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;
650 /* alloc_time includes depth and tag waits */
651 if (blk_queue_rq_alloc_time(q))
652 alloc_time_ns = ktime_get_ns();
655 * If the tag allocator sleeps we could get an allocation for a
656 * different hardware context. No need to complicate the low level
657 * allocator for this for the rare use case of a command tied to
660 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
661 return ERR_PTR(-EINVAL);
663 if (hctx_idx >= q->nr_hw_queues)
664 return ERR_PTR(-EIO);
666 ret = blk_queue_enter(q, flags);
671 * Check if the hardware context is actually mapped to anything.
672 * If not tell the caller that it should skip this queue.
675 data.hctx = xa_load(&q->hctx_table, hctx_idx);
676 if (!blk_mq_hw_queue_mapped(data.hctx))
678 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
679 if (cpu >= nr_cpu_ids)
681 data.ctx = __blk_mq_get_ctx(q, cpu);
684 blk_mq_tag_busy(data.hctx);
686 data.rq_flags |= RQF_ELV;
688 if (flags & BLK_MQ_REQ_RESERVED)
689 data.rq_flags |= RQF_RESV;
692 tag = blk_mq_get_tag(&data);
693 if (tag == BLK_MQ_NO_TAG)
695 return blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag,
702 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
704 static void __blk_mq_free_request(struct request *rq)
706 struct request_queue *q = rq->q;
707 struct blk_mq_ctx *ctx = rq->mq_ctx;
708 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
709 const int sched_tag = rq->internal_tag;
711 blk_crypto_free_request(rq);
712 blk_pm_mark_last_busy(rq);
714 if (rq->tag != BLK_MQ_NO_TAG)
715 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
716 if (sched_tag != BLK_MQ_NO_TAG)
717 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
718 blk_mq_sched_restart(hctx);
722 void blk_mq_free_request(struct request *rq)
724 struct request_queue *q = rq->q;
725 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
727 if ((rq->rq_flags & RQF_ELVPRIV) &&
728 q->elevator->type->ops.finish_request)
729 q->elevator->type->ops.finish_request(rq);
731 if (rq->rq_flags & RQF_MQ_INFLIGHT)
732 __blk_mq_dec_active_requests(hctx);
734 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
735 laptop_io_completion(q->disk->bdi);
739 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
740 if (req_ref_put_and_test(rq))
741 __blk_mq_free_request(rq);
743 EXPORT_SYMBOL_GPL(blk_mq_free_request);
745 void blk_mq_free_plug_rqs(struct blk_plug *plug)
749 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
750 blk_mq_free_request(rq);
753 void blk_dump_rq_flags(struct request *rq, char *msg)
755 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
756 rq->q->disk ? rq->q->disk->disk_name : "?",
757 (__force unsigned long long) rq->cmd_flags);
759 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
760 (unsigned long long)blk_rq_pos(rq),
761 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
762 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
763 rq->bio, rq->biotail, blk_rq_bytes(rq));
765 EXPORT_SYMBOL(blk_dump_rq_flags);
767 static void req_bio_endio(struct request *rq, struct bio *bio,
768 unsigned int nbytes, blk_status_t error)
770 if (unlikely(error)) {
771 bio->bi_status = error;
772 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
774 * Partial zone append completions cannot be supported as the
775 * BIO fragments may end up not being written sequentially.
777 if (bio->bi_iter.bi_size != nbytes)
778 bio->bi_status = BLK_STS_IOERR;
780 bio->bi_iter.bi_sector = rq->__sector;
783 bio_advance(bio, nbytes);
785 if (unlikely(rq->rq_flags & RQF_QUIET))
786 bio_set_flag(bio, BIO_QUIET);
787 /* don't actually finish bio if it's part of flush sequence */
788 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
792 static void blk_account_io_completion(struct request *req, unsigned int bytes)
794 if (req->part && blk_do_io_stat(req)) {
795 const int sgrp = op_stat_group(req_op(req));
798 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
803 static void blk_print_req_error(struct request *req, blk_status_t status)
805 printk_ratelimited(KERN_ERR
806 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
807 "phys_seg %u prio class %u\n",
808 blk_status_to_str(status),
809 req->q->disk ? req->q->disk->disk_name : "?",
810 blk_rq_pos(req), (__force u32)req_op(req),
811 blk_op_str(req_op(req)),
812 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
813 req->nr_phys_segments,
814 IOPRIO_PRIO_CLASS(req->ioprio));
818 * Fully end IO on a request. Does not support partial completions, or
821 static void blk_complete_request(struct request *req)
823 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
824 int total_bytes = blk_rq_bytes(req);
825 struct bio *bio = req->bio;
827 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
832 #ifdef CONFIG_BLK_DEV_INTEGRITY
833 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
834 req->q->integrity.profile->complete_fn(req, total_bytes);
837 blk_account_io_completion(req, total_bytes);
840 struct bio *next = bio->bi_next;
842 /* Completion has already been traced */
843 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
845 if (req_op(req) == REQ_OP_ZONE_APPEND)
846 bio->bi_iter.bi_sector = req->__sector;
854 * Reset counters so that the request stacking driver
855 * can find how many bytes remain in the request
865 * blk_update_request - Complete multiple bytes without completing the request
866 * @req: the request being processed
867 * @error: block status code
868 * @nr_bytes: number of bytes to complete for @req
871 * Ends I/O on a number of bytes attached to @req, but doesn't complete
872 * the request structure even if @req doesn't have leftover.
873 * If @req has leftover, sets it up for the next range of segments.
875 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
876 * %false return from this function.
879 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
880 * except in the consistency check at the end of this function.
883 * %false - this request doesn't have any more data
884 * %true - this request has more data
886 bool blk_update_request(struct request *req, blk_status_t error,
887 unsigned int nr_bytes)
891 trace_block_rq_complete(req, error, nr_bytes);
896 #ifdef CONFIG_BLK_DEV_INTEGRITY
897 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
899 req->q->integrity.profile->complete_fn(req, nr_bytes);
902 if (unlikely(error && !blk_rq_is_passthrough(req) &&
903 !(req->rq_flags & RQF_QUIET)) &&
904 !test_bit(GD_DEAD, &req->q->disk->state)) {
905 blk_print_req_error(req, error);
906 trace_block_rq_error(req, error, nr_bytes);
909 blk_account_io_completion(req, nr_bytes);
913 struct bio *bio = req->bio;
914 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
916 if (bio_bytes == bio->bi_iter.bi_size)
917 req->bio = bio->bi_next;
919 /* Completion has already been traced */
920 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
921 req_bio_endio(req, bio, bio_bytes, error);
923 total_bytes += bio_bytes;
924 nr_bytes -= bio_bytes;
935 * Reset counters so that the request stacking driver
936 * can find how many bytes remain in the request
943 req->__data_len -= total_bytes;
945 /* update sector only for requests with clear definition of sector */
946 if (!blk_rq_is_passthrough(req))
947 req->__sector += total_bytes >> 9;
949 /* mixed attributes always follow the first bio */
950 if (req->rq_flags & RQF_MIXED_MERGE) {
951 req->cmd_flags &= ~REQ_FAILFAST_MASK;
952 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
955 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
957 * If total number of sectors is less than the first segment
958 * size, something has gone terribly wrong.
960 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
961 blk_dump_rq_flags(req, "request botched");
962 req->__data_len = blk_rq_cur_bytes(req);
965 /* recalculate the number of segments */
966 req->nr_phys_segments = blk_recalc_rq_segments(req);
971 EXPORT_SYMBOL_GPL(blk_update_request);
973 static void __blk_account_io_done(struct request *req, u64 now)
975 const int sgrp = op_stat_group(req_op(req));
978 update_io_ticks(req->part, jiffies, true);
979 part_stat_inc(req->part, ios[sgrp]);
980 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
984 static inline void blk_account_io_done(struct request *req, u64 now)
987 * Account IO completion. flush_rq isn't accounted as a
988 * normal IO on queueing nor completion. Accounting the
989 * containing request is enough.
991 if (blk_do_io_stat(req) && req->part &&
992 !(req->rq_flags & RQF_FLUSH_SEQ))
993 __blk_account_io_done(req, now);
996 static void __blk_account_io_start(struct request *rq)
999 * All non-passthrough requests are created from a bio with one
1000 * exception: when a flush command that is part of a flush sequence
1001 * generated by the state machine in blk-flush.c is cloned onto the
1002 * lower device by dm-multipath we can get here without a bio.
1005 rq->part = rq->bio->bi_bdev;
1007 rq->part = rq->q->disk->part0;
1010 update_io_ticks(rq->part, jiffies, false);
1014 static inline void blk_account_io_start(struct request *req)
1016 if (blk_do_io_stat(req))
1017 __blk_account_io_start(req);
1020 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1022 if (rq->rq_flags & RQF_STATS) {
1023 blk_mq_poll_stats_start(rq->q);
1024 blk_stat_add(rq, now);
1027 blk_mq_sched_completed_request(rq, now);
1028 blk_account_io_done(rq, now);
1031 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1033 if (blk_mq_need_time_stamp(rq))
1034 __blk_mq_end_request_acct(rq, ktime_get_ns());
1037 rq_qos_done(rq->q, rq);
1038 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1039 blk_mq_free_request(rq);
1041 blk_mq_free_request(rq);
1044 EXPORT_SYMBOL(__blk_mq_end_request);
1046 void blk_mq_end_request(struct request *rq, blk_status_t error)
1048 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1050 __blk_mq_end_request(rq, error);
1052 EXPORT_SYMBOL(blk_mq_end_request);
1054 #define TAG_COMP_BATCH 32
1056 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1057 int *tag_array, int nr_tags)
1059 struct request_queue *q = hctx->queue;
1062 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1063 * update hctx->nr_active in batch
1065 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1066 __blk_mq_sub_active_requests(hctx, nr_tags);
1068 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1069 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1072 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1074 int tags[TAG_COMP_BATCH], nr_tags = 0;
1075 struct blk_mq_hw_ctx *cur_hctx = NULL;
1080 now = ktime_get_ns();
1082 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1084 prefetch(rq->rq_next);
1086 blk_complete_request(rq);
1088 __blk_mq_end_request_acct(rq, now);
1090 rq_qos_done(rq->q, rq);
1093 * If end_io handler returns NONE, then it still has
1094 * ownership of the request.
1096 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1099 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1100 if (!req_ref_put_and_test(rq))
1103 blk_crypto_free_request(rq);
1104 blk_pm_mark_last_busy(rq);
1106 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1108 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1110 cur_hctx = rq->mq_hctx;
1112 tags[nr_tags++] = rq->tag;
1116 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1118 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1120 static void blk_complete_reqs(struct llist_head *list)
1122 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1123 struct request *rq, *next;
1125 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1126 rq->q->mq_ops->complete(rq);
1129 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1131 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1134 static int blk_softirq_cpu_dead(unsigned int cpu)
1136 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1140 static void __blk_mq_complete_request_remote(void *data)
1142 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1145 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1147 int cpu = raw_smp_processor_id();
1149 if (!IS_ENABLED(CONFIG_SMP) ||
1150 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1153 * With force threaded interrupts enabled, raising softirq from an SMP
1154 * function call will always result in waking the ksoftirqd thread.
1155 * This is probably worse than completing the request on a different
1158 if (force_irqthreads())
1161 /* same CPU or cache domain? Complete locally */
1162 if (cpu == rq->mq_ctx->cpu ||
1163 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1164 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1167 /* don't try to IPI to an offline CPU */
1168 return cpu_online(rq->mq_ctx->cpu);
1171 static void blk_mq_complete_send_ipi(struct request *rq)
1173 struct llist_head *list;
1176 cpu = rq->mq_ctx->cpu;
1177 list = &per_cpu(blk_cpu_done, cpu);
1178 if (llist_add(&rq->ipi_list, list)) {
1179 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1180 smp_call_function_single_async(cpu, &rq->csd);
1184 static void blk_mq_raise_softirq(struct request *rq)
1186 struct llist_head *list;
1189 list = this_cpu_ptr(&blk_cpu_done);
1190 if (llist_add(&rq->ipi_list, list))
1191 raise_softirq(BLOCK_SOFTIRQ);
1195 bool blk_mq_complete_request_remote(struct request *rq)
1197 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1200 * For request which hctx has only one ctx mapping,
1201 * or a polled request, always complete locally,
1202 * it's pointless to redirect the completion.
1204 if (rq->mq_hctx->nr_ctx == 1 ||
1205 rq->cmd_flags & REQ_POLLED)
1208 if (blk_mq_complete_need_ipi(rq)) {
1209 blk_mq_complete_send_ipi(rq);
1213 if (rq->q->nr_hw_queues == 1) {
1214 blk_mq_raise_softirq(rq);
1219 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1222 * blk_mq_complete_request - end I/O on a request
1223 * @rq: the request being processed
1226 * Complete a request by scheduling the ->complete_rq operation.
1228 void blk_mq_complete_request(struct request *rq)
1230 if (!blk_mq_complete_request_remote(rq))
1231 rq->q->mq_ops->complete(rq);
1233 EXPORT_SYMBOL(blk_mq_complete_request);
1236 * blk_mq_start_request - Start processing a request
1237 * @rq: Pointer to request to be started
1239 * Function used by device drivers to notify the block layer that a request
1240 * is going to be processed now, so blk layer can do proper initializations
1241 * such as starting the timeout timer.
1243 void blk_mq_start_request(struct request *rq)
1245 struct request_queue *q = rq->q;
1247 trace_block_rq_issue(rq);
1249 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1250 rq->io_start_time_ns = ktime_get_ns();
1251 rq->stats_sectors = blk_rq_sectors(rq);
1252 rq->rq_flags |= RQF_STATS;
1253 rq_qos_issue(q, rq);
1256 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1259 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1261 #ifdef CONFIG_BLK_DEV_INTEGRITY
1262 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1263 q->integrity.profile->prepare_fn(rq);
1265 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1266 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1268 EXPORT_SYMBOL(blk_mq_start_request);
1271 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1272 * queues. This is important for md arrays to benefit from merging
1275 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1277 if (plug->multiple_queues)
1278 return BLK_MAX_REQUEST_COUNT * 2;
1279 return BLK_MAX_REQUEST_COUNT;
1282 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1284 struct request *last = rq_list_peek(&plug->mq_list);
1286 if (!plug->rq_count) {
1287 trace_block_plug(rq->q);
1288 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1289 (!blk_queue_nomerges(rq->q) &&
1290 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1291 blk_mq_flush_plug_list(plug, false);
1292 trace_block_plug(rq->q);
1295 if (!plug->multiple_queues && last && last->q != rq->q)
1296 plug->multiple_queues = true;
1297 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1298 plug->has_elevator = true;
1300 rq_list_add(&plug->mq_list, rq);
1305 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1306 * @rq: request to insert
1307 * @at_head: insert request at head or tail of queue
1310 * Insert a fully prepared request at the back of the I/O scheduler queue
1311 * for execution. Don't wait for completion.
1314 * This function will invoke @done directly if the queue is dead.
1316 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1318 WARN_ON(irqs_disabled());
1319 WARN_ON(!blk_rq_is_passthrough(rq));
1321 blk_account_io_start(rq);
1324 * As plugging can be enabled for passthrough requests on a zoned
1325 * device, directly accessing the plug instead of using blk_mq_plug()
1326 * should not have any consequences.
1329 blk_add_rq_to_plug(current->plug, rq);
1331 blk_mq_sched_insert_request(rq, at_head, true, false);
1333 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1335 struct blk_rq_wait {
1336 struct completion done;
1340 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1342 struct blk_rq_wait *wait = rq->end_io_data;
1345 complete(&wait->done);
1346 return RQ_END_IO_NONE;
1349 bool blk_rq_is_poll(struct request *rq)
1353 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1355 if (WARN_ON_ONCE(!rq->bio))
1359 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1361 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1364 bio_poll(rq->bio, NULL, 0);
1366 } while (!completion_done(wait));
1370 * blk_execute_rq - insert a request into queue for execution
1371 * @rq: request to insert
1372 * @at_head: insert request at head or tail of queue
1375 * Insert a fully prepared request at the back of the I/O scheduler queue
1376 * for execution and wait for completion.
1377 * Return: The blk_status_t result provided to blk_mq_end_request().
1379 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1381 struct blk_rq_wait wait = {
1382 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1385 WARN_ON(irqs_disabled());
1386 WARN_ON(!blk_rq_is_passthrough(rq));
1388 rq->end_io_data = &wait;
1389 rq->end_io = blk_end_sync_rq;
1391 blk_account_io_start(rq);
1392 blk_mq_sched_insert_request(rq, at_head, true, false);
1394 if (blk_rq_is_poll(rq)) {
1395 blk_rq_poll_completion(rq, &wait.done);
1398 * Prevent hang_check timer from firing at us during very long
1401 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1404 while (!wait_for_completion_io_timeout(&wait.done,
1405 hang_check * (HZ/2)))
1408 wait_for_completion_io(&wait.done);
1413 EXPORT_SYMBOL(blk_execute_rq);
1415 static void __blk_mq_requeue_request(struct request *rq)
1417 struct request_queue *q = rq->q;
1419 blk_mq_put_driver_tag(rq);
1421 trace_block_rq_requeue(rq);
1422 rq_qos_requeue(q, rq);
1424 if (blk_mq_request_started(rq)) {
1425 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1426 rq->rq_flags &= ~RQF_TIMED_OUT;
1430 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1432 __blk_mq_requeue_request(rq);
1434 /* this request will be re-inserted to io scheduler queue */
1435 blk_mq_sched_requeue_request(rq);
1437 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1439 EXPORT_SYMBOL(blk_mq_requeue_request);
1441 static void blk_mq_requeue_work(struct work_struct *work)
1443 struct request_queue *q =
1444 container_of(work, struct request_queue, requeue_work.work);
1446 struct request *rq, *next;
1448 spin_lock_irq(&q->requeue_lock);
1449 list_splice_init(&q->requeue_list, &rq_list);
1450 spin_unlock_irq(&q->requeue_lock);
1452 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1453 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1456 rq->rq_flags &= ~RQF_SOFTBARRIER;
1457 list_del_init(&rq->queuelist);
1459 * If RQF_DONTPREP, rq has contained some driver specific
1460 * data, so insert it to hctx dispatch list to avoid any
1463 if (rq->rq_flags & RQF_DONTPREP)
1464 blk_mq_request_bypass_insert(rq, false, false);
1466 blk_mq_sched_insert_request(rq, true, false, false);
1469 while (!list_empty(&rq_list)) {
1470 rq = list_entry(rq_list.next, struct request, queuelist);
1471 list_del_init(&rq->queuelist);
1472 blk_mq_sched_insert_request(rq, false, false, false);
1475 blk_mq_run_hw_queues(q, false);
1478 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1479 bool kick_requeue_list)
1481 struct request_queue *q = rq->q;
1482 unsigned long flags;
1485 * We abuse this flag that is otherwise used by the I/O scheduler to
1486 * request head insertion from the workqueue.
1488 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1490 spin_lock_irqsave(&q->requeue_lock, flags);
1492 rq->rq_flags |= RQF_SOFTBARRIER;
1493 list_add(&rq->queuelist, &q->requeue_list);
1495 list_add_tail(&rq->queuelist, &q->requeue_list);
1497 spin_unlock_irqrestore(&q->requeue_lock, flags);
1499 if (kick_requeue_list)
1500 blk_mq_kick_requeue_list(q);
1503 void blk_mq_kick_requeue_list(struct request_queue *q)
1505 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1507 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1509 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1510 unsigned long msecs)
1512 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1513 msecs_to_jiffies(msecs));
1515 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1517 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1520 * If we find a request that isn't idle we know the queue is busy
1521 * as it's checked in the iter.
1522 * Return false to stop the iteration.
1524 if (blk_mq_request_started(rq)) {
1534 bool blk_mq_queue_inflight(struct request_queue *q)
1538 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1541 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1543 static void blk_mq_rq_timed_out(struct request *req)
1545 req->rq_flags |= RQF_TIMED_OUT;
1546 if (req->q->mq_ops->timeout) {
1547 enum blk_eh_timer_return ret;
1549 ret = req->q->mq_ops->timeout(req);
1550 if (ret == BLK_EH_DONE)
1552 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1558 struct blk_expired_data {
1559 bool has_timedout_rq;
1561 unsigned long timeout_start;
1564 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1566 unsigned long deadline;
1568 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1570 if (rq->rq_flags & RQF_TIMED_OUT)
1573 deadline = READ_ONCE(rq->deadline);
1574 if (time_after_eq(expired->timeout_start, deadline))
1577 if (expired->next == 0)
1578 expired->next = deadline;
1579 else if (time_after(expired->next, deadline))
1580 expired->next = deadline;
1584 void blk_mq_put_rq_ref(struct request *rq)
1586 if (is_flush_rq(rq)) {
1587 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1588 blk_mq_free_request(rq);
1589 } else if (req_ref_put_and_test(rq)) {
1590 __blk_mq_free_request(rq);
1594 static bool blk_mq_check_expired(struct request *rq, void *priv)
1596 struct blk_expired_data *expired = priv;
1599 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1600 * be reallocated underneath the timeout handler's processing, then
1601 * the expire check is reliable. If the request is not expired, then
1602 * it was completed and reallocated as a new request after returning
1603 * from blk_mq_check_expired().
1605 if (blk_mq_req_expired(rq, expired)) {
1606 expired->has_timedout_rq = true;
1612 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1614 struct blk_expired_data *expired = priv;
1616 if (blk_mq_req_expired(rq, expired))
1617 blk_mq_rq_timed_out(rq);
1621 static void blk_mq_timeout_work(struct work_struct *work)
1623 struct request_queue *q =
1624 container_of(work, struct request_queue, timeout_work);
1625 struct blk_expired_data expired = {
1626 .timeout_start = jiffies,
1628 struct blk_mq_hw_ctx *hctx;
1631 /* A deadlock might occur if a request is stuck requiring a
1632 * timeout at the same time a queue freeze is waiting
1633 * completion, since the timeout code would not be able to
1634 * acquire the queue reference here.
1636 * That's why we don't use blk_queue_enter here; instead, we use
1637 * percpu_ref_tryget directly, because we need to be able to
1638 * obtain a reference even in the short window between the queue
1639 * starting to freeze, by dropping the first reference in
1640 * blk_freeze_queue_start, and the moment the last request is
1641 * consumed, marked by the instant q_usage_counter reaches
1644 if (!percpu_ref_tryget(&q->q_usage_counter))
1647 /* check if there is any timed-out request */
1648 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1649 if (expired.has_timedout_rq) {
1651 * Before walking tags, we must ensure any submit started
1652 * before the current time has finished. Since the submit
1653 * uses srcu or rcu, wait for a synchronization point to
1654 * ensure all running submits have finished
1656 blk_mq_wait_quiesce_done(q->tag_set);
1659 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1662 if (expired.next != 0) {
1663 mod_timer(&q->timeout, expired.next);
1666 * Request timeouts are handled as a forward rolling timer. If
1667 * we end up here it means that no requests are pending and
1668 * also that no request has been pending for a while. Mark
1669 * each hctx as idle.
1671 queue_for_each_hw_ctx(q, hctx, i) {
1672 /* the hctx may be unmapped, so check it here */
1673 if (blk_mq_hw_queue_mapped(hctx))
1674 blk_mq_tag_idle(hctx);
1680 struct flush_busy_ctx_data {
1681 struct blk_mq_hw_ctx *hctx;
1682 struct list_head *list;
1685 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1687 struct flush_busy_ctx_data *flush_data = data;
1688 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1689 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1690 enum hctx_type type = hctx->type;
1692 spin_lock(&ctx->lock);
1693 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1694 sbitmap_clear_bit(sb, bitnr);
1695 spin_unlock(&ctx->lock);
1700 * Process software queues that have been marked busy, splicing them
1701 * to the for-dispatch
1703 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1705 struct flush_busy_ctx_data data = {
1710 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1712 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1714 struct dispatch_rq_data {
1715 struct blk_mq_hw_ctx *hctx;
1719 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1722 struct dispatch_rq_data *dispatch_data = data;
1723 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1724 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1725 enum hctx_type type = hctx->type;
1727 spin_lock(&ctx->lock);
1728 if (!list_empty(&ctx->rq_lists[type])) {
1729 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1730 list_del_init(&dispatch_data->rq->queuelist);
1731 if (list_empty(&ctx->rq_lists[type]))
1732 sbitmap_clear_bit(sb, bitnr);
1734 spin_unlock(&ctx->lock);
1736 return !dispatch_data->rq;
1739 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1740 struct blk_mq_ctx *start)
1742 unsigned off = start ? start->index_hw[hctx->type] : 0;
1743 struct dispatch_rq_data data = {
1748 __sbitmap_for_each_set(&hctx->ctx_map, off,
1749 dispatch_rq_from_ctx, &data);
1754 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1756 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1757 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1760 blk_mq_tag_busy(rq->mq_hctx);
1762 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1763 bt = &rq->mq_hctx->tags->breserved_tags;
1766 if (!hctx_may_queue(rq->mq_hctx, bt))
1770 tag = __sbitmap_queue_get(bt);
1771 if (tag == BLK_MQ_NO_TAG)
1774 rq->tag = tag + tag_offset;
1778 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1780 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1783 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1784 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1785 rq->rq_flags |= RQF_MQ_INFLIGHT;
1786 __blk_mq_inc_active_requests(hctx);
1788 hctx->tags->rqs[rq->tag] = rq;
1792 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1793 int flags, void *key)
1795 struct blk_mq_hw_ctx *hctx;
1797 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1799 spin_lock(&hctx->dispatch_wait_lock);
1800 if (!list_empty(&wait->entry)) {
1801 struct sbitmap_queue *sbq;
1803 list_del_init(&wait->entry);
1804 sbq = &hctx->tags->bitmap_tags;
1805 atomic_dec(&sbq->ws_active);
1807 spin_unlock(&hctx->dispatch_wait_lock);
1809 blk_mq_run_hw_queue(hctx, true);
1814 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1815 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1816 * restart. For both cases, take care to check the condition again after
1817 * marking us as waiting.
1819 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1822 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1823 struct wait_queue_head *wq;
1824 wait_queue_entry_t *wait;
1827 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1828 blk_mq_sched_mark_restart_hctx(hctx);
1831 * It's possible that a tag was freed in the window between the
1832 * allocation failure and adding the hardware queue to the wait
1835 * Don't clear RESTART here, someone else could have set it.
1836 * At most this will cost an extra queue run.
1838 return blk_mq_get_driver_tag(rq);
1841 wait = &hctx->dispatch_wait;
1842 if (!list_empty_careful(&wait->entry))
1845 wq = &bt_wait_ptr(sbq, hctx)->wait;
1847 spin_lock_irq(&wq->lock);
1848 spin_lock(&hctx->dispatch_wait_lock);
1849 if (!list_empty(&wait->entry)) {
1850 spin_unlock(&hctx->dispatch_wait_lock);
1851 spin_unlock_irq(&wq->lock);
1855 atomic_inc(&sbq->ws_active);
1856 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1857 __add_wait_queue(wq, wait);
1860 * It's possible that a tag was freed in the window between the
1861 * allocation failure and adding the hardware queue to the wait
1864 ret = blk_mq_get_driver_tag(rq);
1866 spin_unlock(&hctx->dispatch_wait_lock);
1867 spin_unlock_irq(&wq->lock);
1872 * We got a tag, remove ourselves from the wait queue to ensure
1873 * someone else gets the wakeup.
1875 list_del_init(&wait->entry);
1876 atomic_dec(&sbq->ws_active);
1877 spin_unlock(&hctx->dispatch_wait_lock);
1878 spin_unlock_irq(&wq->lock);
1883 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1884 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1886 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1887 * - EWMA is one simple way to compute running average value
1888 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1889 * - take 4 as factor for avoiding to get too small(0) result, and this
1890 * factor doesn't matter because EWMA decreases exponentially
1892 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1896 ewma = hctx->dispatch_busy;
1901 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1903 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1904 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1906 hctx->dispatch_busy = ewma;
1909 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1911 static void blk_mq_handle_dev_resource(struct request *rq,
1912 struct list_head *list)
1914 struct request *next =
1915 list_first_entry_or_null(list, struct request, queuelist);
1918 * If an I/O scheduler has been configured and we got a driver tag for
1919 * the next request already, free it.
1922 blk_mq_put_driver_tag(next);
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 * Returns true if we did some work AND can potentially do more.
2001 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
2002 unsigned int nr_budgets)
2004 enum prep_dispatch prep;
2005 struct request_queue *q = hctx->queue;
2006 struct request *rq, *nxt;
2008 blk_status_t ret = BLK_STS_OK;
2009 LIST_HEAD(zone_list);
2010 bool needs_resource = false;
2012 if (list_empty(list))
2016 * Now process all the entries, sending them to the driver.
2018 errors = queued = 0;
2020 struct blk_mq_queue_data bd;
2022 rq = list_first_entry(list, struct request, queuelist);
2024 WARN_ON_ONCE(hctx != rq->mq_hctx);
2025 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2026 if (prep != PREP_DISPATCH_OK)
2029 list_del_init(&rq->queuelist);
2034 * Flag last if we have no more requests, or if we have more
2035 * but can't assign a driver tag to it.
2037 if (list_empty(list))
2040 nxt = list_first_entry(list, struct request, queuelist);
2041 bd.last = !blk_mq_get_driver_tag(nxt);
2045 * once the request is queued to lld, no need to cover the
2050 ret = q->mq_ops->queue_rq(hctx, &bd);
2055 case BLK_STS_RESOURCE:
2056 needs_resource = true;
2058 case BLK_STS_DEV_RESOURCE:
2059 blk_mq_handle_dev_resource(rq, list);
2061 case BLK_STS_ZONE_RESOURCE:
2063 * Move the request to zone_list and keep going through
2064 * the dispatch list to find more requests the drive can
2067 blk_mq_handle_zone_resource(rq, &zone_list);
2068 needs_resource = true;
2072 blk_mq_end_request(rq, ret);
2074 } while (!list_empty(list));
2076 if (!list_empty(&zone_list))
2077 list_splice_tail_init(&zone_list, list);
2079 /* If we didn't flush the entire list, we could have told the driver
2080 * there was more coming, but that turned out to be a lie.
2082 if ((!list_empty(list) || errors || needs_resource ||
2083 ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued)
2084 q->mq_ops->commit_rqs(hctx);
2086 * Any items that need requeuing? Stuff them into hctx->dispatch,
2087 * that is where we will continue on next queue run.
2089 if (!list_empty(list)) {
2091 /* For non-shared tags, the RESTART check will suffice */
2092 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2093 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
2096 blk_mq_release_budgets(q, list);
2098 spin_lock(&hctx->lock);
2099 list_splice_tail_init(list, &hctx->dispatch);
2100 spin_unlock(&hctx->lock);
2103 * Order adding requests to hctx->dispatch and checking
2104 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2105 * in blk_mq_sched_restart(). Avoid restart code path to
2106 * miss the new added requests to hctx->dispatch, meantime
2107 * SCHED_RESTART is observed here.
2112 * If SCHED_RESTART was set by the caller of this function and
2113 * it is no longer set that means that it was cleared by another
2114 * thread and hence that a queue rerun is needed.
2116 * If 'no_tag' is set, that means that we failed getting
2117 * a driver tag with an I/O scheduler attached. If our dispatch
2118 * waitqueue is no longer active, ensure that we run the queue
2119 * AFTER adding our entries back to the list.
2121 * If no I/O scheduler has been configured it is possible that
2122 * the hardware queue got stopped and restarted before requests
2123 * were pushed back onto the dispatch list. Rerun the queue to
2124 * avoid starvation. Notes:
2125 * - blk_mq_run_hw_queue() checks whether or not a queue has
2126 * been stopped before rerunning a queue.
2127 * - Some but not all block drivers stop a queue before
2128 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2131 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2132 * bit is set, run queue after a delay to avoid IO stalls
2133 * that could otherwise occur if the queue is idle. We'll do
2134 * similar if we couldn't get budget or couldn't lock a zone
2135 * and SCHED_RESTART is set.
2137 needs_restart = blk_mq_sched_needs_restart(hctx);
2138 if (prep == PREP_DISPATCH_NO_BUDGET)
2139 needs_resource = true;
2140 if (!needs_restart ||
2141 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2142 blk_mq_run_hw_queue(hctx, true);
2143 else if (needs_resource)
2144 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2146 blk_mq_update_dispatch_busy(hctx, true);
2149 blk_mq_update_dispatch_busy(hctx, false);
2151 return (queued + errors) != 0;
2155 * __blk_mq_run_hw_queue - Run a hardware queue.
2156 * @hctx: Pointer to the hardware queue to run.
2158 * Send pending requests to the hardware.
2160 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
2163 * We can't run the queue inline with ints disabled. Ensure that
2164 * we catch bad users of this early.
2166 WARN_ON_ONCE(in_interrupt());
2168 blk_mq_run_dispatch_ops(hctx->queue,
2169 blk_mq_sched_dispatch_requests(hctx));
2172 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2174 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2176 if (cpu >= nr_cpu_ids)
2177 cpu = cpumask_first(hctx->cpumask);
2182 * It'd be great if the workqueue API had a way to pass
2183 * in a mask and had some smarts for more clever placement.
2184 * For now we just round-robin here, switching for every
2185 * BLK_MQ_CPU_WORK_BATCH queued items.
2187 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2190 int next_cpu = hctx->next_cpu;
2192 if (hctx->queue->nr_hw_queues == 1)
2193 return WORK_CPU_UNBOUND;
2195 if (--hctx->next_cpu_batch <= 0) {
2197 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2199 if (next_cpu >= nr_cpu_ids)
2200 next_cpu = blk_mq_first_mapped_cpu(hctx);
2201 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2205 * Do unbound schedule if we can't find a online CPU for this hctx,
2206 * and it should only happen in the path of handling CPU DEAD.
2208 if (!cpu_online(next_cpu)) {
2215 * Make sure to re-select CPU next time once after CPUs
2216 * in hctx->cpumask become online again.
2218 hctx->next_cpu = next_cpu;
2219 hctx->next_cpu_batch = 1;
2220 return WORK_CPU_UNBOUND;
2223 hctx->next_cpu = next_cpu;
2228 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2229 * @hctx: Pointer to the hardware queue to run.
2230 * @async: If we want to run the queue asynchronously.
2231 * @msecs: Milliseconds of delay to wait before running the queue.
2233 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2234 * with a delay of @msecs.
2236 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2237 unsigned long msecs)
2239 if (unlikely(blk_mq_hctx_stopped(hctx)))
2242 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2243 if (cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2244 __blk_mq_run_hw_queue(hctx);
2249 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2250 msecs_to_jiffies(msecs));
2254 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2255 * @hctx: Pointer to the hardware queue to run.
2256 * @msecs: Milliseconds of delay to wait before running the queue.
2258 * Run a hardware queue asynchronously with a delay of @msecs.
2260 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2262 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2264 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2267 * blk_mq_run_hw_queue - Start to run a hardware queue.
2268 * @hctx: Pointer to the hardware queue to run.
2269 * @async: If we want to run the queue asynchronously.
2271 * Check if the request queue is not in a quiesced state and if there are
2272 * pending requests to be sent. If this is true, run the queue to send requests
2275 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2280 * When queue is quiesced, we may be switching io scheduler, or
2281 * updating nr_hw_queues, or other things, and we can't run queue
2282 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2284 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2287 __blk_mq_run_dispatch_ops(hctx->queue, false,
2288 need_run = !blk_queue_quiesced(hctx->queue) &&
2289 blk_mq_hctx_has_pending(hctx));
2292 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2294 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2297 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2300 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2302 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2304 * If the IO scheduler does not respect hardware queues when
2305 * dispatching, we just don't bother with multiple HW queues and
2306 * dispatch from hctx for the current CPU since running multiple queues
2307 * just causes lock contention inside the scheduler and pointless cache
2310 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2312 if (!blk_mq_hctx_stopped(hctx))
2318 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2319 * @q: Pointer to the request queue to run.
2320 * @async: If we want to run the queue asynchronously.
2322 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2324 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2328 if (blk_queue_sq_sched(q))
2329 sq_hctx = blk_mq_get_sq_hctx(q);
2330 queue_for_each_hw_ctx(q, hctx, i) {
2331 if (blk_mq_hctx_stopped(hctx))
2334 * Dispatch from this hctx either if there's no hctx preferred
2335 * by IO scheduler or if it has requests that bypass the
2338 if (!sq_hctx || sq_hctx == hctx ||
2339 !list_empty_careful(&hctx->dispatch))
2340 blk_mq_run_hw_queue(hctx, async);
2343 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2346 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2347 * @q: Pointer to the request queue to run.
2348 * @msecs: Milliseconds of delay to wait before running the queues.
2350 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2352 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2356 if (blk_queue_sq_sched(q))
2357 sq_hctx = blk_mq_get_sq_hctx(q);
2358 queue_for_each_hw_ctx(q, hctx, i) {
2359 if (blk_mq_hctx_stopped(hctx))
2362 * If there is already a run_work pending, leave the
2363 * pending delay untouched. Otherwise, a hctx can stall
2364 * if another hctx is re-delaying the other's work
2365 * before the work executes.
2367 if (delayed_work_pending(&hctx->run_work))
2370 * Dispatch from this hctx either if there's no hctx preferred
2371 * by IO scheduler or if it has requests that bypass the
2374 if (!sq_hctx || sq_hctx == hctx ||
2375 !list_empty_careful(&hctx->dispatch))
2376 blk_mq_delay_run_hw_queue(hctx, msecs);
2379 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2382 * This function is often used for pausing .queue_rq() by driver when
2383 * there isn't enough resource or some conditions aren't satisfied, and
2384 * BLK_STS_RESOURCE is usually returned.
2386 * We do not guarantee that dispatch can be drained or blocked
2387 * after blk_mq_stop_hw_queue() returns. Please use
2388 * blk_mq_quiesce_queue() for that requirement.
2390 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2392 cancel_delayed_work(&hctx->run_work);
2394 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2396 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2399 * This function is often used for pausing .queue_rq() by driver when
2400 * there isn't enough resource or some conditions aren't satisfied, and
2401 * BLK_STS_RESOURCE is usually returned.
2403 * We do not guarantee that dispatch can be drained or blocked
2404 * after blk_mq_stop_hw_queues() returns. Please use
2405 * blk_mq_quiesce_queue() for that requirement.
2407 void blk_mq_stop_hw_queues(struct request_queue *q)
2409 struct blk_mq_hw_ctx *hctx;
2412 queue_for_each_hw_ctx(q, hctx, i)
2413 blk_mq_stop_hw_queue(hctx);
2415 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2417 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2419 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2421 blk_mq_run_hw_queue(hctx, false);
2423 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2425 void blk_mq_start_hw_queues(struct request_queue *q)
2427 struct blk_mq_hw_ctx *hctx;
2430 queue_for_each_hw_ctx(q, hctx, i)
2431 blk_mq_start_hw_queue(hctx);
2433 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2435 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2437 if (!blk_mq_hctx_stopped(hctx))
2440 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2441 blk_mq_run_hw_queue(hctx, async);
2443 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2445 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2447 struct blk_mq_hw_ctx *hctx;
2450 queue_for_each_hw_ctx(q, hctx, i)
2451 blk_mq_start_stopped_hw_queue(hctx, async);
2453 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2455 static void blk_mq_run_work_fn(struct work_struct *work)
2457 struct blk_mq_hw_ctx *hctx;
2459 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2462 * If we are stopped, don't run the queue.
2464 if (blk_mq_hctx_stopped(hctx))
2467 __blk_mq_run_hw_queue(hctx);
2470 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2474 struct blk_mq_ctx *ctx = rq->mq_ctx;
2475 enum hctx_type type = hctx->type;
2477 lockdep_assert_held(&ctx->lock);
2479 trace_block_rq_insert(rq);
2482 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2484 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2487 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2490 struct blk_mq_ctx *ctx = rq->mq_ctx;
2492 lockdep_assert_held(&ctx->lock);
2494 __blk_mq_insert_req_list(hctx, rq, at_head);
2495 blk_mq_hctx_mark_pending(hctx, ctx);
2499 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2500 * @rq: Pointer to request to be inserted.
2501 * @at_head: true if the request should be inserted at the head of the list.
2502 * @run_queue: If we should run the hardware queue after inserting the request.
2504 * Should only be used carefully, when the caller knows we want to
2505 * bypass a potential IO scheduler on the target device.
2507 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2510 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2512 spin_lock(&hctx->lock);
2514 list_add(&rq->queuelist, &hctx->dispatch);
2516 list_add_tail(&rq->queuelist, &hctx->dispatch);
2517 spin_unlock(&hctx->lock);
2520 blk_mq_run_hw_queue(hctx, false);
2523 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2524 struct list_head *list)
2528 enum hctx_type type = hctx->type;
2531 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2534 list_for_each_entry(rq, list, queuelist) {
2535 BUG_ON(rq->mq_ctx != ctx);
2536 trace_block_rq_insert(rq);
2539 spin_lock(&ctx->lock);
2540 list_splice_tail_init(list, &ctx->rq_lists[type]);
2541 blk_mq_hctx_mark_pending(hctx, ctx);
2542 spin_unlock(&ctx->lock);
2545 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2548 if (hctx->queue->mq_ops->commit_rqs) {
2549 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2550 hctx->queue->mq_ops->commit_rqs(hctx);
2555 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2556 unsigned int nr_segs)
2560 if (bio->bi_opf & REQ_RAHEAD)
2561 rq->cmd_flags |= REQ_FAILFAST_MASK;
2563 rq->__sector = bio->bi_iter.bi_sector;
2564 blk_rq_bio_prep(rq, bio, nr_segs);
2566 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2567 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2570 blk_account_io_start(rq);
2573 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2574 struct request *rq, bool last)
2576 struct request_queue *q = rq->q;
2577 struct blk_mq_queue_data bd = {
2584 * For OK queue, we are done. For error, caller may kill it.
2585 * Any other error (busy), just add it to our list as we
2586 * previously would have done.
2588 ret = q->mq_ops->queue_rq(hctx, &bd);
2591 blk_mq_update_dispatch_busy(hctx, false);
2593 case BLK_STS_RESOURCE:
2594 case BLK_STS_DEV_RESOURCE:
2595 blk_mq_update_dispatch_busy(hctx, true);
2596 __blk_mq_requeue_request(rq);
2599 blk_mq_update_dispatch_busy(hctx, false);
2606 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2608 bool bypass_insert, bool last)
2610 struct request_queue *q = rq->q;
2611 bool run_queue = true;
2615 * RCU or SRCU read lock is needed before checking quiesced flag.
2617 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2618 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2619 * and avoid driver to try to dispatch again.
2621 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2623 bypass_insert = false;
2627 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2630 budget_token = blk_mq_get_dispatch_budget(q);
2631 if (budget_token < 0)
2634 blk_mq_set_rq_budget_token(rq, budget_token);
2636 if (!blk_mq_get_driver_tag(rq)) {
2637 blk_mq_put_dispatch_budget(q, budget_token);
2641 return __blk_mq_issue_directly(hctx, rq, last);
2644 return BLK_STS_RESOURCE;
2646 blk_mq_sched_insert_request(rq, false, run_queue, false);
2652 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2653 * @hctx: Pointer of the associated hardware queue.
2654 * @rq: Pointer to request to be sent.
2656 * If the device has enough resources to accept a new request now, send the
2657 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2658 * we can try send it another time in the future. Requests inserted at this
2659 * queue have higher priority.
2661 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2665 __blk_mq_try_issue_directly(hctx, rq, false, true);
2667 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2668 blk_mq_request_bypass_insert(rq, false, true);
2669 else if (ret != BLK_STS_OK)
2670 blk_mq_end_request(rq, ret);
2673 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2675 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2678 static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2680 struct blk_mq_hw_ctx *hctx = NULL;
2685 while ((rq = rq_list_pop(&plug->mq_list))) {
2686 bool last = rq_list_empty(plug->mq_list);
2689 if (hctx != rq->mq_hctx) {
2691 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2695 ret = blk_mq_request_issue_directly(rq, last);
2700 case BLK_STS_RESOURCE:
2701 case BLK_STS_DEV_RESOURCE:
2702 blk_mq_request_bypass_insert(rq, false, true);
2703 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2706 blk_mq_end_request(rq, ret);
2713 * If we didn't flush the entire list, we could have told the driver
2714 * there was more coming, but that turned out to be a lie.
2717 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2720 static void __blk_mq_flush_plug_list(struct request_queue *q,
2721 struct blk_plug *plug)
2723 if (blk_queue_quiesced(q))
2725 q->mq_ops->queue_rqs(&plug->mq_list);
2728 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2730 struct blk_mq_hw_ctx *this_hctx = NULL;
2731 struct blk_mq_ctx *this_ctx = NULL;
2732 struct request *requeue_list = NULL;
2733 unsigned int depth = 0;
2737 struct request *rq = rq_list_pop(&plug->mq_list);
2740 this_hctx = rq->mq_hctx;
2741 this_ctx = rq->mq_ctx;
2742 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2743 rq_list_add(&requeue_list, rq);
2746 list_add_tail(&rq->queuelist, &list);
2748 } while (!rq_list_empty(plug->mq_list));
2750 plug->mq_list = requeue_list;
2751 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2752 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched);
2755 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2759 if (rq_list_empty(plug->mq_list))
2763 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2764 struct request_queue *q;
2766 rq = rq_list_peek(&plug->mq_list);
2770 * Peek first request and see if we have a ->queue_rqs() hook.
2771 * If we do, we can dispatch the whole plug list in one go. We
2772 * already know at this point that all requests belong to the
2773 * same queue, caller must ensure that's the case.
2775 * Since we pass off the full list to the driver at this point,
2776 * we do not increment the active request count for the queue.
2777 * Bypass shared tags for now because of that.
2779 if (q->mq_ops->queue_rqs &&
2780 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2781 blk_mq_run_dispatch_ops(q,
2782 __blk_mq_flush_plug_list(q, plug));
2783 if (rq_list_empty(plug->mq_list))
2787 blk_mq_run_dispatch_ops(q,
2788 blk_mq_plug_issue_direct(plug, false));
2789 if (rq_list_empty(plug->mq_list))
2794 blk_mq_dispatch_plug_list(plug, from_schedule);
2795 } while (!rq_list_empty(plug->mq_list));
2798 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2799 struct list_head *list)
2804 while (!list_empty(list)) {
2806 struct request *rq = list_first_entry(list, struct request,
2809 list_del_init(&rq->queuelist);
2810 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2811 if (ret != BLK_STS_OK) {
2813 if (ret == BLK_STS_RESOURCE ||
2814 ret == BLK_STS_DEV_RESOURCE) {
2815 blk_mq_request_bypass_insert(rq, false,
2819 blk_mq_end_request(rq, ret);
2825 * If we didn't flush the entire list, we could have told
2826 * the driver there was more coming, but that turned out to
2829 if ((!list_empty(list) || errors) &&
2830 hctx->queue->mq_ops->commit_rqs && queued)
2831 hctx->queue->mq_ops->commit_rqs(hctx);
2834 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2835 struct bio *bio, unsigned int nr_segs)
2837 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2838 if (blk_attempt_plug_merge(q, bio, nr_segs))
2840 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2846 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2847 struct blk_plug *plug,
2851 struct blk_mq_alloc_data data = {
2854 .cmd_flags = bio->bi_opf,
2858 if (unlikely(bio_queue_enter(bio)))
2861 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2864 rq_qos_throttle(q, bio);
2867 data.nr_tags = plug->nr_ios;
2869 data.cached_rq = &plug->cached_rq;
2872 rq = __blk_mq_alloc_requests(&data);
2875 rq_qos_cleanup(q, bio);
2876 if (bio->bi_opf & REQ_NOWAIT)
2877 bio_wouldblock_error(bio);
2883 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2884 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2890 rq = rq_list_peek(&plug->cached_rq);
2891 if (!rq || rq->q != q)
2894 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2899 if (blk_mq_get_hctx_type((*bio)->bi_opf) != rq->mq_hctx->type)
2901 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2905 * If any qos ->throttle() end up blocking, we will have flushed the
2906 * plug and hence killed the cached_rq list as well. Pop this entry
2907 * before we throttle.
2909 plug->cached_rq = rq_list_next(rq);
2910 rq_qos_throttle(q, *bio);
2912 rq->cmd_flags = (*bio)->bi_opf;
2913 INIT_LIST_HEAD(&rq->queuelist);
2917 static void bio_set_ioprio(struct bio *bio)
2919 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2920 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2921 bio->bi_ioprio = get_current_ioprio();
2922 blkcg_set_ioprio(bio);
2926 * blk_mq_submit_bio - Create and send a request to block device.
2927 * @bio: Bio pointer.
2929 * Builds up a request structure from @q and @bio and send to the device. The
2930 * request may not be queued directly to hardware if:
2931 * * This request can be merged with another one
2932 * * We want to place request at plug queue for possible future merging
2933 * * There is an IO scheduler active at this queue
2935 * It will not queue the request if there is an error with the bio, or at the
2938 void blk_mq_submit_bio(struct bio *bio)
2940 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2941 struct blk_plug *plug = blk_mq_plug(bio);
2942 const int is_sync = op_is_sync(bio->bi_opf);
2944 unsigned int nr_segs = 1;
2947 bio = blk_queue_bounce(bio, q);
2948 if (bio_may_exceed_limits(bio, &q->limits))
2949 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2951 if (!bio_integrity_prep(bio))
2954 bio_set_ioprio(bio);
2956 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2960 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2965 trace_block_getrq(bio);
2967 rq_qos_track(q, rq, bio);
2969 blk_mq_bio_to_request(rq, bio, nr_segs);
2971 ret = blk_crypto_init_request(rq);
2972 if (ret != BLK_STS_OK) {
2973 bio->bi_status = ret;
2975 blk_mq_free_request(rq);
2979 if (op_is_flush(bio->bi_opf)) {
2980 blk_insert_flush(rq);
2985 blk_add_rq_to_plug(plug, rq);
2986 else if ((rq->rq_flags & RQF_ELV) ||
2987 (rq->mq_hctx->dispatch_busy &&
2988 (q->nr_hw_queues == 1 || !is_sync)))
2989 blk_mq_sched_insert_request(rq, false, true, true);
2991 blk_mq_run_dispatch_ops(rq->q,
2992 blk_mq_try_issue_directly(rq->mq_hctx, rq));
2995 #ifdef CONFIG_BLK_MQ_STACKING
2997 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
2998 * @rq: the request being queued
3000 blk_status_t blk_insert_cloned_request(struct request *rq)
3002 struct request_queue *q = rq->q;
3003 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3006 if (blk_rq_sectors(rq) > max_sectors) {
3008 * SCSI device does not have a good way to return if
3009 * Write Same/Zero is actually supported. If a device rejects
3010 * a non-read/write command (discard, write same,etc.) the
3011 * low-level device driver will set the relevant queue limit to
3012 * 0 to prevent blk-lib from issuing more of the offending
3013 * operations. Commands queued prior to the queue limit being
3014 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3015 * errors being propagated to upper layers.
3017 if (max_sectors == 0)
3018 return BLK_STS_NOTSUPP;
3020 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3021 __func__, blk_rq_sectors(rq), max_sectors);
3022 return BLK_STS_IOERR;
3026 * The queue settings related to segment counting may differ from the
3029 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3030 if (rq->nr_phys_segments > queue_max_segments(q)) {
3031 printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n",
3032 __func__, rq->nr_phys_segments, queue_max_segments(q));
3033 return BLK_STS_IOERR;
3036 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3037 return BLK_STS_IOERR;
3039 if (blk_crypto_insert_cloned_request(rq))
3040 return BLK_STS_IOERR;
3042 blk_account_io_start(rq);
3045 * Since we have a scheduler attached on the top device,
3046 * bypass a potential scheduler on the bottom device for
3049 blk_mq_run_dispatch_ops(q,
3050 ret = blk_mq_request_issue_directly(rq, true));
3052 blk_account_io_done(rq, ktime_get_ns());
3055 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3058 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3059 * @rq: the clone request to be cleaned up
3062 * Free all bios in @rq for a cloned request.
3064 void blk_rq_unprep_clone(struct request *rq)
3068 while ((bio = rq->bio) != NULL) {
3069 rq->bio = bio->bi_next;
3074 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3077 * blk_rq_prep_clone - Helper function to setup clone request
3078 * @rq: the request to be setup
3079 * @rq_src: original request to be cloned
3080 * @bs: bio_set that bios for clone are allocated from
3081 * @gfp_mask: memory allocation mask for bio
3082 * @bio_ctr: setup function to be called for each clone bio.
3083 * Returns %0 for success, non %0 for failure.
3084 * @data: private data to be passed to @bio_ctr
3087 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3088 * Also, pages which the original bios are pointing to are not copied
3089 * and the cloned bios just point same pages.
3090 * So cloned bios must be completed before original bios, which means
3091 * the caller must complete @rq before @rq_src.
3093 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3094 struct bio_set *bs, gfp_t gfp_mask,
3095 int (*bio_ctr)(struct bio *, struct bio *, void *),
3098 struct bio *bio, *bio_src;
3103 __rq_for_each_bio(bio_src, rq_src) {
3104 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3109 if (bio_ctr && bio_ctr(bio, bio_src, data))
3113 rq->biotail->bi_next = bio;
3116 rq->bio = rq->biotail = bio;
3121 /* Copy attributes of the original request to the clone request. */
3122 rq->__sector = blk_rq_pos(rq_src);
3123 rq->__data_len = blk_rq_bytes(rq_src);
3124 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3125 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3126 rq->special_vec = rq_src->special_vec;
3128 rq->nr_phys_segments = rq_src->nr_phys_segments;
3129 rq->ioprio = rq_src->ioprio;
3131 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3139 blk_rq_unprep_clone(rq);
3143 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3144 #endif /* CONFIG_BLK_MQ_STACKING */
3147 * Steal bios from a request and add them to a bio list.
3148 * The request must not have been partially completed before.
3150 void blk_steal_bios(struct bio_list *list, struct request *rq)
3154 list->tail->bi_next = rq->bio;
3156 list->head = rq->bio;
3157 list->tail = rq->biotail;
3165 EXPORT_SYMBOL_GPL(blk_steal_bios);
3167 static size_t order_to_size(unsigned int order)
3169 return (size_t)PAGE_SIZE << order;
3172 /* called before freeing request pool in @tags */
3173 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3174 struct blk_mq_tags *tags)
3177 unsigned long flags;
3180 * There is no need to clear mapping if driver tags is not initialized
3181 * or the mapping belongs to the driver tags.
3183 if (!drv_tags || drv_tags == tags)
3186 list_for_each_entry(page, &tags->page_list, lru) {
3187 unsigned long start = (unsigned long)page_address(page);
3188 unsigned long end = start + order_to_size(page->private);
3191 for (i = 0; i < drv_tags->nr_tags; i++) {
3192 struct request *rq = drv_tags->rqs[i];
3193 unsigned long rq_addr = (unsigned long)rq;
3195 if (rq_addr >= start && rq_addr < end) {
3196 WARN_ON_ONCE(req_ref_read(rq) != 0);
3197 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3203 * Wait until all pending iteration is done.
3205 * Request reference is cleared and it is guaranteed to be observed
3206 * after the ->lock is released.
3208 spin_lock_irqsave(&drv_tags->lock, flags);
3209 spin_unlock_irqrestore(&drv_tags->lock, flags);
3212 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3213 unsigned int hctx_idx)
3215 struct blk_mq_tags *drv_tags;
3218 if (list_empty(&tags->page_list))
3221 if (blk_mq_is_shared_tags(set->flags))
3222 drv_tags = set->shared_tags;
3224 drv_tags = set->tags[hctx_idx];
3226 if (tags->static_rqs && set->ops->exit_request) {
3229 for (i = 0; i < tags->nr_tags; i++) {
3230 struct request *rq = tags->static_rqs[i];
3234 set->ops->exit_request(set, rq, hctx_idx);
3235 tags->static_rqs[i] = NULL;
3239 blk_mq_clear_rq_mapping(drv_tags, tags);
3241 while (!list_empty(&tags->page_list)) {
3242 page = list_first_entry(&tags->page_list, struct page, lru);
3243 list_del_init(&page->lru);
3245 * Remove kmemleak object previously allocated in
3246 * blk_mq_alloc_rqs().
3248 kmemleak_free(page_address(page));
3249 __free_pages(page, page->private);
3253 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3257 kfree(tags->static_rqs);
3258 tags->static_rqs = NULL;
3260 blk_mq_free_tags(tags);
3263 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3264 unsigned int hctx_idx)
3268 for (i = 0; i < set->nr_maps; i++) {
3269 unsigned int start = set->map[i].queue_offset;
3270 unsigned int end = start + set->map[i].nr_queues;
3272 if (hctx_idx >= start && hctx_idx < end)
3276 if (i >= set->nr_maps)
3277 i = HCTX_TYPE_DEFAULT;
3282 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3283 unsigned int hctx_idx)
3285 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3287 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3290 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3291 unsigned int hctx_idx,
3292 unsigned int nr_tags,
3293 unsigned int reserved_tags)
3295 int node = blk_mq_get_hctx_node(set, hctx_idx);
3296 struct blk_mq_tags *tags;
3298 if (node == NUMA_NO_NODE)
3299 node = set->numa_node;
3301 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3302 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3306 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3307 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3312 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3313 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3315 if (!tags->static_rqs)
3323 blk_mq_free_tags(tags);
3327 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3328 unsigned int hctx_idx, int node)
3332 if (set->ops->init_request) {
3333 ret = set->ops->init_request(set, rq, hctx_idx, node);
3338 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3342 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3343 struct blk_mq_tags *tags,
3344 unsigned int hctx_idx, unsigned int depth)
3346 unsigned int i, j, entries_per_page, max_order = 4;
3347 int node = blk_mq_get_hctx_node(set, hctx_idx);
3348 size_t rq_size, left;
3350 if (node == NUMA_NO_NODE)
3351 node = set->numa_node;
3353 INIT_LIST_HEAD(&tags->page_list);
3356 * rq_size is the size of the request plus driver payload, rounded
3357 * to the cacheline size
3359 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3361 left = rq_size * depth;
3363 for (i = 0; i < depth; ) {
3364 int this_order = max_order;
3369 while (this_order && left < order_to_size(this_order - 1))
3373 page = alloc_pages_node(node,
3374 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3380 if (order_to_size(this_order) < rq_size)
3387 page->private = this_order;
3388 list_add_tail(&page->lru, &tags->page_list);
3390 p = page_address(page);
3392 * Allow kmemleak to scan these pages as they contain pointers
3393 * to additional allocations like via ops->init_request().
3395 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3396 entries_per_page = order_to_size(this_order) / rq_size;
3397 to_do = min(entries_per_page, depth - i);
3398 left -= to_do * rq_size;
3399 for (j = 0; j < to_do; j++) {
3400 struct request *rq = p;
3402 tags->static_rqs[i] = rq;
3403 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3404 tags->static_rqs[i] = NULL;
3415 blk_mq_free_rqs(set, tags, hctx_idx);
3419 struct rq_iter_data {
3420 struct blk_mq_hw_ctx *hctx;
3424 static bool blk_mq_has_request(struct request *rq, void *data)
3426 struct rq_iter_data *iter_data = data;
3428 if (rq->mq_hctx != iter_data->hctx)
3430 iter_data->has_rq = true;
3434 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3436 struct blk_mq_tags *tags = hctx->sched_tags ?
3437 hctx->sched_tags : hctx->tags;
3438 struct rq_iter_data data = {
3442 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3446 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3447 struct blk_mq_hw_ctx *hctx)
3449 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3451 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3456 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3458 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3459 struct blk_mq_hw_ctx, cpuhp_online);
3461 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3462 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3466 * Prevent new request from being allocated on the current hctx.
3468 * The smp_mb__after_atomic() Pairs with the implied barrier in
3469 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3470 * seen once we return from the tag allocator.
3472 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3473 smp_mb__after_atomic();
3476 * Try to grab a reference to the queue and wait for any outstanding
3477 * requests. If we could not grab a reference the queue has been
3478 * frozen and there are no requests.
3480 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3481 while (blk_mq_hctx_has_requests(hctx))
3483 percpu_ref_put(&hctx->queue->q_usage_counter);
3489 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3491 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3492 struct blk_mq_hw_ctx, cpuhp_online);
3494 if (cpumask_test_cpu(cpu, hctx->cpumask))
3495 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3500 * 'cpu' is going away. splice any existing rq_list entries from this
3501 * software queue to the hw queue dispatch list, and ensure that it
3504 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3506 struct blk_mq_hw_ctx *hctx;
3507 struct blk_mq_ctx *ctx;
3509 enum hctx_type type;
3511 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3512 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3515 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3518 spin_lock(&ctx->lock);
3519 if (!list_empty(&ctx->rq_lists[type])) {
3520 list_splice_init(&ctx->rq_lists[type], &tmp);
3521 blk_mq_hctx_clear_pending(hctx, ctx);
3523 spin_unlock(&ctx->lock);
3525 if (list_empty(&tmp))
3528 spin_lock(&hctx->lock);
3529 list_splice_tail_init(&tmp, &hctx->dispatch);
3530 spin_unlock(&hctx->lock);
3532 blk_mq_run_hw_queue(hctx, true);
3536 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3538 if (!(hctx->flags & BLK_MQ_F_STACKING))
3539 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3540 &hctx->cpuhp_online);
3541 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3546 * Before freeing hw queue, clearing the flush request reference in
3547 * tags->rqs[] for avoiding potential UAF.
3549 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3550 unsigned int queue_depth, struct request *flush_rq)
3553 unsigned long flags;
3555 /* The hw queue may not be mapped yet */
3559 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3561 for (i = 0; i < queue_depth; i++)
3562 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3565 * Wait until all pending iteration is done.
3567 * Request reference is cleared and it is guaranteed to be observed
3568 * after the ->lock is released.
3570 spin_lock_irqsave(&tags->lock, flags);
3571 spin_unlock_irqrestore(&tags->lock, flags);
3574 /* hctx->ctxs will be freed in queue's release handler */
3575 static void blk_mq_exit_hctx(struct request_queue *q,
3576 struct blk_mq_tag_set *set,
3577 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3579 struct request *flush_rq = hctx->fq->flush_rq;
3581 if (blk_mq_hw_queue_mapped(hctx))
3582 blk_mq_tag_idle(hctx);
3584 if (blk_queue_init_done(q))
3585 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3586 set->queue_depth, flush_rq);
3587 if (set->ops->exit_request)
3588 set->ops->exit_request(set, flush_rq, hctx_idx);
3590 if (set->ops->exit_hctx)
3591 set->ops->exit_hctx(hctx, hctx_idx);
3593 blk_mq_remove_cpuhp(hctx);
3595 xa_erase(&q->hctx_table, hctx_idx);
3597 spin_lock(&q->unused_hctx_lock);
3598 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3599 spin_unlock(&q->unused_hctx_lock);
3602 static void blk_mq_exit_hw_queues(struct request_queue *q,
3603 struct blk_mq_tag_set *set, int nr_queue)
3605 struct blk_mq_hw_ctx *hctx;
3608 queue_for_each_hw_ctx(q, hctx, i) {
3611 blk_mq_exit_hctx(q, set, hctx, i);
3615 static int blk_mq_init_hctx(struct request_queue *q,
3616 struct blk_mq_tag_set *set,
3617 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3619 hctx->queue_num = hctx_idx;
3621 if (!(hctx->flags & BLK_MQ_F_STACKING))
3622 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3623 &hctx->cpuhp_online);
3624 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3626 hctx->tags = set->tags[hctx_idx];
3628 if (set->ops->init_hctx &&
3629 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3630 goto unregister_cpu_notifier;
3632 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3636 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3642 if (set->ops->exit_request)
3643 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3645 if (set->ops->exit_hctx)
3646 set->ops->exit_hctx(hctx, hctx_idx);
3647 unregister_cpu_notifier:
3648 blk_mq_remove_cpuhp(hctx);
3652 static struct blk_mq_hw_ctx *
3653 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3656 struct blk_mq_hw_ctx *hctx;
3657 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3659 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3661 goto fail_alloc_hctx;
3663 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3666 atomic_set(&hctx->nr_active, 0);
3667 if (node == NUMA_NO_NODE)
3668 node = set->numa_node;
3669 hctx->numa_node = node;
3671 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3672 spin_lock_init(&hctx->lock);
3673 INIT_LIST_HEAD(&hctx->dispatch);
3675 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3677 INIT_LIST_HEAD(&hctx->hctx_list);
3680 * Allocate space for all possible cpus to avoid allocation at
3683 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3688 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3689 gfp, node, false, false))
3693 spin_lock_init(&hctx->dispatch_wait_lock);
3694 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3695 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3697 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3701 blk_mq_hctx_kobj_init(hctx);
3706 sbitmap_free(&hctx->ctx_map);
3710 free_cpumask_var(hctx->cpumask);
3717 static void blk_mq_init_cpu_queues(struct request_queue *q,
3718 unsigned int nr_hw_queues)
3720 struct blk_mq_tag_set *set = q->tag_set;
3723 for_each_possible_cpu(i) {
3724 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3725 struct blk_mq_hw_ctx *hctx;
3729 spin_lock_init(&__ctx->lock);
3730 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3731 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3736 * Set local node, IFF we have more than one hw queue. If
3737 * not, we remain on the home node of the device
3739 for (j = 0; j < set->nr_maps; j++) {
3740 hctx = blk_mq_map_queue_type(q, j, i);
3741 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3742 hctx->numa_node = cpu_to_node(i);
3747 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3748 unsigned int hctx_idx,
3751 struct blk_mq_tags *tags;
3754 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3758 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3760 blk_mq_free_rq_map(tags);
3767 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3770 if (blk_mq_is_shared_tags(set->flags)) {
3771 set->tags[hctx_idx] = set->shared_tags;
3776 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3779 return set->tags[hctx_idx];
3782 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3783 struct blk_mq_tags *tags,
3784 unsigned int hctx_idx)
3787 blk_mq_free_rqs(set, tags, hctx_idx);
3788 blk_mq_free_rq_map(tags);
3792 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3793 unsigned int hctx_idx)
3795 if (!blk_mq_is_shared_tags(set->flags))
3796 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3798 set->tags[hctx_idx] = NULL;
3801 static void blk_mq_map_swqueue(struct request_queue *q)
3803 unsigned int j, hctx_idx;
3805 struct blk_mq_hw_ctx *hctx;
3806 struct blk_mq_ctx *ctx;
3807 struct blk_mq_tag_set *set = q->tag_set;
3809 queue_for_each_hw_ctx(q, hctx, i) {
3810 cpumask_clear(hctx->cpumask);
3812 hctx->dispatch_from = NULL;
3816 * Map software to hardware queues.
3818 * If the cpu isn't present, the cpu is mapped to first hctx.
3820 for_each_possible_cpu(i) {
3822 ctx = per_cpu_ptr(q->queue_ctx, i);
3823 for (j = 0; j < set->nr_maps; j++) {
3824 if (!set->map[j].nr_queues) {
3825 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3826 HCTX_TYPE_DEFAULT, i);
3829 hctx_idx = set->map[j].mq_map[i];
3830 /* unmapped hw queue can be remapped after CPU topo changed */
3831 if (!set->tags[hctx_idx] &&
3832 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3834 * If tags initialization fail for some hctx,
3835 * that hctx won't be brought online. In this
3836 * case, remap the current ctx to hctx[0] which
3837 * is guaranteed to always have tags allocated
3839 set->map[j].mq_map[i] = 0;
3842 hctx = blk_mq_map_queue_type(q, j, i);
3843 ctx->hctxs[j] = hctx;
3845 * If the CPU is already set in the mask, then we've
3846 * mapped this one already. This can happen if
3847 * devices share queues across queue maps.
3849 if (cpumask_test_cpu(i, hctx->cpumask))
3852 cpumask_set_cpu(i, hctx->cpumask);
3854 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3855 hctx->ctxs[hctx->nr_ctx++] = ctx;
3858 * If the nr_ctx type overflows, we have exceeded the
3859 * amount of sw queues we can support.
3861 BUG_ON(!hctx->nr_ctx);
3864 for (; j < HCTX_MAX_TYPES; j++)
3865 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3866 HCTX_TYPE_DEFAULT, i);
3869 queue_for_each_hw_ctx(q, hctx, i) {
3871 * If no software queues are mapped to this hardware queue,
3872 * disable it and free the request entries.
3874 if (!hctx->nr_ctx) {
3875 /* Never unmap queue 0. We need it as a
3876 * fallback in case of a new remap fails
3880 __blk_mq_free_map_and_rqs(set, i);
3886 hctx->tags = set->tags[i];
3887 WARN_ON(!hctx->tags);
3890 * Set the map size to the number of mapped software queues.
3891 * This is more accurate and more efficient than looping
3892 * over all possibly mapped software queues.
3894 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3897 * Initialize batch roundrobin counts
3899 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3900 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3905 * Caller needs to ensure that we're either frozen/quiesced, or that
3906 * the queue isn't live yet.
3908 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3910 struct blk_mq_hw_ctx *hctx;
3913 queue_for_each_hw_ctx(q, hctx, i) {
3915 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3917 blk_mq_tag_idle(hctx);
3918 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3923 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3926 struct request_queue *q;
3928 lockdep_assert_held(&set->tag_list_lock);
3930 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3931 blk_mq_freeze_queue(q);
3932 queue_set_hctx_shared(q, shared);
3933 blk_mq_unfreeze_queue(q);
3937 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3939 struct blk_mq_tag_set *set = q->tag_set;
3941 mutex_lock(&set->tag_list_lock);
3942 list_del(&q->tag_set_list);
3943 if (list_is_singular(&set->tag_list)) {
3944 /* just transitioned to unshared */
3945 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3946 /* update existing queue */
3947 blk_mq_update_tag_set_shared(set, false);
3949 mutex_unlock(&set->tag_list_lock);
3950 INIT_LIST_HEAD(&q->tag_set_list);
3953 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3954 struct request_queue *q)
3956 mutex_lock(&set->tag_list_lock);
3959 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3961 if (!list_empty(&set->tag_list) &&
3962 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3963 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3964 /* update existing queue */
3965 blk_mq_update_tag_set_shared(set, true);
3967 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3968 queue_set_hctx_shared(q, true);
3969 list_add_tail(&q->tag_set_list, &set->tag_list);
3971 mutex_unlock(&set->tag_list_lock);
3974 /* All allocations will be freed in release handler of q->mq_kobj */
3975 static int blk_mq_alloc_ctxs(struct request_queue *q)
3977 struct blk_mq_ctxs *ctxs;
3980 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3984 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3985 if (!ctxs->queue_ctx)
3988 for_each_possible_cpu(cpu) {
3989 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3993 q->mq_kobj = &ctxs->kobj;
3994 q->queue_ctx = ctxs->queue_ctx;
4003 * It is the actual release handler for mq, but we do it from
4004 * request queue's release handler for avoiding use-after-free
4005 * and headache because q->mq_kobj shouldn't have been introduced,
4006 * but we can't group ctx/kctx kobj without it.
4008 void blk_mq_release(struct request_queue *q)
4010 struct blk_mq_hw_ctx *hctx, *next;
4013 queue_for_each_hw_ctx(q, hctx, i)
4014 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4016 /* all hctx are in .unused_hctx_list now */
4017 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4018 list_del_init(&hctx->hctx_list);
4019 kobject_put(&hctx->kobj);
4022 xa_destroy(&q->hctx_table);
4025 * release .mq_kobj and sw queue's kobject now because
4026 * both share lifetime with request queue.
4028 blk_mq_sysfs_deinit(q);
4031 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4034 struct request_queue *q;
4037 q = blk_alloc_queue(set->numa_node);
4039 return ERR_PTR(-ENOMEM);
4040 q->queuedata = queuedata;
4041 ret = blk_mq_init_allocated_queue(set, q);
4044 return ERR_PTR(ret);
4049 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4051 return blk_mq_init_queue_data(set, NULL);
4053 EXPORT_SYMBOL(blk_mq_init_queue);
4056 * blk_mq_destroy_queue - shutdown a request queue
4057 * @q: request queue to shutdown
4059 * This shuts down a request queue allocated by blk_mq_init_queue() and drops
4060 * the initial reference. All future requests will failed with -ENODEV.
4062 * Context: can sleep
4064 void blk_mq_destroy_queue(struct request_queue *q)
4066 WARN_ON_ONCE(!queue_is_mq(q));
4067 WARN_ON_ONCE(blk_queue_registered(q));
4071 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4072 blk_queue_start_drain(q);
4073 blk_mq_freeze_queue_wait(q);
4076 blk_mq_cancel_work_sync(q);
4077 blk_mq_exit_queue(q);
4079 EXPORT_SYMBOL(blk_mq_destroy_queue);
4081 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4082 struct lock_class_key *lkclass)
4084 struct request_queue *q;
4085 struct gendisk *disk;
4087 q = blk_mq_init_queue_data(set, queuedata);
4091 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4093 blk_mq_destroy_queue(q);
4095 return ERR_PTR(-ENOMEM);
4097 set_bit(GD_OWNS_QUEUE, &disk->state);
4100 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4102 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4103 struct lock_class_key *lkclass)
4105 if (!blk_get_queue(q))
4107 return __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4109 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4111 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4112 struct blk_mq_tag_set *set, struct request_queue *q,
4113 int hctx_idx, int node)
4115 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4117 /* reuse dead hctx first */
4118 spin_lock(&q->unused_hctx_lock);
4119 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4120 if (tmp->numa_node == node) {
4126 list_del_init(&hctx->hctx_list);
4127 spin_unlock(&q->unused_hctx_lock);
4130 hctx = blk_mq_alloc_hctx(q, set, node);
4134 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4140 kobject_put(&hctx->kobj);
4145 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4146 struct request_queue *q)
4148 struct blk_mq_hw_ctx *hctx;
4151 /* protect against switching io scheduler */
4152 mutex_lock(&q->sysfs_lock);
4153 for (i = 0; i < set->nr_hw_queues; i++) {
4155 int node = blk_mq_get_hctx_node(set, i);
4156 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4159 old_node = old_hctx->numa_node;
4160 blk_mq_exit_hctx(q, set, old_hctx, i);
4163 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4166 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4168 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4169 WARN_ON_ONCE(!hctx);
4173 * Increasing nr_hw_queues fails. Free the newly allocated
4174 * hctxs and keep the previous q->nr_hw_queues.
4176 if (i != set->nr_hw_queues) {
4177 j = q->nr_hw_queues;
4180 q->nr_hw_queues = set->nr_hw_queues;
4183 xa_for_each_start(&q->hctx_table, j, hctx, j)
4184 blk_mq_exit_hctx(q, set, hctx, j);
4185 mutex_unlock(&q->sysfs_lock);
4188 static void blk_mq_update_poll_flag(struct request_queue *q)
4190 struct blk_mq_tag_set *set = q->tag_set;
4192 if (set->nr_maps > HCTX_TYPE_POLL &&
4193 set->map[HCTX_TYPE_POLL].nr_queues)
4194 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4196 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4199 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4200 struct request_queue *q)
4202 /* mark the queue as mq asap */
4203 q->mq_ops = set->ops;
4205 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
4206 blk_mq_poll_stats_bkt,
4207 BLK_MQ_POLL_STATS_BKTS, q);
4211 if (blk_mq_alloc_ctxs(q))
4214 /* init q->mq_kobj and sw queues' kobjects */
4215 blk_mq_sysfs_init(q);
4217 INIT_LIST_HEAD(&q->unused_hctx_list);
4218 spin_lock_init(&q->unused_hctx_lock);
4220 xa_init(&q->hctx_table);
4222 blk_mq_realloc_hw_ctxs(set, q);
4223 if (!q->nr_hw_queues)
4226 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4227 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4231 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4232 blk_mq_update_poll_flag(q);
4234 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4235 INIT_LIST_HEAD(&q->requeue_list);
4236 spin_lock_init(&q->requeue_lock);
4238 q->nr_requests = set->queue_depth;
4241 * Default to classic polling
4243 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
4245 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4246 blk_mq_add_queue_tag_set(set, q);
4247 blk_mq_map_swqueue(q);
4251 xa_destroy(&q->hctx_table);
4252 q->nr_hw_queues = 0;
4253 blk_mq_sysfs_deinit(q);
4255 blk_stat_free_callback(q->poll_cb);
4261 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4263 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4264 void blk_mq_exit_queue(struct request_queue *q)
4266 struct blk_mq_tag_set *set = q->tag_set;
4268 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4269 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4270 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4271 blk_mq_del_queue_tag_set(q);
4274 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4278 if (blk_mq_is_shared_tags(set->flags)) {
4279 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4282 if (!set->shared_tags)
4286 for (i = 0; i < set->nr_hw_queues; i++) {
4287 if (!__blk_mq_alloc_map_and_rqs(set, i))
4296 __blk_mq_free_map_and_rqs(set, i);
4298 if (blk_mq_is_shared_tags(set->flags)) {
4299 blk_mq_free_map_and_rqs(set, set->shared_tags,
4300 BLK_MQ_NO_HCTX_IDX);
4307 * Allocate the request maps associated with this tag_set. Note that this
4308 * may reduce the depth asked for, if memory is tight. set->queue_depth
4309 * will be updated to reflect the allocated depth.
4311 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4316 depth = set->queue_depth;
4318 err = __blk_mq_alloc_rq_maps(set);
4322 set->queue_depth >>= 1;
4323 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4327 } while (set->queue_depth);
4329 if (!set->queue_depth || err) {
4330 pr_err("blk-mq: failed to allocate request map\n");
4334 if (depth != set->queue_depth)
4335 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4336 depth, set->queue_depth);
4341 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4344 * blk_mq_map_queues() and multiple .map_queues() implementations
4345 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4346 * number of hardware queues.
4348 if (set->nr_maps == 1)
4349 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4351 if (set->ops->map_queues && !is_kdump_kernel()) {
4355 * transport .map_queues is usually done in the following
4358 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4359 * mask = get_cpu_mask(queue)
4360 * for_each_cpu(cpu, mask)
4361 * set->map[x].mq_map[cpu] = queue;
4364 * When we need to remap, the table has to be cleared for
4365 * killing stale mapping since one CPU may not be mapped
4368 for (i = 0; i < set->nr_maps; i++)
4369 blk_mq_clear_mq_map(&set->map[i]);
4371 set->ops->map_queues(set);
4373 BUG_ON(set->nr_maps > 1);
4374 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4378 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4379 int new_nr_hw_queues)
4381 struct blk_mq_tags **new_tags;
4383 if (set->nr_hw_queues >= new_nr_hw_queues)
4386 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4387 GFP_KERNEL, set->numa_node);
4392 memcpy(new_tags, set->tags, set->nr_hw_queues *
4393 sizeof(*set->tags));
4395 set->tags = new_tags;
4397 set->nr_hw_queues = new_nr_hw_queues;
4402 * Alloc a tag set to be associated with one or more request queues.
4403 * May fail with EINVAL for various error conditions. May adjust the
4404 * requested depth down, if it's too large. In that case, the set
4405 * value will be stored in set->queue_depth.
4407 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4411 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4413 if (!set->nr_hw_queues)
4415 if (!set->queue_depth)
4417 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4420 if (!set->ops->queue_rq)
4423 if (!set->ops->get_budget ^ !set->ops->put_budget)
4426 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4427 pr_info("blk-mq: reduced tag depth to %u\n",
4429 set->queue_depth = BLK_MQ_MAX_DEPTH;
4434 else if (set->nr_maps > HCTX_MAX_TYPES)
4438 * If a crashdump is active, then we are potentially in a very
4439 * memory constrained environment. Limit us to 1 queue and
4440 * 64 tags to prevent using too much memory.
4442 if (is_kdump_kernel()) {
4443 set->nr_hw_queues = 1;
4445 set->queue_depth = min(64U, set->queue_depth);
4448 * There is no use for more h/w queues than cpus if we just have
4451 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4452 set->nr_hw_queues = nr_cpu_ids;
4454 if (set->flags & BLK_MQ_F_BLOCKING) {
4455 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4458 ret = init_srcu_struct(set->srcu);
4464 set->tags = kcalloc_node(set->nr_hw_queues,
4465 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4468 goto out_cleanup_srcu;
4470 for (i = 0; i < set->nr_maps; i++) {
4471 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4472 sizeof(set->map[i].mq_map[0]),
4473 GFP_KERNEL, set->numa_node);
4474 if (!set->map[i].mq_map)
4475 goto out_free_mq_map;
4476 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4479 blk_mq_update_queue_map(set);
4481 ret = blk_mq_alloc_set_map_and_rqs(set);
4483 goto out_free_mq_map;
4485 mutex_init(&set->tag_list_lock);
4486 INIT_LIST_HEAD(&set->tag_list);
4491 for (i = 0; i < set->nr_maps; i++) {
4492 kfree(set->map[i].mq_map);
4493 set->map[i].mq_map = NULL;
4498 if (set->flags & BLK_MQ_F_BLOCKING)
4499 cleanup_srcu_struct(set->srcu);
4501 if (set->flags & BLK_MQ_F_BLOCKING)
4505 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4507 /* allocate and initialize a tagset for a simple single-queue device */
4508 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4509 const struct blk_mq_ops *ops, unsigned int queue_depth,
4510 unsigned int set_flags)
4512 memset(set, 0, sizeof(*set));
4514 set->nr_hw_queues = 1;
4516 set->queue_depth = queue_depth;
4517 set->numa_node = NUMA_NO_NODE;
4518 set->flags = set_flags;
4519 return blk_mq_alloc_tag_set(set);
4521 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4523 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4527 for (i = 0; i < set->nr_hw_queues; i++)
4528 __blk_mq_free_map_and_rqs(set, i);
4530 if (blk_mq_is_shared_tags(set->flags)) {
4531 blk_mq_free_map_and_rqs(set, set->shared_tags,
4532 BLK_MQ_NO_HCTX_IDX);
4535 for (j = 0; j < set->nr_maps; j++) {
4536 kfree(set->map[j].mq_map);
4537 set->map[j].mq_map = NULL;
4542 if (set->flags & BLK_MQ_F_BLOCKING) {
4543 cleanup_srcu_struct(set->srcu);
4547 EXPORT_SYMBOL(blk_mq_free_tag_set);
4549 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4551 struct blk_mq_tag_set *set = q->tag_set;
4552 struct blk_mq_hw_ctx *hctx;
4559 if (q->nr_requests == nr)
4562 blk_mq_freeze_queue(q);
4563 blk_mq_quiesce_queue(q);
4566 queue_for_each_hw_ctx(q, hctx, i) {
4570 * If we're using an MQ scheduler, just update the scheduler
4571 * queue depth. This is similar to what the old code would do.
4573 if (hctx->sched_tags) {
4574 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4577 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4582 if (q->elevator && q->elevator->type->ops.depth_updated)
4583 q->elevator->type->ops.depth_updated(hctx);
4586 q->nr_requests = nr;
4587 if (blk_mq_is_shared_tags(set->flags)) {
4589 blk_mq_tag_update_sched_shared_tags(q);
4591 blk_mq_tag_resize_shared_tags(set, nr);
4595 blk_mq_unquiesce_queue(q);
4596 blk_mq_unfreeze_queue(q);
4602 * request_queue and elevator_type pair.
4603 * It is just used by __blk_mq_update_nr_hw_queues to cache
4604 * the elevator_type associated with a request_queue.
4606 struct blk_mq_qe_pair {
4607 struct list_head node;
4608 struct request_queue *q;
4609 struct elevator_type *type;
4613 * Cache the elevator_type in qe pair list and switch the
4614 * io scheduler to 'none'
4616 static bool blk_mq_elv_switch_none(struct list_head *head,
4617 struct request_queue *q)
4619 struct blk_mq_qe_pair *qe;
4624 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4628 /* q->elevator needs protection from ->sysfs_lock */
4629 mutex_lock(&q->sysfs_lock);
4631 INIT_LIST_HEAD(&qe->node);
4633 qe->type = q->elevator->type;
4634 /* keep a reference to the elevator module as we'll switch back */
4635 __elevator_get(qe->type);
4636 list_add(&qe->node, head);
4637 elevator_disable(q);
4638 mutex_unlock(&q->sysfs_lock);
4643 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4644 struct request_queue *q)
4646 struct blk_mq_qe_pair *qe;
4648 list_for_each_entry(qe, head, node)
4655 static void blk_mq_elv_switch_back(struct list_head *head,
4656 struct request_queue *q)
4658 struct blk_mq_qe_pair *qe;
4659 struct elevator_type *t;
4661 qe = blk_lookup_qe_pair(head, q);
4665 list_del(&qe->node);
4668 mutex_lock(&q->sysfs_lock);
4669 elevator_switch(q, t);
4670 /* drop the reference acquired in blk_mq_elv_switch_none */
4672 mutex_unlock(&q->sysfs_lock);
4675 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4678 struct request_queue *q;
4680 int prev_nr_hw_queues;
4682 lockdep_assert_held(&set->tag_list_lock);
4684 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4685 nr_hw_queues = nr_cpu_ids;
4686 if (nr_hw_queues < 1)
4688 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4691 list_for_each_entry(q, &set->tag_list, tag_set_list)
4692 blk_mq_freeze_queue(q);
4694 * Switch IO scheduler to 'none', cleaning up the data associated
4695 * with the previous scheduler. We will switch back once we are done
4696 * updating the new sw to hw queue mappings.
4698 list_for_each_entry(q, &set->tag_list, tag_set_list)
4699 if (!blk_mq_elv_switch_none(&head, q))
4702 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4703 blk_mq_debugfs_unregister_hctxs(q);
4704 blk_mq_sysfs_unregister_hctxs(q);
4707 prev_nr_hw_queues = set->nr_hw_queues;
4708 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4712 blk_mq_update_queue_map(set);
4713 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4714 blk_mq_realloc_hw_ctxs(set, q);
4715 blk_mq_update_poll_flag(q);
4716 if (q->nr_hw_queues != set->nr_hw_queues) {
4717 int i = prev_nr_hw_queues;
4719 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4720 nr_hw_queues, prev_nr_hw_queues);
4721 for (; i < set->nr_hw_queues; i++)
4722 __blk_mq_free_map_and_rqs(set, i);
4724 set->nr_hw_queues = prev_nr_hw_queues;
4725 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4728 blk_mq_map_swqueue(q);
4732 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4733 blk_mq_sysfs_register_hctxs(q);
4734 blk_mq_debugfs_register_hctxs(q);
4738 list_for_each_entry(q, &set->tag_list, tag_set_list)
4739 blk_mq_elv_switch_back(&head, q);
4741 list_for_each_entry(q, &set->tag_list, tag_set_list)
4742 blk_mq_unfreeze_queue(q);
4745 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4747 mutex_lock(&set->tag_list_lock);
4748 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4749 mutex_unlock(&set->tag_list_lock);
4751 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4753 /* Enable polling stats and return whether they were already enabled. */
4754 static bool blk_poll_stats_enable(struct request_queue *q)
4759 return blk_stats_alloc_enable(q);
4762 static void blk_mq_poll_stats_start(struct request_queue *q)
4765 * We don't arm the callback if polling stats are not enabled or the
4766 * callback is already active.
4768 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
4771 blk_stat_activate_msecs(q->poll_cb, 100);
4774 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4776 struct request_queue *q = cb->data;
4779 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4780 if (cb->stat[bucket].nr_samples)
4781 q->poll_stat[bucket] = cb->stat[bucket];
4785 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4788 unsigned long ret = 0;
4792 * If stats collection isn't on, don't sleep but turn it on for
4795 if (!blk_poll_stats_enable(q))
4799 * As an optimistic guess, use half of the mean service time
4800 * for this type of request. We can (and should) make this smarter.
4801 * For instance, if the completion latencies are tight, we can
4802 * get closer than just half the mean. This is especially
4803 * important on devices where the completion latencies are longer
4804 * than ~10 usec. We do use the stats for the relevant IO size
4805 * if available which does lead to better estimates.
4807 bucket = blk_mq_poll_stats_bkt(rq);
4811 if (q->poll_stat[bucket].nr_samples)
4812 ret = (q->poll_stat[bucket].mean + 1) / 2;
4817 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4819 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4820 struct request *rq = blk_qc_to_rq(hctx, qc);
4821 struct hrtimer_sleeper hs;
4822 enum hrtimer_mode mode;
4827 * If a request has completed on queue that uses an I/O scheduler, we
4828 * won't get back a request from blk_qc_to_rq.
4830 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4834 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4836 * 0: use half of prev avg
4837 * >0: use this specific value
4839 if (q->poll_nsec > 0)
4840 nsecs = q->poll_nsec;
4842 nsecs = blk_mq_poll_nsecs(q, rq);
4847 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4850 * This will be replaced with the stats tracking code, using
4851 * 'avg_completion_time / 2' as the pre-sleep target.
4855 mode = HRTIMER_MODE_REL;
4856 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4857 hrtimer_set_expires(&hs.timer, kt);
4860 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4862 set_current_state(TASK_UNINTERRUPTIBLE);
4863 hrtimer_sleeper_start_expires(&hs, mode);
4866 hrtimer_cancel(&hs.timer);
4867 mode = HRTIMER_MODE_ABS;
4868 } while (hs.task && !signal_pending(current));
4870 __set_current_state(TASK_RUNNING);
4871 destroy_hrtimer_on_stack(&hs.timer);
4874 * If we sleep, have the caller restart the poll loop to reset the
4875 * state. Like for the other success return cases, the caller is
4876 * responsible for checking if the IO completed. If the IO isn't
4877 * complete, we'll get called again and will go straight to the busy
4883 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4884 struct io_comp_batch *iob, unsigned int flags)
4886 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4887 long state = get_current_state();
4891 ret = q->mq_ops->poll(hctx, iob);
4893 __set_current_state(TASK_RUNNING);
4897 if (signal_pending_state(state, current))
4898 __set_current_state(TASK_RUNNING);
4899 if (task_is_running(current))
4902 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4905 } while (!need_resched());
4907 __set_current_state(TASK_RUNNING);
4911 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4914 if (!(flags & BLK_POLL_NOSLEEP) &&
4915 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4916 if (blk_mq_poll_hybrid(q, cookie))
4919 return blk_mq_poll_classic(q, cookie, iob, flags);
4922 unsigned int blk_mq_rq_cpu(struct request *rq)
4924 return rq->mq_ctx->cpu;
4926 EXPORT_SYMBOL(blk_mq_rq_cpu);
4928 void blk_mq_cancel_work_sync(struct request_queue *q)
4930 struct blk_mq_hw_ctx *hctx;
4933 cancel_delayed_work_sync(&q->requeue_work);
4935 queue_for_each_hw_ctx(q, hctx, i)
4936 cancel_delayed_work_sync(&hctx->run_work);
4939 static int __init blk_mq_init(void)
4943 for_each_possible_cpu(i)
4944 init_llist_head(&per_cpu(blk_cpu_done, i));
4945 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4947 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4948 "block/softirq:dead", NULL,
4949 blk_softirq_cpu_dead);
4950 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4951 blk_mq_hctx_notify_dead);
4952 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4953 blk_mq_hctx_notify_online,
4954 blk_mq_hctx_notify_offline);
4957 subsys_initcall(blk_mq_init);