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/topology.h>
25 #include <linux/sched/signal.h>
26 #include <linux/delay.h>
27 #include <linux/crash_dump.h>
28 #include <linux/prefetch.h>
29 #include <linux/blk-crypto.h>
30 #include <linux/part_stat.h>
31 #include <linux/sched/isolation.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 = blk_time_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 = blk_time_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 = blk_time_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 = blk_time_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, blk_time_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 = blk_time_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 and capacity? 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) &&
1169 cpus_equal_capacity(cpu, rq->mq_ctx->cpu)))
1172 /* don't try to IPI to an offline CPU */
1173 return cpu_online(rq->mq_ctx->cpu);
1176 static void blk_mq_complete_send_ipi(struct request *rq)
1180 cpu = rq->mq_ctx->cpu;
1181 if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu)))
1182 smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu));
1185 static void blk_mq_raise_softirq(struct request *rq)
1187 struct llist_head *list;
1190 list = this_cpu_ptr(&blk_cpu_done);
1191 if (llist_add(&rq->ipi_list, list))
1192 raise_softirq(BLOCK_SOFTIRQ);
1196 bool blk_mq_complete_request_remote(struct request *rq)
1198 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1201 * For request which hctx has only one ctx mapping,
1202 * or a polled request, always complete locally,
1203 * it's pointless to redirect the completion.
1205 if ((rq->mq_hctx->nr_ctx == 1 &&
1206 rq->mq_ctx->cpu == raw_smp_processor_id()) ||
1207 rq->cmd_flags & REQ_POLLED)
1210 if (blk_mq_complete_need_ipi(rq)) {
1211 blk_mq_complete_send_ipi(rq);
1215 if (rq->q->nr_hw_queues == 1) {
1216 blk_mq_raise_softirq(rq);
1221 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1224 * blk_mq_complete_request - end I/O on a request
1225 * @rq: the request being processed
1228 * Complete a request by scheduling the ->complete_rq operation.
1230 void blk_mq_complete_request(struct request *rq)
1232 if (!blk_mq_complete_request_remote(rq))
1233 rq->q->mq_ops->complete(rq);
1235 EXPORT_SYMBOL(blk_mq_complete_request);
1238 * blk_mq_start_request - Start processing a request
1239 * @rq: Pointer to request to be started
1241 * Function used by device drivers to notify the block layer that a request
1242 * is going to be processed now, so blk layer can do proper initializations
1243 * such as starting the timeout timer.
1245 void blk_mq_start_request(struct request *rq)
1247 struct request_queue *q = rq->q;
1249 trace_block_rq_issue(rq);
1251 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) &&
1252 !blk_rq_is_passthrough(rq)) {
1253 rq->io_start_time_ns = blk_time_get_ns();
1254 rq->stats_sectors = blk_rq_sectors(rq);
1255 rq->rq_flags |= RQF_STATS;
1256 rq_qos_issue(q, rq);
1259 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1262 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1263 rq->mq_hctx->tags->rqs[rq->tag] = rq;
1265 #ifdef CONFIG_BLK_DEV_INTEGRITY
1266 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1267 q->integrity.profile->prepare_fn(rq);
1269 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1270 WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num);
1272 EXPORT_SYMBOL(blk_mq_start_request);
1275 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1276 * queues. This is important for md arrays to benefit from merging
1279 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1281 if (plug->multiple_queues)
1282 return BLK_MAX_REQUEST_COUNT * 2;
1283 return BLK_MAX_REQUEST_COUNT;
1286 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1288 struct request *last = rq_list_peek(&plug->mq_list);
1290 if (!plug->rq_count) {
1291 trace_block_plug(rq->q);
1292 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1293 (!blk_queue_nomerges(rq->q) &&
1294 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1295 blk_mq_flush_plug_list(plug, false);
1297 trace_block_plug(rq->q);
1300 if (!plug->multiple_queues && last && last->q != rq->q)
1301 plug->multiple_queues = true;
1303 * Any request allocated from sched tags can't be issued to
1304 * ->queue_rqs() directly
1306 if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))
1307 plug->has_elevator = true;
1309 rq_list_add(&plug->mq_list, rq);
1314 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1315 * @rq: request to insert
1316 * @at_head: insert request at head or tail of queue
1319 * Insert a fully prepared request at the back of the I/O scheduler queue
1320 * for execution. Don't wait for completion.
1323 * This function will invoke @done directly if the queue is dead.
1325 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1327 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1329 WARN_ON(irqs_disabled());
1330 WARN_ON(!blk_rq_is_passthrough(rq));
1332 blk_account_io_start(rq);
1335 * As plugging can be enabled for passthrough requests on a zoned
1336 * device, directly accessing the plug instead of using blk_mq_plug()
1337 * should not have any consequences.
1339 if (current->plug && !at_head) {
1340 blk_add_rq_to_plug(current->plug, rq);
1344 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1345 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
1347 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1349 struct blk_rq_wait {
1350 struct completion done;
1354 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1356 struct blk_rq_wait *wait = rq->end_io_data;
1359 complete(&wait->done);
1360 return RQ_END_IO_NONE;
1363 bool blk_rq_is_poll(struct request *rq)
1367 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1371 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1373 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1376 blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0);
1378 } while (!completion_done(wait));
1382 * blk_execute_rq - insert a request into queue for execution
1383 * @rq: request to insert
1384 * @at_head: insert request at head or tail of queue
1387 * Insert a fully prepared request at the back of the I/O scheduler queue
1388 * for execution and wait for completion.
1389 * Return: The blk_status_t result provided to blk_mq_end_request().
1391 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1393 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1394 struct blk_rq_wait wait = {
1395 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1398 WARN_ON(irqs_disabled());
1399 WARN_ON(!blk_rq_is_passthrough(rq));
1401 rq->end_io_data = &wait;
1402 rq->end_io = blk_end_sync_rq;
1404 blk_account_io_start(rq);
1405 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1406 blk_mq_run_hw_queue(hctx, false);
1408 if (blk_rq_is_poll(rq))
1409 blk_rq_poll_completion(rq, &wait.done);
1411 blk_wait_io(&wait.done);
1415 EXPORT_SYMBOL(blk_execute_rq);
1417 static void __blk_mq_requeue_request(struct request *rq)
1419 struct request_queue *q = rq->q;
1421 blk_mq_put_driver_tag(rq);
1423 trace_block_rq_requeue(rq);
1424 rq_qos_requeue(q, rq);
1426 if (blk_mq_request_started(rq)) {
1427 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1428 rq->rq_flags &= ~RQF_TIMED_OUT;
1432 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1434 struct request_queue *q = rq->q;
1435 unsigned long flags;
1437 __blk_mq_requeue_request(rq);
1439 /* this request will be re-inserted to io scheduler queue */
1440 blk_mq_sched_requeue_request(rq);
1442 spin_lock_irqsave(&q->requeue_lock, flags);
1443 list_add_tail(&rq->queuelist, &q->requeue_list);
1444 spin_unlock_irqrestore(&q->requeue_lock, flags);
1446 if (kick_requeue_list)
1447 blk_mq_kick_requeue_list(q);
1449 EXPORT_SYMBOL(blk_mq_requeue_request);
1451 static void blk_mq_requeue_work(struct work_struct *work)
1453 struct request_queue *q =
1454 container_of(work, struct request_queue, requeue_work.work);
1456 LIST_HEAD(flush_list);
1459 spin_lock_irq(&q->requeue_lock);
1460 list_splice_init(&q->requeue_list, &rq_list);
1461 list_splice_init(&q->flush_list, &flush_list);
1462 spin_unlock_irq(&q->requeue_lock);
1464 while (!list_empty(&rq_list)) {
1465 rq = list_entry(rq_list.next, struct request, queuelist);
1467 * If RQF_DONTPREP ist set, the request has been started by the
1468 * driver already and might have driver-specific data allocated
1469 * already. Insert it into the hctx dispatch list to avoid
1470 * block layer merges for the request.
1472 if (rq->rq_flags & RQF_DONTPREP) {
1473 list_del_init(&rq->queuelist);
1474 blk_mq_request_bypass_insert(rq, 0);
1476 list_del_init(&rq->queuelist);
1477 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1481 while (!list_empty(&flush_list)) {
1482 rq = list_entry(flush_list.next, struct request, queuelist);
1483 list_del_init(&rq->queuelist);
1484 blk_mq_insert_request(rq, 0);
1487 blk_mq_run_hw_queues(q, false);
1490 void blk_mq_kick_requeue_list(struct request_queue *q)
1492 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1494 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1496 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1497 unsigned long msecs)
1499 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1500 msecs_to_jiffies(msecs));
1502 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1504 static bool blk_is_flush_data_rq(struct request *rq)
1506 return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq);
1509 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1512 * If we find a request that isn't idle we know the queue is busy
1513 * as it's checked in the iter.
1514 * Return false to stop the iteration.
1516 * In case of queue quiesce, if one flush data request is completed,
1517 * don't count it as inflight given the flush sequence is suspended,
1518 * and the original flush data request is invisible to driver, just
1519 * like other pending requests because of quiesce
1521 if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) &&
1522 blk_is_flush_data_rq(rq) &&
1523 blk_mq_request_completed(rq))) {
1533 bool blk_mq_queue_inflight(struct request_queue *q)
1537 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1540 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1542 static void blk_mq_rq_timed_out(struct request *req)
1544 req->rq_flags |= RQF_TIMED_OUT;
1545 if (req->q->mq_ops->timeout) {
1546 enum blk_eh_timer_return ret;
1548 ret = req->q->mq_ops->timeout(req);
1549 if (ret == BLK_EH_DONE)
1551 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1557 struct blk_expired_data {
1558 bool has_timedout_rq;
1560 unsigned long timeout_start;
1563 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1565 unsigned long deadline;
1567 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1569 if (rq->rq_flags & RQF_TIMED_OUT)
1572 deadline = READ_ONCE(rq->deadline);
1573 if (time_after_eq(expired->timeout_start, deadline))
1576 if (expired->next == 0)
1577 expired->next = deadline;
1578 else if (time_after(expired->next, deadline))
1579 expired->next = deadline;
1583 void blk_mq_put_rq_ref(struct request *rq)
1585 if (is_flush_rq(rq)) {
1586 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1587 blk_mq_free_request(rq);
1588 } else if (req_ref_put_and_test(rq)) {
1589 __blk_mq_free_request(rq);
1593 static bool blk_mq_check_expired(struct request *rq, void *priv)
1595 struct blk_expired_data *expired = priv;
1598 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1599 * be reallocated underneath the timeout handler's processing, then
1600 * the expire check is reliable. If the request is not expired, then
1601 * it was completed and reallocated as a new request after returning
1602 * from blk_mq_check_expired().
1604 if (blk_mq_req_expired(rq, expired)) {
1605 expired->has_timedout_rq = true;
1611 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1613 struct blk_expired_data *expired = priv;
1615 if (blk_mq_req_expired(rq, expired))
1616 blk_mq_rq_timed_out(rq);
1620 static void blk_mq_timeout_work(struct work_struct *work)
1622 struct request_queue *q =
1623 container_of(work, struct request_queue, timeout_work);
1624 struct blk_expired_data expired = {
1625 .timeout_start = jiffies,
1627 struct blk_mq_hw_ctx *hctx;
1630 /* A deadlock might occur if a request is stuck requiring a
1631 * timeout at the same time a queue freeze is waiting
1632 * completion, since the timeout code would not be able to
1633 * acquire the queue reference here.
1635 * That's why we don't use blk_queue_enter here; instead, we use
1636 * percpu_ref_tryget directly, because we need to be able to
1637 * obtain a reference even in the short window between the queue
1638 * starting to freeze, by dropping the first reference in
1639 * blk_freeze_queue_start, and the moment the last request is
1640 * consumed, marked by the instant q_usage_counter reaches
1643 if (!percpu_ref_tryget(&q->q_usage_counter))
1646 /* check if there is any timed-out request */
1647 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1648 if (expired.has_timedout_rq) {
1650 * Before walking tags, we must ensure any submit started
1651 * before the current time has finished. Since the submit
1652 * uses srcu or rcu, wait for a synchronization point to
1653 * ensure all running submits have finished
1655 blk_mq_wait_quiesce_done(q->tag_set);
1658 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1661 if (expired.next != 0) {
1662 mod_timer(&q->timeout, expired.next);
1665 * Request timeouts are handled as a forward rolling timer. If
1666 * we end up here it means that no requests are pending and
1667 * also that no request has been pending for a while. Mark
1668 * each hctx as idle.
1670 queue_for_each_hw_ctx(q, hctx, i) {
1671 /* the hctx may be unmapped, so check it here */
1672 if (blk_mq_hw_queue_mapped(hctx))
1673 blk_mq_tag_idle(hctx);
1679 struct flush_busy_ctx_data {
1680 struct blk_mq_hw_ctx *hctx;
1681 struct list_head *list;
1684 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1686 struct flush_busy_ctx_data *flush_data = data;
1687 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1688 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1689 enum hctx_type type = hctx->type;
1691 spin_lock(&ctx->lock);
1692 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1693 sbitmap_clear_bit(sb, bitnr);
1694 spin_unlock(&ctx->lock);
1699 * Process software queues that have been marked busy, splicing them
1700 * to the for-dispatch
1702 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1704 struct flush_busy_ctx_data data = {
1709 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1711 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1713 struct dispatch_rq_data {
1714 struct blk_mq_hw_ctx *hctx;
1718 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1721 struct dispatch_rq_data *dispatch_data = data;
1722 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1723 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1724 enum hctx_type type = hctx->type;
1726 spin_lock(&ctx->lock);
1727 if (!list_empty(&ctx->rq_lists[type])) {
1728 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1729 list_del_init(&dispatch_data->rq->queuelist);
1730 if (list_empty(&ctx->rq_lists[type]))
1731 sbitmap_clear_bit(sb, bitnr);
1733 spin_unlock(&ctx->lock);
1735 return !dispatch_data->rq;
1738 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1739 struct blk_mq_ctx *start)
1741 unsigned off = start ? start->index_hw[hctx->type] : 0;
1742 struct dispatch_rq_data data = {
1747 __sbitmap_for_each_set(&hctx->ctx_map, off,
1748 dispatch_rq_from_ctx, &data);
1753 bool __blk_mq_alloc_driver_tag(struct request *rq)
1755 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1756 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1759 blk_mq_tag_busy(rq->mq_hctx);
1761 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1762 bt = &rq->mq_hctx->tags->breserved_tags;
1765 if (!hctx_may_queue(rq->mq_hctx, bt))
1769 tag = __sbitmap_queue_get(bt);
1770 if (tag == BLK_MQ_NO_TAG)
1773 rq->tag = tag + tag_offset;
1774 blk_mq_inc_active_requests(rq->mq_hctx);
1778 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1779 int flags, void *key)
1781 struct blk_mq_hw_ctx *hctx;
1783 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1785 spin_lock(&hctx->dispatch_wait_lock);
1786 if (!list_empty(&wait->entry)) {
1787 struct sbitmap_queue *sbq;
1789 list_del_init(&wait->entry);
1790 sbq = &hctx->tags->bitmap_tags;
1791 atomic_dec(&sbq->ws_active);
1793 spin_unlock(&hctx->dispatch_wait_lock);
1795 blk_mq_run_hw_queue(hctx, true);
1800 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1801 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1802 * restart. For both cases, take care to check the condition again after
1803 * marking us as waiting.
1805 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1808 struct sbitmap_queue *sbq;
1809 struct wait_queue_head *wq;
1810 wait_queue_entry_t *wait;
1813 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1814 !(blk_mq_is_shared_tags(hctx->flags))) {
1815 blk_mq_sched_mark_restart_hctx(hctx);
1818 * It's possible that a tag was freed in the window between the
1819 * allocation failure and adding the hardware queue to the wait
1822 * Don't clear RESTART here, someone else could have set it.
1823 * At most this will cost an extra queue run.
1825 return blk_mq_get_driver_tag(rq);
1828 wait = &hctx->dispatch_wait;
1829 if (!list_empty_careful(&wait->entry))
1832 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1833 sbq = &hctx->tags->breserved_tags;
1835 sbq = &hctx->tags->bitmap_tags;
1836 wq = &bt_wait_ptr(sbq, hctx)->wait;
1838 spin_lock_irq(&wq->lock);
1839 spin_lock(&hctx->dispatch_wait_lock);
1840 if (!list_empty(&wait->entry)) {
1841 spin_unlock(&hctx->dispatch_wait_lock);
1842 spin_unlock_irq(&wq->lock);
1846 atomic_inc(&sbq->ws_active);
1847 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1848 __add_wait_queue(wq, wait);
1851 * Add one explicit barrier since blk_mq_get_driver_tag() may
1852 * not imply barrier in case of failure.
1854 * Order adding us to wait queue and allocating driver tag.
1856 * The pair is the one implied in sbitmap_queue_wake_up() which
1857 * orders clearing sbitmap tag bits and waitqueue_active() in
1858 * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless
1860 * Otherwise, re-order of adding wait queue and getting driver tag
1861 * may cause __sbitmap_queue_wake_up() to wake up nothing because
1862 * the waitqueue_active() may not observe us in wait queue.
1867 * It's possible that a tag was freed in the window between the
1868 * allocation failure and adding the hardware queue to the wait
1871 ret = blk_mq_get_driver_tag(rq);
1873 spin_unlock(&hctx->dispatch_wait_lock);
1874 spin_unlock_irq(&wq->lock);
1879 * We got a tag, remove ourselves from the wait queue to ensure
1880 * someone else gets the wakeup.
1882 list_del_init(&wait->entry);
1883 atomic_dec(&sbq->ws_active);
1884 spin_unlock(&hctx->dispatch_wait_lock);
1885 spin_unlock_irq(&wq->lock);
1890 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1891 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1893 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1894 * - EWMA is one simple way to compute running average value
1895 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1896 * - take 4 as factor for avoiding to get too small(0) result, and this
1897 * factor doesn't matter because EWMA decreases exponentially
1899 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1903 ewma = hctx->dispatch_busy;
1908 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1910 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1911 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1913 hctx->dispatch_busy = ewma;
1916 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1918 static void blk_mq_handle_dev_resource(struct request *rq,
1919 struct list_head *list)
1921 list_add(&rq->queuelist, list);
1922 __blk_mq_requeue_request(rq);
1925 static void blk_mq_handle_zone_resource(struct request *rq,
1926 struct list_head *zone_list)
1929 * If we end up here it is because we cannot dispatch a request to a
1930 * specific zone due to LLD level zone-write locking or other zone
1931 * related resource not being available. In this case, set the request
1932 * aside in zone_list for retrying it later.
1934 list_add(&rq->queuelist, zone_list);
1935 __blk_mq_requeue_request(rq);
1938 enum prep_dispatch {
1940 PREP_DISPATCH_NO_TAG,
1941 PREP_DISPATCH_NO_BUDGET,
1944 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1947 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1948 int budget_token = -1;
1951 budget_token = blk_mq_get_dispatch_budget(rq->q);
1952 if (budget_token < 0) {
1953 blk_mq_put_driver_tag(rq);
1954 return PREP_DISPATCH_NO_BUDGET;
1956 blk_mq_set_rq_budget_token(rq, budget_token);
1959 if (!blk_mq_get_driver_tag(rq)) {
1961 * The initial allocation attempt failed, so we need to
1962 * rerun the hardware queue when a tag is freed. The
1963 * waitqueue takes care of that. If the queue is run
1964 * before we add this entry back on the dispatch list,
1965 * we'll re-run it below.
1967 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1969 * All budgets not got from this function will be put
1970 * together during handling partial dispatch
1973 blk_mq_put_dispatch_budget(rq->q, budget_token);
1974 return PREP_DISPATCH_NO_TAG;
1978 return PREP_DISPATCH_OK;
1981 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1982 static void blk_mq_release_budgets(struct request_queue *q,
1983 struct list_head *list)
1987 list_for_each_entry(rq, list, queuelist) {
1988 int budget_token = blk_mq_get_rq_budget_token(rq);
1990 if (budget_token >= 0)
1991 blk_mq_put_dispatch_budget(q, budget_token);
1996 * blk_mq_commit_rqs will notify driver using bd->last that there is no
1997 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
1999 * Attention, we should explicitly call this in unusual cases:
2000 * 1) did not queue everything initially scheduled to queue
2001 * 2) the last attempt to queue a request failed
2003 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
2006 if (hctx->queue->mq_ops->commit_rqs && queued) {
2007 trace_block_unplug(hctx->queue, queued, !from_schedule);
2008 hctx->queue->mq_ops->commit_rqs(hctx);
2013 * Returns true if we did some work AND can potentially do more.
2015 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
2016 unsigned int nr_budgets)
2018 enum prep_dispatch prep;
2019 struct request_queue *q = hctx->queue;
2022 blk_status_t ret = BLK_STS_OK;
2023 LIST_HEAD(zone_list);
2024 bool needs_resource = false;
2026 if (list_empty(list))
2030 * Now process all the entries, sending them to the driver.
2034 struct blk_mq_queue_data bd;
2036 rq = list_first_entry(list, struct request, queuelist);
2038 WARN_ON_ONCE(hctx != rq->mq_hctx);
2039 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2040 if (prep != PREP_DISPATCH_OK)
2043 list_del_init(&rq->queuelist);
2046 bd.last = list_empty(list);
2049 * once the request is queued to lld, no need to cover the
2054 ret = q->mq_ops->queue_rq(hctx, &bd);
2059 case BLK_STS_RESOURCE:
2060 needs_resource = true;
2062 case BLK_STS_DEV_RESOURCE:
2063 blk_mq_handle_dev_resource(rq, list);
2065 case BLK_STS_ZONE_RESOURCE:
2067 * Move the request to zone_list and keep going through
2068 * the dispatch list to find more requests the drive can
2071 blk_mq_handle_zone_resource(rq, &zone_list);
2072 needs_resource = true;
2075 blk_mq_end_request(rq, ret);
2077 } while (!list_empty(list));
2079 if (!list_empty(&zone_list))
2080 list_splice_tail_init(&zone_list, list);
2082 /* If we didn't flush the entire list, we could have told the driver
2083 * there was more coming, but that turned out to be a lie.
2085 if (!list_empty(list) || ret != BLK_STS_OK)
2086 blk_mq_commit_rqs(hctx, queued, false);
2089 * Any items that need requeuing? Stuff them into hctx->dispatch,
2090 * that is where we will continue on next queue run.
2092 if (!list_empty(list)) {
2094 /* For non-shared tags, the RESTART check will suffice */
2095 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2096 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2097 blk_mq_is_shared_tags(hctx->flags));
2100 blk_mq_release_budgets(q, list);
2102 spin_lock(&hctx->lock);
2103 list_splice_tail_init(list, &hctx->dispatch);
2104 spin_unlock(&hctx->lock);
2107 * Order adding requests to hctx->dispatch and checking
2108 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2109 * in blk_mq_sched_restart(). Avoid restart code path to
2110 * miss the new added requests to hctx->dispatch, meantime
2111 * SCHED_RESTART is observed here.
2116 * If SCHED_RESTART was set by the caller of this function and
2117 * it is no longer set that means that it was cleared by another
2118 * thread and hence that a queue rerun is needed.
2120 * If 'no_tag' is set, that means that we failed getting
2121 * a driver tag with an I/O scheduler attached. If our dispatch
2122 * waitqueue is no longer active, ensure that we run the queue
2123 * AFTER adding our entries back to the list.
2125 * If no I/O scheduler has been configured it is possible that
2126 * the hardware queue got stopped and restarted before requests
2127 * were pushed back onto the dispatch list. Rerun the queue to
2128 * avoid starvation. Notes:
2129 * - blk_mq_run_hw_queue() checks whether or not a queue has
2130 * been stopped before rerunning a queue.
2131 * - Some but not all block drivers stop a queue before
2132 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2135 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2136 * bit is set, run queue after a delay to avoid IO stalls
2137 * that could otherwise occur if the queue is idle. We'll do
2138 * similar if we couldn't get budget or couldn't lock a zone
2139 * and SCHED_RESTART is set.
2141 needs_restart = blk_mq_sched_needs_restart(hctx);
2142 if (prep == PREP_DISPATCH_NO_BUDGET)
2143 needs_resource = true;
2144 if (!needs_restart ||
2145 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2146 blk_mq_run_hw_queue(hctx, true);
2147 else if (needs_resource)
2148 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2150 blk_mq_update_dispatch_busy(hctx, true);
2154 blk_mq_update_dispatch_busy(hctx, false);
2158 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2160 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2162 if (cpu >= nr_cpu_ids)
2163 cpu = cpumask_first(hctx->cpumask);
2168 * ->next_cpu is always calculated from hctx->cpumask, so simply use
2169 * it for speeding up the check
2171 static bool blk_mq_hctx_empty_cpumask(struct blk_mq_hw_ctx *hctx)
2173 return hctx->next_cpu >= nr_cpu_ids;
2177 * It'd be great if the workqueue API had a way to pass
2178 * in a mask and had some smarts for more clever placement.
2179 * For now we just round-robin here, switching for every
2180 * BLK_MQ_CPU_WORK_BATCH queued items.
2182 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2185 int next_cpu = hctx->next_cpu;
2187 /* Switch to unbound if no allowable CPUs in this hctx */
2188 if (hctx->queue->nr_hw_queues == 1 || blk_mq_hctx_empty_cpumask(hctx))
2189 return WORK_CPU_UNBOUND;
2191 if (--hctx->next_cpu_batch <= 0) {
2193 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2195 if (next_cpu >= nr_cpu_ids)
2196 next_cpu = blk_mq_first_mapped_cpu(hctx);
2197 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2201 * Do unbound schedule if we can't find a online CPU for this hctx,
2202 * and it should only happen in the path of handling CPU DEAD.
2204 if (!cpu_online(next_cpu)) {
2211 * Make sure to re-select CPU next time once after CPUs
2212 * in hctx->cpumask become online again.
2214 hctx->next_cpu = next_cpu;
2215 hctx->next_cpu_batch = 1;
2216 return WORK_CPU_UNBOUND;
2219 hctx->next_cpu = next_cpu;
2224 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2225 * @hctx: Pointer to the hardware queue to run.
2226 * @msecs: Milliseconds of delay to wait before running the queue.
2228 * Run a hardware queue asynchronously with a delay of @msecs.
2230 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2232 if (unlikely(blk_mq_hctx_stopped(hctx)))
2234 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2235 msecs_to_jiffies(msecs));
2237 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2240 * blk_mq_run_hw_queue - Start to run a hardware queue.
2241 * @hctx: Pointer to the hardware queue to run.
2242 * @async: If we want to run the queue asynchronously.
2244 * Check if the request queue is not in a quiesced state and if there are
2245 * pending requests to be sent. If this is true, run the queue to send requests
2248 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2253 * We can't run the queue inline with interrupts disabled.
2255 WARN_ON_ONCE(!async && in_interrupt());
2257 might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING);
2260 * When queue is quiesced, we may be switching io scheduler, or
2261 * updating nr_hw_queues, or other things, and we can't run queue
2262 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2264 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2267 __blk_mq_run_dispatch_ops(hctx->queue, false,
2268 need_run = !blk_queue_quiesced(hctx->queue) &&
2269 blk_mq_hctx_has_pending(hctx));
2274 if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2275 blk_mq_delay_run_hw_queue(hctx, 0);
2279 blk_mq_run_dispatch_ops(hctx->queue,
2280 blk_mq_sched_dispatch_requests(hctx));
2282 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2285 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2288 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2290 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2292 * If the IO scheduler does not respect hardware queues when
2293 * dispatching, we just don't bother with multiple HW queues and
2294 * dispatch from hctx for the current CPU since running multiple queues
2295 * just causes lock contention inside the scheduler and pointless cache
2298 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2300 if (!blk_mq_hctx_stopped(hctx))
2306 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2307 * @q: Pointer to the request queue to run.
2308 * @async: If we want to run the queue asynchronously.
2310 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2312 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2316 if (blk_queue_sq_sched(q))
2317 sq_hctx = blk_mq_get_sq_hctx(q);
2318 queue_for_each_hw_ctx(q, hctx, i) {
2319 if (blk_mq_hctx_stopped(hctx))
2322 * Dispatch from this hctx either if there's no hctx preferred
2323 * by IO scheduler or if it has requests that bypass the
2326 if (!sq_hctx || sq_hctx == hctx ||
2327 !list_empty_careful(&hctx->dispatch))
2328 blk_mq_run_hw_queue(hctx, async);
2331 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2334 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2335 * @q: Pointer to the request queue to run.
2336 * @msecs: Milliseconds of delay to wait before running the queues.
2338 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2340 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2344 if (blk_queue_sq_sched(q))
2345 sq_hctx = blk_mq_get_sq_hctx(q);
2346 queue_for_each_hw_ctx(q, hctx, i) {
2347 if (blk_mq_hctx_stopped(hctx))
2350 * If there is already a run_work pending, leave the
2351 * pending delay untouched. Otherwise, a hctx can stall
2352 * if another hctx is re-delaying the other's work
2353 * before the work executes.
2355 if (delayed_work_pending(&hctx->run_work))
2358 * Dispatch from this hctx either if there's no hctx preferred
2359 * by IO scheduler or if it has requests that bypass the
2362 if (!sq_hctx || sq_hctx == hctx ||
2363 !list_empty_careful(&hctx->dispatch))
2364 blk_mq_delay_run_hw_queue(hctx, msecs);
2367 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2370 * This function is often used for pausing .queue_rq() by driver when
2371 * there isn't enough resource or some conditions aren't satisfied, and
2372 * BLK_STS_RESOURCE is usually returned.
2374 * We do not guarantee that dispatch can be drained or blocked
2375 * after blk_mq_stop_hw_queue() returns. Please use
2376 * blk_mq_quiesce_queue() for that requirement.
2378 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2380 cancel_delayed_work(&hctx->run_work);
2382 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2384 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2387 * This function is often used for pausing .queue_rq() by driver when
2388 * there isn't enough resource or some conditions aren't satisfied, and
2389 * BLK_STS_RESOURCE is usually returned.
2391 * We do not guarantee that dispatch can be drained or blocked
2392 * after blk_mq_stop_hw_queues() returns. Please use
2393 * blk_mq_quiesce_queue() for that requirement.
2395 void blk_mq_stop_hw_queues(struct request_queue *q)
2397 struct blk_mq_hw_ctx *hctx;
2400 queue_for_each_hw_ctx(q, hctx, i)
2401 blk_mq_stop_hw_queue(hctx);
2403 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2405 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2407 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2409 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
2411 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2413 void blk_mq_start_hw_queues(struct request_queue *q)
2415 struct blk_mq_hw_ctx *hctx;
2418 queue_for_each_hw_ctx(q, hctx, i)
2419 blk_mq_start_hw_queue(hctx);
2421 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2423 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2425 if (!blk_mq_hctx_stopped(hctx))
2428 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2429 blk_mq_run_hw_queue(hctx, async);
2431 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2433 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2435 struct blk_mq_hw_ctx *hctx;
2438 queue_for_each_hw_ctx(q, hctx, i)
2439 blk_mq_start_stopped_hw_queue(hctx, async ||
2440 (hctx->flags & BLK_MQ_F_BLOCKING));
2442 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2444 static void blk_mq_run_work_fn(struct work_struct *work)
2446 struct blk_mq_hw_ctx *hctx =
2447 container_of(work, struct blk_mq_hw_ctx, run_work.work);
2449 blk_mq_run_dispatch_ops(hctx->queue,
2450 blk_mq_sched_dispatch_requests(hctx));
2454 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2455 * @rq: Pointer to request to be inserted.
2456 * @flags: BLK_MQ_INSERT_*
2458 * Should only be used carefully, when the caller knows we want to
2459 * bypass a potential IO scheduler on the target device.
2461 static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2463 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2465 spin_lock(&hctx->lock);
2466 if (flags & BLK_MQ_INSERT_AT_HEAD)
2467 list_add(&rq->queuelist, &hctx->dispatch);
2469 list_add_tail(&rq->queuelist, &hctx->dispatch);
2470 spin_unlock(&hctx->lock);
2473 static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2474 struct blk_mq_ctx *ctx, struct list_head *list,
2475 bool run_queue_async)
2478 enum hctx_type type = hctx->type;
2481 * Try to issue requests directly if the hw queue isn't busy to save an
2482 * extra enqueue & dequeue to the sw queue.
2484 if (!hctx->dispatch_busy && !run_queue_async) {
2485 blk_mq_run_dispatch_ops(hctx->queue,
2486 blk_mq_try_issue_list_directly(hctx, list));
2487 if (list_empty(list))
2492 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2495 list_for_each_entry(rq, list, queuelist) {
2496 BUG_ON(rq->mq_ctx != ctx);
2497 trace_block_rq_insert(rq);
2498 if (rq->cmd_flags & REQ_NOWAIT)
2499 run_queue_async = true;
2502 spin_lock(&ctx->lock);
2503 list_splice_tail_init(list, &ctx->rq_lists[type]);
2504 blk_mq_hctx_mark_pending(hctx, ctx);
2505 spin_unlock(&ctx->lock);
2507 blk_mq_run_hw_queue(hctx, run_queue_async);
2510 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2512 struct request_queue *q = rq->q;
2513 struct blk_mq_ctx *ctx = rq->mq_ctx;
2514 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2516 if (blk_rq_is_passthrough(rq)) {
2518 * Passthrough request have to be added to hctx->dispatch
2519 * directly. The device may be in a situation where it can't
2520 * handle FS request, and always returns BLK_STS_RESOURCE for
2521 * them, which gets them added to hctx->dispatch.
2523 * If a passthrough request is required to unblock the queues,
2524 * and it is added to the scheduler queue, there is no chance to
2525 * dispatch it given we prioritize requests in hctx->dispatch.
2527 blk_mq_request_bypass_insert(rq, flags);
2528 } else if (req_op(rq) == REQ_OP_FLUSH) {
2530 * Firstly normal IO request is inserted to scheduler queue or
2531 * sw queue, meantime we add flush request to dispatch queue(
2532 * hctx->dispatch) directly and there is at most one in-flight
2533 * flush request for each hw queue, so it doesn't matter to add
2534 * flush request to tail or front of the dispatch queue.
2536 * Secondly in case of NCQ, flush request belongs to non-NCQ
2537 * command, and queueing it will fail when there is any
2538 * in-flight normal IO request(NCQ command). When adding flush
2539 * rq to the front of hctx->dispatch, it is easier to introduce
2540 * extra time to flush rq's latency because of S_SCHED_RESTART
2541 * compared with adding to the tail of dispatch queue, then
2542 * chance of flush merge is increased, and less flush requests
2543 * will be issued to controller. It is observed that ~10% time
2544 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2545 * drive when adding flush rq to the front of hctx->dispatch.
2547 * Simply queue flush rq to the front of hctx->dispatch so that
2548 * intensive flush workloads can benefit in case of NCQ HW.
2550 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2551 } else if (q->elevator) {
2554 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2556 list_add(&rq->queuelist, &list);
2557 q->elevator->type->ops.insert_requests(hctx, &list, flags);
2559 trace_block_rq_insert(rq);
2561 spin_lock(&ctx->lock);
2562 if (flags & BLK_MQ_INSERT_AT_HEAD)
2563 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2565 list_add_tail(&rq->queuelist,
2566 &ctx->rq_lists[hctx->type]);
2567 blk_mq_hctx_mark_pending(hctx, ctx);
2568 spin_unlock(&ctx->lock);
2572 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2573 unsigned int nr_segs)
2577 if (bio->bi_opf & REQ_RAHEAD)
2578 rq->cmd_flags |= REQ_FAILFAST_MASK;
2580 rq->__sector = bio->bi_iter.bi_sector;
2581 rq->write_hint = bio->bi_write_hint;
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 rq_qos_throttle(q, bio);
2892 data.nr_tags = plug->nr_ios;
2894 data.cached_rq = &plug->cached_rq;
2897 rq = __blk_mq_alloc_requests(&data);
2900 rq_qos_cleanup(q, bio);
2901 if (bio->bi_opf & REQ_NOWAIT)
2902 bio_wouldblock_error(bio);
2907 * Check if there is a suitable cached request and return it.
2909 static struct request *blk_mq_peek_cached_request(struct blk_plug *plug,
2910 struct request_queue *q, blk_opf_t opf)
2912 enum hctx_type type = blk_mq_get_hctx_type(opf);
2917 rq = rq_list_peek(&plug->cached_rq);
2918 if (!rq || rq->q != q)
2920 if (type != rq->mq_hctx->type &&
2921 (type != HCTX_TYPE_READ || rq->mq_hctx->type != HCTX_TYPE_DEFAULT))
2923 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
2928 static void blk_mq_use_cached_rq(struct request *rq, struct blk_plug *plug,
2931 WARN_ON_ONCE(rq_list_peek(&plug->cached_rq) != rq);
2934 * If any qos ->throttle() end up blocking, we will have flushed the
2935 * plug and hence killed the cached_rq list as well. Pop this entry
2936 * before we throttle.
2938 plug->cached_rq = rq_list_next(rq);
2939 rq_qos_throttle(rq->q, bio);
2941 blk_mq_rq_time_init(rq, 0);
2942 rq->cmd_flags = bio->bi_opf;
2943 INIT_LIST_HEAD(&rq->queuelist);
2947 * blk_mq_submit_bio - Create and send a request to block device.
2948 * @bio: Bio pointer.
2950 * Builds up a request structure from @q and @bio and send to the device. The
2951 * request may not be queued directly to hardware if:
2952 * * This request can be merged with another one
2953 * * We want to place request at plug queue for possible future merging
2954 * * There is an IO scheduler active at this queue
2956 * It will not queue the request if there is an error with the bio, or at the
2959 void blk_mq_submit_bio(struct bio *bio)
2961 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2962 struct blk_plug *plug = blk_mq_plug(bio);
2963 const int is_sync = op_is_sync(bio->bi_opf);
2964 struct blk_mq_hw_ctx *hctx;
2965 unsigned int nr_segs = 1;
2969 bio = blk_queue_bounce(bio, q);
2972 * If the plug has a cached request for this queue, try use it.
2974 * The cached request already holds a q_usage_counter reference and we
2975 * don't have to acquire a new one if we use it.
2977 rq = blk_mq_peek_cached_request(plug, q, bio->bi_opf);
2979 if (unlikely(bio_queue_enter(bio)))
2983 if (unlikely(bio_may_exceed_limits(bio, &q->limits))) {
2984 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2988 if (!bio_integrity_prep(bio))
2991 if (blk_mq_attempt_bio_merge(q, bio, nr_segs))
2995 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2999 blk_mq_use_cached_rq(rq, plug, bio);
3002 trace_block_getrq(bio);
3004 rq_qos_track(q, rq, bio);
3006 blk_mq_bio_to_request(rq, bio, nr_segs);
3008 ret = blk_crypto_rq_get_keyslot(rq);
3009 if (ret != BLK_STS_OK) {
3010 bio->bi_status = ret;
3012 blk_mq_free_request(rq);
3016 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
3020 blk_add_rq_to_plug(plug, rq);
3025 if ((rq->rq_flags & RQF_USE_SCHED) ||
3026 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
3027 blk_mq_insert_request(rq, 0);
3028 blk_mq_run_hw_queue(hctx, true);
3030 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
3036 * Don't drop the queue reference if we were trying to use a cached
3037 * request and thus didn't acquire one.
3043 #ifdef CONFIG_BLK_MQ_STACKING
3045 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3046 * @rq: the request being queued
3048 blk_status_t blk_insert_cloned_request(struct request *rq)
3050 struct request_queue *q = rq->q;
3051 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3052 unsigned int max_segments = blk_rq_get_max_segments(rq);
3055 if (blk_rq_sectors(rq) > max_sectors) {
3057 * SCSI device does not have a good way to return if
3058 * Write Same/Zero is actually supported. If a device rejects
3059 * a non-read/write command (discard, write same,etc.) the
3060 * low-level device driver will set the relevant queue limit to
3061 * 0 to prevent blk-lib from issuing more of the offending
3062 * operations. Commands queued prior to the queue limit being
3063 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3064 * errors being propagated to upper layers.
3066 if (max_sectors == 0)
3067 return BLK_STS_NOTSUPP;
3069 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3070 __func__, blk_rq_sectors(rq), max_sectors);
3071 return BLK_STS_IOERR;
3075 * The queue settings related to segment counting may differ from the
3078 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3079 if (rq->nr_phys_segments > max_segments) {
3080 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3081 __func__, rq->nr_phys_segments, max_segments);
3082 return BLK_STS_IOERR;
3085 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3086 return BLK_STS_IOERR;
3088 ret = blk_crypto_rq_get_keyslot(rq);
3089 if (ret != BLK_STS_OK)
3092 blk_account_io_start(rq);
3095 * Since we have a scheduler attached on the top device,
3096 * bypass a potential scheduler on the bottom device for
3099 blk_mq_run_dispatch_ops(q,
3100 ret = blk_mq_request_issue_directly(rq, true));
3102 blk_account_io_done(rq, blk_time_get_ns());
3105 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3108 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3109 * @rq: the clone request to be cleaned up
3112 * Free all bios in @rq for a cloned request.
3114 void blk_rq_unprep_clone(struct request *rq)
3118 while ((bio = rq->bio) != NULL) {
3119 rq->bio = bio->bi_next;
3124 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3127 * blk_rq_prep_clone - Helper function to setup clone request
3128 * @rq: the request to be setup
3129 * @rq_src: original request to be cloned
3130 * @bs: bio_set that bios for clone are allocated from
3131 * @gfp_mask: memory allocation mask for bio
3132 * @bio_ctr: setup function to be called for each clone bio.
3133 * Returns %0 for success, non %0 for failure.
3134 * @data: private data to be passed to @bio_ctr
3137 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3138 * Also, pages which the original bios are pointing to are not copied
3139 * and the cloned bios just point same pages.
3140 * So cloned bios must be completed before original bios, which means
3141 * the caller must complete @rq before @rq_src.
3143 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3144 struct bio_set *bs, gfp_t gfp_mask,
3145 int (*bio_ctr)(struct bio *, struct bio *, void *),
3148 struct bio *bio, *bio_src;
3153 __rq_for_each_bio(bio_src, rq_src) {
3154 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3159 if (bio_ctr && bio_ctr(bio, bio_src, data))
3163 rq->biotail->bi_next = bio;
3166 rq->bio = rq->biotail = bio;
3171 /* Copy attributes of the original request to the clone request. */
3172 rq->__sector = blk_rq_pos(rq_src);
3173 rq->__data_len = blk_rq_bytes(rq_src);
3174 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3175 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3176 rq->special_vec = rq_src->special_vec;
3178 rq->nr_phys_segments = rq_src->nr_phys_segments;
3179 rq->ioprio = rq_src->ioprio;
3180 rq->write_hint = rq_src->write_hint;
3182 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3190 blk_rq_unprep_clone(rq);
3194 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3195 #endif /* CONFIG_BLK_MQ_STACKING */
3198 * Steal bios from a request and add them to a bio list.
3199 * The request must not have been partially completed before.
3201 void blk_steal_bios(struct bio_list *list, struct request *rq)
3205 list->tail->bi_next = rq->bio;
3207 list->head = rq->bio;
3208 list->tail = rq->biotail;
3216 EXPORT_SYMBOL_GPL(blk_steal_bios);
3218 static size_t order_to_size(unsigned int order)
3220 return (size_t)PAGE_SIZE << order;
3223 /* called before freeing request pool in @tags */
3224 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3225 struct blk_mq_tags *tags)
3228 unsigned long flags;
3231 * There is no need to clear mapping if driver tags is not initialized
3232 * or the mapping belongs to the driver tags.
3234 if (!drv_tags || drv_tags == tags)
3237 list_for_each_entry(page, &tags->page_list, lru) {
3238 unsigned long start = (unsigned long)page_address(page);
3239 unsigned long end = start + order_to_size(page->private);
3242 for (i = 0; i < drv_tags->nr_tags; i++) {
3243 struct request *rq = drv_tags->rqs[i];
3244 unsigned long rq_addr = (unsigned long)rq;
3246 if (rq_addr >= start && rq_addr < end) {
3247 WARN_ON_ONCE(req_ref_read(rq) != 0);
3248 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3254 * Wait until all pending iteration is done.
3256 * Request reference is cleared and it is guaranteed to be observed
3257 * after the ->lock is released.
3259 spin_lock_irqsave(&drv_tags->lock, flags);
3260 spin_unlock_irqrestore(&drv_tags->lock, flags);
3263 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3264 unsigned int hctx_idx)
3266 struct blk_mq_tags *drv_tags;
3269 if (list_empty(&tags->page_list))
3272 if (blk_mq_is_shared_tags(set->flags))
3273 drv_tags = set->shared_tags;
3275 drv_tags = set->tags[hctx_idx];
3277 if (tags->static_rqs && set->ops->exit_request) {
3280 for (i = 0; i < tags->nr_tags; i++) {
3281 struct request *rq = tags->static_rqs[i];
3285 set->ops->exit_request(set, rq, hctx_idx);
3286 tags->static_rqs[i] = NULL;
3290 blk_mq_clear_rq_mapping(drv_tags, tags);
3292 while (!list_empty(&tags->page_list)) {
3293 page = list_first_entry(&tags->page_list, struct page, lru);
3294 list_del_init(&page->lru);
3296 * Remove kmemleak object previously allocated in
3297 * blk_mq_alloc_rqs().
3299 kmemleak_free(page_address(page));
3300 __free_pages(page, page->private);
3304 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3308 kfree(tags->static_rqs);
3309 tags->static_rqs = NULL;
3311 blk_mq_free_tags(tags);
3314 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3315 unsigned int hctx_idx)
3319 for (i = 0; i < set->nr_maps; i++) {
3320 unsigned int start = set->map[i].queue_offset;
3321 unsigned int end = start + set->map[i].nr_queues;
3323 if (hctx_idx >= start && hctx_idx < end)
3327 if (i >= set->nr_maps)
3328 i = HCTX_TYPE_DEFAULT;
3333 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3334 unsigned int hctx_idx)
3336 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3338 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3341 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3342 unsigned int hctx_idx,
3343 unsigned int nr_tags,
3344 unsigned int reserved_tags)
3346 int node = blk_mq_get_hctx_node(set, hctx_idx);
3347 struct blk_mq_tags *tags;
3349 if (node == NUMA_NO_NODE)
3350 node = set->numa_node;
3352 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3353 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3357 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3358 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3363 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3364 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3366 if (!tags->static_rqs)
3374 blk_mq_free_tags(tags);
3378 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3379 unsigned int hctx_idx, int node)
3383 if (set->ops->init_request) {
3384 ret = set->ops->init_request(set, rq, hctx_idx, node);
3389 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3393 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3394 struct blk_mq_tags *tags,
3395 unsigned int hctx_idx, unsigned int depth)
3397 unsigned int i, j, entries_per_page, max_order = 4;
3398 int node = blk_mq_get_hctx_node(set, hctx_idx);
3399 size_t rq_size, left;
3401 if (node == NUMA_NO_NODE)
3402 node = set->numa_node;
3404 INIT_LIST_HEAD(&tags->page_list);
3407 * rq_size is the size of the request plus driver payload, rounded
3408 * to the cacheline size
3410 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3412 left = rq_size * depth;
3414 for (i = 0; i < depth; ) {
3415 int this_order = max_order;
3420 while (this_order && left < order_to_size(this_order - 1))
3424 page = alloc_pages_node(node,
3425 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3431 if (order_to_size(this_order) < rq_size)
3438 page->private = this_order;
3439 list_add_tail(&page->lru, &tags->page_list);
3441 p = page_address(page);
3443 * Allow kmemleak to scan these pages as they contain pointers
3444 * to additional allocations like via ops->init_request().
3446 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3447 entries_per_page = order_to_size(this_order) / rq_size;
3448 to_do = min(entries_per_page, depth - i);
3449 left -= to_do * rq_size;
3450 for (j = 0; j < to_do; j++) {
3451 struct request *rq = p;
3453 tags->static_rqs[i] = rq;
3454 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3455 tags->static_rqs[i] = NULL;
3466 blk_mq_free_rqs(set, tags, hctx_idx);
3470 struct rq_iter_data {
3471 struct blk_mq_hw_ctx *hctx;
3475 static bool blk_mq_has_request(struct request *rq, void *data)
3477 struct rq_iter_data *iter_data = data;
3479 if (rq->mq_hctx != iter_data->hctx)
3481 iter_data->has_rq = true;
3485 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3487 struct blk_mq_tags *tags = hctx->sched_tags ?
3488 hctx->sched_tags : hctx->tags;
3489 struct rq_iter_data data = {
3493 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3497 static bool blk_mq_hctx_has_online_cpu(struct blk_mq_hw_ctx *hctx,
3498 unsigned int this_cpu)
3500 enum hctx_type type = hctx->type;
3504 * hctx->cpumask has to rule out isolated CPUs, but userspace still
3505 * might submit IOs on these isolated CPUs, so use the queue map to
3506 * check if all CPUs mapped to this hctx are offline
3508 for_each_online_cpu(cpu) {
3509 struct blk_mq_hw_ctx *h = blk_mq_map_queue_type(hctx->queue,
3515 /* this hctx has at least one online CPU */
3516 if (this_cpu != cpu)
3523 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3525 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3526 struct blk_mq_hw_ctx, cpuhp_online);
3528 if (blk_mq_hctx_has_online_cpu(hctx, cpu))
3532 * Prevent new request from being allocated on the current hctx.
3534 * The smp_mb__after_atomic() Pairs with the implied barrier in
3535 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3536 * seen once we return from the tag allocator.
3538 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3539 smp_mb__after_atomic();
3542 * Try to grab a reference to the queue and wait for any outstanding
3543 * requests. If we could not grab a reference the queue has been
3544 * frozen and there are no requests.
3546 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3547 while (blk_mq_hctx_has_requests(hctx))
3549 percpu_ref_put(&hctx->queue->q_usage_counter);
3555 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3557 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3558 struct blk_mq_hw_ctx, cpuhp_online);
3560 if (cpumask_test_cpu(cpu, hctx->cpumask))
3561 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3566 * 'cpu' is going away. splice any existing rq_list entries from this
3567 * software queue to the hw queue dispatch list, and ensure that it
3570 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3572 struct blk_mq_hw_ctx *hctx;
3573 struct blk_mq_ctx *ctx;
3575 enum hctx_type type;
3577 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3578 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3581 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3584 spin_lock(&ctx->lock);
3585 if (!list_empty(&ctx->rq_lists[type])) {
3586 list_splice_init(&ctx->rq_lists[type], &tmp);
3587 blk_mq_hctx_clear_pending(hctx, ctx);
3589 spin_unlock(&ctx->lock);
3591 if (list_empty(&tmp))
3594 spin_lock(&hctx->lock);
3595 list_splice_tail_init(&tmp, &hctx->dispatch);
3596 spin_unlock(&hctx->lock);
3598 blk_mq_run_hw_queue(hctx, true);
3602 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3604 if (!(hctx->flags & BLK_MQ_F_STACKING))
3605 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3606 &hctx->cpuhp_online);
3607 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3612 * Before freeing hw queue, clearing the flush request reference in
3613 * tags->rqs[] for avoiding potential UAF.
3615 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3616 unsigned int queue_depth, struct request *flush_rq)
3619 unsigned long flags;
3621 /* The hw queue may not be mapped yet */
3625 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3627 for (i = 0; i < queue_depth; i++)
3628 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3631 * Wait until all pending iteration is done.
3633 * Request reference is cleared and it is guaranteed to be observed
3634 * after the ->lock is released.
3636 spin_lock_irqsave(&tags->lock, flags);
3637 spin_unlock_irqrestore(&tags->lock, flags);
3640 /* hctx->ctxs will be freed in queue's release handler */
3641 static void blk_mq_exit_hctx(struct request_queue *q,
3642 struct blk_mq_tag_set *set,
3643 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3645 struct request *flush_rq = hctx->fq->flush_rq;
3647 if (blk_mq_hw_queue_mapped(hctx))
3648 blk_mq_tag_idle(hctx);
3650 if (blk_queue_init_done(q))
3651 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3652 set->queue_depth, flush_rq);
3653 if (set->ops->exit_request)
3654 set->ops->exit_request(set, flush_rq, hctx_idx);
3656 if (set->ops->exit_hctx)
3657 set->ops->exit_hctx(hctx, hctx_idx);
3659 blk_mq_remove_cpuhp(hctx);
3661 xa_erase(&q->hctx_table, hctx_idx);
3663 spin_lock(&q->unused_hctx_lock);
3664 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3665 spin_unlock(&q->unused_hctx_lock);
3668 static void blk_mq_exit_hw_queues(struct request_queue *q,
3669 struct blk_mq_tag_set *set, int nr_queue)
3671 struct blk_mq_hw_ctx *hctx;
3674 queue_for_each_hw_ctx(q, hctx, i) {
3677 blk_mq_exit_hctx(q, set, hctx, i);
3681 static int blk_mq_init_hctx(struct request_queue *q,
3682 struct blk_mq_tag_set *set,
3683 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3685 hctx->queue_num = hctx_idx;
3687 if (!(hctx->flags & BLK_MQ_F_STACKING))
3688 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3689 &hctx->cpuhp_online);
3690 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3692 hctx->tags = set->tags[hctx_idx];
3694 if (set->ops->init_hctx &&
3695 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3696 goto unregister_cpu_notifier;
3698 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3702 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3708 if (set->ops->exit_request)
3709 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3711 if (set->ops->exit_hctx)
3712 set->ops->exit_hctx(hctx, hctx_idx);
3713 unregister_cpu_notifier:
3714 blk_mq_remove_cpuhp(hctx);
3718 static struct blk_mq_hw_ctx *
3719 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3722 struct blk_mq_hw_ctx *hctx;
3723 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3725 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3727 goto fail_alloc_hctx;
3729 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3732 atomic_set(&hctx->nr_active, 0);
3733 if (node == NUMA_NO_NODE)
3734 node = set->numa_node;
3735 hctx->numa_node = node;
3737 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3738 spin_lock_init(&hctx->lock);
3739 INIT_LIST_HEAD(&hctx->dispatch);
3741 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3743 INIT_LIST_HEAD(&hctx->hctx_list);
3746 * Allocate space for all possible cpus to avoid allocation at
3749 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3754 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3755 gfp, node, false, false))
3759 spin_lock_init(&hctx->dispatch_wait_lock);
3760 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3761 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3763 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3767 blk_mq_hctx_kobj_init(hctx);
3772 sbitmap_free(&hctx->ctx_map);
3776 free_cpumask_var(hctx->cpumask);
3783 static void blk_mq_init_cpu_queues(struct request_queue *q,
3784 unsigned int nr_hw_queues)
3786 struct blk_mq_tag_set *set = q->tag_set;
3789 for_each_possible_cpu(i) {
3790 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3791 struct blk_mq_hw_ctx *hctx;
3795 spin_lock_init(&__ctx->lock);
3796 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3797 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3802 * Set local node, IFF we have more than one hw queue. If
3803 * not, we remain on the home node of the device
3805 for (j = 0; j < set->nr_maps; j++) {
3806 hctx = blk_mq_map_queue_type(q, j, i);
3807 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3808 hctx->numa_node = cpu_to_node(i);
3813 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3814 unsigned int hctx_idx,
3817 struct blk_mq_tags *tags;
3820 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3824 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3826 blk_mq_free_rq_map(tags);
3833 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3836 if (blk_mq_is_shared_tags(set->flags)) {
3837 set->tags[hctx_idx] = set->shared_tags;
3842 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3845 return set->tags[hctx_idx];
3848 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3849 struct blk_mq_tags *tags,
3850 unsigned int hctx_idx)
3853 blk_mq_free_rqs(set, tags, hctx_idx);
3854 blk_mq_free_rq_map(tags);
3858 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3859 unsigned int hctx_idx)
3861 if (!blk_mq_is_shared_tags(set->flags))
3862 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3864 set->tags[hctx_idx] = NULL;
3867 static void blk_mq_map_swqueue(struct request_queue *q)
3869 unsigned int j, hctx_idx;
3871 struct blk_mq_hw_ctx *hctx;
3872 struct blk_mq_ctx *ctx;
3873 struct blk_mq_tag_set *set = q->tag_set;
3875 queue_for_each_hw_ctx(q, hctx, i) {
3876 cpumask_clear(hctx->cpumask);
3878 hctx->dispatch_from = NULL;
3882 * Map software to hardware queues.
3884 * If the cpu isn't present, the cpu is mapped to first hctx.
3886 for_each_possible_cpu(i) {
3888 ctx = per_cpu_ptr(q->queue_ctx, i);
3889 for (j = 0; j < set->nr_maps; j++) {
3890 if (!set->map[j].nr_queues) {
3891 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3892 HCTX_TYPE_DEFAULT, i);
3895 hctx_idx = set->map[j].mq_map[i];
3896 /* unmapped hw queue can be remapped after CPU topo changed */
3897 if (!set->tags[hctx_idx] &&
3898 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3900 * If tags initialization fail for some hctx,
3901 * that hctx won't be brought online. In this
3902 * case, remap the current ctx to hctx[0] which
3903 * is guaranteed to always have tags allocated
3905 set->map[j].mq_map[i] = 0;
3908 hctx = blk_mq_map_queue_type(q, j, i);
3909 ctx->hctxs[j] = hctx;
3911 * If the CPU is already set in the mask, then we've
3912 * mapped this one already. This can happen if
3913 * devices share queues across queue maps.
3915 if (cpumask_test_cpu(i, hctx->cpumask))
3918 cpumask_set_cpu(i, hctx->cpumask);
3920 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3921 hctx->ctxs[hctx->nr_ctx++] = ctx;
3924 * If the nr_ctx type overflows, we have exceeded the
3925 * amount of sw queues we can support.
3927 BUG_ON(!hctx->nr_ctx);
3930 for (; j < HCTX_MAX_TYPES; j++)
3931 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3932 HCTX_TYPE_DEFAULT, i);
3935 queue_for_each_hw_ctx(q, hctx, i) {
3939 * If no software queues are mapped to this hardware queue,
3940 * disable it and free the request entries.
3942 if (!hctx->nr_ctx) {
3943 /* Never unmap queue 0. We need it as a
3944 * fallback in case of a new remap fails
3948 __blk_mq_free_map_and_rqs(set, i);
3954 hctx->tags = set->tags[i];
3955 WARN_ON(!hctx->tags);
3958 * Set the map size to the number of mapped software queues.
3959 * This is more accurate and more efficient than looping
3960 * over all possibly mapped software queues.
3962 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3965 * Rule out isolated CPUs from hctx->cpumask to avoid
3966 * running block kworker on isolated CPUs
3968 for_each_cpu(cpu, hctx->cpumask) {
3969 if (cpu_is_isolated(cpu))
3970 cpumask_clear_cpu(cpu, hctx->cpumask);
3974 * Initialize batch roundrobin counts
3976 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3977 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3982 * Caller needs to ensure that we're either frozen/quiesced, or that
3983 * the queue isn't live yet.
3985 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3987 struct blk_mq_hw_ctx *hctx;
3990 queue_for_each_hw_ctx(q, hctx, i) {
3992 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3994 blk_mq_tag_idle(hctx);
3995 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
4000 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
4003 struct request_queue *q;
4005 lockdep_assert_held(&set->tag_list_lock);
4007 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4008 blk_mq_freeze_queue(q);
4009 queue_set_hctx_shared(q, shared);
4010 blk_mq_unfreeze_queue(q);
4014 static void blk_mq_del_queue_tag_set(struct request_queue *q)
4016 struct blk_mq_tag_set *set = q->tag_set;
4018 mutex_lock(&set->tag_list_lock);
4019 list_del(&q->tag_set_list);
4020 if (list_is_singular(&set->tag_list)) {
4021 /* just transitioned to unshared */
4022 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
4023 /* update existing queue */
4024 blk_mq_update_tag_set_shared(set, false);
4026 mutex_unlock(&set->tag_list_lock);
4027 INIT_LIST_HEAD(&q->tag_set_list);
4030 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
4031 struct request_queue *q)
4033 mutex_lock(&set->tag_list_lock);
4036 * Check to see if we're transitioning to shared (from 1 to 2 queues).
4038 if (!list_empty(&set->tag_list) &&
4039 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
4040 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
4041 /* update existing queue */
4042 blk_mq_update_tag_set_shared(set, true);
4044 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
4045 queue_set_hctx_shared(q, true);
4046 list_add_tail(&q->tag_set_list, &set->tag_list);
4048 mutex_unlock(&set->tag_list_lock);
4051 /* All allocations will be freed in release handler of q->mq_kobj */
4052 static int blk_mq_alloc_ctxs(struct request_queue *q)
4054 struct blk_mq_ctxs *ctxs;
4057 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
4061 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
4062 if (!ctxs->queue_ctx)
4065 for_each_possible_cpu(cpu) {
4066 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4070 q->mq_kobj = &ctxs->kobj;
4071 q->queue_ctx = ctxs->queue_ctx;
4080 * It is the actual release handler for mq, but we do it from
4081 * request queue's release handler for avoiding use-after-free
4082 * and headache because q->mq_kobj shouldn't have been introduced,
4083 * but we can't group ctx/kctx kobj without it.
4085 void blk_mq_release(struct request_queue *q)
4087 struct blk_mq_hw_ctx *hctx, *next;
4090 queue_for_each_hw_ctx(q, hctx, i)
4091 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4093 /* all hctx are in .unused_hctx_list now */
4094 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4095 list_del_init(&hctx->hctx_list);
4096 kobject_put(&hctx->kobj);
4099 xa_destroy(&q->hctx_table);
4102 * release .mq_kobj and sw queue's kobject now because
4103 * both share lifetime with request queue.
4105 blk_mq_sysfs_deinit(q);
4108 struct request_queue *blk_mq_alloc_queue(struct blk_mq_tag_set *set,
4109 struct queue_limits *lim, void *queuedata)
4111 struct queue_limits default_lim = { };
4112 struct request_queue *q;
4115 q = blk_alloc_queue(lim ? lim : &default_lim, set->numa_node);
4118 q->queuedata = queuedata;
4119 ret = blk_mq_init_allocated_queue(set, q);
4122 return ERR_PTR(ret);
4126 EXPORT_SYMBOL(blk_mq_alloc_queue);
4129 * blk_mq_destroy_queue - shutdown a request queue
4130 * @q: request queue to shutdown
4132 * This shuts down a request queue allocated by blk_mq_alloc_queue(). All future
4133 * requests will be failed with -ENODEV. The caller is responsible for dropping
4134 * the reference from blk_mq_alloc_queue() by calling blk_put_queue().
4136 * Context: can sleep
4138 void blk_mq_destroy_queue(struct request_queue *q)
4140 WARN_ON_ONCE(!queue_is_mq(q));
4141 WARN_ON_ONCE(blk_queue_registered(q));
4145 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4146 blk_queue_start_drain(q);
4147 blk_mq_freeze_queue_wait(q);
4150 blk_mq_cancel_work_sync(q);
4151 blk_mq_exit_queue(q);
4153 EXPORT_SYMBOL(blk_mq_destroy_queue);
4155 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set,
4156 struct queue_limits *lim, void *queuedata,
4157 struct lock_class_key *lkclass)
4159 struct request_queue *q;
4160 struct gendisk *disk;
4162 q = blk_mq_alloc_queue(set, lim, queuedata);
4166 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4168 blk_mq_destroy_queue(q);
4170 return ERR_PTR(-ENOMEM);
4172 set_bit(GD_OWNS_QUEUE, &disk->state);
4175 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4177 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4178 struct lock_class_key *lkclass)
4180 struct gendisk *disk;
4182 if (!blk_get_queue(q))
4184 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4189 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4191 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4192 struct blk_mq_tag_set *set, struct request_queue *q,
4193 int hctx_idx, int node)
4195 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4197 /* reuse dead hctx first */
4198 spin_lock(&q->unused_hctx_lock);
4199 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4200 if (tmp->numa_node == node) {
4206 list_del_init(&hctx->hctx_list);
4207 spin_unlock(&q->unused_hctx_lock);
4210 hctx = blk_mq_alloc_hctx(q, set, node);
4214 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4220 kobject_put(&hctx->kobj);
4225 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4226 struct request_queue *q)
4228 struct blk_mq_hw_ctx *hctx;
4231 /* protect against switching io scheduler */
4232 mutex_lock(&q->sysfs_lock);
4233 for (i = 0; i < set->nr_hw_queues; i++) {
4235 int node = blk_mq_get_hctx_node(set, i);
4236 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4239 old_node = old_hctx->numa_node;
4240 blk_mq_exit_hctx(q, set, old_hctx, i);
4243 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4246 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4248 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4249 WARN_ON_ONCE(!hctx);
4253 * Increasing nr_hw_queues fails. Free the newly allocated
4254 * hctxs and keep the previous q->nr_hw_queues.
4256 if (i != set->nr_hw_queues) {
4257 j = q->nr_hw_queues;
4260 q->nr_hw_queues = set->nr_hw_queues;
4263 xa_for_each_start(&q->hctx_table, j, hctx, j)
4264 blk_mq_exit_hctx(q, set, hctx, j);
4265 mutex_unlock(&q->sysfs_lock);
4268 static void blk_mq_update_poll_flag(struct request_queue *q)
4270 struct blk_mq_tag_set *set = q->tag_set;
4272 if (set->nr_maps > HCTX_TYPE_POLL &&
4273 set->map[HCTX_TYPE_POLL].nr_queues)
4274 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4276 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4279 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4280 struct request_queue *q)
4282 /* mark the queue as mq asap */
4283 q->mq_ops = set->ops;
4285 if (blk_mq_alloc_ctxs(q))
4288 /* init q->mq_kobj and sw queues' kobjects */
4289 blk_mq_sysfs_init(q);
4291 INIT_LIST_HEAD(&q->unused_hctx_list);
4292 spin_lock_init(&q->unused_hctx_lock);
4294 xa_init(&q->hctx_table);
4296 blk_mq_realloc_hw_ctxs(set, q);
4297 if (!q->nr_hw_queues)
4300 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4301 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4305 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4306 blk_mq_update_poll_flag(q);
4308 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4309 INIT_LIST_HEAD(&q->flush_list);
4310 INIT_LIST_HEAD(&q->requeue_list);
4311 spin_lock_init(&q->requeue_lock);
4313 q->nr_requests = set->queue_depth;
4315 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4316 blk_mq_add_queue_tag_set(set, q);
4317 blk_mq_map_swqueue(q);
4326 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4328 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4329 void blk_mq_exit_queue(struct request_queue *q)
4331 struct blk_mq_tag_set *set = q->tag_set;
4333 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4334 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4335 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4336 blk_mq_del_queue_tag_set(q);
4339 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4343 if (blk_mq_is_shared_tags(set->flags)) {
4344 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4347 if (!set->shared_tags)
4351 for (i = 0; i < set->nr_hw_queues; i++) {
4352 if (!__blk_mq_alloc_map_and_rqs(set, i))
4361 __blk_mq_free_map_and_rqs(set, i);
4363 if (blk_mq_is_shared_tags(set->flags)) {
4364 blk_mq_free_map_and_rqs(set, set->shared_tags,
4365 BLK_MQ_NO_HCTX_IDX);
4372 * Allocate the request maps associated with this tag_set. Note that this
4373 * may reduce the depth asked for, if memory is tight. set->queue_depth
4374 * will be updated to reflect the allocated depth.
4376 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4381 depth = set->queue_depth;
4383 err = __blk_mq_alloc_rq_maps(set);
4387 set->queue_depth >>= 1;
4388 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4392 } while (set->queue_depth);
4394 if (!set->queue_depth || err) {
4395 pr_err("blk-mq: failed to allocate request map\n");
4399 if (depth != set->queue_depth)
4400 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4401 depth, set->queue_depth);
4406 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4409 * blk_mq_map_queues() and multiple .map_queues() implementations
4410 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4411 * number of hardware queues.
4413 if (set->nr_maps == 1)
4414 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4416 if (set->ops->map_queues) {
4420 * transport .map_queues is usually done in the following
4423 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4424 * mask = get_cpu_mask(queue)
4425 * for_each_cpu(cpu, mask)
4426 * set->map[x].mq_map[cpu] = queue;
4429 * When we need to remap, the table has to be cleared for
4430 * killing stale mapping since one CPU may not be mapped
4433 for (i = 0; i < set->nr_maps; i++)
4434 blk_mq_clear_mq_map(&set->map[i]);
4436 set->ops->map_queues(set);
4438 BUG_ON(set->nr_maps > 1);
4439 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4443 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4444 int new_nr_hw_queues)
4446 struct blk_mq_tags **new_tags;
4449 if (set->nr_hw_queues >= new_nr_hw_queues)
4452 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4453 GFP_KERNEL, set->numa_node);
4458 memcpy(new_tags, set->tags, set->nr_hw_queues *
4459 sizeof(*set->tags));
4461 set->tags = new_tags;
4463 for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) {
4464 if (!__blk_mq_alloc_map_and_rqs(set, i)) {
4465 while (--i >= set->nr_hw_queues)
4466 __blk_mq_free_map_and_rqs(set, i);
4473 set->nr_hw_queues = new_nr_hw_queues;
4478 * Alloc a tag set to be associated with one or more request queues.
4479 * May fail with EINVAL for various error conditions. May adjust the
4480 * requested depth down, if it's too large. In that case, the set
4481 * value will be stored in set->queue_depth.
4483 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4487 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4489 if (!set->nr_hw_queues)
4491 if (!set->queue_depth)
4493 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4496 if (!set->ops->queue_rq)
4499 if (!set->ops->get_budget ^ !set->ops->put_budget)
4502 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4503 pr_info("blk-mq: reduced tag depth to %u\n",
4505 set->queue_depth = BLK_MQ_MAX_DEPTH;
4510 else if (set->nr_maps > HCTX_MAX_TYPES)
4514 * If a crashdump is active, then we are potentially in a very
4515 * memory constrained environment. Limit us to 64 tags to prevent
4516 * using too much memory.
4518 if (is_kdump_kernel())
4519 set->queue_depth = min(64U, set->queue_depth);
4522 * There is no use for more h/w queues than cpus if we just have
4525 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4526 set->nr_hw_queues = nr_cpu_ids;
4528 if (set->flags & BLK_MQ_F_BLOCKING) {
4529 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4532 ret = init_srcu_struct(set->srcu);
4538 set->tags = kcalloc_node(set->nr_hw_queues,
4539 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4542 goto out_cleanup_srcu;
4544 for (i = 0; i < set->nr_maps; i++) {
4545 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4546 sizeof(set->map[i].mq_map[0]),
4547 GFP_KERNEL, set->numa_node);
4548 if (!set->map[i].mq_map)
4549 goto out_free_mq_map;
4550 set->map[i].nr_queues = set->nr_hw_queues;
4553 blk_mq_update_queue_map(set);
4555 ret = blk_mq_alloc_set_map_and_rqs(set);
4557 goto out_free_mq_map;
4559 mutex_init(&set->tag_list_lock);
4560 INIT_LIST_HEAD(&set->tag_list);
4565 for (i = 0; i < set->nr_maps; i++) {
4566 kfree(set->map[i].mq_map);
4567 set->map[i].mq_map = NULL;
4572 if (set->flags & BLK_MQ_F_BLOCKING)
4573 cleanup_srcu_struct(set->srcu);
4575 if (set->flags & BLK_MQ_F_BLOCKING)
4579 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4581 /* allocate and initialize a tagset for a simple single-queue device */
4582 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4583 const struct blk_mq_ops *ops, unsigned int queue_depth,
4584 unsigned int set_flags)
4586 memset(set, 0, sizeof(*set));
4588 set->nr_hw_queues = 1;
4590 set->queue_depth = queue_depth;
4591 set->numa_node = NUMA_NO_NODE;
4592 set->flags = set_flags;
4593 return blk_mq_alloc_tag_set(set);
4595 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4597 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4601 for (i = 0; i < set->nr_hw_queues; i++)
4602 __blk_mq_free_map_and_rqs(set, i);
4604 if (blk_mq_is_shared_tags(set->flags)) {
4605 blk_mq_free_map_and_rqs(set, set->shared_tags,
4606 BLK_MQ_NO_HCTX_IDX);
4609 for (j = 0; j < set->nr_maps; j++) {
4610 kfree(set->map[j].mq_map);
4611 set->map[j].mq_map = NULL;
4616 if (set->flags & BLK_MQ_F_BLOCKING) {
4617 cleanup_srcu_struct(set->srcu);
4621 EXPORT_SYMBOL(blk_mq_free_tag_set);
4623 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4625 struct blk_mq_tag_set *set = q->tag_set;
4626 struct blk_mq_hw_ctx *hctx;
4633 if (q->nr_requests == nr)
4636 blk_mq_freeze_queue(q);
4637 blk_mq_quiesce_queue(q);
4640 queue_for_each_hw_ctx(q, hctx, i) {
4644 * If we're using an MQ scheduler, just update the scheduler
4645 * queue depth. This is similar to what the old code would do.
4647 if (hctx->sched_tags) {
4648 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4651 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4656 if (q->elevator && q->elevator->type->ops.depth_updated)
4657 q->elevator->type->ops.depth_updated(hctx);
4660 q->nr_requests = nr;
4661 if (blk_mq_is_shared_tags(set->flags)) {
4663 blk_mq_tag_update_sched_shared_tags(q);
4665 blk_mq_tag_resize_shared_tags(set, nr);
4669 blk_mq_unquiesce_queue(q);
4670 blk_mq_unfreeze_queue(q);
4676 * request_queue and elevator_type pair.
4677 * It is just used by __blk_mq_update_nr_hw_queues to cache
4678 * the elevator_type associated with a request_queue.
4680 struct blk_mq_qe_pair {
4681 struct list_head node;
4682 struct request_queue *q;
4683 struct elevator_type *type;
4687 * Cache the elevator_type in qe pair list and switch the
4688 * io scheduler to 'none'
4690 static bool blk_mq_elv_switch_none(struct list_head *head,
4691 struct request_queue *q)
4693 struct blk_mq_qe_pair *qe;
4695 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4699 /* q->elevator needs protection from ->sysfs_lock */
4700 mutex_lock(&q->sysfs_lock);
4702 /* the check has to be done with holding sysfs_lock */
4708 INIT_LIST_HEAD(&qe->node);
4710 qe->type = q->elevator->type;
4711 /* keep a reference to the elevator module as we'll switch back */
4712 __elevator_get(qe->type);
4713 list_add(&qe->node, head);
4714 elevator_disable(q);
4716 mutex_unlock(&q->sysfs_lock);
4721 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4722 struct request_queue *q)
4724 struct blk_mq_qe_pair *qe;
4726 list_for_each_entry(qe, head, node)
4733 static void blk_mq_elv_switch_back(struct list_head *head,
4734 struct request_queue *q)
4736 struct blk_mq_qe_pair *qe;
4737 struct elevator_type *t;
4739 qe = blk_lookup_qe_pair(head, q);
4743 list_del(&qe->node);
4746 mutex_lock(&q->sysfs_lock);
4747 elevator_switch(q, t);
4748 /* drop the reference acquired in blk_mq_elv_switch_none */
4750 mutex_unlock(&q->sysfs_lock);
4753 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4756 struct request_queue *q;
4758 int prev_nr_hw_queues = set->nr_hw_queues;
4761 lockdep_assert_held(&set->tag_list_lock);
4763 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4764 nr_hw_queues = nr_cpu_ids;
4765 if (nr_hw_queues < 1)
4767 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4770 list_for_each_entry(q, &set->tag_list, tag_set_list)
4771 blk_mq_freeze_queue(q);
4773 * Switch IO scheduler to 'none', cleaning up the data associated
4774 * with the previous scheduler. We will switch back once we are done
4775 * updating the new sw to hw queue mappings.
4777 list_for_each_entry(q, &set->tag_list, tag_set_list)
4778 if (!blk_mq_elv_switch_none(&head, q))
4781 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4782 blk_mq_debugfs_unregister_hctxs(q);
4783 blk_mq_sysfs_unregister_hctxs(q);
4786 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4790 blk_mq_update_queue_map(set);
4791 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4792 blk_mq_realloc_hw_ctxs(set, q);
4793 blk_mq_update_poll_flag(q);
4794 if (q->nr_hw_queues != set->nr_hw_queues) {
4795 int i = prev_nr_hw_queues;
4797 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4798 nr_hw_queues, prev_nr_hw_queues);
4799 for (; i < set->nr_hw_queues; i++)
4800 __blk_mq_free_map_and_rqs(set, i);
4802 set->nr_hw_queues = prev_nr_hw_queues;
4805 blk_mq_map_swqueue(q);
4809 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4810 blk_mq_sysfs_register_hctxs(q);
4811 blk_mq_debugfs_register_hctxs(q);
4815 list_for_each_entry(q, &set->tag_list, tag_set_list)
4816 blk_mq_elv_switch_back(&head, q);
4818 list_for_each_entry(q, &set->tag_list, tag_set_list)
4819 blk_mq_unfreeze_queue(q);
4821 /* Free the excess tags when nr_hw_queues shrink. */
4822 for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)
4823 __blk_mq_free_map_and_rqs(set, i);
4826 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4828 mutex_lock(&set->tag_list_lock);
4829 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4830 mutex_unlock(&set->tag_list_lock);
4832 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4834 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
4835 struct io_comp_batch *iob, unsigned int flags)
4837 long state = get_current_state();
4841 ret = q->mq_ops->poll(hctx, iob);
4843 __set_current_state(TASK_RUNNING);
4847 if (signal_pending_state(state, current))
4848 __set_current_state(TASK_RUNNING);
4849 if (task_is_running(current))
4852 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4855 } while (!need_resched());
4857 __set_current_state(TASK_RUNNING);
4861 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
4862 struct io_comp_batch *iob, unsigned int flags)
4864 struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie);
4866 return blk_hctx_poll(q, hctx, iob, flags);
4869 int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
4870 unsigned int poll_flags)
4872 struct request_queue *q = rq->q;
4875 if (!blk_rq_is_poll(rq))
4877 if (!percpu_ref_tryget(&q->q_usage_counter))
4880 ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags);
4885 EXPORT_SYMBOL_GPL(blk_rq_poll);
4887 unsigned int blk_mq_rq_cpu(struct request *rq)
4889 return rq->mq_ctx->cpu;
4891 EXPORT_SYMBOL(blk_mq_rq_cpu);
4893 void blk_mq_cancel_work_sync(struct request_queue *q)
4895 struct blk_mq_hw_ctx *hctx;
4898 cancel_delayed_work_sync(&q->requeue_work);
4900 queue_for_each_hw_ctx(q, hctx, i)
4901 cancel_delayed_work_sync(&hctx->run_work);
4904 static int __init blk_mq_init(void)
4908 for_each_possible_cpu(i)
4909 init_llist_head(&per_cpu(blk_cpu_done, i));
4910 for_each_possible_cpu(i)
4911 INIT_CSD(&per_cpu(blk_cpu_csd, i),
4912 __blk_mq_complete_request_remote, NULL);
4913 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4915 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4916 "block/softirq:dead", NULL,
4917 blk_softirq_cpu_dead);
4918 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4919 blk_mq_hctx_notify_dead);
4920 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4921 blk_mq_hctx_notify_online,
4922 blk_mq_hctx_notify_offline);
4925 subsys_initcall(blk_mq_init);