1 // SPDX-License-Identifier: GPL-2.0
3 * Block multiqueue core code
5 * Copyright (C) 2013-2014 Jens Axboe
6 * Copyright (C) 2013-2014 Christoph Hellwig
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/backing-dev.h>
11 #include <linux/bio.h>
12 #include <linux/blkdev.h>
13 #include <linux/blk-integrity.h>
14 #include <linux/kmemleak.h>
16 #include <linux/init.h>
17 #include <linux/slab.h>
18 #include <linux/workqueue.h>
19 #include <linux/smp.h>
20 #include <linux/interrupt.h>
21 #include <linux/llist.h>
22 #include <linux/cpu.h>
23 #include <linux/cache.h>
24 #include <linux/sched/sysctl.h>
25 #include <linux/sched/topology.h>
26 #include <linux/sched/signal.h>
27 #include <linux/delay.h>
28 #include <linux/crash_dump.h>
29 #include <linux/prefetch.h>
30 #include <linux/blk-crypto.h>
31 #include <linux/part_stat.h>
33 #include <trace/events/block.h>
35 #include <linux/t10-pi.h>
38 #include "blk-mq-debugfs.h"
41 #include "blk-mq-sched.h"
42 #include "blk-rq-qos.h"
43 #include "blk-ioprio.h"
45 static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
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);
53 static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q,
56 return xa_load(&q->hctx_table, qc);
59 static inline blk_qc_t blk_rq_to_qc(struct request *rq)
61 return rq->mq_hctx->queue_num;
65 * Check if any of the ctx, dispatch list or elevator
66 * have pending work in this hardware queue.
68 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
70 return !list_empty_careful(&hctx->dispatch) ||
71 sbitmap_any_bit_set(&hctx->ctx_map) ||
72 blk_mq_sched_has_work(hctx);
76 * Mark this ctx as having pending work in this hardware queue
78 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
79 struct blk_mq_ctx *ctx)
81 const int bit = ctx->index_hw[hctx->type];
83 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
84 sbitmap_set_bit(&hctx->ctx_map, bit);
87 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
88 struct blk_mq_ctx *ctx)
90 const int bit = ctx->index_hw[hctx->type];
92 sbitmap_clear_bit(&hctx->ctx_map, bit);
96 struct block_device *part;
97 unsigned int inflight[2];
100 static bool blk_mq_check_inflight(struct request *rq, void *priv)
102 struct mq_inflight *mi = priv;
104 if (rq->part && blk_do_io_stat(rq) &&
105 (!mi->part->bd_partno || rq->part == mi->part) &&
106 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
107 mi->inflight[rq_data_dir(rq)]++;
112 unsigned int blk_mq_in_flight(struct request_queue *q,
113 struct block_device *part)
115 struct mq_inflight mi = { .part = part };
117 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
119 return mi.inflight[0] + mi.inflight[1];
122 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
123 unsigned int inflight[2])
125 struct mq_inflight mi = { .part = part };
127 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
128 inflight[0] = mi.inflight[0];
129 inflight[1] = mi.inflight[1];
132 void blk_freeze_queue_start(struct request_queue *q)
134 mutex_lock(&q->mq_freeze_lock);
135 if (++q->mq_freeze_depth == 1) {
136 percpu_ref_kill(&q->q_usage_counter);
137 mutex_unlock(&q->mq_freeze_lock);
139 blk_mq_run_hw_queues(q, false);
141 mutex_unlock(&q->mq_freeze_lock);
144 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
146 void blk_mq_freeze_queue_wait(struct request_queue *q)
148 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
150 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
152 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
153 unsigned long timeout)
155 return wait_event_timeout(q->mq_freeze_wq,
156 percpu_ref_is_zero(&q->q_usage_counter),
159 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
162 * Guarantee no request is in use, so we can change any data structure of
163 * the queue afterward.
165 void blk_freeze_queue(struct request_queue *q)
168 * In the !blk_mq case we are only calling this to kill the
169 * q_usage_counter, otherwise this increases the freeze depth
170 * and waits for it to return to zero. For this reason there is
171 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
172 * exported to drivers as the only user for unfreeze is blk_mq.
174 blk_freeze_queue_start(q);
175 blk_mq_freeze_queue_wait(q);
178 void blk_mq_freeze_queue(struct request_queue *q)
181 * ...just an alias to keep freeze and unfreeze actions balanced
182 * in the blk_mq_* namespace
186 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
188 void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
190 mutex_lock(&q->mq_freeze_lock);
192 q->q_usage_counter.data->force_atomic = true;
193 q->mq_freeze_depth--;
194 WARN_ON_ONCE(q->mq_freeze_depth < 0);
195 if (!q->mq_freeze_depth) {
196 percpu_ref_resurrect(&q->q_usage_counter);
197 wake_up_all(&q->mq_freeze_wq);
199 mutex_unlock(&q->mq_freeze_lock);
202 void blk_mq_unfreeze_queue(struct request_queue *q)
204 __blk_mq_unfreeze_queue(q, false);
206 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
209 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
210 * mpt3sas driver such that this function can be removed.
212 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
216 spin_lock_irqsave(&q->queue_lock, flags);
217 if (!q->quiesce_depth++)
218 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
219 spin_unlock_irqrestore(&q->queue_lock, flags);
221 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
224 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
225 * @set: tag_set to wait on
227 * Note: it is driver's responsibility for making sure that quiesce has
228 * been started on or more of the request_queues of the tag_set. This
229 * function only waits for the quiesce on those request_queues that had
230 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
232 void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
234 if (set->flags & BLK_MQ_F_BLOCKING)
235 synchronize_srcu(set->srcu);
239 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
242 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
245 * Note: this function does not prevent that the struct request end_io()
246 * callback function is invoked. Once this function is returned, we make
247 * sure no dispatch can happen until the queue is unquiesced via
248 * blk_mq_unquiesce_queue().
250 void blk_mq_quiesce_queue(struct request_queue *q)
252 blk_mq_quiesce_queue_nowait(q);
253 /* nothing to wait for non-mq queues */
255 blk_mq_wait_quiesce_done(q->tag_set);
257 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
260 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
263 * This function recovers queue into the state before quiescing
264 * which is done by blk_mq_quiesce_queue.
266 void blk_mq_unquiesce_queue(struct request_queue *q)
269 bool run_queue = false;
271 spin_lock_irqsave(&q->queue_lock, flags);
272 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
274 } else if (!--q->quiesce_depth) {
275 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
278 spin_unlock_irqrestore(&q->queue_lock, flags);
280 /* dispatch requests which are inserted during quiescing */
282 blk_mq_run_hw_queues(q, true);
284 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
286 void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
288 struct request_queue *q;
290 mutex_lock(&set->tag_list_lock);
291 list_for_each_entry(q, &set->tag_list, tag_set_list) {
292 if (!blk_queue_skip_tagset_quiesce(q))
293 blk_mq_quiesce_queue_nowait(q);
295 blk_mq_wait_quiesce_done(set);
296 mutex_unlock(&set->tag_list_lock);
298 EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
300 void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
302 struct request_queue *q;
304 mutex_lock(&set->tag_list_lock);
305 list_for_each_entry(q, &set->tag_list, tag_set_list) {
306 if (!blk_queue_skip_tagset_quiesce(q))
307 blk_mq_unquiesce_queue(q);
309 mutex_unlock(&set->tag_list_lock);
311 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
313 void blk_mq_wake_waiters(struct request_queue *q)
315 struct blk_mq_hw_ctx *hctx;
318 queue_for_each_hw_ctx(q, hctx, i)
319 if (blk_mq_hw_queue_mapped(hctx))
320 blk_mq_tag_wakeup_all(hctx->tags, true);
323 void blk_rq_init(struct request_queue *q, struct request *rq)
325 memset(rq, 0, sizeof(*rq));
327 INIT_LIST_HEAD(&rq->queuelist);
329 rq->__sector = (sector_t) -1;
330 INIT_HLIST_NODE(&rq->hash);
331 RB_CLEAR_NODE(&rq->rb_node);
332 rq->tag = BLK_MQ_NO_TAG;
333 rq->internal_tag = BLK_MQ_NO_TAG;
334 rq->start_time_ns = ktime_get_ns();
336 blk_crypto_rq_set_defaults(rq);
338 EXPORT_SYMBOL(blk_rq_init);
340 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
341 struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns)
343 struct blk_mq_ctx *ctx = data->ctx;
344 struct blk_mq_hw_ctx *hctx = data->hctx;
345 struct request_queue *q = data->q;
346 struct request *rq = tags->static_rqs[tag];
351 rq->cmd_flags = data->cmd_flags;
353 if (data->flags & BLK_MQ_REQ_PM)
354 data->rq_flags |= RQF_PM;
355 if (blk_queue_io_stat(q))
356 data->rq_flags |= RQF_IO_STAT;
357 rq->rq_flags = data->rq_flags;
359 if (data->rq_flags & RQF_SCHED_TAGS) {
360 rq->tag = BLK_MQ_NO_TAG;
361 rq->internal_tag = tag;
364 rq->internal_tag = BLK_MQ_NO_TAG;
368 if (blk_mq_need_time_stamp(rq))
369 rq->start_time_ns = ktime_get_ns();
371 rq->start_time_ns = 0;
373 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
374 rq->alloc_time_ns = alloc_time_ns;
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,
408 unsigned int tag, tag_offset;
409 struct blk_mq_tags *tags;
411 unsigned long tag_mask;
414 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
415 if (unlikely(!tag_mask))
418 tags = blk_mq_tags_from_data(data);
419 for (i = 0; tag_mask; i++) {
420 if (!(tag_mask & (1UL << i)))
422 tag = tag_offset + i;
423 prefetch(tags->static_rqs[tag]);
424 tag_mask &= ~(1UL << i);
425 rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns);
426 rq_list_add(data->cached_rq, rq);
429 /* caller already holds a reference, add for remainder */
430 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
433 return rq_list_pop(data->cached_rq);
436 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
438 struct request_queue *q = data->q;
439 u64 alloc_time_ns = 0;
443 /* alloc_time includes depth and tag waits */
444 if (blk_queue_rq_alloc_time(q))
445 alloc_time_ns = ktime_get_ns();
447 if (data->cmd_flags & REQ_NOWAIT)
448 data->flags |= BLK_MQ_REQ_NOWAIT;
452 * All requests use scheduler tags when an I/O scheduler is
453 * enabled for the queue.
455 data->rq_flags |= RQF_SCHED_TAGS;
458 * Flush/passthrough requests are special and go directly to the
461 if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&
462 !blk_op_is_passthrough(data->cmd_flags)) {
463 struct elevator_mq_ops *ops = &q->elevator->type->ops;
465 WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
467 data->rq_flags |= RQF_USE_SCHED;
468 if (ops->limit_depth)
469 ops->limit_depth(data->cmd_flags, data);
474 data->ctx = blk_mq_get_ctx(q);
475 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
476 if (!(data->rq_flags & RQF_SCHED_TAGS))
477 blk_mq_tag_busy(data->hctx);
479 if (data->flags & BLK_MQ_REQ_RESERVED)
480 data->rq_flags |= RQF_RESV;
483 * Try batched alloc if we want more than 1 tag.
485 if (data->nr_tags > 1) {
486 rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns);
493 * Waiting allocations only fail because of an inactive hctx. In that
494 * case just retry the hctx assignment and tag allocation as CPU hotplug
495 * should have migrated us to an online CPU by now.
497 tag = blk_mq_get_tag(data);
498 if (tag == BLK_MQ_NO_TAG) {
499 if (data->flags & BLK_MQ_REQ_NOWAIT)
502 * Give up the CPU and sleep for a random short time to
503 * ensure that thread using a realtime scheduling class
504 * are migrated off the CPU, and thus off the hctx that
511 return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag,
515 static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
516 struct blk_plug *plug,
518 blk_mq_req_flags_t flags)
520 struct blk_mq_alloc_data data = {
524 .nr_tags = plug->nr_ios,
525 .cached_rq = &plug->cached_rq,
529 if (blk_queue_enter(q, flags))
534 rq = __blk_mq_alloc_requests(&data);
540 static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
542 blk_mq_req_flags_t flags)
544 struct blk_plug *plug = current->plug;
550 if (rq_list_empty(plug->cached_rq)) {
551 if (plug->nr_ios == 1)
553 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
557 rq = rq_list_peek(&plug->cached_rq);
558 if (!rq || rq->q != q)
561 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
563 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
566 plug->cached_rq = rq_list_next(rq);
570 INIT_LIST_HEAD(&rq->queuelist);
574 struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
575 blk_mq_req_flags_t flags)
579 rq = blk_mq_alloc_cached_request(q, opf, flags);
581 struct blk_mq_alloc_data data = {
589 ret = blk_queue_enter(q, flags);
593 rq = __blk_mq_alloc_requests(&data);
598 rq->__sector = (sector_t) -1;
599 rq->bio = rq->biotail = NULL;
603 return ERR_PTR(-EWOULDBLOCK);
605 EXPORT_SYMBOL(blk_mq_alloc_request);
607 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
608 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
610 struct blk_mq_alloc_data data = {
616 u64 alloc_time_ns = 0;
622 /* alloc_time includes depth and tag waits */
623 if (blk_queue_rq_alloc_time(q))
624 alloc_time_ns = ktime_get_ns();
627 * If the tag allocator sleeps we could get an allocation for a
628 * different hardware context. No need to complicate the low level
629 * allocator for this for the rare use case of a command tied to
632 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
633 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
634 return ERR_PTR(-EINVAL);
636 if (hctx_idx >= q->nr_hw_queues)
637 return ERR_PTR(-EIO);
639 ret = blk_queue_enter(q, flags);
644 * Check if the hardware context is actually mapped to anything.
645 * If not tell the caller that it should skip this queue.
648 data.hctx = xa_load(&q->hctx_table, hctx_idx);
649 if (!blk_mq_hw_queue_mapped(data.hctx))
651 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
652 if (cpu >= nr_cpu_ids)
654 data.ctx = __blk_mq_get_ctx(q, cpu);
657 data.rq_flags |= RQF_SCHED_TAGS;
659 blk_mq_tag_busy(data.hctx);
661 if (flags & BLK_MQ_REQ_RESERVED)
662 data.rq_flags |= RQF_RESV;
665 tag = blk_mq_get_tag(&data);
666 if (tag == BLK_MQ_NO_TAG)
668 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag,
671 rq->__sector = (sector_t) -1;
672 rq->bio = rq->biotail = NULL;
679 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
681 static void __blk_mq_free_request(struct request *rq)
683 struct request_queue *q = rq->q;
684 struct blk_mq_ctx *ctx = rq->mq_ctx;
685 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
686 const int sched_tag = rq->internal_tag;
688 blk_crypto_free_request(rq);
689 blk_pm_mark_last_busy(rq);
691 if (rq->tag != BLK_MQ_NO_TAG)
692 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
693 if (sched_tag != BLK_MQ_NO_TAG)
694 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
695 blk_mq_sched_restart(hctx);
699 void blk_mq_free_request(struct request *rq)
701 struct request_queue *q = rq->q;
702 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
704 if ((rq->rq_flags & RQF_USE_SCHED) &&
705 q->elevator->type->ops.finish_request)
706 q->elevator->type->ops.finish_request(rq);
708 if (rq->rq_flags & RQF_MQ_INFLIGHT)
709 __blk_mq_dec_active_requests(hctx);
711 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
712 laptop_io_completion(q->disk->bdi);
716 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
717 if (req_ref_put_and_test(rq))
718 __blk_mq_free_request(rq);
720 EXPORT_SYMBOL_GPL(blk_mq_free_request);
722 void blk_mq_free_plug_rqs(struct blk_plug *plug)
726 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
727 blk_mq_free_request(rq);
730 void blk_dump_rq_flags(struct request *rq, char *msg)
732 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
733 rq->q->disk ? rq->q->disk->disk_name : "?",
734 (__force unsigned long long) rq->cmd_flags);
736 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
737 (unsigned long long)blk_rq_pos(rq),
738 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
739 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
740 rq->bio, rq->biotail, blk_rq_bytes(rq));
742 EXPORT_SYMBOL(blk_dump_rq_flags);
744 static void req_bio_endio(struct request *rq, struct bio *bio,
745 unsigned int nbytes, blk_status_t error)
747 if (unlikely(error)) {
748 bio->bi_status = error;
749 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
751 * Partial zone append completions cannot be supported as the
752 * BIO fragments may end up not being written sequentially.
754 if (bio->bi_iter.bi_size != nbytes)
755 bio->bi_status = BLK_STS_IOERR;
757 bio->bi_iter.bi_sector = rq->__sector;
760 bio_advance(bio, nbytes);
762 if (unlikely(rq->rq_flags & RQF_QUIET))
763 bio_set_flag(bio, BIO_QUIET);
764 /* don't actually finish bio if it's part of flush sequence */
765 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
769 static void blk_account_io_completion(struct request *req, unsigned int bytes)
771 if (req->part && blk_do_io_stat(req)) {
772 const int sgrp = op_stat_group(req_op(req));
775 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
780 static void blk_print_req_error(struct request *req, blk_status_t status)
782 printk_ratelimited(KERN_ERR
783 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
784 "phys_seg %u prio class %u\n",
785 blk_status_to_str(status),
786 req->q->disk ? req->q->disk->disk_name : "?",
787 blk_rq_pos(req), (__force u32)req_op(req),
788 blk_op_str(req_op(req)),
789 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
790 req->nr_phys_segments,
791 IOPRIO_PRIO_CLASS(req->ioprio));
795 * Fully end IO on a request. Does not support partial completions, or
798 static void blk_complete_request(struct request *req)
800 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
801 int total_bytes = blk_rq_bytes(req);
802 struct bio *bio = req->bio;
804 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
809 #ifdef CONFIG_BLK_DEV_INTEGRITY
810 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
811 req->q->integrity.profile->complete_fn(req, total_bytes);
815 * Upper layers may call blk_crypto_evict_key() anytime after the last
816 * bio_endio(). Therefore, the keyslot must be released before that.
818 blk_crypto_rq_put_keyslot(req);
820 blk_account_io_completion(req, total_bytes);
823 struct bio *next = bio->bi_next;
825 /* Completion has already been traced */
826 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
828 if (req_op(req) == REQ_OP_ZONE_APPEND)
829 bio->bi_iter.bi_sector = req->__sector;
837 * Reset counters so that the request stacking driver
838 * can find how many bytes remain in the request
848 * blk_update_request - Complete multiple bytes without completing the request
849 * @req: the request being processed
850 * @error: block status code
851 * @nr_bytes: number of bytes to complete for @req
854 * Ends I/O on a number of bytes attached to @req, but doesn't complete
855 * the request structure even if @req doesn't have leftover.
856 * If @req has leftover, sets it up for the next range of segments.
858 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
859 * %false return from this function.
862 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
863 * except in the consistency check at the end of this function.
866 * %false - this request doesn't have any more data
867 * %true - this request has more data
869 bool blk_update_request(struct request *req, blk_status_t error,
870 unsigned int nr_bytes)
874 trace_block_rq_complete(req, error, nr_bytes);
879 #ifdef CONFIG_BLK_DEV_INTEGRITY
880 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
882 req->q->integrity.profile->complete_fn(req, nr_bytes);
886 * Upper layers may call blk_crypto_evict_key() anytime after the last
887 * bio_endio(). Therefore, the keyslot must be released before that.
889 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
890 __blk_crypto_rq_put_keyslot(req);
892 if (unlikely(error && !blk_rq_is_passthrough(req) &&
893 !(req->rq_flags & RQF_QUIET)) &&
894 !test_bit(GD_DEAD, &req->q->disk->state)) {
895 blk_print_req_error(req, error);
896 trace_block_rq_error(req, error, nr_bytes);
899 blk_account_io_completion(req, nr_bytes);
903 struct bio *bio = req->bio;
904 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
906 if (bio_bytes == bio->bi_iter.bi_size)
907 req->bio = bio->bi_next;
909 /* Completion has already been traced */
910 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
911 req_bio_endio(req, bio, bio_bytes, error);
913 total_bytes += bio_bytes;
914 nr_bytes -= bio_bytes;
925 * Reset counters so that the request stacking driver
926 * can find how many bytes remain in the request
933 req->__data_len -= total_bytes;
935 /* update sector only for requests with clear definition of sector */
936 if (!blk_rq_is_passthrough(req))
937 req->__sector += total_bytes >> 9;
939 /* mixed attributes always follow the first bio */
940 if (req->rq_flags & RQF_MIXED_MERGE) {
941 req->cmd_flags &= ~REQ_FAILFAST_MASK;
942 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
945 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
947 * If total number of sectors is less than the first segment
948 * size, something has gone terribly wrong.
950 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
951 blk_dump_rq_flags(req, "request botched");
952 req->__data_len = blk_rq_cur_bytes(req);
955 /* recalculate the number of segments */
956 req->nr_phys_segments = blk_recalc_rq_segments(req);
961 EXPORT_SYMBOL_GPL(blk_update_request);
963 static inline void blk_account_io_done(struct request *req, u64 now)
966 * Account IO completion. flush_rq isn't accounted as a
967 * normal IO on queueing nor completion. Accounting the
968 * containing request is enough.
970 if (blk_do_io_stat(req) && req->part &&
971 !(req->rq_flags & RQF_FLUSH_SEQ)) {
972 const int sgrp = op_stat_group(req_op(req));
975 update_io_ticks(req->part, jiffies, true);
976 part_stat_inc(req->part, ios[sgrp]);
977 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
982 static inline void blk_account_io_start(struct request *req)
984 if (blk_do_io_stat(req)) {
986 * All non-passthrough requests are created from a bio with one
987 * exception: when a flush command that is part of a flush sequence
988 * generated by the state machine in blk-flush.c is cloned onto the
989 * lower device by dm-multipath we can get here without a bio.
992 req->part = req->bio->bi_bdev;
994 req->part = req->q->disk->part0;
997 update_io_ticks(req->part, jiffies, false);
1002 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1004 if (rq->rq_flags & RQF_STATS)
1005 blk_stat_add(rq, now);
1007 blk_mq_sched_completed_request(rq, now);
1008 blk_account_io_done(rq, now);
1011 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1013 if (blk_mq_need_time_stamp(rq))
1014 __blk_mq_end_request_acct(rq, ktime_get_ns());
1017 rq_qos_done(rq->q, rq);
1018 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1019 blk_mq_free_request(rq);
1021 blk_mq_free_request(rq);
1024 EXPORT_SYMBOL(__blk_mq_end_request);
1026 void blk_mq_end_request(struct request *rq, blk_status_t error)
1028 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1030 __blk_mq_end_request(rq, error);
1032 EXPORT_SYMBOL(blk_mq_end_request);
1034 #define TAG_COMP_BATCH 32
1036 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1037 int *tag_array, int nr_tags)
1039 struct request_queue *q = hctx->queue;
1042 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1043 * update hctx->nr_active in batch
1045 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1046 __blk_mq_sub_active_requests(hctx, nr_tags);
1048 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1049 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1052 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1054 int tags[TAG_COMP_BATCH], nr_tags = 0;
1055 struct blk_mq_hw_ctx *cur_hctx = NULL;
1060 now = ktime_get_ns();
1062 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1064 prefetch(rq->rq_next);
1066 blk_complete_request(rq);
1068 __blk_mq_end_request_acct(rq, now);
1070 rq_qos_done(rq->q, rq);
1073 * If end_io handler returns NONE, then it still has
1074 * ownership of the request.
1076 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1079 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1080 if (!req_ref_put_and_test(rq))
1083 blk_crypto_free_request(rq);
1084 blk_pm_mark_last_busy(rq);
1086 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1088 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1090 cur_hctx = rq->mq_hctx;
1092 tags[nr_tags++] = rq->tag;
1096 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1098 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1100 static void blk_complete_reqs(struct llist_head *list)
1102 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1103 struct request *rq, *next;
1105 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1106 rq->q->mq_ops->complete(rq);
1109 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1111 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1114 static int blk_softirq_cpu_dead(unsigned int cpu)
1116 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1120 static void __blk_mq_complete_request_remote(void *data)
1122 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1125 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1127 int cpu = raw_smp_processor_id();
1129 if (!IS_ENABLED(CONFIG_SMP) ||
1130 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1133 * With force threaded interrupts enabled, raising softirq from an SMP
1134 * function call will always result in waking the ksoftirqd thread.
1135 * This is probably worse than completing the request on a different
1138 if (force_irqthreads())
1141 /* same CPU or cache domain? Complete locally */
1142 if (cpu == rq->mq_ctx->cpu ||
1143 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1144 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1147 /* don't try to IPI to an offline CPU */
1148 return cpu_online(rq->mq_ctx->cpu);
1151 static void blk_mq_complete_send_ipi(struct request *rq)
1153 struct llist_head *list;
1156 cpu = rq->mq_ctx->cpu;
1157 list = &per_cpu(blk_cpu_done, cpu);
1158 if (llist_add(&rq->ipi_list, list)) {
1159 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1160 smp_call_function_single_async(cpu, &rq->csd);
1164 static void blk_mq_raise_softirq(struct request *rq)
1166 struct llist_head *list;
1169 list = this_cpu_ptr(&blk_cpu_done);
1170 if (llist_add(&rq->ipi_list, list))
1171 raise_softirq(BLOCK_SOFTIRQ);
1175 bool blk_mq_complete_request_remote(struct request *rq)
1177 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1180 * For request which hctx has only one ctx mapping,
1181 * or a polled request, always complete locally,
1182 * it's pointless to redirect the completion.
1184 if (rq->mq_hctx->nr_ctx == 1 ||
1185 rq->cmd_flags & REQ_POLLED)
1188 if (blk_mq_complete_need_ipi(rq)) {
1189 blk_mq_complete_send_ipi(rq);
1193 if (rq->q->nr_hw_queues == 1) {
1194 blk_mq_raise_softirq(rq);
1199 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1202 * blk_mq_complete_request - end I/O on a request
1203 * @rq: the request being processed
1206 * Complete a request by scheduling the ->complete_rq operation.
1208 void blk_mq_complete_request(struct request *rq)
1210 if (!blk_mq_complete_request_remote(rq))
1211 rq->q->mq_ops->complete(rq);
1213 EXPORT_SYMBOL(blk_mq_complete_request);
1216 * blk_mq_start_request - Start processing a request
1217 * @rq: Pointer to request to be started
1219 * Function used by device drivers to notify the block layer that a request
1220 * is going to be processed now, so blk layer can do proper initializations
1221 * such as starting the timeout timer.
1223 void blk_mq_start_request(struct request *rq)
1225 struct request_queue *q = rq->q;
1227 trace_block_rq_issue(rq);
1229 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1230 rq->io_start_time_ns = ktime_get_ns();
1231 rq->stats_sectors = blk_rq_sectors(rq);
1232 rq->rq_flags |= RQF_STATS;
1233 rq_qos_issue(q, rq);
1236 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1239 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1241 #ifdef CONFIG_BLK_DEV_INTEGRITY
1242 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1243 q->integrity.profile->prepare_fn(rq);
1245 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1246 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1248 EXPORT_SYMBOL(blk_mq_start_request);
1251 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1252 * queues. This is important for md arrays to benefit from merging
1255 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1257 if (plug->multiple_queues)
1258 return BLK_MAX_REQUEST_COUNT * 2;
1259 return BLK_MAX_REQUEST_COUNT;
1262 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1264 struct request *last = rq_list_peek(&plug->mq_list);
1266 if (!plug->rq_count) {
1267 trace_block_plug(rq->q);
1268 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1269 (!blk_queue_nomerges(rq->q) &&
1270 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1271 blk_mq_flush_plug_list(plug, false);
1273 trace_block_plug(rq->q);
1276 if (!plug->multiple_queues && last && last->q != rq->q)
1277 plug->multiple_queues = true;
1278 if (!plug->has_elevator && (rq->rq_flags & RQF_USE_SCHED))
1279 plug->has_elevator = true;
1281 rq_list_add(&plug->mq_list, rq);
1286 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1287 * @rq: request to insert
1288 * @at_head: insert request at head or tail of queue
1291 * Insert a fully prepared request at the back of the I/O scheduler queue
1292 * for execution. Don't wait for completion.
1295 * This function will invoke @done directly if the queue is dead.
1297 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1299 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1301 WARN_ON(irqs_disabled());
1302 WARN_ON(!blk_rq_is_passthrough(rq));
1304 blk_account_io_start(rq);
1307 * As plugging can be enabled for passthrough requests on a zoned
1308 * device, directly accessing the plug instead of using blk_mq_plug()
1309 * should not have any consequences.
1311 if (current->plug && !at_head) {
1312 blk_add_rq_to_plug(current->plug, rq);
1316 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1317 blk_mq_run_hw_queue(hctx, false);
1319 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1321 struct blk_rq_wait {
1322 struct completion done;
1326 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1328 struct blk_rq_wait *wait = rq->end_io_data;
1331 complete(&wait->done);
1332 return RQ_END_IO_NONE;
1335 bool blk_rq_is_poll(struct request *rq)
1339 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1343 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1345 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1348 blk_mq_poll(rq->q, blk_rq_to_qc(rq), NULL, 0);
1350 } while (!completion_done(wait));
1354 * blk_execute_rq - insert a request into queue for execution
1355 * @rq: request to insert
1356 * @at_head: insert request at head or tail of queue
1359 * Insert a fully prepared request at the back of the I/O scheduler queue
1360 * for execution and wait for completion.
1361 * Return: The blk_status_t result provided to blk_mq_end_request().
1363 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1365 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1366 struct blk_rq_wait wait = {
1367 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1370 WARN_ON(irqs_disabled());
1371 WARN_ON(!blk_rq_is_passthrough(rq));
1373 rq->end_io_data = &wait;
1374 rq->end_io = blk_end_sync_rq;
1376 blk_account_io_start(rq);
1377 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1378 blk_mq_run_hw_queue(hctx, false);
1380 if (blk_rq_is_poll(rq)) {
1381 blk_rq_poll_completion(rq, &wait.done);
1384 * Prevent hang_check timer from firing at us during very long
1387 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1390 while (!wait_for_completion_io_timeout(&wait.done,
1391 hang_check * (HZ/2)))
1394 wait_for_completion_io(&wait.done);
1399 EXPORT_SYMBOL(blk_execute_rq);
1401 static void __blk_mq_requeue_request(struct request *rq)
1403 struct request_queue *q = rq->q;
1405 blk_mq_put_driver_tag(rq);
1407 trace_block_rq_requeue(rq);
1408 rq_qos_requeue(q, rq);
1410 if (blk_mq_request_started(rq)) {
1411 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1412 rq->rq_flags &= ~RQF_TIMED_OUT;
1416 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1418 struct request_queue *q = rq->q;
1420 __blk_mq_requeue_request(rq);
1422 /* this request will be re-inserted to io scheduler queue */
1423 blk_mq_sched_requeue_request(rq);
1425 blk_mq_add_to_requeue_list(rq, BLK_MQ_INSERT_AT_HEAD);
1427 if (kick_requeue_list)
1428 blk_mq_kick_requeue_list(q);
1430 EXPORT_SYMBOL(blk_mq_requeue_request);
1432 static void blk_mq_requeue_work(struct work_struct *work)
1434 struct request_queue *q =
1435 container_of(work, struct request_queue, requeue_work.work);
1437 struct request *rq, *next;
1439 spin_lock_irq(&q->requeue_lock);
1440 list_splice_init(&q->requeue_list, &rq_list);
1441 spin_unlock_irq(&q->requeue_lock);
1443 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1445 * If RQF_DONTPREP ist set, the request has been started by the
1446 * driver already and might have driver-specific data allocated
1447 * already. Insert it into the hctx dispatch list to avoid
1448 * block layer merges for the request.
1450 if (rq->rq_flags & RQF_DONTPREP) {
1451 rq->rq_flags &= ~RQF_SOFTBARRIER;
1452 list_del_init(&rq->queuelist);
1453 blk_mq_request_bypass_insert(rq, 0);
1454 } else if (rq->rq_flags & RQF_SOFTBARRIER) {
1455 rq->rq_flags &= ~RQF_SOFTBARRIER;
1456 list_del_init(&rq->queuelist);
1457 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1461 while (!list_empty(&rq_list)) {
1462 rq = list_entry(rq_list.next, struct request, queuelist);
1463 list_del_init(&rq->queuelist);
1464 blk_mq_insert_request(rq, 0);
1467 blk_mq_run_hw_queues(q, false);
1470 void blk_mq_add_to_requeue_list(struct request *rq, blk_insert_t insert_flags)
1472 struct request_queue *q = rq->q;
1473 unsigned long flags;
1476 * We abuse this flag that is otherwise used by the I/O scheduler to
1477 * request head insertion from the workqueue.
1479 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1481 spin_lock_irqsave(&q->requeue_lock, flags);
1482 if (insert_flags & BLK_MQ_INSERT_AT_HEAD) {
1483 rq->rq_flags |= RQF_SOFTBARRIER;
1484 list_add(&rq->queuelist, &q->requeue_list);
1486 list_add_tail(&rq->queuelist, &q->requeue_list);
1488 spin_unlock_irqrestore(&q->requeue_lock, flags);
1491 void blk_mq_kick_requeue_list(struct request_queue *q)
1493 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1495 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1497 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1498 unsigned long msecs)
1500 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1501 msecs_to_jiffies(msecs));
1503 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1505 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1508 * If we find a request that isn't idle we know the queue is busy
1509 * as it's checked in the iter.
1510 * Return false to stop the iteration.
1512 if (blk_mq_request_started(rq)) {
1522 bool blk_mq_queue_inflight(struct request_queue *q)
1526 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1529 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1531 static void blk_mq_rq_timed_out(struct request *req)
1533 req->rq_flags |= RQF_TIMED_OUT;
1534 if (req->q->mq_ops->timeout) {
1535 enum blk_eh_timer_return ret;
1537 ret = req->q->mq_ops->timeout(req);
1538 if (ret == BLK_EH_DONE)
1540 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1546 struct blk_expired_data {
1547 bool has_timedout_rq;
1549 unsigned long timeout_start;
1552 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1554 unsigned long deadline;
1556 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1558 if (rq->rq_flags & RQF_TIMED_OUT)
1561 deadline = READ_ONCE(rq->deadline);
1562 if (time_after_eq(expired->timeout_start, deadline))
1565 if (expired->next == 0)
1566 expired->next = deadline;
1567 else if (time_after(expired->next, deadline))
1568 expired->next = deadline;
1572 void blk_mq_put_rq_ref(struct request *rq)
1574 if (is_flush_rq(rq)) {
1575 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1576 blk_mq_free_request(rq);
1577 } else if (req_ref_put_and_test(rq)) {
1578 __blk_mq_free_request(rq);
1582 static bool blk_mq_check_expired(struct request *rq, void *priv)
1584 struct blk_expired_data *expired = priv;
1587 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1588 * be reallocated underneath the timeout handler's processing, then
1589 * the expire check is reliable. If the request is not expired, then
1590 * it was completed and reallocated as a new request after returning
1591 * from blk_mq_check_expired().
1593 if (blk_mq_req_expired(rq, expired)) {
1594 expired->has_timedout_rq = true;
1600 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1602 struct blk_expired_data *expired = priv;
1604 if (blk_mq_req_expired(rq, expired))
1605 blk_mq_rq_timed_out(rq);
1609 static void blk_mq_timeout_work(struct work_struct *work)
1611 struct request_queue *q =
1612 container_of(work, struct request_queue, timeout_work);
1613 struct blk_expired_data expired = {
1614 .timeout_start = jiffies,
1616 struct blk_mq_hw_ctx *hctx;
1619 /* A deadlock might occur if a request is stuck requiring a
1620 * timeout at the same time a queue freeze is waiting
1621 * completion, since the timeout code would not be able to
1622 * acquire the queue reference here.
1624 * That's why we don't use blk_queue_enter here; instead, we use
1625 * percpu_ref_tryget directly, because we need to be able to
1626 * obtain a reference even in the short window between the queue
1627 * starting to freeze, by dropping the first reference in
1628 * blk_freeze_queue_start, and the moment the last request is
1629 * consumed, marked by the instant q_usage_counter reaches
1632 if (!percpu_ref_tryget(&q->q_usage_counter))
1635 /* check if there is any timed-out request */
1636 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1637 if (expired.has_timedout_rq) {
1639 * Before walking tags, we must ensure any submit started
1640 * before the current time has finished. Since the submit
1641 * uses srcu or rcu, wait for a synchronization point to
1642 * ensure all running submits have finished
1644 blk_mq_wait_quiesce_done(q->tag_set);
1647 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1650 if (expired.next != 0) {
1651 mod_timer(&q->timeout, expired.next);
1654 * Request timeouts are handled as a forward rolling timer. If
1655 * we end up here it means that no requests are pending and
1656 * also that no request has been pending for a while. Mark
1657 * each hctx as idle.
1659 queue_for_each_hw_ctx(q, hctx, i) {
1660 /* the hctx may be unmapped, so check it here */
1661 if (blk_mq_hw_queue_mapped(hctx))
1662 blk_mq_tag_idle(hctx);
1668 struct flush_busy_ctx_data {
1669 struct blk_mq_hw_ctx *hctx;
1670 struct list_head *list;
1673 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1675 struct flush_busy_ctx_data *flush_data = data;
1676 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1677 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1678 enum hctx_type type = hctx->type;
1680 spin_lock(&ctx->lock);
1681 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1682 sbitmap_clear_bit(sb, bitnr);
1683 spin_unlock(&ctx->lock);
1688 * Process software queues that have been marked busy, splicing them
1689 * to the for-dispatch
1691 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1693 struct flush_busy_ctx_data data = {
1698 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1700 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1702 struct dispatch_rq_data {
1703 struct blk_mq_hw_ctx *hctx;
1707 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1710 struct dispatch_rq_data *dispatch_data = data;
1711 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1712 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1713 enum hctx_type type = hctx->type;
1715 spin_lock(&ctx->lock);
1716 if (!list_empty(&ctx->rq_lists[type])) {
1717 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1718 list_del_init(&dispatch_data->rq->queuelist);
1719 if (list_empty(&ctx->rq_lists[type]))
1720 sbitmap_clear_bit(sb, bitnr);
1722 spin_unlock(&ctx->lock);
1724 return !dispatch_data->rq;
1727 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1728 struct blk_mq_ctx *start)
1730 unsigned off = start ? start->index_hw[hctx->type] : 0;
1731 struct dispatch_rq_data data = {
1736 __sbitmap_for_each_set(&hctx->ctx_map, off,
1737 dispatch_rq_from_ctx, &data);
1742 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1744 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1745 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1748 blk_mq_tag_busy(rq->mq_hctx);
1750 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1751 bt = &rq->mq_hctx->tags->breserved_tags;
1754 if (!hctx_may_queue(rq->mq_hctx, bt))
1758 tag = __sbitmap_queue_get(bt);
1759 if (tag == BLK_MQ_NO_TAG)
1762 rq->tag = tag + tag_offset;
1766 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1768 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1771 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1772 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1773 rq->rq_flags |= RQF_MQ_INFLIGHT;
1774 __blk_mq_inc_active_requests(hctx);
1776 hctx->tags->rqs[rq->tag] = rq;
1780 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1781 int flags, void *key)
1783 struct blk_mq_hw_ctx *hctx;
1785 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1787 spin_lock(&hctx->dispatch_wait_lock);
1788 if (!list_empty(&wait->entry)) {
1789 struct sbitmap_queue *sbq;
1791 list_del_init(&wait->entry);
1792 sbq = &hctx->tags->bitmap_tags;
1793 atomic_dec(&sbq->ws_active);
1795 spin_unlock(&hctx->dispatch_wait_lock);
1797 blk_mq_run_hw_queue(hctx, true);
1802 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1803 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1804 * restart. For both cases, take care to check the condition again after
1805 * marking us as waiting.
1807 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1810 struct sbitmap_queue *sbq;
1811 struct wait_queue_head *wq;
1812 wait_queue_entry_t *wait;
1815 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1816 !(blk_mq_is_shared_tags(hctx->flags))) {
1817 blk_mq_sched_mark_restart_hctx(hctx);
1820 * It's possible that a tag was freed in the window between the
1821 * allocation failure and adding the hardware queue to the wait
1824 * Don't clear RESTART here, someone else could have set it.
1825 * At most this will cost an extra queue run.
1827 return blk_mq_get_driver_tag(rq);
1830 wait = &hctx->dispatch_wait;
1831 if (!list_empty_careful(&wait->entry))
1834 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1835 sbq = &hctx->tags->breserved_tags;
1837 sbq = &hctx->tags->bitmap_tags;
1838 wq = &bt_wait_ptr(sbq, hctx)->wait;
1840 spin_lock_irq(&wq->lock);
1841 spin_lock(&hctx->dispatch_wait_lock);
1842 if (!list_empty(&wait->entry)) {
1843 spin_unlock(&hctx->dispatch_wait_lock);
1844 spin_unlock_irq(&wq->lock);
1848 atomic_inc(&sbq->ws_active);
1849 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1850 __add_wait_queue(wq, wait);
1853 * It's possible that a tag was freed in the window between the
1854 * allocation failure and adding the hardware queue to the wait
1857 ret = blk_mq_get_driver_tag(rq);
1859 spin_unlock(&hctx->dispatch_wait_lock);
1860 spin_unlock_irq(&wq->lock);
1865 * We got a tag, remove ourselves from the wait queue to ensure
1866 * someone else gets the wakeup.
1868 list_del_init(&wait->entry);
1869 atomic_dec(&sbq->ws_active);
1870 spin_unlock(&hctx->dispatch_wait_lock);
1871 spin_unlock_irq(&wq->lock);
1876 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1877 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1879 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1880 * - EWMA is one simple way to compute running average value
1881 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1882 * - take 4 as factor for avoiding to get too small(0) result, and this
1883 * factor doesn't matter because EWMA decreases exponentially
1885 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1889 ewma = hctx->dispatch_busy;
1894 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1896 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1897 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1899 hctx->dispatch_busy = ewma;
1902 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1904 static void blk_mq_handle_dev_resource(struct request *rq,
1905 struct list_head *list)
1907 list_add(&rq->queuelist, list);
1908 __blk_mq_requeue_request(rq);
1911 static void blk_mq_handle_zone_resource(struct request *rq,
1912 struct list_head *zone_list)
1915 * If we end up here it is because we cannot dispatch a request to a
1916 * specific zone due to LLD level zone-write locking or other zone
1917 * related resource not being available. In this case, set the request
1918 * aside in zone_list for retrying it later.
1920 list_add(&rq->queuelist, zone_list);
1921 __blk_mq_requeue_request(rq);
1924 enum prep_dispatch {
1926 PREP_DISPATCH_NO_TAG,
1927 PREP_DISPATCH_NO_BUDGET,
1930 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1933 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1934 int budget_token = -1;
1937 budget_token = blk_mq_get_dispatch_budget(rq->q);
1938 if (budget_token < 0) {
1939 blk_mq_put_driver_tag(rq);
1940 return PREP_DISPATCH_NO_BUDGET;
1942 blk_mq_set_rq_budget_token(rq, budget_token);
1945 if (!blk_mq_get_driver_tag(rq)) {
1947 * The initial allocation attempt failed, so we need to
1948 * rerun the hardware queue when a tag is freed. The
1949 * waitqueue takes care of that. If the queue is run
1950 * before we add this entry back on the dispatch list,
1951 * we'll re-run it below.
1953 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1955 * All budgets not got from this function will be put
1956 * together during handling partial dispatch
1959 blk_mq_put_dispatch_budget(rq->q, budget_token);
1960 return PREP_DISPATCH_NO_TAG;
1964 return PREP_DISPATCH_OK;
1967 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1968 static void blk_mq_release_budgets(struct request_queue *q,
1969 struct list_head *list)
1973 list_for_each_entry(rq, list, queuelist) {
1974 int budget_token = blk_mq_get_rq_budget_token(rq);
1976 if (budget_token >= 0)
1977 blk_mq_put_dispatch_budget(q, budget_token);
1982 * blk_mq_commit_rqs will notify driver using bd->last that there is no
1983 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
1985 * Attention, we should explicitly call this in unusual cases:
1986 * 1) did not queue everything initially scheduled to queue
1987 * 2) the last attempt to queue a request failed
1989 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
1992 if (hctx->queue->mq_ops->commit_rqs && queued) {
1993 trace_block_unplug(hctx->queue, queued, !from_schedule);
1994 hctx->queue->mq_ops->commit_rqs(hctx);
1999 * Returns true if we did some work AND can potentially do more.
2001 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
2002 unsigned int nr_budgets)
2004 enum prep_dispatch prep;
2005 struct request_queue *q = hctx->queue;
2008 blk_status_t ret = BLK_STS_OK;
2009 LIST_HEAD(zone_list);
2010 bool needs_resource = false;
2012 if (list_empty(list))
2016 * Now process all the entries, sending them to the driver.
2020 struct blk_mq_queue_data bd;
2022 rq = list_first_entry(list, struct request, queuelist);
2024 WARN_ON_ONCE(hctx != rq->mq_hctx);
2025 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2026 if (prep != PREP_DISPATCH_OK)
2029 list_del_init(&rq->queuelist);
2032 bd.last = list_empty(list);
2035 * once the request is queued to lld, no need to cover the
2040 ret = q->mq_ops->queue_rq(hctx, &bd);
2045 case BLK_STS_RESOURCE:
2046 needs_resource = true;
2048 case BLK_STS_DEV_RESOURCE:
2049 blk_mq_handle_dev_resource(rq, list);
2051 case BLK_STS_ZONE_RESOURCE:
2053 * Move the request to zone_list and keep going through
2054 * the dispatch list to find more requests the drive can
2057 blk_mq_handle_zone_resource(rq, &zone_list);
2058 needs_resource = true;
2061 blk_mq_end_request(rq, ret);
2063 } while (!list_empty(list));
2065 if (!list_empty(&zone_list))
2066 list_splice_tail_init(&zone_list, list);
2068 /* If we didn't flush the entire list, we could have told the driver
2069 * there was more coming, but that turned out to be a lie.
2071 if (!list_empty(list) || ret != BLK_STS_OK)
2072 blk_mq_commit_rqs(hctx, queued, false);
2075 * Any items that need requeuing? Stuff them into hctx->dispatch,
2076 * that is where we will continue on next queue run.
2078 if (!list_empty(list)) {
2080 /* For non-shared tags, the RESTART check will suffice */
2081 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2082 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2083 blk_mq_is_shared_tags(hctx->flags));
2086 blk_mq_release_budgets(q, list);
2088 spin_lock(&hctx->lock);
2089 list_splice_tail_init(list, &hctx->dispatch);
2090 spin_unlock(&hctx->lock);
2093 * Order adding requests to hctx->dispatch and checking
2094 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2095 * in blk_mq_sched_restart(). Avoid restart code path to
2096 * miss the new added requests to hctx->dispatch, meantime
2097 * SCHED_RESTART is observed here.
2102 * If SCHED_RESTART was set by the caller of this function and
2103 * it is no longer set that means that it was cleared by another
2104 * thread and hence that a queue rerun is needed.
2106 * If 'no_tag' is set, that means that we failed getting
2107 * a driver tag with an I/O scheduler attached. If our dispatch
2108 * waitqueue is no longer active, ensure that we run the queue
2109 * AFTER adding our entries back to the list.
2111 * If no I/O scheduler has been configured it is possible that
2112 * the hardware queue got stopped and restarted before requests
2113 * were pushed back onto the dispatch list. Rerun the queue to
2114 * avoid starvation. Notes:
2115 * - blk_mq_run_hw_queue() checks whether or not a queue has
2116 * been stopped before rerunning a queue.
2117 * - Some but not all block drivers stop a queue before
2118 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2121 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2122 * bit is set, run queue after a delay to avoid IO stalls
2123 * that could otherwise occur if the queue is idle. We'll do
2124 * similar if we couldn't get budget or couldn't lock a zone
2125 * and SCHED_RESTART is set.
2127 needs_restart = blk_mq_sched_needs_restart(hctx);
2128 if (prep == PREP_DISPATCH_NO_BUDGET)
2129 needs_resource = true;
2130 if (!needs_restart ||
2131 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2132 blk_mq_run_hw_queue(hctx, true);
2133 else if (needs_resource)
2134 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2136 blk_mq_update_dispatch_busy(hctx, true);
2140 blk_mq_update_dispatch_busy(hctx, false);
2144 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2146 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2148 if (cpu >= nr_cpu_ids)
2149 cpu = cpumask_first(hctx->cpumask);
2154 * It'd be great if the workqueue API had a way to pass
2155 * in a mask and had some smarts for more clever placement.
2156 * For now we just round-robin here, switching for every
2157 * BLK_MQ_CPU_WORK_BATCH queued items.
2159 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2162 int next_cpu = hctx->next_cpu;
2164 if (hctx->queue->nr_hw_queues == 1)
2165 return WORK_CPU_UNBOUND;
2167 if (--hctx->next_cpu_batch <= 0) {
2169 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2171 if (next_cpu >= nr_cpu_ids)
2172 next_cpu = blk_mq_first_mapped_cpu(hctx);
2173 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2177 * Do unbound schedule if we can't find a online CPU for this hctx,
2178 * and it should only happen in the path of handling CPU DEAD.
2180 if (!cpu_online(next_cpu)) {
2187 * Make sure to re-select CPU next time once after CPUs
2188 * in hctx->cpumask become online again.
2190 hctx->next_cpu = next_cpu;
2191 hctx->next_cpu_batch = 1;
2192 return WORK_CPU_UNBOUND;
2195 hctx->next_cpu = next_cpu;
2200 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2201 * @hctx: Pointer to the hardware queue to run.
2202 * @msecs: Milliseconds of delay to wait before running the queue.
2204 * Run a hardware queue asynchronously with a delay of @msecs.
2206 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2208 if (unlikely(blk_mq_hctx_stopped(hctx)))
2210 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2211 msecs_to_jiffies(msecs));
2213 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2216 * blk_mq_run_hw_queue - Start to run a hardware queue.
2217 * @hctx: Pointer to the hardware queue to run.
2218 * @async: If we want to run the queue asynchronously.
2220 * Check if the request queue is not in a quiesced state and if there are
2221 * pending requests to be sent. If this is true, run the queue to send requests
2224 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2229 * We can't run the queue inline with interrupts disabled.
2231 WARN_ON_ONCE(!async && in_interrupt());
2234 * When queue is quiesced, we may be switching io scheduler, or
2235 * updating nr_hw_queues, or other things, and we can't run queue
2236 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2238 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2241 __blk_mq_run_dispatch_ops(hctx->queue, false,
2242 need_run = !blk_queue_quiesced(hctx->queue) &&
2243 blk_mq_hctx_has_pending(hctx));
2248 if (async || (hctx->flags & BLK_MQ_F_BLOCKING) ||
2249 !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2250 blk_mq_delay_run_hw_queue(hctx, 0);
2254 blk_mq_run_dispatch_ops(hctx->queue,
2255 blk_mq_sched_dispatch_requests(hctx));
2257 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2260 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2263 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2265 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2267 * If the IO scheduler does not respect hardware queues when
2268 * dispatching, we just don't bother with multiple HW queues and
2269 * dispatch from hctx for the current CPU since running multiple queues
2270 * just causes lock contention inside the scheduler and pointless cache
2273 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2275 if (!blk_mq_hctx_stopped(hctx))
2281 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2282 * @q: Pointer to the request queue to run.
2283 * @async: If we want to run the queue asynchronously.
2285 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2287 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2291 if (blk_queue_sq_sched(q))
2292 sq_hctx = blk_mq_get_sq_hctx(q);
2293 queue_for_each_hw_ctx(q, hctx, i) {
2294 if (blk_mq_hctx_stopped(hctx))
2297 * Dispatch from this hctx either if there's no hctx preferred
2298 * by IO scheduler or if it has requests that bypass the
2301 if (!sq_hctx || sq_hctx == hctx ||
2302 !list_empty_careful(&hctx->dispatch))
2303 blk_mq_run_hw_queue(hctx, async);
2306 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2309 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2310 * @q: Pointer to the request queue to run.
2311 * @msecs: Milliseconds of delay to wait before running the queues.
2313 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2315 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2319 if (blk_queue_sq_sched(q))
2320 sq_hctx = blk_mq_get_sq_hctx(q);
2321 queue_for_each_hw_ctx(q, hctx, i) {
2322 if (blk_mq_hctx_stopped(hctx))
2325 * If there is already a run_work pending, leave the
2326 * pending delay untouched. Otherwise, a hctx can stall
2327 * if another hctx is re-delaying the other's work
2328 * before the work executes.
2330 if (delayed_work_pending(&hctx->run_work))
2333 * Dispatch from this hctx either if there's no hctx preferred
2334 * by IO scheduler or if it has requests that bypass the
2337 if (!sq_hctx || sq_hctx == hctx ||
2338 !list_empty_careful(&hctx->dispatch))
2339 blk_mq_delay_run_hw_queue(hctx, msecs);
2342 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2345 * This function is often used for pausing .queue_rq() by driver when
2346 * there isn't enough resource or some conditions aren't satisfied, and
2347 * BLK_STS_RESOURCE is usually returned.
2349 * We do not guarantee that dispatch can be drained or blocked
2350 * after blk_mq_stop_hw_queue() returns. Please use
2351 * blk_mq_quiesce_queue() for that requirement.
2353 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2355 cancel_delayed_work(&hctx->run_work);
2357 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2359 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2362 * This function is often used for pausing .queue_rq() by driver when
2363 * there isn't enough resource or some conditions aren't satisfied, and
2364 * BLK_STS_RESOURCE is usually returned.
2366 * We do not guarantee that dispatch can be drained or blocked
2367 * after blk_mq_stop_hw_queues() returns. Please use
2368 * blk_mq_quiesce_queue() for that requirement.
2370 void blk_mq_stop_hw_queues(struct request_queue *q)
2372 struct blk_mq_hw_ctx *hctx;
2375 queue_for_each_hw_ctx(q, hctx, i)
2376 blk_mq_stop_hw_queue(hctx);
2378 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2380 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2382 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2384 blk_mq_run_hw_queue(hctx, false);
2386 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2388 void blk_mq_start_hw_queues(struct request_queue *q)
2390 struct blk_mq_hw_ctx *hctx;
2393 queue_for_each_hw_ctx(q, hctx, i)
2394 blk_mq_start_hw_queue(hctx);
2396 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2398 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2400 if (!blk_mq_hctx_stopped(hctx))
2403 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2404 blk_mq_run_hw_queue(hctx, async);
2406 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2408 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2410 struct blk_mq_hw_ctx *hctx;
2413 queue_for_each_hw_ctx(q, hctx, i)
2414 blk_mq_start_stopped_hw_queue(hctx, async);
2416 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2418 static void blk_mq_run_work_fn(struct work_struct *work)
2420 struct blk_mq_hw_ctx *hctx =
2421 container_of(work, struct blk_mq_hw_ctx, run_work.work);
2423 blk_mq_run_dispatch_ops(hctx->queue,
2424 blk_mq_sched_dispatch_requests(hctx));
2428 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2429 * @rq: Pointer to request to be inserted.
2430 * @flags: BLK_MQ_INSERT_*
2432 * Should only be used carefully, when the caller knows we want to
2433 * bypass a potential IO scheduler on the target device.
2435 static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2437 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2439 spin_lock(&hctx->lock);
2440 if (flags & BLK_MQ_INSERT_AT_HEAD)
2441 list_add(&rq->queuelist, &hctx->dispatch);
2443 list_add_tail(&rq->queuelist, &hctx->dispatch);
2444 spin_unlock(&hctx->lock);
2447 static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2448 struct blk_mq_ctx *ctx, struct list_head *list,
2449 bool run_queue_async)
2452 enum hctx_type type = hctx->type;
2455 * Try to issue requests directly if the hw queue isn't busy to save an
2456 * extra enqueue & dequeue to the sw queue.
2458 if (!hctx->dispatch_busy && !run_queue_async) {
2459 blk_mq_run_dispatch_ops(hctx->queue,
2460 blk_mq_try_issue_list_directly(hctx, list));
2461 if (list_empty(list))
2466 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2469 list_for_each_entry(rq, list, queuelist) {
2470 BUG_ON(rq->mq_ctx != ctx);
2471 trace_block_rq_insert(rq);
2474 spin_lock(&ctx->lock);
2475 list_splice_tail_init(list, &ctx->rq_lists[type]);
2476 blk_mq_hctx_mark_pending(hctx, ctx);
2477 spin_unlock(&ctx->lock);
2479 blk_mq_run_hw_queue(hctx, run_queue_async);
2482 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2484 struct request_queue *q = rq->q;
2485 struct blk_mq_ctx *ctx = rq->mq_ctx;
2486 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2488 if (blk_rq_is_passthrough(rq)) {
2490 * Passthrough request have to be added to hctx->dispatch
2491 * directly. The device may be in a situation where it can't
2492 * handle FS request, and always returns BLK_STS_RESOURCE for
2493 * them, which gets them added to hctx->dispatch.
2495 * If a passthrough request is required to unblock the queues,
2496 * and it is added to the scheduler queue, there is no chance to
2497 * dispatch it given we prioritize requests in hctx->dispatch.
2499 blk_mq_request_bypass_insert(rq, flags);
2500 } else if (req_op(rq) == REQ_OP_FLUSH) {
2502 * Firstly normal IO request is inserted to scheduler queue or
2503 * sw queue, meantime we add flush request to dispatch queue(
2504 * hctx->dispatch) directly and there is at most one in-flight
2505 * flush request for each hw queue, so it doesn't matter to add
2506 * flush request to tail or front of the dispatch queue.
2508 * Secondly in case of NCQ, flush request belongs to non-NCQ
2509 * command, and queueing it will fail when there is any
2510 * in-flight normal IO request(NCQ command). When adding flush
2511 * rq to the front of hctx->dispatch, it is easier to introduce
2512 * extra time to flush rq's latency because of S_SCHED_RESTART
2513 * compared with adding to the tail of dispatch queue, then
2514 * chance of flush merge is increased, and less flush requests
2515 * will be issued to controller. It is observed that ~10% time
2516 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2517 * drive when adding flush rq to the front of hctx->dispatch.
2519 * Simply queue flush rq to the front of hctx->dispatch so that
2520 * intensive flush workloads can benefit in case of NCQ HW.
2522 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2523 } else if (q->elevator) {
2526 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2528 list_add(&rq->queuelist, &list);
2529 q->elevator->type->ops.insert_requests(hctx, &list, flags);
2531 trace_block_rq_insert(rq);
2533 spin_lock(&ctx->lock);
2534 if (flags & BLK_MQ_INSERT_AT_HEAD)
2535 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2537 list_add_tail(&rq->queuelist,
2538 &ctx->rq_lists[hctx->type]);
2539 blk_mq_hctx_mark_pending(hctx, ctx);
2540 spin_unlock(&ctx->lock);
2544 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2545 unsigned int nr_segs)
2549 if (bio->bi_opf & REQ_RAHEAD)
2550 rq->cmd_flags |= REQ_FAILFAST_MASK;
2552 rq->__sector = bio->bi_iter.bi_sector;
2553 blk_rq_bio_prep(rq, bio, nr_segs);
2555 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2556 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2559 blk_account_io_start(rq);
2562 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2563 struct request *rq, bool last)
2565 struct request_queue *q = rq->q;
2566 struct blk_mq_queue_data bd = {
2573 * For OK queue, we are done. For error, caller may kill it.
2574 * Any other error (busy), just add it to our list as we
2575 * previously would have done.
2577 ret = q->mq_ops->queue_rq(hctx, &bd);
2580 blk_mq_update_dispatch_busy(hctx, false);
2582 case BLK_STS_RESOURCE:
2583 case BLK_STS_DEV_RESOURCE:
2584 blk_mq_update_dispatch_busy(hctx, true);
2585 __blk_mq_requeue_request(rq);
2588 blk_mq_update_dispatch_busy(hctx, false);
2595 static bool blk_mq_get_budget_and_tag(struct request *rq)
2599 budget_token = blk_mq_get_dispatch_budget(rq->q);
2600 if (budget_token < 0)
2602 blk_mq_set_rq_budget_token(rq, budget_token);
2603 if (!blk_mq_get_driver_tag(rq)) {
2604 blk_mq_put_dispatch_budget(rq->q, budget_token);
2611 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2612 * @hctx: Pointer of the associated hardware queue.
2613 * @rq: Pointer to request to be sent.
2615 * If the device has enough resources to accept a new request now, send the
2616 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2617 * we can try send it another time in the future. Requests inserted at this
2618 * queue have higher priority.
2620 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2625 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2626 blk_mq_insert_request(rq, 0);
2630 if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
2631 blk_mq_insert_request(rq, 0);
2632 blk_mq_run_hw_queue(hctx, false);
2636 ret = __blk_mq_issue_directly(hctx, rq, true);
2640 case BLK_STS_RESOURCE:
2641 case BLK_STS_DEV_RESOURCE:
2642 blk_mq_request_bypass_insert(rq, 0);
2643 blk_mq_run_hw_queue(hctx, false);
2646 blk_mq_end_request(rq, ret);
2651 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2653 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2655 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2656 blk_mq_insert_request(rq, 0);
2660 if (!blk_mq_get_budget_and_tag(rq))
2661 return BLK_STS_RESOURCE;
2662 return __blk_mq_issue_directly(hctx, rq, last);
2665 static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2667 struct blk_mq_hw_ctx *hctx = NULL;
2670 blk_status_t ret = BLK_STS_OK;
2672 while ((rq = rq_list_pop(&plug->mq_list))) {
2673 bool last = rq_list_empty(plug->mq_list);
2675 if (hctx != rq->mq_hctx) {
2677 blk_mq_commit_rqs(hctx, queued, false);
2683 ret = blk_mq_request_issue_directly(rq, last);
2688 case BLK_STS_RESOURCE:
2689 case BLK_STS_DEV_RESOURCE:
2690 blk_mq_request_bypass_insert(rq, 0);
2691 blk_mq_run_hw_queue(hctx, false);
2694 blk_mq_end_request(rq, ret);
2700 if (ret != BLK_STS_OK)
2701 blk_mq_commit_rqs(hctx, queued, false);
2704 static void __blk_mq_flush_plug_list(struct request_queue *q,
2705 struct blk_plug *plug)
2707 if (blk_queue_quiesced(q))
2709 q->mq_ops->queue_rqs(&plug->mq_list);
2712 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2714 struct blk_mq_hw_ctx *this_hctx = NULL;
2715 struct blk_mq_ctx *this_ctx = NULL;
2716 struct request *requeue_list = NULL;
2717 struct request **requeue_lastp = &requeue_list;
2718 unsigned int depth = 0;
2719 bool is_passthrough = false;
2723 struct request *rq = rq_list_pop(&plug->mq_list);
2726 this_hctx = rq->mq_hctx;
2727 this_ctx = rq->mq_ctx;
2728 is_passthrough = blk_rq_is_passthrough(rq);
2729 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2730 is_passthrough != blk_rq_is_passthrough(rq)) {
2731 rq_list_add_tail(&requeue_lastp, rq);
2734 list_add(&rq->queuelist, &list);
2736 } while (!rq_list_empty(plug->mq_list));
2738 plug->mq_list = requeue_list;
2739 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2741 percpu_ref_get(&this_hctx->queue->q_usage_counter);
2742 /* passthrough requests should never be issued to the I/O scheduler */
2743 if (this_hctx->queue->elevator && !is_passthrough) {
2744 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2746 blk_mq_run_hw_queue(this_hctx, from_sched);
2748 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2750 percpu_ref_put(&this_hctx->queue->q_usage_counter);
2753 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2757 if (rq_list_empty(plug->mq_list))
2761 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2762 struct request_queue *q;
2764 rq = rq_list_peek(&plug->mq_list);
2768 * Peek first request and see if we have a ->queue_rqs() hook.
2769 * If we do, we can dispatch the whole plug list in one go. We
2770 * already know at this point that all requests belong to the
2771 * same queue, caller must ensure that's the case.
2773 * Since we pass off the full list to the driver at this point,
2774 * we do not increment the active request count for the queue.
2775 * Bypass shared tags for now because of that.
2777 if (q->mq_ops->queue_rqs &&
2778 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2779 blk_mq_run_dispatch_ops(q,
2780 __blk_mq_flush_plug_list(q, plug));
2781 if (rq_list_empty(plug->mq_list))
2785 blk_mq_run_dispatch_ops(q,
2786 blk_mq_plug_issue_direct(plug));
2787 if (rq_list_empty(plug->mq_list))
2792 blk_mq_dispatch_plug_list(plug, from_schedule);
2793 } while (!rq_list_empty(plug->mq_list));
2796 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2797 struct list_head *list)
2800 blk_status_t ret = BLK_STS_OK;
2802 while (!list_empty(list)) {
2803 struct request *rq = list_first_entry(list, struct request,
2806 list_del_init(&rq->queuelist);
2807 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2812 case BLK_STS_RESOURCE:
2813 case BLK_STS_DEV_RESOURCE:
2814 blk_mq_request_bypass_insert(rq, 0);
2815 if (list_empty(list))
2816 blk_mq_run_hw_queue(hctx, false);
2819 blk_mq_end_request(rq, ret);
2825 if (ret != BLK_STS_OK)
2826 blk_mq_commit_rqs(hctx, queued, false);
2829 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2830 struct bio *bio, unsigned int nr_segs)
2832 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2833 if (blk_attempt_plug_merge(q, bio, nr_segs))
2835 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2841 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2842 struct blk_plug *plug,
2846 struct blk_mq_alloc_data data = {
2849 .cmd_flags = bio->bi_opf,
2853 if (unlikely(bio_queue_enter(bio)))
2856 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2859 rq_qos_throttle(q, bio);
2862 data.nr_tags = plug->nr_ios;
2864 data.cached_rq = &plug->cached_rq;
2867 rq = __blk_mq_alloc_requests(&data);
2870 rq_qos_cleanup(q, bio);
2871 if (bio->bi_opf & REQ_NOWAIT)
2872 bio_wouldblock_error(bio);
2878 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2879 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2882 enum hctx_type type, hctx_type;
2886 rq = rq_list_peek(&plug->cached_rq);
2887 if (!rq || rq->q != q)
2890 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2895 type = blk_mq_get_hctx_type((*bio)->bi_opf);
2896 hctx_type = rq->mq_hctx->type;
2897 if (type != hctx_type &&
2898 !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2900 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2904 * If any qos ->throttle() end up blocking, we will have flushed the
2905 * plug and hence killed the cached_rq list as well. Pop this entry
2906 * before we throttle.
2908 plug->cached_rq = rq_list_next(rq);
2909 rq_qos_throttle(q, *bio);
2911 rq->cmd_flags = (*bio)->bi_opf;
2912 INIT_LIST_HEAD(&rq->queuelist);
2916 static void bio_set_ioprio(struct bio *bio)
2918 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2919 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2920 bio->bi_ioprio = get_current_ioprio();
2921 blkcg_set_ioprio(bio);
2925 * blk_mq_submit_bio - Create and send a request to block device.
2926 * @bio: Bio pointer.
2928 * Builds up a request structure from @q and @bio and send to the device. The
2929 * request may not be queued directly to hardware if:
2930 * * This request can be merged with another one
2931 * * We want to place request at plug queue for possible future merging
2932 * * There is an IO scheduler active at this queue
2934 * It will not queue the request if there is an error with the bio, or at the
2937 void blk_mq_submit_bio(struct bio *bio)
2939 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2940 struct blk_plug *plug = blk_mq_plug(bio);
2941 const int is_sync = op_is_sync(bio->bi_opf);
2942 struct blk_mq_hw_ctx *hctx;
2944 unsigned int nr_segs = 1;
2947 bio = blk_queue_bounce(bio, q);
2948 if (bio_may_exceed_limits(bio, &q->limits)) {
2949 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2954 if (!bio_integrity_prep(bio))
2957 bio_set_ioprio(bio);
2959 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2963 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2968 trace_block_getrq(bio);
2970 rq_qos_track(q, rq, bio);
2972 blk_mq_bio_to_request(rq, bio, nr_segs);
2974 ret = blk_crypto_rq_get_keyslot(rq);
2975 if (ret != BLK_STS_OK) {
2976 bio->bi_status = ret;
2978 blk_mq_free_request(rq);
2982 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
2986 blk_add_rq_to_plug(plug, rq);
2991 if ((rq->rq_flags & RQF_USE_SCHED) ||
2992 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
2993 blk_mq_insert_request(rq, 0);
2994 blk_mq_run_hw_queue(hctx, true);
2996 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
3000 #ifdef CONFIG_BLK_MQ_STACKING
3002 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3003 * @rq: the request being queued
3005 blk_status_t blk_insert_cloned_request(struct request *rq)
3007 struct request_queue *q = rq->q;
3008 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3009 unsigned int max_segments = blk_rq_get_max_segments(rq);
3012 if (blk_rq_sectors(rq) > max_sectors) {
3014 * SCSI device does not have a good way to return if
3015 * Write Same/Zero is actually supported. If a device rejects
3016 * a non-read/write command (discard, write same,etc.) the
3017 * low-level device driver will set the relevant queue limit to
3018 * 0 to prevent blk-lib from issuing more of the offending
3019 * operations. Commands queued prior to the queue limit being
3020 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3021 * errors being propagated to upper layers.
3023 if (max_sectors == 0)
3024 return BLK_STS_NOTSUPP;
3026 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3027 __func__, blk_rq_sectors(rq), max_sectors);
3028 return BLK_STS_IOERR;
3032 * The queue settings related to segment counting may differ from the
3035 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3036 if (rq->nr_phys_segments > max_segments) {
3037 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3038 __func__, rq->nr_phys_segments, max_segments);
3039 return BLK_STS_IOERR;
3042 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3043 return BLK_STS_IOERR;
3045 ret = blk_crypto_rq_get_keyslot(rq);
3046 if (ret != BLK_STS_OK)
3049 blk_account_io_start(rq);
3052 * Since we have a scheduler attached on the top device,
3053 * bypass a potential scheduler on the bottom device for
3056 blk_mq_run_dispatch_ops(q,
3057 ret = blk_mq_request_issue_directly(rq, true));
3059 blk_account_io_done(rq, ktime_get_ns());
3062 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3065 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3066 * @rq: the clone request to be cleaned up
3069 * Free all bios in @rq for a cloned request.
3071 void blk_rq_unprep_clone(struct request *rq)
3075 while ((bio = rq->bio) != NULL) {
3076 rq->bio = bio->bi_next;
3081 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3084 * blk_rq_prep_clone - Helper function to setup clone request
3085 * @rq: the request to be setup
3086 * @rq_src: original request to be cloned
3087 * @bs: bio_set that bios for clone are allocated from
3088 * @gfp_mask: memory allocation mask for bio
3089 * @bio_ctr: setup function to be called for each clone bio.
3090 * Returns %0 for success, non %0 for failure.
3091 * @data: private data to be passed to @bio_ctr
3094 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3095 * Also, pages which the original bios are pointing to are not copied
3096 * and the cloned bios just point same pages.
3097 * So cloned bios must be completed before original bios, which means
3098 * the caller must complete @rq before @rq_src.
3100 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3101 struct bio_set *bs, gfp_t gfp_mask,
3102 int (*bio_ctr)(struct bio *, struct bio *, void *),
3105 struct bio *bio, *bio_src;
3110 __rq_for_each_bio(bio_src, rq_src) {
3111 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3116 if (bio_ctr && bio_ctr(bio, bio_src, data))
3120 rq->biotail->bi_next = bio;
3123 rq->bio = rq->biotail = bio;
3128 /* Copy attributes of the original request to the clone request. */
3129 rq->__sector = blk_rq_pos(rq_src);
3130 rq->__data_len = blk_rq_bytes(rq_src);
3131 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3132 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3133 rq->special_vec = rq_src->special_vec;
3135 rq->nr_phys_segments = rq_src->nr_phys_segments;
3136 rq->ioprio = rq_src->ioprio;
3138 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3146 blk_rq_unprep_clone(rq);
3150 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3151 #endif /* CONFIG_BLK_MQ_STACKING */
3154 * Steal bios from a request and add them to a bio list.
3155 * The request must not have been partially completed before.
3157 void blk_steal_bios(struct bio_list *list, struct request *rq)
3161 list->tail->bi_next = rq->bio;
3163 list->head = rq->bio;
3164 list->tail = rq->biotail;
3172 EXPORT_SYMBOL_GPL(blk_steal_bios);
3174 static size_t order_to_size(unsigned int order)
3176 return (size_t)PAGE_SIZE << order;
3179 /* called before freeing request pool in @tags */
3180 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3181 struct blk_mq_tags *tags)
3184 unsigned long flags;
3187 * There is no need to clear mapping if driver tags is not initialized
3188 * or the mapping belongs to the driver tags.
3190 if (!drv_tags || drv_tags == tags)
3193 list_for_each_entry(page, &tags->page_list, lru) {
3194 unsigned long start = (unsigned long)page_address(page);
3195 unsigned long end = start + order_to_size(page->private);
3198 for (i = 0; i < drv_tags->nr_tags; i++) {
3199 struct request *rq = drv_tags->rqs[i];
3200 unsigned long rq_addr = (unsigned long)rq;
3202 if (rq_addr >= start && rq_addr < end) {
3203 WARN_ON_ONCE(req_ref_read(rq) != 0);
3204 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3210 * Wait until all pending iteration is done.
3212 * Request reference is cleared and it is guaranteed to be observed
3213 * after the ->lock is released.
3215 spin_lock_irqsave(&drv_tags->lock, flags);
3216 spin_unlock_irqrestore(&drv_tags->lock, flags);
3219 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3220 unsigned int hctx_idx)
3222 struct blk_mq_tags *drv_tags;
3225 if (list_empty(&tags->page_list))
3228 if (blk_mq_is_shared_tags(set->flags))
3229 drv_tags = set->shared_tags;
3231 drv_tags = set->tags[hctx_idx];
3233 if (tags->static_rqs && set->ops->exit_request) {
3236 for (i = 0; i < tags->nr_tags; i++) {
3237 struct request *rq = tags->static_rqs[i];
3241 set->ops->exit_request(set, rq, hctx_idx);
3242 tags->static_rqs[i] = NULL;
3246 blk_mq_clear_rq_mapping(drv_tags, tags);
3248 while (!list_empty(&tags->page_list)) {
3249 page = list_first_entry(&tags->page_list, struct page, lru);
3250 list_del_init(&page->lru);
3252 * Remove kmemleak object previously allocated in
3253 * blk_mq_alloc_rqs().
3255 kmemleak_free(page_address(page));
3256 __free_pages(page, page->private);
3260 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3264 kfree(tags->static_rqs);
3265 tags->static_rqs = NULL;
3267 blk_mq_free_tags(tags);
3270 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3271 unsigned int hctx_idx)
3275 for (i = 0; i < set->nr_maps; i++) {
3276 unsigned int start = set->map[i].queue_offset;
3277 unsigned int end = start + set->map[i].nr_queues;
3279 if (hctx_idx >= start && hctx_idx < end)
3283 if (i >= set->nr_maps)
3284 i = HCTX_TYPE_DEFAULT;
3289 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3290 unsigned int hctx_idx)
3292 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3294 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3297 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3298 unsigned int hctx_idx,
3299 unsigned int nr_tags,
3300 unsigned int reserved_tags)
3302 int node = blk_mq_get_hctx_node(set, hctx_idx);
3303 struct blk_mq_tags *tags;
3305 if (node == NUMA_NO_NODE)
3306 node = set->numa_node;
3308 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3309 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3313 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3314 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3319 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3320 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3322 if (!tags->static_rqs)
3330 blk_mq_free_tags(tags);
3334 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3335 unsigned int hctx_idx, int node)
3339 if (set->ops->init_request) {
3340 ret = set->ops->init_request(set, rq, hctx_idx, node);
3345 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3349 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3350 struct blk_mq_tags *tags,
3351 unsigned int hctx_idx, unsigned int depth)
3353 unsigned int i, j, entries_per_page, max_order = 4;
3354 int node = blk_mq_get_hctx_node(set, hctx_idx);
3355 size_t rq_size, left;
3357 if (node == NUMA_NO_NODE)
3358 node = set->numa_node;
3360 INIT_LIST_HEAD(&tags->page_list);
3363 * rq_size is the size of the request plus driver payload, rounded
3364 * to the cacheline size
3366 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3368 left = rq_size * depth;
3370 for (i = 0; i < depth; ) {
3371 int this_order = max_order;
3376 while (this_order && left < order_to_size(this_order - 1))
3380 page = alloc_pages_node(node,
3381 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3387 if (order_to_size(this_order) < rq_size)
3394 page->private = this_order;
3395 list_add_tail(&page->lru, &tags->page_list);
3397 p = page_address(page);
3399 * Allow kmemleak to scan these pages as they contain pointers
3400 * to additional allocations like via ops->init_request().
3402 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3403 entries_per_page = order_to_size(this_order) / rq_size;
3404 to_do = min(entries_per_page, depth - i);
3405 left -= to_do * rq_size;
3406 for (j = 0; j < to_do; j++) {
3407 struct request *rq = p;
3409 tags->static_rqs[i] = rq;
3410 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3411 tags->static_rqs[i] = NULL;
3422 blk_mq_free_rqs(set, tags, hctx_idx);
3426 struct rq_iter_data {
3427 struct blk_mq_hw_ctx *hctx;
3431 static bool blk_mq_has_request(struct request *rq, void *data)
3433 struct rq_iter_data *iter_data = data;
3435 if (rq->mq_hctx != iter_data->hctx)
3437 iter_data->has_rq = true;
3441 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3443 struct blk_mq_tags *tags = hctx->sched_tags ?
3444 hctx->sched_tags : hctx->tags;
3445 struct rq_iter_data data = {
3449 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3453 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3454 struct blk_mq_hw_ctx *hctx)
3456 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3458 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3463 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3465 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3466 struct blk_mq_hw_ctx, cpuhp_online);
3468 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3469 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3473 * Prevent new request from being allocated on the current hctx.
3475 * The smp_mb__after_atomic() Pairs with the implied barrier in
3476 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3477 * seen once we return from the tag allocator.
3479 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3480 smp_mb__after_atomic();
3483 * Try to grab a reference to the queue and wait for any outstanding
3484 * requests. If we could not grab a reference the queue has been
3485 * frozen and there are no requests.
3487 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3488 while (blk_mq_hctx_has_requests(hctx))
3490 percpu_ref_put(&hctx->queue->q_usage_counter);
3496 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3498 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3499 struct blk_mq_hw_ctx, cpuhp_online);
3501 if (cpumask_test_cpu(cpu, hctx->cpumask))
3502 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3507 * 'cpu' is going away. splice any existing rq_list entries from this
3508 * software queue to the hw queue dispatch list, and ensure that it
3511 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3513 struct blk_mq_hw_ctx *hctx;
3514 struct blk_mq_ctx *ctx;
3516 enum hctx_type type;
3518 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3519 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3522 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3525 spin_lock(&ctx->lock);
3526 if (!list_empty(&ctx->rq_lists[type])) {
3527 list_splice_init(&ctx->rq_lists[type], &tmp);
3528 blk_mq_hctx_clear_pending(hctx, ctx);
3530 spin_unlock(&ctx->lock);
3532 if (list_empty(&tmp))
3535 spin_lock(&hctx->lock);
3536 list_splice_tail_init(&tmp, &hctx->dispatch);
3537 spin_unlock(&hctx->lock);
3539 blk_mq_run_hw_queue(hctx, true);
3543 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3545 if (!(hctx->flags & BLK_MQ_F_STACKING))
3546 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3547 &hctx->cpuhp_online);
3548 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3553 * Before freeing hw queue, clearing the flush request reference in
3554 * tags->rqs[] for avoiding potential UAF.
3556 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3557 unsigned int queue_depth, struct request *flush_rq)
3560 unsigned long flags;
3562 /* The hw queue may not be mapped yet */
3566 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3568 for (i = 0; i < queue_depth; i++)
3569 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3572 * Wait until all pending iteration is done.
3574 * Request reference is cleared and it is guaranteed to be observed
3575 * after the ->lock is released.
3577 spin_lock_irqsave(&tags->lock, flags);
3578 spin_unlock_irqrestore(&tags->lock, flags);
3581 /* hctx->ctxs will be freed in queue's release handler */
3582 static void blk_mq_exit_hctx(struct request_queue *q,
3583 struct blk_mq_tag_set *set,
3584 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3586 struct request *flush_rq = hctx->fq->flush_rq;
3588 if (blk_mq_hw_queue_mapped(hctx))
3589 blk_mq_tag_idle(hctx);
3591 if (blk_queue_init_done(q))
3592 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3593 set->queue_depth, flush_rq);
3594 if (set->ops->exit_request)
3595 set->ops->exit_request(set, flush_rq, hctx_idx);
3597 if (set->ops->exit_hctx)
3598 set->ops->exit_hctx(hctx, hctx_idx);
3600 blk_mq_remove_cpuhp(hctx);
3602 xa_erase(&q->hctx_table, hctx_idx);
3604 spin_lock(&q->unused_hctx_lock);
3605 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3606 spin_unlock(&q->unused_hctx_lock);
3609 static void blk_mq_exit_hw_queues(struct request_queue *q,
3610 struct blk_mq_tag_set *set, int nr_queue)
3612 struct blk_mq_hw_ctx *hctx;
3615 queue_for_each_hw_ctx(q, hctx, i) {
3618 blk_mq_exit_hctx(q, set, hctx, i);
3622 static int blk_mq_init_hctx(struct request_queue *q,
3623 struct blk_mq_tag_set *set,
3624 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3626 hctx->queue_num = hctx_idx;
3628 if (!(hctx->flags & BLK_MQ_F_STACKING))
3629 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3630 &hctx->cpuhp_online);
3631 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3633 hctx->tags = set->tags[hctx_idx];
3635 if (set->ops->init_hctx &&
3636 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3637 goto unregister_cpu_notifier;
3639 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3643 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3649 if (set->ops->exit_request)
3650 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3652 if (set->ops->exit_hctx)
3653 set->ops->exit_hctx(hctx, hctx_idx);
3654 unregister_cpu_notifier:
3655 blk_mq_remove_cpuhp(hctx);
3659 static struct blk_mq_hw_ctx *
3660 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3663 struct blk_mq_hw_ctx *hctx;
3664 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3666 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3668 goto fail_alloc_hctx;
3670 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3673 atomic_set(&hctx->nr_active, 0);
3674 if (node == NUMA_NO_NODE)
3675 node = set->numa_node;
3676 hctx->numa_node = node;
3678 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3679 spin_lock_init(&hctx->lock);
3680 INIT_LIST_HEAD(&hctx->dispatch);
3682 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3684 INIT_LIST_HEAD(&hctx->hctx_list);
3687 * Allocate space for all possible cpus to avoid allocation at
3690 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3695 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3696 gfp, node, false, false))
3700 spin_lock_init(&hctx->dispatch_wait_lock);
3701 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3702 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3704 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3708 blk_mq_hctx_kobj_init(hctx);
3713 sbitmap_free(&hctx->ctx_map);
3717 free_cpumask_var(hctx->cpumask);
3724 static void blk_mq_init_cpu_queues(struct request_queue *q,
3725 unsigned int nr_hw_queues)
3727 struct blk_mq_tag_set *set = q->tag_set;
3730 for_each_possible_cpu(i) {
3731 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3732 struct blk_mq_hw_ctx *hctx;
3736 spin_lock_init(&__ctx->lock);
3737 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3738 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3743 * Set local node, IFF we have more than one hw queue. If
3744 * not, we remain on the home node of the device
3746 for (j = 0; j < set->nr_maps; j++) {
3747 hctx = blk_mq_map_queue_type(q, j, i);
3748 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3749 hctx->numa_node = cpu_to_node(i);
3754 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3755 unsigned int hctx_idx,
3758 struct blk_mq_tags *tags;
3761 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3765 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3767 blk_mq_free_rq_map(tags);
3774 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3777 if (blk_mq_is_shared_tags(set->flags)) {
3778 set->tags[hctx_idx] = set->shared_tags;
3783 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3786 return set->tags[hctx_idx];
3789 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3790 struct blk_mq_tags *tags,
3791 unsigned int hctx_idx)
3794 blk_mq_free_rqs(set, tags, hctx_idx);
3795 blk_mq_free_rq_map(tags);
3799 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3800 unsigned int hctx_idx)
3802 if (!blk_mq_is_shared_tags(set->flags))
3803 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3805 set->tags[hctx_idx] = NULL;
3808 static void blk_mq_map_swqueue(struct request_queue *q)
3810 unsigned int j, hctx_idx;
3812 struct blk_mq_hw_ctx *hctx;
3813 struct blk_mq_ctx *ctx;
3814 struct blk_mq_tag_set *set = q->tag_set;
3816 queue_for_each_hw_ctx(q, hctx, i) {
3817 cpumask_clear(hctx->cpumask);
3819 hctx->dispatch_from = NULL;
3823 * Map software to hardware queues.
3825 * If the cpu isn't present, the cpu is mapped to first hctx.
3827 for_each_possible_cpu(i) {
3829 ctx = per_cpu_ptr(q->queue_ctx, i);
3830 for (j = 0; j < set->nr_maps; j++) {
3831 if (!set->map[j].nr_queues) {
3832 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3833 HCTX_TYPE_DEFAULT, i);
3836 hctx_idx = set->map[j].mq_map[i];
3837 /* unmapped hw queue can be remapped after CPU topo changed */
3838 if (!set->tags[hctx_idx] &&
3839 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3841 * If tags initialization fail for some hctx,
3842 * that hctx won't be brought online. In this
3843 * case, remap the current ctx to hctx[0] which
3844 * is guaranteed to always have tags allocated
3846 set->map[j].mq_map[i] = 0;
3849 hctx = blk_mq_map_queue_type(q, j, i);
3850 ctx->hctxs[j] = hctx;
3852 * If the CPU is already set in the mask, then we've
3853 * mapped this one already. This can happen if
3854 * devices share queues across queue maps.
3856 if (cpumask_test_cpu(i, hctx->cpumask))
3859 cpumask_set_cpu(i, hctx->cpumask);
3861 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3862 hctx->ctxs[hctx->nr_ctx++] = ctx;
3865 * If the nr_ctx type overflows, we have exceeded the
3866 * amount of sw queues we can support.
3868 BUG_ON(!hctx->nr_ctx);
3871 for (; j < HCTX_MAX_TYPES; j++)
3872 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3873 HCTX_TYPE_DEFAULT, i);
3876 queue_for_each_hw_ctx(q, hctx, i) {
3878 * If no software queues are mapped to this hardware queue,
3879 * disable it and free the request entries.
3881 if (!hctx->nr_ctx) {
3882 /* Never unmap queue 0. We need it as a
3883 * fallback in case of a new remap fails
3887 __blk_mq_free_map_and_rqs(set, i);
3893 hctx->tags = set->tags[i];
3894 WARN_ON(!hctx->tags);
3897 * Set the map size to the number of mapped software queues.
3898 * This is more accurate and more efficient than looping
3899 * over all possibly mapped software queues.
3901 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3904 * Initialize batch roundrobin counts
3906 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3907 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3912 * Caller needs to ensure that we're either frozen/quiesced, or that
3913 * the queue isn't live yet.
3915 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3917 struct blk_mq_hw_ctx *hctx;
3920 queue_for_each_hw_ctx(q, hctx, i) {
3922 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3924 blk_mq_tag_idle(hctx);
3925 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3930 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3933 struct request_queue *q;
3935 lockdep_assert_held(&set->tag_list_lock);
3937 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3938 blk_mq_freeze_queue(q);
3939 queue_set_hctx_shared(q, shared);
3940 blk_mq_unfreeze_queue(q);
3944 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3946 struct blk_mq_tag_set *set = q->tag_set;
3948 mutex_lock(&set->tag_list_lock);
3949 list_del(&q->tag_set_list);
3950 if (list_is_singular(&set->tag_list)) {
3951 /* just transitioned to unshared */
3952 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3953 /* update existing queue */
3954 blk_mq_update_tag_set_shared(set, false);
3956 mutex_unlock(&set->tag_list_lock);
3957 INIT_LIST_HEAD(&q->tag_set_list);
3960 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3961 struct request_queue *q)
3963 mutex_lock(&set->tag_list_lock);
3966 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3968 if (!list_empty(&set->tag_list) &&
3969 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3970 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3971 /* update existing queue */
3972 blk_mq_update_tag_set_shared(set, true);
3974 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3975 queue_set_hctx_shared(q, true);
3976 list_add_tail(&q->tag_set_list, &set->tag_list);
3978 mutex_unlock(&set->tag_list_lock);
3981 /* All allocations will be freed in release handler of q->mq_kobj */
3982 static int blk_mq_alloc_ctxs(struct request_queue *q)
3984 struct blk_mq_ctxs *ctxs;
3987 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3991 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3992 if (!ctxs->queue_ctx)
3995 for_each_possible_cpu(cpu) {
3996 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4000 q->mq_kobj = &ctxs->kobj;
4001 q->queue_ctx = ctxs->queue_ctx;
4010 * It is the actual release handler for mq, but we do it from
4011 * request queue's release handler for avoiding use-after-free
4012 * and headache because q->mq_kobj shouldn't have been introduced,
4013 * but we can't group ctx/kctx kobj without it.
4015 void blk_mq_release(struct request_queue *q)
4017 struct blk_mq_hw_ctx *hctx, *next;
4020 queue_for_each_hw_ctx(q, hctx, i)
4021 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4023 /* all hctx are in .unused_hctx_list now */
4024 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4025 list_del_init(&hctx->hctx_list);
4026 kobject_put(&hctx->kobj);
4029 xa_destroy(&q->hctx_table);
4032 * release .mq_kobj and sw queue's kobject now because
4033 * both share lifetime with request queue.
4035 blk_mq_sysfs_deinit(q);
4038 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4041 struct request_queue *q;
4044 q = blk_alloc_queue(set->numa_node);
4046 return ERR_PTR(-ENOMEM);
4047 q->queuedata = queuedata;
4048 ret = blk_mq_init_allocated_queue(set, q);
4051 return ERR_PTR(ret);
4056 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4058 return blk_mq_init_queue_data(set, NULL);
4060 EXPORT_SYMBOL(blk_mq_init_queue);
4063 * blk_mq_destroy_queue - shutdown a request queue
4064 * @q: request queue to shutdown
4066 * This shuts down a request queue allocated by blk_mq_init_queue(). All future
4067 * requests will be failed with -ENODEV. The caller is responsible for dropping
4068 * the reference from blk_mq_init_queue() by calling blk_put_queue().
4070 * Context: can sleep
4072 void blk_mq_destroy_queue(struct request_queue *q)
4074 WARN_ON_ONCE(!queue_is_mq(q));
4075 WARN_ON_ONCE(blk_queue_registered(q));
4079 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4080 blk_queue_start_drain(q);
4081 blk_mq_freeze_queue_wait(q);
4084 blk_mq_cancel_work_sync(q);
4085 blk_mq_exit_queue(q);
4087 EXPORT_SYMBOL(blk_mq_destroy_queue);
4089 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4090 struct lock_class_key *lkclass)
4092 struct request_queue *q;
4093 struct gendisk *disk;
4095 q = blk_mq_init_queue_data(set, queuedata);
4099 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4101 blk_mq_destroy_queue(q);
4103 return ERR_PTR(-ENOMEM);
4105 set_bit(GD_OWNS_QUEUE, &disk->state);
4108 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4110 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4111 struct lock_class_key *lkclass)
4113 struct gendisk *disk;
4115 if (!blk_get_queue(q))
4117 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4122 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4124 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4125 struct blk_mq_tag_set *set, struct request_queue *q,
4126 int hctx_idx, int node)
4128 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4130 /* reuse dead hctx first */
4131 spin_lock(&q->unused_hctx_lock);
4132 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4133 if (tmp->numa_node == node) {
4139 list_del_init(&hctx->hctx_list);
4140 spin_unlock(&q->unused_hctx_lock);
4143 hctx = blk_mq_alloc_hctx(q, set, node);
4147 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4153 kobject_put(&hctx->kobj);
4158 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4159 struct request_queue *q)
4161 struct blk_mq_hw_ctx *hctx;
4164 /* protect against switching io scheduler */
4165 mutex_lock(&q->sysfs_lock);
4166 for (i = 0; i < set->nr_hw_queues; i++) {
4168 int node = blk_mq_get_hctx_node(set, i);
4169 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4172 old_node = old_hctx->numa_node;
4173 blk_mq_exit_hctx(q, set, old_hctx, i);
4176 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4179 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4181 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4182 WARN_ON_ONCE(!hctx);
4186 * Increasing nr_hw_queues fails. Free the newly allocated
4187 * hctxs and keep the previous q->nr_hw_queues.
4189 if (i != set->nr_hw_queues) {
4190 j = q->nr_hw_queues;
4193 q->nr_hw_queues = set->nr_hw_queues;
4196 xa_for_each_start(&q->hctx_table, j, hctx, j)
4197 blk_mq_exit_hctx(q, set, hctx, j);
4198 mutex_unlock(&q->sysfs_lock);
4201 static void blk_mq_update_poll_flag(struct request_queue *q)
4203 struct blk_mq_tag_set *set = q->tag_set;
4205 if (set->nr_maps > HCTX_TYPE_POLL &&
4206 set->map[HCTX_TYPE_POLL].nr_queues)
4207 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4209 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4212 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4213 struct request_queue *q)
4215 /* mark the queue as mq asap */
4216 q->mq_ops = set->ops;
4218 if (blk_mq_alloc_ctxs(q))
4221 /* init q->mq_kobj and sw queues' kobjects */
4222 blk_mq_sysfs_init(q);
4224 INIT_LIST_HEAD(&q->unused_hctx_list);
4225 spin_lock_init(&q->unused_hctx_lock);
4227 xa_init(&q->hctx_table);
4229 blk_mq_realloc_hw_ctxs(set, q);
4230 if (!q->nr_hw_queues)
4233 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4234 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4238 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4239 blk_mq_update_poll_flag(q);
4241 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4242 INIT_LIST_HEAD(&q->requeue_list);
4243 spin_lock_init(&q->requeue_lock);
4245 q->nr_requests = set->queue_depth;
4247 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4248 blk_mq_add_queue_tag_set(set, q);
4249 blk_mq_map_swqueue(q);
4258 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4260 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4261 void blk_mq_exit_queue(struct request_queue *q)
4263 struct blk_mq_tag_set *set = q->tag_set;
4265 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4266 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4267 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4268 blk_mq_del_queue_tag_set(q);
4271 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4275 if (blk_mq_is_shared_tags(set->flags)) {
4276 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4279 if (!set->shared_tags)
4283 for (i = 0; i < set->nr_hw_queues; i++) {
4284 if (!__blk_mq_alloc_map_and_rqs(set, i))
4293 __blk_mq_free_map_and_rqs(set, i);
4295 if (blk_mq_is_shared_tags(set->flags)) {
4296 blk_mq_free_map_and_rqs(set, set->shared_tags,
4297 BLK_MQ_NO_HCTX_IDX);
4304 * Allocate the request maps associated with this tag_set. Note that this
4305 * may reduce the depth asked for, if memory is tight. set->queue_depth
4306 * will be updated to reflect the allocated depth.
4308 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4313 depth = set->queue_depth;
4315 err = __blk_mq_alloc_rq_maps(set);
4319 set->queue_depth >>= 1;
4320 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4324 } while (set->queue_depth);
4326 if (!set->queue_depth || err) {
4327 pr_err("blk-mq: failed to allocate request map\n");
4331 if (depth != set->queue_depth)
4332 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4333 depth, set->queue_depth);
4338 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4341 * blk_mq_map_queues() and multiple .map_queues() implementations
4342 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4343 * number of hardware queues.
4345 if (set->nr_maps == 1)
4346 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4348 if (set->ops->map_queues && !is_kdump_kernel()) {
4352 * transport .map_queues is usually done in the following
4355 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4356 * mask = get_cpu_mask(queue)
4357 * for_each_cpu(cpu, mask)
4358 * set->map[x].mq_map[cpu] = queue;
4361 * When we need to remap, the table has to be cleared for
4362 * killing stale mapping since one CPU may not be mapped
4365 for (i = 0; i < set->nr_maps; i++)
4366 blk_mq_clear_mq_map(&set->map[i]);
4368 set->ops->map_queues(set);
4370 BUG_ON(set->nr_maps > 1);
4371 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4375 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4376 int new_nr_hw_queues)
4378 struct blk_mq_tags **new_tags;
4380 if (set->nr_hw_queues >= new_nr_hw_queues)
4383 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4384 GFP_KERNEL, set->numa_node);
4389 memcpy(new_tags, set->tags, set->nr_hw_queues *
4390 sizeof(*set->tags));
4392 set->tags = new_tags;
4394 set->nr_hw_queues = new_nr_hw_queues;
4399 * Alloc a tag set to be associated with one or more request queues.
4400 * May fail with EINVAL for various error conditions. May adjust the
4401 * requested depth down, if it's too large. In that case, the set
4402 * value will be stored in set->queue_depth.
4404 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4408 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4410 if (!set->nr_hw_queues)
4412 if (!set->queue_depth)
4414 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4417 if (!set->ops->queue_rq)
4420 if (!set->ops->get_budget ^ !set->ops->put_budget)
4423 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4424 pr_info("blk-mq: reduced tag depth to %u\n",
4426 set->queue_depth = BLK_MQ_MAX_DEPTH;
4431 else if (set->nr_maps > HCTX_MAX_TYPES)
4435 * If a crashdump is active, then we are potentially in a very
4436 * memory constrained environment. Limit us to 1 queue and
4437 * 64 tags to prevent using too much memory.
4439 if (is_kdump_kernel()) {
4440 set->nr_hw_queues = 1;
4442 set->queue_depth = min(64U, set->queue_depth);
4445 * There is no use for more h/w queues than cpus if we just have
4448 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4449 set->nr_hw_queues = nr_cpu_ids;
4451 if (set->flags & BLK_MQ_F_BLOCKING) {
4452 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4455 ret = init_srcu_struct(set->srcu);
4461 set->tags = kcalloc_node(set->nr_hw_queues,
4462 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4465 goto out_cleanup_srcu;
4467 for (i = 0; i < set->nr_maps; i++) {
4468 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4469 sizeof(set->map[i].mq_map[0]),
4470 GFP_KERNEL, set->numa_node);
4471 if (!set->map[i].mq_map)
4472 goto out_free_mq_map;
4473 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4476 blk_mq_update_queue_map(set);
4478 ret = blk_mq_alloc_set_map_and_rqs(set);
4480 goto out_free_mq_map;
4482 mutex_init(&set->tag_list_lock);
4483 INIT_LIST_HEAD(&set->tag_list);
4488 for (i = 0; i < set->nr_maps; i++) {
4489 kfree(set->map[i].mq_map);
4490 set->map[i].mq_map = NULL;
4495 if (set->flags & BLK_MQ_F_BLOCKING)
4496 cleanup_srcu_struct(set->srcu);
4498 if (set->flags & BLK_MQ_F_BLOCKING)
4502 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4504 /* allocate and initialize a tagset for a simple single-queue device */
4505 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4506 const struct blk_mq_ops *ops, unsigned int queue_depth,
4507 unsigned int set_flags)
4509 memset(set, 0, sizeof(*set));
4511 set->nr_hw_queues = 1;
4513 set->queue_depth = queue_depth;
4514 set->numa_node = NUMA_NO_NODE;
4515 set->flags = set_flags;
4516 return blk_mq_alloc_tag_set(set);
4518 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4520 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4524 for (i = 0; i < set->nr_hw_queues; i++)
4525 __blk_mq_free_map_and_rqs(set, i);
4527 if (blk_mq_is_shared_tags(set->flags)) {
4528 blk_mq_free_map_and_rqs(set, set->shared_tags,
4529 BLK_MQ_NO_HCTX_IDX);
4532 for (j = 0; j < set->nr_maps; j++) {
4533 kfree(set->map[j].mq_map);
4534 set->map[j].mq_map = NULL;
4539 if (set->flags & BLK_MQ_F_BLOCKING) {
4540 cleanup_srcu_struct(set->srcu);
4544 EXPORT_SYMBOL(blk_mq_free_tag_set);
4546 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4548 struct blk_mq_tag_set *set = q->tag_set;
4549 struct blk_mq_hw_ctx *hctx;
4556 if (q->nr_requests == nr)
4559 blk_mq_freeze_queue(q);
4560 blk_mq_quiesce_queue(q);
4563 queue_for_each_hw_ctx(q, hctx, i) {
4567 * If we're using an MQ scheduler, just update the scheduler
4568 * queue depth. This is similar to what the old code would do.
4570 if (hctx->sched_tags) {
4571 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4574 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4579 if (q->elevator && q->elevator->type->ops.depth_updated)
4580 q->elevator->type->ops.depth_updated(hctx);
4583 q->nr_requests = nr;
4584 if (blk_mq_is_shared_tags(set->flags)) {
4586 blk_mq_tag_update_sched_shared_tags(q);
4588 blk_mq_tag_resize_shared_tags(set, nr);
4592 blk_mq_unquiesce_queue(q);
4593 blk_mq_unfreeze_queue(q);
4599 * request_queue and elevator_type pair.
4600 * It is just used by __blk_mq_update_nr_hw_queues to cache
4601 * the elevator_type associated with a request_queue.
4603 struct blk_mq_qe_pair {
4604 struct list_head node;
4605 struct request_queue *q;
4606 struct elevator_type *type;
4610 * Cache the elevator_type in qe pair list and switch the
4611 * io scheduler to 'none'
4613 static bool blk_mq_elv_switch_none(struct list_head *head,
4614 struct request_queue *q)
4616 struct blk_mq_qe_pair *qe;
4621 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4625 /* q->elevator needs protection from ->sysfs_lock */
4626 mutex_lock(&q->sysfs_lock);
4628 INIT_LIST_HEAD(&qe->node);
4630 qe->type = q->elevator->type;
4631 /* keep a reference to the elevator module as we'll switch back */
4632 __elevator_get(qe->type);
4633 list_add(&qe->node, head);
4634 elevator_disable(q);
4635 mutex_unlock(&q->sysfs_lock);
4640 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4641 struct request_queue *q)
4643 struct blk_mq_qe_pair *qe;
4645 list_for_each_entry(qe, head, node)
4652 static void blk_mq_elv_switch_back(struct list_head *head,
4653 struct request_queue *q)
4655 struct blk_mq_qe_pair *qe;
4656 struct elevator_type *t;
4658 qe = blk_lookup_qe_pair(head, q);
4662 list_del(&qe->node);
4665 mutex_lock(&q->sysfs_lock);
4666 elevator_switch(q, t);
4667 /* drop the reference acquired in blk_mq_elv_switch_none */
4669 mutex_unlock(&q->sysfs_lock);
4672 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4675 struct request_queue *q;
4677 int prev_nr_hw_queues;
4679 lockdep_assert_held(&set->tag_list_lock);
4681 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4682 nr_hw_queues = nr_cpu_ids;
4683 if (nr_hw_queues < 1)
4685 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4688 list_for_each_entry(q, &set->tag_list, tag_set_list)
4689 blk_mq_freeze_queue(q);
4691 * Switch IO scheduler to 'none', cleaning up the data associated
4692 * with the previous scheduler. We will switch back once we are done
4693 * updating the new sw to hw queue mappings.
4695 list_for_each_entry(q, &set->tag_list, tag_set_list)
4696 if (!blk_mq_elv_switch_none(&head, q))
4699 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4700 blk_mq_debugfs_unregister_hctxs(q);
4701 blk_mq_sysfs_unregister_hctxs(q);
4704 prev_nr_hw_queues = set->nr_hw_queues;
4705 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4709 blk_mq_update_queue_map(set);
4710 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4711 blk_mq_realloc_hw_ctxs(set, q);
4712 blk_mq_update_poll_flag(q);
4713 if (q->nr_hw_queues != set->nr_hw_queues) {
4714 int i = prev_nr_hw_queues;
4716 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4717 nr_hw_queues, prev_nr_hw_queues);
4718 for (; i < set->nr_hw_queues; i++)
4719 __blk_mq_free_map_and_rqs(set, i);
4721 set->nr_hw_queues = prev_nr_hw_queues;
4722 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4725 blk_mq_map_swqueue(q);
4729 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4730 blk_mq_sysfs_register_hctxs(q);
4731 blk_mq_debugfs_register_hctxs(q);
4735 list_for_each_entry(q, &set->tag_list, tag_set_list)
4736 blk_mq_elv_switch_back(&head, q);
4738 list_for_each_entry(q, &set->tag_list, tag_set_list)
4739 blk_mq_unfreeze_queue(q);
4742 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4744 mutex_lock(&set->tag_list_lock);
4745 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4746 mutex_unlock(&set->tag_list_lock);
4748 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4750 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4753 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4754 long state = get_current_state();
4758 ret = q->mq_ops->poll(hctx, iob);
4760 __set_current_state(TASK_RUNNING);
4764 if (signal_pending_state(state, current))
4765 __set_current_state(TASK_RUNNING);
4766 if (task_is_running(current))
4769 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4772 } while (!need_resched());
4774 __set_current_state(TASK_RUNNING);
4778 unsigned int blk_mq_rq_cpu(struct request *rq)
4780 return rq->mq_ctx->cpu;
4782 EXPORT_SYMBOL(blk_mq_rq_cpu);
4784 void blk_mq_cancel_work_sync(struct request_queue *q)
4786 struct blk_mq_hw_ctx *hctx;
4789 cancel_delayed_work_sync(&q->requeue_work);
4791 queue_for_each_hw_ctx(q, hctx, i)
4792 cancel_delayed_work_sync(&hctx->run_work);
4795 static int __init blk_mq_init(void)
4799 for_each_possible_cpu(i)
4800 init_llist_head(&per_cpu(blk_cpu_done, i));
4801 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4803 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4804 "block/softirq:dead", NULL,
4805 blk_softirq_cpu_dead);
4806 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4807 blk_mq_hctx_notify_dead);
4808 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4809 blk_mq_hctx_notify_online,
4810 blk_mq_hctx_notify_offline);
4813 subsys_initcall(blk_mq_init);