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/t10-pi.h>
38 #include "blk-mq-debugfs.h"
41 #include "blk-mq-sched.h"
42 #include "blk-rq-qos.h"
44 static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
45 static DEFINE_PER_CPU(call_single_data_t, blk_cpu_csd);
47 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags);
48 static void blk_mq_request_bypass_insert(struct request *rq,
50 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
51 struct list_head *list);
52 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
53 struct io_comp_batch *iob, unsigned int flags);
56 * Check if any of the ctx, dispatch list or elevator
57 * have pending work in this hardware queue.
59 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
61 return !list_empty_careful(&hctx->dispatch) ||
62 sbitmap_any_bit_set(&hctx->ctx_map) ||
63 blk_mq_sched_has_work(hctx);
67 * Mark this ctx as having pending work in this hardware queue
69 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
70 struct blk_mq_ctx *ctx)
72 const int bit = ctx->index_hw[hctx->type];
74 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
75 sbitmap_set_bit(&hctx->ctx_map, bit);
78 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
79 struct blk_mq_ctx *ctx)
81 const int bit = ctx->index_hw[hctx->type];
83 sbitmap_clear_bit(&hctx->ctx_map, bit);
87 struct block_device *part;
88 unsigned int inflight[2];
91 static bool blk_mq_check_inflight(struct request *rq, void *priv)
93 struct mq_inflight *mi = priv;
95 if (rq->part && blk_do_io_stat(rq) &&
96 (!mi->part->bd_partno || rq->part == mi->part) &&
97 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
98 mi->inflight[rq_data_dir(rq)]++;
103 unsigned int blk_mq_in_flight(struct request_queue *q,
104 struct block_device *part)
106 struct mq_inflight mi = { .part = part };
108 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
110 return mi.inflight[0] + mi.inflight[1];
113 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
114 unsigned int inflight[2])
116 struct mq_inflight mi = { .part = part };
118 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
119 inflight[0] = mi.inflight[0];
120 inflight[1] = mi.inflight[1];
123 void blk_freeze_queue_start(struct request_queue *q)
125 mutex_lock(&q->mq_freeze_lock);
126 if (++q->mq_freeze_depth == 1) {
127 percpu_ref_kill(&q->q_usage_counter);
128 mutex_unlock(&q->mq_freeze_lock);
130 blk_mq_run_hw_queues(q, false);
132 mutex_unlock(&q->mq_freeze_lock);
135 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
137 void blk_mq_freeze_queue_wait(struct request_queue *q)
139 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
141 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
143 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
144 unsigned long timeout)
146 return wait_event_timeout(q->mq_freeze_wq,
147 percpu_ref_is_zero(&q->q_usage_counter),
150 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
153 * Guarantee no request is in use, so we can change any data structure of
154 * the queue afterward.
156 void blk_freeze_queue(struct request_queue *q)
159 * In the !blk_mq case we are only calling this to kill the
160 * q_usage_counter, otherwise this increases the freeze depth
161 * and waits for it to return to zero. For this reason there is
162 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
163 * exported to drivers as the only user for unfreeze is blk_mq.
165 blk_freeze_queue_start(q);
166 blk_mq_freeze_queue_wait(q);
169 void blk_mq_freeze_queue(struct request_queue *q)
172 * ...just an alias to keep freeze and unfreeze actions balanced
173 * in the blk_mq_* namespace
177 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
179 void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
181 mutex_lock(&q->mq_freeze_lock);
183 q->q_usage_counter.data->force_atomic = true;
184 q->mq_freeze_depth--;
185 WARN_ON_ONCE(q->mq_freeze_depth < 0);
186 if (!q->mq_freeze_depth) {
187 percpu_ref_resurrect(&q->q_usage_counter);
188 wake_up_all(&q->mq_freeze_wq);
190 mutex_unlock(&q->mq_freeze_lock);
193 void blk_mq_unfreeze_queue(struct request_queue *q)
195 __blk_mq_unfreeze_queue(q, false);
197 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
200 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
201 * mpt3sas driver such that this function can be removed.
203 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
207 spin_lock_irqsave(&q->queue_lock, flags);
208 if (!q->quiesce_depth++)
209 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
210 spin_unlock_irqrestore(&q->queue_lock, flags);
212 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
215 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
216 * @set: tag_set to wait on
218 * Note: it is driver's responsibility for making sure that quiesce has
219 * been started on or more of the request_queues of the tag_set. This
220 * function only waits for the quiesce on those request_queues that had
221 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
223 void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
225 if (set->flags & BLK_MQ_F_BLOCKING)
226 synchronize_srcu(set->srcu);
230 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
233 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
236 * Note: this function does not prevent that the struct request end_io()
237 * callback function is invoked. Once this function is returned, we make
238 * sure no dispatch can happen until the queue is unquiesced via
239 * blk_mq_unquiesce_queue().
241 void blk_mq_quiesce_queue(struct request_queue *q)
243 blk_mq_quiesce_queue_nowait(q);
244 /* nothing to wait for non-mq queues */
246 blk_mq_wait_quiesce_done(q->tag_set);
248 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
251 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
254 * This function recovers queue into the state before quiescing
255 * which is done by blk_mq_quiesce_queue.
257 void blk_mq_unquiesce_queue(struct request_queue *q)
260 bool run_queue = false;
262 spin_lock_irqsave(&q->queue_lock, flags);
263 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
265 } else if (!--q->quiesce_depth) {
266 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
269 spin_unlock_irqrestore(&q->queue_lock, flags);
271 /* dispatch requests which are inserted during quiescing */
273 blk_mq_run_hw_queues(q, true);
275 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
277 void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
279 struct request_queue *q;
281 mutex_lock(&set->tag_list_lock);
282 list_for_each_entry(q, &set->tag_list, tag_set_list) {
283 if (!blk_queue_skip_tagset_quiesce(q))
284 blk_mq_quiesce_queue_nowait(q);
286 blk_mq_wait_quiesce_done(set);
287 mutex_unlock(&set->tag_list_lock);
289 EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
291 void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
293 struct request_queue *q;
295 mutex_lock(&set->tag_list_lock);
296 list_for_each_entry(q, &set->tag_list, tag_set_list) {
297 if (!blk_queue_skip_tagset_quiesce(q))
298 blk_mq_unquiesce_queue(q);
300 mutex_unlock(&set->tag_list_lock);
302 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
304 void blk_mq_wake_waiters(struct request_queue *q)
306 struct blk_mq_hw_ctx *hctx;
309 queue_for_each_hw_ctx(q, hctx, i)
310 if (blk_mq_hw_queue_mapped(hctx))
311 blk_mq_tag_wakeup_all(hctx->tags, true);
314 void blk_rq_init(struct request_queue *q, struct request *rq)
316 memset(rq, 0, sizeof(*rq));
318 INIT_LIST_HEAD(&rq->queuelist);
320 rq->__sector = (sector_t) -1;
321 INIT_HLIST_NODE(&rq->hash);
322 RB_CLEAR_NODE(&rq->rb_node);
323 rq->tag = BLK_MQ_NO_TAG;
324 rq->internal_tag = BLK_MQ_NO_TAG;
325 rq->start_time_ns = ktime_get_ns();
327 blk_crypto_rq_set_defaults(rq);
329 EXPORT_SYMBOL(blk_rq_init);
331 /* Set start and alloc time when the allocated request is actually used */
332 static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns)
334 if (blk_mq_need_time_stamp(rq))
335 rq->start_time_ns = ktime_get_ns();
337 rq->start_time_ns = 0;
339 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
340 if (blk_queue_rq_alloc_time(rq->q))
341 rq->alloc_time_ns = alloc_time_ns ?: rq->start_time_ns;
343 rq->alloc_time_ns = 0;
347 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
348 struct blk_mq_tags *tags, unsigned int tag)
350 struct blk_mq_ctx *ctx = data->ctx;
351 struct blk_mq_hw_ctx *hctx = data->hctx;
352 struct request_queue *q = data->q;
353 struct request *rq = tags->static_rqs[tag];
358 rq->cmd_flags = data->cmd_flags;
360 if (data->flags & BLK_MQ_REQ_PM)
361 data->rq_flags |= RQF_PM;
362 if (blk_queue_io_stat(q))
363 data->rq_flags |= RQF_IO_STAT;
364 rq->rq_flags = data->rq_flags;
366 if (data->rq_flags & RQF_SCHED_TAGS) {
367 rq->tag = BLK_MQ_NO_TAG;
368 rq->internal_tag = tag;
371 rq->internal_tag = BLK_MQ_NO_TAG;
376 rq->io_start_time_ns = 0;
377 rq->stats_sectors = 0;
378 rq->nr_phys_segments = 0;
379 #if defined(CONFIG_BLK_DEV_INTEGRITY)
380 rq->nr_integrity_segments = 0;
383 rq->end_io_data = NULL;
385 blk_crypto_rq_set_defaults(rq);
386 INIT_LIST_HEAD(&rq->queuelist);
387 /* tag was already set */
388 WRITE_ONCE(rq->deadline, 0);
391 if (rq->rq_flags & RQF_USE_SCHED) {
392 struct elevator_queue *e = data->q->elevator;
394 INIT_HLIST_NODE(&rq->hash);
395 RB_CLEAR_NODE(&rq->rb_node);
397 if (e->type->ops.prepare_request)
398 e->type->ops.prepare_request(rq);
404 static inline struct request *
405 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data)
407 unsigned int tag, tag_offset;
408 struct blk_mq_tags *tags;
410 unsigned long tag_mask;
413 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
414 if (unlikely(!tag_mask))
417 tags = blk_mq_tags_from_data(data);
418 for (i = 0; tag_mask; i++) {
419 if (!(tag_mask & (1UL << i)))
421 tag = tag_offset + i;
422 prefetch(tags->static_rqs[tag]);
423 tag_mask &= ~(1UL << i);
424 rq = blk_mq_rq_ctx_init(data, tags, tag);
425 rq_list_add(data->cached_rq, rq);
428 if (!(data->rq_flags & RQF_SCHED_TAGS))
429 blk_mq_add_active_requests(data->hctx, nr);
430 /* caller already holds a reference, add for remainder */
431 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
434 return rq_list_pop(data->cached_rq);
437 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
439 struct request_queue *q = data->q;
440 u64 alloc_time_ns = 0;
444 /* alloc_time includes depth and tag waits */
445 if (blk_queue_rq_alloc_time(q))
446 alloc_time_ns = ktime_get_ns();
448 if (data->cmd_flags & REQ_NOWAIT)
449 data->flags |= BLK_MQ_REQ_NOWAIT;
453 * All requests use scheduler tags when an I/O scheduler is
454 * enabled for the queue.
456 data->rq_flags |= RQF_SCHED_TAGS;
459 * Flush/passthrough requests are special and go directly to the
462 if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&
463 !blk_op_is_passthrough(data->cmd_flags)) {
464 struct elevator_mq_ops *ops = &q->elevator->type->ops;
466 WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
468 data->rq_flags |= RQF_USE_SCHED;
469 if (ops->limit_depth)
470 ops->limit_depth(data->cmd_flags, data);
475 data->ctx = blk_mq_get_ctx(q);
476 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
477 if (!(data->rq_flags & RQF_SCHED_TAGS))
478 blk_mq_tag_busy(data->hctx);
480 if (data->flags & BLK_MQ_REQ_RESERVED)
481 data->rq_flags |= RQF_RESV;
484 * Try batched alloc if we want more than 1 tag.
486 if (data->nr_tags > 1) {
487 rq = __blk_mq_alloc_requests_batch(data);
489 blk_mq_rq_time_init(rq, alloc_time_ns);
496 * Waiting allocations only fail because of an inactive hctx. In that
497 * case just retry the hctx assignment and tag allocation as CPU hotplug
498 * should have migrated us to an online CPU by now.
500 tag = blk_mq_get_tag(data);
501 if (tag == BLK_MQ_NO_TAG) {
502 if (data->flags & BLK_MQ_REQ_NOWAIT)
505 * Give up the CPU and sleep for a random short time to
506 * ensure that thread using a realtime scheduling class
507 * are migrated off the CPU, and thus off the hctx that
514 if (!(data->rq_flags & RQF_SCHED_TAGS))
515 blk_mq_inc_active_requests(data->hctx);
516 rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag);
517 blk_mq_rq_time_init(rq, alloc_time_ns);
521 static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
522 struct blk_plug *plug,
524 blk_mq_req_flags_t flags)
526 struct blk_mq_alloc_data data = {
530 .nr_tags = plug->nr_ios,
531 .cached_rq = &plug->cached_rq,
535 if (blk_queue_enter(q, flags))
540 rq = __blk_mq_alloc_requests(&data);
546 static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
548 blk_mq_req_flags_t flags)
550 struct blk_plug *plug = current->plug;
556 if (rq_list_empty(plug->cached_rq)) {
557 if (plug->nr_ios == 1)
559 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
563 rq = rq_list_peek(&plug->cached_rq);
564 if (!rq || rq->q != q)
567 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
569 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
572 plug->cached_rq = rq_list_next(rq);
573 blk_mq_rq_time_init(rq, 0);
577 INIT_LIST_HEAD(&rq->queuelist);
581 struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
582 blk_mq_req_flags_t flags)
586 rq = blk_mq_alloc_cached_request(q, opf, flags);
588 struct blk_mq_alloc_data data = {
596 ret = blk_queue_enter(q, flags);
600 rq = __blk_mq_alloc_requests(&data);
605 rq->__sector = (sector_t) -1;
606 rq->bio = rq->biotail = NULL;
610 return ERR_PTR(-EWOULDBLOCK);
612 EXPORT_SYMBOL(blk_mq_alloc_request);
614 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
615 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
617 struct blk_mq_alloc_data data = {
623 u64 alloc_time_ns = 0;
629 /* alloc_time includes depth and tag waits */
630 if (blk_queue_rq_alloc_time(q))
631 alloc_time_ns = ktime_get_ns();
634 * If the tag allocator sleeps we could get an allocation for a
635 * different hardware context. No need to complicate the low level
636 * allocator for this for the rare use case of a command tied to
639 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
640 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
641 return ERR_PTR(-EINVAL);
643 if (hctx_idx >= q->nr_hw_queues)
644 return ERR_PTR(-EIO);
646 ret = blk_queue_enter(q, flags);
651 * Check if the hardware context is actually mapped to anything.
652 * If not tell the caller that it should skip this queue.
655 data.hctx = xa_load(&q->hctx_table, hctx_idx);
656 if (!blk_mq_hw_queue_mapped(data.hctx))
658 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
659 if (cpu >= nr_cpu_ids)
661 data.ctx = __blk_mq_get_ctx(q, cpu);
664 data.rq_flags |= RQF_SCHED_TAGS;
666 blk_mq_tag_busy(data.hctx);
668 if (flags & BLK_MQ_REQ_RESERVED)
669 data.rq_flags |= RQF_RESV;
672 tag = blk_mq_get_tag(&data);
673 if (tag == BLK_MQ_NO_TAG)
675 if (!(data.rq_flags & RQF_SCHED_TAGS))
676 blk_mq_inc_active_requests(data.hctx);
677 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag);
678 blk_mq_rq_time_init(rq, alloc_time_ns);
680 rq->__sector = (sector_t) -1;
681 rq->bio = rq->biotail = NULL;
688 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
690 static void blk_mq_finish_request(struct request *rq)
692 struct request_queue *q = rq->q;
694 if (rq->rq_flags & RQF_USE_SCHED) {
695 q->elevator->type->ops.finish_request(rq);
697 * For postflush request that may need to be
698 * completed twice, we should clear this flag
699 * to avoid double finish_request() on the rq.
701 rq->rq_flags &= ~RQF_USE_SCHED;
705 static void __blk_mq_free_request(struct request *rq)
707 struct request_queue *q = rq->q;
708 struct blk_mq_ctx *ctx = rq->mq_ctx;
709 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
710 const int sched_tag = rq->internal_tag;
712 blk_crypto_free_request(rq);
713 blk_pm_mark_last_busy(rq);
716 if (rq->tag != BLK_MQ_NO_TAG) {
717 blk_mq_dec_active_requests(hctx);
718 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
720 if (sched_tag != BLK_MQ_NO_TAG)
721 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
722 blk_mq_sched_restart(hctx);
726 void blk_mq_free_request(struct request *rq)
728 struct request_queue *q = rq->q;
730 blk_mq_finish_request(rq);
732 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
733 laptop_io_completion(q->disk->bdi);
737 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
738 if (req_ref_put_and_test(rq))
739 __blk_mq_free_request(rq);
741 EXPORT_SYMBOL_GPL(blk_mq_free_request);
743 void blk_mq_free_plug_rqs(struct blk_plug *plug)
747 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
748 blk_mq_free_request(rq);
751 void blk_dump_rq_flags(struct request *rq, char *msg)
753 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
754 rq->q->disk ? rq->q->disk->disk_name : "?",
755 (__force unsigned long long) rq->cmd_flags);
757 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
758 (unsigned long long)blk_rq_pos(rq),
759 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
760 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
761 rq->bio, rq->biotail, blk_rq_bytes(rq));
763 EXPORT_SYMBOL(blk_dump_rq_flags);
765 static void req_bio_endio(struct request *rq, struct bio *bio,
766 unsigned int nbytes, blk_status_t error)
768 if (unlikely(error)) {
769 bio->bi_status = error;
770 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
772 * Partial zone append completions cannot be supported as the
773 * BIO fragments may end up not being written sequentially.
775 if (bio->bi_iter.bi_size != nbytes)
776 bio->bi_status = BLK_STS_IOERR;
778 bio->bi_iter.bi_sector = rq->__sector;
781 bio_advance(bio, nbytes);
783 if (unlikely(rq->rq_flags & RQF_QUIET))
784 bio_set_flag(bio, BIO_QUIET);
785 /* don't actually finish bio if it's part of flush sequence */
786 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
790 static void blk_account_io_completion(struct request *req, unsigned int bytes)
792 if (req->part && blk_do_io_stat(req)) {
793 const int sgrp = op_stat_group(req_op(req));
796 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
801 static void blk_print_req_error(struct request *req, blk_status_t status)
803 printk_ratelimited(KERN_ERR
804 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
805 "phys_seg %u prio class %u\n",
806 blk_status_to_str(status),
807 req->q->disk ? req->q->disk->disk_name : "?",
808 blk_rq_pos(req), (__force u32)req_op(req),
809 blk_op_str(req_op(req)),
810 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
811 req->nr_phys_segments,
812 IOPRIO_PRIO_CLASS(req->ioprio));
816 * Fully end IO on a request. Does not support partial completions, or
819 static void blk_complete_request(struct request *req)
821 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
822 int total_bytes = blk_rq_bytes(req);
823 struct bio *bio = req->bio;
825 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
830 #ifdef CONFIG_BLK_DEV_INTEGRITY
831 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
832 req->q->integrity.profile->complete_fn(req, total_bytes);
836 * Upper layers may call blk_crypto_evict_key() anytime after the last
837 * bio_endio(). Therefore, the keyslot must be released before that.
839 blk_crypto_rq_put_keyslot(req);
841 blk_account_io_completion(req, total_bytes);
844 struct bio *next = bio->bi_next;
846 /* Completion has already been traced */
847 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
849 if (req_op(req) == REQ_OP_ZONE_APPEND)
850 bio->bi_iter.bi_sector = req->__sector;
858 * Reset counters so that the request stacking driver
859 * can find how many bytes remain in the request
869 * blk_update_request - Complete multiple bytes without completing the request
870 * @req: the request being processed
871 * @error: block status code
872 * @nr_bytes: number of bytes to complete for @req
875 * Ends I/O on a number of bytes attached to @req, but doesn't complete
876 * the request structure even if @req doesn't have leftover.
877 * If @req has leftover, sets it up for the next range of segments.
879 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
880 * %false return from this function.
883 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
884 * except in the consistency check at the end of this function.
887 * %false - this request doesn't have any more data
888 * %true - this request has more data
890 bool blk_update_request(struct request *req, blk_status_t error,
891 unsigned int nr_bytes)
895 trace_block_rq_complete(req, error, nr_bytes);
900 #ifdef CONFIG_BLK_DEV_INTEGRITY
901 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
903 req->q->integrity.profile->complete_fn(req, nr_bytes);
907 * Upper layers may call blk_crypto_evict_key() anytime after the last
908 * bio_endio(). Therefore, the keyslot must be released before that.
910 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
911 __blk_crypto_rq_put_keyslot(req);
913 if (unlikely(error && !blk_rq_is_passthrough(req) &&
914 !(req->rq_flags & RQF_QUIET)) &&
915 !test_bit(GD_DEAD, &req->q->disk->state)) {
916 blk_print_req_error(req, error);
917 trace_block_rq_error(req, error, nr_bytes);
920 blk_account_io_completion(req, nr_bytes);
924 struct bio *bio = req->bio;
925 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
927 if (bio_bytes == bio->bi_iter.bi_size)
928 req->bio = bio->bi_next;
930 /* Completion has already been traced */
931 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
932 req_bio_endio(req, bio, bio_bytes, error);
934 total_bytes += bio_bytes;
935 nr_bytes -= bio_bytes;
946 * Reset counters so that the request stacking driver
947 * can find how many bytes remain in the request
954 req->__data_len -= total_bytes;
956 /* update sector only for requests with clear definition of sector */
957 if (!blk_rq_is_passthrough(req))
958 req->__sector += total_bytes >> 9;
960 /* mixed attributes always follow the first bio */
961 if (req->rq_flags & RQF_MIXED_MERGE) {
962 req->cmd_flags &= ~REQ_FAILFAST_MASK;
963 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
966 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
968 * If total number of sectors is less than the first segment
969 * size, something has gone terribly wrong.
971 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
972 blk_dump_rq_flags(req, "request botched");
973 req->__data_len = blk_rq_cur_bytes(req);
976 /* recalculate the number of segments */
977 req->nr_phys_segments = blk_recalc_rq_segments(req);
982 EXPORT_SYMBOL_GPL(blk_update_request);
984 static inline void blk_account_io_done(struct request *req, u64 now)
986 trace_block_io_done(req);
989 * Account IO completion. flush_rq isn't accounted as a
990 * normal IO on queueing nor completion. Accounting the
991 * containing request is enough.
993 if (blk_do_io_stat(req) && req->part &&
994 !(req->rq_flags & RQF_FLUSH_SEQ)) {
995 const int sgrp = op_stat_group(req_op(req));
998 update_io_ticks(req->part, jiffies, true);
999 part_stat_inc(req->part, ios[sgrp]);
1000 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
1005 static inline void blk_account_io_start(struct request *req)
1007 trace_block_io_start(req);
1009 if (blk_do_io_stat(req)) {
1011 * All non-passthrough requests are created from a bio with one
1012 * exception: when a flush command that is part of a flush sequence
1013 * generated by the state machine in blk-flush.c is cloned onto the
1014 * lower device by dm-multipath we can get here without a bio.
1017 req->part = req->bio->bi_bdev;
1019 req->part = req->q->disk->part0;
1022 update_io_ticks(req->part, jiffies, false);
1027 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1029 if (rq->rq_flags & RQF_STATS)
1030 blk_stat_add(rq, now);
1032 blk_mq_sched_completed_request(rq, now);
1033 blk_account_io_done(rq, now);
1036 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1038 if (blk_mq_need_time_stamp(rq))
1039 __blk_mq_end_request_acct(rq, ktime_get_ns());
1041 blk_mq_finish_request(rq);
1044 rq_qos_done(rq->q, rq);
1045 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1046 blk_mq_free_request(rq);
1048 blk_mq_free_request(rq);
1051 EXPORT_SYMBOL(__blk_mq_end_request);
1053 void blk_mq_end_request(struct request *rq, blk_status_t error)
1055 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1057 __blk_mq_end_request(rq, error);
1059 EXPORT_SYMBOL(blk_mq_end_request);
1061 #define TAG_COMP_BATCH 32
1063 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1064 int *tag_array, int nr_tags)
1066 struct request_queue *q = hctx->queue;
1068 blk_mq_sub_active_requests(hctx, nr_tags);
1070 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1071 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1074 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1076 int tags[TAG_COMP_BATCH], nr_tags = 0;
1077 struct blk_mq_hw_ctx *cur_hctx = NULL;
1082 now = ktime_get_ns();
1084 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1086 prefetch(rq->rq_next);
1088 blk_complete_request(rq);
1090 __blk_mq_end_request_acct(rq, now);
1092 blk_mq_finish_request(rq);
1094 rq_qos_done(rq->q, rq);
1097 * If end_io handler returns NONE, then it still has
1098 * ownership of the request.
1100 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1103 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1104 if (!req_ref_put_and_test(rq))
1107 blk_crypto_free_request(rq);
1108 blk_pm_mark_last_busy(rq);
1110 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1112 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1114 cur_hctx = rq->mq_hctx;
1116 tags[nr_tags++] = rq->tag;
1120 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1122 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1124 static void blk_complete_reqs(struct llist_head *list)
1126 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1127 struct request *rq, *next;
1129 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1130 rq->q->mq_ops->complete(rq);
1133 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1135 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1138 static int blk_softirq_cpu_dead(unsigned int cpu)
1140 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1144 static void __blk_mq_complete_request_remote(void *data)
1146 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1149 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1151 int cpu = raw_smp_processor_id();
1153 if (!IS_ENABLED(CONFIG_SMP) ||
1154 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1157 * With force threaded interrupts enabled, raising softirq from an SMP
1158 * function call will always result in waking the ksoftirqd thread.
1159 * This is probably worse than completing the request on a different
1162 if (force_irqthreads())
1165 /* same CPU or cache domain? Complete locally */
1166 if (cpu == rq->mq_ctx->cpu ||
1167 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1168 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1171 /* don't try to IPI to an offline CPU */
1172 return cpu_online(rq->mq_ctx->cpu);
1175 static void blk_mq_complete_send_ipi(struct request *rq)
1179 cpu = rq->mq_ctx->cpu;
1180 if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu)))
1181 smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu));
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->mq_ctx->cpu == raw_smp_processor_id()) ||
1206 rq->cmd_flags & REQ_POLLED)
1209 if (blk_mq_complete_need_ipi(rq)) {
1210 blk_mq_complete_send_ipi(rq);
1214 if (rq->q->nr_hw_queues == 1) {
1215 blk_mq_raise_softirq(rq);
1220 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1223 * blk_mq_complete_request - end I/O on a request
1224 * @rq: the request being processed
1227 * Complete a request by scheduling the ->complete_rq operation.
1229 void blk_mq_complete_request(struct request *rq)
1231 if (!blk_mq_complete_request_remote(rq))
1232 rq->q->mq_ops->complete(rq);
1234 EXPORT_SYMBOL(blk_mq_complete_request);
1237 * blk_mq_start_request - Start processing a request
1238 * @rq: Pointer to request to be started
1240 * Function used by device drivers to notify the block layer that a request
1241 * is going to be processed now, so blk layer can do proper initializations
1242 * such as starting the timeout timer.
1244 void blk_mq_start_request(struct request *rq)
1246 struct request_queue *q = rq->q;
1248 trace_block_rq_issue(rq);
1250 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) &&
1251 !blk_rq_is_passthrough(rq)) {
1252 rq->io_start_time_ns = ktime_get_ns();
1253 rq->stats_sectors = blk_rq_sectors(rq);
1254 rq->rq_flags |= RQF_STATS;
1255 rq_qos_issue(q, rq);
1258 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1261 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1262 rq->mq_hctx->tags->rqs[rq->tag] = rq;
1264 #ifdef CONFIG_BLK_DEV_INTEGRITY
1265 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1266 q->integrity.profile->prepare_fn(rq);
1268 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1269 WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num);
1271 EXPORT_SYMBOL(blk_mq_start_request);
1274 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1275 * queues. This is important for md arrays to benefit from merging
1278 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1280 if (plug->multiple_queues)
1281 return BLK_MAX_REQUEST_COUNT * 2;
1282 return BLK_MAX_REQUEST_COUNT;
1285 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1287 struct request *last = rq_list_peek(&plug->mq_list);
1289 if (!plug->rq_count) {
1290 trace_block_plug(rq->q);
1291 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1292 (!blk_queue_nomerges(rq->q) &&
1293 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1294 blk_mq_flush_plug_list(plug, false);
1296 trace_block_plug(rq->q);
1299 if (!plug->multiple_queues && last && last->q != rq->q)
1300 plug->multiple_queues = true;
1302 * Any request allocated from sched tags can't be issued to
1303 * ->queue_rqs() directly
1305 if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))
1306 plug->has_elevator = true;
1308 rq_list_add(&plug->mq_list, rq);
1313 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1314 * @rq: request to insert
1315 * @at_head: insert request at head or tail of queue
1318 * Insert a fully prepared request at the back of the I/O scheduler queue
1319 * for execution. Don't wait for completion.
1322 * This function will invoke @done directly if the queue is dead.
1324 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1326 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1328 WARN_ON(irqs_disabled());
1329 WARN_ON(!blk_rq_is_passthrough(rq));
1331 blk_account_io_start(rq);
1334 * As plugging can be enabled for passthrough requests on a zoned
1335 * device, directly accessing the plug instead of using blk_mq_plug()
1336 * should not have any consequences.
1338 if (current->plug && !at_head) {
1339 blk_add_rq_to_plug(current->plug, rq);
1343 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1344 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
1346 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1348 struct blk_rq_wait {
1349 struct completion done;
1353 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1355 struct blk_rq_wait *wait = rq->end_io_data;
1358 complete(&wait->done);
1359 return RQ_END_IO_NONE;
1362 bool blk_rq_is_poll(struct request *rq)
1366 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1370 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1372 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1375 blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0);
1377 } while (!completion_done(wait));
1381 * blk_execute_rq - insert a request into queue for execution
1382 * @rq: request to insert
1383 * @at_head: insert request at head or tail of queue
1386 * Insert a fully prepared request at the back of the I/O scheduler queue
1387 * for execution and wait for completion.
1388 * Return: The blk_status_t result provided to blk_mq_end_request().
1390 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1392 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1393 struct blk_rq_wait wait = {
1394 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1397 WARN_ON(irqs_disabled());
1398 WARN_ON(!blk_rq_is_passthrough(rq));
1400 rq->end_io_data = &wait;
1401 rq->end_io = blk_end_sync_rq;
1403 blk_account_io_start(rq);
1404 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1405 blk_mq_run_hw_queue(hctx, false);
1407 if (blk_rq_is_poll(rq)) {
1408 blk_rq_poll_completion(rq, &wait.done);
1411 * Prevent hang_check timer from firing at us during very long
1414 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1417 while (!wait_for_completion_io_timeout(&wait.done,
1418 hang_check * (HZ/2)))
1421 wait_for_completion_io(&wait.done);
1426 EXPORT_SYMBOL(blk_execute_rq);
1428 static void __blk_mq_requeue_request(struct request *rq)
1430 struct request_queue *q = rq->q;
1432 blk_mq_put_driver_tag(rq);
1434 trace_block_rq_requeue(rq);
1435 rq_qos_requeue(q, rq);
1437 if (blk_mq_request_started(rq)) {
1438 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1439 rq->rq_flags &= ~RQF_TIMED_OUT;
1443 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1445 struct request_queue *q = rq->q;
1446 unsigned long flags;
1448 __blk_mq_requeue_request(rq);
1450 /* this request will be re-inserted to io scheduler queue */
1451 blk_mq_sched_requeue_request(rq);
1453 spin_lock_irqsave(&q->requeue_lock, flags);
1454 list_add_tail(&rq->queuelist, &q->requeue_list);
1455 spin_unlock_irqrestore(&q->requeue_lock, flags);
1457 if (kick_requeue_list)
1458 blk_mq_kick_requeue_list(q);
1460 EXPORT_SYMBOL(blk_mq_requeue_request);
1462 static void blk_mq_requeue_work(struct work_struct *work)
1464 struct request_queue *q =
1465 container_of(work, struct request_queue, requeue_work.work);
1467 LIST_HEAD(flush_list);
1470 spin_lock_irq(&q->requeue_lock);
1471 list_splice_init(&q->requeue_list, &rq_list);
1472 list_splice_init(&q->flush_list, &flush_list);
1473 spin_unlock_irq(&q->requeue_lock);
1475 while (!list_empty(&rq_list)) {
1476 rq = list_entry(rq_list.next, struct request, queuelist);
1478 * If RQF_DONTPREP ist set, the request has been started by the
1479 * driver already and might have driver-specific data allocated
1480 * already. Insert it into the hctx dispatch list to avoid
1481 * block layer merges for the request.
1483 if (rq->rq_flags & RQF_DONTPREP) {
1484 list_del_init(&rq->queuelist);
1485 blk_mq_request_bypass_insert(rq, 0);
1487 list_del_init(&rq->queuelist);
1488 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1492 while (!list_empty(&flush_list)) {
1493 rq = list_entry(flush_list.next, struct request, queuelist);
1494 list_del_init(&rq->queuelist);
1495 blk_mq_insert_request(rq, 0);
1498 blk_mq_run_hw_queues(q, false);
1501 void blk_mq_kick_requeue_list(struct request_queue *q)
1503 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1505 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1507 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1508 unsigned long msecs)
1510 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1511 msecs_to_jiffies(msecs));
1513 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1515 static bool blk_is_flush_data_rq(struct request *rq)
1517 return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq);
1520 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1523 * If we find a request that isn't idle we know the queue is busy
1524 * as it's checked in the iter.
1525 * Return false to stop the iteration.
1527 * In case of queue quiesce, if one flush data request is completed,
1528 * don't count it as inflight given the flush sequence is suspended,
1529 * and the original flush data request is invisible to driver, just
1530 * like other pending requests because of quiesce
1532 if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) &&
1533 blk_is_flush_data_rq(rq) &&
1534 blk_mq_request_completed(rq))) {
1544 bool blk_mq_queue_inflight(struct request_queue *q)
1548 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1551 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1553 static void blk_mq_rq_timed_out(struct request *req)
1555 req->rq_flags |= RQF_TIMED_OUT;
1556 if (req->q->mq_ops->timeout) {
1557 enum blk_eh_timer_return ret;
1559 ret = req->q->mq_ops->timeout(req);
1560 if (ret == BLK_EH_DONE)
1562 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1568 struct blk_expired_data {
1569 bool has_timedout_rq;
1571 unsigned long timeout_start;
1574 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1576 unsigned long deadline;
1578 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1580 if (rq->rq_flags & RQF_TIMED_OUT)
1583 deadline = READ_ONCE(rq->deadline);
1584 if (time_after_eq(expired->timeout_start, deadline))
1587 if (expired->next == 0)
1588 expired->next = deadline;
1589 else if (time_after(expired->next, deadline))
1590 expired->next = deadline;
1594 void blk_mq_put_rq_ref(struct request *rq)
1596 if (is_flush_rq(rq)) {
1597 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1598 blk_mq_free_request(rq);
1599 } else if (req_ref_put_and_test(rq)) {
1600 __blk_mq_free_request(rq);
1604 static bool blk_mq_check_expired(struct request *rq, void *priv)
1606 struct blk_expired_data *expired = priv;
1609 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1610 * be reallocated underneath the timeout handler's processing, then
1611 * the expire check is reliable. If the request is not expired, then
1612 * it was completed and reallocated as a new request after returning
1613 * from blk_mq_check_expired().
1615 if (blk_mq_req_expired(rq, expired)) {
1616 expired->has_timedout_rq = true;
1622 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1624 struct blk_expired_data *expired = priv;
1626 if (blk_mq_req_expired(rq, expired))
1627 blk_mq_rq_timed_out(rq);
1631 static void blk_mq_timeout_work(struct work_struct *work)
1633 struct request_queue *q =
1634 container_of(work, struct request_queue, timeout_work);
1635 struct blk_expired_data expired = {
1636 .timeout_start = jiffies,
1638 struct blk_mq_hw_ctx *hctx;
1641 /* A deadlock might occur if a request is stuck requiring a
1642 * timeout at the same time a queue freeze is waiting
1643 * completion, since the timeout code would not be able to
1644 * acquire the queue reference here.
1646 * That's why we don't use blk_queue_enter here; instead, we use
1647 * percpu_ref_tryget directly, because we need to be able to
1648 * obtain a reference even in the short window between the queue
1649 * starting to freeze, by dropping the first reference in
1650 * blk_freeze_queue_start, and the moment the last request is
1651 * consumed, marked by the instant q_usage_counter reaches
1654 if (!percpu_ref_tryget(&q->q_usage_counter))
1657 /* check if there is any timed-out request */
1658 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1659 if (expired.has_timedout_rq) {
1661 * Before walking tags, we must ensure any submit started
1662 * before the current time has finished. Since the submit
1663 * uses srcu or rcu, wait for a synchronization point to
1664 * ensure all running submits have finished
1666 blk_mq_wait_quiesce_done(q->tag_set);
1669 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1672 if (expired.next != 0) {
1673 mod_timer(&q->timeout, expired.next);
1676 * Request timeouts are handled as a forward rolling timer. If
1677 * we end up here it means that no requests are pending and
1678 * also that no request has been pending for a while. Mark
1679 * each hctx as idle.
1681 queue_for_each_hw_ctx(q, hctx, i) {
1682 /* the hctx may be unmapped, so check it here */
1683 if (blk_mq_hw_queue_mapped(hctx))
1684 blk_mq_tag_idle(hctx);
1690 struct flush_busy_ctx_data {
1691 struct blk_mq_hw_ctx *hctx;
1692 struct list_head *list;
1695 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1697 struct flush_busy_ctx_data *flush_data = data;
1698 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1699 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1700 enum hctx_type type = hctx->type;
1702 spin_lock(&ctx->lock);
1703 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1704 sbitmap_clear_bit(sb, bitnr);
1705 spin_unlock(&ctx->lock);
1710 * Process software queues that have been marked busy, splicing them
1711 * to the for-dispatch
1713 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1715 struct flush_busy_ctx_data data = {
1720 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1722 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1724 struct dispatch_rq_data {
1725 struct blk_mq_hw_ctx *hctx;
1729 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1732 struct dispatch_rq_data *dispatch_data = data;
1733 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1734 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1735 enum hctx_type type = hctx->type;
1737 spin_lock(&ctx->lock);
1738 if (!list_empty(&ctx->rq_lists[type])) {
1739 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1740 list_del_init(&dispatch_data->rq->queuelist);
1741 if (list_empty(&ctx->rq_lists[type]))
1742 sbitmap_clear_bit(sb, bitnr);
1744 spin_unlock(&ctx->lock);
1746 return !dispatch_data->rq;
1749 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1750 struct blk_mq_ctx *start)
1752 unsigned off = start ? start->index_hw[hctx->type] : 0;
1753 struct dispatch_rq_data data = {
1758 __sbitmap_for_each_set(&hctx->ctx_map, off,
1759 dispatch_rq_from_ctx, &data);
1764 bool __blk_mq_alloc_driver_tag(struct request *rq)
1766 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1767 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1770 blk_mq_tag_busy(rq->mq_hctx);
1772 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1773 bt = &rq->mq_hctx->tags->breserved_tags;
1776 if (!hctx_may_queue(rq->mq_hctx, bt))
1780 tag = __sbitmap_queue_get(bt);
1781 if (tag == BLK_MQ_NO_TAG)
1784 rq->tag = tag + tag_offset;
1785 blk_mq_inc_active_requests(rq->mq_hctx);
1789 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1790 int flags, void *key)
1792 struct blk_mq_hw_ctx *hctx;
1794 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1796 spin_lock(&hctx->dispatch_wait_lock);
1797 if (!list_empty(&wait->entry)) {
1798 struct sbitmap_queue *sbq;
1800 list_del_init(&wait->entry);
1801 sbq = &hctx->tags->bitmap_tags;
1802 atomic_dec(&sbq->ws_active);
1804 spin_unlock(&hctx->dispatch_wait_lock);
1806 blk_mq_run_hw_queue(hctx, true);
1811 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1812 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1813 * restart. For both cases, take care to check the condition again after
1814 * marking us as waiting.
1816 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1819 struct sbitmap_queue *sbq;
1820 struct wait_queue_head *wq;
1821 wait_queue_entry_t *wait;
1824 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1825 !(blk_mq_is_shared_tags(hctx->flags))) {
1826 blk_mq_sched_mark_restart_hctx(hctx);
1829 * It's possible that a tag was freed in the window between the
1830 * allocation failure and adding the hardware queue to the wait
1833 * Don't clear RESTART here, someone else could have set it.
1834 * At most this will cost an extra queue run.
1836 return blk_mq_get_driver_tag(rq);
1839 wait = &hctx->dispatch_wait;
1840 if (!list_empty_careful(&wait->entry))
1843 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1844 sbq = &hctx->tags->breserved_tags;
1846 sbq = &hctx->tags->bitmap_tags;
1847 wq = &bt_wait_ptr(sbq, hctx)->wait;
1849 spin_lock_irq(&wq->lock);
1850 spin_lock(&hctx->dispatch_wait_lock);
1851 if (!list_empty(&wait->entry)) {
1852 spin_unlock(&hctx->dispatch_wait_lock);
1853 spin_unlock_irq(&wq->lock);
1857 atomic_inc(&sbq->ws_active);
1858 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1859 __add_wait_queue(wq, wait);
1862 * Add one explicit barrier since blk_mq_get_driver_tag() may
1863 * not imply barrier in case of failure.
1865 * Order adding us to wait queue and allocating driver tag.
1867 * The pair is the one implied in sbitmap_queue_wake_up() which
1868 * orders clearing sbitmap tag bits and waitqueue_active() in
1869 * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless
1871 * Otherwise, re-order of adding wait queue and getting driver tag
1872 * may cause __sbitmap_queue_wake_up() to wake up nothing because
1873 * the waitqueue_active() may not observe us in wait queue.
1878 * It's possible that a tag was freed in the window between the
1879 * allocation failure and adding the hardware queue to the wait
1882 ret = blk_mq_get_driver_tag(rq);
1884 spin_unlock(&hctx->dispatch_wait_lock);
1885 spin_unlock_irq(&wq->lock);
1890 * We got a tag, remove ourselves from the wait queue to ensure
1891 * someone else gets the wakeup.
1893 list_del_init(&wait->entry);
1894 atomic_dec(&sbq->ws_active);
1895 spin_unlock(&hctx->dispatch_wait_lock);
1896 spin_unlock_irq(&wq->lock);
1901 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1902 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1904 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1905 * - EWMA is one simple way to compute running average value
1906 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1907 * - take 4 as factor for avoiding to get too small(0) result, and this
1908 * factor doesn't matter because EWMA decreases exponentially
1910 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1914 ewma = hctx->dispatch_busy;
1919 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1921 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1922 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1924 hctx->dispatch_busy = ewma;
1927 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1929 static void blk_mq_handle_dev_resource(struct request *rq,
1930 struct list_head *list)
1932 list_add(&rq->queuelist, list);
1933 __blk_mq_requeue_request(rq);
1936 static void blk_mq_handle_zone_resource(struct request *rq,
1937 struct list_head *zone_list)
1940 * If we end up here it is because we cannot dispatch a request to a
1941 * specific zone due to LLD level zone-write locking or other zone
1942 * related resource not being available. In this case, set the request
1943 * aside in zone_list for retrying it later.
1945 list_add(&rq->queuelist, zone_list);
1946 __blk_mq_requeue_request(rq);
1949 enum prep_dispatch {
1951 PREP_DISPATCH_NO_TAG,
1952 PREP_DISPATCH_NO_BUDGET,
1955 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1958 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1959 int budget_token = -1;
1962 budget_token = blk_mq_get_dispatch_budget(rq->q);
1963 if (budget_token < 0) {
1964 blk_mq_put_driver_tag(rq);
1965 return PREP_DISPATCH_NO_BUDGET;
1967 blk_mq_set_rq_budget_token(rq, budget_token);
1970 if (!blk_mq_get_driver_tag(rq)) {
1972 * The initial allocation attempt failed, so we need to
1973 * rerun the hardware queue when a tag is freed. The
1974 * waitqueue takes care of that. If the queue is run
1975 * before we add this entry back on the dispatch list,
1976 * we'll re-run it below.
1978 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1980 * All budgets not got from this function will be put
1981 * together during handling partial dispatch
1984 blk_mq_put_dispatch_budget(rq->q, budget_token);
1985 return PREP_DISPATCH_NO_TAG;
1989 return PREP_DISPATCH_OK;
1992 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1993 static void blk_mq_release_budgets(struct request_queue *q,
1994 struct list_head *list)
1998 list_for_each_entry(rq, list, queuelist) {
1999 int budget_token = blk_mq_get_rq_budget_token(rq);
2001 if (budget_token >= 0)
2002 blk_mq_put_dispatch_budget(q, budget_token);
2007 * blk_mq_commit_rqs will notify driver using bd->last that there is no
2008 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
2010 * Attention, we should explicitly call this in unusual cases:
2011 * 1) did not queue everything initially scheduled to queue
2012 * 2) the last attempt to queue a request failed
2014 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
2017 if (hctx->queue->mq_ops->commit_rqs && queued) {
2018 trace_block_unplug(hctx->queue, queued, !from_schedule);
2019 hctx->queue->mq_ops->commit_rqs(hctx);
2024 * Returns true if we did some work AND can potentially do more.
2026 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
2027 unsigned int nr_budgets)
2029 enum prep_dispatch prep;
2030 struct request_queue *q = hctx->queue;
2033 blk_status_t ret = BLK_STS_OK;
2034 LIST_HEAD(zone_list);
2035 bool needs_resource = false;
2037 if (list_empty(list))
2041 * Now process all the entries, sending them to the driver.
2045 struct blk_mq_queue_data bd;
2047 rq = list_first_entry(list, struct request, queuelist);
2049 WARN_ON_ONCE(hctx != rq->mq_hctx);
2050 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2051 if (prep != PREP_DISPATCH_OK)
2054 list_del_init(&rq->queuelist);
2057 bd.last = list_empty(list);
2060 * once the request is queued to lld, no need to cover the
2065 ret = q->mq_ops->queue_rq(hctx, &bd);
2070 case BLK_STS_RESOURCE:
2071 needs_resource = true;
2073 case BLK_STS_DEV_RESOURCE:
2074 blk_mq_handle_dev_resource(rq, list);
2076 case BLK_STS_ZONE_RESOURCE:
2078 * Move the request to zone_list and keep going through
2079 * the dispatch list to find more requests the drive can
2082 blk_mq_handle_zone_resource(rq, &zone_list);
2083 needs_resource = true;
2086 blk_mq_end_request(rq, ret);
2088 } while (!list_empty(list));
2090 if (!list_empty(&zone_list))
2091 list_splice_tail_init(&zone_list, list);
2093 /* If we didn't flush the entire list, we could have told the driver
2094 * there was more coming, but that turned out to be a lie.
2096 if (!list_empty(list) || ret != BLK_STS_OK)
2097 blk_mq_commit_rqs(hctx, queued, false);
2100 * Any items that need requeuing? Stuff them into hctx->dispatch,
2101 * that is where we will continue on next queue run.
2103 if (!list_empty(list)) {
2105 /* For non-shared tags, the RESTART check will suffice */
2106 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2107 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2108 blk_mq_is_shared_tags(hctx->flags));
2111 blk_mq_release_budgets(q, list);
2113 spin_lock(&hctx->lock);
2114 list_splice_tail_init(list, &hctx->dispatch);
2115 spin_unlock(&hctx->lock);
2118 * Order adding requests to hctx->dispatch and checking
2119 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2120 * in blk_mq_sched_restart(). Avoid restart code path to
2121 * miss the new added requests to hctx->dispatch, meantime
2122 * SCHED_RESTART is observed here.
2127 * If SCHED_RESTART was set by the caller of this function and
2128 * it is no longer set that means that it was cleared by another
2129 * thread and hence that a queue rerun is needed.
2131 * If 'no_tag' is set, that means that we failed getting
2132 * a driver tag with an I/O scheduler attached. If our dispatch
2133 * waitqueue is no longer active, ensure that we run the queue
2134 * AFTER adding our entries back to the list.
2136 * If no I/O scheduler has been configured it is possible that
2137 * the hardware queue got stopped and restarted before requests
2138 * were pushed back onto the dispatch list. Rerun the queue to
2139 * avoid starvation. Notes:
2140 * - blk_mq_run_hw_queue() checks whether or not a queue has
2141 * been stopped before rerunning a queue.
2142 * - Some but not all block drivers stop a queue before
2143 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2146 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2147 * bit is set, run queue after a delay to avoid IO stalls
2148 * that could otherwise occur if the queue is idle. We'll do
2149 * similar if we couldn't get budget or couldn't lock a zone
2150 * and SCHED_RESTART is set.
2152 needs_restart = blk_mq_sched_needs_restart(hctx);
2153 if (prep == PREP_DISPATCH_NO_BUDGET)
2154 needs_resource = true;
2155 if (!needs_restart ||
2156 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2157 blk_mq_run_hw_queue(hctx, true);
2158 else if (needs_resource)
2159 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2161 blk_mq_update_dispatch_busy(hctx, true);
2165 blk_mq_update_dispatch_busy(hctx, false);
2169 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2171 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2173 if (cpu >= nr_cpu_ids)
2174 cpu = cpumask_first(hctx->cpumask);
2179 * It'd be great if the workqueue API had a way to pass
2180 * in a mask and had some smarts for more clever placement.
2181 * For now we just round-robin here, switching for every
2182 * BLK_MQ_CPU_WORK_BATCH queued items.
2184 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2187 int next_cpu = hctx->next_cpu;
2189 if (hctx->queue->nr_hw_queues == 1)
2190 return WORK_CPU_UNBOUND;
2192 if (--hctx->next_cpu_batch <= 0) {
2194 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2196 if (next_cpu >= nr_cpu_ids)
2197 next_cpu = blk_mq_first_mapped_cpu(hctx);
2198 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2202 * Do unbound schedule if we can't find a online CPU for this hctx,
2203 * and it should only happen in the path of handling CPU DEAD.
2205 if (!cpu_online(next_cpu)) {
2212 * Make sure to re-select CPU next time once after CPUs
2213 * in hctx->cpumask become online again.
2215 hctx->next_cpu = next_cpu;
2216 hctx->next_cpu_batch = 1;
2217 return WORK_CPU_UNBOUND;
2220 hctx->next_cpu = next_cpu;
2225 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2226 * @hctx: Pointer to the hardware queue to run.
2227 * @msecs: Milliseconds of delay to wait before running the queue.
2229 * Run a hardware queue asynchronously with a delay of @msecs.
2231 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2233 if (unlikely(blk_mq_hctx_stopped(hctx)))
2235 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2236 msecs_to_jiffies(msecs));
2238 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2241 * blk_mq_run_hw_queue - Start to run a hardware queue.
2242 * @hctx: Pointer to the hardware queue to run.
2243 * @async: If we want to run the queue asynchronously.
2245 * Check if the request queue is not in a quiesced state and if there are
2246 * pending requests to be sent. If this is true, run the queue to send requests
2249 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2254 * We can't run the queue inline with interrupts disabled.
2256 WARN_ON_ONCE(!async && in_interrupt());
2258 might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING);
2261 * When queue is quiesced, we may be switching io scheduler, or
2262 * updating nr_hw_queues, or other things, and we can't run queue
2263 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2265 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2268 __blk_mq_run_dispatch_ops(hctx->queue, false,
2269 need_run = !blk_queue_quiesced(hctx->queue) &&
2270 blk_mq_hctx_has_pending(hctx));
2275 if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2276 blk_mq_delay_run_hw_queue(hctx, 0);
2280 blk_mq_run_dispatch_ops(hctx->queue,
2281 blk_mq_sched_dispatch_requests(hctx));
2283 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2286 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2289 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2291 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2293 * If the IO scheduler does not respect hardware queues when
2294 * dispatching, we just don't bother with multiple HW queues and
2295 * dispatch from hctx for the current CPU since running multiple queues
2296 * just causes lock contention inside the scheduler and pointless cache
2299 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2301 if (!blk_mq_hctx_stopped(hctx))
2307 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2308 * @q: Pointer to the request queue to run.
2309 * @async: If we want to run the queue asynchronously.
2311 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2313 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2317 if (blk_queue_sq_sched(q))
2318 sq_hctx = blk_mq_get_sq_hctx(q);
2319 queue_for_each_hw_ctx(q, hctx, i) {
2320 if (blk_mq_hctx_stopped(hctx))
2323 * Dispatch from this hctx either if there's no hctx preferred
2324 * by IO scheduler or if it has requests that bypass the
2327 if (!sq_hctx || sq_hctx == hctx ||
2328 !list_empty_careful(&hctx->dispatch))
2329 blk_mq_run_hw_queue(hctx, async);
2332 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2335 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2336 * @q: Pointer to the request queue to run.
2337 * @msecs: Milliseconds of delay to wait before running the queues.
2339 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2341 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2345 if (blk_queue_sq_sched(q))
2346 sq_hctx = blk_mq_get_sq_hctx(q);
2347 queue_for_each_hw_ctx(q, hctx, i) {
2348 if (blk_mq_hctx_stopped(hctx))
2351 * If there is already a run_work pending, leave the
2352 * pending delay untouched. Otherwise, a hctx can stall
2353 * if another hctx is re-delaying the other's work
2354 * before the work executes.
2356 if (delayed_work_pending(&hctx->run_work))
2359 * Dispatch from this hctx either if there's no hctx preferred
2360 * by IO scheduler or if it has requests that bypass the
2363 if (!sq_hctx || sq_hctx == hctx ||
2364 !list_empty_careful(&hctx->dispatch))
2365 blk_mq_delay_run_hw_queue(hctx, msecs);
2368 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2371 * This function is often used for pausing .queue_rq() by driver when
2372 * there isn't enough resource or some conditions aren't satisfied, and
2373 * BLK_STS_RESOURCE is usually returned.
2375 * We do not guarantee that dispatch can be drained or blocked
2376 * after blk_mq_stop_hw_queue() returns. Please use
2377 * blk_mq_quiesce_queue() for that requirement.
2379 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2381 cancel_delayed_work(&hctx->run_work);
2383 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2385 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2388 * This function is often used for pausing .queue_rq() by driver when
2389 * there isn't enough resource or some conditions aren't satisfied, and
2390 * BLK_STS_RESOURCE is usually returned.
2392 * We do not guarantee that dispatch can be drained or blocked
2393 * after blk_mq_stop_hw_queues() returns. Please use
2394 * blk_mq_quiesce_queue() for that requirement.
2396 void blk_mq_stop_hw_queues(struct request_queue *q)
2398 struct blk_mq_hw_ctx *hctx;
2401 queue_for_each_hw_ctx(q, hctx, i)
2402 blk_mq_stop_hw_queue(hctx);
2404 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2406 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2408 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2410 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
2412 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2414 void blk_mq_start_hw_queues(struct request_queue *q)
2416 struct blk_mq_hw_ctx *hctx;
2419 queue_for_each_hw_ctx(q, hctx, i)
2420 blk_mq_start_hw_queue(hctx);
2422 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2424 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2426 if (!blk_mq_hctx_stopped(hctx))
2429 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2430 blk_mq_run_hw_queue(hctx, async);
2432 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2434 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2436 struct blk_mq_hw_ctx *hctx;
2439 queue_for_each_hw_ctx(q, hctx, i)
2440 blk_mq_start_stopped_hw_queue(hctx, async ||
2441 (hctx->flags & BLK_MQ_F_BLOCKING));
2443 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2445 static void blk_mq_run_work_fn(struct work_struct *work)
2447 struct blk_mq_hw_ctx *hctx =
2448 container_of(work, struct blk_mq_hw_ctx, run_work.work);
2450 blk_mq_run_dispatch_ops(hctx->queue,
2451 blk_mq_sched_dispatch_requests(hctx));
2455 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2456 * @rq: Pointer to request to be inserted.
2457 * @flags: BLK_MQ_INSERT_*
2459 * Should only be used carefully, when the caller knows we want to
2460 * bypass a potential IO scheduler on the target device.
2462 static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2464 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2466 spin_lock(&hctx->lock);
2467 if (flags & BLK_MQ_INSERT_AT_HEAD)
2468 list_add(&rq->queuelist, &hctx->dispatch);
2470 list_add_tail(&rq->queuelist, &hctx->dispatch);
2471 spin_unlock(&hctx->lock);
2474 static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2475 struct blk_mq_ctx *ctx, struct list_head *list,
2476 bool run_queue_async)
2479 enum hctx_type type = hctx->type;
2482 * Try to issue requests directly if the hw queue isn't busy to save an
2483 * extra enqueue & dequeue to the sw queue.
2485 if (!hctx->dispatch_busy && !run_queue_async) {
2486 blk_mq_run_dispatch_ops(hctx->queue,
2487 blk_mq_try_issue_list_directly(hctx, list));
2488 if (list_empty(list))
2493 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2496 list_for_each_entry(rq, list, queuelist) {
2497 BUG_ON(rq->mq_ctx != ctx);
2498 trace_block_rq_insert(rq);
2499 if (rq->cmd_flags & REQ_NOWAIT)
2500 run_queue_async = true;
2503 spin_lock(&ctx->lock);
2504 list_splice_tail_init(list, &ctx->rq_lists[type]);
2505 blk_mq_hctx_mark_pending(hctx, ctx);
2506 spin_unlock(&ctx->lock);
2508 blk_mq_run_hw_queue(hctx, run_queue_async);
2511 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2513 struct request_queue *q = rq->q;
2514 struct blk_mq_ctx *ctx = rq->mq_ctx;
2515 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2517 if (blk_rq_is_passthrough(rq)) {
2519 * Passthrough request have to be added to hctx->dispatch
2520 * directly. The device may be in a situation where it can't
2521 * handle FS request, and always returns BLK_STS_RESOURCE for
2522 * them, which gets them added to hctx->dispatch.
2524 * If a passthrough request is required to unblock the queues,
2525 * and it is added to the scheduler queue, there is no chance to
2526 * dispatch it given we prioritize requests in hctx->dispatch.
2528 blk_mq_request_bypass_insert(rq, flags);
2529 } else if (req_op(rq) == REQ_OP_FLUSH) {
2531 * Firstly normal IO request is inserted to scheduler queue or
2532 * sw queue, meantime we add flush request to dispatch queue(
2533 * hctx->dispatch) directly and there is at most one in-flight
2534 * flush request for each hw queue, so it doesn't matter to add
2535 * flush request to tail or front of the dispatch queue.
2537 * Secondly in case of NCQ, flush request belongs to non-NCQ
2538 * command, and queueing it will fail when there is any
2539 * in-flight normal IO request(NCQ command). When adding flush
2540 * rq to the front of hctx->dispatch, it is easier to introduce
2541 * extra time to flush rq's latency because of S_SCHED_RESTART
2542 * compared with adding to the tail of dispatch queue, then
2543 * chance of flush merge is increased, and less flush requests
2544 * will be issued to controller. It is observed that ~10% time
2545 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2546 * drive when adding flush rq to the front of hctx->dispatch.
2548 * Simply queue flush rq to the front of hctx->dispatch so that
2549 * intensive flush workloads can benefit in case of NCQ HW.
2551 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2552 } else if (q->elevator) {
2555 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2557 list_add(&rq->queuelist, &list);
2558 q->elevator->type->ops.insert_requests(hctx, &list, flags);
2560 trace_block_rq_insert(rq);
2562 spin_lock(&ctx->lock);
2563 if (flags & BLK_MQ_INSERT_AT_HEAD)
2564 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2566 list_add_tail(&rq->queuelist,
2567 &ctx->rq_lists[hctx->type]);
2568 blk_mq_hctx_mark_pending(hctx, ctx);
2569 spin_unlock(&ctx->lock);
2573 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2574 unsigned int nr_segs)
2578 if (bio->bi_opf & REQ_RAHEAD)
2579 rq->cmd_flags |= REQ_FAILFAST_MASK;
2581 rq->__sector = bio->bi_iter.bi_sector;
2582 blk_rq_bio_prep(rq, bio, nr_segs);
2584 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2585 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2588 blk_account_io_start(rq);
2591 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2592 struct request *rq, bool last)
2594 struct request_queue *q = rq->q;
2595 struct blk_mq_queue_data bd = {
2602 * For OK queue, we are done. For error, caller may kill it.
2603 * Any other error (busy), just add it to our list as we
2604 * previously would have done.
2606 ret = q->mq_ops->queue_rq(hctx, &bd);
2609 blk_mq_update_dispatch_busy(hctx, false);
2611 case BLK_STS_RESOURCE:
2612 case BLK_STS_DEV_RESOURCE:
2613 blk_mq_update_dispatch_busy(hctx, true);
2614 __blk_mq_requeue_request(rq);
2617 blk_mq_update_dispatch_busy(hctx, false);
2624 static bool blk_mq_get_budget_and_tag(struct request *rq)
2628 budget_token = blk_mq_get_dispatch_budget(rq->q);
2629 if (budget_token < 0)
2631 blk_mq_set_rq_budget_token(rq, budget_token);
2632 if (!blk_mq_get_driver_tag(rq)) {
2633 blk_mq_put_dispatch_budget(rq->q, budget_token);
2640 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2641 * @hctx: Pointer of the associated hardware queue.
2642 * @rq: Pointer to request to be sent.
2644 * If the device has enough resources to accept a new request now, send the
2645 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2646 * we can try send it another time in the future. Requests inserted at this
2647 * queue have higher priority.
2649 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2654 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2655 blk_mq_insert_request(rq, 0);
2659 if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
2660 blk_mq_insert_request(rq, 0);
2661 blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT);
2665 ret = __blk_mq_issue_directly(hctx, rq, true);
2669 case BLK_STS_RESOURCE:
2670 case BLK_STS_DEV_RESOURCE:
2671 blk_mq_request_bypass_insert(rq, 0);
2672 blk_mq_run_hw_queue(hctx, false);
2675 blk_mq_end_request(rq, ret);
2680 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2682 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2684 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2685 blk_mq_insert_request(rq, 0);
2689 if (!blk_mq_get_budget_and_tag(rq))
2690 return BLK_STS_RESOURCE;
2691 return __blk_mq_issue_directly(hctx, rq, last);
2694 static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2696 struct blk_mq_hw_ctx *hctx = NULL;
2699 blk_status_t ret = BLK_STS_OK;
2701 while ((rq = rq_list_pop(&plug->mq_list))) {
2702 bool last = rq_list_empty(plug->mq_list);
2704 if (hctx != rq->mq_hctx) {
2706 blk_mq_commit_rqs(hctx, queued, false);
2712 ret = blk_mq_request_issue_directly(rq, last);
2717 case BLK_STS_RESOURCE:
2718 case BLK_STS_DEV_RESOURCE:
2719 blk_mq_request_bypass_insert(rq, 0);
2720 blk_mq_run_hw_queue(hctx, false);
2723 blk_mq_end_request(rq, ret);
2729 if (ret != BLK_STS_OK)
2730 blk_mq_commit_rqs(hctx, queued, false);
2733 static void __blk_mq_flush_plug_list(struct request_queue *q,
2734 struct blk_plug *plug)
2736 if (blk_queue_quiesced(q))
2738 q->mq_ops->queue_rqs(&plug->mq_list);
2741 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2743 struct blk_mq_hw_ctx *this_hctx = NULL;
2744 struct blk_mq_ctx *this_ctx = NULL;
2745 struct request *requeue_list = NULL;
2746 struct request **requeue_lastp = &requeue_list;
2747 unsigned int depth = 0;
2748 bool is_passthrough = false;
2752 struct request *rq = rq_list_pop(&plug->mq_list);
2755 this_hctx = rq->mq_hctx;
2756 this_ctx = rq->mq_ctx;
2757 is_passthrough = blk_rq_is_passthrough(rq);
2758 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2759 is_passthrough != blk_rq_is_passthrough(rq)) {
2760 rq_list_add_tail(&requeue_lastp, rq);
2763 list_add(&rq->queuelist, &list);
2765 } while (!rq_list_empty(plug->mq_list));
2767 plug->mq_list = requeue_list;
2768 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2770 percpu_ref_get(&this_hctx->queue->q_usage_counter);
2771 /* passthrough requests should never be issued to the I/O scheduler */
2772 if (is_passthrough) {
2773 spin_lock(&this_hctx->lock);
2774 list_splice_tail_init(&list, &this_hctx->dispatch);
2775 spin_unlock(&this_hctx->lock);
2776 blk_mq_run_hw_queue(this_hctx, from_sched);
2777 } else if (this_hctx->queue->elevator) {
2778 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2780 blk_mq_run_hw_queue(this_hctx, from_sched);
2782 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2784 percpu_ref_put(&this_hctx->queue->q_usage_counter);
2787 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2792 * We may have been called recursively midway through handling
2793 * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2794 * To avoid mq_list changing under our feet, clear rq_count early and
2795 * bail out specifically if rq_count is 0 rather than checking
2796 * whether the mq_list is empty.
2798 if (plug->rq_count == 0)
2802 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2803 struct request_queue *q;
2805 rq = rq_list_peek(&plug->mq_list);
2809 * Peek first request and see if we have a ->queue_rqs() hook.
2810 * If we do, we can dispatch the whole plug list in one go. We
2811 * already know at this point that all requests belong to the
2812 * same queue, caller must ensure that's the case.
2814 if (q->mq_ops->queue_rqs) {
2815 blk_mq_run_dispatch_ops(q,
2816 __blk_mq_flush_plug_list(q, plug));
2817 if (rq_list_empty(plug->mq_list))
2821 blk_mq_run_dispatch_ops(q,
2822 blk_mq_plug_issue_direct(plug));
2823 if (rq_list_empty(plug->mq_list))
2828 blk_mq_dispatch_plug_list(plug, from_schedule);
2829 } while (!rq_list_empty(plug->mq_list));
2832 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2833 struct list_head *list)
2836 blk_status_t ret = BLK_STS_OK;
2838 while (!list_empty(list)) {
2839 struct request *rq = list_first_entry(list, struct request,
2842 list_del_init(&rq->queuelist);
2843 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2848 case BLK_STS_RESOURCE:
2849 case BLK_STS_DEV_RESOURCE:
2850 blk_mq_request_bypass_insert(rq, 0);
2851 if (list_empty(list))
2852 blk_mq_run_hw_queue(hctx, false);
2855 blk_mq_end_request(rq, ret);
2861 if (ret != BLK_STS_OK)
2862 blk_mq_commit_rqs(hctx, queued, false);
2865 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2866 struct bio *bio, unsigned int nr_segs)
2868 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2869 if (blk_attempt_plug_merge(q, bio, nr_segs))
2871 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2877 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2878 struct blk_plug *plug,
2882 struct blk_mq_alloc_data data = {
2885 .cmd_flags = bio->bi_opf,
2889 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2892 rq_qos_throttle(q, bio);
2895 data.nr_tags = plug->nr_ios;
2897 data.cached_rq = &plug->cached_rq;
2900 rq = __blk_mq_alloc_requests(&data);
2903 rq_qos_cleanup(q, bio);
2904 if (bio->bi_opf & REQ_NOWAIT)
2905 bio_wouldblock_error(bio);
2910 * Check if we can use the passed on request for submitting the passed in bio,
2911 * and remove it from the request list if it can be used.
2913 static bool blk_mq_use_cached_rq(struct request *rq, struct blk_plug *plug,
2916 enum hctx_type type = blk_mq_get_hctx_type(bio->bi_opf);
2917 enum hctx_type hctx_type = rq->mq_hctx->type;
2919 WARN_ON_ONCE(rq_list_peek(&plug->cached_rq) != rq);
2921 if (type != hctx_type &&
2922 !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2924 if (op_is_flush(rq->cmd_flags) != op_is_flush(bio->bi_opf))
2928 * If any qos ->throttle() end up blocking, we will have flushed the
2929 * plug and hence killed the cached_rq list as well. Pop this entry
2930 * before we throttle.
2932 plug->cached_rq = rq_list_next(rq);
2933 rq_qos_throttle(rq->q, bio);
2935 blk_mq_rq_time_init(rq, 0);
2936 rq->cmd_flags = bio->bi_opf;
2937 INIT_LIST_HEAD(&rq->queuelist);
2942 * blk_mq_submit_bio - Create and send a request to block device.
2943 * @bio: Bio pointer.
2945 * Builds up a request structure from @q and @bio and send to the device. The
2946 * request may not be queued directly to hardware if:
2947 * * This request can be merged with another one
2948 * * We want to place request at plug queue for possible future merging
2949 * * There is an IO scheduler active at this queue
2951 * It will not queue the request if there is an error with the bio, or at the
2954 void blk_mq_submit_bio(struct bio *bio)
2956 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2957 struct blk_plug *plug = blk_mq_plug(bio);
2958 const int is_sync = op_is_sync(bio->bi_opf);
2959 struct blk_mq_hw_ctx *hctx;
2960 struct request *rq = NULL;
2961 unsigned int nr_segs = 1;
2964 bio = blk_queue_bounce(bio, q);
2967 rq = rq_list_peek(&plug->cached_rq);
2968 if (rq && rq->q != q)
2972 if (unlikely(bio_may_exceed_limits(bio, &q->limits))) {
2973 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2977 if (!bio_integrity_prep(bio))
2979 if (blk_mq_attempt_bio_merge(q, bio, nr_segs))
2981 if (blk_mq_use_cached_rq(rq, plug, bio))
2983 percpu_ref_get(&q->q_usage_counter);
2985 if (unlikely(bio_queue_enter(bio)))
2987 if (unlikely(bio_may_exceed_limits(bio, &q->limits))) {
2988 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2992 if (!bio_integrity_prep(bio))
2996 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2997 if (unlikely(!rq)) {
3004 trace_block_getrq(bio);
3006 rq_qos_track(q, rq, bio);
3008 blk_mq_bio_to_request(rq, bio, nr_segs);
3010 ret = blk_crypto_rq_get_keyslot(rq);
3011 if (ret != BLK_STS_OK) {
3012 bio->bi_status = ret;
3014 blk_mq_free_request(rq);
3018 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
3022 blk_add_rq_to_plug(plug, rq);
3027 if ((rq->rq_flags & RQF_USE_SCHED) ||
3028 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
3029 blk_mq_insert_request(rq, 0);
3030 blk_mq_run_hw_queue(hctx, true);
3032 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
3036 #ifdef CONFIG_BLK_MQ_STACKING
3038 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3039 * @rq: the request being queued
3041 blk_status_t blk_insert_cloned_request(struct request *rq)
3043 struct request_queue *q = rq->q;
3044 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3045 unsigned int max_segments = blk_rq_get_max_segments(rq);
3048 if (blk_rq_sectors(rq) > max_sectors) {
3050 * SCSI device does not have a good way to return if
3051 * Write Same/Zero is actually supported. If a device rejects
3052 * a non-read/write command (discard, write same,etc.) the
3053 * low-level device driver will set the relevant queue limit to
3054 * 0 to prevent blk-lib from issuing more of the offending
3055 * operations. Commands queued prior to the queue limit being
3056 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3057 * errors being propagated to upper layers.
3059 if (max_sectors == 0)
3060 return BLK_STS_NOTSUPP;
3062 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3063 __func__, blk_rq_sectors(rq), max_sectors);
3064 return BLK_STS_IOERR;
3068 * The queue settings related to segment counting may differ from the
3071 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3072 if (rq->nr_phys_segments > max_segments) {
3073 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3074 __func__, rq->nr_phys_segments, max_segments);
3075 return BLK_STS_IOERR;
3078 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3079 return BLK_STS_IOERR;
3081 ret = blk_crypto_rq_get_keyslot(rq);
3082 if (ret != BLK_STS_OK)
3085 blk_account_io_start(rq);
3088 * Since we have a scheduler attached on the top device,
3089 * bypass a potential scheduler on the bottom device for
3092 blk_mq_run_dispatch_ops(q,
3093 ret = blk_mq_request_issue_directly(rq, true));
3095 blk_account_io_done(rq, ktime_get_ns());
3098 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3101 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3102 * @rq: the clone request to be cleaned up
3105 * Free all bios in @rq for a cloned request.
3107 void blk_rq_unprep_clone(struct request *rq)
3111 while ((bio = rq->bio) != NULL) {
3112 rq->bio = bio->bi_next;
3117 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3120 * blk_rq_prep_clone - Helper function to setup clone request
3121 * @rq: the request to be setup
3122 * @rq_src: original request to be cloned
3123 * @bs: bio_set that bios for clone are allocated from
3124 * @gfp_mask: memory allocation mask for bio
3125 * @bio_ctr: setup function to be called for each clone bio.
3126 * Returns %0 for success, non %0 for failure.
3127 * @data: private data to be passed to @bio_ctr
3130 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3131 * Also, pages which the original bios are pointing to are not copied
3132 * and the cloned bios just point same pages.
3133 * So cloned bios must be completed before original bios, which means
3134 * the caller must complete @rq before @rq_src.
3136 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3137 struct bio_set *bs, gfp_t gfp_mask,
3138 int (*bio_ctr)(struct bio *, struct bio *, void *),
3141 struct bio *bio, *bio_src;
3146 __rq_for_each_bio(bio_src, rq_src) {
3147 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3152 if (bio_ctr && bio_ctr(bio, bio_src, data))
3156 rq->biotail->bi_next = bio;
3159 rq->bio = rq->biotail = bio;
3164 /* Copy attributes of the original request to the clone request. */
3165 rq->__sector = blk_rq_pos(rq_src);
3166 rq->__data_len = blk_rq_bytes(rq_src);
3167 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3168 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3169 rq->special_vec = rq_src->special_vec;
3171 rq->nr_phys_segments = rq_src->nr_phys_segments;
3172 rq->ioprio = rq_src->ioprio;
3174 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3182 blk_rq_unprep_clone(rq);
3186 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3187 #endif /* CONFIG_BLK_MQ_STACKING */
3190 * Steal bios from a request and add them to a bio list.
3191 * The request must not have been partially completed before.
3193 void blk_steal_bios(struct bio_list *list, struct request *rq)
3197 list->tail->bi_next = rq->bio;
3199 list->head = rq->bio;
3200 list->tail = rq->biotail;
3208 EXPORT_SYMBOL_GPL(blk_steal_bios);
3210 static size_t order_to_size(unsigned int order)
3212 return (size_t)PAGE_SIZE << order;
3215 /* called before freeing request pool in @tags */
3216 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3217 struct blk_mq_tags *tags)
3220 unsigned long flags;
3223 * There is no need to clear mapping if driver tags is not initialized
3224 * or the mapping belongs to the driver tags.
3226 if (!drv_tags || drv_tags == tags)
3229 list_for_each_entry(page, &tags->page_list, lru) {
3230 unsigned long start = (unsigned long)page_address(page);
3231 unsigned long end = start + order_to_size(page->private);
3234 for (i = 0; i < drv_tags->nr_tags; i++) {
3235 struct request *rq = drv_tags->rqs[i];
3236 unsigned long rq_addr = (unsigned long)rq;
3238 if (rq_addr >= start && rq_addr < end) {
3239 WARN_ON_ONCE(req_ref_read(rq) != 0);
3240 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3246 * Wait until all pending iteration is done.
3248 * Request reference is cleared and it is guaranteed to be observed
3249 * after the ->lock is released.
3251 spin_lock_irqsave(&drv_tags->lock, flags);
3252 spin_unlock_irqrestore(&drv_tags->lock, flags);
3255 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3256 unsigned int hctx_idx)
3258 struct blk_mq_tags *drv_tags;
3261 if (list_empty(&tags->page_list))
3264 if (blk_mq_is_shared_tags(set->flags))
3265 drv_tags = set->shared_tags;
3267 drv_tags = set->tags[hctx_idx];
3269 if (tags->static_rqs && set->ops->exit_request) {
3272 for (i = 0; i < tags->nr_tags; i++) {
3273 struct request *rq = tags->static_rqs[i];
3277 set->ops->exit_request(set, rq, hctx_idx);
3278 tags->static_rqs[i] = NULL;
3282 blk_mq_clear_rq_mapping(drv_tags, tags);
3284 while (!list_empty(&tags->page_list)) {
3285 page = list_first_entry(&tags->page_list, struct page, lru);
3286 list_del_init(&page->lru);
3288 * Remove kmemleak object previously allocated in
3289 * blk_mq_alloc_rqs().
3291 kmemleak_free(page_address(page));
3292 __free_pages(page, page->private);
3296 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3300 kfree(tags->static_rqs);
3301 tags->static_rqs = NULL;
3303 blk_mq_free_tags(tags);
3306 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3307 unsigned int hctx_idx)
3311 for (i = 0; i < set->nr_maps; i++) {
3312 unsigned int start = set->map[i].queue_offset;
3313 unsigned int end = start + set->map[i].nr_queues;
3315 if (hctx_idx >= start && hctx_idx < end)
3319 if (i >= set->nr_maps)
3320 i = HCTX_TYPE_DEFAULT;
3325 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3326 unsigned int hctx_idx)
3328 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3330 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3333 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3334 unsigned int hctx_idx,
3335 unsigned int nr_tags,
3336 unsigned int reserved_tags)
3338 int node = blk_mq_get_hctx_node(set, hctx_idx);
3339 struct blk_mq_tags *tags;
3341 if (node == NUMA_NO_NODE)
3342 node = set->numa_node;
3344 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3345 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3349 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3350 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3355 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3356 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3358 if (!tags->static_rqs)
3366 blk_mq_free_tags(tags);
3370 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3371 unsigned int hctx_idx, int node)
3375 if (set->ops->init_request) {
3376 ret = set->ops->init_request(set, rq, hctx_idx, node);
3381 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3385 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3386 struct blk_mq_tags *tags,
3387 unsigned int hctx_idx, unsigned int depth)
3389 unsigned int i, j, entries_per_page, max_order = 4;
3390 int node = blk_mq_get_hctx_node(set, hctx_idx);
3391 size_t rq_size, left;
3393 if (node == NUMA_NO_NODE)
3394 node = set->numa_node;
3396 INIT_LIST_HEAD(&tags->page_list);
3399 * rq_size is the size of the request plus driver payload, rounded
3400 * to the cacheline size
3402 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3404 left = rq_size * depth;
3406 for (i = 0; i < depth; ) {
3407 int this_order = max_order;
3412 while (this_order && left < order_to_size(this_order - 1))
3416 page = alloc_pages_node(node,
3417 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3423 if (order_to_size(this_order) < rq_size)
3430 page->private = this_order;
3431 list_add_tail(&page->lru, &tags->page_list);
3433 p = page_address(page);
3435 * Allow kmemleak to scan these pages as they contain pointers
3436 * to additional allocations like via ops->init_request().
3438 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3439 entries_per_page = order_to_size(this_order) / rq_size;
3440 to_do = min(entries_per_page, depth - i);
3441 left -= to_do * rq_size;
3442 for (j = 0; j < to_do; j++) {
3443 struct request *rq = p;
3445 tags->static_rqs[i] = rq;
3446 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3447 tags->static_rqs[i] = NULL;
3458 blk_mq_free_rqs(set, tags, hctx_idx);
3462 struct rq_iter_data {
3463 struct blk_mq_hw_ctx *hctx;
3467 static bool blk_mq_has_request(struct request *rq, void *data)
3469 struct rq_iter_data *iter_data = data;
3471 if (rq->mq_hctx != iter_data->hctx)
3473 iter_data->has_rq = true;
3477 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3479 struct blk_mq_tags *tags = hctx->sched_tags ?
3480 hctx->sched_tags : hctx->tags;
3481 struct rq_iter_data data = {
3485 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3489 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3490 struct blk_mq_hw_ctx *hctx)
3492 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3494 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3499 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3501 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3502 struct blk_mq_hw_ctx, cpuhp_online);
3504 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3505 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3509 * Prevent new request from being allocated on the current hctx.
3511 * The smp_mb__after_atomic() Pairs with the implied barrier in
3512 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3513 * seen once we return from the tag allocator.
3515 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3516 smp_mb__after_atomic();
3519 * Try to grab a reference to the queue and wait for any outstanding
3520 * requests. If we could not grab a reference the queue has been
3521 * frozen and there are no requests.
3523 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3524 while (blk_mq_hctx_has_requests(hctx))
3526 percpu_ref_put(&hctx->queue->q_usage_counter);
3532 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3534 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3535 struct blk_mq_hw_ctx, cpuhp_online);
3537 if (cpumask_test_cpu(cpu, hctx->cpumask))
3538 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3543 * 'cpu' is going away. splice any existing rq_list entries from this
3544 * software queue to the hw queue dispatch list, and ensure that it
3547 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3549 struct blk_mq_hw_ctx *hctx;
3550 struct blk_mq_ctx *ctx;
3552 enum hctx_type type;
3554 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3555 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3558 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3561 spin_lock(&ctx->lock);
3562 if (!list_empty(&ctx->rq_lists[type])) {
3563 list_splice_init(&ctx->rq_lists[type], &tmp);
3564 blk_mq_hctx_clear_pending(hctx, ctx);
3566 spin_unlock(&ctx->lock);
3568 if (list_empty(&tmp))
3571 spin_lock(&hctx->lock);
3572 list_splice_tail_init(&tmp, &hctx->dispatch);
3573 spin_unlock(&hctx->lock);
3575 blk_mq_run_hw_queue(hctx, true);
3579 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3581 if (!(hctx->flags & BLK_MQ_F_STACKING))
3582 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3583 &hctx->cpuhp_online);
3584 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3589 * Before freeing hw queue, clearing the flush request reference in
3590 * tags->rqs[] for avoiding potential UAF.
3592 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3593 unsigned int queue_depth, struct request *flush_rq)
3596 unsigned long flags;
3598 /* The hw queue may not be mapped yet */
3602 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3604 for (i = 0; i < queue_depth; i++)
3605 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3608 * Wait until all pending iteration is done.
3610 * Request reference is cleared and it is guaranteed to be observed
3611 * after the ->lock is released.
3613 spin_lock_irqsave(&tags->lock, flags);
3614 spin_unlock_irqrestore(&tags->lock, flags);
3617 /* hctx->ctxs will be freed in queue's release handler */
3618 static void blk_mq_exit_hctx(struct request_queue *q,
3619 struct blk_mq_tag_set *set,
3620 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3622 struct request *flush_rq = hctx->fq->flush_rq;
3624 if (blk_mq_hw_queue_mapped(hctx))
3625 blk_mq_tag_idle(hctx);
3627 if (blk_queue_init_done(q))
3628 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3629 set->queue_depth, flush_rq);
3630 if (set->ops->exit_request)
3631 set->ops->exit_request(set, flush_rq, hctx_idx);
3633 if (set->ops->exit_hctx)
3634 set->ops->exit_hctx(hctx, hctx_idx);
3636 blk_mq_remove_cpuhp(hctx);
3638 xa_erase(&q->hctx_table, hctx_idx);
3640 spin_lock(&q->unused_hctx_lock);
3641 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3642 spin_unlock(&q->unused_hctx_lock);
3645 static void blk_mq_exit_hw_queues(struct request_queue *q,
3646 struct blk_mq_tag_set *set, int nr_queue)
3648 struct blk_mq_hw_ctx *hctx;
3651 queue_for_each_hw_ctx(q, hctx, i) {
3654 blk_mq_exit_hctx(q, set, hctx, i);
3658 static int blk_mq_init_hctx(struct request_queue *q,
3659 struct blk_mq_tag_set *set,
3660 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3662 hctx->queue_num = hctx_idx;
3664 if (!(hctx->flags & BLK_MQ_F_STACKING))
3665 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3666 &hctx->cpuhp_online);
3667 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3669 hctx->tags = set->tags[hctx_idx];
3671 if (set->ops->init_hctx &&
3672 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3673 goto unregister_cpu_notifier;
3675 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3679 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3685 if (set->ops->exit_request)
3686 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3688 if (set->ops->exit_hctx)
3689 set->ops->exit_hctx(hctx, hctx_idx);
3690 unregister_cpu_notifier:
3691 blk_mq_remove_cpuhp(hctx);
3695 static struct blk_mq_hw_ctx *
3696 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3699 struct blk_mq_hw_ctx *hctx;
3700 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3702 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3704 goto fail_alloc_hctx;
3706 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3709 atomic_set(&hctx->nr_active, 0);
3710 if (node == NUMA_NO_NODE)
3711 node = set->numa_node;
3712 hctx->numa_node = node;
3714 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3715 spin_lock_init(&hctx->lock);
3716 INIT_LIST_HEAD(&hctx->dispatch);
3718 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3720 INIT_LIST_HEAD(&hctx->hctx_list);
3723 * Allocate space for all possible cpus to avoid allocation at
3726 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3731 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3732 gfp, node, false, false))
3736 spin_lock_init(&hctx->dispatch_wait_lock);
3737 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3738 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3740 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3744 blk_mq_hctx_kobj_init(hctx);
3749 sbitmap_free(&hctx->ctx_map);
3753 free_cpumask_var(hctx->cpumask);
3760 static void blk_mq_init_cpu_queues(struct request_queue *q,
3761 unsigned int nr_hw_queues)
3763 struct blk_mq_tag_set *set = q->tag_set;
3766 for_each_possible_cpu(i) {
3767 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3768 struct blk_mq_hw_ctx *hctx;
3772 spin_lock_init(&__ctx->lock);
3773 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3774 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3779 * Set local node, IFF we have more than one hw queue. If
3780 * not, we remain on the home node of the device
3782 for (j = 0; j < set->nr_maps; j++) {
3783 hctx = blk_mq_map_queue_type(q, j, i);
3784 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3785 hctx->numa_node = cpu_to_node(i);
3790 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3791 unsigned int hctx_idx,
3794 struct blk_mq_tags *tags;
3797 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3801 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3803 blk_mq_free_rq_map(tags);
3810 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3813 if (blk_mq_is_shared_tags(set->flags)) {
3814 set->tags[hctx_idx] = set->shared_tags;
3819 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3822 return set->tags[hctx_idx];
3825 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3826 struct blk_mq_tags *tags,
3827 unsigned int hctx_idx)
3830 blk_mq_free_rqs(set, tags, hctx_idx);
3831 blk_mq_free_rq_map(tags);
3835 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3836 unsigned int hctx_idx)
3838 if (!blk_mq_is_shared_tags(set->flags))
3839 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3841 set->tags[hctx_idx] = NULL;
3844 static void blk_mq_map_swqueue(struct request_queue *q)
3846 unsigned int j, hctx_idx;
3848 struct blk_mq_hw_ctx *hctx;
3849 struct blk_mq_ctx *ctx;
3850 struct blk_mq_tag_set *set = q->tag_set;
3852 queue_for_each_hw_ctx(q, hctx, i) {
3853 cpumask_clear(hctx->cpumask);
3855 hctx->dispatch_from = NULL;
3859 * Map software to hardware queues.
3861 * If the cpu isn't present, the cpu is mapped to first hctx.
3863 for_each_possible_cpu(i) {
3865 ctx = per_cpu_ptr(q->queue_ctx, i);
3866 for (j = 0; j < set->nr_maps; j++) {
3867 if (!set->map[j].nr_queues) {
3868 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3869 HCTX_TYPE_DEFAULT, i);
3872 hctx_idx = set->map[j].mq_map[i];
3873 /* unmapped hw queue can be remapped after CPU topo changed */
3874 if (!set->tags[hctx_idx] &&
3875 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3877 * If tags initialization fail for some hctx,
3878 * that hctx won't be brought online. In this
3879 * case, remap the current ctx to hctx[0] which
3880 * is guaranteed to always have tags allocated
3882 set->map[j].mq_map[i] = 0;
3885 hctx = blk_mq_map_queue_type(q, j, i);
3886 ctx->hctxs[j] = hctx;
3888 * If the CPU is already set in the mask, then we've
3889 * mapped this one already. This can happen if
3890 * devices share queues across queue maps.
3892 if (cpumask_test_cpu(i, hctx->cpumask))
3895 cpumask_set_cpu(i, hctx->cpumask);
3897 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3898 hctx->ctxs[hctx->nr_ctx++] = ctx;
3901 * If the nr_ctx type overflows, we have exceeded the
3902 * amount of sw queues we can support.
3904 BUG_ON(!hctx->nr_ctx);
3907 for (; j < HCTX_MAX_TYPES; j++)
3908 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3909 HCTX_TYPE_DEFAULT, i);
3912 queue_for_each_hw_ctx(q, hctx, i) {
3914 * If no software queues are mapped to this hardware queue,
3915 * disable it and free the request entries.
3917 if (!hctx->nr_ctx) {
3918 /* Never unmap queue 0. We need it as a
3919 * fallback in case of a new remap fails
3923 __blk_mq_free_map_and_rqs(set, i);
3929 hctx->tags = set->tags[i];
3930 WARN_ON(!hctx->tags);
3933 * Set the map size to the number of mapped software queues.
3934 * This is more accurate and more efficient than looping
3935 * over all possibly mapped software queues.
3937 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3940 * Initialize batch roundrobin counts
3942 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3943 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3948 * Caller needs to ensure that we're either frozen/quiesced, or that
3949 * the queue isn't live yet.
3951 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3953 struct blk_mq_hw_ctx *hctx;
3956 queue_for_each_hw_ctx(q, hctx, i) {
3958 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3960 blk_mq_tag_idle(hctx);
3961 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3966 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3969 struct request_queue *q;
3971 lockdep_assert_held(&set->tag_list_lock);
3973 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3974 blk_mq_freeze_queue(q);
3975 queue_set_hctx_shared(q, shared);
3976 blk_mq_unfreeze_queue(q);
3980 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3982 struct blk_mq_tag_set *set = q->tag_set;
3984 mutex_lock(&set->tag_list_lock);
3985 list_del(&q->tag_set_list);
3986 if (list_is_singular(&set->tag_list)) {
3987 /* just transitioned to unshared */
3988 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3989 /* update existing queue */
3990 blk_mq_update_tag_set_shared(set, false);
3992 mutex_unlock(&set->tag_list_lock);
3993 INIT_LIST_HEAD(&q->tag_set_list);
3996 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3997 struct request_queue *q)
3999 mutex_lock(&set->tag_list_lock);
4002 * Check to see if we're transitioning to shared (from 1 to 2 queues).
4004 if (!list_empty(&set->tag_list) &&
4005 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
4006 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
4007 /* update existing queue */
4008 blk_mq_update_tag_set_shared(set, true);
4010 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
4011 queue_set_hctx_shared(q, true);
4012 list_add_tail(&q->tag_set_list, &set->tag_list);
4014 mutex_unlock(&set->tag_list_lock);
4017 /* All allocations will be freed in release handler of q->mq_kobj */
4018 static int blk_mq_alloc_ctxs(struct request_queue *q)
4020 struct blk_mq_ctxs *ctxs;
4023 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
4027 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
4028 if (!ctxs->queue_ctx)
4031 for_each_possible_cpu(cpu) {
4032 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4036 q->mq_kobj = &ctxs->kobj;
4037 q->queue_ctx = ctxs->queue_ctx;
4046 * It is the actual release handler for mq, but we do it from
4047 * request queue's release handler for avoiding use-after-free
4048 * and headache because q->mq_kobj shouldn't have been introduced,
4049 * but we can't group ctx/kctx kobj without it.
4051 void blk_mq_release(struct request_queue *q)
4053 struct blk_mq_hw_ctx *hctx, *next;
4056 queue_for_each_hw_ctx(q, hctx, i)
4057 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4059 /* all hctx are in .unused_hctx_list now */
4060 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4061 list_del_init(&hctx->hctx_list);
4062 kobject_put(&hctx->kobj);
4065 xa_destroy(&q->hctx_table);
4068 * release .mq_kobj and sw queue's kobject now because
4069 * both share lifetime with request queue.
4071 blk_mq_sysfs_deinit(q);
4074 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4077 struct request_queue *q;
4080 q = blk_alloc_queue(set->numa_node);
4082 return ERR_PTR(-ENOMEM);
4083 q->queuedata = queuedata;
4084 ret = blk_mq_init_allocated_queue(set, q);
4087 return ERR_PTR(ret);
4092 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4094 return blk_mq_init_queue_data(set, NULL);
4096 EXPORT_SYMBOL(blk_mq_init_queue);
4099 * blk_mq_destroy_queue - shutdown a request queue
4100 * @q: request queue to shutdown
4102 * This shuts down a request queue allocated by blk_mq_init_queue(). All future
4103 * requests will be failed with -ENODEV. The caller is responsible for dropping
4104 * the reference from blk_mq_init_queue() by calling blk_put_queue().
4106 * Context: can sleep
4108 void blk_mq_destroy_queue(struct request_queue *q)
4110 WARN_ON_ONCE(!queue_is_mq(q));
4111 WARN_ON_ONCE(blk_queue_registered(q));
4115 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4116 blk_queue_start_drain(q);
4117 blk_mq_freeze_queue_wait(q);
4120 blk_mq_cancel_work_sync(q);
4121 blk_mq_exit_queue(q);
4123 EXPORT_SYMBOL(blk_mq_destroy_queue);
4125 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4126 struct lock_class_key *lkclass)
4128 struct request_queue *q;
4129 struct gendisk *disk;
4131 q = blk_mq_init_queue_data(set, queuedata);
4135 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4137 blk_mq_destroy_queue(q);
4139 return ERR_PTR(-ENOMEM);
4141 set_bit(GD_OWNS_QUEUE, &disk->state);
4144 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4146 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4147 struct lock_class_key *lkclass)
4149 struct gendisk *disk;
4151 if (!blk_get_queue(q))
4153 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4158 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4160 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4161 struct blk_mq_tag_set *set, struct request_queue *q,
4162 int hctx_idx, int node)
4164 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4166 /* reuse dead hctx first */
4167 spin_lock(&q->unused_hctx_lock);
4168 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4169 if (tmp->numa_node == node) {
4175 list_del_init(&hctx->hctx_list);
4176 spin_unlock(&q->unused_hctx_lock);
4179 hctx = blk_mq_alloc_hctx(q, set, node);
4183 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4189 kobject_put(&hctx->kobj);
4194 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4195 struct request_queue *q)
4197 struct blk_mq_hw_ctx *hctx;
4200 /* protect against switching io scheduler */
4201 mutex_lock(&q->sysfs_lock);
4202 for (i = 0; i < set->nr_hw_queues; i++) {
4204 int node = blk_mq_get_hctx_node(set, i);
4205 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4208 old_node = old_hctx->numa_node;
4209 blk_mq_exit_hctx(q, set, old_hctx, i);
4212 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4215 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4217 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4218 WARN_ON_ONCE(!hctx);
4222 * Increasing nr_hw_queues fails. Free the newly allocated
4223 * hctxs and keep the previous q->nr_hw_queues.
4225 if (i != set->nr_hw_queues) {
4226 j = q->nr_hw_queues;
4229 q->nr_hw_queues = set->nr_hw_queues;
4232 xa_for_each_start(&q->hctx_table, j, hctx, j)
4233 blk_mq_exit_hctx(q, set, hctx, j);
4234 mutex_unlock(&q->sysfs_lock);
4237 static void blk_mq_update_poll_flag(struct request_queue *q)
4239 struct blk_mq_tag_set *set = q->tag_set;
4241 if (set->nr_maps > HCTX_TYPE_POLL &&
4242 set->map[HCTX_TYPE_POLL].nr_queues)
4243 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4245 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4248 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4249 struct request_queue *q)
4251 /* mark the queue as mq asap */
4252 q->mq_ops = set->ops;
4254 if (blk_mq_alloc_ctxs(q))
4257 /* init q->mq_kobj and sw queues' kobjects */
4258 blk_mq_sysfs_init(q);
4260 INIT_LIST_HEAD(&q->unused_hctx_list);
4261 spin_lock_init(&q->unused_hctx_lock);
4263 xa_init(&q->hctx_table);
4265 blk_mq_realloc_hw_ctxs(set, q);
4266 if (!q->nr_hw_queues)
4269 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4270 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4274 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4275 blk_mq_update_poll_flag(q);
4277 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4278 INIT_LIST_HEAD(&q->flush_list);
4279 INIT_LIST_HEAD(&q->requeue_list);
4280 spin_lock_init(&q->requeue_lock);
4282 q->nr_requests = set->queue_depth;
4284 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4285 blk_mq_add_queue_tag_set(set, q);
4286 blk_mq_map_swqueue(q);
4295 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4297 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4298 void blk_mq_exit_queue(struct request_queue *q)
4300 struct blk_mq_tag_set *set = q->tag_set;
4302 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4303 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4304 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4305 blk_mq_del_queue_tag_set(q);
4308 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4312 if (blk_mq_is_shared_tags(set->flags)) {
4313 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4316 if (!set->shared_tags)
4320 for (i = 0; i < set->nr_hw_queues; i++) {
4321 if (!__blk_mq_alloc_map_and_rqs(set, i))
4330 __blk_mq_free_map_and_rqs(set, i);
4332 if (blk_mq_is_shared_tags(set->flags)) {
4333 blk_mq_free_map_and_rqs(set, set->shared_tags,
4334 BLK_MQ_NO_HCTX_IDX);
4341 * Allocate the request maps associated with this tag_set. Note that this
4342 * may reduce the depth asked for, if memory is tight. set->queue_depth
4343 * will be updated to reflect the allocated depth.
4345 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4350 depth = set->queue_depth;
4352 err = __blk_mq_alloc_rq_maps(set);
4356 set->queue_depth >>= 1;
4357 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4361 } while (set->queue_depth);
4363 if (!set->queue_depth || err) {
4364 pr_err("blk-mq: failed to allocate request map\n");
4368 if (depth != set->queue_depth)
4369 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4370 depth, set->queue_depth);
4375 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4378 * blk_mq_map_queues() and multiple .map_queues() implementations
4379 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4380 * number of hardware queues.
4382 if (set->nr_maps == 1)
4383 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4385 if (set->ops->map_queues && !is_kdump_kernel()) {
4389 * transport .map_queues is usually done in the following
4392 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4393 * mask = get_cpu_mask(queue)
4394 * for_each_cpu(cpu, mask)
4395 * set->map[x].mq_map[cpu] = queue;
4398 * When we need to remap, the table has to be cleared for
4399 * killing stale mapping since one CPU may not be mapped
4402 for (i = 0; i < set->nr_maps; i++)
4403 blk_mq_clear_mq_map(&set->map[i]);
4405 set->ops->map_queues(set);
4407 BUG_ON(set->nr_maps > 1);
4408 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4412 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4413 int new_nr_hw_queues)
4415 struct blk_mq_tags **new_tags;
4418 if (set->nr_hw_queues >= new_nr_hw_queues)
4421 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4422 GFP_KERNEL, set->numa_node);
4427 memcpy(new_tags, set->tags, set->nr_hw_queues *
4428 sizeof(*set->tags));
4430 set->tags = new_tags;
4432 for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) {
4433 if (!__blk_mq_alloc_map_and_rqs(set, i)) {
4434 while (--i >= set->nr_hw_queues)
4435 __blk_mq_free_map_and_rqs(set, i);
4442 set->nr_hw_queues = new_nr_hw_queues;
4447 * Alloc a tag set to be associated with one or more request queues.
4448 * May fail with EINVAL for various error conditions. May adjust the
4449 * requested depth down, if it's too large. In that case, the set
4450 * value will be stored in set->queue_depth.
4452 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4456 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4458 if (!set->nr_hw_queues)
4460 if (!set->queue_depth)
4462 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4465 if (!set->ops->queue_rq)
4468 if (!set->ops->get_budget ^ !set->ops->put_budget)
4471 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4472 pr_info("blk-mq: reduced tag depth to %u\n",
4474 set->queue_depth = BLK_MQ_MAX_DEPTH;
4479 else if (set->nr_maps > HCTX_MAX_TYPES)
4483 * If a crashdump is active, then we are potentially in a very
4484 * memory constrained environment. Limit us to 1 queue and
4485 * 64 tags to prevent using too much memory.
4487 if (is_kdump_kernel()) {
4488 set->nr_hw_queues = 1;
4490 set->queue_depth = min(64U, set->queue_depth);
4493 * There is no use for more h/w queues than cpus if we just have
4496 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4497 set->nr_hw_queues = nr_cpu_ids;
4499 if (set->flags & BLK_MQ_F_BLOCKING) {
4500 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4503 ret = init_srcu_struct(set->srcu);
4509 set->tags = kcalloc_node(set->nr_hw_queues,
4510 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4513 goto out_cleanup_srcu;
4515 for (i = 0; i < set->nr_maps; i++) {
4516 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4517 sizeof(set->map[i].mq_map[0]),
4518 GFP_KERNEL, set->numa_node);
4519 if (!set->map[i].mq_map)
4520 goto out_free_mq_map;
4521 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4524 blk_mq_update_queue_map(set);
4526 ret = blk_mq_alloc_set_map_and_rqs(set);
4528 goto out_free_mq_map;
4530 mutex_init(&set->tag_list_lock);
4531 INIT_LIST_HEAD(&set->tag_list);
4536 for (i = 0; i < set->nr_maps; i++) {
4537 kfree(set->map[i].mq_map);
4538 set->map[i].mq_map = NULL;
4543 if (set->flags & BLK_MQ_F_BLOCKING)
4544 cleanup_srcu_struct(set->srcu);
4546 if (set->flags & BLK_MQ_F_BLOCKING)
4550 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4552 /* allocate and initialize a tagset for a simple single-queue device */
4553 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4554 const struct blk_mq_ops *ops, unsigned int queue_depth,
4555 unsigned int set_flags)
4557 memset(set, 0, sizeof(*set));
4559 set->nr_hw_queues = 1;
4561 set->queue_depth = queue_depth;
4562 set->numa_node = NUMA_NO_NODE;
4563 set->flags = set_flags;
4564 return blk_mq_alloc_tag_set(set);
4566 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4568 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4572 for (i = 0; i < set->nr_hw_queues; i++)
4573 __blk_mq_free_map_and_rqs(set, i);
4575 if (blk_mq_is_shared_tags(set->flags)) {
4576 blk_mq_free_map_and_rqs(set, set->shared_tags,
4577 BLK_MQ_NO_HCTX_IDX);
4580 for (j = 0; j < set->nr_maps; j++) {
4581 kfree(set->map[j].mq_map);
4582 set->map[j].mq_map = NULL;
4587 if (set->flags & BLK_MQ_F_BLOCKING) {
4588 cleanup_srcu_struct(set->srcu);
4592 EXPORT_SYMBOL(blk_mq_free_tag_set);
4594 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4596 struct blk_mq_tag_set *set = q->tag_set;
4597 struct blk_mq_hw_ctx *hctx;
4604 if (q->nr_requests == nr)
4607 blk_mq_freeze_queue(q);
4608 blk_mq_quiesce_queue(q);
4611 queue_for_each_hw_ctx(q, hctx, i) {
4615 * If we're using an MQ scheduler, just update the scheduler
4616 * queue depth. This is similar to what the old code would do.
4618 if (hctx->sched_tags) {
4619 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4622 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4627 if (q->elevator && q->elevator->type->ops.depth_updated)
4628 q->elevator->type->ops.depth_updated(hctx);
4631 q->nr_requests = nr;
4632 if (blk_mq_is_shared_tags(set->flags)) {
4634 blk_mq_tag_update_sched_shared_tags(q);
4636 blk_mq_tag_resize_shared_tags(set, nr);
4640 blk_mq_unquiesce_queue(q);
4641 blk_mq_unfreeze_queue(q);
4647 * request_queue and elevator_type pair.
4648 * It is just used by __blk_mq_update_nr_hw_queues to cache
4649 * the elevator_type associated with a request_queue.
4651 struct blk_mq_qe_pair {
4652 struct list_head node;
4653 struct request_queue *q;
4654 struct elevator_type *type;
4658 * Cache the elevator_type in qe pair list and switch the
4659 * io scheduler to 'none'
4661 static bool blk_mq_elv_switch_none(struct list_head *head,
4662 struct request_queue *q)
4664 struct blk_mq_qe_pair *qe;
4666 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4670 /* q->elevator needs protection from ->sysfs_lock */
4671 mutex_lock(&q->sysfs_lock);
4673 /* the check has to be done with holding sysfs_lock */
4679 INIT_LIST_HEAD(&qe->node);
4681 qe->type = q->elevator->type;
4682 /* keep a reference to the elevator module as we'll switch back */
4683 __elevator_get(qe->type);
4684 list_add(&qe->node, head);
4685 elevator_disable(q);
4687 mutex_unlock(&q->sysfs_lock);
4692 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4693 struct request_queue *q)
4695 struct blk_mq_qe_pair *qe;
4697 list_for_each_entry(qe, head, node)
4704 static void blk_mq_elv_switch_back(struct list_head *head,
4705 struct request_queue *q)
4707 struct blk_mq_qe_pair *qe;
4708 struct elevator_type *t;
4710 qe = blk_lookup_qe_pair(head, q);
4714 list_del(&qe->node);
4717 mutex_lock(&q->sysfs_lock);
4718 elevator_switch(q, t);
4719 /* drop the reference acquired in blk_mq_elv_switch_none */
4721 mutex_unlock(&q->sysfs_lock);
4724 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4727 struct request_queue *q;
4729 int prev_nr_hw_queues = set->nr_hw_queues;
4732 lockdep_assert_held(&set->tag_list_lock);
4734 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4735 nr_hw_queues = nr_cpu_ids;
4736 if (nr_hw_queues < 1)
4738 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4741 list_for_each_entry(q, &set->tag_list, tag_set_list)
4742 blk_mq_freeze_queue(q);
4744 * Switch IO scheduler to 'none', cleaning up the data associated
4745 * with the previous scheduler. We will switch back once we are done
4746 * updating the new sw to hw queue mappings.
4748 list_for_each_entry(q, &set->tag_list, tag_set_list)
4749 if (!blk_mq_elv_switch_none(&head, q))
4752 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4753 blk_mq_debugfs_unregister_hctxs(q);
4754 blk_mq_sysfs_unregister_hctxs(q);
4757 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4761 blk_mq_update_queue_map(set);
4762 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4763 blk_mq_realloc_hw_ctxs(set, q);
4764 blk_mq_update_poll_flag(q);
4765 if (q->nr_hw_queues != set->nr_hw_queues) {
4766 int i = prev_nr_hw_queues;
4768 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4769 nr_hw_queues, prev_nr_hw_queues);
4770 for (; i < set->nr_hw_queues; i++)
4771 __blk_mq_free_map_and_rqs(set, i);
4773 set->nr_hw_queues = prev_nr_hw_queues;
4776 blk_mq_map_swqueue(q);
4780 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4781 blk_mq_sysfs_register_hctxs(q);
4782 blk_mq_debugfs_register_hctxs(q);
4786 list_for_each_entry(q, &set->tag_list, tag_set_list)
4787 blk_mq_elv_switch_back(&head, q);
4789 list_for_each_entry(q, &set->tag_list, tag_set_list)
4790 blk_mq_unfreeze_queue(q);
4792 /* Free the excess tags when nr_hw_queues shrink. */
4793 for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)
4794 __blk_mq_free_map_and_rqs(set, i);
4797 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4799 mutex_lock(&set->tag_list_lock);
4800 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4801 mutex_unlock(&set->tag_list_lock);
4803 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4805 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
4806 struct io_comp_batch *iob, unsigned int flags)
4808 long state = get_current_state();
4812 ret = q->mq_ops->poll(hctx, iob);
4814 __set_current_state(TASK_RUNNING);
4818 if (signal_pending_state(state, current))
4819 __set_current_state(TASK_RUNNING);
4820 if (task_is_running(current))
4823 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4826 } while (!need_resched());
4828 __set_current_state(TASK_RUNNING);
4832 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
4833 struct io_comp_batch *iob, unsigned int flags)
4835 struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie);
4837 return blk_hctx_poll(q, hctx, iob, flags);
4840 int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
4841 unsigned int poll_flags)
4843 struct request_queue *q = rq->q;
4846 if (!blk_rq_is_poll(rq))
4848 if (!percpu_ref_tryget(&q->q_usage_counter))
4851 ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags);
4856 EXPORT_SYMBOL_GPL(blk_rq_poll);
4858 unsigned int blk_mq_rq_cpu(struct request *rq)
4860 return rq->mq_ctx->cpu;
4862 EXPORT_SYMBOL(blk_mq_rq_cpu);
4864 void blk_mq_cancel_work_sync(struct request_queue *q)
4866 struct blk_mq_hw_ctx *hctx;
4869 cancel_delayed_work_sync(&q->requeue_work);
4871 queue_for_each_hw_ctx(q, hctx, i)
4872 cancel_delayed_work_sync(&hctx->run_work);
4875 static int __init blk_mq_init(void)
4879 for_each_possible_cpu(i)
4880 init_llist_head(&per_cpu(blk_cpu_done, i));
4881 for_each_possible_cpu(i)
4882 INIT_CSD(&per_cpu(blk_cpu_csd, i),
4883 __blk_mq_complete_request_remote, NULL);
4884 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4886 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4887 "block/softirq:dead", NULL,
4888 blk_softirq_cpu_dead);
4889 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4890 blk_mq_hctx_notify_dead);
4891 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4892 blk_mq_hctx_notify_online,
4893 blk_mq_hctx_notify_offline);
4896 subsys_initcall(blk_mq_init);