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
257 * @set: tag_set to wait on
259 * Note: it is driver's responsibility for making sure that quiesce has
260 * been started on or more of the request_queues of the tag_set. This
261 * function only waits for the quiesce on those request_queues that had
262 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
264 void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
266 if (set->flags & BLK_MQ_F_BLOCKING)
267 synchronize_srcu(set->srcu);
271 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
274 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
277 * Note: this function does not prevent that the struct request end_io()
278 * callback function is invoked. Once this function is returned, we make
279 * sure no dispatch can happen until the queue is unquiesced via
280 * blk_mq_unquiesce_queue().
282 void blk_mq_quiesce_queue(struct request_queue *q)
284 blk_mq_quiesce_queue_nowait(q);
285 /* nothing to wait for non-mq queues */
287 blk_mq_wait_quiesce_done(q->tag_set);
289 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
292 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
295 * This function recovers queue into the state before quiescing
296 * which is done by blk_mq_quiesce_queue.
298 void blk_mq_unquiesce_queue(struct request_queue *q)
301 bool run_queue = false;
303 spin_lock_irqsave(&q->queue_lock, flags);
304 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
306 } else if (!--q->quiesce_depth) {
307 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
310 spin_unlock_irqrestore(&q->queue_lock, flags);
312 /* dispatch requests which are inserted during quiescing */
314 blk_mq_run_hw_queues(q, true);
316 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
318 void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
320 struct request_queue *q;
322 mutex_lock(&set->tag_list_lock);
323 list_for_each_entry(q, &set->tag_list, tag_set_list) {
324 if (!blk_queue_skip_tagset_quiesce(q))
325 blk_mq_quiesce_queue_nowait(q);
327 blk_mq_wait_quiesce_done(set);
328 mutex_unlock(&set->tag_list_lock);
330 EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
332 void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
334 struct request_queue *q;
336 mutex_lock(&set->tag_list_lock);
337 list_for_each_entry(q, &set->tag_list, tag_set_list) {
338 if (!blk_queue_skip_tagset_quiesce(q))
339 blk_mq_unquiesce_queue(q);
341 mutex_unlock(&set->tag_list_lock);
343 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
345 void blk_mq_wake_waiters(struct request_queue *q)
347 struct blk_mq_hw_ctx *hctx;
350 queue_for_each_hw_ctx(q, hctx, i)
351 if (blk_mq_hw_queue_mapped(hctx))
352 blk_mq_tag_wakeup_all(hctx->tags, true);
355 void blk_rq_init(struct request_queue *q, struct request *rq)
357 memset(rq, 0, sizeof(*rq));
359 INIT_LIST_HEAD(&rq->queuelist);
361 rq->__sector = (sector_t) -1;
362 INIT_HLIST_NODE(&rq->hash);
363 RB_CLEAR_NODE(&rq->rb_node);
364 rq->tag = BLK_MQ_NO_TAG;
365 rq->internal_tag = BLK_MQ_NO_TAG;
366 rq->start_time_ns = ktime_get_ns();
368 blk_crypto_rq_set_defaults(rq);
370 EXPORT_SYMBOL(blk_rq_init);
372 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
373 struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns)
375 struct blk_mq_ctx *ctx = data->ctx;
376 struct blk_mq_hw_ctx *hctx = data->hctx;
377 struct request_queue *q = data->q;
378 struct request *rq = tags->static_rqs[tag];
383 rq->cmd_flags = data->cmd_flags;
385 if (data->flags & BLK_MQ_REQ_PM)
386 data->rq_flags |= RQF_PM;
387 if (blk_queue_io_stat(q))
388 data->rq_flags |= RQF_IO_STAT;
389 rq->rq_flags = data->rq_flags;
391 if (!(data->rq_flags & RQF_ELV)) {
393 rq->internal_tag = BLK_MQ_NO_TAG;
395 rq->tag = BLK_MQ_NO_TAG;
396 rq->internal_tag = tag;
400 if (blk_mq_need_time_stamp(rq))
401 rq->start_time_ns = ktime_get_ns();
403 rq->start_time_ns = 0;
405 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
406 rq->alloc_time_ns = alloc_time_ns;
408 rq->io_start_time_ns = 0;
409 rq->stats_sectors = 0;
410 rq->nr_phys_segments = 0;
411 #if defined(CONFIG_BLK_DEV_INTEGRITY)
412 rq->nr_integrity_segments = 0;
415 rq->end_io_data = NULL;
417 blk_crypto_rq_set_defaults(rq);
418 INIT_LIST_HEAD(&rq->queuelist);
419 /* tag was already set */
420 WRITE_ONCE(rq->deadline, 0);
423 if (rq->rq_flags & RQF_ELV) {
424 struct elevator_queue *e = data->q->elevator;
426 INIT_HLIST_NODE(&rq->hash);
427 RB_CLEAR_NODE(&rq->rb_node);
429 if (!op_is_flush(data->cmd_flags) &&
430 e->type->ops.prepare_request) {
431 e->type->ops.prepare_request(rq);
432 rq->rq_flags |= RQF_ELVPRIV;
439 static inline struct request *
440 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data,
443 unsigned int tag, tag_offset;
444 struct blk_mq_tags *tags;
446 unsigned long tag_mask;
449 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
450 if (unlikely(!tag_mask))
453 tags = blk_mq_tags_from_data(data);
454 for (i = 0; tag_mask; i++) {
455 if (!(tag_mask & (1UL << i)))
457 tag = tag_offset + i;
458 prefetch(tags->static_rqs[tag]);
459 tag_mask &= ~(1UL << i);
460 rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns);
461 rq_list_add(data->cached_rq, rq);
464 /* caller already holds a reference, add for remainder */
465 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
468 return rq_list_pop(data->cached_rq);
471 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
473 struct request_queue *q = data->q;
474 u64 alloc_time_ns = 0;
478 /* alloc_time includes depth and tag waits */
479 if (blk_queue_rq_alloc_time(q))
480 alloc_time_ns = ktime_get_ns();
482 if (data->cmd_flags & REQ_NOWAIT)
483 data->flags |= BLK_MQ_REQ_NOWAIT;
486 struct elevator_queue *e = q->elevator;
488 data->rq_flags |= RQF_ELV;
491 * Flush/passthrough requests are special and go directly to the
492 * dispatch list. Don't include reserved tags in the
493 * limiting, as it isn't useful.
495 if (!op_is_flush(data->cmd_flags) &&
496 !blk_op_is_passthrough(data->cmd_flags) &&
497 e->type->ops.limit_depth &&
498 !(data->flags & BLK_MQ_REQ_RESERVED))
499 e->type->ops.limit_depth(data->cmd_flags, data);
503 data->ctx = blk_mq_get_ctx(q);
504 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
505 if (!(data->rq_flags & RQF_ELV))
506 blk_mq_tag_busy(data->hctx);
508 if (data->flags & BLK_MQ_REQ_RESERVED)
509 data->rq_flags |= RQF_RESV;
512 * Try batched alloc if we want more than 1 tag.
514 if (data->nr_tags > 1) {
515 rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns);
522 * Waiting allocations only fail because of an inactive hctx. In that
523 * case just retry the hctx assignment and tag allocation as CPU hotplug
524 * should have migrated us to an online CPU by now.
526 tag = blk_mq_get_tag(data);
527 if (tag == BLK_MQ_NO_TAG) {
528 if (data->flags & BLK_MQ_REQ_NOWAIT)
531 * Give up the CPU and sleep for a random short time to
532 * ensure that thread using a realtime scheduling class
533 * are migrated off the CPU, and thus off the hctx that
540 return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag,
544 static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
545 struct blk_plug *plug,
547 blk_mq_req_flags_t flags)
549 struct blk_mq_alloc_data data = {
553 .nr_tags = plug->nr_ios,
554 .cached_rq = &plug->cached_rq,
558 if (blk_queue_enter(q, flags))
563 rq = __blk_mq_alloc_requests(&data);
569 static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
571 blk_mq_req_flags_t flags)
573 struct blk_plug *plug = current->plug;
579 if (rq_list_empty(plug->cached_rq)) {
580 if (plug->nr_ios == 1)
582 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
586 rq = rq_list_peek(&plug->cached_rq);
587 if (!rq || rq->q != q)
590 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
592 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
595 plug->cached_rq = rq_list_next(rq);
599 INIT_LIST_HEAD(&rq->queuelist);
603 struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
604 blk_mq_req_flags_t flags)
608 rq = blk_mq_alloc_cached_request(q, opf, flags);
610 struct blk_mq_alloc_data data = {
618 ret = blk_queue_enter(q, flags);
622 rq = __blk_mq_alloc_requests(&data);
627 rq->__sector = (sector_t) -1;
628 rq->bio = rq->biotail = NULL;
632 return ERR_PTR(-EWOULDBLOCK);
634 EXPORT_SYMBOL(blk_mq_alloc_request);
636 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
637 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
639 struct blk_mq_alloc_data data = {
645 u64 alloc_time_ns = 0;
651 /* alloc_time includes depth and tag waits */
652 if (blk_queue_rq_alloc_time(q))
653 alloc_time_ns = ktime_get_ns();
656 * If the tag allocator sleeps we could get an allocation for a
657 * different hardware context. No need to complicate the low level
658 * allocator for this for the rare use case of a command tied to
661 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
662 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
663 return ERR_PTR(-EINVAL);
665 if (hctx_idx >= q->nr_hw_queues)
666 return ERR_PTR(-EIO);
668 ret = blk_queue_enter(q, flags);
673 * Check if the hardware context is actually mapped to anything.
674 * If not tell the caller that it should skip this queue.
677 data.hctx = xa_load(&q->hctx_table, hctx_idx);
678 if (!blk_mq_hw_queue_mapped(data.hctx))
680 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
681 if (cpu >= nr_cpu_ids)
683 data.ctx = __blk_mq_get_ctx(q, cpu);
686 blk_mq_tag_busy(data.hctx);
688 data.rq_flags |= RQF_ELV;
690 if (flags & BLK_MQ_REQ_RESERVED)
691 data.rq_flags |= RQF_RESV;
694 tag = blk_mq_get_tag(&data);
695 if (tag == BLK_MQ_NO_TAG)
697 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag,
700 rq->__sector = (sector_t) -1;
701 rq->bio = rq->biotail = NULL;
708 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
710 static void __blk_mq_free_request(struct request *rq)
712 struct request_queue *q = rq->q;
713 struct blk_mq_ctx *ctx = rq->mq_ctx;
714 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
715 const int sched_tag = rq->internal_tag;
717 blk_crypto_free_request(rq);
718 blk_pm_mark_last_busy(rq);
720 if (rq->tag != BLK_MQ_NO_TAG)
721 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
722 if (sched_tag != BLK_MQ_NO_TAG)
723 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
724 blk_mq_sched_restart(hctx);
728 void blk_mq_free_request(struct request *rq)
730 struct request_queue *q = rq->q;
731 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
733 if ((rq->rq_flags & RQF_ELVPRIV) &&
734 q->elevator->type->ops.finish_request)
735 q->elevator->type->ops.finish_request(rq);
737 if (rq->rq_flags & RQF_MQ_INFLIGHT)
738 __blk_mq_dec_active_requests(hctx);
740 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
741 laptop_io_completion(q->disk->bdi);
745 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
746 if (req_ref_put_and_test(rq))
747 __blk_mq_free_request(rq);
749 EXPORT_SYMBOL_GPL(blk_mq_free_request);
751 void blk_mq_free_plug_rqs(struct blk_plug *plug)
755 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
756 blk_mq_free_request(rq);
759 void blk_dump_rq_flags(struct request *rq, char *msg)
761 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
762 rq->q->disk ? rq->q->disk->disk_name : "?",
763 (__force unsigned long long) rq->cmd_flags);
765 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
766 (unsigned long long)blk_rq_pos(rq),
767 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
768 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
769 rq->bio, rq->biotail, blk_rq_bytes(rq));
771 EXPORT_SYMBOL(blk_dump_rq_flags);
773 static void req_bio_endio(struct request *rq, struct bio *bio,
774 unsigned int nbytes, blk_status_t error)
776 if (unlikely(error)) {
777 bio->bi_status = error;
778 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
780 * Partial zone append completions cannot be supported as the
781 * BIO fragments may end up not being written sequentially.
783 if (bio->bi_iter.bi_size != nbytes)
784 bio->bi_status = BLK_STS_IOERR;
786 bio->bi_iter.bi_sector = rq->__sector;
789 bio_advance(bio, nbytes);
791 if (unlikely(rq->rq_flags & RQF_QUIET))
792 bio_set_flag(bio, BIO_QUIET);
793 /* don't actually finish bio if it's part of flush sequence */
794 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
798 static void blk_account_io_completion(struct request *req, unsigned int bytes)
800 if (req->part && blk_do_io_stat(req)) {
801 const int sgrp = op_stat_group(req_op(req));
804 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
809 static void blk_print_req_error(struct request *req, blk_status_t status)
811 printk_ratelimited(KERN_ERR
812 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
813 "phys_seg %u prio class %u\n",
814 blk_status_to_str(status),
815 req->q->disk ? req->q->disk->disk_name : "?",
816 blk_rq_pos(req), (__force u32)req_op(req),
817 blk_op_str(req_op(req)),
818 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
819 req->nr_phys_segments,
820 IOPRIO_PRIO_CLASS(req->ioprio));
824 * Fully end IO on a request. Does not support partial completions, or
827 static void blk_complete_request(struct request *req)
829 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
830 int total_bytes = blk_rq_bytes(req);
831 struct bio *bio = req->bio;
833 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
838 #ifdef CONFIG_BLK_DEV_INTEGRITY
839 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
840 req->q->integrity.profile->complete_fn(req, total_bytes);
843 blk_account_io_completion(req, total_bytes);
846 struct bio *next = bio->bi_next;
848 /* Completion has already been traced */
849 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
851 if (req_op(req) == REQ_OP_ZONE_APPEND)
852 bio->bi_iter.bi_sector = req->__sector;
860 * Reset counters so that the request stacking driver
861 * can find how many bytes remain in the request
871 * blk_update_request - Complete multiple bytes without completing the request
872 * @req: the request being processed
873 * @error: block status code
874 * @nr_bytes: number of bytes to complete for @req
877 * Ends I/O on a number of bytes attached to @req, but doesn't complete
878 * the request structure even if @req doesn't have leftover.
879 * If @req has leftover, sets it up for the next range of segments.
881 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
882 * %false return from this function.
885 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
886 * except in the consistency check at the end of this function.
889 * %false - this request doesn't have any more data
890 * %true - this request has more data
892 bool blk_update_request(struct request *req, blk_status_t error,
893 unsigned int nr_bytes)
897 trace_block_rq_complete(req, error, nr_bytes);
902 #ifdef CONFIG_BLK_DEV_INTEGRITY
903 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
905 req->q->integrity.profile->complete_fn(req, nr_bytes);
908 if (unlikely(error && !blk_rq_is_passthrough(req) &&
909 !(req->rq_flags & RQF_QUIET)) &&
910 !test_bit(GD_DEAD, &req->q->disk->state)) {
911 blk_print_req_error(req, error);
912 trace_block_rq_error(req, error, nr_bytes);
915 blk_account_io_completion(req, nr_bytes);
919 struct bio *bio = req->bio;
920 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
922 if (bio_bytes == bio->bi_iter.bi_size)
923 req->bio = bio->bi_next;
925 /* Completion has already been traced */
926 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
927 req_bio_endio(req, bio, bio_bytes, error);
929 total_bytes += bio_bytes;
930 nr_bytes -= bio_bytes;
941 * Reset counters so that the request stacking driver
942 * can find how many bytes remain in the request
949 req->__data_len -= total_bytes;
951 /* update sector only for requests with clear definition of sector */
952 if (!blk_rq_is_passthrough(req))
953 req->__sector += total_bytes >> 9;
955 /* mixed attributes always follow the first bio */
956 if (req->rq_flags & RQF_MIXED_MERGE) {
957 req->cmd_flags &= ~REQ_FAILFAST_MASK;
958 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
961 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
963 * If total number of sectors is less than the first segment
964 * size, something has gone terribly wrong.
966 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
967 blk_dump_rq_flags(req, "request botched");
968 req->__data_len = blk_rq_cur_bytes(req);
971 /* recalculate the number of segments */
972 req->nr_phys_segments = blk_recalc_rq_segments(req);
977 EXPORT_SYMBOL_GPL(blk_update_request);
979 static void __blk_account_io_done(struct request *req, u64 now)
981 const int sgrp = op_stat_group(req_op(req));
984 update_io_ticks(req->part, jiffies, true);
985 part_stat_inc(req->part, ios[sgrp]);
986 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
990 static inline void blk_account_io_done(struct request *req, u64 now)
993 * Account IO completion. flush_rq isn't accounted as a
994 * normal IO on queueing nor completion. Accounting the
995 * containing request is enough.
997 if (blk_do_io_stat(req) && req->part &&
998 !(req->rq_flags & RQF_FLUSH_SEQ))
999 __blk_account_io_done(req, now);
1002 static void __blk_account_io_start(struct request *rq)
1005 * All non-passthrough requests are created from a bio with one
1006 * exception: when a flush command that is part of a flush sequence
1007 * generated by the state machine in blk-flush.c is cloned onto the
1008 * lower device by dm-multipath we can get here without a bio.
1011 rq->part = rq->bio->bi_bdev;
1013 rq->part = rq->q->disk->part0;
1016 update_io_ticks(rq->part, jiffies, false);
1020 static inline void blk_account_io_start(struct request *req)
1022 if (blk_do_io_stat(req))
1023 __blk_account_io_start(req);
1026 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1028 if (rq->rq_flags & RQF_STATS) {
1029 blk_mq_poll_stats_start(rq->q);
1030 blk_stat_add(rq, now);
1033 blk_mq_sched_completed_request(rq, now);
1034 blk_account_io_done(rq, now);
1037 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1039 if (blk_mq_need_time_stamp(rq))
1040 __blk_mq_end_request_acct(rq, ktime_get_ns());
1043 rq_qos_done(rq->q, rq);
1044 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1045 blk_mq_free_request(rq);
1047 blk_mq_free_request(rq);
1050 EXPORT_SYMBOL(__blk_mq_end_request);
1052 void blk_mq_end_request(struct request *rq, blk_status_t error)
1054 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1056 __blk_mq_end_request(rq, error);
1058 EXPORT_SYMBOL(blk_mq_end_request);
1060 #define TAG_COMP_BATCH 32
1062 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1063 int *tag_array, int nr_tags)
1065 struct request_queue *q = hctx->queue;
1068 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1069 * update hctx->nr_active in batch
1071 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1072 __blk_mq_sub_active_requests(hctx, nr_tags);
1074 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1075 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1078 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1080 int tags[TAG_COMP_BATCH], nr_tags = 0;
1081 struct blk_mq_hw_ctx *cur_hctx = NULL;
1086 now = ktime_get_ns();
1088 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1090 prefetch(rq->rq_next);
1092 blk_complete_request(rq);
1094 __blk_mq_end_request_acct(rq, now);
1096 rq_qos_done(rq->q, rq);
1099 * If end_io handler returns NONE, then it still has
1100 * ownership of the request.
1102 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1105 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1106 if (!req_ref_put_and_test(rq))
1109 blk_crypto_free_request(rq);
1110 blk_pm_mark_last_busy(rq);
1112 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1114 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1116 cur_hctx = rq->mq_hctx;
1118 tags[nr_tags++] = rq->tag;
1122 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1124 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1126 static void blk_complete_reqs(struct llist_head *list)
1128 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1129 struct request *rq, *next;
1131 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1132 rq->q->mq_ops->complete(rq);
1135 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1137 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1140 static int blk_softirq_cpu_dead(unsigned int cpu)
1142 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1146 static void __blk_mq_complete_request_remote(void *data)
1148 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1151 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1153 int cpu = raw_smp_processor_id();
1155 if (!IS_ENABLED(CONFIG_SMP) ||
1156 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1159 * With force threaded interrupts enabled, raising softirq from an SMP
1160 * function call will always result in waking the ksoftirqd thread.
1161 * This is probably worse than completing the request on a different
1164 if (force_irqthreads())
1167 /* same CPU or cache domain? Complete locally */
1168 if (cpu == rq->mq_ctx->cpu ||
1169 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1170 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1173 /* don't try to IPI to an offline CPU */
1174 return cpu_online(rq->mq_ctx->cpu);
1177 static void blk_mq_complete_send_ipi(struct request *rq)
1179 struct llist_head *list;
1182 cpu = rq->mq_ctx->cpu;
1183 list = &per_cpu(blk_cpu_done, cpu);
1184 if (llist_add(&rq->ipi_list, list)) {
1185 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1186 smp_call_function_single_async(cpu, &rq->csd);
1190 static void blk_mq_raise_softirq(struct request *rq)
1192 struct llist_head *list;
1195 list = this_cpu_ptr(&blk_cpu_done);
1196 if (llist_add(&rq->ipi_list, list))
1197 raise_softirq(BLOCK_SOFTIRQ);
1201 bool blk_mq_complete_request_remote(struct request *rq)
1203 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1206 * For request which hctx has only one ctx mapping,
1207 * or a polled request, always complete locally,
1208 * it's pointless to redirect the completion.
1210 if (rq->mq_hctx->nr_ctx == 1 ||
1211 rq->cmd_flags & REQ_POLLED)
1214 if (blk_mq_complete_need_ipi(rq)) {
1215 blk_mq_complete_send_ipi(rq);
1219 if (rq->q->nr_hw_queues == 1) {
1220 blk_mq_raise_softirq(rq);
1225 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1228 * blk_mq_complete_request - end I/O on a request
1229 * @rq: the request being processed
1232 * Complete a request by scheduling the ->complete_rq operation.
1234 void blk_mq_complete_request(struct request *rq)
1236 if (!blk_mq_complete_request_remote(rq))
1237 rq->q->mq_ops->complete(rq);
1239 EXPORT_SYMBOL(blk_mq_complete_request);
1242 * blk_mq_start_request - Start processing a request
1243 * @rq: Pointer to request to be started
1245 * Function used by device drivers to notify the block layer that a request
1246 * is going to be processed now, so blk layer can do proper initializations
1247 * such as starting the timeout timer.
1249 void blk_mq_start_request(struct request *rq)
1251 struct request_queue *q = rq->q;
1253 trace_block_rq_issue(rq);
1255 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1256 rq->io_start_time_ns = ktime_get_ns();
1257 rq->stats_sectors = blk_rq_sectors(rq);
1258 rq->rq_flags |= RQF_STATS;
1259 rq_qos_issue(q, rq);
1262 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1265 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1267 #ifdef CONFIG_BLK_DEV_INTEGRITY
1268 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1269 q->integrity.profile->prepare_fn(rq);
1271 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1272 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1274 EXPORT_SYMBOL(blk_mq_start_request);
1277 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1278 * queues. This is important for md arrays to benefit from merging
1281 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1283 if (plug->multiple_queues)
1284 return BLK_MAX_REQUEST_COUNT * 2;
1285 return BLK_MAX_REQUEST_COUNT;
1288 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1290 struct request *last = rq_list_peek(&plug->mq_list);
1292 if (!plug->rq_count) {
1293 trace_block_plug(rq->q);
1294 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1295 (!blk_queue_nomerges(rq->q) &&
1296 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1297 blk_mq_flush_plug_list(plug, false);
1299 trace_block_plug(rq->q);
1302 if (!plug->multiple_queues && last && last->q != rq->q)
1303 plug->multiple_queues = true;
1304 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1305 plug->has_elevator = true;
1307 rq_list_add(&plug->mq_list, rq);
1312 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1313 * @rq: request to insert
1314 * @at_head: insert request at head or tail of queue
1317 * Insert a fully prepared request at the back of the I/O scheduler queue
1318 * for execution. Don't wait for completion.
1321 * This function will invoke @done directly if the queue is dead.
1323 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1325 WARN_ON(irqs_disabled());
1326 WARN_ON(!blk_rq_is_passthrough(rq));
1328 blk_account_io_start(rq);
1331 * As plugging can be enabled for passthrough requests on a zoned
1332 * device, directly accessing the plug instead of using blk_mq_plug()
1333 * should not have any consequences.
1336 blk_add_rq_to_plug(current->plug, rq);
1338 blk_mq_sched_insert_request(rq, at_head, true, false);
1340 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1342 struct blk_rq_wait {
1343 struct completion done;
1347 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1349 struct blk_rq_wait *wait = rq->end_io_data;
1352 complete(&wait->done);
1353 return RQ_END_IO_NONE;
1356 bool blk_rq_is_poll(struct request *rq)
1360 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1364 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1366 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1369 blk_mq_poll(rq->q, blk_rq_to_qc(rq), NULL, 0);
1371 } while (!completion_done(wait));
1375 * blk_execute_rq - insert a request into queue for execution
1376 * @rq: request to insert
1377 * @at_head: insert request at head or tail of queue
1380 * Insert a fully prepared request at the back of the I/O scheduler queue
1381 * for execution and wait for completion.
1382 * Return: The blk_status_t result provided to blk_mq_end_request().
1384 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1386 struct blk_rq_wait wait = {
1387 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1390 WARN_ON(irqs_disabled());
1391 WARN_ON(!blk_rq_is_passthrough(rq));
1393 rq->end_io_data = &wait;
1394 rq->end_io = blk_end_sync_rq;
1396 blk_account_io_start(rq);
1397 blk_mq_sched_insert_request(rq, at_head, true, false);
1399 if (blk_rq_is_poll(rq)) {
1400 blk_rq_poll_completion(rq, &wait.done);
1403 * Prevent hang_check timer from firing at us during very long
1406 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1409 while (!wait_for_completion_io_timeout(&wait.done,
1410 hang_check * (HZ/2)))
1413 wait_for_completion_io(&wait.done);
1418 EXPORT_SYMBOL(blk_execute_rq);
1420 static void __blk_mq_requeue_request(struct request *rq)
1422 struct request_queue *q = rq->q;
1424 blk_mq_put_driver_tag(rq);
1426 trace_block_rq_requeue(rq);
1427 rq_qos_requeue(q, rq);
1429 if (blk_mq_request_started(rq)) {
1430 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1431 rq->rq_flags &= ~RQF_TIMED_OUT;
1435 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1437 __blk_mq_requeue_request(rq);
1439 /* this request will be re-inserted to io scheduler queue */
1440 blk_mq_sched_requeue_request(rq);
1442 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1444 EXPORT_SYMBOL(blk_mq_requeue_request);
1446 static void blk_mq_requeue_work(struct work_struct *work)
1448 struct request_queue *q =
1449 container_of(work, struct request_queue, requeue_work.work);
1451 struct request *rq, *next;
1453 spin_lock_irq(&q->requeue_lock);
1454 list_splice_init(&q->requeue_list, &rq_list);
1455 spin_unlock_irq(&q->requeue_lock);
1457 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1458 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1461 rq->rq_flags &= ~RQF_SOFTBARRIER;
1462 list_del_init(&rq->queuelist);
1464 * If RQF_DONTPREP, rq has contained some driver specific
1465 * data, so insert it to hctx dispatch list to avoid any
1468 if (rq->rq_flags & RQF_DONTPREP)
1469 blk_mq_request_bypass_insert(rq, false, false);
1471 blk_mq_sched_insert_request(rq, true, false, false);
1474 while (!list_empty(&rq_list)) {
1475 rq = list_entry(rq_list.next, struct request, queuelist);
1476 list_del_init(&rq->queuelist);
1477 blk_mq_sched_insert_request(rq, false, false, false);
1480 blk_mq_run_hw_queues(q, false);
1483 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1484 bool kick_requeue_list)
1486 struct request_queue *q = rq->q;
1487 unsigned long flags;
1490 * We abuse this flag that is otherwise used by the I/O scheduler to
1491 * request head insertion from the workqueue.
1493 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1495 spin_lock_irqsave(&q->requeue_lock, flags);
1497 rq->rq_flags |= RQF_SOFTBARRIER;
1498 list_add(&rq->queuelist, &q->requeue_list);
1500 list_add_tail(&rq->queuelist, &q->requeue_list);
1502 spin_unlock_irqrestore(&q->requeue_lock, flags);
1504 if (kick_requeue_list)
1505 blk_mq_kick_requeue_list(q);
1508 void blk_mq_kick_requeue_list(struct request_queue *q)
1510 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1512 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1514 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1515 unsigned long msecs)
1517 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1518 msecs_to_jiffies(msecs));
1520 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1522 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1525 * If we find a request that isn't idle we know the queue is busy
1526 * as it's checked in the iter.
1527 * Return false to stop the iteration.
1529 if (blk_mq_request_started(rq)) {
1539 bool blk_mq_queue_inflight(struct request_queue *q)
1543 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1546 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1548 static void blk_mq_rq_timed_out(struct request *req)
1550 req->rq_flags |= RQF_TIMED_OUT;
1551 if (req->q->mq_ops->timeout) {
1552 enum blk_eh_timer_return ret;
1554 ret = req->q->mq_ops->timeout(req);
1555 if (ret == BLK_EH_DONE)
1557 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1563 struct blk_expired_data {
1564 bool has_timedout_rq;
1566 unsigned long timeout_start;
1569 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1571 unsigned long deadline;
1573 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1575 if (rq->rq_flags & RQF_TIMED_OUT)
1578 deadline = READ_ONCE(rq->deadline);
1579 if (time_after_eq(expired->timeout_start, deadline))
1582 if (expired->next == 0)
1583 expired->next = deadline;
1584 else if (time_after(expired->next, deadline))
1585 expired->next = deadline;
1589 void blk_mq_put_rq_ref(struct request *rq)
1591 if (is_flush_rq(rq)) {
1592 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1593 blk_mq_free_request(rq);
1594 } else if (req_ref_put_and_test(rq)) {
1595 __blk_mq_free_request(rq);
1599 static bool blk_mq_check_expired(struct request *rq, void *priv)
1601 struct blk_expired_data *expired = priv;
1604 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1605 * be reallocated underneath the timeout handler's processing, then
1606 * the expire check is reliable. If the request is not expired, then
1607 * it was completed and reallocated as a new request after returning
1608 * from blk_mq_check_expired().
1610 if (blk_mq_req_expired(rq, expired)) {
1611 expired->has_timedout_rq = true;
1617 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1619 struct blk_expired_data *expired = priv;
1621 if (blk_mq_req_expired(rq, expired))
1622 blk_mq_rq_timed_out(rq);
1626 static void blk_mq_timeout_work(struct work_struct *work)
1628 struct request_queue *q =
1629 container_of(work, struct request_queue, timeout_work);
1630 struct blk_expired_data expired = {
1631 .timeout_start = jiffies,
1633 struct blk_mq_hw_ctx *hctx;
1636 /* A deadlock might occur if a request is stuck requiring a
1637 * timeout at the same time a queue freeze is waiting
1638 * completion, since the timeout code would not be able to
1639 * acquire the queue reference here.
1641 * That's why we don't use blk_queue_enter here; instead, we use
1642 * percpu_ref_tryget directly, because we need to be able to
1643 * obtain a reference even in the short window between the queue
1644 * starting to freeze, by dropping the first reference in
1645 * blk_freeze_queue_start, and the moment the last request is
1646 * consumed, marked by the instant q_usage_counter reaches
1649 if (!percpu_ref_tryget(&q->q_usage_counter))
1652 /* check if there is any timed-out request */
1653 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1654 if (expired.has_timedout_rq) {
1656 * Before walking tags, we must ensure any submit started
1657 * before the current time has finished. Since the submit
1658 * uses srcu or rcu, wait for a synchronization point to
1659 * ensure all running submits have finished
1661 blk_mq_wait_quiesce_done(q->tag_set);
1664 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1667 if (expired.next != 0) {
1668 mod_timer(&q->timeout, expired.next);
1671 * Request timeouts are handled as a forward rolling timer. If
1672 * we end up here it means that no requests are pending and
1673 * also that no request has been pending for a while. Mark
1674 * each hctx as idle.
1676 queue_for_each_hw_ctx(q, hctx, i) {
1677 /* the hctx may be unmapped, so check it here */
1678 if (blk_mq_hw_queue_mapped(hctx))
1679 blk_mq_tag_idle(hctx);
1685 struct flush_busy_ctx_data {
1686 struct blk_mq_hw_ctx *hctx;
1687 struct list_head *list;
1690 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1692 struct flush_busy_ctx_data *flush_data = data;
1693 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1694 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1695 enum hctx_type type = hctx->type;
1697 spin_lock(&ctx->lock);
1698 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1699 sbitmap_clear_bit(sb, bitnr);
1700 spin_unlock(&ctx->lock);
1705 * Process software queues that have been marked busy, splicing them
1706 * to the for-dispatch
1708 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1710 struct flush_busy_ctx_data data = {
1715 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1717 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1719 struct dispatch_rq_data {
1720 struct blk_mq_hw_ctx *hctx;
1724 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1727 struct dispatch_rq_data *dispatch_data = data;
1728 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1729 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1730 enum hctx_type type = hctx->type;
1732 spin_lock(&ctx->lock);
1733 if (!list_empty(&ctx->rq_lists[type])) {
1734 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1735 list_del_init(&dispatch_data->rq->queuelist);
1736 if (list_empty(&ctx->rq_lists[type]))
1737 sbitmap_clear_bit(sb, bitnr);
1739 spin_unlock(&ctx->lock);
1741 return !dispatch_data->rq;
1744 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1745 struct blk_mq_ctx *start)
1747 unsigned off = start ? start->index_hw[hctx->type] : 0;
1748 struct dispatch_rq_data data = {
1753 __sbitmap_for_each_set(&hctx->ctx_map, off,
1754 dispatch_rq_from_ctx, &data);
1759 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1761 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1762 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1765 blk_mq_tag_busy(rq->mq_hctx);
1767 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1768 bt = &rq->mq_hctx->tags->breserved_tags;
1771 if (!hctx_may_queue(rq->mq_hctx, bt))
1775 tag = __sbitmap_queue_get(bt);
1776 if (tag == BLK_MQ_NO_TAG)
1779 rq->tag = tag + tag_offset;
1783 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1785 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1788 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1789 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1790 rq->rq_flags |= RQF_MQ_INFLIGHT;
1791 __blk_mq_inc_active_requests(hctx);
1793 hctx->tags->rqs[rq->tag] = rq;
1797 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1798 int flags, void *key)
1800 struct blk_mq_hw_ctx *hctx;
1802 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1804 spin_lock(&hctx->dispatch_wait_lock);
1805 if (!list_empty(&wait->entry)) {
1806 struct sbitmap_queue *sbq;
1808 list_del_init(&wait->entry);
1809 sbq = &hctx->tags->bitmap_tags;
1810 atomic_dec(&sbq->ws_active);
1812 spin_unlock(&hctx->dispatch_wait_lock);
1814 blk_mq_run_hw_queue(hctx, true);
1819 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1820 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1821 * restart. For both cases, take care to check the condition again after
1822 * marking us as waiting.
1824 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1827 struct sbitmap_queue *sbq;
1828 struct wait_queue_head *wq;
1829 wait_queue_entry_t *wait;
1832 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1833 !(blk_mq_is_shared_tags(hctx->flags))) {
1834 blk_mq_sched_mark_restart_hctx(hctx);
1837 * It's possible that a tag was freed in the window between the
1838 * allocation failure and adding the hardware queue to the wait
1841 * Don't clear RESTART here, someone else could have set it.
1842 * At most this will cost an extra queue run.
1844 return blk_mq_get_driver_tag(rq);
1847 wait = &hctx->dispatch_wait;
1848 if (!list_empty_careful(&wait->entry))
1851 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1852 sbq = &hctx->tags->breserved_tags;
1854 sbq = &hctx->tags->bitmap_tags;
1855 wq = &bt_wait_ptr(sbq, hctx)->wait;
1857 spin_lock_irq(&wq->lock);
1858 spin_lock(&hctx->dispatch_wait_lock);
1859 if (!list_empty(&wait->entry)) {
1860 spin_unlock(&hctx->dispatch_wait_lock);
1861 spin_unlock_irq(&wq->lock);
1865 atomic_inc(&sbq->ws_active);
1866 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1867 __add_wait_queue(wq, wait);
1870 * It's possible that a tag was freed in the window between the
1871 * allocation failure and adding the hardware queue to the wait
1874 ret = blk_mq_get_driver_tag(rq);
1876 spin_unlock(&hctx->dispatch_wait_lock);
1877 spin_unlock_irq(&wq->lock);
1882 * We got a tag, remove ourselves from the wait queue to ensure
1883 * someone else gets the wakeup.
1885 list_del_init(&wait->entry);
1886 atomic_dec(&sbq->ws_active);
1887 spin_unlock(&hctx->dispatch_wait_lock);
1888 spin_unlock_irq(&wq->lock);
1893 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1894 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1896 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1897 * - EWMA is one simple way to compute running average value
1898 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1899 * - take 4 as factor for avoiding to get too small(0) result, and this
1900 * factor doesn't matter because EWMA decreases exponentially
1902 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1906 ewma = hctx->dispatch_busy;
1911 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1913 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1914 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1916 hctx->dispatch_busy = ewma;
1919 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1921 static void blk_mq_handle_dev_resource(struct request *rq,
1922 struct list_head *list)
1924 list_add(&rq->queuelist, list);
1925 __blk_mq_requeue_request(rq);
1928 static void blk_mq_handle_zone_resource(struct request *rq,
1929 struct list_head *zone_list)
1932 * If we end up here it is because we cannot dispatch a request to a
1933 * specific zone due to LLD level zone-write locking or other zone
1934 * related resource not being available. In this case, set the request
1935 * aside in zone_list for retrying it later.
1937 list_add(&rq->queuelist, zone_list);
1938 __blk_mq_requeue_request(rq);
1941 enum prep_dispatch {
1943 PREP_DISPATCH_NO_TAG,
1944 PREP_DISPATCH_NO_BUDGET,
1947 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1950 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1951 int budget_token = -1;
1954 budget_token = blk_mq_get_dispatch_budget(rq->q);
1955 if (budget_token < 0) {
1956 blk_mq_put_driver_tag(rq);
1957 return PREP_DISPATCH_NO_BUDGET;
1959 blk_mq_set_rq_budget_token(rq, budget_token);
1962 if (!blk_mq_get_driver_tag(rq)) {
1964 * The initial allocation attempt failed, so we need to
1965 * rerun the hardware queue when a tag is freed. The
1966 * waitqueue takes care of that. If the queue is run
1967 * before we add this entry back on the dispatch list,
1968 * we'll re-run it below.
1970 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1972 * All budgets not got from this function will be put
1973 * together during handling partial dispatch
1976 blk_mq_put_dispatch_budget(rq->q, budget_token);
1977 return PREP_DISPATCH_NO_TAG;
1981 return PREP_DISPATCH_OK;
1984 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1985 static void blk_mq_release_budgets(struct request_queue *q,
1986 struct list_head *list)
1990 list_for_each_entry(rq, list, queuelist) {
1991 int budget_token = blk_mq_get_rq_budget_token(rq);
1993 if (budget_token >= 0)
1994 blk_mq_put_dispatch_budget(q, budget_token);
1999 * blk_mq_commit_rqs will notify driver using bd->last that there is no
2000 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
2002 * Attention, we should explicitly call this in unusual cases:
2003 * 1) did not queue everything initially scheduled to queue
2004 * 2) the last attempt to queue a request failed
2006 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
2009 if (hctx->queue->mq_ops->commit_rqs && queued) {
2010 trace_block_unplug(hctx->queue, queued, !from_schedule);
2011 hctx->queue->mq_ops->commit_rqs(hctx);
2016 * Returns true if we did some work AND can potentially do more.
2018 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
2019 unsigned int nr_budgets)
2021 enum prep_dispatch prep;
2022 struct request_queue *q = hctx->queue;
2025 blk_status_t ret = BLK_STS_OK;
2026 LIST_HEAD(zone_list);
2027 bool needs_resource = false;
2029 if (list_empty(list))
2033 * Now process all the entries, sending them to the driver.
2037 struct blk_mq_queue_data bd;
2039 rq = list_first_entry(list, struct request, queuelist);
2041 WARN_ON_ONCE(hctx != rq->mq_hctx);
2042 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2043 if (prep != PREP_DISPATCH_OK)
2046 list_del_init(&rq->queuelist);
2049 bd.last = list_empty(list);
2052 * once the request is queued to lld, no need to cover the
2057 ret = q->mq_ops->queue_rq(hctx, &bd);
2062 case BLK_STS_RESOURCE:
2063 needs_resource = true;
2065 case BLK_STS_DEV_RESOURCE:
2066 blk_mq_handle_dev_resource(rq, list);
2068 case BLK_STS_ZONE_RESOURCE:
2070 * Move the request to zone_list and keep going through
2071 * the dispatch list to find more requests the drive can
2074 blk_mq_handle_zone_resource(rq, &zone_list);
2075 needs_resource = true;
2078 blk_mq_end_request(rq, ret);
2080 } while (!list_empty(list));
2082 if (!list_empty(&zone_list))
2083 list_splice_tail_init(&zone_list, list);
2085 /* If we didn't flush the entire list, we could have told the driver
2086 * there was more coming, but that turned out to be a lie.
2088 if (!list_empty(list) || ret != BLK_STS_OK)
2089 blk_mq_commit_rqs(hctx, queued, false);
2092 * Any items that need requeuing? Stuff them into hctx->dispatch,
2093 * that is where we will continue on next queue run.
2095 if (!list_empty(list)) {
2097 /* For non-shared tags, the RESTART check will suffice */
2098 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2099 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2100 blk_mq_is_shared_tags(hctx->flags));
2103 blk_mq_release_budgets(q, list);
2105 spin_lock(&hctx->lock);
2106 list_splice_tail_init(list, &hctx->dispatch);
2107 spin_unlock(&hctx->lock);
2110 * Order adding requests to hctx->dispatch and checking
2111 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2112 * in blk_mq_sched_restart(). Avoid restart code path to
2113 * miss the new added requests to hctx->dispatch, meantime
2114 * SCHED_RESTART is observed here.
2119 * If SCHED_RESTART was set by the caller of this function and
2120 * it is no longer set that means that it was cleared by another
2121 * thread and hence that a queue rerun is needed.
2123 * If 'no_tag' is set, that means that we failed getting
2124 * a driver tag with an I/O scheduler attached. If our dispatch
2125 * waitqueue is no longer active, ensure that we run the queue
2126 * AFTER adding our entries back to the list.
2128 * If no I/O scheduler has been configured it is possible that
2129 * the hardware queue got stopped and restarted before requests
2130 * were pushed back onto the dispatch list. Rerun the queue to
2131 * avoid starvation. Notes:
2132 * - blk_mq_run_hw_queue() checks whether or not a queue has
2133 * been stopped before rerunning a queue.
2134 * - Some but not all block drivers stop a queue before
2135 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2138 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2139 * bit is set, run queue after a delay to avoid IO stalls
2140 * that could otherwise occur if the queue is idle. We'll do
2141 * similar if we couldn't get budget or couldn't lock a zone
2142 * and SCHED_RESTART is set.
2144 needs_restart = blk_mq_sched_needs_restart(hctx);
2145 if (prep == PREP_DISPATCH_NO_BUDGET)
2146 needs_resource = true;
2147 if (!needs_restart ||
2148 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2149 blk_mq_run_hw_queue(hctx, true);
2150 else if (needs_resource)
2151 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2153 blk_mq_update_dispatch_busy(hctx, true);
2157 blk_mq_update_dispatch_busy(hctx, false);
2162 * __blk_mq_run_hw_queue - Run a hardware queue.
2163 * @hctx: Pointer to the hardware queue to run.
2165 * Send pending requests to the hardware.
2167 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
2170 * We can't run the queue inline with ints disabled. Ensure that
2171 * we catch bad users of this early.
2173 WARN_ON_ONCE(in_interrupt());
2175 blk_mq_run_dispatch_ops(hctx->queue,
2176 blk_mq_sched_dispatch_requests(hctx));
2179 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2181 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2183 if (cpu >= nr_cpu_ids)
2184 cpu = cpumask_first(hctx->cpumask);
2189 * It'd be great if the workqueue API had a way to pass
2190 * in a mask and had some smarts for more clever placement.
2191 * For now we just round-robin here, switching for every
2192 * BLK_MQ_CPU_WORK_BATCH queued items.
2194 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2197 int next_cpu = hctx->next_cpu;
2199 if (hctx->queue->nr_hw_queues == 1)
2200 return WORK_CPU_UNBOUND;
2202 if (--hctx->next_cpu_batch <= 0) {
2204 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2206 if (next_cpu >= nr_cpu_ids)
2207 next_cpu = blk_mq_first_mapped_cpu(hctx);
2208 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2212 * Do unbound schedule if we can't find a online CPU for this hctx,
2213 * and it should only happen in the path of handling CPU DEAD.
2215 if (!cpu_online(next_cpu)) {
2222 * Make sure to re-select CPU next time once after CPUs
2223 * in hctx->cpumask become online again.
2225 hctx->next_cpu = next_cpu;
2226 hctx->next_cpu_batch = 1;
2227 return WORK_CPU_UNBOUND;
2230 hctx->next_cpu = next_cpu;
2235 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2236 * @hctx: Pointer to the hardware queue to run.
2237 * @async: If we want to run the queue asynchronously.
2238 * @msecs: Milliseconds of delay to wait before running the queue.
2240 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2241 * with a delay of @msecs.
2243 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2244 unsigned long msecs)
2246 if (unlikely(blk_mq_hctx_stopped(hctx)))
2249 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2250 if (cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2251 __blk_mq_run_hw_queue(hctx);
2256 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2257 msecs_to_jiffies(msecs));
2261 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2262 * @hctx: Pointer to the hardware queue to run.
2263 * @msecs: Milliseconds of delay to wait before running the queue.
2265 * Run a hardware queue asynchronously with a delay of @msecs.
2267 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2269 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2271 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2274 * blk_mq_run_hw_queue - Start to run a hardware queue.
2275 * @hctx: Pointer to the hardware queue to run.
2276 * @async: If we want to run the queue asynchronously.
2278 * Check if the request queue is not in a quiesced state and if there are
2279 * pending requests to be sent. If this is true, run the queue to send requests
2282 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2287 * When queue is quiesced, we may be switching io scheduler, or
2288 * updating nr_hw_queues, or other things, and we can't run queue
2289 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2291 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2294 __blk_mq_run_dispatch_ops(hctx->queue, false,
2295 need_run = !blk_queue_quiesced(hctx->queue) &&
2296 blk_mq_hctx_has_pending(hctx));
2299 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2301 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2304 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2307 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2309 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2311 * If the IO scheduler does not respect hardware queues when
2312 * dispatching, we just don't bother with multiple HW queues and
2313 * dispatch from hctx for the current CPU since running multiple queues
2314 * just causes lock contention inside the scheduler and pointless cache
2317 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2319 if (!blk_mq_hctx_stopped(hctx))
2325 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2326 * @q: Pointer to the request queue to run.
2327 * @async: If we want to run the queue asynchronously.
2329 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2331 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2335 if (blk_queue_sq_sched(q))
2336 sq_hctx = blk_mq_get_sq_hctx(q);
2337 queue_for_each_hw_ctx(q, hctx, i) {
2338 if (blk_mq_hctx_stopped(hctx))
2341 * Dispatch from this hctx either if there's no hctx preferred
2342 * by IO scheduler or if it has requests that bypass the
2345 if (!sq_hctx || sq_hctx == hctx ||
2346 !list_empty_careful(&hctx->dispatch))
2347 blk_mq_run_hw_queue(hctx, async);
2350 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2353 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2354 * @q: Pointer to the request queue to run.
2355 * @msecs: Milliseconds of delay to wait before running the queues.
2357 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2359 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2363 if (blk_queue_sq_sched(q))
2364 sq_hctx = blk_mq_get_sq_hctx(q);
2365 queue_for_each_hw_ctx(q, hctx, i) {
2366 if (blk_mq_hctx_stopped(hctx))
2369 * If there is already a run_work pending, leave the
2370 * pending delay untouched. Otherwise, a hctx can stall
2371 * if another hctx is re-delaying the other's work
2372 * before the work executes.
2374 if (delayed_work_pending(&hctx->run_work))
2377 * Dispatch from this hctx either if there's no hctx preferred
2378 * by IO scheduler or if it has requests that bypass the
2381 if (!sq_hctx || sq_hctx == hctx ||
2382 !list_empty_careful(&hctx->dispatch))
2383 blk_mq_delay_run_hw_queue(hctx, msecs);
2386 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2389 * This function is often used for pausing .queue_rq() by driver when
2390 * there isn't enough resource or some conditions aren't satisfied, and
2391 * BLK_STS_RESOURCE is usually returned.
2393 * We do not guarantee that dispatch can be drained or blocked
2394 * after blk_mq_stop_hw_queue() returns. Please use
2395 * blk_mq_quiesce_queue() for that requirement.
2397 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2399 cancel_delayed_work(&hctx->run_work);
2401 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2403 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2406 * This function is often used for pausing .queue_rq() by driver when
2407 * there isn't enough resource or some conditions aren't satisfied, and
2408 * BLK_STS_RESOURCE is usually returned.
2410 * We do not guarantee that dispatch can be drained or blocked
2411 * after blk_mq_stop_hw_queues() returns. Please use
2412 * blk_mq_quiesce_queue() for that requirement.
2414 void blk_mq_stop_hw_queues(struct request_queue *q)
2416 struct blk_mq_hw_ctx *hctx;
2419 queue_for_each_hw_ctx(q, hctx, i)
2420 blk_mq_stop_hw_queue(hctx);
2422 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2424 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2426 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2428 blk_mq_run_hw_queue(hctx, false);
2430 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2432 void blk_mq_start_hw_queues(struct request_queue *q)
2434 struct blk_mq_hw_ctx *hctx;
2437 queue_for_each_hw_ctx(q, hctx, i)
2438 blk_mq_start_hw_queue(hctx);
2440 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2442 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2444 if (!blk_mq_hctx_stopped(hctx))
2447 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2448 blk_mq_run_hw_queue(hctx, async);
2450 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2452 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2454 struct blk_mq_hw_ctx *hctx;
2457 queue_for_each_hw_ctx(q, hctx, i)
2458 blk_mq_start_stopped_hw_queue(hctx, async);
2460 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2462 static void blk_mq_run_work_fn(struct work_struct *work)
2464 struct blk_mq_hw_ctx *hctx;
2466 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2469 * If we are stopped, don't run the queue.
2471 if (blk_mq_hctx_stopped(hctx))
2474 __blk_mq_run_hw_queue(hctx);
2477 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2481 struct blk_mq_ctx *ctx = rq->mq_ctx;
2482 enum hctx_type type = hctx->type;
2484 lockdep_assert_held(&ctx->lock);
2486 trace_block_rq_insert(rq);
2489 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2491 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2494 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2497 struct blk_mq_ctx *ctx = rq->mq_ctx;
2499 lockdep_assert_held(&ctx->lock);
2501 __blk_mq_insert_req_list(hctx, rq, at_head);
2502 blk_mq_hctx_mark_pending(hctx, ctx);
2506 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2507 * @rq: Pointer to request to be inserted.
2508 * @at_head: true if the request should be inserted at the head of the list.
2509 * @run_queue: If we should run the hardware queue after inserting the request.
2511 * Should only be used carefully, when the caller knows we want to
2512 * bypass a potential IO scheduler on the target device.
2514 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2517 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2519 spin_lock(&hctx->lock);
2521 list_add(&rq->queuelist, &hctx->dispatch);
2523 list_add_tail(&rq->queuelist, &hctx->dispatch);
2524 spin_unlock(&hctx->lock);
2527 blk_mq_run_hw_queue(hctx, false);
2530 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2531 struct list_head *list)
2535 enum hctx_type type = hctx->type;
2538 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2541 list_for_each_entry(rq, list, queuelist) {
2542 BUG_ON(rq->mq_ctx != ctx);
2543 trace_block_rq_insert(rq);
2546 spin_lock(&ctx->lock);
2547 list_splice_tail_init(list, &ctx->rq_lists[type]);
2548 blk_mq_hctx_mark_pending(hctx, ctx);
2549 spin_unlock(&ctx->lock);
2552 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2553 unsigned int nr_segs)
2557 if (bio->bi_opf & REQ_RAHEAD)
2558 rq->cmd_flags |= REQ_FAILFAST_MASK;
2560 rq->__sector = bio->bi_iter.bi_sector;
2561 blk_rq_bio_prep(rq, bio, nr_segs);
2563 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2564 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2567 blk_account_io_start(rq);
2570 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2571 struct request *rq, bool last)
2573 struct request_queue *q = rq->q;
2574 struct blk_mq_queue_data bd = {
2581 * For OK queue, we are done. For error, caller may kill it.
2582 * Any other error (busy), just add it to our list as we
2583 * previously would have done.
2585 ret = q->mq_ops->queue_rq(hctx, &bd);
2588 blk_mq_update_dispatch_busy(hctx, false);
2590 case BLK_STS_RESOURCE:
2591 case BLK_STS_DEV_RESOURCE:
2592 blk_mq_update_dispatch_busy(hctx, true);
2593 __blk_mq_requeue_request(rq);
2596 blk_mq_update_dispatch_busy(hctx, false);
2603 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2605 bool bypass_insert, bool last)
2607 struct request_queue *q = rq->q;
2608 bool run_queue = true;
2612 * RCU or SRCU read lock is needed before checking quiesced flag.
2614 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2615 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2616 * and avoid driver to try to dispatch again.
2618 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2620 bypass_insert = false;
2624 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2627 budget_token = blk_mq_get_dispatch_budget(q);
2628 if (budget_token < 0)
2631 blk_mq_set_rq_budget_token(rq, budget_token);
2633 if (!blk_mq_get_driver_tag(rq)) {
2634 blk_mq_put_dispatch_budget(q, budget_token);
2638 return __blk_mq_issue_directly(hctx, rq, last);
2641 return BLK_STS_RESOURCE;
2643 blk_mq_sched_insert_request(rq, false, run_queue, false);
2649 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2650 * @hctx: Pointer of the associated hardware queue.
2651 * @rq: Pointer to request to be sent.
2653 * If the device has enough resources to accept a new request now, send the
2654 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2655 * we can try send it another time in the future. Requests inserted at this
2656 * queue have higher priority.
2658 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2662 __blk_mq_try_issue_directly(hctx, rq, false, true);
2664 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2665 blk_mq_request_bypass_insert(rq, false, true);
2666 else if (ret != BLK_STS_OK)
2667 blk_mq_end_request(rq, ret);
2670 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2672 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2675 static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2677 struct blk_mq_hw_ctx *hctx = NULL;
2680 blk_status_t ret = BLK_STS_OK;
2682 while ((rq = rq_list_pop(&plug->mq_list))) {
2683 bool last = rq_list_empty(plug->mq_list);
2685 if (hctx != rq->mq_hctx) {
2687 blk_mq_commit_rqs(hctx, queued, false);
2693 ret = blk_mq_request_issue_directly(rq, last);
2698 case BLK_STS_RESOURCE:
2699 case BLK_STS_DEV_RESOURCE:
2700 blk_mq_request_bypass_insert(rq, false, true);
2703 blk_mq_end_request(rq, ret);
2709 if (ret != BLK_STS_OK)
2710 blk_mq_commit_rqs(hctx, queued, false);
2713 static void __blk_mq_flush_plug_list(struct request_queue *q,
2714 struct blk_plug *plug)
2716 if (blk_queue_quiesced(q))
2718 q->mq_ops->queue_rqs(&plug->mq_list);
2721 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2723 struct blk_mq_hw_ctx *this_hctx = NULL;
2724 struct blk_mq_ctx *this_ctx = NULL;
2725 struct request *requeue_list = NULL;
2726 struct request **requeue_lastp = &requeue_list;
2727 unsigned int depth = 0;
2731 struct request *rq = rq_list_pop(&plug->mq_list);
2734 this_hctx = rq->mq_hctx;
2735 this_ctx = rq->mq_ctx;
2736 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2737 rq_list_add_tail(&requeue_lastp, rq);
2740 list_add(&rq->queuelist, &list);
2742 } while (!rq_list_empty(plug->mq_list));
2744 plug->mq_list = requeue_list;
2745 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2746 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched);
2749 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2753 if (rq_list_empty(plug->mq_list))
2757 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2758 struct request_queue *q;
2760 rq = rq_list_peek(&plug->mq_list);
2764 * Peek first request and see if we have a ->queue_rqs() hook.
2765 * If we do, we can dispatch the whole plug list in one go. We
2766 * already know at this point that all requests belong to the
2767 * same queue, caller must ensure that's the case.
2769 * Since we pass off the full list to the driver at this point,
2770 * we do not increment the active request count for the queue.
2771 * Bypass shared tags for now because of that.
2773 if (q->mq_ops->queue_rqs &&
2774 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2775 blk_mq_run_dispatch_ops(q,
2776 __blk_mq_flush_plug_list(q, plug));
2777 if (rq_list_empty(plug->mq_list))
2781 blk_mq_run_dispatch_ops(q,
2782 blk_mq_plug_issue_direct(plug));
2783 if (rq_list_empty(plug->mq_list))
2788 blk_mq_dispatch_plug_list(plug, from_schedule);
2789 } while (!rq_list_empty(plug->mq_list));
2792 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2793 struct list_head *list)
2796 blk_status_t ret = BLK_STS_OK;
2798 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));
2808 case BLK_STS_RESOURCE:
2809 case BLK_STS_DEV_RESOURCE:
2810 blk_mq_request_bypass_insert(rq, false,
2814 blk_mq_end_request(rq, ret);
2820 if (ret != BLK_STS_OK)
2821 blk_mq_commit_rqs(hctx, queued, false);
2824 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2825 struct bio *bio, unsigned int nr_segs)
2827 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2828 if (blk_attempt_plug_merge(q, bio, nr_segs))
2830 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2836 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2837 struct blk_plug *plug,
2841 struct blk_mq_alloc_data data = {
2844 .cmd_flags = bio->bi_opf,
2848 if (unlikely(bio_queue_enter(bio)))
2851 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2854 rq_qos_throttle(q, bio);
2857 data.nr_tags = plug->nr_ios;
2859 data.cached_rq = &plug->cached_rq;
2862 rq = __blk_mq_alloc_requests(&data);
2865 rq_qos_cleanup(q, bio);
2866 if (bio->bi_opf & REQ_NOWAIT)
2867 bio_wouldblock_error(bio);
2873 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2874 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2877 enum hctx_type type, hctx_type;
2881 rq = rq_list_peek(&plug->cached_rq);
2882 if (!rq || rq->q != q)
2885 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2890 type = blk_mq_get_hctx_type((*bio)->bi_opf);
2891 hctx_type = rq->mq_hctx->type;
2892 if (type != hctx_type &&
2893 !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2895 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2899 * If any qos ->throttle() end up blocking, we will have flushed the
2900 * plug and hence killed the cached_rq list as well. Pop this entry
2901 * before we throttle.
2903 plug->cached_rq = rq_list_next(rq);
2904 rq_qos_throttle(q, *bio);
2906 rq->cmd_flags = (*bio)->bi_opf;
2907 INIT_LIST_HEAD(&rq->queuelist);
2911 static void bio_set_ioprio(struct bio *bio)
2913 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2914 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2915 bio->bi_ioprio = get_current_ioprio();
2916 blkcg_set_ioprio(bio);
2920 * blk_mq_submit_bio - Create and send a request to block device.
2921 * @bio: Bio pointer.
2923 * Builds up a request structure from @q and @bio and send to the device. The
2924 * request may not be queued directly to hardware if:
2925 * * This request can be merged with another one
2926 * * We want to place request at plug queue for possible future merging
2927 * * There is an IO scheduler active at this queue
2929 * It will not queue the request if there is an error with the bio, or at the
2932 void blk_mq_submit_bio(struct bio *bio)
2934 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2935 struct blk_plug *plug = blk_mq_plug(bio);
2936 const int is_sync = op_is_sync(bio->bi_opf);
2938 unsigned int nr_segs = 1;
2941 bio = blk_queue_bounce(bio, q);
2942 if (bio_may_exceed_limits(bio, &q->limits)) {
2943 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2948 if (!bio_integrity_prep(bio))
2951 bio_set_ioprio(bio);
2953 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2957 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2962 trace_block_getrq(bio);
2964 rq_qos_track(q, rq, bio);
2966 blk_mq_bio_to_request(rq, bio, nr_segs);
2968 ret = blk_crypto_init_request(rq);
2969 if (ret != BLK_STS_OK) {
2970 bio->bi_status = ret;
2972 blk_mq_free_request(rq);
2976 if (op_is_flush(bio->bi_opf)) {
2977 blk_insert_flush(rq);
2982 blk_add_rq_to_plug(plug, rq);
2983 else if ((rq->rq_flags & RQF_ELV) ||
2984 (rq->mq_hctx->dispatch_busy &&
2985 (q->nr_hw_queues == 1 || !is_sync)))
2986 blk_mq_sched_insert_request(rq, false, true, true);
2988 blk_mq_run_dispatch_ops(rq->q,
2989 blk_mq_try_issue_directly(rq->mq_hctx, rq));
2992 #ifdef CONFIG_BLK_MQ_STACKING
2994 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
2995 * @rq: the request being queued
2997 blk_status_t blk_insert_cloned_request(struct request *rq)
2999 struct request_queue *q = rq->q;
3000 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3001 unsigned int max_segments = blk_rq_get_max_segments(rq);
3004 if (blk_rq_sectors(rq) > max_sectors) {
3006 * SCSI device does not have a good way to return if
3007 * Write Same/Zero is actually supported. If a device rejects
3008 * a non-read/write command (discard, write same,etc.) the
3009 * low-level device driver will set the relevant queue limit to
3010 * 0 to prevent blk-lib from issuing more of the offending
3011 * operations. Commands queued prior to the queue limit being
3012 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3013 * errors being propagated to upper layers.
3015 if (max_sectors == 0)
3016 return BLK_STS_NOTSUPP;
3018 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3019 __func__, blk_rq_sectors(rq), max_sectors);
3020 return BLK_STS_IOERR;
3024 * The queue settings related to segment counting may differ from the
3027 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3028 if (rq->nr_phys_segments > max_segments) {
3029 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3030 __func__, rq->nr_phys_segments, max_segments);
3031 return BLK_STS_IOERR;
3034 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3035 return BLK_STS_IOERR;
3037 if (blk_crypto_insert_cloned_request(rq))
3038 return BLK_STS_IOERR;
3040 blk_account_io_start(rq);
3043 * Since we have a scheduler attached on the top device,
3044 * bypass a potential scheduler on the bottom device for
3047 blk_mq_run_dispatch_ops(q,
3048 ret = blk_mq_request_issue_directly(rq, true));
3050 blk_account_io_done(rq, ktime_get_ns());
3053 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3056 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3057 * @rq: the clone request to be cleaned up
3060 * Free all bios in @rq for a cloned request.
3062 void blk_rq_unprep_clone(struct request *rq)
3066 while ((bio = rq->bio) != NULL) {
3067 rq->bio = bio->bi_next;
3072 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3075 * blk_rq_prep_clone - Helper function to setup clone request
3076 * @rq: the request to be setup
3077 * @rq_src: original request to be cloned
3078 * @bs: bio_set that bios for clone are allocated from
3079 * @gfp_mask: memory allocation mask for bio
3080 * @bio_ctr: setup function to be called for each clone bio.
3081 * Returns %0 for success, non %0 for failure.
3082 * @data: private data to be passed to @bio_ctr
3085 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3086 * Also, pages which the original bios are pointing to are not copied
3087 * and the cloned bios just point same pages.
3088 * So cloned bios must be completed before original bios, which means
3089 * the caller must complete @rq before @rq_src.
3091 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3092 struct bio_set *bs, gfp_t gfp_mask,
3093 int (*bio_ctr)(struct bio *, struct bio *, void *),
3096 struct bio *bio, *bio_src;
3101 __rq_for_each_bio(bio_src, rq_src) {
3102 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3107 if (bio_ctr && bio_ctr(bio, bio_src, data))
3111 rq->biotail->bi_next = bio;
3114 rq->bio = rq->biotail = bio;
3119 /* Copy attributes of the original request to the clone request. */
3120 rq->__sector = blk_rq_pos(rq_src);
3121 rq->__data_len = blk_rq_bytes(rq_src);
3122 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3123 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3124 rq->special_vec = rq_src->special_vec;
3126 rq->nr_phys_segments = rq_src->nr_phys_segments;
3127 rq->ioprio = rq_src->ioprio;
3129 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3137 blk_rq_unprep_clone(rq);
3141 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3142 #endif /* CONFIG_BLK_MQ_STACKING */
3145 * Steal bios from a request and add them to a bio list.
3146 * The request must not have been partially completed before.
3148 void blk_steal_bios(struct bio_list *list, struct request *rq)
3152 list->tail->bi_next = rq->bio;
3154 list->head = rq->bio;
3155 list->tail = rq->biotail;
3163 EXPORT_SYMBOL_GPL(blk_steal_bios);
3165 static size_t order_to_size(unsigned int order)
3167 return (size_t)PAGE_SIZE << order;
3170 /* called before freeing request pool in @tags */
3171 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3172 struct blk_mq_tags *tags)
3175 unsigned long flags;
3178 * There is no need to clear mapping if driver tags is not initialized
3179 * or the mapping belongs to the driver tags.
3181 if (!drv_tags || drv_tags == tags)
3184 list_for_each_entry(page, &tags->page_list, lru) {
3185 unsigned long start = (unsigned long)page_address(page);
3186 unsigned long end = start + order_to_size(page->private);
3189 for (i = 0; i < drv_tags->nr_tags; i++) {
3190 struct request *rq = drv_tags->rqs[i];
3191 unsigned long rq_addr = (unsigned long)rq;
3193 if (rq_addr >= start && rq_addr < end) {
3194 WARN_ON_ONCE(req_ref_read(rq) != 0);
3195 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3201 * Wait until all pending iteration is done.
3203 * Request reference is cleared and it is guaranteed to be observed
3204 * after the ->lock is released.
3206 spin_lock_irqsave(&drv_tags->lock, flags);
3207 spin_unlock_irqrestore(&drv_tags->lock, flags);
3210 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3211 unsigned int hctx_idx)
3213 struct blk_mq_tags *drv_tags;
3216 if (list_empty(&tags->page_list))
3219 if (blk_mq_is_shared_tags(set->flags))
3220 drv_tags = set->shared_tags;
3222 drv_tags = set->tags[hctx_idx];
3224 if (tags->static_rqs && set->ops->exit_request) {
3227 for (i = 0; i < tags->nr_tags; i++) {
3228 struct request *rq = tags->static_rqs[i];
3232 set->ops->exit_request(set, rq, hctx_idx);
3233 tags->static_rqs[i] = NULL;
3237 blk_mq_clear_rq_mapping(drv_tags, tags);
3239 while (!list_empty(&tags->page_list)) {
3240 page = list_first_entry(&tags->page_list, struct page, lru);
3241 list_del_init(&page->lru);
3243 * Remove kmemleak object previously allocated in
3244 * blk_mq_alloc_rqs().
3246 kmemleak_free(page_address(page));
3247 __free_pages(page, page->private);
3251 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3255 kfree(tags->static_rqs);
3256 tags->static_rqs = NULL;
3258 blk_mq_free_tags(tags);
3261 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3262 unsigned int hctx_idx)
3266 for (i = 0; i < set->nr_maps; i++) {
3267 unsigned int start = set->map[i].queue_offset;
3268 unsigned int end = start + set->map[i].nr_queues;
3270 if (hctx_idx >= start && hctx_idx < end)
3274 if (i >= set->nr_maps)
3275 i = HCTX_TYPE_DEFAULT;
3280 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3281 unsigned int hctx_idx)
3283 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3285 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3288 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3289 unsigned int hctx_idx,
3290 unsigned int nr_tags,
3291 unsigned int reserved_tags)
3293 int node = blk_mq_get_hctx_node(set, hctx_idx);
3294 struct blk_mq_tags *tags;
3296 if (node == NUMA_NO_NODE)
3297 node = set->numa_node;
3299 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3300 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3304 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3305 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3310 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3311 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3313 if (!tags->static_rqs)
3321 blk_mq_free_tags(tags);
3325 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3326 unsigned int hctx_idx, int node)
3330 if (set->ops->init_request) {
3331 ret = set->ops->init_request(set, rq, hctx_idx, node);
3336 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3340 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3341 struct blk_mq_tags *tags,
3342 unsigned int hctx_idx, unsigned int depth)
3344 unsigned int i, j, entries_per_page, max_order = 4;
3345 int node = blk_mq_get_hctx_node(set, hctx_idx);
3346 size_t rq_size, left;
3348 if (node == NUMA_NO_NODE)
3349 node = set->numa_node;
3351 INIT_LIST_HEAD(&tags->page_list);
3354 * rq_size is the size of the request plus driver payload, rounded
3355 * to the cacheline size
3357 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3359 left = rq_size * depth;
3361 for (i = 0; i < depth; ) {
3362 int this_order = max_order;
3367 while (this_order && left < order_to_size(this_order - 1))
3371 page = alloc_pages_node(node,
3372 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3378 if (order_to_size(this_order) < rq_size)
3385 page->private = this_order;
3386 list_add_tail(&page->lru, &tags->page_list);
3388 p = page_address(page);
3390 * Allow kmemleak to scan these pages as they contain pointers
3391 * to additional allocations like via ops->init_request().
3393 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3394 entries_per_page = order_to_size(this_order) / rq_size;
3395 to_do = min(entries_per_page, depth - i);
3396 left -= to_do * rq_size;
3397 for (j = 0; j < to_do; j++) {
3398 struct request *rq = p;
3400 tags->static_rqs[i] = rq;
3401 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3402 tags->static_rqs[i] = NULL;
3413 blk_mq_free_rqs(set, tags, hctx_idx);
3417 struct rq_iter_data {
3418 struct blk_mq_hw_ctx *hctx;
3422 static bool blk_mq_has_request(struct request *rq, void *data)
3424 struct rq_iter_data *iter_data = data;
3426 if (rq->mq_hctx != iter_data->hctx)
3428 iter_data->has_rq = true;
3432 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3434 struct blk_mq_tags *tags = hctx->sched_tags ?
3435 hctx->sched_tags : hctx->tags;
3436 struct rq_iter_data data = {
3440 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3444 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3445 struct blk_mq_hw_ctx *hctx)
3447 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3449 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3454 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3456 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3457 struct blk_mq_hw_ctx, cpuhp_online);
3459 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3460 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3464 * Prevent new request from being allocated on the current hctx.
3466 * The smp_mb__after_atomic() Pairs with the implied barrier in
3467 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3468 * seen once we return from the tag allocator.
3470 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3471 smp_mb__after_atomic();
3474 * Try to grab a reference to the queue and wait for any outstanding
3475 * requests. If we could not grab a reference the queue has been
3476 * frozen and there are no requests.
3478 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3479 while (blk_mq_hctx_has_requests(hctx))
3481 percpu_ref_put(&hctx->queue->q_usage_counter);
3487 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3489 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3490 struct blk_mq_hw_ctx, cpuhp_online);
3492 if (cpumask_test_cpu(cpu, hctx->cpumask))
3493 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3498 * 'cpu' is going away. splice any existing rq_list entries from this
3499 * software queue to the hw queue dispatch list, and ensure that it
3502 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3504 struct blk_mq_hw_ctx *hctx;
3505 struct blk_mq_ctx *ctx;
3507 enum hctx_type type;
3509 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3510 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3513 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3516 spin_lock(&ctx->lock);
3517 if (!list_empty(&ctx->rq_lists[type])) {
3518 list_splice_init(&ctx->rq_lists[type], &tmp);
3519 blk_mq_hctx_clear_pending(hctx, ctx);
3521 spin_unlock(&ctx->lock);
3523 if (list_empty(&tmp))
3526 spin_lock(&hctx->lock);
3527 list_splice_tail_init(&tmp, &hctx->dispatch);
3528 spin_unlock(&hctx->lock);
3530 blk_mq_run_hw_queue(hctx, true);
3534 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3536 if (!(hctx->flags & BLK_MQ_F_STACKING))
3537 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3538 &hctx->cpuhp_online);
3539 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3544 * Before freeing hw queue, clearing the flush request reference in
3545 * tags->rqs[] for avoiding potential UAF.
3547 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3548 unsigned int queue_depth, struct request *flush_rq)
3551 unsigned long flags;
3553 /* The hw queue may not be mapped yet */
3557 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3559 for (i = 0; i < queue_depth; i++)
3560 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3563 * Wait until all pending iteration is done.
3565 * Request reference is cleared and it is guaranteed to be observed
3566 * after the ->lock is released.
3568 spin_lock_irqsave(&tags->lock, flags);
3569 spin_unlock_irqrestore(&tags->lock, flags);
3572 /* hctx->ctxs will be freed in queue's release handler */
3573 static void blk_mq_exit_hctx(struct request_queue *q,
3574 struct blk_mq_tag_set *set,
3575 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3577 struct request *flush_rq = hctx->fq->flush_rq;
3579 if (blk_mq_hw_queue_mapped(hctx))
3580 blk_mq_tag_idle(hctx);
3582 if (blk_queue_init_done(q))
3583 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3584 set->queue_depth, flush_rq);
3585 if (set->ops->exit_request)
3586 set->ops->exit_request(set, flush_rq, hctx_idx);
3588 if (set->ops->exit_hctx)
3589 set->ops->exit_hctx(hctx, hctx_idx);
3591 blk_mq_remove_cpuhp(hctx);
3593 xa_erase(&q->hctx_table, hctx_idx);
3595 spin_lock(&q->unused_hctx_lock);
3596 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3597 spin_unlock(&q->unused_hctx_lock);
3600 static void blk_mq_exit_hw_queues(struct request_queue *q,
3601 struct blk_mq_tag_set *set, int nr_queue)
3603 struct blk_mq_hw_ctx *hctx;
3606 queue_for_each_hw_ctx(q, hctx, i) {
3609 blk_mq_exit_hctx(q, set, hctx, i);
3613 static int blk_mq_init_hctx(struct request_queue *q,
3614 struct blk_mq_tag_set *set,
3615 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3617 hctx->queue_num = hctx_idx;
3619 if (!(hctx->flags & BLK_MQ_F_STACKING))
3620 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3621 &hctx->cpuhp_online);
3622 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3624 hctx->tags = set->tags[hctx_idx];
3626 if (set->ops->init_hctx &&
3627 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3628 goto unregister_cpu_notifier;
3630 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3634 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3640 if (set->ops->exit_request)
3641 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3643 if (set->ops->exit_hctx)
3644 set->ops->exit_hctx(hctx, hctx_idx);
3645 unregister_cpu_notifier:
3646 blk_mq_remove_cpuhp(hctx);
3650 static struct blk_mq_hw_ctx *
3651 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3654 struct blk_mq_hw_ctx *hctx;
3655 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3657 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3659 goto fail_alloc_hctx;
3661 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3664 atomic_set(&hctx->nr_active, 0);
3665 if (node == NUMA_NO_NODE)
3666 node = set->numa_node;
3667 hctx->numa_node = node;
3669 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3670 spin_lock_init(&hctx->lock);
3671 INIT_LIST_HEAD(&hctx->dispatch);
3673 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3675 INIT_LIST_HEAD(&hctx->hctx_list);
3678 * Allocate space for all possible cpus to avoid allocation at
3681 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3686 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3687 gfp, node, false, false))
3691 spin_lock_init(&hctx->dispatch_wait_lock);
3692 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3693 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3695 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3699 blk_mq_hctx_kobj_init(hctx);
3704 sbitmap_free(&hctx->ctx_map);
3708 free_cpumask_var(hctx->cpumask);
3715 static void blk_mq_init_cpu_queues(struct request_queue *q,
3716 unsigned int nr_hw_queues)
3718 struct blk_mq_tag_set *set = q->tag_set;
3721 for_each_possible_cpu(i) {
3722 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3723 struct blk_mq_hw_ctx *hctx;
3727 spin_lock_init(&__ctx->lock);
3728 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3729 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3734 * Set local node, IFF we have more than one hw queue. If
3735 * not, we remain on the home node of the device
3737 for (j = 0; j < set->nr_maps; j++) {
3738 hctx = blk_mq_map_queue_type(q, j, i);
3739 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3740 hctx->numa_node = cpu_to_node(i);
3745 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3746 unsigned int hctx_idx,
3749 struct blk_mq_tags *tags;
3752 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3756 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3758 blk_mq_free_rq_map(tags);
3765 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3768 if (blk_mq_is_shared_tags(set->flags)) {
3769 set->tags[hctx_idx] = set->shared_tags;
3774 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3777 return set->tags[hctx_idx];
3780 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3781 struct blk_mq_tags *tags,
3782 unsigned int hctx_idx)
3785 blk_mq_free_rqs(set, tags, hctx_idx);
3786 blk_mq_free_rq_map(tags);
3790 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3791 unsigned int hctx_idx)
3793 if (!blk_mq_is_shared_tags(set->flags))
3794 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3796 set->tags[hctx_idx] = NULL;
3799 static void blk_mq_map_swqueue(struct request_queue *q)
3801 unsigned int j, hctx_idx;
3803 struct blk_mq_hw_ctx *hctx;
3804 struct blk_mq_ctx *ctx;
3805 struct blk_mq_tag_set *set = q->tag_set;
3807 queue_for_each_hw_ctx(q, hctx, i) {
3808 cpumask_clear(hctx->cpumask);
3810 hctx->dispatch_from = NULL;
3814 * Map software to hardware queues.
3816 * If the cpu isn't present, the cpu is mapped to first hctx.
3818 for_each_possible_cpu(i) {
3820 ctx = per_cpu_ptr(q->queue_ctx, i);
3821 for (j = 0; j < set->nr_maps; j++) {
3822 if (!set->map[j].nr_queues) {
3823 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3824 HCTX_TYPE_DEFAULT, i);
3827 hctx_idx = set->map[j].mq_map[i];
3828 /* unmapped hw queue can be remapped after CPU topo changed */
3829 if (!set->tags[hctx_idx] &&
3830 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3832 * If tags initialization fail for some hctx,
3833 * that hctx won't be brought online. In this
3834 * case, remap the current ctx to hctx[0] which
3835 * is guaranteed to always have tags allocated
3837 set->map[j].mq_map[i] = 0;
3840 hctx = blk_mq_map_queue_type(q, j, i);
3841 ctx->hctxs[j] = hctx;
3843 * If the CPU is already set in the mask, then we've
3844 * mapped this one already. This can happen if
3845 * devices share queues across queue maps.
3847 if (cpumask_test_cpu(i, hctx->cpumask))
3850 cpumask_set_cpu(i, hctx->cpumask);
3852 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3853 hctx->ctxs[hctx->nr_ctx++] = ctx;
3856 * If the nr_ctx type overflows, we have exceeded the
3857 * amount of sw queues we can support.
3859 BUG_ON(!hctx->nr_ctx);
3862 for (; j < HCTX_MAX_TYPES; j++)
3863 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3864 HCTX_TYPE_DEFAULT, i);
3867 queue_for_each_hw_ctx(q, hctx, i) {
3869 * If no software queues are mapped to this hardware queue,
3870 * disable it and free the request entries.
3872 if (!hctx->nr_ctx) {
3873 /* Never unmap queue 0. We need it as a
3874 * fallback in case of a new remap fails
3878 __blk_mq_free_map_and_rqs(set, i);
3884 hctx->tags = set->tags[i];
3885 WARN_ON(!hctx->tags);
3888 * Set the map size to the number of mapped software queues.
3889 * This is more accurate and more efficient than looping
3890 * over all possibly mapped software queues.
3892 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3895 * Initialize batch roundrobin counts
3897 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3898 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3903 * Caller needs to ensure that we're either frozen/quiesced, or that
3904 * the queue isn't live yet.
3906 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3908 struct blk_mq_hw_ctx *hctx;
3911 queue_for_each_hw_ctx(q, hctx, i) {
3913 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3915 blk_mq_tag_idle(hctx);
3916 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3921 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3924 struct request_queue *q;
3926 lockdep_assert_held(&set->tag_list_lock);
3928 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3929 blk_mq_freeze_queue(q);
3930 queue_set_hctx_shared(q, shared);
3931 blk_mq_unfreeze_queue(q);
3935 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3937 struct blk_mq_tag_set *set = q->tag_set;
3939 mutex_lock(&set->tag_list_lock);
3940 list_del(&q->tag_set_list);
3941 if (list_is_singular(&set->tag_list)) {
3942 /* just transitioned to unshared */
3943 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3944 /* update existing queue */
3945 blk_mq_update_tag_set_shared(set, false);
3947 mutex_unlock(&set->tag_list_lock);
3948 INIT_LIST_HEAD(&q->tag_set_list);
3951 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3952 struct request_queue *q)
3954 mutex_lock(&set->tag_list_lock);
3957 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3959 if (!list_empty(&set->tag_list) &&
3960 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3961 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3962 /* update existing queue */
3963 blk_mq_update_tag_set_shared(set, true);
3965 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3966 queue_set_hctx_shared(q, true);
3967 list_add_tail(&q->tag_set_list, &set->tag_list);
3969 mutex_unlock(&set->tag_list_lock);
3972 /* All allocations will be freed in release handler of q->mq_kobj */
3973 static int blk_mq_alloc_ctxs(struct request_queue *q)
3975 struct blk_mq_ctxs *ctxs;
3978 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3982 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3983 if (!ctxs->queue_ctx)
3986 for_each_possible_cpu(cpu) {
3987 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3991 q->mq_kobj = &ctxs->kobj;
3992 q->queue_ctx = ctxs->queue_ctx;
4001 * It is the actual release handler for mq, but we do it from
4002 * request queue's release handler for avoiding use-after-free
4003 * and headache because q->mq_kobj shouldn't have been introduced,
4004 * but we can't group ctx/kctx kobj without it.
4006 void blk_mq_release(struct request_queue *q)
4008 struct blk_mq_hw_ctx *hctx, *next;
4011 queue_for_each_hw_ctx(q, hctx, i)
4012 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4014 /* all hctx are in .unused_hctx_list now */
4015 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4016 list_del_init(&hctx->hctx_list);
4017 kobject_put(&hctx->kobj);
4020 xa_destroy(&q->hctx_table);
4023 * release .mq_kobj and sw queue's kobject now because
4024 * both share lifetime with request queue.
4026 blk_mq_sysfs_deinit(q);
4029 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4032 struct request_queue *q;
4035 q = blk_alloc_queue(set->numa_node);
4037 return ERR_PTR(-ENOMEM);
4038 q->queuedata = queuedata;
4039 ret = blk_mq_init_allocated_queue(set, q);
4042 return ERR_PTR(ret);
4047 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4049 return blk_mq_init_queue_data(set, NULL);
4051 EXPORT_SYMBOL(blk_mq_init_queue);
4054 * blk_mq_destroy_queue - shutdown a request queue
4055 * @q: request queue to shutdown
4057 * This shuts down a request queue allocated by blk_mq_init_queue(). All future
4058 * requests will be failed with -ENODEV. The caller is responsible for dropping
4059 * the reference from blk_mq_init_queue() by calling blk_put_queue().
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_mq_freeze_queue_wait(q);
4075 blk_mq_cancel_work_sync(q);
4076 blk_mq_exit_queue(q);
4078 EXPORT_SYMBOL(blk_mq_destroy_queue);
4080 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4081 struct lock_class_key *lkclass)
4083 struct request_queue *q;
4084 struct gendisk *disk;
4086 q = blk_mq_init_queue_data(set, queuedata);
4090 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4092 blk_mq_destroy_queue(q);
4094 return ERR_PTR(-ENOMEM);
4096 set_bit(GD_OWNS_QUEUE, &disk->state);
4099 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4101 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4102 struct lock_class_key *lkclass)
4104 struct gendisk *disk;
4106 if (!blk_get_queue(q))
4108 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4113 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4115 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4116 struct blk_mq_tag_set *set, struct request_queue *q,
4117 int hctx_idx, int node)
4119 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4121 /* reuse dead hctx first */
4122 spin_lock(&q->unused_hctx_lock);
4123 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4124 if (tmp->numa_node == node) {
4130 list_del_init(&hctx->hctx_list);
4131 spin_unlock(&q->unused_hctx_lock);
4134 hctx = blk_mq_alloc_hctx(q, set, node);
4138 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4144 kobject_put(&hctx->kobj);
4149 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4150 struct request_queue *q)
4152 struct blk_mq_hw_ctx *hctx;
4155 /* protect against switching io scheduler */
4156 mutex_lock(&q->sysfs_lock);
4157 for (i = 0; i < set->nr_hw_queues; i++) {
4159 int node = blk_mq_get_hctx_node(set, i);
4160 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4163 old_node = old_hctx->numa_node;
4164 blk_mq_exit_hctx(q, set, old_hctx, i);
4167 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4170 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4172 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4173 WARN_ON_ONCE(!hctx);
4177 * Increasing nr_hw_queues fails. Free the newly allocated
4178 * hctxs and keep the previous q->nr_hw_queues.
4180 if (i != set->nr_hw_queues) {
4181 j = q->nr_hw_queues;
4184 q->nr_hw_queues = set->nr_hw_queues;
4187 xa_for_each_start(&q->hctx_table, j, hctx, j)
4188 blk_mq_exit_hctx(q, set, hctx, j);
4189 mutex_unlock(&q->sysfs_lock);
4192 static void blk_mq_update_poll_flag(struct request_queue *q)
4194 struct blk_mq_tag_set *set = q->tag_set;
4196 if (set->nr_maps > HCTX_TYPE_POLL &&
4197 set->map[HCTX_TYPE_POLL].nr_queues)
4198 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4200 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4203 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4204 struct request_queue *q)
4206 /* mark the queue as mq asap */
4207 q->mq_ops = set->ops;
4209 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
4210 blk_mq_poll_stats_bkt,
4211 BLK_MQ_POLL_STATS_BKTS, q);
4215 if (blk_mq_alloc_ctxs(q))
4218 /* init q->mq_kobj and sw queues' kobjects */
4219 blk_mq_sysfs_init(q);
4221 INIT_LIST_HEAD(&q->unused_hctx_list);
4222 spin_lock_init(&q->unused_hctx_lock);
4224 xa_init(&q->hctx_table);
4226 blk_mq_realloc_hw_ctxs(set, q);
4227 if (!q->nr_hw_queues)
4230 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4231 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4235 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4236 blk_mq_update_poll_flag(q);
4238 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4239 INIT_LIST_HEAD(&q->requeue_list);
4240 spin_lock_init(&q->requeue_lock);
4242 q->nr_requests = set->queue_depth;
4245 * Default to classic polling
4247 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
4249 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4250 blk_mq_add_queue_tag_set(set, q);
4251 blk_mq_map_swqueue(q);
4257 blk_stat_free_callback(q->poll_cb);
4263 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4265 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4266 void blk_mq_exit_queue(struct request_queue *q)
4268 struct blk_mq_tag_set *set = q->tag_set;
4270 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4271 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4272 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4273 blk_mq_del_queue_tag_set(q);
4276 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4280 if (blk_mq_is_shared_tags(set->flags)) {
4281 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4284 if (!set->shared_tags)
4288 for (i = 0; i < set->nr_hw_queues; i++) {
4289 if (!__blk_mq_alloc_map_and_rqs(set, i))
4298 __blk_mq_free_map_and_rqs(set, i);
4300 if (blk_mq_is_shared_tags(set->flags)) {
4301 blk_mq_free_map_and_rqs(set, set->shared_tags,
4302 BLK_MQ_NO_HCTX_IDX);
4309 * Allocate the request maps associated with this tag_set. Note that this
4310 * may reduce the depth asked for, if memory is tight. set->queue_depth
4311 * will be updated to reflect the allocated depth.
4313 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4318 depth = set->queue_depth;
4320 err = __blk_mq_alloc_rq_maps(set);
4324 set->queue_depth >>= 1;
4325 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4329 } while (set->queue_depth);
4331 if (!set->queue_depth || err) {
4332 pr_err("blk-mq: failed to allocate request map\n");
4336 if (depth != set->queue_depth)
4337 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4338 depth, set->queue_depth);
4343 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4346 * blk_mq_map_queues() and multiple .map_queues() implementations
4347 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4348 * number of hardware queues.
4350 if (set->nr_maps == 1)
4351 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4353 if (set->ops->map_queues && !is_kdump_kernel()) {
4357 * transport .map_queues is usually done in the following
4360 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4361 * mask = get_cpu_mask(queue)
4362 * for_each_cpu(cpu, mask)
4363 * set->map[x].mq_map[cpu] = queue;
4366 * When we need to remap, the table has to be cleared for
4367 * killing stale mapping since one CPU may not be mapped
4370 for (i = 0; i < set->nr_maps; i++)
4371 blk_mq_clear_mq_map(&set->map[i]);
4373 set->ops->map_queues(set);
4375 BUG_ON(set->nr_maps > 1);
4376 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4380 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4381 int new_nr_hw_queues)
4383 struct blk_mq_tags **new_tags;
4385 if (set->nr_hw_queues >= new_nr_hw_queues)
4388 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4389 GFP_KERNEL, set->numa_node);
4394 memcpy(new_tags, set->tags, set->nr_hw_queues *
4395 sizeof(*set->tags));
4397 set->tags = new_tags;
4399 set->nr_hw_queues = new_nr_hw_queues;
4404 * Alloc a tag set to be associated with one or more request queues.
4405 * May fail with EINVAL for various error conditions. May adjust the
4406 * requested depth down, if it's too large. In that case, the set
4407 * value will be stored in set->queue_depth.
4409 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4413 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4415 if (!set->nr_hw_queues)
4417 if (!set->queue_depth)
4419 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4422 if (!set->ops->queue_rq)
4425 if (!set->ops->get_budget ^ !set->ops->put_budget)
4428 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4429 pr_info("blk-mq: reduced tag depth to %u\n",
4431 set->queue_depth = BLK_MQ_MAX_DEPTH;
4436 else if (set->nr_maps > HCTX_MAX_TYPES)
4440 * If a crashdump is active, then we are potentially in a very
4441 * memory constrained environment. Limit us to 1 queue and
4442 * 64 tags to prevent using too much memory.
4444 if (is_kdump_kernel()) {
4445 set->nr_hw_queues = 1;
4447 set->queue_depth = min(64U, set->queue_depth);
4450 * There is no use for more h/w queues than cpus if we just have
4453 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4454 set->nr_hw_queues = nr_cpu_ids;
4456 if (set->flags & BLK_MQ_F_BLOCKING) {
4457 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4460 ret = init_srcu_struct(set->srcu);
4466 set->tags = kcalloc_node(set->nr_hw_queues,
4467 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4470 goto out_cleanup_srcu;
4472 for (i = 0; i < set->nr_maps; i++) {
4473 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4474 sizeof(set->map[i].mq_map[0]),
4475 GFP_KERNEL, set->numa_node);
4476 if (!set->map[i].mq_map)
4477 goto out_free_mq_map;
4478 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4481 blk_mq_update_queue_map(set);
4483 ret = blk_mq_alloc_set_map_and_rqs(set);
4485 goto out_free_mq_map;
4487 mutex_init(&set->tag_list_lock);
4488 INIT_LIST_HEAD(&set->tag_list);
4493 for (i = 0; i < set->nr_maps; i++) {
4494 kfree(set->map[i].mq_map);
4495 set->map[i].mq_map = NULL;
4500 if (set->flags & BLK_MQ_F_BLOCKING)
4501 cleanup_srcu_struct(set->srcu);
4503 if (set->flags & BLK_MQ_F_BLOCKING)
4507 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4509 /* allocate and initialize a tagset for a simple single-queue device */
4510 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4511 const struct blk_mq_ops *ops, unsigned int queue_depth,
4512 unsigned int set_flags)
4514 memset(set, 0, sizeof(*set));
4516 set->nr_hw_queues = 1;
4518 set->queue_depth = queue_depth;
4519 set->numa_node = NUMA_NO_NODE;
4520 set->flags = set_flags;
4521 return blk_mq_alloc_tag_set(set);
4523 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4525 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4529 for (i = 0; i < set->nr_hw_queues; i++)
4530 __blk_mq_free_map_and_rqs(set, i);
4532 if (blk_mq_is_shared_tags(set->flags)) {
4533 blk_mq_free_map_and_rqs(set, set->shared_tags,
4534 BLK_MQ_NO_HCTX_IDX);
4537 for (j = 0; j < set->nr_maps; j++) {
4538 kfree(set->map[j].mq_map);
4539 set->map[j].mq_map = NULL;
4544 if (set->flags & BLK_MQ_F_BLOCKING) {
4545 cleanup_srcu_struct(set->srcu);
4549 EXPORT_SYMBOL(blk_mq_free_tag_set);
4551 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4553 struct blk_mq_tag_set *set = q->tag_set;
4554 struct blk_mq_hw_ctx *hctx;
4561 if (q->nr_requests == nr)
4564 blk_mq_freeze_queue(q);
4565 blk_mq_quiesce_queue(q);
4568 queue_for_each_hw_ctx(q, hctx, i) {
4572 * If we're using an MQ scheduler, just update the scheduler
4573 * queue depth. This is similar to what the old code would do.
4575 if (hctx->sched_tags) {
4576 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4579 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4584 if (q->elevator && q->elevator->type->ops.depth_updated)
4585 q->elevator->type->ops.depth_updated(hctx);
4588 q->nr_requests = nr;
4589 if (blk_mq_is_shared_tags(set->flags)) {
4591 blk_mq_tag_update_sched_shared_tags(q);
4593 blk_mq_tag_resize_shared_tags(set, nr);
4597 blk_mq_unquiesce_queue(q);
4598 blk_mq_unfreeze_queue(q);
4604 * request_queue and elevator_type pair.
4605 * It is just used by __blk_mq_update_nr_hw_queues to cache
4606 * the elevator_type associated with a request_queue.
4608 struct blk_mq_qe_pair {
4609 struct list_head node;
4610 struct request_queue *q;
4611 struct elevator_type *type;
4615 * Cache the elevator_type in qe pair list and switch the
4616 * io scheduler to 'none'
4618 static bool blk_mq_elv_switch_none(struct list_head *head,
4619 struct request_queue *q)
4621 struct blk_mq_qe_pair *qe;
4626 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4630 /* q->elevator needs protection from ->sysfs_lock */
4631 mutex_lock(&q->sysfs_lock);
4633 INIT_LIST_HEAD(&qe->node);
4635 qe->type = q->elevator->type;
4636 /* keep a reference to the elevator module as we'll switch back */
4637 __elevator_get(qe->type);
4638 list_add(&qe->node, head);
4639 elevator_disable(q);
4640 mutex_unlock(&q->sysfs_lock);
4645 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4646 struct request_queue *q)
4648 struct blk_mq_qe_pair *qe;
4650 list_for_each_entry(qe, head, node)
4657 static void blk_mq_elv_switch_back(struct list_head *head,
4658 struct request_queue *q)
4660 struct blk_mq_qe_pair *qe;
4661 struct elevator_type *t;
4663 qe = blk_lookup_qe_pair(head, q);
4667 list_del(&qe->node);
4670 mutex_lock(&q->sysfs_lock);
4671 elevator_switch(q, t);
4672 /* drop the reference acquired in blk_mq_elv_switch_none */
4674 mutex_unlock(&q->sysfs_lock);
4677 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4680 struct request_queue *q;
4682 int prev_nr_hw_queues;
4684 lockdep_assert_held(&set->tag_list_lock);
4686 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4687 nr_hw_queues = nr_cpu_ids;
4688 if (nr_hw_queues < 1)
4690 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4693 list_for_each_entry(q, &set->tag_list, tag_set_list)
4694 blk_mq_freeze_queue(q);
4696 * Switch IO scheduler to 'none', cleaning up the data associated
4697 * with the previous scheduler. We will switch back once we are done
4698 * updating the new sw to hw queue mappings.
4700 list_for_each_entry(q, &set->tag_list, tag_set_list)
4701 if (!blk_mq_elv_switch_none(&head, q))
4704 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4705 blk_mq_debugfs_unregister_hctxs(q);
4706 blk_mq_sysfs_unregister_hctxs(q);
4709 prev_nr_hw_queues = set->nr_hw_queues;
4710 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4714 blk_mq_update_queue_map(set);
4715 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4716 blk_mq_realloc_hw_ctxs(set, q);
4717 blk_mq_update_poll_flag(q);
4718 if (q->nr_hw_queues != set->nr_hw_queues) {
4719 int i = prev_nr_hw_queues;
4721 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4722 nr_hw_queues, prev_nr_hw_queues);
4723 for (; i < set->nr_hw_queues; i++)
4724 __blk_mq_free_map_and_rqs(set, i);
4726 set->nr_hw_queues = prev_nr_hw_queues;
4727 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4730 blk_mq_map_swqueue(q);
4734 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4735 blk_mq_sysfs_register_hctxs(q);
4736 blk_mq_debugfs_register_hctxs(q);
4740 list_for_each_entry(q, &set->tag_list, tag_set_list)
4741 blk_mq_elv_switch_back(&head, q);
4743 list_for_each_entry(q, &set->tag_list, tag_set_list)
4744 blk_mq_unfreeze_queue(q);
4747 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4749 mutex_lock(&set->tag_list_lock);
4750 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4751 mutex_unlock(&set->tag_list_lock);
4753 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4755 /* Enable polling stats and return whether they were already enabled. */
4756 static bool blk_poll_stats_enable(struct request_queue *q)
4761 return blk_stats_alloc_enable(q);
4764 static void blk_mq_poll_stats_start(struct request_queue *q)
4767 * We don't arm the callback if polling stats are not enabled or the
4768 * callback is already active.
4770 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
4773 blk_stat_activate_msecs(q->poll_cb, 100);
4776 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4778 struct request_queue *q = cb->data;
4781 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4782 if (cb->stat[bucket].nr_samples)
4783 q->poll_stat[bucket] = cb->stat[bucket];
4787 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4790 unsigned long ret = 0;
4794 * If stats collection isn't on, don't sleep but turn it on for
4797 if (!blk_poll_stats_enable(q))
4801 * As an optimistic guess, use half of the mean service time
4802 * for this type of request. We can (and should) make this smarter.
4803 * For instance, if the completion latencies are tight, we can
4804 * get closer than just half the mean. This is especially
4805 * important on devices where the completion latencies are longer
4806 * than ~10 usec. We do use the stats for the relevant IO size
4807 * if available which does lead to better estimates.
4809 bucket = blk_mq_poll_stats_bkt(rq);
4813 if (q->poll_stat[bucket].nr_samples)
4814 ret = (q->poll_stat[bucket].mean + 1) / 2;
4819 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4821 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4822 struct request *rq = blk_qc_to_rq(hctx, qc);
4823 struct hrtimer_sleeper hs;
4824 enum hrtimer_mode mode;
4829 * If a request has completed on queue that uses an I/O scheduler, we
4830 * won't get back a request from blk_qc_to_rq.
4832 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4836 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4838 * 0: use half of prev avg
4839 * >0: use this specific value
4841 if (q->poll_nsec > 0)
4842 nsecs = q->poll_nsec;
4844 nsecs = blk_mq_poll_nsecs(q, rq);
4849 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4852 * This will be replaced with the stats tracking code, using
4853 * 'avg_completion_time / 2' as the pre-sleep target.
4857 mode = HRTIMER_MODE_REL;
4858 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4859 hrtimer_set_expires(&hs.timer, kt);
4862 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4864 set_current_state(TASK_UNINTERRUPTIBLE);
4865 hrtimer_sleeper_start_expires(&hs, mode);
4868 hrtimer_cancel(&hs.timer);
4869 mode = HRTIMER_MODE_ABS;
4870 } while (hs.task && !signal_pending(current));
4872 __set_current_state(TASK_RUNNING);
4873 destroy_hrtimer_on_stack(&hs.timer);
4876 * If we sleep, have the caller restart the poll loop to reset the
4877 * state. Like for the other success return cases, the caller is
4878 * responsible for checking if the IO completed. If the IO isn't
4879 * complete, we'll get called again and will go straight to the busy
4885 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4886 struct io_comp_batch *iob, unsigned int flags)
4888 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4889 long state = get_current_state();
4893 ret = q->mq_ops->poll(hctx, iob);
4895 __set_current_state(TASK_RUNNING);
4899 if (signal_pending_state(state, current))
4900 __set_current_state(TASK_RUNNING);
4901 if (task_is_running(current))
4904 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4907 } while (!need_resched());
4909 __set_current_state(TASK_RUNNING);
4913 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4916 if (!(flags & BLK_POLL_NOSLEEP) &&
4917 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4918 if (blk_mq_poll_hybrid(q, cookie))
4921 return blk_mq_poll_classic(q, cookie, iob, flags);
4924 unsigned int blk_mq_rq_cpu(struct request *rq)
4926 return rq->mq_ctx->cpu;
4928 EXPORT_SYMBOL(blk_mq_rq_cpu);
4930 void blk_mq_cancel_work_sync(struct request_queue *q)
4932 struct blk_mq_hw_ctx *hctx;
4935 cancel_delayed_work_sync(&q->requeue_work);
4937 queue_for_each_hw_ctx(q, hctx, i)
4938 cancel_delayed_work_sync(&hctx->run_work);
4941 static int __init blk_mq_init(void)
4945 for_each_possible_cpu(i)
4946 init_llist_head(&per_cpu(blk_cpu_done, i));
4947 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4949 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4950 "block/softirq:dead", NULL,
4951 blk_softirq_cpu_dead);
4952 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4953 blk_mq_hctx_notify_dead);
4954 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4955 blk_mq_hctx_notify_online,
4956 blk_mq_hctx_notify_offline);
4959 subsys_initcall(blk_mq_init);