1 // SPDX-License-Identifier: GPL-2.0
3 * Block multiqueue core code
5 * Copyright (C) 2013-2014 Jens Axboe
6 * Copyright (C) 2013-2014 Christoph Hellwig
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/backing-dev.h>
11 #include <linux/bio.h>
12 #include <linux/blkdev.h>
13 #include <linux/blk-integrity.h>
14 #include <linux/kmemleak.h>
16 #include <linux/init.h>
17 #include <linux/slab.h>
18 #include <linux/workqueue.h>
19 #include <linux/smp.h>
20 #include <linux/interrupt.h>
21 #include <linux/llist.h>
22 #include <linux/cpu.h>
23 #include <linux/cache.h>
24 #include <linux/sched/sysctl.h>
25 #include <linux/sched/topology.h>
26 #include <linux/sched/signal.h>
27 #include <linux/delay.h>
28 #include <linux/crash_dump.h>
29 #include <linux/prefetch.h>
30 #include <linux/blk-crypto.h>
31 #include <linux/part_stat.h>
33 #include <trace/events/block.h>
35 #include <linux/blk-mq.h>
36 #include <linux/t10-pi.h>
39 #include "blk-mq-debugfs.h"
40 #include "blk-mq-tag.h"
43 #include "blk-mq-sched.h"
44 #include "blk-rq-qos.h"
45 #include "blk-ioprio.h"
47 static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
49 static void blk_mq_poll_stats_start(struct request_queue *q);
50 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
52 static int blk_mq_poll_stats_bkt(const struct request *rq)
54 int ddir, sectors, bucket;
56 ddir = rq_data_dir(rq);
57 sectors = blk_rq_stats_sectors(rq);
59 bucket = ddir + 2 * ilog2(sectors);
63 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
64 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
69 #define BLK_QC_T_SHIFT 16
70 #define BLK_QC_T_INTERNAL (1U << 31)
72 static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q,
75 return xa_load(&q->hctx_table,
76 (qc & ~BLK_QC_T_INTERNAL) >> BLK_QC_T_SHIFT);
79 static inline struct request *blk_qc_to_rq(struct blk_mq_hw_ctx *hctx,
82 unsigned int tag = qc & ((1U << BLK_QC_T_SHIFT) - 1);
84 if (qc & BLK_QC_T_INTERNAL)
85 return blk_mq_tag_to_rq(hctx->sched_tags, tag);
86 return blk_mq_tag_to_rq(hctx->tags, tag);
89 static inline blk_qc_t blk_rq_to_qc(struct request *rq)
91 return (rq->mq_hctx->queue_num << BLK_QC_T_SHIFT) |
93 rq->tag : (rq->internal_tag | BLK_QC_T_INTERNAL));
97 * Check if any of the ctx, dispatch list or elevator
98 * have pending work in this hardware queue.
100 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
102 return !list_empty_careful(&hctx->dispatch) ||
103 sbitmap_any_bit_set(&hctx->ctx_map) ||
104 blk_mq_sched_has_work(hctx);
108 * Mark this ctx as having pending work in this hardware queue
110 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
111 struct blk_mq_ctx *ctx)
113 const int bit = ctx->index_hw[hctx->type];
115 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
116 sbitmap_set_bit(&hctx->ctx_map, bit);
119 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
120 struct blk_mq_ctx *ctx)
122 const int bit = ctx->index_hw[hctx->type];
124 sbitmap_clear_bit(&hctx->ctx_map, bit);
128 struct block_device *part;
129 unsigned int inflight[2];
132 static bool blk_mq_check_inflight(struct request *rq, void *priv)
134 struct mq_inflight *mi = priv;
136 if (rq->part && blk_do_io_stat(rq) &&
137 (!mi->part->bd_partno || rq->part == mi->part) &&
138 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
139 mi->inflight[rq_data_dir(rq)]++;
144 unsigned int blk_mq_in_flight(struct request_queue *q,
145 struct block_device *part)
147 struct mq_inflight mi = { .part = part };
149 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
151 return mi.inflight[0] + mi.inflight[1];
154 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
155 unsigned int inflight[2])
157 struct mq_inflight mi = { .part = part };
159 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
160 inflight[0] = mi.inflight[0];
161 inflight[1] = mi.inflight[1];
164 void blk_freeze_queue_start(struct request_queue *q)
166 mutex_lock(&q->mq_freeze_lock);
167 if (++q->mq_freeze_depth == 1) {
168 percpu_ref_kill(&q->q_usage_counter);
169 mutex_unlock(&q->mq_freeze_lock);
171 blk_mq_run_hw_queues(q, false);
173 mutex_unlock(&q->mq_freeze_lock);
176 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
178 void blk_mq_freeze_queue_wait(struct request_queue *q)
180 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
182 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
184 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
185 unsigned long timeout)
187 return wait_event_timeout(q->mq_freeze_wq,
188 percpu_ref_is_zero(&q->q_usage_counter),
191 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
194 * Guarantee no request is in use, so we can change any data structure of
195 * the queue afterward.
197 void blk_freeze_queue(struct request_queue *q)
200 * In the !blk_mq case we are only calling this to kill the
201 * q_usage_counter, otherwise this increases the freeze depth
202 * and waits for it to return to zero. For this reason there is
203 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
204 * exported to drivers as the only user for unfreeze is blk_mq.
206 blk_freeze_queue_start(q);
207 blk_mq_freeze_queue_wait(q);
210 void blk_mq_freeze_queue(struct request_queue *q)
213 * ...just an alias to keep freeze and unfreeze actions balanced
214 * in the blk_mq_* namespace
218 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
220 void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
222 mutex_lock(&q->mq_freeze_lock);
224 q->q_usage_counter.data->force_atomic = true;
225 q->mq_freeze_depth--;
226 WARN_ON_ONCE(q->mq_freeze_depth < 0);
227 if (!q->mq_freeze_depth) {
228 percpu_ref_resurrect(&q->q_usage_counter);
229 wake_up_all(&q->mq_freeze_wq);
231 mutex_unlock(&q->mq_freeze_lock);
234 void blk_mq_unfreeze_queue(struct request_queue *q)
236 __blk_mq_unfreeze_queue(q, false);
238 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
241 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
242 * mpt3sas driver such that this function can be removed.
244 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
248 spin_lock_irqsave(&q->queue_lock, flags);
249 if (!q->quiesce_depth++)
250 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
251 spin_unlock_irqrestore(&q->queue_lock, flags);
253 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
256 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
259 * Note: it is driver's responsibility for making sure that quiesce has
262 void blk_mq_wait_quiesce_done(struct request_queue *q)
264 if (blk_queue_has_srcu(q))
265 synchronize_srcu(q->srcu);
269 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
272 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
275 * Note: this function does not prevent that the struct request end_io()
276 * callback function is invoked. Once this function is returned, we make
277 * sure no dispatch can happen until the queue is unquiesced via
278 * blk_mq_unquiesce_queue().
280 void blk_mq_quiesce_queue(struct request_queue *q)
282 blk_mq_quiesce_queue_nowait(q);
283 blk_mq_wait_quiesce_done(q);
285 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
288 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
291 * This function recovers queue into the state before quiescing
292 * which is done by blk_mq_quiesce_queue.
294 void blk_mq_unquiesce_queue(struct request_queue *q)
297 bool run_queue = false;
299 spin_lock_irqsave(&q->queue_lock, flags);
300 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
302 } else if (!--q->quiesce_depth) {
303 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
306 spin_unlock_irqrestore(&q->queue_lock, flags);
308 /* dispatch requests which are inserted during quiescing */
310 blk_mq_run_hw_queues(q, true);
312 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
314 void blk_mq_wake_waiters(struct request_queue *q)
316 struct blk_mq_hw_ctx *hctx;
319 queue_for_each_hw_ctx(q, hctx, i)
320 if (blk_mq_hw_queue_mapped(hctx))
321 blk_mq_tag_wakeup_all(hctx->tags, true);
324 void blk_rq_init(struct request_queue *q, struct request *rq)
326 memset(rq, 0, sizeof(*rq));
328 INIT_LIST_HEAD(&rq->queuelist);
330 rq->__sector = (sector_t) -1;
331 INIT_HLIST_NODE(&rq->hash);
332 RB_CLEAR_NODE(&rq->rb_node);
333 rq->tag = BLK_MQ_NO_TAG;
334 rq->internal_tag = BLK_MQ_NO_TAG;
335 rq->start_time_ns = ktime_get_ns();
337 blk_crypto_rq_set_defaults(rq);
339 EXPORT_SYMBOL(blk_rq_init);
341 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
342 struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns)
344 struct blk_mq_ctx *ctx = data->ctx;
345 struct blk_mq_hw_ctx *hctx = data->hctx;
346 struct request_queue *q = data->q;
347 struct request *rq = tags->static_rqs[tag];
352 rq->cmd_flags = data->cmd_flags;
354 if (data->flags & BLK_MQ_REQ_PM)
355 data->rq_flags |= RQF_PM;
356 if (blk_queue_io_stat(q))
357 data->rq_flags |= RQF_IO_STAT;
358 rq->rq_flags = data->rq_flags;
360 if (!(data->rq_flags & RQF_ELV)) {
362 rq->internal_tag = BLK_MQ_NO_TAG;
364 rq->tag = BLK_MQ_NO_TAG;
365 rq->internal_tag = tag;
369 if (blk_mq_need_time_stamp(rq))
370 rq->start_time_ns = ktime_get_ns();
372 rq->start_time_ns = 0;
374 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
375 rq->alloc_time_ns = alloc_time_ns;
377 rq->io_start_time_ns = 0;
378 rq->stats_sectors = 0;
379 rq->nr_phys_segments = 0;
380 #if defined(CONFIG_BLK_DEV_INTEGRITY)
381 rq->nr_integrity_segments = 0;
384 rq->end_io_data = NULL;
386 blk_crypto_rq_set_defaults(rq);
387 INIT_LIST_HEAD(&rq->queuelist);
388 /* tag was already set */
389 WRITE_ONCE(rq->deadline, 0);
392 if (rq->rq_flags & RQF_ELV) {
393 struct elevator_queue *e = data->q->elevator;
395 INIT_HLIST_NODE(&rq->hash);
396 RB_CLEAR_NODE(&rq->rb_node);
398 if (!op_is_flush(data->cmd_flags) &&
399 e->type->ops.prepare_request) {
400 e->type->ops.prepare_request(rq);
401 rq->rq_flags |= RQF_ELVPRIV;
408 static inline struct request *
409 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data,
412 unsigned int tag, tag_offset;
413 struct blk_mq_tags *tags;
415 unsigned long tag_mask;
418 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
419 if (unlikely(!tag_mask))
422 tags = blk_mq_tags_from_data(data);
423 for (i = 0; tag_mask; i++) {
424 if (!(tag_mask & (1UL << i)))
426 tag = tag_offset + i;
427 prefetch(tags->static_rqs[tag]);
428 tag_mask &= ~(1UL << i);
429 rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns);
430 rq_list_add(data->cached_rq, rq);
433 /* caller already holds a reference, add for remainder */
434 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
437 return rq_list_pop(data->cached_rq);
440 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
442 struct request_queue *q = data->q;
443 u64 alloc_time_ns = 0;
447 /* alloc_time includes depth and tag waits */
448 if (blk_queue_rq_alloc_time(q))
449 alloc_time_ns = ktime_get_ns();
451 if (data->cmd_flags & REQ_NOWAIT)
452 data->flags |= BLK_MQ_REQ_NOWAIT;
455 struct elevator_queue *e = q->elevator;
457 data->rq_flags |= RQF_ELV;
460 * Flush/passthrough requests are special and go directly to the
461 * dispatch list. Don't include reserved tags in the
462 * limiting, as it isn't useful.
464 if (!op_is_flush(data->cmd_flags) &&
465 !blk_op_is_passthrough(data->cmd_flags) &&
466 e->type->ops.limit_depth &&
467 !(data->flags & BLK_MQ_REQ_RESERVED))
468 e->type->ops.limit_depth(data->cmd_flags, data);
472 data->ctx = blk_mq_get_ctx(q);
473 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
474 if (!(data->rq_flags & RQF_ELV))
475 blk_mq_tag_busy(data->hctx);
477 if (data->flags & BLK_MQ_REQ_RESERVED)
478 data->rq_flags |= RQF_RESV;
481 * Try batched alloc if we want more than 1 tag.
483 if (data->nr_tags > 1) {
484 rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns);
491 * Waiting allocations only fail because of an inactive hctx. In that
492 * case just retry the hctx assignment and tag allocation as CPU hotplug
493 * should have migrated us to an online CPU by now.
495 tag = blk_mq_get_tag(data);
496 if (tag == BLK_MQ_NO_TAG) {
497 if (data->flags & BLK_MQ_REQ_NOWAIT)
500 * Give up the CPU and sleep for a random short time to
501 * ensure that thread using a realtime scheduling class
502 * are migrated off the CPU, and thus off the hctx that
509 return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag,
513 static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
514 struct blk_plug *plug,
516 blk_mq_req_flags_t flags)
518 struct blk_mq_alloc_data data = {
522 .nr_tags = plug->nr_ios,
523 .cached_rq = &plug->cached_rq,
527 if (blk_queue_enter(q, flags))
532 rq = __blk_mq_alloc_requests(&data);
538 static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
540 blk_mq_req_flags_t flags)
542 struct blk_plug *plug = current->plug;
547 if (rq_list_empty(plug->cached_rq)) {
548 if (plug->nr_ios == 1)
550 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
555 rq = rq_list_peek(&plug->cached_rq);
556 if (!rq || rq->q != q)
559 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
561 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
564 plug->cached_rq = rq_list_next(rq);
567 INIT_LIST_HEAD(&rq->queuelist);
571 struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
572 blk_mq_req_flags_t flags)
576 rq = blk_mq_alloc_cached_request(q, opf, flags);
578 struct blk_mq_alloc_data data = {
586 ret = blk_queue_enter(q, flags);
590 rq = __blk_mq_alloc_requests(&data);
595 rq->__sector = (sector_t) -1;
596 rq->bio = rq->biotail = NULL;
600 return ERR_PTR(-EWOULDBLOCK);
602 EXPORT_SYMBOL(blk_mq_alloc_request);
604 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
605 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
607 struct blk_mq_alloc_data data = {
613 u64 alloc_time_ns = 0;
619 /* alloc_time includes depth and tag waits */
620 if (blk_queue_rq_alloc_time(q))
621 alloc_time_ns = ktime_get_ns();
624 * If the tag allocator sleeps we could get an allocation for a
625 * different hardware context. No need to complicate the low level
626 * allocator for this for the rare use case of a command tied to
629 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
630 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
631 return ERR_PTR(-EINVAL);
633 if (hctx_idx >= q->nr_hw_queues)
634 return ERR_PTR(-EIO);
636 ret = blk_queue_enter(q, flags);
641 * Check if the hardware context is actually mapped to anything.
642 * If not tell the caller that it should skip this queue.
645 data.hctx = xa_load(&q->hctx_table, hctx_idx);
646 if (!blk_mq_hw_queue_mapped(data.hctx))
648 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
649 if (cpu >= nr_cpu_ids)
651 data.ctx = __blk_mq_get_ctx(q, cpu);
654 blk_mq_tag_busy(data.hctx);
656 data.rq_flags |= RQF_ELV;
658 if (flags & BLK_MQ_REQ_RESERVED)
659 data.rq_flags |= RQF_RESV;
662 tag = blk_mq_get_tag(&data);
663 if (tag == BLK_MQ_NO_TAG)
665 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag,
668 rq->__sector = (sector_t) -1;
669 rq->bio = rq->biotail = NULL;
676 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
678 static void __blk_mq_free_request(struct request *rq)
680 struct request_queue *q = rq->q;
681 struct blk_mq_ctx *ctx = rq->mq_ctx;
682 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
683 const int sched_tag = rq->internal_tag;
685 blk_crypto_free_request(rq);
686 blk_pm_mark_last_busy(rq);
688 if (rq->tag != BLK_MQ_NO_TAG)
689 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
690 if (sched_tag != BLK_MQ_NO_TAG)
691 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
692 blk_mq_sched_restart(hctx);
696 void blk_mq_free_request(struct request *rq)
698 struct request_queue *q = rq->q;
699 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
701 if ((rq->rq_flags & RQF_ELVPRIV) &&
702 q->elevator->type->ops.finish_request)
703 q->elevator->type->ops.finish_request(rq);
705 if (rq->rq_flags & RQF_MQ_INFLIGHT)
706 __blk_mq_dec_active_requests(hctx);
708 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
709 laptop_io_completion(q->disk->bdi);
713 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
714 if (req_ref_put_and_test(rq))
715 __blk_mq_free_request(rq);
717 EXPORT_SYMBOL_GPL(blk_mq_free_request);
719 void blk_mq_free_plug_rqs(struct blk_plug *plug)
723 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
724 blk_mq_free_request(rq);
727 void blk_dump_rq_flags(struct request *rq, char *msg)
729 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
730 rq->q->disk ? rq->q->disk->disk_name : "?",
731 (__force unsigned long long) rq->cmd_flags);
733 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
734 (unsigned long long)blk_rq_pos(rq),
735 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
736 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
737 rq->bio, rq->biotail, blk_rq_bytes(rq));
739 EXPORT_SYMBOL(blk_dump_rq_flags);
741 static void req_bio_endio(struct request *rq, struct bio *bio,
742 unsigned int nbytes, blk_status_t error)
744 if (unlikely(error)) {
745 bio->bi_status = error;
746 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
748 * Partial zone append completions cannot be supported as the
749 * BIO fragments may end up not being written sequentially.
751 if (bio->bi_iter.bi_size != nbytes)
752 bio->bi_status = BLK_STS_IOERR;
754 bio->bi_iter.bi_sector = rq->__sector;
757 bio_advance(bio, nbytes);
759 if (unlikely(rq->rq_flags & RQF_QUIET))
760 bio_set_flag(bio, BIO_QUIET);
761 /* don't actually finish bio if it's part of flush sequence */
762 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
766 static void blk_account_io_completion(struct request *req, unsigned int bytes)
768 if (req->part && blk_do_io_stat(req)) {
769 const int sgrp = op_stat_group(req_op(req));
772 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
777 static void blk_print_req_error(struct request *req, blk_status_t status)
779 printk_ratelimited(KERN_ERR
780 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
781 "phys_seg %u prio class %u\n",
782 blk_status_to_str(status),
783 req->q->disk ? req->q->disk->disk_name : "?",
784 blk_rq_pos(req), (__force u32)req_op(req),
785 blk_op_str(req_op(req)),
786 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
787 req->nr_phys_segments,
788 IOPRIO_PRIO_CLASS(req->ioprio));
792 * Fully end IO on a request. Does not support partial completions, or
795 static void blk_complete_request(struct request *req)
797 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
798 int total_bytes = blk_rq_bytes(req);
799 struct bio *bio = req->bio;
801 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
806 #ifdef CONFIG_BLK_DEV_INTEGRITY
807 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
808 req->q->integrity.profile->complete_fn(req, total_bytes);
811 blk_account_io_completion(req, total_bytes);
814 struct bio *next = bio->bi_next;
816 /* Completion has already been traced */
817 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
819 if (req_op(req) == REQ_OP_ZONE_APPEND)
820 bio->bi_iter.bi_sector = req->__sector;
828 * Reset counters so that the request stacking driver
829 * can find how many bytes remain in the request
839 * blk_update_request - Complete multiple bytes without completing the request
840 * @req: the request being processed
841 * @error: block status code
842 * @nr_bytes: number of bytes to complete for @req
845 * Ends I/O on a number of bytes attached to @req, but doesn't complete
846 * the request structure even if @req doesn't have leftover.
847 * If @req has leftover, sets it up for the next range of segments.
849 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
850 * %false return from this function.
853 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
854 * except in the consistency check at the end of this function.
857 * %false - this request doesn't have any more data
858 * %true - this request has more data
860 bool blk_update_request(struct request *req, blk_status_t error,
861 unsigned int nr_bytes)
865 trace_block_rq_complete(req, error, nr_bytes);
870 #ifdef CONFIG_BLK_DEV_INTEGRITY
871 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
873 req->q->integrity.profile->complete_fn(req, nr_bytes);
876 if (unlikely(error && !blk_rq_is_passthrough(req) &&
877 !(req->rq_flags & RQF_QUIET)) &&
878 !test_bit(GD_DEAD, &req->q->disk->state)) {
879 blk_print_req_error(req, error);
880 trace_block_rq_error(req, error, nr_bytes);
883 blk_account_io_completion(req, nr_bytes);
887 struct bio *bio = req->bio;
888 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
890 if (bio_bytes == bio->bi_iter.bi_size)
891 req->bio = bio->bi_next;
893 /* Completion has already been traced */
894 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
895 req_bio_endio(req, bio, bio_bytes, error);
897 total_bytes += bio_bytes;
898 nr_bytes -= bio_bytes;
909 * Reset counters so that the request stacking driver
910 * can find how many bytes remain in the request
917 req->__data_len -= total_bytes;
919 /* update sector only for requests with clear definition of sector */
920 if (!blk_rq_is_passthrough(req))
921 req->__sector += total_bytes >> 9;
923 /* mixed attributes always follow the first bio */
924 if (req->rq_flags & RQF_MIXED_MERGE) {
925 req->cmd_flags &= ~REQ_FAILFAST_MASK;
926 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
929 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
931 * If total number of sectors is less than the first segment
932 * size, something has gone terribly wrong.
934 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
935 blk_dump_rq_flags(req, "request botched");
936 req->__data_len = blk_rq_cur_bytes(req);
939 /* recalculate the number of segments */
940 req->nr_phys_segments = blk_recalc_rq_segments(req);
945 EXPORT_SYMBOL_GPL(blk_update_request);
947 static void __blk_account_io_done(struct request *req, u64 now)
949 const int sgrp = op_stat_group(req_op(req));
952 update_io_ticks(req->part, jiffies, true);
953 part_stat_inc(req->part, ios[sgrp]);
954 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
958 static inline void blk_account_io_done(struct request *req, u64 now)
961 * Account IO completion. flush_rq isn't accounted as a
962 * normal IO on queueing nor completion. Accounting the
963 * containing request is enough.
965 if (blk_do_io_stat(req) && req->part &&
966 !(req->rq_flags & RQF_FLUSH_SEQ))
967 __blk_account_io_done(req, now);
970 static void __blk_account_io_start(struct request *rq)
973 * All non-passthrough requests are created from a bio with one
974 * exception: when a flush command that is part of a flush sequence
975 * generated by the state machine in blk-flush.c is cloned onto the
976 * lower device by dm-multipath we can get here without a bio.
979 rq->part = rq->bio->bi_bdev;
981 rq->part = rq->q->disk->part0;
984 update_io_ticks(rq->part, jiffies, false);
988 static inline void blk_account_io_start(struct request *req)
990 if (blk_do_io_stat(req))
991 __blk_account_io_start(req);
994 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
996 if (rq->rq_flags & RQF_STATS) {
997 blk_mq_poll_stats_start(rq->q);
998 blk_stat_add(rq, now);
1001 blk_mq_sched_completed_request(rq, now);
1002 blk_account_io_done(rq, now);
1005 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1007 if (blk_mq_need_time_stamp(rq))
1008 __blk_mq_end_request_acct(rq, ktime_get_ns());
1011 rq_qos_done(rq->q, rq);
1012 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1013 blk_mq_free_request(rq);
1015 blk_mq_free_request(rq);
1018 EXPORT_SYMBOL(__blk_mq_end_request);
1020 void blk_mq_end_request(struct request *rq, blk_status_t error)
1022 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1024 __blk_mq_end_request(rq, error);
1026 EXPORT_SYMBOL(blk_mq_end_request);
1028 #define TAG_COMP_BATCH 32
1030 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1031 int *tag_array, int nr_tags)
1033 struct request_queue *q = hctx->queue;
1036 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1037 * update hctx->nr_active in batch
1039 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1040 __blk_mq_sub_active_requests(hctx, nr_tags);
1042 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1043 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1046 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1048 int tags[TAG_COMP_BATCH], nr_tags = 0;
1049 struct blk_mq_hw_ctx *cur_hctx = NULL;
1054 now = ktime_get_ns();
1056 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1058 prefetch(rq->rq_next);
1060 blk_complete_request(rq);
1062 __blk_mq_end_request_acct(rq, now);
1064 rq_qos_done(rq->q, rq);
1067 * If end_io handler returns NONE, then it still has
1068 * ownership of the request.
1070 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1073 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1074 if (!req_ref_put_and_test(rq))
1077 blk_crypto_free_request(rq);
1078 blk_pm_mark_last_busy(rq);
1080 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1082 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1084 cur_hctx = rq->mq_hctx;
1086 tags[nr_tags++] = rq->tag;
1090 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1092 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1094 static void blk_complete_reqs(struct llist_head *list)
1096 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1097 struct request *rq, *next;
1099 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1100 rq->q->mq_ops->complete(rq);
1103 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1105 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1108 static int blk_softirq_cpu_dead(unsigned int cpu)
1110 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1114 static void __blk_mq_complete_request_remote(void *data)
1116 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1119 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1121 int cpu = raw_smp_processor_id();
1123 if (!IS_ENABLED(CONFIG_SMP) ||
1124 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1127 * With force threaded interrupts enabled, raising softirq from an SMP
1128 * function call will always result in waking the ksoftirqd thread.
1129 * This is probably worse than completing the request on a different
1132 if (force_irqthreads())
1135 /* same CPU or cache domain? Complete locally */
1136 if (cpu == rq->mq_ctx->cpu ||
1137 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1138 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1141 /* don't try to IPI to an offline CPU */
1142 return cpu_online(rq->mq_ctx->cpu);
1145 static void blk_mq_complete_send_ipi(struct request *rq)
1147 struct llist_head *list;
1150 cpu = rq->mq_ctx->cpu;
1151 list = &per_cpu(blk_cpu_done, cpu);
1152 if (llist_add(&rq->ipi_list, list)) {
1153 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1154 smp_call_function_single_async(cpu, &rq->csd);
1158 static void blk_mq_raise_softirq(struct request *rq)
1160 struct llist_head *list;
1163 list = this_cpu_ptr(&blk_cpu_done);
1164 if (llist_add(&rq->ipi_list, list))
1165 raise_softirq(BLOCK_SOFTIRQ);
1169 bool blk_mq_complete_request_remote(struct request *rq)
1171 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1174 * For request which hctx has only one ctx mapping,
1175 * or a polled request, always complete locally,
1176 * it's pointless to redirect the completion.
1178 if (rq->mq_hctx->nr_ctx == 1 ||
1179 rq->cmd_flags & REQ_POLLED)
1182 if (blk_mq_complete_need_ipi(rq)) {
1183 blk_mq_complete_send_ipi(rq);
1187 if (rq->q->nr_hw_queues == 1) {
1188 blk_mq_raise_softirq(rq);
1193 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1196 * blk_mq_complete_request - end I/O on a request
1197 * @rq: the request being processed
1200 * Complete a request by scheduling the ->complete_rq operation.
1202 void blk_mq_complete_request(struct request *rq)
1204 if (!blk_mq_complete_request_remote(rq))
1205 rq->q->mq_ops->complete(rq);
1207 EXPORT_SYMBOL(blk_mq_complete_request);
1210 * blk_mq_start_request - Start processing a request
1211 * @rq: Pointer to request to be started
1213 * Function used by device drivers to notify the block layer that a request
1214 * is going to be processed now, so blk layer can do proper initializations
1215 * such as starting the timeout timer.
1217 void blk_mq_start_request(struct request *rq)
1219 struct request_queue *q = rq->q;
1221 trace_block_rq_issue(rq);
1223 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1224 rq->io_start_time_ns = ktime_get_ns();
1225 rq->stats_sectors = blk_rq_sectors(rq);
1226 rq->rq_flags |= RQF_STATS;
1227 rq_qos_issue(q, rq);
1230 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1233 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1235 #ifdef CONFIG_BLK_DEV_INTEGRITY
1236 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1237 q->integrity.profile->prepare_fn(rq);
1239 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1240 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1242 EXPORT_SYMBOL(blk_mq_start_request);
1245 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1246 * queues. This is important for md arrays to benefit from merging
1249 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1251 if (plug->multiple_queues)
1252 return BLK_MAX_REQUEST_COUNT * 2;
1253 return BLK_MAX_REQUEST_COUNT;
1256 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1258 struct request *last = rq_list_peek(&plug->mq_list);
1260 if (!plug->rq_count) {
1261 trace_block_plug(rq->q);
1262 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1263 (!blk_queue_nomerges(rq->q) &&
1264 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1265 blk_mq_flush_plug_list(plug, false);
1267 trace_block_plug(rq->q);
1270 if (!plug->multiple_queues && last && last->q != rq->q)
1271 plug->multiple_queues = true;
1272 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1273 plug->has_elevator = true;
1275 rq_list_add(&plug->mq_list, rq);
1280 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1281 * @rq: request to insert
1282 * @at_head: insert request at head or tail of queue
1285 * Insert a fully prepared request at the back of the I/O scheduler queue
1286 * for execution. Don't wait for completion.
1289 * This function will invoke @done directly if the queue is dead.
1291 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1293 WARN_ON(irqs_disabled());
1294 WARN_ON(!blk_rq_is_passthrough(rq));
1296 blk_account_io_start(rq);
1299 * As plugging can be enabled for passthrough requests on a zoned
1300 * device, directly accessing the plug instead of using blk_mq_plug()
1301 * should not have any consequences.
1304 blk_add_rq_to_plug(current->plug, rq);
1306 blk_mq_sched_insert_request(rq, at_head, true, false);
1308 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1310 struct blk_rq_wait {
1311 struct completion done;
1315 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1317 struct blk_rq_wait *wait = rq->end_io_data;
1320 complete(&wait->done);
1321 return RQ_END_IO_NONE;
1324 bool blk_rq_is_poll(struct request *rq)
1328 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1332 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1334 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1337 blk_mq_poll(rq->q, blk_rq_to_qc(rq), NULL, 0);
1339 } while (!completion_done(wait));
1343 * blk_execute_rq - insert a request into queue for execution
1344 * @rq: request to insert
1345 * @at_head: insert request at head or tail of queue
1348 * Insert a fully prepared request at the back of the I/O scheduler queue
1349 * for execution and wait for completion.
1350 * Return: The blk_status_t result provided to blk_mq_end_request().
1352 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1354 struct blk_rq_wait wait = {
1355 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1358 WARN_ON(irqs_disabled());
1359 WARN_ON(!blk_rq_is_passthrough(rq));
1361 rq->end_io_data = &wait;
1362 rq->end_io = blk_end_sync_rq;
1364 blk_account_io_start(rq);
1365 blk_mq_sched_insert_request(rq, at_head, true, false);
1367 if (blk_rq_is_poll(rq)) {
1368 blk_rq_poll_completion(rq, &wait.done);
1371 * Prevent hang_check timer from firing at us during very long
1374 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1377 while (!wait_for_completion_io_timeout(&wait.done,
1378 hang_check * (HZ/2)))
1381 wait_for_completion_io(&wait.done);
1386 EXPORT_SYMBOL(blk_execute_rq);
1388 static void __blk_mq_requeue_request(struct request *rq)
1390 struct request_queue *q = rq->q;
1392 blk_mq_put_driver_tag(rq);
1394 trace_block_rq_requeue(rq);
1395 rq_qos_requeue(q, rq);
1397 if (blk_mq_request_started(rq)) {
1398 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1399 rq->rq_flags &= ~RQF_TIMED_OUT;
1403 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1405 __blk_mq_requeue_request(rq);
1407 /* this request will be re-inserted to io scheduler queue */
1408 blk_mq_sched_requeue_request(rq);
1410 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1412 EXPORT_SYMBOL(blk_mq_requeue_request);
1414 static void blk_mq_requeue_work(struct work_struct *work)
1416 struct request_queue *q =
1417 container_of(work, struct request_queue, requeue_work.work);
1419 struct request *rq, *next;
1421 spin_lock_irq(&q->requeue_lock);
1422 list_splice_init(&q->requeue_list, &rq_list);
1423 spin_unlock_irq(&q->requeue_lock);
1425 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1426 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1429 rq->rq_flags &= ~RQF_SOFTBARRIER;
1430 list_del_init(&rq->queuelist);
1432 * If RQF_DONTPREP, rq has contained some driver specific
1433 * data, so insert it to hctx dispatch list to avoid any
1436 if (rq->rq_flags & RQF_DONTPREP)
1437 blk_mq_request_bypass_insert(rq, false, false);
1439 blk_mq_sched_insert_request(rq, true, false, false);
1442 while (!list_empty(&rq_list)) {
1443 rq = list_entry(rq_list.next, struct request, queuelist);
1444 list_del_init(&rq->queuelist);
1445 blk_mq_sched_insert_request(rq, false, false, false);
1448 blk_mq_run_hw_queues(q, false);
1451 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1452 bool kick_requeue_list)
1454 struct request_queue *q = rq->q;
1455 unsigned long flags;
1458 * We abuse this flag that is otherwise used by the I/O scheduler to
1459 * request head insertion from the workqueue.
1461 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1463 spin_lock_irqsave(&q->requeue_lock, flags);
1465 rq->rq_flags |= RQF_SOFTBARRIER;
1466 list_add(&rq->queuelist, &q->requeue_list);
1468 list_add_tail(&rq->queuelist, &q->requeue_list);
1470 spin_unlock_irqrestore(&q->requeue_lock, flags);
1472 if (kick_requeue_list)
1473 blk_mq_kick_requeue_list(q);
1476 void blk_mq_kick_requeue_list(struct request_queue *q)
1478 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1480 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1482 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1483 unsigned long msecs)
1485 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1486 msecs_to_jiffies(msecs));
1488 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1490 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1493 * If we find a request that isn't idle we know the queue is busy
1494 * as it's checked in the iter.
1495 * Return false to stop the iteration.
1497 if (blk_mq_request_started(rq)) {
1507 bool blk_mq_queue_inflight(struct request_queue *q)
1511 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1514 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1516 static void blk_mq_rq_timed_out(struct request *req)
1518 req->rq_flags |= RQF_TIMED_OUT;
1519 if (req->q->mq_ops->timeout) {
1520 enum blk_eh_timer_return ret;
1522 ret = req->q->mq_ops->timeout(req);
1523 if (ret == BLK_EH_DONE)
1525 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1531 struct blk_expired_data {
1532 bool has_timedout_rq;
1534 unsigned long timeout_start;
1537 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1539 unsigned long deadline;
1541 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1543 if (rq->rq_flags & RQF_TIMED_OUT)
1546 deadline = READ_ONCE(rq->deadline);
1547 if (time_after_eq(expired->timeout_start, deadline))
1550 if (expired->next == 0)
1551 expired->next = deadline;
1552 else if (time_after(expired->next, deadline))
1553 expired->next = deadline;
1557 void blk_mq_put_rq_ref(struct request *rq)
1559 if (is_flush_rq(rq)) {
1560 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1561 blk_mq_free_request(rq);
1562 } else if (req_ref_put_and_test(rq)) {
1563 __blk_mq_free_request(rq);
1567 static bool blk_mq_check_expired(struct request *rq, void *priv)
1569 struct blk_expired_data *expired = priv;
1572 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1573 * be reallocated underneath the timeout handler's processing, then
1574 * the expire check is reliable. If the request is not expired, then
1575 * it was completed and reallocated as a new request after returning
1576 * from blk_mq_check_expired().
1578 if (blk_mq_req_expired(rq, expired)) {
1579 expired->has_timedout_rq = true;
1585 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1587 struct blk_expired_data *expired = priv;
1589 if (blk_mq_req_expired(rq, expired))
1590 blk_mq_rq_timed_out(rq);
1594 static void blk_mq_timeout_work(struct work_struct *work)
1596 struct request_queue *q =
1597 container_of(work, struct request_queue, timeout_work);
1598 struct blk_expired_data expired = {
1599 .timeout_start = jiffies,
1601 struct blk_mq_hw_ctx *hctx;
1604 /* A deadlock might occur if a request is stuck requiring a
1605 * timeout at the same time a queue freeze is waiting
1606 * completion, since the timeout code would not be able to
1607 * acquire the queue reference here.
1609 * That's why we don't use blk_queue_enter here; instead, we use
1610 * percpu_ref_tryget directly, because we need to be able to
1611 * obtain a reference even in the short window between the queue
1612 * starting to freeze, by dropping the first reference in
1613 * blk_freeze_queue_start, and the moment the last request is
1614 * consumed, marked by the instant q_usage_counter reaches
1617 if (!percpu_ref_tryget(&q->q_usage_counter))
1620 /* check if there is any timed-out request */
1621 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1622 if (expired.has_timedout_rq) {
1624 * Before walking tags, we must ensure any submit started
1625 * before the current time has finished. Since the submit
1626 * uses srcu or rcu, wait for a synchronization point to
1627 * ensure all running submits have finished
1629 blk_mq_wait_quiesce_done(q);
1632 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1635 if (expired.next != 0) {
1636 mod_timer(&q->timeout, expired.next);
1639 * Request timeouts are handled as a forward rolling timer. If
1640 * we end up here it means that no requests are pending and
1641 * also that no request has been pending for a while. Mark
1642 * each hctx as idle.
1644 queue_for_each_hw_ctx(q, hctx, i) {
1645 /* the hctx may be unmapped, so check it here */
1646 if (blk_mq_hw_queue_mapped(hctx))
1647 blk_mq_tag_idle(hctx);
1653 struct flush_busy_ctx_data {
1654 struct blk_mq_hw_ctx *hctx;
1655 struct list_head *list;
1658 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1660 struct flush_busy_ctx_data *flush_data = data;
1661 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1662 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1663 enum hctx_type type = hctx->type;
1665 spin_lock(&ctx->lock);
1666 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1667 sbitmap_clear_bit(sb, bitnr);
1668 spin_unlock(&ctx->lock);
1673 * Process software queues that have been marked busy, splicing them
1674 * to the for-dispatch
1676 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1678 struct flush_busy_ctx_data data = {
1683 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1685 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1687 struct dispatch_rq_data {
1688 struct blk_mq_hw_ctx *hctx;
1692 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1695 struct dispatch_rq_data *dispatch_data = data;
1696 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1697 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1698 enum hctx_type type = hctx->type;
1700 spin_lock(&ctx->lock);
1701 if (!list_empty(&ctx->rq_lists[type])) {
1702 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1703 list_del_init(&dispatch_data->rq->queuelist);
1704 if (list_empty(&ctx->rq_lists[type]))
1705 sbitmap_clear_bit(sb, bitnr);
1707 spin_unlock(&ctx->lock);
1709 return !dispatch_data->rq;
1712 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1713 struct blk_mq_ctx *start)
1715 unsigned off = start ? start->index_hw[hctx->type] : 0;
1716 struct dispatch_rq_data data = {
1721 __sbitmap_for_each_set(&hctx->ctx_map, off,
1722 dispatch_rq_from_ctx, &data);
1727 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1729 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1730 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1733 blk_mq_tag_busy(rq->mq_hctx);
1735 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1736 bt = &rq->mq_hctx->tags->breserved_tags;
1739 if (!hctx_may_queue(rq->mq_hctx, bt))
1743 tag = __sbitmap_queue_get(bt);
1744 if (tag == BLK_MQ_NO_TAG)
1747 rq->tag = tag + tag_offset;
1751 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1753 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1756 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1757 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1758 rq->rq_flags |= RQF_MQ_INFLIGHT;
1759 __blk_mq_inc_active_requests(hctx);
1761 hctx->tags->rqs[rq->tag] = rq;
1765 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1766 int flags, void *key)
1768 struct blk_mq_hw_ctx *hctx;
1770 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1772 spin_lock(&hctx->dispatch_wait_lock);
1773 if (!list_empty(&wait->entry)) {
1774 struct sbitmap_queue *sbq;
1776 list_del_init(&wait->entry);
1777 sbq = &hctx->tags->bitmap_tags;
1778 atomic_dec(&sbq->ws_active);
1780 spin_unlock(&hctx->dispatch_wait_lock);
1782 blk_mq_run_hw_queue(hctx, true);
1787 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1788 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1789 * restart. For both cases, take care to check the condition again after
1790 * marking us as waiting.
1792 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1795 struct sbitmap_queue *sbq;
1796 struct wait_queue_head *wq;
1797 wait_queue_entry_t *wait;
1800 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1801 !(blk_mq_is_shared_tags(hctx->flags))) {
1802 blk_mq_sched_mark_restart_hctx(hctx);
1805 * It's possible that a tag was freed in the window between the
1806 * allocation failure and adding the hardware queue to the wait
1809 * Don't clear RESTART here, someone else could have set it.
1810 * At most this will cost an extra queue run.
1812 return blk_mq_get_driver_tag(rq);
1815 wait = &hctx->dispatch_wait;
1816 if (!list_empty_careful(&wait->entry))
1819 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1820 sbq = &hctx->tags->breserved_tags;
1822 sbq = &hctx->tags->bitmap_tags;
1823 wq = &bt_wait_ptr(sbq, hctx)->wait;
1825 spin_lock_irq(&wq->lock);
1826 spin_lock(&hctx->dispatch_wait_lock);
1827 if (!list_empty(&wait->entry)) {
1828 spin_unlock(&hctx->dispatch_wait_lock);
1829 spin_unlock_irq(&wq->lock);
1833 atomic_inc(&sbq->ws_active);
1834 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1835 __add_wait_queue(wq, wait);
1838 * It's possible that a tag was freed in the window between the
1839 * allocation failure and adding the hardware queue to the wait
1842 ret = blk_mq_get_driver_tag(rq);
1844 spin_unlock(&hctx->dispatch_wait_lock);
1845 spin_unlock_irq(&wq->lock);
1850 * We got a tag, remove ourselves from the wait queue to ensure
1851 * someone else gets the wakeup.
1853 list_del_init(&wait->entry);
1854 atomic_dec(&sbq->ws_active);
1855 spin_unlock(&hctx->dispatch_wait_lock);
1856 spin_unlock_irq(&wq->lock);
1861 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1862 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1864 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1865 * - EWMA is one simple way to compute running average value
1866 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1867 * - take 4 as factor for avoiding to get too small(0) result, and this
1868 * factor doesn't matter because EWMA decreases exponentially
1870 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1874 ewma = hctx->dispatch_busy;
1879 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1881 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1882 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1884 hctx->dispatch_busy = ewma;
1887 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1889 static void blk_mq_handle_dev_resource(struct request *rq,
1890 struct list_head *list)
1892 struct request *next =
1893 list_first_entry_or_null(list, struct request, queuelist);
1896 * If an I/O scheduler has been configured and we got a driver tag for
1897 * the next request already, free it.
1900 blk_mq_put_driver_tag(next);
1902 list_add(&rq->queuelist, list);
1903 __blk_mq_requeue_request(rq);
1906 static void blk_mq_handle_zone_resource(struct request *rq,
1907 struct list_head *zone_list)
1910 * If we end up here it is because we cannot dispatch a request to a
1911 * specific zone due to LLD level zone-write locking or other zone
1912 * related resource not being available. In this case, set the request
1913 * aside in zone_list for retrying it later.
1915 list_add(&rq->queuelist, zone_list);
1916 __blk_mq_requeue_request(rq);
1919 enum prep_dispatch {
1921 PREP_DISPATCH_NO_TAG,
1922 PREP_DISPATCH_NO_BUDGET,
1925 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1928 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1929 int budget_token = -1;
1932 budget_token = blk_mq_get_dispatch_budget(rq->q);
1933 if (budget_token < 0) {
1934 blk_mq_put_driver_tag(rq);
1935 return PREP_DISPATCH_NO_BUDGET;
1937 blk_mq_set_rq_budget_token(rq, budget_token);
1940 if (!blk_mq_get_driver_tag(rq)) {
1942 * The initial allocation attempt failed, so we need to
1943 * rerun the hardware queue when a tag is freed. The
1944 * waitqueue takes care of that. If the queue is run
1945 * before we add this entry back on the dispatch list,
1946 * we'll re-run it below.
1948 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1950 * All budgets not got from this function will be put
1951 * together during handling partial dispatch
1954 blk_mq_put_dispatch_budget(rq->q, budget_token);
1955 return PREP_DISPATCH_NO_TAG;
1959 return PREP_DISPATCH_OK;
1962 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1963 static void blk_mq_release_budgets(struct request_queue *q,
1964 struct list_head *list)
1968 list_for_each_entry(rq, list, queuelist) {
1969 int budget_token = blk_mq_get_rq_budget_token(rq);
1971 if (budget_token >= 0)
1972 blk_mq_put_dispatch_budget(q, budget_token);
1977 * Returns true if we did some work AND can potentially do more.
1979 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1980 unsigned int nr_budgets)
1982 enum prep_dispatch prep;
1983 struct request_queue *q = hctx->queue;
1984 struct request *rq, *nxt;
1986 blk_status_t ret = BLK_STS_OK;
1987 LIST_HEAD(zone_list);
1988 bool needs_resource = false;
1990 if (list_empty(list))
1994 * Now process all the entries, sending them to the driver.
1996 errors = queued = 0;
1998 struct blk_mq_queue_data bd;
2000 rq = list_first_entry(list, struct request, queuelist);
2002 WARN_ON_ONCE(hctx != rq->mq_hctx);
2003 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2004 if (prep != PREP_DISPATCH_OK)
2007 list_del_init(&rq->queuelist);
2012 * Flag last if we have no more requests, or if we have more
2013 * but can't assign a driver tag to it.
2015 if (list_empty(list))
2018 nxt = list_first_entry(list, struct request, queuelist);
2019 bd.last = !blk_mq_get_driver_tag(nxt);
2023 * once the request is queued to lld, no need to cover the
2028 ret = q->mq_ops->queue_rq(hctx, &bd);
2033 case BLK_STS_RESOURCE:
2034 needs_resource = true;
2036 case BLK_STS_DEV_RESOURCE:
2037 blk_mq_handle_dev_resource(rq, list);
2039 case BLK_STS_ZONE_RESOURCE:
2041 * Move the request to zone_list and keep going through
2042 * the dispatch list to find more requests the drive can
2045 blk_mq_handle_zone_resource(rq, &zone_list);
2046 needs_resource = true;
2050 blk_mq_end_request(rq, ret);
2052 } while (!list_empty(list));
2054 if (!list_empty(&zone_list))
2055 list_splice_tail_init(&zone_list, list);
2057 /* If we didn't flush the entire list, we could have told the driver
2058 * there was more coming, but that turned out to be a lie.
2060 if ((!list_empty(list) || errors || needs_resource ||
2061 ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued)
2062 q->mq_ops->commit_rqs(hctx);
2064 * Any items that need requeuing? Stuff them into hctx->dispatch,
2065 * that is where we will continue on next queue run.
2067 if (!list_empty(list)) {
2069 /* For non-shared tags, the RESTART check will suffice */
2070 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2071 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2072 blk_mq_is_shared_tags(hctx->flags));
2075 blk_mq_release_budgets(q, list);
2077 spin_lock(&hctx->lock);
2078 list_splice_tail_init(list, &hctx->dispatch);
2079 spin_unlock(&hctx->lock);
2082 * Order adding requests to hctx->dispatch and checking
2083 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2084 * in blk_mq_sched_restart(). Avoid restart code path to
2085 * miss the new added requests to hctx->dispatch, meantime
2086 * SCHED_RESTART is observed here.
2091 * If SCHED_RESTART was set by the caller of this function and
2092 * it is no longer set that means that it was cleared by another
2093 * thread and hence that a queue rerun is needed.
2095 * If 'no_tag' is set, that means that we failed getting
2096 * a driver tag with an I/O scheduler attached. If our dispatch
2097 * waitqueue is no longer active, ensure that we run the queue
2098 * AFTER adding our entries back to the list.
2100 * If no I/O scheduler has been configured it is possible that
2101 * the hardware queue got stopped and restarted before requests
2102 * were pushed back onto the dispatch list. Rerun the queue to
2103 * avoid starvation. Notes:
2104 * - blk_mq_run_hw_queue() checks whether or not a queue has
2105 * been stopped before rerunning a queue.
2106 * - Some but not all block drivers stop a queue before
2107 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2110 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2111 * bit is set, run queue after a delay to avoid IO stalls
2112 * that could otherwise occur if the queue is idle. We'll do
2113 * similar if we couldn't get budget or couldn't lock a zone
2114 * and SCHED_RESTART is set.
2116 needs_restart = blk_mq_sched_needs_restart(hctx);
2117 if (prep == PREP_DISPATCH_NO_BUDGET)
2118 needs_resource = true;
2119 if (!needs_restart ||
2120 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2121 blk_mq_run_hw_queue(hctx, true);
2122 else if (needs_resource)
2123 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2125 blk_mq_update_dispatch_busy(hctx, true);
2128 blk_mq_update_dispatch_busy(hctx, false);
2130 return (queued + errors) != 0;
2134 * __blk_mq_run_hw_queue - Run a hardware queue.
2135 * @hctx: Pointer to the hardware queue to run.
2137 * Send pending requests to the hardware.
2139 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
2142 * We can't run the queue inline with ints disabled. Ensure that
2143 * we catch bad users of this early.
2145 WARN_ON_ONCE(in_interrupt());
2147 blk_mq_run_dispatch_ops(hctx->queue,
2148 blk_mq_sched_dispatch_requests(hctx));
2151 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2153 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2155 if (cpu >= nr_cpu_ids)
2156 cpu = cpumask_first(hctx->cpumask);
2161 * It'd be great if the workqueue API had a way to pass
2162 * in a mask and had some smarts for more clever placement.
2163 * For now we just round-robin here, switching for every
2164 * BLK_MQ_CPU_WORK_BATCH queued items.
2166 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2169 int next_cpu = hctx->next_cpu;
2171 if (hctx->queue->nr_hw_queues == 1)
2172 return WORK_CPU_UNBOUND;
2174 if (--hctx->next_cpu_batch <= 0) {
2176 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2178 if (next_cpu >= nr_cpu_ids)
2179 next_cpu = blk_mq_first_mapped_cpu(hctx);
2180 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2184 * Do unbound schedule if we can't find a online CPU for this hctx,
2185 * and it should only happen in the path of handling CPU DEAD.
2187 if (!cpu_online(next_cpu)) {
2194 * Make sure to re-select CPU next time once after CPUs
2195 * in hctx->cpumask become online again.
2197 hctx->next_cpu = next_cpu;
2198 hctx->next_cpu_batch = 1;
2199 return WORK_CPU_UNBOUND;
2202 hctx->next_cpu = next_cpu;
2207 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2208 * @hctx: Pointer to the hardware queue to run.
2209 * @async: If we want to run the queue asynchronously.
2210 * @msecs: Milliseconds of delay to wait before running the queue.
2212 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2213 * with a delay of @msecs.
2215 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2216 unsigned long msecs)
2218 if (unlikely(blk_mq_hctx_stopped(hctx)))
2221 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2222 if (cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2223 __blk_mq_run_hw_queue(hctx);
2228 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2229 msecs_to_jiffies(msecs));
2233 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2234 * @hctx: Pointer to the hardware queue to run.
2235 * @msecs: Milliseconds of delay to wait before running the queue.
2237 * Run a hardware queue asynchronously with a delay of @msecs.
2239 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2241 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2243 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2246 * blk_mq_run_hw_queue - Start to run a hardware queue.
2247 * @hctx: Pointer to the hardware queue to run.
2248 * @async: If we want to run the queue asynchronously.
2250 * Check if the request queue is not in a quiesced state and if there are
2251 * pending requests to be sent. If this is true, run the queue to send requests
2254 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2259 * When queue is quiesced, we may be switching io scheduler, or
2260 * updating nr_hw_queues, or other things, and we can't run queue
2261 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2263 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2266 __blk_mq_run_dispatch_ops(hctx->queue, false,
2267 need_run = !blk_queue_quiesced(hctx->queue) &&
2268 blk_mq_hctx_has_pending(hctx));
2271 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2273 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2276 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2279 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2281 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2283 * If the IO scheduler does not respect hardware queues when
2284 * dispatching, we just don't bother with multiple HW queues and
2285 * dispatch from hctx for the current CPU since running multiple queues
2286 * just causes lock contention inside the scheduler and pointless cache
2289 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2291 if (!blk_mq_hctx_stopped(hctx))
2297 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2298 * @q: Pointer to the request queue to run.
2299 * @async: If we want to run the queue asynchronously.
2301 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2303 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2307 if (blk_queue_sq_sched(q))
2308 sq_hctx = blk_mq_get_sq_hctx(q);
2309 queue_for_each_hw_ctx(q, hctx, i) {
2310 if (blk_mq_hctx_stopped(hctx))
2313 * Dispatch from this hctx either if there's no hctx preferred
2314 * by IO scheduler or if it has requests that bypass the
2317 if (!sq_hctx || sq_hctx == hctx ||
2318 !list_empty_careful(&hctx->dispatch))
2319 blk_mq_run_hw_queue(hctx, async);
2322 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2325 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2326 * @q: Pointer to the request queue to run.
2327 * @msecs: Milliseconds of delay to wait before running the queues.
2329 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2331 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2335 if (blk_queue_sq_sched(q))
2336 sq_hctx = blk_mq_get_sq_hctx(q);
2337 queue_for_each_hw_ctx(q, hctx, i) {
2338 if (blk_mq_hctx_stopped(hctx))
2341 * If there is already a run_work pending, leave the
2342 * pending delay untouched. Otherwise, a hctx can stall
2343 * if another hctx is re-delaying the other's work
2344 * before the work executes.
2346 if (delayed_work_pending(&hctx->run_work))
2349 * Dispatch from this hctx either if there's no hctx preferred
2350 * by IO scheduler or if it has requests that bypass the
2353 if (!sq_hctx || sq_hctx == hctx ||
2354 !list_empty_careful(&hctx->dispatch))
2355 blk_mq_delay_run_hw_queue(hctx, msecs);
2358 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2361 * This function is often used for pausing .queue_rq() by driver when
2362 * there isn't enough resource or some conditions aren't satisfied, and
2363 * BLK_STS_RESOURCE is usually returned.
2365 * We do not guarantee that dispatch can be drained or blocked
2366 * after blk_mq_stop_hw_queue() returns. Please use
2367 * blk_mq_quiesce_queue() for that requirement.
2369 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2371 cancel_delayed_work(&hctx->run_work);
2373 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2375 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2378 * This function is often used for pausing .queue_rq() by driver when
2379 * there isn't enough resource or some conditions aren't satisfied, and
2380 * BLK_STS_RESOURCE is usually returned.
2382 * We do not guarantee that dispatch can be drained or blocked
2383 * after blk_mq_stop_hw_queues() returns. Please use
2384 * blk_mq_quiesce_queue() for that requirement.
2386 void blk_mq_stop_hw_queues(struct request_queue *q)
2388 struct blk_mq_hw_ctx *hctx;
2391 queue_for_each_hw_ctx(q, hctx, i)
2392 blk_mq_stop_hw_queue(hctx);
2394 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2396 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2398 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2400 blk_mq_run_hw_queue(hctx, false);
2402 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2404 void blk_mq_start_hw_queues(struct request_queue *q)
2406 struct blk_mq_hw_ctx *hctx;
2409 queue_for_each_hw_ctx(q, hctx, i)
2410 blk_mq_start_hw_queue(hctx);
2412 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2414 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2416 if (!blk_mq_hctx_stopped(hctx))
2419 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2420 blk_mq_run_hw_queue(hctx, async);
2422 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2424 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2426 struct blk_mq_hw_ctx *hctx;
2429 queue_for_each_hw_ctx(q, hctx, i)
2430 blk_mq_start_stopped_hw_queue(hctx, async);
2432 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2434 static void blk_mq_run_work_fn(struct work_struct *work)
2436 struct blk_mq_hw_ctx *hctx;
2438 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2441 * If we are stopped, don't run the queue.
2443 if (blk_mq_hctx_stopped(hctx))
2446 __blk_mq_run_hw_queue(hctx);
2449 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2453 struct blk_mq_ctx *ctx = rq->mq_ctx;
2454 enum hctx_type type = hctx->type;
2456 lockdep_assert_held(&ctx->lock);
2458 trace_block_rq_insert(rq);
2461 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2463 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2466 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2469 struct blk_mq_ctx *ctx = rq->mq_ctx;
2471 lockdep_assert_held(&ctx->lock);
2473 __blk_mq_insert_req_list(hctx, rq, at_head);
2474 blk_mq_hctx_mark_pending(hctx, ctx);
2478 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2479 * @rq: Pointer to request to be inserted.
2480 * @at_head: true if the request should be inserted at the head of the list.
2481 * @run_queue: If we should run the hardware queue after inserting the request.
2483 * Should only be used carefully, when the caller knows we want to
2484 * bypass a potential IO scheduler on the target device.
2486 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2489 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2491 spin_lock(&hctx->lock);
2493 list_add(&rq->queuelist, &hctx->dispatch);
2495 list_add_tail(&rq->queuelist, &hctx->dispatch);
2496 spin_unlock(&hctx->lock);
2499 blk_mq_run_hw_queue(hctx, false);
2502 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2503 struct list_head *list)
2507 enum hctx_type type = hctx->type;
2510 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2513 list_for_each_entry(rq, list, queuelist) {
2514 BUG_ON(rq->mq_ctx != ctx);
2515 trace_block_rq_insert(rq);
2518 spin_lock(&ctx->lock);
2519 list_splice_tail_init(list, &ctx->rq_lists[type]);
2520 blk_mq_hctx_mark_pending(hctx, ctx);
2521 spin_unlock(&ctx->lock);
2524 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2527 if (hctx->queue->mq_ops->commit_rqs) {
2528 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2529 hctx->queue->mq_ops->commit_rqs(hctx);
2534 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2535 unsigned int nr_segs)
2539 if (bio->bi_opf & REQ_RAHEAD)
2540 rq->cmd_flags |= REQ_FAILFAST_MASK;
2542 rq->__sector = bio->bi_iter.bi_sector;
2543 blk_rq_bio_prep(rq, bio, nr_segs);
2545 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2546 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2549 blk_account_io_start(rq);
2552 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2553 struct request *rq, bool last)
2555 struct request_queue *q = rq->q;
2556 struct blk_mq_queue_data bd = {
2563 * For OK queue, we are done. For error, caller may kill it.
2564 * Any other error (busy), just add it to our list as we
2565 * previously would have done.
2567 ret = q->mq_ops->queue_rq(hctx, &bd);
2570 blk_mq_update_dispatch_busy(hctx, false);
2572 case BLK_STS_RESOURCE:
2573 case BLK_STS_DEV_RESOURCE:
2574 blk_mq_update_dispatch_busy(hctx, true);
2575 __blk_mq_requeue_request(rq);
2578 blk_mq_update_dispatch_busy(hctx, false);
2585 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2587 bool bypass_insert, bool last)
2589 struct request_queue *q = rq->q;
2590 bool run_queue = true;
2594 * RCU or SRCU read lock is needed before checking quiesced flag.
2596 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2597 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2598 * and avoid driver to try to dispatch again.
2600 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2602 bypass_insert = false;
2606 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2609 budget_token = blk_mq_get_dispatch_budget(q);
2610 if (budget_token < 0)
2613 blk_mq_set_rq_budget_token(rq, budget_token);
2615 if (!blk_mq_get_driver_tag(rq)) {
2616 blk_mq_put_dispatch_budget(q, budget_token);
2620 return __blk_mq_issue_directly(hctx, rq, last);
2623 return BLK_STS_RESOURCE;
2625 blk_mq_sched_insert_request(rq, false, run_queue, false);
2631 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2632 * @hctx: Pointer of the associated hardware queue.
2633 * @rq: Pointer to request to be sent.
2635 * If the device has enough resources to accept a new request now, send the
2636 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2637 * we can try send it another time in the future. Requests inserted at this
2638 * queue have higher priority.
2640 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2644 __blk_mq_try_issue_directly(hctx, rq, false, true);
2646 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2647 blk_mq_request_bypass_insert(rq, false, true);
2648 else if (ret != BLK_STS_OK)
2649 blk_mq_end_request(rq, ret);
2652 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2654 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2657 static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2659 struct blk_mq_hw_ctx *hctx = NULL;
2664 while ((rq = rq_list_pop(&plug->mq_list))) {
2665 bool last = rq_list_empty(plug->mq_list);
2668 if (hctx != rq->mq_hctx) {
2670 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2674 ret = blk_mq_request_issue_directly(rq, last);
2679 case BLK_STS_RESOURCE:
2680 case BLK_STS_DEV_RESOURCE:
2681 blk_mq_request_bypass_insert(rq, false, true);
2682 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2685 blk_mq_end_request(rq, ret);
2692 * If we didn't flush the entire list, we could have told the driver
2693 * there was more coming, but that turned out to be a lie.
2696 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2699 static void __blk_mq_flush_plug_list(struct request_queue *q,
2700 struct blk_plug *plug)
2702 if (blk_queue_quiesced(q))
2704 q->mq_ops->queue_rqs(&plug->mq_list);
2707 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2709 struct blk_mq_hw_ctx *this_hctx = NULL;
2710 struct blk_mq_ctx *this_ctx = NULL;
2711 struct request *requeue_list = NULL;
2712 struct request **requeue_lastp = &requeue_list;
2713 unsigned int depth = 0;
2717 struct request *rq = rq_list_pop(&plug->mq_list);
2720 this_hctx = rq->mq_hctx;
2721 this_ctx = rq->mq_ctx;
2722 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2723 rq_list_add_tail(&requeue_lastp, rq);
2726 list_add(&rq->queuelist, &list);
2728 } while (!rq_list_empty(plug->mq_list));
2730 plug->mq_list = requeue_list;
2731 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2732 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched);
2735 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2739 if (rq_list_empty(plug->mq_list))
2743 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2744 struct request_queue *q;
2746 rq = rq_list_peek(&plug->mq_list);
2750 * Peek first request and see if we have a ->queue_rqs() hook.
2751 * If we do, we can dispatch the whole plug list in one go. We
2752 * already know at this point that all requests belong to the
2753 * same queue, caller must ensure that's the case.
2755 * Since we pass off the full list to the driver at this point,
2756 * we do not increment the active request count for the queue.
2757 * Bypass shared tags for now because of that.
2759 if (q->mq_ops->queue_rqs &&
2760 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2761 blk_mq_run_dispatch_ops(q,
2762 __blk_mq_flush_plug_list(q, plug));
2763 if (rq_list_empty(plug->mq_list))
2767 blk_mq_run_dispatch_ops(q,
2768 blk_mq_plug_issue_direct(plug, false));
2769 if (rq_list_empty(plug->mq_list))
2774 blk_mq_dispatch_plug_list(plug, from_schedule);
2775 } while (!rq_list_empty(plug->mq_list));
2778 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2779 struct list_head *list)
2784 while (!list_empty(list)) {
2786 struct request *rq = list_first_entry(list, struct request,
2789 list_del_init(&rq->queuelist);
2790 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2791 if (ret != BLK_STS_OK) {
2793 if (ret == BLK_STS_RESOURCE ||
2794 ret == BLK_STS_DEV_RESOURCE) {
2795 blk_mq_request_bypass_insert(rq, false,
2799 blk_mq_end_request(rq, ret);
2805 * If we didn't flush the entire list, we could have told
2806 * the driver there was more coming, but that turned out to
2809 if ((!list_empty(list) || errors) &&
2810 hctx->queue->mq_ops->commit_rqs && queued)
2811 hctx->queue->mq_ops->commit_rqs(hctx);
2814 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2815 struct bio *bio, unsigned int nr_segs)
2817 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2818 if (blk_attempt_plug_merge(q, bio, nr_segs))
2820 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2826 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2827 struct blk_plug *plug,
2831 struct blk_mq_alloc_data data = {
2834 .cmd_flags = bio->bi_opf,
2838 if (unlikely(bio_queue_enter(bio)))
2841 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2844 rq_qos_throttle(q, bio);
2847 data.nr_tags = plug->nr_ios;
2849 data.cached_rq = &plug->cached_rq;
2852 rq = __blk_mq_alloc_requests(&data);
2855 rq_qos_cleanup(q, bio);
2856 if (bio->bi_opf & REQ_NOWAIT)
2857 bio_wouldblock_error(bio);
2863 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2864 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2867 enum hctx_type type, hctx_type;
2871 rq = rq_list_peek(&plug->cached_rq);
2872 if (!rq || rq->q != q)
2875 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2880 type = blk_mq_get_hctx_type((*bio)->bi_opf);
2881 hctx_type = rq->mq_hctx->type;
2882 if (type != hctx_type &&
2883 !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2885 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2889 * If any qos ->throttle() end up blocking, we will have flushed the
2890 * plug and hence killed the cached_rq list as well. Pop this entry
2891 * before we throttle.
2893 plug->cached_rq = rq_list_next(rq);
2894 rq_qos_throttle(q, *bio);
2896 rq->cmd_flags = (*bio)->bi_opf;
2897 INIT_LIST_HEAD(&rq->queuelist);
2901 static void bio_set_ioprio(struct bio *bio)
2903 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2904 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2905 bio->bi_ioprio = get_current_ioprio();
2906 blkcg_set_ioprio(bio);
2910 * blk_mq_submit_bio - Create and send a request to block device.
2911 * @bio: Bio pointer.
2913 * Builds up a request structure from @q and @bio and send to the device. The
2914 * request may not be queued directly to hardware if:
2915 * * This request can be merged with another one
2916 * * We want to place request at plug queue for possible future merging
2917 * * There is an IO scheduler active at this queue
2919 * It will not queue the request if there is an error with the bio, or at the
2922 void blk_mq_submit_bio(struct bio *bio)
2924 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2925 struct blk_plug *plug = blk_mq_plug(bio);
2926 const int is_sync = op_is_sync(bio->bi_opf);
2928 unsigned int nr_segs = 1;
2931 bio = blk_queue_bounce(bio, q);
2932 if (bio_may_exceed_limits(bio, &q->limits)) {
2933 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2938 if (!bio_integrity_prep(bio))
2941 bio_set_ioprio(bio);
2943 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2947 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2952 trace_block_getrq(bio);
2954 rq_qos_track(q, rq, bio);
2956 blk_mq_bio_to_request(rq, bio, nr_segs);
2958 ret = blk_crypto_init_request(rq);
2959 if (ret != BLK_STS_OK) {
2960 bio->bi_status = ret;
2962 blk_mq_free_request(rq);
2966 if (op_is_flush(bio->bi_opf)) {
2967 blk_insert_flush(rq);
2972 blk_add_rq_to_plug(plug, rq);
2973 else if ((rq->rq_flags & RQF_ELV) ||
2974 (rq->mq_hctx->dispatch_busy &&
2975 (q->nr_hw_queues == 1 || !is_sync)))
2976 blk_mq_sched_insert_request(rq, false, true, true);
2978 blk_mq_run_dispatch_ops(rq->q,
2979 blk_mq_try_issue_directly(rq->mq_hctx, rq));
2982 #ifdef CONFIG_BLK_MQ_STACKING
2984 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
2985 * @rq: the request being queued
2987 blk_status_t blk_insert_cloned_request(struct request *rq)
2989 struct request_queue *q = rq->q;
2990 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
2993 if (blk_rq_sectors(rq) > max_sectors) {
2995 * SCSI device does not have a good way to return if
2996 * Write Same/Zero is actually supported. If a device rejects
2997 * a non-read/write command (discard, write same,etc.) the
2998 * low-level device driver will set the relevant queue limit to
2999 * 0 to prevent blk-lib from issuing more of the offending
3000 * operations. Commands queued prior to the queue limit being
3001 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3002 * errors being propagated to upper layers.
3004 if (max_sectors == 0)
3005 return BLK_STS_NOTSUPP;
3007 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3008 __func__, blk_rq_sectors(rq), max_sectors);
3009 return BLK_STS_IOERR;
3013 * The queue settings related to segment counting may differ from the
3016 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3017 if (rq->nr_phys_segments > queue_max_segments(q)) {
3018 printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n",
3019 __func__, rq->nr_phys_segments, queue_max_segments(q));
3020 return BLK_STS_IOERR;
3023 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3024 return BLK_STS_IOERR;
3026 if (blk_crypto_insert_cloned_request(rq))
3027 return BLK_STS_IOERR;
3029 blk_account_io_start(rq);
3032 * Since we have a scheduler attached on the top device,
3033 * bypass a potential scheduler on the bottom device for
3036 blk_mq_run_dispatch_ops(q,
3037 ret = blk_mq_request_issue_directly(rq, true));
3039 blk_account_io_done(rq, ktime_get_ns());
3042 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3045 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3046 * @rq: the clone request to be cleaned up
3049 * Free all bios in @rq for a cloned request.
3051 void blk_rq_unprep_clone(struct request *rq)
3055 while ((bio = rq->bio) != NULL) {
3056 rq->bio = bio->bi_next;
3061 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3064 * blk_rq_prep_clone - Helper function to setup clone request
3065 * @rq: the request to be setup
3066 * @rq_src: original request to be cloned
3067 * @bs: bio_set that bios for clone are allocated from
3068 * @gfp_mask: memory allocation mask for bio
3069 * @bio_ctr: setup function to be called for each clone bio.
3070 * Returns %0 for success, non %0 for failure.
3071 * @data: private data to be passed to @bio_ctr
3074 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3075 * Also, pages which the original bios are pointing to are not copied
3076 * and the cloned bios just point same pages.
3077 * So cloned bios must be completed before original bios, which means
3078 * the caller must complete @rq before @rq_src.
3080 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3081 struct bio_set *bs, gfp_t gfp_mask,
3082 int (*bio_ctr)(struct bio *, struct bio *, void *),
3085 struct bio *bio, *bio_src;
3090 __rq_for_each_bio(bio_src, rq_src) {
3091 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3096 if (bio_ctr && bio_ctr(bio, bio_src, data))
3100 rq->biotail->bi_next = bio;
3103 rq->bio = rq->biotail = bio;
3108 /* Copy attributes of the original request to the clone request. */
3109 rq->__sector = blk_rq_pos(rq_src);
3110 rq->__data_len = blk_rq_bytes(rq_src);
3111 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3112 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3113 rq->special_vec = rq_src->special_vec;
3115 rq->nr_phys_segments = rq_src->nr_phys_segments;
3116 rq->ioprio = rq_src->ioprio;
3118 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3126 blk_rq_unprep_clone(rq);
3130 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3131 #endif /* CONFIG_BLK_MQ_STACKING */
3134 * Steal bios from a request and add them to a bio list.
3135 * The request must not have been partially completed before.
3137 void blk_steal_bios(struct bio_list *list, struct request *rq)
3141 list->tail->bi_next = rq->bio;
3143 list->head = rq->bio;
3144 list->tail = rq->biotail;
3152 EXPORT_SYMBOL_GPL(blk_steal_bios);
3154 static size_t order_to_size(unsigned int order)
3156 return (size_t)PAGE_SIZE << order;
3159 /* called before freeing request pool in @tags */
3160 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3161 struct blk_mq_tags *tags)
3164 unsigned long flags;
3167 * There is no need to clear mapping if driver tags is not initialized
3168 * or the mapping belongs to the driver tags.
3170 if (!drv_tags || drv_tags == tags)
3173 list_for_each_entry(page, &tags->page_list, lru) {
3174 unsigned long start = (unsigned long)page_address(page);
3175 unsigned long end = start + order_to_size(page->private);
3178 for (i = 0; i < drv_tags->nr_tags; i++) {
3179 struct request *rq = drv_tags->rqs[i];
3180 unsigned long rq_addr = (unsigned long)rq;
3182 if (rq_addr >= start && rq_addr < end) {
3183 WARN_ON_ONCE(req_ref_read(rq) != 0);
3184 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3190 * Wait until all pending iteration is done.
3192 * Request reference is cleared and it is guaranteed to be observed
3193 * after the ->lock is released.
3195 spin_lock_irqsave(&drv_tags->lock, flags);
3196 spin_unlock_irqrestore(&drv_tags->lock, flags);
3199 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3200 unsigned int hctx_idx)
3202 struct blk_mq_tags *drv_tags;
3205 if (list_empty(&tags->page_list))
3208 if (blk_mq_is_shared_tags(set->flags))
3209 drv_tags = set->shared_tags;
3211 drv_tags = set->tags[hctx_idx];
3213 if (tags->static_rqs && set->ops->exit_request) {
3216 for (i = 0; i < tags->nr_tags; i++) {
3217 struct request *rq = tags->static_rqs[i];
3221 set->ops->exit_request(set, rq, hctx_idx);
3222 tags->static_rqs[i] = NULL;
3226 blk_mq_clear_rq_mapping(drv_tags, tags);
3228 while (!list_empty(&tags->page_list)) {
3229 page = list_first_entry(&tags->page_list, struct page, lru);
3230 list_del_init(&page->lru);
3232 * Remove kmemleak object previously allocated in
3233 * blk_mq_alloc_rqs().
3235 kmemleak_free(page_address(page));
3236 __free_pages(page, page->private);
3240 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3244 kfree(tags->static_rqs);
3245 tags->static_rqs = NULL;
3247 blk_mq_free_tags(tags);
3250 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3251 unsigned int hctx_idx)
3255 for (i = 0; i < set->nr_maps; i++) {
3256 unsigned int start = set->map[i].queue_offset;
3257 unsigned int end = start + set->map[i].nr_queues;
3259 if (hctx_idx >= start && hctx_idx < end)
3263 if (i >= set->nr_maps)
3264 i = HCTX_TYPE_DEFAULT;
3269 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3270 unsigned int hctx_idx)
3272 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3274 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3277 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3278 unsigned int hctx_idx,
3279 unsigned int nr_tags,
3280 unsigned int reserved_tags)
3282 int node = blk_mq_get_hctx_node(set, hctx_idx);
3283 struct blk_mq_tags *tags;
3285 if (node == NUMA_NO_NODE)
3286 node = set->numa_node;
3288 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3289 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3293 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3294 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3297 blk_mq_free_tags(tags);
3301 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3302 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3304 if (!tags->static_rqs) {
3306 blk_mq_free_tags(tags);
3313 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3314 unsigned int hctx_idx, int node)
3318 if (set->ops->init_request) {
3319 ret = set->ops->init_request(set, rq, hctx_idx, node);
3324 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3328 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3329 struct blk_mq_tags *tags,
3330 unsigned int hctx_idx, unsigned int depth)
3332 unsigned int i, j, entries_per_page, max_order = 4;
3333 int node = blk_mq_get_hctx_node(set, hctx_idx);
3334 size_t rq_size, left;
3336 if (node == NUMA_NO_NODE)
3337 node = set->numa_node;
3339 INIT_LIST_HEAD(&tags->page_list);
3342 * rq_size is the size of the request plus driver payload, rounded
3343 * to the cacheline size
3345 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3347 left = rq_size * depth;
3349 for (i = 0; i < depth; ) {
3350 int this_order = max_order;
3355 while (this_order && left < order_to_size(this_order - 1))
3359 page = alloc_pages_node(node,
3360 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3366 if (order_to_size(this_order) < rq_size)
3373 page->private = this_order;
3374 list_add_tail(&page->lru, &tags->page_list);
3376 p = page_address(page);
3378 * Allow kmemleak to scan these pages as they contain pointers
3379 * to additional allocations like via ops->init_request().
3381 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3382 entries_per_page = order_to_size(this_order) / rq_size;
3383 to_do = min(entries_per_page, depth - i);
3384 left -= to_do * rq_size;
3385 for (j = 0; j < to_do; j++) {
3386 struct request *rq = p;
3388 tags->static_rqs[i] = rq;
3389 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3390 tags->static_rqs[i] = NULL;
3401 blk_mq_free_rqs(set, tags, hctx_idx);
3405 struct rq_iter_data {
3406 struct blk_mq_hw_ctx *hctx;
3410 static bool blk_mq_has_request(struct request *rq, void *data)
3412 struct rq_iter_data *iter_data = data;
3414 if (rq->mq_hctx != iter_data->hctx)
3416 iter_data->has_rq = true;
3420 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3422 struct blk_mq_tags *tags = hctx->sched_tags ?
3423 hctx->sched_tags : hctx->tags;
3424 struct rq_iter_data data = {
3428 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3432 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3433 struct blk_mq_hw_ctx *hctx)
3435 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3437 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3442 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3444 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3445 struct blk_mq_hw_ctx, cpuhp_online);
3447 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3448 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3452 * Prevent new request from being allocated on the current hctx.
3454 * The smp_mb__after_atomic() Pairs with the implied barrier in
3455 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3456 * seen once we return from the tag allocator.
3458 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3459 smp_mb__after_atomic();
3462 * Try to grab a reference to the queue and wait for any outstanding
3463 * requests. If we could not grab a reference the queue has been
3464 * frozen and there are no requests.
3466 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3467 while (blk_mq_hctx_has_requests(hctx))
3469 percpu_ref_put(&hctx->queue->q_usage_counter);
3475 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3477 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3478 struct blk_mq_hw_ctx, cpuhp_online);
3480 if (cpumask_test_cpu(cpu, hctx->cpumask))
3481 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3486 * 'cpu' is going away. splice any existing rq_list entries from this
3487 * software queue to the hw queue dispatch list, and ensure that it
3490 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3492 struct blk_mq_hw_ctx *hctx;
3493 struct blk_mq_ctx *ctx;
3495 enum hctx_type type;
3497 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3498 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3501 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3504 spin_lock(&ctx->lock);
3505 if (!list_empty(&ctx->rq_lists[type])) {
3506 list_splice_init(&ctx->rq_lists[type], &tmp);
3507 blk_mq_hctx_clear_pending(hctx, ctx);
3509 spin_unlock(&ctx->lock);
3511 if (list_empty(&tmp))
3514 spin_lock(&hctx->lock);
3515 list_splice_tail_init(&tmp, &hctx->dispatch);
3516 spin_unlock(&hctx->lock);
3518 blk_mq_run_hw_queue(hctx, true);
3522 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3524 if (!(hctx->flags & BLK_MQ_F_STACKING))
3525 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3526 &hctx->cpuhp_online);
3527 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3532 * Before freeing hw queue, clearing the flush request reference in
3533 * tags->rqs[] for avoiding potential UAF.
3535 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3536 unsigned int queue_depth, struct request *flush_rq)
3539 unsigned long flags;
3541 /* The hw queue may not be mapped yet */
3545 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3547 for (i = 0; i < queue_depth; i++)
3548 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3551 * Wait until all pending iteration is done.
3553 * Request reference is cleared and it is guaranteed to be observed
3554 * after the ->lock is released.
3556 spin_lock_irqsave(&tags->lock, flags);
3557 spin_unlock_irqrestore(&tags->lock, flags);
3560 /* hctx->ctxs will be freed in queue's release handler */
3561 static void blk_mq_exit_hctx(struct request_queue *q,
3562 struct blk_mq_tag_set *set,
3563 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3565 struct request *flush_rq = hctx->fq->flush_rq;
3567 if (blk_mq_hw_queue_mapped(hctx))
3568 blk_mq_tag_idle(hctx);
3570 if (blk_queue_init_done(q))
3571 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3572 set->queue_depth, flush_rq);
3573 if (set->ops->exit_request)
3574 set->ops->exit_request(set, flush_rq, hctx_idx);
3576 if (set->ops->exit_hctx)
3577 set->ops->exit_hctx(hctx, hctx_idx);
3579 blk_mq_remove_cpuhp(hctx);
3581 xa_erase(&q->hctx_table, hctx_idx);
3583 spin_lock(&q->unused_hctx_lock);
3584 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3585 spin_unlock(&q->unused_hctx_lock);
3588 static void blk_mq_exit_hw_queues(struct request_queue *q,
3589 struct blk_mq_tag_set *set, int nr_queue)
3591 struct blk_mq_hw_ctx *hctx;
3594 queue_for_each_hw_ctx(q, hctx, i) {
3597 blk_mq_exit_hctx(q, set, hctx, i);
3601 static int blk_mq_init_hctx(struct request_queue *q,
3602 struct blk_mq_tag_set *set,
3603 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3605 hctx->queue_num = hctx_idx;
3607 if (!(hctx->flags & BLK_MQ_F_STACKING))
3608 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3609 &hctx->cpuhp_online);
3610 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3612 hctx->tags = set->tags[hctx_idx];
3614 if (set->ops->init_hctx &&
3615 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3616 goto unregister_cpu_notifier;
3618 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3622 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3628 if (set->ops->exit_request)
3629 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3631 if (set->ops->exit_hctx)
3632 set->ops->exit_hctx(hctx, hctx_idx);
3633 unregister_cpu_notifier:
3634 blk_mq_remove_cpuhp(hctx);
3638 static struct blk_mq_hw_ctx *
3639 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3642 struct blk_mq_hw_ctx *hctx;
3643 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3645 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3647 goto fail_alloc_hctx;
3649 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3652 atomic_set(&hctx->nr_active, 0);
3653 if (node == NUMA_NO_NODE)
3654 node = set->numa_node;
3655 hctx->numa_node = node;
3657 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3658 spin_lock_init(&hctx->lock);
3659 INIT_LIST_HEAD(&hctx->dispatch);
3661 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3663 INIT_LIST_HEAD(&hctx->hctx_list);
3666 * Allocate space for all possible cpus to avoid allocation at
3669 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3674 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3675 gfp, node, false, false))
3679 spin_lock_init(&hctx->dispatch_wait_lock);
3680 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3681 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3683 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3687 blk_mq_hctx_kobj_init(hctx);
3692 sbitmap_free(&hctx->ctx_map);
3696 free_cpumask_var(hctx->cpumask);
3703 static void blk_mq_init_cpu_queues(struct request_queue *q,
3704 unsigned int nr_hw_queues)
3706 struct blk_mq_tag_set *set = q->tag_set;
3709 for_each_possible_cpu(i) {
3710 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3711 struct blk_mq_hw_ctx *hctx;
3715 spin_lock_init(&__ctx->lock);
3716 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3717 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3722 * Set local node, IFF we have more than one hw queue. If
3723 * not, we remain on the home node of the device
3725 for (j = 0; j < set->nr_maps; j++) {
3726 hctx = blk_mq_map_queue_type(q, j, i);
3727 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3728 hctx->numa_node = cpu_to_node(i);
3733 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3734 unsigned int hctx_idx,
3737 struct blk_mq_tags *tags;
3740 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3744 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3746 blk_mq_free_rq_map(tags);
3753 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3756 if (blk_mq_is_shared_tags(set->flags)) {
3757 set->tags[hctx_idx] = set->shared_tags;
3762 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3765 return set->tags[hctx_idx];
3768 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3769 struct blk_mq_tags *tags,
3770 unsigned int hctx_idx)
3773 blk_mq_free_rqs(set, tags, hctx_idx);
3774 blk_mq_free_rq_map(tags);
3778 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3779 unsigned int hctx_idx)
3781 if (!blk_mq_is_shared_tags(set->flags))
3782 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3784 set->tags[hctx_idx] = NULL;
3787 static void blk_mq_map_swqueue(struct request_queue *q)
3789 unsigned int j, hctx_idx;
3791 struct blk_mq_hw_ctx *hctx;
3792 struct blk_mq_ctx *ctx;
3793 struct blk_mq_tag_set *set = q->tag_set;
3795 queue_for_each_hw_ctx(q, hctx, i) {
3796 cpumask_clear(hctx->cpumask);
3798 hctx->dispatch_from = NULL;
3802 * Map software to hardware queues.
3804 * If the cpu isn't present, the cpu is mapped to first hctx.
3806 for_each_possible_cpu(i) {
3808 ctx = per_cpu_ptr(q->queue_ctx, i);
3809 for (j = 0; j < set->nr_maps; j++) {
3810 if (!set->map[j].nr_queues) {
3811 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3812 HCTX_TYPE_DEFAULT, i);
3815 hctx_idx = set->map[j].mq_map[i];
3816 /* unmapped hw queue can be remapped after CPU topo changed */
3817 if (!set->tags[hctx_idx] &&
3818 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3820 * If tags initialization fail for some hctx,
3821 * that hctx won't be brought online. In this
3822 * case, remap the current ctx to hctx[0] which
3823 * is guaranteed to always have tags allocated
3825 set->map[j].mq_map[i] = 0;
3828 hctx = blk_mq_map_queue_type(q, j, i);
3829 ctx->hctxs[j] = hctx;
3831 * If the CPU is already set in the mask, then we've
3832 * mapped this one already. This can happen if
3833 * devices share queues across queue maps.
3835 if (cpumask_test_cpu(i, hctx->cpumask))
3838 cpumask_set_cpu(i, hctx->cpumask);
3840 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3841 hctx->ctxs[hctx->nr_ctx++] = ctx;
3844 * If the nr_ctx type overflows, we have exceeded the
3845 * amount of sw queues we can support.
3847 BUG_ON(!hctx->nr_ctx);
3850 for (; j < HCTX_MAX_TYPES; j++)
3851 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3852 HCTX_TYPE_DEFAULT, i);
3855 queue_for_each_hw_ctx(q, hctx, i) {
3857 * If no software queues are mapped to this hardware queue,
3858 * disable it and free the request entries.
3860 if (!hctx->nr_ctx) {
3861 /* Never unmap queue 0. We need it as a
3862 * fallback in case of a new remap fails
3866 __blk_mq_free_map_and_rqs(set, i);
3872 hctx->tags = set->tags[i];
3873 WARN_ON(!hctx->tags);
3876 * Set the map size to the number of mapped software queues.
3877 * This is more accurate and more efficient than looping
3878 * over all possibly mapped software queues.
3880 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3883 * Initialize batch roundrobin counts
3885 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3886 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3891 * Caller needs to ensure that we're either frozen/quiesced, or that
3892 * the queue isn't live yet.
3894 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3896 struct blk_mq_hw_ctx *hctx;
3899 queue_for_each_hw_ctx(q, hctx, i) {
3901 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3903 blk_mq_tag_idle(hctx);
3904 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3909 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3912 struct request_queue *q;
3914 lockdep_assert_held(&set->tag_list_lock);
3916 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3917 blk_mq_freeze_queue(q);
3918 queue_set_hctx_shared(q, shared);
3919 blk_mq_unfreeze_queue(q);
3923 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3925 struct blk_mq_tag_set *set = q->tag_set;
3927 mutex_lock(&set->tag_list_lock);
3928 list_del(&q->tag_set_list);
3929 if (list_is_singular(&set->tag_list)) {
3930 /* just transitioned to unshared */
3931 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3932 /* update existing queue */
3933 blk_mq_update_tag_set_shared(set, false);
3935 mutex_unlock(&set->tag_list_lock);
3936 INIT_LIST_HEAD(&q->tag_set_list);
3939 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3940 struct request_queue *q)
3942 mutex_lock(&set->tag_list_lock);
3945 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3947 if (!list_empty(&set->tag_list) &&
3948 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3949 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3950 /* update existing queue */
3951 blk_mq_update_tag_set_shared(set, true);
3953 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3954 queue_set_hctx_shared(q, true);
3955 list_add_tail(&q->tag_set_list, &set->tag_list);
3957 mutex_unlock(&set->tag_list_lock);
3960 /* All allocations will be freed in release handler of q->mq_kobj */
3961 static int blk_mq_alloc_ctxs(struct request_queue *q)
3963 struct blk_mq_ctxs *ctxs;
3966 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3970 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3971 if (!ctxs->queue_ctx)
3974 for_each_possible_cpu(cpu) {
3975 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3979 q->mq_kobj = &ctxs->kobj;
3980 q->queue_ctx = ctxs->queue_ctx;
3989 * It is the actual release handler for mq, but we do it from
3990 * request queue's release handler for avoiding use-after-free
3991 * and headache because q->mq_kobj shouldn't have been introduced,
3992 * but we can't group ctx/kctx kobj without it.
3994 void blk_mq_release(struct request_queue *q)
3996 struct blk_mq_hw_ctx *hctx, *next;
3999 queue_for_each_hw_ctx(q, hctx, i)
4000 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4002 /* all hctx are in .unused_hctx_list now */
4003 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4004 list_del_init(&hctx->hctx_list);
4005 kobject_put(&hctx->kobj);
4008 xa_destroy(&q->hctx_table);
4011 * release .mq_kobj and sw queue's kobject now because
4012 * both share lifetime with request queue.
4014 blk_mq_sysfs_deinit(q);
4017 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4020 struct request_queue *q;
4023 q = blk_alloc_queue(set->numa_node, set->flags & BLK_MQ_F_BLOCKING);
4025 return ERR_PTR(-ENOMEM);
4026 q->queuedata = queuedata;
4027 ret = blk_mq_init_allocated_queue(set, q);
4030 return ERR_PTR(ret);
4035 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4037 return blk_mq_init_queue_data(set, NULL);
4039 EXPORT_SYMBOL(blk_mq_init_queue);
4042 * blk_mq_destroy_queue - shutdown a request queue
4043 * @q: request queue to shutdown
4045 * This shuts down a request queue allocated by blk_mq_init_queue() and drops
4046 * the initial reference. All future requests will failed with -ENODEV.
4048 * Context: can sleep
4050 void blk_mq_destroy_queue(struct request_queue *q)
4052 WARN_ON_ONCE(!queue_is_mq(q));
4053 WARN_ON_ONCE(blk_queue_registered(q));
4057 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4058 blk_queue_start_drain(q);
4059 blk_freeze_queue(q);
4062 blk_mq_cancel_work_sync(q);
4063 blk_mq_exit_queue(q);
4065 /* @q is and will stay empty, shutdown and put */
4068 EXPORT_SYMBOL(blk_mq_destroy_queue);
4070 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4071 struct lock_class_key *lkclass)
4073 struct request_queue *q;
4074 struct gendisk *disk;
4076 q = blk_mq_init_queue_data(set, queuedata);
4080 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4082 blk_mq_destroy_queue(q);
4083 return ERR_PTR(-ENOMEM);
4085 set_bit(GD_OWNS_QUEUE, &disk->state);
4088 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4090 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4091 struct lock_class_key *lkclass)
4093 struct gendisk *disk;
4095 if (!blk_get_queue(q))
4097 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4102 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4104 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4105 struct blk_mq_tag_set *set, struct request_queue *q,
4106 int hctx_idx, int node)
4108 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4110 /* reuse dead hctx first */
4111 spin_lock(&q->unused_hctx_lock);
4112 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4113 if (tmp->numa_node == node) {
4119 list_del_init(&hctx->hctx_list);
4120 spin_unlock(&q->unused_hctx_lock);
4123 hctx = blk_mq_alloc_hctx(q, set, node);
4127 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4133 kobject_put(&hctx->kobj);
4138 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4139 struct request_queue *q)
4141 struct blk_mq_hw_ctx *hctx;
4144 /* protect against switching io scheduler */
4145 mutex_lock(&q->sysfs_lock);
4146 for (i = 0; i < set->nr_hw_queues; i++) {
4148 int node = blk_mq_get_hctx_node(set, i);
4149 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4152 old_node = old_hctx->numa_node;
4153 blk_mq_exit_hctx(q, set, old_hctx, i);
4156 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4159 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4161 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4162 WARN_ON_ONCE(!hctx);
4166 * Increasing nr_hw_queues fails. Free the newly allocated
4167 * hctxs and keep the previous q->nr_hw_queues.
4169 if (i != set->nr_hw_queues) {
4170 j = q->nr_hw_queues;
4173 q->nr_hw_queues = set->nr_hw_queues;
4176 xa_for_each_start(&q->hctx_table, j, hctx, j)
4177 blk_mq_exit_hctx(q, set, hctx, j);
4178 mutex_unlock(&q->sysfs_lock);
4181 static void blk_mq_update_poll_flag(struct request_queue *q)
4183 struct blk_mq_tag_set *set = q->tag_set;
4185 if (set->nr_maps > HCTX_TYPE_POLL &&
4186 set->map[HCTX_TYPE_POLL].nr_queues)
4187 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4189 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4192 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4193 struct request_queue *q)
4195 WARN_ON_ONCE(blk_queue_has_srcu(q) !=
4196 !!(set->flags & BLK_MQ_F_BLOCKING));
4198 /* mark the queue as mq asap */
4199 q->mq_ops = set->ops;
4201 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
4202 blk_mq_poll_stats_bkt,
4203 BLK_MQ_POLL_STATS_BKTS, q);
4207 if (blk_mq_alloc_ctxs(q))
4210 /* init q->mq_kobj and sw queues' kobjects */
4211 blk_mq_sysfs_init(q);
4213 INIT_LIST_HEAD(&q->unused_hctx_list);
4214 spin_lock_init(&q->unused_hctx_lock);
4216 xa_init(&q->hctx_table);
4218 blk_mq_realloc_hw_ctxs(set, q);
4219 if (!q->nr_hw_queues)
4222 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4223 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4227 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4228 blk_mq_update_poll_flag(q);
4230 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4231 INIT_LIST_HEAD(&q->requeue_list);
4232 spin_lock_init(&q->requeue_lock);
4234 q->nr_requests = set->queue_depth;
4237 * Default to classic polling
4239 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
4241 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4242 blk_mq_add_queue_tag_set(set, q);
4243 blk_mq_map_swqueue(q);
4249 blk_stat_free_callback(q->poll_cb);
4255 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4257 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4258 void blk_mq_exit_queue(struct request_queue *q)
4260 struct blk_mq_tag_set *set = q->tag_set;
4262 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4263 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4264 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4265 blk_mq_del_queue_tag_set(q);
4268 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4272 if (blk_mq_is_shared_tags(set->flags)) {
4273 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4276 if (!set->shared_tags)
4280 for (i = 0; i < set->nr_hw_queues; i++) {
4281 if (!__blk_mq_alloc_map_and_rqs(set, i))
4290 __blk_mq_free_map_and_rqs(set, i);
4292 if (blk_mq_is_shared_tags(set->flags)) {
4293 blk_mq_free_map_and_rqs(set, set->shared_tags,
4294 BLK_MQ_NO_HCTX_IDX);
4301 * Allocate the request maps associated with this tag_set. Note that this
4302 * may reduce the depth asked for, if memory is tight. set->queue_depth
4303 * will be updated to reflect the allocated depth.
4305 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4310 depth = set->queue_depth;
4312 err = __blk_mq_alloc_rq_maps(set);
4316 set->queue_depth >>= 1;
4317 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4321 } while (set->queue_depth);
4323 if (!set->queue_depth || err) {
4324 pr_err("blk-mq: failed to allocate request map\n");
4328 if (depth != set->queue_depth)
4329 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4330 depth, set->queue_depth);
4335 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4338 * blk_mq_map_queues() and multiple .map_queues() implementations
4339 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4340 * number of hardware queues.
4342 if (set->nr_maps == 1)
4343 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4345 if (set->ops->map_queues && !is_kdump_kernel()) {
4349 * transport .map_queues is usually done in the following
4352 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4353 * mask = get_cpu_mask(queue)
4354 * for_each_cpu(cpu, mask)
4355 * set->map[x].mq_map[cpu] = queue;
4358 * When we need to remap, the table has to be cleared for
4359 * killing stale mapping since one CPU may not be mapped
4362 for (i = 0; i < set->nr_maps; i++)
4363 blk_mq_clear_mq_map(&set->map[i]);
4365 set->ops->map_queues(set);
4367 BUG_ON(set->nr_maps > 1);
4368 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4372 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4373 int cur_nr_hw_queues, int new_nr_hw_queues)
4375 struct blk_mq_tags **new_tags;
4377 if (cur_nr_hw_queues >= new_nr_hw_queues)
4380 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4381 GFP_KERNEL, set->numa_node);
4386 memcpy(new_tags, set->tags, cur_nr_hw_queues *
4387 sizeof(*set->tags));
4389 set->tags = new_tags;
4390 set->nr_hw_queues = new_nr_hw_queues;
4395 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
4396 int new_nr_hw_queues)
4398 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
4402 * Alloc a tag set to be associated with one or more request queues.
4403 * May fail with EINVAL for various error conditions. May adjust the
4404 * requested depth down, if it's too large. In that case, the set
4405 * value will be stored in set->queue_depth.
4407 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4411 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4413 if (!set->nr_hw_queues)
4415 if (!set->queue_depth)
4417 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4420 if (!set->ops->queue_rq)
4423 if (!set->ops->get_budget ^ !set->ops->put_budget)
4426 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4427 pr_info("blk-mq: reduced tag depth to %u\n",
4429 set->queue_depth = BLK_MQ_MAX_DEPTH;
4434 else if (set->nr_maps > HCTX_MAX_TYPES)
4438 * If a crashdump is active, then we are potentially in a very
4439 * memory constrained environment. Limit us to 1 queue and
4440 * 64 tags to prevent using too much memory.
4442 if (is_kdump_kernel()) {
4443 set->nr_hw_queues = 1;
4445 set->queue_depth = min(64U, set->queue_depth);
4448 * There is no use for more h/w queues than cpus if we just have
4451 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4452 set->nr_hw_queues = nr_cpu_ids;
4454 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
4458 for (i = 0; i < set->nr_maps; i++) {
4459 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4460 sizeof(set->map[i].mq_map[0]),
4461 GFP_KERNEL, set->numa_node);
4462 if (!set->map[i].mq_map)
4463 goto out_free_mq_map;
4464 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4467 blk_mq_update_queue_map(set);
4469 ret = blk_mq_alloc_set_map_and_rqs(set);
4471 goto out_free_mq_map;
4473 mutex_init(&set->tag_list_lock);
4474 INIT_LIST_HEAD(&set->tag_list);
4479 for (i = 0; i < set->nr_maps; i++) {
4480 kfree(set->map[i].mq_map);
4481 set->map[i].mq_map = NULL;
4487 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4489 /* allocate and initialize a tagset for a simple single-queue device */
4490 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4491 const struct blk_mq_ops *ops, unsigned int queue_depth,
4492 unsigned int set_flags)
4494 memset(set, 0, sizeof(*set));
4496 set->nr_hw_queues = 1;
4498 set->queue_depth = queue_depth;
4499 set->numa_node = NUMA_NO_NODE;
4500 set->flags = set_flags;
4501 return blk_mq_alloc_tag_set(set);
4503 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4505 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4509 for (i = 0; i < set->nr_hw_queues; i++)
4510 __blk_mq_free_map_and_rqs(set, i);
4512 if (blk_mq_is_shared_tags(set->flags)) {
4513 blk_mq_free_map_and_rqs(set, set->shared_tags,
4514 BLK_MQ_NO_HCTX_IDX);
4517 for (j = 0; j < set->nr_maps; j++) {
4518 kfree(set->map[j].mq_map);
4519 set->map[j].mq_map = NULL;
4525 EXPORT_SYMBOL(blk_mq_free_tag_set);
4527 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4529 struct blk_mq_tag_set *set = q->tag_set;
4530 struct blk_mq_hw_ctx *hctx;
4537 if (q->nr_requests == nr)
4540 blk_mq_freeze_queue(q);
4541 blk_mq_quiesce_queue(q);
4544 queue_for_each_hw_ctx(q, hctx, i) {
4548 * If we're using an MQ scheduler, just update the scheduler
4549 * queue depth. This is similar to what the old code would do.
4551 if (hctx->sched_tags) {
4552 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4555 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4560 if (q->elevator && q->elevator->type->ops.depth_updated)
4561 q->elevator->type->ops.depth_updated(hctx);
4564 q->nr_requests = nr;
4565 if (blk_mq_is_shared_tags(set->flags)) {
4567 blk_mq_tag_update_sched_shared_tags(q);
4569 blk_mq_tag_resize_shared_tags(set, nr);
4573 blk_mq_unquiesce_queue(q);
4574 blk_mq_unfreeze_queue(q);
4580 * request_queue and elevator_type pair.
4581 * It is just used by __blk_mq_update_nr_hw_queues to cache
4582 * the elevator_type associated with a request_queue.
4584 struct blk_mq_qe_pair {
4585 struct list_head node;
4586 struct request_queue *q;
4587 struct elevator_type *type;
4591 * Cache the elevator_type in qe pair list and switch the
4592 * io scheduler to 'none'
4594 static bool blk_mq_elv_switch_none(struct list_head *head,
4595 struct request_queue *q)
4597 struct blk_mq_qe_pair *qe;
4602 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4606 /* q->elevator needs protection from ->sysfs_lock */
4607 mutex_lock(&q->sysfs_lock);
4609 INIT_LIST_HEAD(&qe->node);
4611 qe->type = q->elevator->type;
4612 list_add(&qe->node, head);
4615 * After elevator_switch, the previous elevator_queue will be
4616 * released by elevator_release. The reference of the io scheduler
4617 * module get by elevator_get will also be put. So we need to get
4618 * a reference of the io scheduler module here to prevent it to be
4621 __module_get(qe->type->elevator_owner);
4622 elevator_switch(q, NULL);
4623 mutex_unlock(&q->sysfs_lock);
4628 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4629 struct request_queue *q)
4631 struct blk_mq_qe_pair *qe;
4633 list_for_each_entry(qe, head, node)
4640 static void blk_mq_elv_switch_back(struct list_head *head,
4641 struct request_queue *q)
4643 struct blk_mq_qe_pair *qe;
4644 struct elevator_type *t;
4646 qe = blk_lookup_qe_pair(head, q);
4650 list_del(&qe->node);
4653 mutex_lock(&q->sysfs_lock);
4654 elevator_switch(q, t);
4655 mutex_unlock(&q->sysfs_lock);
4658 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4661 struct request_queue *q;
4663 int prev_nr_hw_queues;
4665 lockdep_assert_held(&set->tag_list_lock);
4667 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4668 nr_hw_queues = nr_cpu_ids;
4669 if (nr_hw_queues < 1)
4671 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4674 list_for_each_entry(q, &set->tag_list, tag_set_list)
4675 blk_mq_freeze_queue(q);
4677 * Switch IO scheduler to 'none', cleaning up the data associated
4678 * with the previous scheduler. We will switch back once we are done
4679 * updating the new sw to hw queue mappings.
4681 list_for_each_entry(q, &set->tag_list, tag_set_list)
4682 if (!blk_mq_elv_switch_none(&head, q))
4685 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4686 blk_mq_debugfs_unregister_hctxs(q);
4687 blk_mq_sysfs_unregister_hctxs(q);
4690 prev_nr_hw_queues = set->nr_hw_queues;
4691 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
4695 set->nr_hw_queues = nr_hw_queues;
4697 blk_mq_update_queue_map(set);
4698 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4699 blk_mq_realloc_hw_ctxs(set, q);
4700 blk_mq_update_poll_flag(q);
4701 if (q->nr_hw_queues != set->nr_hw_queues) {
4702 int i = prev_nr_hw_queues;
4704 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4705 nr_hw_queues, prev_nr_hw_queues);
4706 for (; i < set->nr_hw_queues; i++)
4707 __blk_mq_free_map_and_rqs(set, i);
4709 set->nr_hw_queues = prev_nr_hw_queues;
4710 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4713 blk_mq_map_swqueue(q);
4717 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4718 blk_mq_sysfs_register_hctxs(q);
4719 blk_mq_debugfs_register_hctxs(q);
4723 list_for_each_entry(q, &set->tag_list, tag_set_list)
4724 blk_mq_elv_switch_back(&head, q);
4726 list_for_each_entry(q, &set->tag_list, tag_set_list)
4727 blk_mq_unfreeze_queue(q);
4730 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4732 mutex_lock(&set->tag_list_lock);
4733 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4734 mutex_unlock(&set->tag_list_lock);
4736 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4738 /* Enable polling stats and return whether they were already enabled. */
4739 static bool blk_poll_stats_enable(struct request_queue *q)
4744 return blk_stats_alloc_enable(q);
4747 static void blk_mq_poll_stats_start(struct request_queue *q)
4750 * We don't arm the callback if polling stats are not enabled or the
4751 * callback is already active.
4753 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
4756 blk_stat_activate_msecs(q->poll_cb, 100);
4759 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4761 struct request_queue *q = cb->data;
4764 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4765 if (cb->stat[bucket].nr_samples)
4766 q->poll_stat[bucket] = cb->stat[bucket];
4770 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4773 unsigned long ret = 0;
4777 * If stats collection isn't on, don't sleep but turn it on for
4780 if (!blk_poll_stats_enable(q))
4784 * As an optimistic guess, use half of the mean service time
4785 * for this type of request. We can (and should) make this smarter.
4786 * For instance, if the completion latencies are tight, we can
4787 * get closer than just half the mean. This is especially
4788 * important on devices where the completion latencies are longer
4789 * than ~10 usec. We do use the stats for the relevant IO size
4790 * if available which does lead to better estimates.
4792 bucket = blk_mq_poll_stats_bkt(rq);
4796 if (q->poll_stat[bucket].nr_samples)
4797 ret = (q->poll_stat[bucket].mean + 1) / 2;
4802 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4804 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4805 struct request *rq = blk_qc_to_rq(hctx, qc);
4806 struct hrtimer_sleeper hs;
4807 enum hrtimer_mode mode;
4812 * If a request has completed on queue that uses an I/O scheduler, we
4813 * won't get back a request from blk_qc_to_rq.
4815 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4819 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4821 * 0: use half of prev avg
4822 * >0: use this specific value
4824 if (q->poll_nsec > 0)
4825 nsecs = q->poll_nsec;
4827 nsecs = blk_mq_poll_nsecs(q, rq);
4832 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4835 * This will be replaced with the stats tracking code, using
4836 * 'avg_completion_time / 2' as the pre-sleep target.
4840 mode = HRTIMER_MODE_REL;
4841 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4842 hrtimer_set_expires(&hs.timer, kt);
4845 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4847 set_current_state(TASK_UNINTERRUPTIBLE);
4848 hrtimer_sleeper_start_expires(&hs, mode);
4851 hrtimer_cancel(&hs.timer);
4852 mode = HRTIMER_MODE_ABS;
4853 } while (hs.task && !signal_pending(current));
4855 __set_current_state(TASK_RUNNING);
4856 destroy_hrtimer_on_stack(&hs.timer);
4859 * If we sleep, have the caller restart the poll loop to reset the
4860 * state. Like for the other success return cases, the caller is
4861 * responsible for checking if the IO completed. If the IO isn't
4862 * complete, we'll get called again and will go straight to the busy
4868 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4869 struct io_comp_batch *iob, unsigned int flags)
4871 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4872 long state = get_current_state();
4876 ret = q->mq_ops->poll(hctx, iob);
4878 __set_current_state(TASK_RUNNING);
4882 if (signal_pending_state(state, current))
4883 __set_current_state(TASK_RUNNING);
4884 if (task_is_running(current))
4887 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4890 } while (!need_resched());
4892 __set_current_state(TASK_RUNNING);
4896 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4899 if (!(flags & BLK_POLL_NOSLEEP) &&
4900 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4901 if (blk_mq_poll_hybrid(q, cookie))
4904 return blk_mq_poll_classic(q, cookie, iob, flags);
4907 unsigned int blk_mq_rq_cpu(struct request *rq)
4909 return rq->mq_ctx->cpu;
4911 EXPORT_SYMBOL(blk_mq_rq_cpu);
4913 void blk_mq_cancel_work_sync(struct request_queue *q)
4915 if (queue_is_mq(q)) {
4916 struct blk_mq_hw_ctx *hctx;
4919 cancel_delayed_work_sync(&q->requeue_work);
4921 queue_for_each_hw_ctx(q, hctx, i)
4922 cancel_delayed_work_sync(&hctx->run_work);
4926 static int __init blk_mq_init(void)
4930 for_each_possible_cpu(i)
4931 init_llist_head(&per_cpu(blk_cpu_done, i));
4932 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4934 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4935 "block/softirq:dead", NULL,
4936 blk_softirq_cpu_dead);
4937 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4938 blk_mq_hctx_notify_dead);
4939 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4940 blk_mq_hctx_notify_online,
4941 blk_mq_hctx_notify_offline);
4944 subsys_initcall(blk_mq_init);