lock, flags);
3616 /* hctx->ctxs will be freed in queue's release handler */
3617 static void blk_mq_exit_hctx(struct request_queue *q,
3618 struct blk_mq_tag_set *set,
3619 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3621 struct request *flush_rq = hctx->fq->flush_rq;
3623 if (blk_mq_hw_queue_mapped(hctx))
3624 blk_mq_tag_idle(hctx);
3626 if (blk_queue_init_done(q))
3627 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3628 set->queue_depth, flush_rq);
3629 if (set->ops->exit_request)
3630 set->ops->exit_request(set, flush_rq, hctx_idx);
3632 if (set->ops->exit_hctx)
3633 set->ops->exit_hctx(hctx, hctx_idx);
3635 blk_mq_remove_cpuhp(hctx);
3637 xa_erase(&q->hctx_table, hctx_idx);
3639 spin_lock(&q->unused_hctx_lock);
3640 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3641 spin_unlock(&q->unused_hctx_lock);
3644 static void blk_mq_exit_hw_queues(struct request_queue *q,
3645 struct blk_mq_tag_set *set, int nr_queue)
3647 struct blk_mq_hw_ctx *hctx;
3650 queue_for_each_hw_ctx(q, hctx, i) {
3653 blk_mq_exit_hctx(q, set, hctx, i);
3657 static int blk_mq_init_hctx(struct request_queue *q,
3658 struct blk_mq_tag_set *set,
3659 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3661 hctx->queue_num = hctx_idx;
3663 if (!(hctx->flags & BLK_MQ_F_STACKING))
3664 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3665 &hctx->cpuhp_online);
3666 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3668 hctx->tags = set->tags[hctx_idx];
3670 if (set->ops->init_hctx &&
3671 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3672 goto unregister_cpu_notifier;
3674 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3678 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3684 if (set->ops->exit_request)
3685 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3687 if (set->ops->exit_hctx)
3688 set->ops->exit_hctx(hctx, hctx_idx);
3689 unregister_cpu_notifier:
3690 blk_mq_remove_cpuhp(hctx);
3694 static struct blk_mq_hw_ctx *
3695 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3698 struct blk_mq_hw_ctx *hctx;
3699 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3701 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3703 goto fail_alloc_hctx;
3705 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3708 atomic_set(&hctx->nr_active, 0);
3709 if (node == NUMA_NO_NODE)
3710 node = set->numa_node;
3711 hctx->numa_node = node;
3713 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3714 spin_lock_init(&hctx->lock);
3715 INIT_LIST_HEAD(&hctx->dispatch);
3717 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3719 INIT_LIST_HEAD(&hctx->hctx_list);
3722 * Allocate space for all possible cpus to avoid allocation at
3725 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3730 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3731 gfp, node, false, false))
3735 spin_lock_init(&hctx->dispatch_wait_lock);
3736 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3737 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3739 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3743 blk_mq_hctx_kobj_init(hctx);
3748 sbitmap_free(&hctx->ctx_map);
3752 free_cpumask_var(hctx->cpumask);
3759 static void blk_mq_init_cpu_queues(struct request_queue *q,
3760 unsigned int nr_hw_queues)
3762 struct blk_mq_tag_set *set = q->tag_set;
3765 for_each_possible_cpu(i) {
3766 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3767 struct blk_mq_hw_ctx *hctx;
3771 spin_lock_init(&__ctx->lock);
3772 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3773 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3778 * Set local node, IFF we have more than one hw queue. If
3779 * not, we remain on the home node of the device
3781 for (j = 0; j < set->nr_maps; j++) {
3782 hctx = blk_mq_map_queue_type(q, j, i);
3783 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3784 hctx->numa_node = cpu_to_node(i);
3789 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3790 unsigned int hctx_idx,
3793 struct blk_mq_tags *tags;
3796 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3800 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3802 blk_mq_free_rq_map(tags);
3809 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3812 if (blk_mq_is_shared_tags(set->flags)) {
3813 set->tags[hctx_idx] = set->shared_tags;
3818 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3821 return set->tags[hctx_idx];
3824 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3825 struct blk_mq_tags *tags,
3826 unsigned int hctx_idx)
3829 blk_mq_free_rqs(set, tags, hctx_idx);
3830 blk_mq_free_rq_map(tags);
3834 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3835 unsigned int hctx_idx)
3837 if (!blk_mq_is_shared_tags(set->flags))
3838 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3840 set->tags[hctx_idx] = NULL;
3843 static void blk_mq_map_swqueue(struct request_queue *q)
3845 unsigned int j, hctx_idx;
3847 struct blk_mq_hw_ctx *hctx;
3848 struct blk_mq_ctx *ctx;
3849 struct blk_mq_tag_set *set = q->tag_set;
3851 queue_for_each_hw_ctx(q, hctx, i) {
3852 cpumask_clear(hctx->cpumask);
3854 hctx->dispatch_from = NULL;
3858 * Map software to hardware queues.
3860 * If the cpu isn't present, the cpu is mapped to first hctx.
3862 for_each_possible_cpu(i) {
3864 ctx = per_cpu_ptr(q->queue_ctx, i);
3865 for (j = 0; j < set->nr_maps; j++) {
3866 if (!set->map[j].nr_queues) {
3867 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3868 HCTX_TYPE_DEFAULT, i);
3871 hctx_idx = set->map[j].mq_map[i];
3872 /* unmapped hw queue can be remapped after CPU topo changed */
3873 if (!set->tags[hctx_idx] &&
3874 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3876 * If tags initialization fail for some hctx,
3877 * that hctx won't be brought online. In this
3878 * case, remap the current ctx to hctx[0] which
3879 * is guaranteed to always have tags allocated
3881 set->map[j].mq_map[i] = 0;
3884 hctx = blk_mq_map_queue_type(q, j, i);
3885 ctx->hctxs[j] = hctx;
3887 * If the CPU is already set in the mask, then we've
3888 * mapped this one already. This can happen if
3889 * devices share queues across queue maps.
3891 if (cpumask_test_cpu(i, hctx->cpumask))
3894 cpumask_set_cpu(i, hctx->cpumask);
3896 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3897 hctx->ctxs[hctx->nr_ctx++] = ctx;
3900 * If the nr_ctx type overflows, we have exceeded the
3901 * amount of sw queues we can support.
3903 BUG_ON(!hctx->nr_ctx);
3906 for (; j < HCTX_MAX_TYPES; j++)
3907 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3908 HCTX_TYPE_DEFAULT, i);
3911 queue_for_each_hw_ctx(q, hctx, i) {
3913 * If no software queues are mapped to this hardware queue,
3914 * disable it and free the request entries.
3916 if (!hctx->nr_ctx) {
3917 /* Never unmap queue 0. We need it as a
3918 * fallback in case of a new remap fails
3922 __blk_mq_free_map_and_rqs(set, i);
3928 hctx->tags = set->tags[i];
3929 WARN_ON(!hctx->tags);
3932 * Set the map size to the number of mapped software queues.
3933 * This is more accurate and more efficient than looping
3934 * over all possibly mapped software queues.
3936 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3939 * Initialize batch roundrobin counts
3941 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3942 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3947 * Caller needs to ensure that we're either frozen/quiesced, or that
3948 * the queue isn't live yet.
3950 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3952 struct blk_mq_hw_ctx *hctx;
3955 queue_for_each_hw_ctx(q, hctx, i) {
3957 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3959 blk_mq_tag_idle(hctx);
3960 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3965 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3968 struct request_queue *q;
3970 lockdep_assert_held(&set->tag_list_lock);
3972 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3973 blk_mq_freeze_queue(q);
3974 queue_set_hctx_shared(q, shared);
3975 blk_mq_unfreeze_queue(q);
3979 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3981 struct blk_mq_tag_set *set = q->tag_set;
3983 mutex_lock(&set->tag_list_lock);
3984 list_del(&q->tag_set_list);
3985 if (list_is_singular(&set->tag_list)) {
3986 /* just transitioned to unshared */
3987 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3988 /* update existing queue */
3989 blk_mq_update_tag_set_shared(set, false);
3991 mutex_unlock(&set->tag_list_lock);
3992 INIT_LIST_HEAD(&q->tag_set_list);
3995 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3996 struct request_queue *q)
3998 mutex_lock(&set->tag_list_lock);
4001 * Check to see if we're transitioning to shared (from 1 to 2 queues).
4003 if (!list_empty(&set->tag_list) &&
4004 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
4005 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
4006 /* update existing queue */
4007 blk_mq_update_tag_set_shared(set, true);
4009 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
4010 queue_set_hctx_shared(q, true);
4011 list_add_tail(&q->tag_set_list, &set->tag_list);
4013 mutex_unlock(&set->tag_list_lock);
4016 /* All allocations will be freed in release handler of q->mq_kobj */
4017 static int blk_mq_alloc_ctxs(struct request_queue *q)
4019 struct blk_mq_ctxs *ctxs;
4022 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
4026 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
4027 if (!ctxs->queue_ctx)
4030 for_each_possible_cpu(cpu) {
4031 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4035 q->mq_kobj = &ctxs->kobj;
4036 q->queue_ctx = ctxs->queue_ctx;
4045 * It is the actual release handler for mq, but we do it from
4046 * request queue's release handler for avoiding use-after-free
4047 * and headache because q->mq_kobj shouldn't have been introduced,
4048 * but we can't group ctx/kctx kobj without it.
4050 void blk_mq_release(struct request_queue *q)
4052 struct blk_mq_hw_ctx *hctx, *next;
4055 queue_for_each_hw_ctx(q, hctx, i)
4056 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4058 /* all hctx are in .unused_hctx_list now */
4059 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4060 list_del_init(&hctx->hctx_list);
4061 kobject_put(&hctx->kobj);
4064 xa_destroy(&q->hctx_table);
4067 * release .mq_kobj and sw queue's kobject now because
4068 * both share lifetime with request queue.
4070 blk_mq_sysfs_deinit(q);
4073 struct request_queue *blk_mq_alloc_queue(struct blk_mq_tag_set *set,
4074 struct queue_limits *lim, void *queuedata)
4076 struct queue_limits default_lim = { };
4077 struct request_queue *q;
4080 q = blk_alloc_queue(lim ? lim : &default_lim, set->numa_node);
4083 q->queuedata = queuedata;
4084 ret = blk_mq_init_allocated_queue(set, q);
4087 return ERR_PTR(ret);
4091 EXPORT_SYMBOL(blk_mq_alloc_queue);
4094 * blk_mq_destroy_queue - shutdown a request queue
4095 * @q: request queue to shutdown
4097 * This shuts down a request queue allocated by blk_mq_alloc_queue(). All future
4098 * requests will be failed with -ENODEV. The caller is responsible for dropping
4099 * the reference from blk_mq_alloc_queue() by calling blk_put_queue().
4101 * Context: can sleep
4103 void blk_mq_destroy_queue(struct request_queue *q)
4105 WARN_ON_ONCE(!queue_is_mq(q));
4106 WARN_ON_ONCE(blk_queue_registered(q));
4110 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4111 blk_queue_start_drain(q);
4112 blk_mq_freeze_queue_wait(q);
4115 blk_mq_cancel_work_sync(q);
4116 blk_mq_exit_queue(q);
4118 EXPORT_SYMBOL(blk_mq_destroy_queue);
4120 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set,
4121 struct queue_limits *lim, void *queuedata,
4122 struct lock_class_key *lkclass)
4124 struct request_queue *q;
4125 struct gendisk *disk;
4127 q = blk_mq_alloc_queue(set, lim, queuedata);
4131 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4133 blk_mq_destroy_queue(q);
4135 return ERR_PTR(-ENOMEM);
4137 set_bit(GD_OWNS_QUEUE, &disk->state);
4140 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4142 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4143 struct lock_class_key *lkclass)
4145 struct gendisk *disk;
4147 if (!blk_get_queue(q))
4149 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4154 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4156 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4157 struct blk_mq_tag_set *set, struct request_queue *q,
4158 int hctx_idx, int node)
4160 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4162 /* reuse dead hctx first */
4163 spin_lock(&q->unused_hctx_lock);
4164 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4165 if (tmp->numa_node == node) {
4171 list_del_init(&hctx->hctx_list);
4172 spin_unlock(&q->unused_hctx_lock);
4175 hctx = blk_mq_alloc_hctx(q, set, node);
4179 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4185 kobject_put(&hctx->kobj);
4190 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4191 struct request_queue *q)
4193 struct blk_mq_hw_ctx *hctx;
4196 /* protect against switching io scheduler */
4197 mutex_lock(&q->sysfs_lock);
4198 for (i = 0; i < set->nr_hw_queues; i++) {
4200 int node = blk_mq_get_hctx_node(set, i);
4201 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4204 old_node = old_hctx->numa_node;
4205 blk_mq_exit_hctx(q, set, old_hctx, i);
4208 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4211 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4213 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4214 WARN_ON_ONCE(!hctx);
4218 * Increasing nr_hw_queues fails. Free the newly allocated
4219 * hctxs and keep the previous q->nr_hw_queues.
4221 if (i != set->nr_hw_queues) {
4222 j = q->nr_hw_queues;
4225 q->nr_hw_queues = set->nr_hw_queues;
4228 xa_for_each_start(&q->hctx_table, j, hctx, j)
4229 blk_mq_exit_hctx(q, set, hctx, j);
4230 mutex_unlock(&q->sysfs_lock);
4233 static void blk_mq_update_poll_flag(struct request_queue *q)
4235 struct blk_mq_tag_set *set = q->tag_set;
4237 if (set->nr_maps > HCTX_TYPE_POLL &&
4238 set->map[HCTX_TYPE_POLL].nr_queues)
4239 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4241 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4244 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4245 struct request_queue *q)
4247 /* mark the queue as mq asap */
4248 q->mq_ops = set->ops;
4250 if (blk_mq_alloc_ctxs(q))
4253 /* init q->mq_kobj and sw queues' kobjects */
4254 blk_mq_sysfs_init(q);
4256 INIT_LIST_HEAD(&q->unused_hctx_list);
4257 spin_lock_init(&q->unused_hctx_lock);
4259 xa_init(&q->hctx_table);
4261 blk_mq_realloc_hw_ctxs(set, q);
4262 if (!q->nr_hw_queues)
4265 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4266 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4270 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4271 blk_mq_update_poll_flag(q);
4273 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4274 INIT_LIST_HEAD(&q->flush_list);
4275 INIT_LIST_HEAD(&q->requeue_list);
4276 spin_lock_init(&q->requeue_lock);
4278 q->nr_requests = set->queue_depth;
4280 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4281 blk_mq_add_queue_tag_set(set, q);
4282 blk_mq_map_swqueue(q);
4291 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4293 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4294 void blk_mq_exit_queue(struct request_queue *q)
4296 struct blk_mq_tag_set *set = q->tag_set;
4298 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4299 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4300 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4301 blk_mq_del_queue_tag_set(q);
4304 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4308 if (blk_mq_is_shared_tags(set->flags)) {
4309 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4312 if (!set->shared_tags)
4316 for (i = 0; i < set->nr_hw_queues; i++) {
4317 if (!__blk_mq_alloc_map_and_rqs(set, i))
4326 __blk_mq_free_map_and_rqs(set, i);
4328 if (blk_mq_is_shared_tags(set->flags)) {
4329 blk_mq_free_map_and_rqs(set, set->shared_tags,
4330 BLK_MQ_NO_HCTX_IDX);
4337 * Allocate the request maps associated with this tag_set. Note that this
4338 * may reduce the depth asked for, if memory is tight. set->queue_depth
4339 * will be updated to reflect the allocated depth.
4341 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4346 depth = set->queue_depth;
4348 err = __blk_mq_alloc_rq_maps(set);
4352 set->queue_depth >>= 1;
4353 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4357 } while (set->queue_depth);
4359 if (!set->queue_depth || err) {
4360 pr_err("blk-mq: failed to allocate request map\n");
4364 if (depth != set->queue_depth)
4365 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4366 depth, set->queue_depth);
4371 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4374 * blk_mq_map_queues() and multiple .map_queues() implementations
4375 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4376 * number of hardware queues.
4378 if (set->nr_maps == 1)
4379 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4381 if (set->ops->map_queues && !is_kdump_kernel()) {
4385 * transport .map_queues is usually done in the following
4388 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4389 * mask = get_cpu_mask(queue)
4390 * for_each_cpu(cpu, mask)
4391 * set->map[x].mq_map[cpu] = queue;
4394 * When we need to remap, the table has to be cleared for
4395 * killing stale mapping since one CPU may not be mapped
4398 for (i = 0; i < set->nr_maps; i++)
4399 blk_mq_clear_mq_map(&set->map[i]);
4401 set->ops->map_queues(set);
4403 BUG_ON(set->nr_maps > 1);
4404 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4408 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4409 int new_nr_hw_queues)
4411 struct blk_mq_tags **new_tags;
4414 if (set->nr_hw_queues >= new_nr_hw_queues)
4417 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4418 GFP_KERNEL, set->numa_node);
4423 memcpy(new_tags, set->tags, set->nr_hw_queues *
4424 sizeof(*set->tags));
4426 set->tags = new_tags;
4428 for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) {
4429 if (!__blk_mq_alloc_map_and_rqs(set, i)) {
4430 while (--i >= set->nr_hw_queues)
4431 __blk_mq_free_map_and_rqs(set, i);
4438 set->nr_hw_queues = new_nr_hw_queues;
4443 * Alloc a tag set to be associated with one or more request queues.
4444 * May fail with EINVAL for various error conditions. May adjust the
4445 * requested depth down, if it's too large. In that case, the set
4446 * value will be stored in set->queue_depth.
4448 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4452 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4454 if (!set->nr_hw_queues)
4456 if (!set->queue_depth)
4458 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4461 if (!set->ops->queue_rq)
4464 if (!set->ops->get_budget ^ !set->ops->put_budget)
4467 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4468 pr_info("blk-mq: reduced tag depth to %u\n",
4470 set->queue_depth = BLK_MQ_MAX_DEPTH;
4475 else if (set->nr_maps > HCTX_MAX_TYPES)
4479 * If a crashdump is active, then we are potentially in a very
4480 * memory constrained environment. Limit us to 1 queue and
4481 * 64 tags to prevent using too much memory.
4483 if (is_kdump_kernel()) {
4484 set->nr_hw_queues = 1;
4486 set->queue_depth = min(64U, set->queue_depth);
4489 * There is no use for more h/w queues than cpus if we just have
4492 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4493 set->nr_hw_queues = nr_cpu_ids;
4495 if (set->flags & BLK_MQ_F_BLOCKING) {
4496 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4499 ret = init_srcu_struct(set->srcu);
4505 set->tags = kcalloc_node(set->nr_hw_queues,
4506 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4509 goto out_cleanup_srcu;
4511 for (i = 0; i < set->nr_maps; i++) {
4512 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4513 sizeof(set->map[i].mq_map[0]),
4514 GFP_KERNEL, set->numa_node);
4515 if (!set->map[i].mq_map)
4516 goto out_free_mq_map;
4517 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4520 blk_mq_update_queue_map(set);
4522 ret = blk_mq_alloc_set_map_and_rqs(set);
4524 goto out_free_mq_map;
4526 mutex_init(&set->tag_list_lock);
4527 INIT_LIST_HEAD(&set->tag_list);
4532 for (i = 0; i < set->nr_maps; i++) {
4533 kfree(set->map[i].mq_map);
4534 set->map[i].mq_map = NULL;
4539 if (set->flags & BLK_MQ_F_BLOCKING)
4540 cleanup_srcu_struct(set->srcu);
4542 if (set->flags & BLK_MQ_F_BLOCKING)
4546 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4548 /* allocate and initialize a tagset for a simple single-queue device */
4549 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4550 const struct blk_mq_ops *ops, unsigned int queue_depth,
4551 unsigned int set_flags)
4553 memset(set, 0, sizeof(*set));
4555 set->nr_hw_queues = 1;
4557 set->queue_depth = queue_depth;
4558 set->numa_node = NUMA_NO_NODE;
4559 set->flags = set_flags;
4560 return blk_mq_alloc_tag_set(set);
4562 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4564 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4568 for (i = 0; i < set->nr_hw_queues; i++)
4569 __blk_mq_free_map_and_rqs(set, i);
4571 if (blk_mq_is_shared_tags(set->flags)) {
4572 blk_mq_free_map_and_rqs(set, set->shared_tags,
4573 BLK_MQ_NO_HCTX_IDX);
4576 for (j = 0; j < set->nr_maps; j++) {
4577 kfree(set->map[j].mq_map);
4578 set->map[j].mq_map = NULL;
4583 if (set->flags & BLK_MQ_F_BLOCKING) {
4584 cleanup_srcu_struct(set->srcu);
4588 EXPORT_SYMBOL(blk_mq_free_tag_set);
4590 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4592 struct blk_mq_tag_set *set = q->tag_set;
4593 struct blk_mq_hw_ctx *hctx;
4600 if (q->nr_requests == nr)
4603 blk_mq_freeze_queue(q);
4604 blk_mq_quiesce_queue(q);
4607 queue_for_each_hw_ctx(q, hctx, i) {
4611 * If we're using an MQ scheduler, just update the scheduler
4612 * queue depth. This is similar to what the old code would do.
4614 if (hctx->sched_tags) {
4615 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4618 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4623 if (q->elevator && q->elevator->type->ops.depth_updated)
4624 q->elevator->type->ops.depth_updated(hctx);
4627 q->nr_requests = nr;
4628 if (blk_mq_is_shared_tags(set->flags)) {
4630 blk_mq_tag_update_sched_shared_tags(q);
4632 blk_mq_tag_resize_shared_tags(set, nr);
4636 blk_mq_unquiesce_queue(q);
4637 blk_mq_unfreeze_queue(q);
4643 * request_queue and elevator_type pair.
4644 * It is just used by __blk_mq_update_nr_hw_queues to cache
4645 * the elevator_type associated with a request_queue.
4647 struct blk_mq_qe_pair {
4648 struct list_head node;
4649 struct request_queue *q;
4650 struct elevator_type *type;
4654 * Cache the elevator_type in qe pair list and switch the
4655 * io scheduler to 'none'
4657 static bool blk_mq_elv_switch_none(struct list_head *head,
4658 struct request_queue *q)
4660 struct blk_mq_qe_pair *qe;
4662 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4666 /* q->elevator needs protection from ->sysfs_lock */
4667 mutex_lock(&q->sysfs_lock);
4669 /* the check has to be done with holding sysfs_lock */
4675 INIT_LIST_HEAD(&qe->node);
4677 qe->type = q->elevator->type;
4678 /* keep a reference to the elevator module as we'll switch back */
4679 __elevator_get(qe->type);
4680 list_add(&qe->node, head);
4681 elevator_disable(q);
4683 mutex_unlock(&q->sysfs_lock);
4688 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4689 struct request_queue *q)
4691 struct blk_mq_qe_pair *qe;
4693 list_for_each_entry(qe, head, node)
4700 static void blk_mq_elv_switch_back(struct list_head *head,
4701 struct request_queue *q)
4703 struct blk_mq_qe_pair *qe;
4704 struct elevator_type *t;
4706 qe = blk_lookup_qe_pair(head, q);
4710 list_del(&qe->node);
4713 mutex_lock(&q->sysfs_lock);
4714 elevator_switch(q, t);
4715 /* drop the reference acquired in blk_mq_elv_switch_none */
4717 mutex_unlock(&q->sysfs_lock);
4720 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4723 struct request_queue *q;
4725 int prev_nr_hw_queues = set->nr_hw_queues;
4728 lockdep_assert_held(&set->tag_list_lock);
4730 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4731 nr_hw_queues = nr_cpu_ids;
4732 if (nr_hw_queues < 1)
4734 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4737 list_for_each_entry(q, &set->tag_list, tag_set_list)
4738 blk_mq_freeze_queue(q);
4740 * Switch IO scheduler to 'none', cleaning up the data associated
4741 * with the previous scheduler. We will switch back once we are done
4742 * updating the new sw to hw queue mappings.
4744 list_for_each_entry(q, &set->tag_list, tag_set_list)
4745 if (!blk_mq_elv_switch_none(&head, q))
4748 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4749 blk_mq_debugfs_unregister_hctxs(q);
4750 blk_mq_sysfs_unregister_hctxs(q);
4753 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4757 blk_mq_update_queue_map(set);
4758 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4759 blk_mq_realloc_hw_ctxs(set, q);
4760 blk_mq_update_poll_flag(q);
4761 if (q->nr_hw_queues != set->nr_hw_queues) {
4762 int i = prev_nr_hw_queues;
4764 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4765 nr_hw_queues, prev_nr_hw_queues);
4766 for (; i < set->nr_hw_queues; i++)
4767 __blk_mq_free_map_and_rqs(set, i);
4769 set->nr_hw_queues = prev_nr_hw_queues;
4772 blk_mq_map_swqueue(q);
4776 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4777 blk_mq_sysfs_register_hctxs(q);
4778 blk_mq_debugfs_register_hctxs(q);
4782 list_for_each_entry(q, &set->tag_list, tag_set_list)
4783 blk_mq_elv_switch_back(&head, q);
4785 list_for_each_entry(q, &set->tag_list, tag_set_list)
4786 blk_mq_unfreeze_queue(q);
4788 /* Free the excess tags when nr_hw_queues shrink. */
4789 for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)
4790 __blk_mq_free_map_and_rqs(set, i);
4793 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4795 mutex_lock(&set->tag_list_lock);
4796 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4797 mutex_unlock(&set->tag_list_lock);
4799 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4801 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
4802 struct io_comp_batch *iob, unsigned int flags)
4804 long state = get_current_state();
4808 ret = q->mq_ops->poll(hctx, iob);
4810 __set_current_state(TASK_RUNNING);
4814 if (signal_pending_state(state, current))
4815 __set_current_state(TASK_RUNNING);
4816 if (task_is_running(current))
4819 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4822 } while (!need_resched());
4824 __set_current_state(TASK_RUNNING);
4828 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
4829 struct io_comp_batch *iob, unsigned int flags)
4831 struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie);
4833 return blk_hctx_poll(q, hctx, iob, flags);
4836 int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
4837 unsigned int poll_flags)
4839 struct request_queue *q = rq->q;
4842 if (!blk_rq_is_poll(rq))
4844 if (!percpu_ref_tryget(&q->q_usage_counter))
4847 ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags);
4852 EXPORT_SYMBOL_GPL(blk_rq_poll);
4854 unsigned int blk_mq_rq_cpu(struct request *rq)
4856 return rq->mq_ctx->cpu;
4858 EXPORT_SYMBOL(blk_mq_rq_cpu);
4860 void blk_mq_cancel_work_sync(struct request_queue *q)
4862 struct blk_mq_hw_ctx *hctx;
4865 cancel_delayed_work_sync(&q->requeue_work);
4867 queue_for_each_hw_ctx(q, hctx, i)
4868 cancel_delayed_work_sync(&hctx->run_work);
4871 static int __init blk_mq_init(void)
4875 for_each_possible_cpu(i)
4876 init_llist_head(&per_cpu(blk_cpu_done, i));
4877 for_each_possible_cpu(i)
4878 INIT_CSD(&per_cpu(blk_cpu_csd, i),
4879 __blk_mq_complete_request_remote, NULL);
4880 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4882 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4883 "block/softirq:dead", NULL,
4884 blk_softirq_cpu_dead);
4885 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4886 blk_mq_hctx_notify_dead);
4887 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4888 blk_mq_hctx_notify_online,
4889 blk_mq_hctx_notify_offline);
4892 subsys_initcall(blk_mq_init);