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);
812 * Upper layers may call blk_crypto_evict_key() anytime after the last
813 * bio_endio(). Therefore, the keyslot must be released before that.
815 blk_crypto_rq_put_keyslot(req);
817 blk_account_io_completion(req, total_bytes);
820 struct bio *next = bio->bi_next;
822 /* Completion has already been traced */
823 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
825 if (req_op(req) == REQ_OP_ZONE_APPEND)
826 bio->bi_iter.bi_sector = req->__sector;
834 * Reset counters so that the request stacking driver
835 * can find how many bytes remain in the request
845 * blk_update_request - Complete multiple bytes without completing the request
846 * @req: the request being processed
847 * @error: block status code
848 * @nr_bytes: number of bytes to complete for @req
851 * Ends I/O on a number of bytes attached to @req, but doesn't complete
852 * the request structure even if @req doesn't have leftover.
853 * If @req has leftover, sets it up for the next range of segments.
855 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
856 * %false return from this function.
859 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
860 * except in the consistency check at the end of this function.
863 * %false - this request doesn't have any more data
864 * %true - this request has more data
866 bool blk_update_request(struct request *req, blk_status_t error,
867 unsigned int nr_bytes)
871 trace_block_rq_complete(req, error, nr_bytes);
876 #ifdef CONFIG_BLK_DEV_INTEGRITY
877 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
879 req->q->integrity.profile->complete_fn(req, nr_bytes);
883 * Upper layers may call blk_crypto_evict_key() anytime after the last
884 * bio_endio(). Therefore, the keyslot must be released before that.
886 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
887 __blk_crypto_rq_put_keyslot(req);
889 if (unlikely(error && !blk_rq_is_passthrough(req) &&
890 !(req->rq_flags & RQF_QUIET)) &&
891 !test_bit(GD_DEAD, &req->q->disk->state)) {
892 blk_print_req_error(req, error);
893 trace_block_rq_error(req, error, nr_bytes);
896 blk_account_io_completion(req, nr_bytes);
900 struct bio *bio = req->bio;
901 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
903 if (bio_bytes == bio->bi_iter.bi_size)
904 req->bio = bio->bi_next;
906 /* Completion has already been traced */
907 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
908 req_bio_endio(req, bio, bio_bytes, error);
910 total_bytes += bio_bytes;
911 nr_bytes -= bio_bytes;
922 * Reset counters so that the request stacking driver
923 * can find how many bytes remain in the request
930 req->__data_len -= total_bytes;
932 /* update sector only for requests with clear definition of sector */
933 if (!blk_rq_is_passthrough(req))
934 req->__sector += total_bytes >> 9;
936 /* mixed attributes always follow the first bio */
937 if (req->rq_flags & RQF_MIXED_MERGE) {
938 req->cmd_flags &= ~REQ_FAILFAST_MASK;
939 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
942 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
944 * If total number of sectors is less than the first segment
945 * size, something has gone terribly wrong.
947 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
948 blk_dump_rq_flags(req, "request botched");
949 req->__data_len = blk_rq_cur_bytes(req);
952 /* recalculate the number of segments */
953 req->nr_phys_segments = blk_recalc_rq_segments(req);
958 EXPORT_SYMBOL_GPL(blk_update_request);
960 static void __blk_account_io_done(struct request *req, u64 now)
962 const int sgrp = op_stat_group(req_op(req));
965 update_io_ticks(req->part, jiffies, true);
966 part_stat_inc(req->part, ios[sgrp]);
967 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
971 static inline void blk_account_io_done(struct request *req, u64 now)
974 * Account IO completion. flush_rq isn't accounted as a
975 * normal IO on queueing nor completion. Accounting the
976 * containing request is enough.
978 if (blk_do_io_stat(req) && req->part &&
979 !(req->rq_flags & RQF_FLUSH_SEQ))
980 __blk_account_io_done(req, now);
983 static void __blk_account_io_start(struct request *rq)
986 * All non-passthrough requests are created from a bio with one
987 * exception: when a flush command that is part of a flush sequence
988 * generated by the state machine in blk-flush.c is cloned onto the
989 * lower device by dm-multipath we can get here without a bio.
992 rq->part = rq->bio->bi_bdev;
994 rq->part = rq->q->disk->part0;
997 update_io_ticks(rq->part, jiffies, false);
1001 static inline void blk_account_io_start(struct request *req)
1003 if (blk_do_io_stat(req))
1004 __blk_account_io_start(req);
1007 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1009 if (rq->rq_flags & RQF_STATS) {
1010 blk_mq_poll_stats_start(rq->q);
1011 blk_stat_add(rq, now);
1014 blk_mq_sched_completed_request(rq, now);
1015 blk_account_io_done(rq, now);
1018 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1020 if (blk_mq_need_time_stamp(rq))
1021 __blk_mq_end_request_acct(rq, ktime_get_ns());
1024 rq_qos_done(rq->q, rq);
1025 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1026 blk_mq_free_request(rq);
1028 blk_mq_free_request(rq);
1031 EXPORT_SYMBOL(__blk_mq_end_request);
1033 void blk_mq_end_request(struct request *rq, blk_status_t error)
1035 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1037 __blk_mq_end_request(rq, error);
1039 EXPORT_SYMBOL(blk_mq_end_request);
1041 #define TAG_COMP_BATCH 32
1043 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1044 int *tag_array, int nr_tags)
1046 struct request_queue *q = hctx->queue;
1049 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1050 * update hctx->nr_active in batch
1052 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1053 __blk_mq_sub_active_requests(hctx, nr_tags);
1055 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1056 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1059 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1061 int tags[TAG_COMP_BATCH], nr_tags = 0;
1062 struct blk_mq_hw_ctx *cur_hctx = NULL;
1067 now = ktime_get_ns();
1069 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1071 prefetch(rq->rq_next);
1073 blk_complete_request(rq);
1075 __blk_mq_end_request_acct(rq, now);
1077 rq_qos_done(rq->q, rq);
1080 * If end_io handler returns NONE, then it still has
1081 * ownership of the request.
1083 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1086 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1087 if (!req_ref_put_and_test(rq))
1090 blk_crypto_free_request(rq);
1091 blk_pm_mark_last_busy(rq);
1093 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1095 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1097 cur_hctx = rq->mq_hctx;
1099 tags[nr_tags++] = rq->tag;
1103 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1105 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1107 static void blk_complete_reqs(struct llist_head *list)
1109 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1110 struct request *rq, *next;
1112 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1113 rq->q->mq_ops->complete(rq);
1116 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1118 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1121 static int blk_softirq_cpu_dead(unsigned int cpu)
1123 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1127 static void __blk_mq_complete_request_remote(void *data)
1129 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1132 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1134 int cpu = raw_smp_processor_id();
1136 if (!IS_ENABLED(CONFIG_SMP) ||
1137 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1140 * With force threaded interrupts enabled, raising softirq from an SMP
1141 * function call will always result in waking the ksoftirqd thread.
1142 * This is probably worse than completing the request on a different
1145 if (force_irqthreads())
1148 /* same CPU or cache domain? Complete locally */
1149 if (cpu == rq->mq_ctx->cpu ||
1150 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1151 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1154 /* don't try to IPI to an offline CPU */
1155 return cpu_online(rq->mq_ctx->cpu);
1158 static void blk_mq_complete_send_ipi(struct request *rq)
1160 struct llist_head *list;
1163 cpu = rq->mq_ctx->cpu;
1164 list = &per_cpu(blk_cpu_done, cpu);
1165 if (llist_add(&rq->ipi_list, list)) {
1166 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1167 smp_call_function_single_async(cpu, &rq->csd);
1171 static void blk_mq_raise_softirq(struct request *rq)
1173 struct llist_head *list;
1176 list = this_cpu_ptr(&blk_cpu_done);
1177 if (llist_add(&rq->ipi_list, list))
1178 raise_softirq(BLOCK_SOFTIRQ);
1182 bool blk_mq_complete_request_remote(struct request *rq)
1184 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1187 * For request which hctx has only one ctx mapping,
1188 * or a polled request, always complete locally,
1189 * it's pointless to redirect the completion.
1191 if (rq->mq_hctx->nr_ctx == 1 ||
1192 rq->cmd_flags & REQ_POLLED)
1195 if (blk_mq_complete_need_ipi(rq)) {
1196 blk_mq_complete_send_ipi(rq);
1200 if (rq->q->nr_hw_queues == 1) {
1201 blk_mq_raise_softirq(rq);
1206 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1209 * blk_mq_complete_request - end I/O on a request
1210 * @rq: the request being processed
1213 * Complete a request by scheduling the ->complete_rq operation.
1215 void blk_mq_complete_request(struct request *rq)
1217 if (!blk_mq_complete_request_remote(rq))
1218 rq->q->mq_ops->complete(rq);
1220 EXPORT_SYMBOL(blk_mq_complete_request);
1223 * blk_mq_start_request - Start processing a request
1224 * @rq: Pointer to request to be started
1226 * Function used by device drivers to notify the block layer that a request
1227 * is going to be processed now, so blk layer can do proper initializations
1228 * such as starting the timeout timer.
1230 void blk_mq_start_request(struct request *rq)
1232 struct request_queue *q = rq->q;
1234 trace_block_rq_issue(rq);
1236 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1237 rq->io_start_time_ns = ktime_get_ns();
1238 rq->stats_sectors = blk_rq_sectors(rq);
1239 rq->rq_flags |= RQF_STATS;
1240 rq_qos_issue(q, rq);
1243 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1246 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1248 #ifdef CONFIG_BLK_DEV_INTEGRITY
1249 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1250 q->integrity.profile->prepare_fn(rq);
1252 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1253 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1255 EXPORT_SYMBOL(blk_mq_start_request);
1258 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1259 * queues. This is important for md arrays to benefit from merging
1262 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1264 if (plug->multiple_queues)
1265 return BLK_MAX_REQUEST_COUNT * 2;
1266 return BLK_MAX_REQUEST_COUNT;
1269 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1271 struct request *last = rq_list_peek(&plug->mq_list);
1273 if (!plug->rq_count) {
1274 trace_block_plug(rq->q);
1275 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1276 (!blk_queue_nomerges(rq->q) &&
1277 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1278 blk_mq_flush_plug_list(plug, false);
1280 trace_block_plug(rq->q);
1283 if (!plug->multiple_queues && last && last->q != rq->q)
1284 plug->multiple_queues = true;
1285 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1286 plug->has_elevator = true;
1288 rq_list_add(&plug->mq_list, rq);
1293 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1294 * @rq: request to insert
1295 * @at_head: insert request at head or tail of queue
1298 * Insert a fully prepared request at the back of the I/O scheduler queue
1299 * for execution. Don't wait for completion.
1302 * This function will invoke @done directly if the queue is dead.
1304 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1306 WARN_ON(irqs_disabled());
1307 WARN_ON(!blk_rq_is_passthrough(rq));
1309 blk_account_io_start(rq);
1312 * As plugging can be enabled for passthrough requests on a zoned
1313 * device, directly accessing the plug instead of using blk_mq_plug()
1314 * should not have any consequences.
1316 if (current->plug && !at_head)
1317 blk_add_rq_to_plug(current->plug, rq);
1319 blk_mq_sched_insert_request(rq, at_head, true, false);
1321 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1323 struct blk_rq_wait {
1324 struct completion done;
1328 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1330 struct blk_rq_wait *wait = rq->end_io_data;
1333 complete(&wait->done);
1334 return RQ_END_IO_NONE;
1337 bool blk_rq_is_poll(struct request *rq)
1341 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1345 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1347 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1350 blk_mq_poll(rq->q, blk_rq_to_qc(rq), NULL, 0);
1352 } while (!completion_done(wait));
1356 * blk_execute_rq - insert a request into queue for execution
1357 * @rq: request to insert
1358 * @at_head: insert request at head or tail of queue
1361 * Insert a fully prepared request at the back of the I/O scheduler queue
1362 * for execution and wait for completion.
1363 * Return: The blk_status_t result provided to blk_mq_end_request().
1365 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1367 struct blk_rq_wait wait = {
1368 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1371 WARN_ON(irqs_disabled());
1372 WARN_ON(!blk_rq_is_passthrough(rq));
1374 rq->end_io_data = &wait;
1375 rq->end_io = blk_end_sync_rq;
1377 blk_account_io_start(rq);
1378 blk_mq_sched_insert_request(rq, at_head, true, false);
1380 if (blk_rq_is_poll(rq)) {
1381 blk_rq_poll_completion(rq, &wait.done);
1384 * Prevent hang_check timer from firing at us during very long
1387 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1390 while (!wait_for_completion_io_timeout(&wait.done,
1391 hang_check * (HZ/2)))
1394 wait_for_completion_io(&wait.done);
1399 EXPORT_SYMBOL(blk_execute_rq);
1401 static void __blk_mq_requeue_request(struct request *rq)
1403 struct request_queue *q = rq->q;
1405 blk_mq_put_driver_tag(rq);
1407 trace_block_rq_requeue(rq);
1408 rq_qos_requeue(q, rq);
1410 if (blk_mq_request_started(rq)) {
1411 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1412 rq->rq_flags &= ~RQF_TIMED_OUT;
1416 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1418 __blk_mq_requeue_request(rq);
1420 /* this request will be re-inserted to io scheduler queue */
1421 blk_mq_sched_requeue_request(rq);
1423 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1425 EXPORT_SYMBOL(blk_mq_requeue_request);
1427 static void blk_mq_requeue_work(struct work_struct *work)
1429 struct request_queue *q =
1430 container_of(work, struct request_queue, requeue_work.work);
1432 struct request *rq, *next;
1434 spin_lock_irq(&q->requeue_lock);
1435 list_splice_init(&q->requeue_list, &rq_list);
1436 spin_unlock_irq(&q->requeue_lock);
1438 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1439 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1442 rq->rq_flags &= ~RQF_SOFTBARRIER;
1443 list_del_init(&rq->queuelist);
1445 * If RQF_DONTPREP, rq has contained some driver specific
1446 * data, so insert it to hctx dispatch list to avoid any
1449 if (rq->rq_flags & RQF_DONTPREP)
1450 blk_mq_request_bypass_insert(rq, false, false);
1452 blk_mq_sched_insert_request(rq, true, false, false);
1455 while (!list_empty(&rq_list)) {
1456 rq = list_entry(rq_list.next, struct request, queuelist);
1457 list_del_init(&rq->queuelist);
1458 blk_mq_sched_insert_request(rq, false, false, false);
1461 blk_mq_run_hw_queues(q, false);
1464 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1465 bool kick_requeue_list)
1467 struct request_queue *q = rq->q;
1468 unsigned long flags;
1471 * We abuse this flag that is otherwise used by the I/O scheduler to
1472 * request head insertion from the workqueue.
1474 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1476 spin_lock_irqsave(&q->requeue_lock, flags);
1478 rq->rq_flags |= RQF_SOFTBARRIER;
1479 list_add(&rq->queuelist, &q->requeue_list);
1481 list_add_tail(&rq->queuelist, &q->requeue_list);
1483 spin_unlock_irqrestore(&q->requeue_lock, flags);
1485 if (kick_requeue_list)
1486 blk_mq_kick_requeue_list(q);
1489 void blk_mq_kick_requeue_list(struct request_queue *q)
1491 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1493 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1495 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1496 unsigned long msecs)
1498 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1499 msecs_to_jiffies(msecs));
1501 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1503 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1506 * If we find a request that isn't idle we know the queue is busy
1507 * as it's checked in the iter.
1508 * Return false to stop the iteration.
1510 if (blk_mq_request_started(rq)) {
1520 bool blk_mq_queue_inflight(struct request_queue *q)
1524 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1527 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1529 static void blk_mq_rq_timed_out(struct request *req)
1531 req->rq_flags |= RQF_TIMED_OUT;
1532 if (req->q->mq_ops->timeout) {
1533 enum blk_eh_timer_return ret;
1535 ret = req->q->mq_ops->timeout(req);
1536 if (ret == BLK_EH_DONE)
1538 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1544 struct blk_expired_data {
1545 bool has_timedout_rq;
1547 unsigned long timeout_start;
1550 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1552 unsigned long deadline;
1554 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1556 if (rq->rq_flags & RQF_TIMED_OUT)
1559 deadline = READ_ONCE(rq->deadline);
1560 if (time_after_eq(expired->timeout_start, deadline))
1563 if (expired->next == 0)
1564 expired->next = deadline;
1565 else if (time_after(expired->next, deadline))
1566 expired->next = deadline;
1570 void blk_mq_put_rq_ref(struct request *rq)
1572 if (is_flush_rq(rq)) {
1573 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1574 blk_mq_free_request(rq);
1575 } else if (req_ref_put_and_test(rq)) {
1576 __blk_mq_free_request(rq);
1580 static bool blk_mq_check_expired(struct request *rq, void *priv)
1582 struct blk_expired_data *expired = priv;
1585 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1586 * be reallocated underneath the timeout handler's processing, then
1587 * the expire check is reliable. If the request is not expired, then
1588 * it was completed and reallocated as a new request after returning
1589 * from blk_mq_check_expired().
1591 if (blk_mq_req_expired(rq, expired)) {
1592 expired->has_timedout_rq = true;
1598 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1600 struct blk_expired_data *expired = priv;
1602 if (blk_mq_req_expired(rq, expired))
1603 blk_mq_rq_timed_out(rq);
1607 static void blk_mq_timeout_work(struct work_struct *work)
1609 struct request_queue *q =
1610 container_of(work, struct request_queue, timeout_work);
1611 struct blk_expired_data expired = {
1612 .timeout_start = jiffies,
1614 struct blk_mq_hw_ctx *hctx;
1617 /* A deadlock might occur if a request is stuck requiring a
1618 * timeout at the same time a queue freeze is waiting
1619 * completion, since the timeout code would not be able to
1620 * acquire the queue reference here.
1622 * That's why we don't use blk_queue_enter here; instead, we use
1623 * percpu_ref_tryget directly, because we need to be able to
1624 * obtain a reference even in the short window between the queue
1625 * starting to freeze, by dropping the first reference in
1626 * blk_freeze_queue_start, and the moment the last request is
1627 * consumed, marked by the instant q_usage_counter reaches
1630 if (!percpu_ref_tryget(&q->q_usage_counter))
1633 /* check if there is any timed-out request */
1634 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1635 if (expired.has_timedout_rq) {
1637 * Before walking tags, we must ensure any submit started
1638 * before the current time has finished. Since the submit
1639 * uses srcu or rcu, wait for a synchronization point to
1640 * ensure all running submits have finished
1642 blk_mq_wait_quiesce_done(q);
1645 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1648 if (expired.next != 0) {
1649 mod_timer(&q->timeout, expired.next);
1652 * Request timeouts are handled as a forward rolling timer. If
1653 * we end up here it means that no requests are pending and
1654 * also that no request has been pending for a while. Mark
1655 * each hctx as idle.
1657 queue_for_each_hw_ctx(q, hctx, i) {
1658 /* the hctx may be unmapped, so check it here */
1659 if (blk_mq_hw_queue_mapped(hctx))
1660 blk_mq_tag_idle(hctx);
1666 struct flush_busy_ctx_data {
1667 struct blk_mq_hw_ctx *hctx;
1668 struct list_head *list;
1671 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1673 struct flush_busy_ctx_data *flush_data = data;
1674 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1675 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1676 enum hctx_type type = hctx->type;
1678 spin_lock(&ctx->lock);
1679 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1680 sbitmap_clear_bit(sb, bitnr);
1681 spin_unlock(&ctx->lock);
1686 * Process software queues that have been marked busy, splicing them
1687 * to the for-dispatch
1689 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1691 struct flush_busy_ctx_data data = {
1696 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1698 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1700 struct dispatch_rq_data {
1701 struct blk_mq_hw_ctx *hctx;
1705 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1708 struct dispatch_rq_data *dispatch_data = data;
1709 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1710 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1711 enum hctx_type type = hctx->type;
1713 spin_lock(&ctx->lock);
1714 if (!list_empty(&ctx->rq_lists[type])) {
1715 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1716 list_del_init(&dispatch_data->rq->queuelist);
1717 if (list_empty(&ctx->rq_lists[type]))
1718 sbitmap_clear_bit(sb, bitnr);
1720 spin_unlock(&ctx->lock);
1722 return !dispatch_data->rq;
1725 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1726 struct blk_mq_ctx *start)
1728 unsigned off = start ? start->index_hw[hctx->type] : 0;
1729 struct dispatch_rq_data data = {
1734 __sbitmap_for_each_set(&hctx->ctx_map, off,
1735 dispatch_rq_from_ctx, &data);
1740 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1742 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1743 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1746 blk_mq_tag_busy(rq->mq_hctx);
1748 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1749 bt = &rq->mq_hctx->tags->breserved_tags;
1752 if (!hctx_may_queue(rq->mq_hctx, bt))
1756 tag = __sbitmap_queue_get(bt);
1757 if (tag == BLK_MQ_NO_TAG)
1760 rq->tag = tag + tag_offset;
1764 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1766 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1769 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1770 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1771 rq->rq_flags |= RQF_MQ_INFLIGHT;
1772 __blk_mq_inc_active_requests(hctx);
1774 hctx->tags->rqs[rq->tag] = rq;
1778 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1779 int flags, void *key)
1781 struct blk_mq_hw_ctx *hctx;
1783 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1785 spin_lock(&hctx->dispatch_wait_lock);
1786 if (!list_empty(&wait->entry)) {
1787 struct sbitmap_queue *sbq;
1789 list_del_init(&wait->entry);
1790 sbq = &hctx->tags->bitmap_tags;
1791 atomic_dec(&sbq->ws_active);
1793 spin_unlock(&hctx->dispatch_wait_lock);
1795 blk_mq_run_hw_queue(hctx, true);
1800 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1801 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1802 * restart. For both cases, take care to check the condition again after
1803 * marking us as waiting.
1805 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1808 struct sbitmap_queue *sbq;
1809 struct wait_queue_head *wq;
1810 wait_queue_entry_t *wait;
1813 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1814 !(blk_mq_is_shared_tags(hctx->flags))) {
1815 blk_mq_sched_mark_restart_hctx(hctx);
1818 * It's possible that a tag was freed in the window between the
1819 * allocation failure and adding the hardware queue to the wait
1822 * Don't clear RESTART here, someone else could have set it.
1823 * At most this will cost an extra queue run.
1825 return blk_mq_get_driver_tag(rq);
1828 wait = &hctx->dispatch_wait;
1829 if (!list_empty_careful(&wait->entry))
1832 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1833 sbq = &hctx->tags->breserved_tags;
1835 sbq = &hctx->tags->bitmap_tags;
1836 wq = &bt_wait_ptr(sbq, hctx)->wait;
1838 spin_lock_irq(&wq->lock);
1839 spin_lock(&hctx->dispatch_wait_lock);
1840 if (!list_empty(&wait->entry)) {
1841 spin_unlock(&hctx->dispatch_wait_lock);
1842 spin_unlock_irq(&wq->lock);
1846 atomic_inc(&sbq->ws_active);
1847 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1848 __add_wait_queue(wq, wait);
1851 * It's possible that a tag was freed in the window between the
1852 * allocation failure and adding the hardware queue to the wait
1855 ret = blk_mq_get_driver_tag(rq);
1857 spin_unlock(&hctx->dispatch_wait_lock);
1858 spin_unlock_irq(&wq->lock);
1863 * We got a tag, remove ourselves from the wait queue to ensure
1864 * someone else gets the wakeup.
1866 list_del_init(&wait->entry);
1867 atomic_dec(&sbq->ws_active);
1868 spin_unlock(&hctx->dispatch_wait_lock);
1869 spin_unlock_irq(&wq->lock);
1874 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1875 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1877 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1878 * - EWMA is one simple way to compute running average value
1879 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1880 * - take 4 as factor for avoiding to get too small(0) result, and this
1881 * factor doesn't matter because EWMA decreases exponentially
1883 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1887 ewma = hctx->dispatch_busy;
1892 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1894 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1895 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1897 hctx->dispatch_busy = ewma;
1900 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1902 static void blk_mq_handle_dev_resource(struct request *rq,
1903 struct list_head *list)
1905 struct request *next =
1906 list_first_entry_or_null(list, struct request, queuelist);
1909 * If an I/O scheduler has been configured and we got a driver tag for
1910 * the next request already, free it.
1913 blk_mq_put_driver_tag(next);
1915 list_add(&rq->queuelist, list);
1916 __blk_mq_requeue_request(rq);
1919 static void blk_mq_handle_zone_resource(struct request *rq,
1920 struct list_head *zone_list)
1923 * If we end up here it is because we cannot dispatch a request to a
1924 * specific zone due to LLD level zone-write locking or other zone
1925 * related resource not being available. In this case, set the request
1926 * aside in zone_list for retrying it later.
1928 list_add(&rq->queuelist, zone_list);
1929 __blk_mq_requeue_request(rq);
1932 enum prep_dispatch {
1934 PREP_DISPATCH_NO_TAG,
1935 PREP_DISPATCH_NO_BUDGET,
1938 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1941 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1942 int budget_token = -1;
1945 budget_token = blk_mq_get_dispatch_budget(rq->q);
1946 if (budget_token < 0) {
1947 blk_mq_put_driver_tag(rq);
1948 return PREP_DISPATCH_NO_BUDGET;
1950 blk_mq_set_rq_budget_token(rq, budget_token);
1953 if (!blk_mq_get_driver_tag(rq)) {
1955 * The initial allocation attempt failed, so we need to
1956 * rerun the hardware queue when a tag is freed. The
1957 * waitqueue takes care of that. If the queue is run
1958 * before we add this entry back on the dispatch list,
1959 * we'll re-run it below.
1961 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1963 * All budgets not got from this function will be put
1964 * together during handling partial dispatch
1967 blk_mq_put_dispatch_budget(rq->q, budget_token);
1968 return PREP_DISPATCH_NO_TAG;
1972 return PREP_DISPATCH_OK;
1975 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1976 static void blk_mq_release_budgets(struct request_queue *q,
1977 struct list_head *list)
1981 list_for_each_entry(rq, list, queuelist) {
1982 int budget_token = blk_mq_get_rq_budget_token(rq);
1984 if (budget_token >= 0)
1985 blk_mq_put_dispatch_budget(q, budget_token);
1990 * Returns true if we did some work AND can potentially do more.
1992 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1993 unsigned int nr_budgets)
1995 enum prep_dispatch prep;
1996 struct request_queue *q = hctx->queue;
1997 struct request *rq, *nxt;
1999 blk_status_t ret = BLK_STS_OK;
2000 LIST_HEAD(zone_list);
2001 bool needs_resource = false;
2003 if (list_empty(list))
2007 * Now process all the entries, sending them to the driver.
2009 errors = queued = 0;
2011 struct blk_mq_queue_data bd;
2013 rq = list_first_entry(list, struct request, queuelist);
2015 WARN_ON_ONCE(hctx != rq->mq_hctx);
2016 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2017 if (prep != PREP_DISPATCH_OK)
2020 list_del_init(&rq->queuelist);
2025 * Flag last if we have no more requests, or if we have more
2026 * but can't assign a driver tag to it.
2028 if (list_empty(list))
2031 nxt = list_first_entry(list, struct request, queuelist);
2032 bd.last = !blk_mq_get_driver_tag(nxt);
2036 * once the request is queued to lld, no need to cover the
2041 ret = q->mq_ops->queue_rq(hctx, &bd);
2046 case BLK_STS_RESOURCE:
2047 needs_resource = true;
2049 case BLK_STS_DEV_RESOURCE:
2050 blk_mq_handle_dev_resource(rq, list);
2052 case BLK_STS_ZONE_RESOURCE:
2054 * Move the request to zone_list and keep going through
2055 * the dispatch list to find more requests the drive can
2058 blk_mq_handle_zone_resource(rq, &zone_list);
2059 needs_resource = true;
2063 blk_mq_end_request(rq, ret);
2065 } while (!list_empty(list));
2067 if (!list_empty(&zone_list))
2068 list_splice_tail_init(&zone_list, list);
2070 /* If we didn't flush the entire list, we could have told the driver
2071 * there was more coming, but that turned out to be a lie.
2073 if ((!list_empty(list) || errors || needs_resource ||
2074 ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued)
2075 q->mq_ops->commit_rqs(hctx);
2077 * Any items that need requeuing? Stuff them into hctx->dispatch,
2078 * that is where we will continue on next queue run.
2080 if (!list_empty(list)) {
2082 /* For non-shared tags, the RESTART check will suffice */
2083 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2084 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2085 blk_mq_is_shared_tags(hctx->flags));
2088 blk_mq_release_budgets(q, list);
2090 spin_lock(&hctx->lock);
2091 list_splice_tail_init(list, &hctx->dispatch);
2092 spin_unlock(&hctx->lock);
2095 * Order adding requests to hctx->dispatch and checking
2096 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2097 * in blk_mq_sched_restart(). Avoid restart code path to
2098 * miss the new added requests to hctx->dispatch, meantime
2099 * SCHED_RESTART is observed here.
2104 * If SCHED_RESTART was set by the caller of this function and
2105 * it is no longer set that means that it was cleared by another
2106 * thread and hence that a queue rerun is needed.
2108 * If 'no_tag' is set, that means that we failed getting
2109 * a driver tag with an I/O scheduler attached. If our dispatch
2110 * waitqueue is no longer active, ensure that we run the queue
2111 * AFTER adding our entries back to the list.
2113 * If no I/O scheduler has been configured it is possible that
2114 * the hardware queue got stopped and restarted before requests
2115 * were pushed back onto the dispatch list. Rerun the queue to
2116 * avoid starvation. Notes:
2117 * - blk_mq_run_hw_queue() checks whether or not a queue has
2118 * been stopped before rerunning a queue.
2119 * - Some but not all block drivers stop a queue before
2120 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2123 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2124 * bit is set, run queue after a delay to avoid IO stalls
2125 * that could otherwise occur if the queue is idle. We'll do
2126 * similar if we couldn't get budget or couldn't lock a zone
2127 * and SCHED_RESTART is set.
2129 needs_restart = blk_mq_sched_needs_restart(hctx);
2130 if (prep == PREP_DISPATCH_NO_BUDGET)
2131 needs_resource = true;
2132 if (!needs_restart ||
2133 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2134 blk_mq_run_hw_queue(hctx, true);
2135 else if (needs_resource)
2136 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2138 blk_mq_update_dispatch_busy(hctx, true);
2141 blk_mq_update_dispatch_busy(hctx, false);
2143 return (queued + errors) != 0;
2147 * __blk_mq_run_hw_queue - Run a hardware queue.
2148 * @hctx: Pointer to the hardware queue to run.
2150 * Send pending requests to the hardware.
2152 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
2155 * We can't run the queue inline with ints disabled. Ensure that
2156 * we catch bad users of this early.
2158 WARN_ON_ONCE(in_interrupt());
2160 blk_mq_run_dispatch_ops(hctx->queue,
2161 blk_mq_sched_dispatch_requests(hctx));
2164 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2166 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2168 if (cpu >= nr_cpu_ids)
2169 cpu = cpumask_first(hctx->cpumask);
2174 * It'd be great if the workqueue API had a way to pass
2175 * in a mask and had some smarts for more clever placement.
2176 * For now we just round-robin here, switching for every
2177 * BLK_MQ_CPU_WORK_BATCH queued items.
2179 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2182 int next_cpu = hctx->next_cpu;
2184 if (hctx->queue->nr_hw_queues == 1)
2185 return WORK_CPU_UNBOUND;
2187 if (--hctx->next_cpu_batch <= 0) {
2189 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2191 if (next_cpu >= nr_cpu_ids)
2192 next_cpu = blk_mq_first_mapped_cpu(hctx);
2193 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2197 * Do unbound schedule if we can't find a online CPU for this hctx,
2198 * and it should only happen in the path of handling CPU DEAD.
2200 if (!cpu_online(next_cpu)) {
2207 * Make sure to re-select CPU next time once after CPUs
2208 * in hctx->cpumask become online again.
2210 hctx->next_cpu = next_cpu;
2211 hctx->next_cpu_batch = 1;
2212 return WORK_CPU_UNBOUND;
2215 hctx->next_cpu = next_cpu;
2220 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2221 * @hctx: Pointer to the hardware queue to run.
2222 * @async: If we want to run the queue asynchronously.
2223 * @msecs: Milliseconds of delay to wait before running the queue.
2225 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2226 * with a delay of @msecs.
2228 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2229 unsigned long msecs)
2231 if (unlikely(blk_mq_hctx_stopped(hctx)))
2234 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2235 if (cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2236 __blk_mq_run_hw_queue(hctx);
2241 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2242 msecs_to_jiffies(msecs));
2246 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2247 * @hctx: Pointer to the hardware queue to run.
2248 * @msecs: Milliseconds of delay to wait before running the queue.
2250 * Run a hardware queue asynchronously with a delay of @msecs.
2252 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2254 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2256 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2259 * blk_mq_run_hw_queue - Start to run a hardware queue.
2260 * @hctx: Pointer to the hardware queue to run.
2261 * @async: If we want to run the queue asynchronously.
2263 * Check if the request queue is not in a quiesced state and if there are
2264 * pending requests to be sent. If this is true, run the queue to send requests
2267 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2272 * When queue is quiesced, we may be switching io scheduler, or
2273 * updating nr_hw_queues, or other things, and we can't run queue
2274 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2276 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2279 __blk_mq_run_dispatch_ops(hctx->queue, false,
2280 need_run = !blk_queue_quiesced(hctx->queue) &&
2281 blk_mq_hctx_has_pending(hctx));
2284 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2286 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2289 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2292 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2294 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2296 * If the IO scheduler does not respect hardware queues when
2297 * dispatching, we just don't bother with multiple HW queues and
2298 * dispatch from hctx for the current CPU since running multiple queues
2299 * just causes lock contention inside the scheduler and pointless cache
2302 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2304 if (!blk_mq_hctx_stopped(hctx))
2310 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2311 * @q: Pointer to the request queue to run.
2312 * @async: If we want to run the queue asynchronously.
2314 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2316 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2320 if (blk_queue_sq_sched(q))
2321 sq_hctx = blk_mq_get_sq_hctx(q);
2322 queue_for_each_hw_ctx(q, hctx, i) {
2323 if (blk_mq_hctx_stopped(hctx))
2326 * Dispatch from this hctx either if there's no hctx preferred
2327 * by IO scheduler or if it has requests that bypass the
2330 if (!sq_hctx || sq_hctx == hctx ||
2331 !list_empty_careful(&hctx->dispatch))
2332 blk_mq_run_hw_queue(hctx, async);
2335 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2338 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2339 * @q: Pointer to the request queue to run.
2340 * @msecs: Milliseconds of delay to wait before running the queues.
2342 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2344 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2348 if (blk_queue_sq_sched(q))
2349 sq_hctx = blk_mq_get_sq_hctx(q);
2350 queue_for_each_hw_ctx(q, hctx, i) {
2351 if (blk_mq_hctx_stopped(hctx))
2354 * If there is already a run_work pending, leave the
2355 * pending delay untouched. Otherwise, a hctx can stall
2356 * if another hctx is re-delaying the other's work
2357 * before the work executes.
2359 if (delayed_work_pending(&hctx->run_work))
2362 * Dispatch from this hctx either if there's no hctx preferred
2363 * by IO scheduler or if it has requests that bypass the
2366 if (!sq_hctx || sq_hctx == hctx ||
2367 !list_empty_careful(&hctx->dispatch))
2368 blk_mq_delay_run_hw_queue(hctx, msecs);
2371 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2374 * This function is often used for pausing .queue_rq() by driver when
2375 * there isn't enough resource or some conditions aren't satisfied, and
2376 * BLK_STS_RESOURCE is usually returned.
2378 * We do not guarantee that dispatch can be drained or blocked
2379 * after blk_mq_stop_hw_queue() returns. Please use
2380 * blk_mq_quiesce_queue() for that requirement.
2382 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2384 cancel_delayed_work(&hctx->run_work);
2386 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2388 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2391 * This function is often used for pausing .queue_rq() by driver when
2392 * there isn't enough resource or some conditions aren't satisfied, and
2393 * BLK_STS_RESOURCE is usually returned.
2395 * We do not guarantee that dispatch can be drained or blocked
2396 * after blk_mq_stop_hw_queues() returns. Please use
2397 * blk_mq_quiesce_queue() for that requirement.
2399 void blk_mq_stop_hw_queues(struct request_queue *q)
2401 struct blk_mq_hw_ctx *hctx;
2404 queue_for_each_hw_ctx(q, hctx, i)
2405 blk_mq_stop_hw_queue(hctx);
2407 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2409 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2411 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2413 blk_mq_run_hw_queue(hctx, false);
2415 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2417 void blk_mq_start_hw_queues(struct request_queue *q)
2419 struct blk_mq_hw_ctx *hctx;
2422 queue_for_each_hw_ctx(q, hctx, i)
2423 blk_mq_start_hw_queue(hctx);
2425 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2427 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2429 if (!blk_mq_hctx_stopped(hctx))
2432 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2433 blk_mq_run_hw_queue(hctx, async);
2435 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2437 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2439 struct blk_mq_hw_ctx *hctx;
2442 queue_for_each_hw_ctx(q, hctx, i)
2443 blk_mq_start_stopped_hw_queue(hctx, async);
2445 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2447 static void blk_mq_run_work_fn(struct work_struct *work)
2449 struct blk_mq_hw_ctx *hctx;
2451 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2454 * If we are stopped, don't run the queue.
2456 if (blk_mq_hctx_stopped(hctx))
2459 __blk_mq_run_hw_queue(hctx);
2462 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2466 struct blk_mq_ctx *ctx = rq->mq_ctx;
2467 enum hctx_type type = hctx->type;
2469 lockdep_assert_held(&ctx->lock);
2471 trace_block_rq_insert(rq);
2474 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2476 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2479 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2482 struct blk_mq_ctx *ctx = rq->mq_ctx;
2484 lockdep_assert_held(&ctx->lock);
2486 __blk_mq_insert_req_list(hctx, rq, at_head);
2487 blk_mq_hctx_mark_pending(hctx, ctx);
2491 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2492 * @rq: Pointer to request to be inserted.
2493 * @at_head: true if the request should be inserted at the head of the list.
2494 * @run_queue: If we should run the hardware queue after inserting the request.
2496 * Should only be used carefully, when the caller knows we want to
2497 * bypass a potential IO scheduler on the target device.
2499 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2502 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2504 spin_lock(&hctx->lock);
2506 list_add(&rq->queuelist, &hctx->dispatch);
2508 list_add_tail(&rq->queuelist, &hctx->dispatch);
2509 spin_unlock(&hctx->lock);
2512 blk_mq_run_hw_queue(hctx, false);
2515 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2516 struct list_head *list)
2520 enum hctx_type type = hctx->type;
2523 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2526 list_for_each_entry(rq, list, queuelist) {
2527 BUG_ON(rq->mq_ctx != ctx);
2528 trace_block_rq_insert(rq);
2531 spin_lock(&ctx->lock);
2532 list_splice_tail_init(list, &ctx->rq_lists[type]);
2533 blk_mq_hctx_mark_pending(hctx, ctx);
2534 spin_unlock(&ctx->lock);
2537 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2540 if (hctx->queue->mq_ops->commit_rqs) {
2541 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2542 hctx->queue->mq_ops->commit_rqs(hctx);
2547 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2548 unsigned int nr_segs)
2552 if (bio->bi_opf & REQ_RAHEAD)
2553 rq->cmd_flags |= REQ_FAILFAST_MASK;
2555 rq->__sector = bio->bi_iter.bi_sector;
2556 blk_rq_bio_prep(rq, bio, nr_segs);
2558 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2559 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2562 blk_account_io_start(rq);
2565 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2566 struct request *rq, bool last)
2568 struct request_queue *q = rq->q;
2569 struct blk_mq_queue_data bd = {
2576 * For OK queue, we are done. For error, caller may kill it.
2577 * Any other error (busy), just add it to our list as we
2578 * previously would have done.
2580 ret = q->mq_ops->queue_rq(hctx, &bd);
2583 blk_mq_update_dispatch_busy(hctx, false);
2585 case BLK_STS_RESOURCE:
2586 case BLK_STS_DEV_RESOURCE:
2587 blk_mq_update_dispatch_busy(hctx, true);
2588 __blk_mq_requeue_request(rq);
2591 blk_mq_update_dispatch_busy(hctx, false);
2598 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2600 bool bypass_insert, bool last)
2602 struct request_queue *q = rq->q;
2603 bool run_queue = true;
2607 * RCU or SRCU read lock is needed before checking quiesced flag.
2609 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2610 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2611 * and avoid driver to try to dispatch again.
2613 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2615 bypass_insert = false;
2619 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2622 budget_token = blk_mq_get_dispatch_budget(q);
2623 if (budget_token < 0)
2626 blk_mq_set_rq_budget_token(rq, budget_token);
2628 if (!blk_mq_get_driver_tag(rq)) {
2629 blk_mq_put_dispatch_budget(q, budget_token);
2633 return __blk_mq_issue_directly(hctx, rq, last);
2636 return BLK_STS_RESOURCE;
2638 blk_mq_sched_insert_request(rq, false, run_queue, false);
2644 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2645 * @hctx: Pointer of the associated hardware queue.
2646 * @rq: Pointer to request to be sent.
2648 * If the device has enough resources to accept a new request now, send the
2649 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2650 * we can try send it another time in the future. Requests inserted at this
2651 * queue have higher priority.
2653 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2657 __blk_mq_try_issue_directly(hctx, rq, false, true);
2659 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2660 blk_mq_request_bypass_insert(rq, false, true);
2661 else if (ret != BLK_STS_OK)
2662 blk_mq_end_request(rq, ret);
2665 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2667 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2670 static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2672 struct blk_mq_hw_ctx *hctx = NULL;
2677 while ((rq = rq_list_pop(&plug->mq_list))) {
2678 bool last = rq_list_empty(plug->mq_list);
2681 if (hctx != rq->mq_hctx) {
2683 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2687 ret = blk_mq_request_issue_directly(rq, last);
2692 case BLK_STS_RESOURCE:
2693 case BLK_STS_DEV_RESOURCE:
2694 blk_mq_request_bypass_insert(rq, false, true);
2695 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2698 blk_mq_end_request(rq, ret);
2705 * If we didn't flush the entire list, we could have told the driver
2706 * there was more coming, but that turned out to be a lie.
2709 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2712 static void __blk_mq_flush_plug_list(struct request_queue *q,
2713 struct blk_plug *plug)
2715 if (blk_queue_quiesced(q))
2717 q->mq_ops->queue_rqs(&plug->mq_list);
2720 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2722 struct blk_mq_hw_ctx *this_hctx = NULL;
2723 struct blk_mq_ctx *this_ctx = NULL;
2724 struct request *requeue_list = NULL;
2725 struct request **requeue_lastp = &requeue_list;
2726 unsigned int depth = 0;
2730 struct request *rq = rq_list_pop(&plug->mq_list);
2733 this_hctx = rq->mq_hctx;
2734 this_ctx = rq->mq_ctx;
2735 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2736 rq_list_add_tail(&requeue_lastp, rq);
2739 list_add(&rq->queuelist, &list);
2741 } while (!rq_list_empty(plug->mq_list));
2743 plug->mq_list = requeue_list;
2744 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2745 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched);
2748 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2752 if (rq_list_empty(plug->mq_list))
2756 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2757 struct request_queue *q;
2759 rq = rq_list_peek(&plug->mq_list);
2763 * Peek first request and see if we have a ->queue_rqs() hook.
2764 * If we do, we can dispatch the whole plug list in one go. We
2765 * already know at this point that all requests belong to the
2766 * same queue, caller must ensure that's the case.
2768 * Since we pass off the full list to the driver at this point,
2769 * we do not increment the active request count for the queue.
2770 * Bypass shared tags for now because of that.
2772 if (q->mq_ops->queue_rqs &&
2773 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2774 blk_mq_run_dispatch_ops(q,
2775 __blk_mq_flush_plug_list(q, plug));
2776 if (rq_list_empty(plug->mq_list))
2780 blk_mq_run_dispatch_ops(q,
2781 blk_mq_plug_issue_direct(plug, false));
2782 if (rq_list_empty(plug->mq_list))
2787 blk_mq_dispatch_plug_list(plug, from_schedule);
2788 } while (!rq_list_empty(plug->mq_list));
2791 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2792 struct list_head *list)
2797 while (!list_empty(list)) {
2799 struct request *rq = list_first_entry(list, struct request,
2802 list_del_init(&rq->queuelist);
2803 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2804 if (ret != BLK_STS_OK) {
2806 if (ret == BLK_STS_RESOURCE ||
2807 ret == BLK_STS_DEV_RESOURCE) {
2808 blk_mq_request_bypass_insert(rq, false,
2812 blk_mq_end_request(rq, ret);
2818 * If we didn't flush the entire list, we could have told
2819 * the driver there was more coming, but that turned out to
2822 if ((!list_empty(list) || errors) &&
2823 hctx->queue->mq_ops->commit_rqs && queued)
2824 hctx->queue->mq_ops->commit_rqs(hctx);
2827 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2828 struct bio *bio, unsigned int nr_segs)
2830 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2831 if (blk_attempt_plug_merge(q, bio, nr_segs))
2833 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2839 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2840 struct blk_plug *plug,
2844 struct blk_mq_alloc_data data = {
2847 .cmd_flags = bio->bi_opf,
2851 if (unlikely(bio_queue_enter(bio)))
2854 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2857 rq_qos_throttle(q, bio);
2860 data.nr_tags = plug->nr_ios;
2862 data.cached_rq = &plug->cached_rq;
2865 rq = __blk_mq_alloc_requests(&data);
2868 rq_qos_cleanup(q, bio);
2869 if (bio->bi_opf & REQ_NOWAIT)
2870 bio_wouldblock_error(bio);
2876 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2877 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2880 enum hctx_type type, hctx_type;
2884 rq = rq_list_peek(&plug->cached_rq);
2885 if (!rq || rq->q != q)
2888 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2893 type = blk_mq_get_hctx_type((*bio)->bi_opf);
2894 hctx_type = rq->mq_hctx->type;
2895 if (type != hctx_type &&
2896 !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2898 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2902 * If any qos ->throttle() end up blocking, we will have flushed the
2903 * plug and hence killed the cached_rq list as well. Pop this entry
2904 * before we throttle.
2906 plug->cached_rq = rq_list_next(rq);
2907 rq_qos_throttle(q, *bio);
2909 rq->cmd_flags = (*bio)->bi_opf;
2910 INIT_LIST_HEAD(&rq->queuelist);
2914 static void bio_set_ioprio(struct bio *bio)
2916 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2917 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2918 bio->bi_ioprio = get_current_ioprio();
2919 blkcg_set_ioprio(bio);
2923 * blk_mq_submit_bio - Create and send a request to block device.
2924 * @bio: Bio pointer.
2926 * Builds up a request structure from @q and @bio and send to the device. The
2927 * request may not be queued directly to hardware if:
2928 * * This request can be merged with another one
2929 * * We want to place request at plug queue for possible future merging
2930 * * There is an IO scheduler active at this queue
2932 * It will not queue the request if there is an error with the bio, or at the
2935 void blk_mq_submit_bio(struct bio *bio)
2937 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2938 struct blk_plug *plug = blk_mq_plug(bio);
2939 const int is_sync = op_is_sync(bio->bi_opf);
2941 unsigned int nr_segs = 1;
2944 bio = blk_queue_bounce(bio, q);
2945 if (bio_may_exceed_limits(bio, &q->limits)) {
2946 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2951 if (!bio_integrity_prep(bio))
2954 bio_set_ioprio(bio);
2956 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2960 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2965 trace_block_getrq(bio);
2967 rq_qos_track(q, rq, bio);
2969 blk_mq_bio_to_request(rq, bio, nr_segs);
2971 ret = blk_crypto_rq_get_keyslot(rq);
2972 if (ret != BLK_STS_OK) {
2973 bio->bi_status = ret;
2975 blk_mq_free_request(rq);
2979 if (op_is_flush(bio->bi_opf)) {
2980 blk_insert_flush(rq);
2985 blk_add_rq_to_plug(plug, rq);
2986 else if ((rq->rq_flags & RQF_ELV) ||
2987 (rq->mq_hctx->dispatch_busy &&
2988 (q->nr_hw_queues == 1 || !is_sync)))
2989 blk_mq_sched_insert_request(rq, false, true, true);
2991 blk_mq_run_dispatch_ops(rq->q,
2992 blk_mq_try_issue_directly(rq->mq_hctx, rq));
2995 #ifdef CONFIG_BLK_MQ_STACKING
2997 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
2998 * @rq: the request being queued
3000 blk_status_t blk_insert_cloned_request(struct request *rq)
3002 struct request_queue *q = rq->q;
3003 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3006 if (blk_rq_sectors(rq) > max_sectors) {
3008 * SCSI device does not have a good way to return if
3009 * Write Same/Zero is actually supported. If a device rejects
3010 * a non-read/write command (discard, write same,etc.) the
3011 * low-level device driver will set the relevant queue limit to
3012 * 0 to prevent blk-lib from issuing more of the offending
3013 * operations. Commands queued prior to the queue limit being
3014 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3015 * errors being propagated to upper layers.
3017 if (max_sectors == 0)
3018 return BLK_STS_NOTSUPP;
3020 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3021 __func__, blk_rq_sectors(rq), max_sectors);
3022 return BLK_STS_IOERR;
3026 * The queue settings related to segment counting may differ from the
3029 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3030 if (rq->nr_phys_segments > queue_max_segments(q)) {
3031 printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n",
3032 __func__, rq->nr_phys_segments, queue_max_segments(q));
3033 return BLK_STS_IOERR;
3036 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3037 return BLK_STS_IOERR;
3039 if (blk_crypto_insert_cloned_request(rq))
3040 return BLK_STS_IOERR;
3042 blk_account_io_start(rq);
3045 * Since we have a scheduler attached on the top device,
3046 * bypass a potential scheduler on the bottom device for
3049 blk_mq_run_dispatch_ops(q,
3050 ret = blk_mq_request_issue_directly(rq, true));
3052 blk_account_io_done(rq, ktime_get_ns());
3055 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3058 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3059 * @rq: the clone request to be cleaned up
3062 * Free all bios in @rq for a cloned request.
3064 void blk_rq_unprep_clone(struct request *rq)
3068 while ((bio = rq->bio) != NULL) {
3069 rq->bio = bio->bi_next;
3074 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3077 * blk_rq_prep_clone - Helper function to setup clone request
3078 * @rq: the request to be setup
3079 * @rq_src: original request to be cloned
3080 * @bs: bio_set that bios for clone are allocated from
3081 * @gfp_mask: memory allocation mask for bio
3082 * @bio_ctr: setup function to be called for each clone bio.
3083 * Returns %0 for success, non %0 for failure.
3084 * @data: private data to be passed to @bio_ctr
3087 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3088 * Also, pages which the original bios are pointing to are not copied
3089 * and the cloned bios just point same pages.
3090 * So cloned bios must be completed before original bios, which means
3091 * the caller must complete @rq before @rq_src.
3093 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3094 struct bio_set *bs, gfp_t gfp_mask,
3095 int (*bio_ctr)(struct bio *, struct bio *, void *),
3098 struct bio *bio, *bio_src;
3103 __rq_for_each_bio(bio_src, rq_src) {
3104 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3109 if (bio_ctr && bio_ctr(bio, bio_src, data))
3113 rq->biotail->bi_next = bio;
3116 rq->bio = rq->biotail = bio;
3121 /* Copy attributes of the original request to the clone request. */
3122 rq->__sector = blk_rq_pos(rq_src);
3123 rq->__data_len = blk_rq_bytes(rq_src);
3124 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3125 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3126 rq->special_vec = rq_src->special_vec;
3128 rq->nr_phys_segments = rq_src->nr_phys_segments;
3129 rq->ioprio = rq_src->ioprio;
3131 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3139 blk_rq_unprep_clone(rq);
3143 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3144 #endif /* CONFIG_BLK_MQ_STACKING */
3147 * Steal bios from a request and add them to a bio list.
3148 * The request must not have been partially completed before.
3150 void blk_steal_bios(struct bio_list *list, struct request *rq)
3154 list->tail->bi_next = rq->bio;
3156 list->head = rq->bio;
3157 list->tail = rq->biotail;
3165 EXPORT_SYMBOL_GPL(blk_steal_bios);
3167 static size_t order_to_size(unsigned int order)
3169 return (size_t)PAGE_SIZE << order;
3172 /* called before freeing request pool in @tags */
3173 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3174 struct blk_mq_tags *tags)
3177 unsigned long flags;
3180 * There is no need to clear mapping if driver tags is not initialized
3181 * or the mapping belongs to the driver tags.
3183 if (!drv_tags || drv_tags == tags)
3186 list_for_each_entry(page, &tags->page_list, lru) {
3187 unsigned long start = (unsigned long)page_address(page);
3188 unsigned long end = start + order_to_size(page->private);
3191 for (i = 0; i < drv_tags->nr_tags; i++) {
3192 struct request *rq = drv_tags->rqs[i];
3193 unsigned long rq_addr = (unsigned long)rq;
3195 if (rq_addr >= start && rq_addr < end) {
3196 WARN_ON_ONCE(req_ref_read(rq) != 0);
3197 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3203 * Wait until all pending iteration is done.
3205 * Request reference is cleared and it is guaranteed to be observed
3206 * after the ->lock is released.
3208 spin_lock_irqsave(&drv_tags->lock, flags);
3209 spin_unlock_irqrestore(&drv_tags->lock, flags);
3212 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3213 unsigned int hctx_idx)
3215 struct blk_mq_tags *drv_tags;
3218 if (list_empty(&tags->page_list))
3221 if (blk_mq_is_shared_tags(set->flags))
3222 drv_tags = set->shared_tags;
3224 drv_tags = set->tags[hctx_idx];
3226 if (tags->static_rqs && set->ops->exit_request) {
3229 for (i = 0; i < tags->nr_tags; i++) {
3230 struct request *rq = tags->static_rqs[i];
3234 set->ops->exit_request(set, rq, hctx_idx);
3235 tags->static_rqs[i] = NULL;
3239 blk_mq_clear_rq_mapping(drv_tags, tags);
3241 while (!list_empty(&tags->page_list)) {
3242 page = list_first_entry(&tags->page_list, struct page, lru);
3243 list_del_init(&page->lru);
3245 * Remove kmemleak object previously allocated in
3246 * blk_mq_alloc_rqs().
3248 kmemleak_free(page_address(page));
3249 __free_pages(page, page->private);
3253 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3257 kfree(tags->static_rqs);
3258 tags->static_rqs = NULL;
3260 blk_mq_free_tags(tags);
3263 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3264 unsigned int hctx_idx)
3268 for (i = 0; i < set->nr_maps; i++) {
3269 unsigned int start = set->map[i].queue_offset;
3270 unsigned int end = start + set->map[i].nr_queues;
3272 if (hctx_idx >= start && hctx_idx < end)
3276 if (i >= set->nr_maps)
3277 i = HCTX_TYPE_DEFAULT;
3282 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3283 unsigned int hctx_idx)
3285 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3287 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3290 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3291 unsigned int hctx_idx,
3292 unsigned int nr_tags,
3293 unsigned int reserved_tags)
3295 int node = blk_mq_get_hctx_node(set, hctx_idx);
3296 struct blk_mq_tags *tags;
3298 if (node == NUMA_NO_NODE)
3299 node = set->numa_node;
3301 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3302 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3306 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3307 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3310 blk_mq_free_tags(tags);
3314 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3315 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3317 if (!tags->static_rqs) {
3319 blk_mq_free_tags(tags);
3326 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3327 unsigned int hctx_idx, int node)
3331 if (set->ops->init_request) {
3332 ret = set->ops->init_request(set, rq, hctx_idx, node);
3337 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3341 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3342 struct blk_mq_tags *tags,
3343 unsigned int hctx_idx, unsigned int depth)
3345 unsigned int i, j, entries_per_page, max_order = 4;
3346 int node = blk_mq_get_hctx_node(set, hctx_idx);
3347 size_t rq_size, left;
3349 if (node == NUMA_NO_NODE)
3350 node = set->numa_node;
3352 INIT_LIST_HEAD(&tags->page_list);
3355 * rq_size is the size of the request plus driver payload, rounded
3356 * to the cacheline size
3358 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3360 left = rq_size * depth;
3362 for (i = 0; i < depth; ) {
3363 int this_order = max_order;
3368 while (this_order && left < order_to_size(this_order - 1))
3372 page = alloc_pages_node(node,
3373 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3379 if (order_to_size(this_order) < rq_size)
3386 page->private = this_order;
3387 list_add_tail(&page->lru, &tags->page_list);
3389 p = page_address(page);
3391 * Allow kmemleak to scan these pages as they contain pointers
3392 * to additional allocations like via ops->init_request().
3394 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3395 entries_per_page = order_to_size(this_order) / rq_size;
3396 to_do = min(entries_per_page, depth - i);
3397 left -= to_do * rq_size;
3398 for (j = 0; j < to_do; j++) {
3399 struct request *rq = p;
3401 tags->static_rqs[i] = rq;
3402 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3403 tags->static_rqs[i] = NULL;
3414 blk_mq_free_rqs(set, tags, hctx_idx);
3418 struct rq_iter_data {
3419 struct blk_mq_hw_ctx *hctx;
3423 static bool blk_mq_has_request(struct request *rq, void *data)
3425 struct rq_iter_data *iter_data = data;
3427 if (rq->mq_hctx != iter_data->hctx)
3429 iter_data->has_rq = true;
3433 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3435 struct blk_mq_tags *tags = hctx->sched_tags ?
3436 hctx->sched_tags : hctx->tags;
3437 struct rq_iter_data data = {
3441 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3445 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3446 struct blk_mq_hw_ctx *hctx)
3448 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3450 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3455 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3457 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3458 struct blk_mq_hw_ctx, cpuhp_online);
3460 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3461 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3465 * Prevent new request from being allocated on the current hctx.
3467 * The smp_mb__after_atomic() Pairs with the implied barrier in
3468 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3469 * seen once we return from the tag allocator.
3471 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3472 smp_mb__after_atomic();
3475 * Try to grab a reference to the queue and wait for any outstanding
3476 * requests. If we could not grab a reference the queue has been
3477 * frozen and there are no requests.
3479 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3480 while (blk_mq_hctx_has_requests(hctx))
3482 percpu_ref_put(&hctx->queue->q_usage_counter);
3488 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3490 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3491 struct blk_mq_hw_ctx, cpuhp_online);
3493 if (cpumask_test_cpu(cpu, hctx->cpumask))
3494 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3499 * 'cpu' is going away. splice any existing rq_list entries from this
3500 * software queue to the hw queue dispatch list, and ensure that it
3503 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3505 struct blk_mq_hw_ctx *hctx;
3506 struct blk_mq_ctx *ctx;
3508 enum hctx_type type;
3510 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3511 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3514 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3517 spin_lock(&ctx->lock);
3518 if (!list_empty(&ctx->rq_lists[type])) {
3519 list_splice_init(&ctx->rq_lists[type], &tmp);
3520 blk_mq_hctx_clear_pending(hctx, ctx);
3522 spin_unlock(&ctx->lock);
3524 if (list_empty(&tmp))
3527 spin_lock(&hctx->lock);
3528 list_splice_tail_init(&tmp, &hctx->dispatch);
3529 spin_unlock(&hctx->lock);
3531 blk_mq_run_hw_queue(hctx, true);
3535 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3537 if (!(hctx->flags & BLK_MQ_F_STACKING))
3538 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3539 &hctx->cpuhp_online);
3540 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3545 * Before freeing hw queue, clearing the flush request reference in
3546 * tags->rqs[] for avoiding potential UAF.
3548 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3549 unsigned int queue_depth, struct request *flush_rq)
3552 unsigned long flags;
3554 /* The hw queue may not be mapped yet */
3558 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3560 for (i = 0; i < queue_depth; i++)
3561 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3564 * Wait until all pending iteration is done.
3566 * Request reference is cleared and it is guaranteed to be observed
3567 * after the ->lock is released.
3569 spin_lock_irqsave(&tags->lock, flags);
3570 spin_unlock_irqrestore(&tags->lock, flags);
3573 /* hctx->ctxs will be freed in queue's release handler */
3574 static void blk_mq_exit_hctx(struct request_queue *q,
3575 struct blk_mq_tag_set *set,
3576 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3578 struct request *flush_rq = hctx->fq->flush_rq;
3580 if (blk_mq_hw_queue_mapped(hctx))
3581 blk_mq_tag_idle(hctx);
3583 if (blk_queue_init_done(q))
3584 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3585 set->queue_depth, flush_rq);
3586 if (set->ops->exit_request)
3587 set->ops->exit_request(set, flush_rq, hctx_idx);
3589 if (set->ops->exit_hctx)
3590 set->ops->exit_hctx(hctx, hctx_idx);
3592 blk_mq_remove_cpuhp(hctx);
3594 xa_erase(&q->hctx_table, hctx_idx);
3596 spin_lock(&q->unused_hctx_lock);
3597 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3598 spin_unlock(&q->unused_hctx_lock);
3601 static void blk_mq_exit_hw_queues(struct request_queue *q,
3602 struct blk_mq_tag_set *set, int nr_queue)
3604 struct blk_mq_hw_ctx *hctx;
3607 queue_for_each_hw_ctx(q, hctx, i) {
3610 blk_mq_exit_hctx(q, set, hctx, i);
3614 static int blk_mq_init_hctx(struct request_queue *q,
3615 struct blk_mq_tag_set *set,
3616 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3618 hctx->queue_num = hctx_idx;
3620 if (!(hctx->flags & BLK_MQ_F_STACKING))
3621 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3622 &hctx->cpuhp_online);
3623 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3625 hctx->tags = set->tags[hctx_idx];
3627 if (set->ops->init_hctx &&
3628 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3629 goto unregister_cpu_notifier;
3631 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3635 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3641 if (set->ops->exit_request)
3642 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3644 if (set->ops->exit_hctx)
3645 set->ops->exit_hctx(hctx, hctx_idx);
3646 unregister_cpu_notifier:
3647 blk_mq_remove_cpuhp(hctx);
3651 static struct blk_mq_hw_ctx *
3652 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3655 struct blk_mq_hw_ctx *hctx;
3656 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3658 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3660 goto fail_alloc_hctx;
3662 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3665 atomic_set(&hctx->nr_active, 0);
3666 if (node == NUMA_NO_NODE)
3667 node = set->numa_node;
3668 hctx->numa_node = node;
3670 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3671 spin_lock_init(&hctx->lock);
3672 INIT_LIST_HEAD(&hctx->dispatch);
3674 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3676 INIT_LIST_HEAD(&hctx->hctx_list);
3679 * Allocate space for all possible cpus to avoid allocation at
3682 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3687 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3688 gfp, node, false, false))
3692 spin_lock_init(&hctx->dispatch_wait_lock);
3693 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3694 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3696 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3700 blk_mq_hctx_kobj_init(hctx);
3705 sbitmap_free(&hctx->ctx_map);
3709 free_cpumask_var(hctx->cpumask);
3716 static void blk_mq_init_cpu_queues(struct request_queue *q,
3717 unsigned int nr_hw_queues)
3719 struct blk_mq_tag_set *set = q->tag_set;
3722 for_each_possible_cpu(i) {
3723 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3724 struct blk_mq_hw_ctx *hctx;
3728 spin_lock_init(&__ctx->lock);
3729 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3730 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3735 * Set local node, IFF we have more than one hw queue. If
3736 * not, we remain on the home node of the device
3738 for (j = 0; j < set->nr_maps; j++) {
3739 hctx = blk_mq_map_queue_type(q, j, i);
3740 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3741 hctx->numa_node = cpu_to_node(i);
3746 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3747 unsigned int hctx_idx,
3750 struct blk_mq_tags *tags;
3753 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3757 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3759 blk_mq_free_rq_map(tags);
3766 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3769 if (blk_mq_is_shared_tags(set->flags)) {
3770 set->tags[hctx_idx] = set->shared_tags;
3775 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3778 return set->tags[hctx_idx];
3781 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3782 struct blk_mq_tags *tags,
3783 unsigned int hctx_idx)
3786 blk_mq_free_rqs(set, tags, hctx_idx);
3787 blk_mq_free_rq_map(tags);
3791 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3792 unsigned int hctx_idx)
3794 if (!blk_mq_is_shared_tags(set->flags))
3795 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3797 set->tags[hctx_idx] = NULL;
3800 static void blk_mq_map_swqueue(struct request_queue *q)
3802 unsigned int j, hctx_idx;
3804 struct blk_mq_hw_ctx *hctx;
3805 struct blk_mq_ctx *ctx;
3806 struct blk_mq_tag_set *set = q->tag_set;
3808 queue_for_each_hw_ctx(q, hctx, i) {
3809 cpumask_clear(hctx->cpumask);
3811 hctx->dispatch_from = NULL;
3815 * Map software to hardware queues.
3817 * If the cpu isn't present, the cpu is mapped to first hctx.
3819 for_each_possible_cpu(i) {
3821 ctx = per_cpu_ptr(q->queue_ctx, i);
3822 for (j = 0; j < set->nr_maps; j++) {
3823 if (!set->map[j].nr_queues) {
3824 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3825 HCTX_TYPE_DEFAULT, i);
3828 hctx_idx = set->map[j].mq_map[i];
3829 /* unmapped hw queue can be remapped after CPU topo changed */
3830 if (!set->tags[hctx_idx] &&
3831 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3833 * If tags initialization fail for some hctx,
3834 * that hctx won't be brought online. In this
3835 * case, remap the current ctx to hctx[0] which
3836 * is guaranteed to always have tags allocated
3838 set->map[j].mq_map[i] = 0;
3841 hctx = blk_mq_map_queue_type(q, j, i);
3842 ctx->hctxs[j] = hctx;
3844 * If the CPU is already set in the mask, then we've
3845 * mapped this one already. This can happen if
3846 * devices share queues across queue maps.
3848 if (cpumask_test_cpu(i, hctx->cpumask))
3851 cpumask_set_cpu(i, hctx->cpumask);
3853 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3854 hctx->ctxs[hctx->nr_ctx++] = ctx;
3857 * If the nr_ctx type overflows, we have exceeded the
3858 * amount of sw queues we can support.
3860 BUG_ON(!hctx->nr_ctx);
3863 for (; j < HCTX_MAX_TYPES; j++)
3864 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3865 HCTX_TYPE_DEFAULT, i);
3868 queue_for_each_hw_ctx(q, hctx, i) {
3870 * If no software queues are mapped to this hardware queue,
3871 * disable it and free the request entries.
3873 if (!hctx->nr_ctx) {
3874 /* Never unmap queue 0. We need it as a
3875 * fallback in case of a new remap fails
3879 __blk_mq_free_map_and_rqs(set, i);
3885 hctx->tags = set->tags[i];
3886 WARN_ON(!hctx->tags);
3889 * Set the map size to the number of mapped software queues.
3890 * This is more accurate and more efficient than looping
3891 * over all possibly mapped software queues.
3893 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3896 * Initialize batch roundrobin counts
3898 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3899 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3904 * Caller needs to ensure that we're either frozen/quiesced, or that
3905 * the queue isn't live yet.
3907 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3909 struct blk_mq_hw_ctx *hctx;
3912 queue_for_each_hw_ctx(q, hctx, i) {
3914 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3916 blk_mq_tag_idle(hctx);
3917 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3922 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3925 struct request_queue *q;
3927 lockdep_assert_held(&set->tag_list_lock);
3929 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3930 blk_mq_freeze_queue(q);
3931 queue_set_hctx_shared(q, shared);
3932 blk_mq_unfreeze_queue(q);
3936 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3938 struct blk_mq_tag_set *set = q->tag_set;
3940 mutex_lock(&set->tag_list_lock);
3941 list_del(&q->tag_set_list);
3942 if (list_is_singular(&set->tag_list)) {
3943 /* just transitioned to unshared */
3944 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3945 /* update existing queue */
3946 blk_mq_update_tag_set_shared(set, false);
3948 mutex_unlock(&set->tag_list_lock);
3949 INIT_LIST_HEAD(&q->tag_set_list);
3952 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3953 struct request_queue *q)
3955 mutex_lock(&set->tag_list_lock);
3958 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3960 if (!list_empty(&set->tag_list) &&
3961 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3962 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3963 /* update existing queue */
3964 blk_mq_update_tag_set_shared(set, true);
3966 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3967 queue_set_hctx_shared(q, true);
3968 list_add_tail(&q->tag_set_list, &set->tag_list);
3970 mutex_unlock(&set->tag_list_lock);
3973 /* All allocations will be freed in release handler of q->mq_kobj */
3974 static int blk_mq_alloc_ctxs(struct request_queue *q)
3976 struct blk_mq_ctxs *ctxs;
3979 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3983 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3984 if (!ctxs->queue_ctx)
3987 for_each_possible_cpu(cpu) {
3988 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3992 q->mq_kobj = &ctxs->kobj;
3993 q->queue_ctx = ctxs->queue_ctx;
4002 * It is the actual release handler for mq, but we do it from
4003 * request queue's release handler for avoiding use-after-free
4004 * and headache because q->mq_kobj shouldn't have been introduced,
4005 * but we can't group ctx/kctx kobj without it.
4007 void blk_mq_release(struct request_queue *q)
4009 struct blk_mq_hw_ctx *hctx, *next;
4012 queue_for_each_hw_ctx(q, hctx, i)
4013 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4015 /* all hctx are in .unused_hctx_list now */
4016 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4017 list_del_init(&hctx->hctx_list);
4018 kobject_put(&hctx->kobj);
4021 xa_destroy(&q->hctx_table);
4024 * release .mq_kobj and sw queue's kobject now because
4025 * both share lifetime with request queue.
4027 blk_mq_sysfs_deinit(q);
4030 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4033 struct request_queue *q;
4036 q = blk_alloc_queue(set->numa_node, set->flags & BLK_MQ_F_BLOCKING);
4038 return ERR_PTR(-ENOMEM);
4039 q->queuedata = queuedata;
4040 ret = blk_mq_init_allocated_queue(set, q);
4043 return ERR_PTR(ret);
4048 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4050 return blk_mq_init_queue_data(set, NULL);
4052 EXPORT_SYMBOL(blk_mq_init_queue);
4055 * blk_mq_destroy_queue - shutdown a request queue
4056 * @q: request queue to shutdown
4058 * This shuts down a request queue allocated by blk_mq_init_queue() and drops
4059 * the initial reference. All future requests will failed with -ENODEV.
4061 * Context: can sleep
4063 void blk_mq_destroy_queue(struct request_queue *q)
4065 WARN_ON_ONCE(!queue_is_mq(q));
4066 WARN_ON_ONCE(blk_queue_registered(q));
4070 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4071 blk_queue_start_drain(q);
4072 blk_freeze_queue(q);
4075 blk_mq_cancel_work_sync(q);
4076 blk_mq_exit_queue(q);
4078 /* @q is and will stay empty, shutdown and put */
4081 EXPORT_SYMBOL(blk_mq_destroy_queue);
4083 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4084 struct lock_class_key *lkclass)
4086 struct request_queue *q;
4087 struct gendisk *disk;
4089 q = blk_mq_init_queue_data(set, queuedata);
4093 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4095 blk_mq_destroy_queue(q);
4096 return ERR_PTR(-ENOMEM);
4098 set_bit(GD_OWNS_QUEUE, &disk->state);
4101 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4103 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4104 struct lock_class_key *lkclass)
4106 struct gendisk *disk;
4108 if (!blk_get_queue(q))
4110 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4115 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4117 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4118 struct blk_mq_tag_set *set, struct request_queue *q,
4119 int hctx_idx, int node)
4121 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4123 /* reuse dead hctx first */
4124 spin_lock(&q->unused_hctx_lock);
4125 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4126 if (tmp->numa_node == node) {
4132 list_del_init(&hctx->hctx_list);
4133 spin_unlock(&q->unused_hctx_lock);
4136 hctx = blk_mq_alloc_hctx(q, set, node);
4140 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4146 kobject_put(&hctx->kobj);
4151 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4152 struct request_queue *q)
4154 struct blk_mq_hw_ctx *hctx;
4157 /* protect against switching io scheduler */
4158 mutex_lock(&q->sysfs_lock);
4159 for (i = 0; i < set->nr_hw_queues; i++) {
4161 int node = blk_mq_get_hctx_node(set, i);
4162 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4165 old_node = old_hctx->numa_node;
4166 blk_mq_exit_hctx(q, set, old_hctx, i);
4169 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4172 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4174 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4175 WARN_ON_ONCE(!hctx);
4179 * Increasing nr_hw_queues fails. Free the newly allocated
4180 * hctxs and keep the previous q->nr_hw_queues.
4182 if (i != set->nr_hw_queues) {
4183 j = q->nr_hw_queues;
4186 q->nr_hw_queues = set->nr_hw_queues;
4189 xa_for_each_start(&q->hctx_table, j, hctx, j)
4190 blk_mq_exit_hctx(q, set, hctx, j);
4191 mutex_unlock(&q->sysfs_lock);
4194 static void blk_mq_update_poll_flag(struct request_queue *q)
4196 struct blk_mq_tag_set *set = q->tag_set;
4198 if (set->nr_maps > HCTX_TYPE_POLL &&
4199 set->map[HCTX_TYPE_POLL].nr_queues)
4200 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4202 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4205 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4206 struct request_queue *q)
4208 WARN_ON_ONCE(blk_queue_has_srcu(q) !=
4209 !!(set->flags & BLK_MQ_F_BLOCKING));
4211 /* mark the queue as mq asap */
4212 q->mq_ops = set->ops;
4214 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
4215 blk_mq_poll_stats_bkt,
4216 BLK_MQ_POLL_STATS_BKTS, q);
4220 if (blk_mq_alloc_ctxs(q))
4223 /* init q->mq_kobj and sw queues' kobjects */
4224 blk_mq_sysfs_init(q);
4226 INIT_LIST_HEAD(&q->unused_hctx_list);
4227 spin_lock_init(&q->unused_hctx_lock);
4229 xa_init(&q->hctx_table);
4231 blk_mq_realloc_hw_ctxs(set, q);
4232 if (!q->nr_hw_queues)
4235 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4236 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4240 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4241 blk_mq_update_poll_flag(q);
4243 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4244 INIT_LIST_HEAD(&q->requeue_list);
4245 spin_lock_init(&q->requeue_lock);
4247 q->nr_requests = set->queue_depth;
4250 * Default to classic polling
4252 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
4254 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4255 blk_mq_add_queue_tag_set(set, q);
4256 blk_mq_map_swqueue(q);
4262 blk_stat_free_callback(q->poll_cb);
4268 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4270 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4271 void blk_mq_exit_queue(struct request_queue *q)
4273 struct blk_mq_tag_set *set = q->tag_set;
4275 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4276 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4277 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4278 blk_mq_del_queue_tag_set(q);
4281 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4285 if (blk_mq_is_shared_tags(set->flags)) {
4286 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4289 if (!set->shared_tags)
4293 for (i = 0; i < set->nr_hw_queues; i++) {
4294 if (!__blk_mq_alloc_map_and_rqs(set, i))
4303 __blk_mq_free_map_and_rqs(set, i);
4305 if (blk_mq_is_shared_tags(set->flags)) {
4306 blk_mq_free_map_and_rqs(set, set->shared_tags,
4307 BLK_MQ_NO_HCTX_IDX);
4314 * Allocate the request maps associated with this tag_set. Note that this
4315 * may reduce the depth asked for, if memory is tight. set->queue_depth
4316 * will be updated to reflect the allocated depth.
4318 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4323 depth = set->queue_depth;
4325 err = __blk_mq_alloc_rq_maps(set);
4329 set->queue_depth >>= 1;
4330 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4334 } while (set->queue_depth);
4336 if (!set->queue_depth || err) {
4337 pr_err("blk-mq: failed to allocate request map\n");
4341 if (depth != set->queue_depth)
4342 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4343 depth, set->queue_depth);
4348 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4351 * blk_mq_map_queues() and multiple .map_queues() implementations
4352 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4353 * number of hardware queues.
4355 if (set->nr_maps == 1)
4356 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4358 if (set->ops->map_queues && !is_kdump_kernel()) {
4362 * transport .map_queues is usually done in the following
4365 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4366 * mask = get_cpu_mask(queue)
4367 * for_each_cpu(cpu, mask)
4368 * set->map[x].mq_map[cpu] = queue;
4371 * When we need to remap, the table has to be cleared for
4372 * killing stale mapping since one CPU may not be mapped
4375 for (i = 0; i < set->nr_maps; i++)
4376 blk_mq_clear_mq_map(&set->map[i]);
4378 set->ops->map_queues(set);
4380 BUG_ON(set->nr_maps > 1);
4381 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4385 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4386 int cur_nr_hw_queues, int new_nr_hw_queues)
4388 struct blk_mq_tags **new_tags;
4390 if (cur_nr_hw_queues >= new_nr_hw_queues)
4393 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4394 GFP_KERNEL, set->numa_node);
4399 memcpy(new_tags, set->tags, cur_nr_hw_queues *
4400 sizeof(*set->tags));
4402 set->tags = new_tags;
4403 set->nr_hw_queues = new_nr_hw_queues;
4408 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
4409 int new_nr_hw_queues)
4411 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
4415 * Alloc a tag set to be associated with one or more request queues.
4416 * May fail with EINVAL for various error conditions. May adjust the
4417 * requested depth down, if it's too large. In that case, the set
4418 * value will be stored in set->queue_depth.
4420 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4424 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4426 if (!set->nr_hw_queues)
4428 if (!set->queue_depth)
4430 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4433 if (!set->ops->queue_rq)
4436 if (!set->ops->get_budget ^ !set->ops->put_budget)
4439 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4440 pr_info("blk-mq: reduced tag depth to %u\n",
4442 set->queue_depth = BLK_MQ_MAX_DEPTH;
4447 else if (set->nr_maps > HCTX_MAX_TYPES)
4451 * If a crashdump is active, then we are potentially in a very
4452 * memory constrained environment. Limit us to 1 queue and
4453 * 64 tags to prevent using too much memory.
4455 if (is_kdump_kernel()) {
4456 set->nr_hw_queues = 1;
4458 set->queue_depth = min(64U, set->queue_depth);
4461 * There is no use for more h/w queues than cpus if we just have
4464 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4465 set->nr_hw_queues = nr_cpu_ids;
4467 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
4471 for (i = 0; i < set->nr_maps; i++) {
4472 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4473 sizeof(set->map[i].mq_map[0]),
4474 GFP_KERNEL, set->numa_node);
4475 if (!set->map[i].mq_map)
4476 goto out_free_mq_map;
4477 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4480 blk_mq_update_queue_map(set);
4482 ret = blk_mq_alloc_set_map_and_rqs(set);
4484 goto out_free_mq_map;
4486 mutex_init(&set->tag_list_lock);
4487 INIT_LIST_HEAD(&set->tag_list);
4492 for (i = 0; i < set->nr_maps; i++) {
4493 kfree(set->map[i].mq_map);
4494 set->map[i].mq_map = NULL;
4500 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4502 /* allocate and initialize a tagset for a simple single-queue device */
4503 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4504 const struct blk_mq_ops *ops, unsigned int queue_depth,
4505 unsigned int set_flags)
4507 memset(set, 0, sizeof(*set));
4509 set->nr_hw_queues = 1;
4511 set->queue_depth = queue_depth;
4512 set->numa_node = NUMA_NO_NODE;
4513 set->flags = set_flags;
4514 return blk_mq_alloc_tag_set(set);
4516 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4518 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4522 for (i = 0; i < set->nr_hw_queues; i++)
4523 __blk_mq_free_map_and_rqs(set, i);
4525 if (blk_mq_is_shared_tags(set->flags)) {
4526 blk_mq_free_map_and_rqs(set, set->shared_tags,
4527 BLK_MQ_NO_HCTX_IDX);
4530 for (j = 0; j < set->nr_maps; j++) {
4531 kfree(set->map[j].mq_map);
4532 set->map[j].mq_map = NULL;
4538 EXPORT_SYMBOL(blk_mq_free_tag_set);
4540 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4542 struct blk_mq_tag_set *set = q->tag_set;
4543 struct blk_mq_hw_ctx *hctx;
4550 if (q->nr_requests == nr)
4553 blk_mq_freeze_queue(q);
4554 blk_mq_quiesce_queue(q);
4557 queue_for_each_hw_ctx(q, hctx, i) {
4561 * If we're using an MQ scheduler, just update the scheduler
4562 * queue depth. This is similar to what the old code would do.
4564 if (hctx->sched_tags) {
4565 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4568 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4573 if (q->elevator && q->elevator->type->ops.depth_updated)
4574 q->elevator->type->ops.depth_updated(hctx);
4577 q->nr_requests = nr;
4578 if (blk_mq_is_shared_tags(set->flags)) {
4580 blk_mq_tag_update_sched_shared_tags(q);
4582 blk_mq_tag_resize_shared_tags(set, nr);
4586 blk_mq_unquiesce_queue(q);
4587 blk_mq_unfreeze_queue(q);
4593 * request_queue and elevator_type pair.
4594 * It is just used by __blk_mq_update_nr_hw_queues to cache
4595 * the elevator_type associated with a request_queue.
4597 struct blk_mq_qe_pair {
4598 struct list_head node;
4599 struct request_queue *q;
4600 struct elevator_type *type;
4604 * Cache the elevator_type in qe pair list and switch the
4605 * io scheduler to 'none'
4607 static bool blk_mq_elv_switch_none(struct list_head *head,
4608 struct request_queue *q)
4610 struct blk_mq_qe_pair *qe;
4615 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4619 /* q->elevator needs protection from ->sysfs_lock */
4620 mutex_lock(&q->sysfs_lock);
4622 INIT_LIST_HEAD(&qe->node);
4624 qe->type = q->elevator->type;
4625 list_add(&qe->node, head);
4628 * After elevator_switch, the previous elevator_queue will be
4629 * released by elevator_release. The reference of the io scheduler
4630 * module get by elevator_get will also be put. So we need to get
4631 * a reference of the io scheduler module here to prevent it to be
4634 __module_get(qe->type->elevator_owner);
4635 elevator_switch(q, NULL);
4636 mutex_unlock(&q->sysfs_lock);
4641 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4642 struct request_queue *q)
4644 struct blk_mq_qe_pair *qe;
4646 list_for_each_entry(qe, head, node)
4653 static void blk_mq_elv_switch_back(struct list_head *head,
4654 struct request_queue *q)
4656 struct blk_mq_qe_pair *qe;
4657 struct elevator_type *t;
4659 qe = blk_lookup_qe_pair(head, q);
4663 list_del(&qe->node);
4666 mutex_lock(&q->sysfs_lock);
4667 elevator_switch(q, t);
4668 mutex_unlock(&q->sysfs_lock);
4671 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4674 struct request_queue *q;
4676 int prev_nr_hw_queues;
4678 lockdep_assert_held(&set->tag_list_lock);
4680 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4681 nr_hw_queues = nr_cpu_ids;
4682 if (nr_hw_queues < 1)
4684 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4687 list_for_each_entry(q, &set->tag_list, tag_set_list)
4688 blk_mq_freeze_queue(q);
4690 * Switch IO scheduler to 'none', cleaning up the data associated
4691 * with the previous scheduler. We will switch back once we are done
4692 * updating the new sw to hw queue mappings.
4694 list_for_each_entry(q, &set->tag_list, tag_set_list)
4695 if (!blk_mq_elv_switch_none(&head, q))
4698 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4699 blk_mq_debugfs_unregister_hctxs(q);
4700 blk_mq_sysfs_unregister_hctxs(q);
4703 prev_nr_hw_queues = set->nr_hw_queues;
4704 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
4708 set->nr_hw_queues = nr_hw_queues;
4710 blk_mq_update_queue_map(set);
4711 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4712 blk_mq_realloc_hw_ctxs(set, q);
4713 blk_mq_update_poll_flag(q);
4714 if (q->nr_hw_queues != set->nr_hw_queues) {
4715 int i = prev_nr_hw_queues;
4717 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4718 nr_hw_queues, prev_nr_hw_queues);
4719 for (; i < set->nr_hw_queues; i++)
4720 __blk_mq_free_map_and_rqs(set, i);
4722 set->nr_hw_queues = prev_nr_hw_queues;
4723 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4726 blk_mq_map_swqueue(q);
4730 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4731 blk_mq_sysfs_register_hctxs(q);
4732 blk_mq_debugfs_register_hctxs(q);
4736 list_for_each_entry(q, &set->tag_list, tag_set_list)
4737 blk_mq_elv_switch_back(&head, q);
4739 list_for_each_entry(q, &set->tag_list, tag_set_list)
4740 blk_mq_unfreeze_queue(q);
4743 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4745 mutex_lock(&set->tag_list_lock);
4746 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4747 mutex_unlock(&set->tag_list_lock);
4749 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4751 /* Enable polling stats and return whether they were already enabled. */
4752 static bool blk_poll_stats_enable(struct request_queue *q)
4757 return blk_stats_alloc_enable(q);
4760 static void blk_mq_poll_stats_start(struct request_queue *q)
4763 * We don't arm the callback if polling stats are not enabled or the
4764 * callback is already active.
4766 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
4769 blk_stat_activate_msecs(q->poll_cb, 100);
4772 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4774 struct request_queue *q = cb->data;
4777 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4778 if (cb->stat[bucket].nr_samples)
4779 q->poll_stat[bucket] = cb->stat[bucket];
4783 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4786 unsigned long ret = 0;
4790 * If stats collection isn't on, don't sleep but turn it on for
4793 if (!blk_poll_stats_enable(q))
4797 * As an optimistic guess, use half of the mean service time
4798 * for this type of request. We can (and should) make this smarter.
4799 * For instance, if the completion latencies are tight, we can
4800 * get closer than just half the mean. This is especially
4801 * important on devices where the completion latencies are longer
4802 * than ~10 usec. We do use the stats for the relevant IO size
4803 * if available which does lead to better estimates.
4805 bucket = blk_mq_poll_stats_bkt(rq);
4809 if (q->poll_stat[bucket].nr_samples)
4810 ret = (q->poll_stat[bucket].mean + 1) / 2;
4815 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4817 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4818 struct request *rq = blk_qc_to_rq(hctx, qc);
4819 struct hrtimer_sleeper hs;
4820 enum hrtimer_mode mode;
4825 * If a request has completed on queue that uses an I/O scheduler, we
4826 * won't get back a request from blk_qc_to_rq.
4828 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4832 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4834 * 0: use half of prev avg
4835 * >0: use this specific value
4837 if (q->poll_nsec > 0)
4838 nsecs = q->poll_nsec;
4840 nsecs = blk_mq_poll_nsecs(q, rq);
4845 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4848 * This will be replaced with the stats tracking code, using
4849 * 'avg_completion_time / 2' as the pre-sleep target.
4853 mode = HRTIMER_MODE_REL;
4854 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4855 hrtimer_set_expires(&hs.timer, kt);
4858 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4860 set_current_state(TASK_UNINTERRUPTIBLE);
4861 hrtimer_sleeper_start_expires(&hs, mode);
4864 hrtimer_cancel(&hs.timer);
4865 mode = HRTIMER_MODE_ABS;
4866 } while (hs.task && !signal_pending(current));
4868 __set_current_state(TASK_RUNNING);
4869 destroy_hrtimer_on_stack(&hs.timer);
4872 * If we sleep, have the caller restart the poll loop to reset the
4873 * state. Like for the other success return cases, the caller is
4874 * responsible for checking if the IO completed. If the IO isn't
4875 * complete, we'll get called again and will go straight to the busy
4881 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4882 struct io_comp_batch *iob, unsigned int flags)
4884 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4885 long state = get_current_state();
4889 ret = q->mq_ops->poll(hctx, iob);
4891 __set_current_state(TASK_RUNNING);
4895 if (signal_pending_state(state, current))
4896 __set_current_state(TASK_RUNNING);
4897 if (task_is_running(current))
4900 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4903 } while (!need_resched());
4905 __set_current_state(TASK_RUNNING);
4909 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4912 if (!(flags & BLK_POLL_NOSLEEP) &&
4913 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4914 if (blk_mq_poll_hybrid(q, cookie))
4917 return blk_mq_poll_classic(q, cookie, iob, flags);
4920 unsigned int blk_mq_rq_cpu(struct request *rq)
4922 return rq->mq_ctx->cpu;
4924 EXPORT_SYMBOL(blk_mq_rq_cpu);
4926 void blk_mq_cancel_work_sync(struct request_queue *q)
4928 if (queue_is_mq(q)) {
4929 struct blk_mq_hw_ctx *hctx;
4932 cancel_delayed_work_sync(&q->requeue_work);
4934 queue_for_each_hw_ctx(q, hctx, i)
4935 cancel_delayed_work_sync(&hctx->run_work);
4939 static int __init blk_mq_init(void)
4943 for_each_possible_cpu(i)
4944 init_llist_head(&per_cpu(blk_cpu_done, i));
4945 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4947 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4948 "block/softirq:dead", NULL,
4949 blk_softirq_cpu_dead);
4950 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4951 blk_mq_hctx_notify_dead);
4952 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4953 blk_mq_hctx_notify_online,
4954 blk_mq_hctx_notify_offline);
4957 subsys_initcall(blk_mq_init);