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)
965 trace_block_io_done(req);
968 * Account IO completion. flush_rq isn't accounted as a
969 * normal IO on queueing nor completion. Accounting the
970 * containing request is enough.
972 if (blk_do_io_stat(req) && req->part &&
973 !(req->rq_flags & RQF_FLUSH_SEQ)) {
974 const int sgrp = op_stat_group(req_op(req));
977 update_io_ticks(req->part, jiffies, true);
978 part_stat_inc(req->part, ios[sgrp]);
979 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
984 static inline void blk_account_io_start(struct request *req)
986 trace_block_io_start(req);
988 if (blk_do_io_stat(req)) {
990 * All non-passthrough requests are created from a bio with one
991 * exception: when a flush command that is part of a flush sequence
992 * generated by the state machine in blk-flush.c is cloned onto the
993 * lower device by dm-multipath we can get here without a bio.
996 req->part = req->bio->bi_bdev;
998 req->part = req->q->disk->part0;
1001 update_io_ticks(req->part, jiffies, false);
1006 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1008 if (rq->rq_flags & RQF_STATS)
1009 blk_stat_add(rq, now);
1011 blk_mq_sched_completed_request(rq, now);
1012 blk_account_io_done(rq, now);
1015 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1017 if (blk_mq_need_time_stamp(rq))
1018 __blk_mq_end_request_acct(rq, ktime_get_ns());
1021 rq_qos_done(rq->q, rq);
1022 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1023 blk_mq_free_request(rq);
1025 blk_mq_free_request(rq);
1028 EXPORT_SYMBOL(__blk_mq_end_request);
1030 void blk_mq_end_request(struct request *rq, blk_status_t error)
1032 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1034 __blk_mq_end_request(rq, error);
1036 EXPORT_SYMBOL(blk_mq_end_request);
1038 #define TAG_COMP_BATCH 32
1040 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1041 int *tag_array, int nr_tags)
1043 struct request_queue *q = hctx->queue;
1046 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1047 * update hctx->nr_active in batch
1049 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1050 __blk_mq_sub_active_requests(hctx, nr_tags);
1052 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1053 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1056 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1058 int tags[TAG_COMP_BATCH], nr_tags = 0;
1059 struct blk_mq_hw_ctx *cur_hctx = NULL;
1064 now = ktime_get_ns();
1066 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1068 prefetch(rq->rq_next);
1070 blk_complete_request(rq);
1072 __blk_mq_end_request_acct(rq, now);
1074 rq_qos_done(rq->q, rq);
1077 * If end_io handler returns NONE, then it still has
1078 * ownership of the request.
1080 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1083 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1084 if (!req_ref_put_and_test(rq))
1087 blk_crypto_free_request(rq);
1088 blk_pm_mark_last_busy(rq);
1090 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1092 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1094 cur_hctx = rq->mq_hctx;
1096 tags[nr_tags++] = rq->tag;
1100 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1102 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1104 static void blk_complete_reqs(struct llist_head *list)
1106 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1107 struct request *rq, *next;
1109 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1110 rq->q->mq_ops->complete(rq);
1113 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1115 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1118 static int blk_softirq_cpu_dead(unsigned int cpu)
1120 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1124 static void __blk_mq_complete_request_remote(void *data)
1126 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1129 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1131 int cpu = raw_smp_processor_id();
1133 if (!IS_ENABLED(CONFIG_SMP) ||
1134 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1137 * With force threaded interrupts enabled, raising softirq from an SMP
1138 * function call will always result in waking the ksoftirqd thread.
1139 * This is probably worse than completing the request on a different
1142 if (force_irqthreads())
1145 /* same CPU or cache domain? Complete locally */
1146 if (cpu == rq->mq_ctx->cpu ||
1147 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1148 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1151 /* don't try to IPI to an offline CPU */
1152 return cpu_online(rq->mq_ctx->cpu);
1155 static void blk_mq_complete_send_ipi(struct request *rq)
1157 struct llist_head *list;
1160 cpu = rq->mq_ctx->cpu;
1161 list = &per_cpu(blk_cpu_done, cpu);
1162 if (llist_add(&rq->ipi_list, list)) {
1163 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1164 smp_call_function_single_async(cpu, &rq->csd);
1168 static void blk_mq_raise_softirq(struct request *rq)
1170 struct llist_head *list;
1173 list = this_cpu_ptr(&blk_cpu_done);
1174 if (llist_add(&rq->ipi_list, list))
1175 raise_softirq(BLOCK_SOFTIRQ);
1179 bool blk_mq_complete_request_remote(struct request *rq)
1181 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1184 * For request which hctx has only one ctx mapping,
1185 * or a polled request, always complete locally,
1186 * it's pointless to redirect the completion.
1188 if (rq->mq_hctx->nr_ctx == 1 ||
1189 rq->cmd_flags & REQ_POLLED)
1192 if (blk_mq_complete_need_ipi(rq)) {
1193 blk_mq_complete_send_ipi(rq);
1197 if (rq->q->nr_hw_queues == 1) {
1198 blk_mq_raise_softirq(rq);
1203 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1206 * blk_mq_complete_request - end I/O on a request
1207 * @rq: the request being processed
1210 * Complete a request by scheduling the ->complete_rq operation.
1212 void blk_mq_complete_request(struct request *rq)
1214 if (!blk_mq_complete_request_remote(rq))
1215 rq->q->mq_ops->complete(rq);
1217 EXPORT_SYMBOL(blk_mq_complete_request);
1220 * blk_mq_start_request - Start processing a request
1221 * @rq: Pointer to request to be started
1223 * Function used by device drivers to notify the block layer that a request
1224 * is going to be processed now, so blk layer can do proper initializations
1225 * such as starting the timeout timer.
1227 void blk_mq_start_request(struct request *rq)
1229 struct request_queue *q = rq->q;
1231 trace_block_rq_issue(rq);
1233 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1234 rq->io_start_time_ns = ktime_get_ns();
1235 rq->stats_sectors = blk_rq_sectors(rq);
1236 rq->rq_flags |= RQF_STATS;
1237 rq_qos_issue(q, rq);
1240 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1243 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1245 #ifdef CONFIG_BLK_DEV_INTEGRITY
1246 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1247 q->integrity.profile->prepare_fn(rq);
1249 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1250 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1252 EXPORT_SYMBOL(blk_mq_start_request);
1255 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1256 * queues. This is important for md arrays to benefit from merging
1259 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1261 if (plug->multiple_queues)
1262 return BLK_MAX_REQUEST_COUNT * 2;
1263 return BLK_MAX_REQUEST_COUNT;
1266 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1268 struct request *last = rq_list_peek(&plug->mq_list);
1270 if (!plug->rq_count) {
1271 trace_block_plug(rq->q);
1272 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1273 (!blk_queue_nomerges(rq->q) &&
1274 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1275 blk_mq_flush_plug_list(plug, false);
1277 trace_block_plug(rq->q);
1280 if (!plug->multiple_queues && last && last->q != rq->q)
1281 plug->multiple_queues = true;
1282 if (!plug->has_elevator && (rq->rq_flags & RQF_USE_SCHED))
1283 plug->has_elevator = true;
1285 rq_list_add(&plug->mq_list, rq);
1290 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1291 * @rq: request to insert
1292 * @at_head: insert request at head or tail of queue
1295 * Insert a fully prepared request at the back of the I/O scheduler queue
1296 * for execution. Don't wait for completion.
1299 * This function will invoke @done directly if the queue is dead.
1301 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1303 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1305 WARN_ON(irqs_disabled());
1306 WARN_ON(!blk_rq_is_passthrough(rq));
1308 blk_account_io_start(rq);
1311 * As plugging can be enabled for passthrough requests on a zoned
1312 * device, directly accessing the plug instead of using blk_mq_plug()
1313 * should not have any consequences.
1315 if (current->plug && !at_head) {
1316 blk_add_rq_to_plug(current->plug, rq);
1320 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1321 blk_mq_run_hw_queue(hctx, false);
1323 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1325 struct blk_rq_wait {
1326 struct completion done;
1330 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1332 struct blk_rq_wait *wait = rq->end_io_data;
1335 complete(&wait->done);
1336 return RQ_END_IO_NONE;
1339 bool blk_rq_is_poll(struct request *rq)
1343 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1347 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1349 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1352 blk_mq_poll(rq->q, blk_rq_to_qc(rq), NULL, 0);
1354 } while (!completion_done(wait));
1358 * blk_execute_rq - insert a request into queue for execution
1359 * @rq: request to insert
1360 * @at_head: insert request at head or tail of queue
1363 * Insert a fully prepared request at the back of the I/O scheduler queue
1364 * for execution and wait for completion.
1365 * Return: The blk_status_t result provided to blk_mq_end_request().
1367 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1369 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1370 struct blk_rq_wait wait = {
1371 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1374 WARN_ON(irqs_disabled());
1375 WARN_ON(!blk_rq_is_passthrough(rq));
1377 rq->end_io_data = &wait;
1378 rq->end_io = blk_end_sync_rq;
1380 blk_account_io_start(rq);
1381 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1382 blk_mq_run_hw_queue(hctx, false);
1384 if (blk_rq_is_poll(rq)) {
1385 blk_rq_poll_completion(rq, &wait.done);
1388 * Prevent hang_check timer from firing at us during very long
1391 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1394 while (!wait_for_completion_io_timeout(&wait.done,
1395 hang_check * (HZ/2)))
1398 wait_for_completion_io(&wait.done);
1403 EXPORT_SYMBOL(blk_execute_rq);
1405 static void __blk_mq_requeue_request(struct request *rq)
1407 struct request_queue *q = rq->q;
1409 blk_mq_put_driver_tag(rq);
1411 trace_block_rq_requeue(rq);
1412 rq_qos_requeue(q, rq);
1414 if (blk_mq_request_started(rq)) {
1415 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1416 rq->rq_flags &= ~RQF_TIMED_OUT;
1420 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1422 struct request_queue *q = rq->q;
1423 unsigned long flags;
1425 __blk_mq_requeue_request(rq);
1427 /* this request will be re-inserted to io scheduler queue */
1428 blk_mq_sched_requeue_request(rq);
1430 spin_lock_irqsave(&q->requeue_lock, flags);
1431 list_add_tail(&rq->queuelist, &q->requeue_list);
1432 spin_unlock_irqrestore(&q->requeue_lock, flags);
1434 if (kick_requeue_list)
1435 blk_mq_kick_requeue_list(q);
1437 EXPORT_SYMBOL(blk_mq_requeue_request);
1439 static void blk_mq_requeue_work(struct work_struct *work)
1441 struct request_queue *q =
1442 container_of(work, struct request_queue, requeue_work.work);
1444 LIST_HEAD(flush_list);
1447 spin_lock_irq(&q->requeue_lock);
1448 list_splice_init(&q->requeue_list, &rq_list);
1449 list_splice_init(&q->flush_list, &flush_list);
1450 spin_unlock_irq(&q->requeue_lock);
1452 while (!list_empty(&rq_list)) {
1453 rq = list_entry(rq_list.next, struct request, queuelist);
1455 * If RQF_DONTPREP ist set, the request has been started by the
1456 * driver already and might have driver-specific data allocated
1457 * already. Insert it into the hctx dispatch list to avoid
1458 * block layer merges for the request.
1460 if (rq->rq_flags & RQF_DONTPREP) {
1461 list_del_init(&rq->queuelist);
1462 blk_mq_request_bypass_insert(rq, 0);
1464 list_del_init(&rq->queuelist);
1465 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1469 while (!list_empty(&flush_list)) {
1470 rq = list_entry(flush_list.next, struct request, queuelist);
1471 list_del_init(&rq->queuelist);
1472 blk_mq_insert_request(rq, 0);
1475 blk_mq_run_hw_queues(q, false);
1478 void blk_mq_kick_requeue_list(struct request_queue *q)
1480 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1482 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1484 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1485 unsigned long msecs)
1487 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1488 msecs_to_jiffies(msecs));
1490 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1492 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1495 * If we find a request that isn't idle we know the queue is busy
1496 * as it's checked in the iter.
1497 * Return false to stop the iteration.
1499 if (blk_mq_request_started(rq)) {
1509 bool blk_mq_queue_inflight(struct request_queue *q)
1513 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1516 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1518 static void blk_mq_rq_timed_out(struct request *req)
1520 req->rq_flags |= RQF_TIMED_OUT;
1521 if (req->q->mq_ops->timeout) {
1522 enum blk_eh_timer_return ret;
1524 ret = req->q->mq_ops->timeout(req);
1525 if (ret == BLK_EH_DONE)
1527 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1533 struct blk_expired_data {
1534 bool has_timedout_rq;
1536 unsigned long timeout_start;
1539 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1541 unsigned long deadline;
1543 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1545 if (rq->rq_flags & RQF_TIMED_OUT)
1548 deadline = READ_ONCE(rq->deadline);
1549 if (time_after_eq(expired->timeout_start, deadline))
1552 if (expired->next == 0)
1553 expired->next = deadline;
1554 else if (time_after(expired->next, deadline))
1555 expired->next = deadline;
1559 void blk_mq_put_rq_ref(struct request *rq)
1561 if (is_flush_rq(rq)) {
1562 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1563 blk_mq_free_request(rq);
1564 } else if (req_ref_put_and_test(rq)) {
1565 __blk_mq_free_request(rq);
1569 static bool blk_mq_check_expired(struct request *rq, void *priv)
1571 struct blk_expired_data *expired = priv;
1574 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1575 * be reallocated underneath the timeout handler's processing, then
1576 * the expire check is reliable. If the request is not expired, then
1577 * it was completed and reallocated as a new request after returning
1578 * from blk_mq_check_expired().
1580 if (blk_mq_req_expired(rq, expired)) {
1581 expired->has_timedout_rq = true;
1587 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1589 struct blk_expired_data *expired = priv;
1591 if (blk_mq_req_expired(rq, expired))
1592 blk_mq_rq_timed_out(rq);
1596 static void blk_mq_timeout_work(struct work_struct *work)
1598 struct request_queue *q =
1599 container_of(work, struct request_queue, timeout_work);
1600 struct blk_expired_data expired = {
1601 .timeout_start = jiffies,
1603 struct blk_mq_hw_ctx *hctx;
1606 /* A deadlock might occur if a request is stuck requiring a
1607 * timeout at the same time a queue freeze is waiting
1608 * completion, since the timeout code would not be able to
1609 * acquire the queue reference here.
1611 * That's why we don't use blk_queue_enter here; instead, we use
1612 * percpu_ref_tryget directly, because we need to be able to
1613 * obtain a reference even in the short window between the queue
1614 * starting to freeze, by dropping the first reference in
1615 * blk_freeze_queue_start, and the moment the last request is
1616 * consumed, marked by the instant q_usage_counter reaches
1619 if (!percpu_ref_tryget(&q->q_usage_counter))
1622 /* check if there is any timed-out request */
1623 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1624 if (expired.has_timedout_rq) {
1626 * Before walking tags, we must ensure any submit started
1627 * before the current time has finished. Since the submit
1628 * uses srcu or rcu, wait for a synchronization point to
1629 * ensure all running submits have finished
1631 blk_mq_wait_quiesce_done(q->tag_set);
1634 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1637 if (expired.next != 0) {
1638 mod_timer(&q->timeout, expired.next);
1641 * Request timeouts are handled as a forward rolling timer. If
1642 * we end up here it means that no requests are pending and
1643 * also that no request has been pending for a while. Mark
1644 * each hctx as idle.
1646 queue_for_each_hw_ctx(q, hctx, i) {
1647 /* the hctx may be unmapped, so check it here */
1648 if (blk_mq_hw_queue_mapped(hctx))
1649 blk_mq_tag_idle(hctx);
1655 struct flush_busy_ctx_data {
1656 struct blk_mq_hw_ctx *hctx;
1657 struct list_head *list;
1660 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1662 struct flush_busy_ctx_data *flush_data = data;
1663 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1664 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1665 enum hctx_type type = hctx->type;
1667 spin_lock(&ctx->lock);
1668 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1669 sbitmap_clear_bit(sb, bitnr);
1670 spin_unlock(&ctx->lock);
1675 * Process software queues that have been marked busy, splicing them
1676 * to the for-dispatch
1678 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1680 struct flush_busy_ctx_data data = {
1685 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1687 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1689 struct dispatch_rq_data {
1690 struct blk_mq_hw_ctx *hctx;
1694 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1697 struct dispatch_rq_data *dispatch_data = data;
1698 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1699 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1700 enum hctx_type type = hctx->type;
1702 spin_lock(&ctx->lock);
1703 if (!list_empty(&ctx->rq_lists[type])) {
1704 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1705 list_del_init(&dispatch_data->rq->queuelist);
1706 if (list_empty(&ctx->rq_lists[type]))
1707 sbitmap_clear_bit(sb, bitnr);
1709 spin_unlock(&ctx->lock);
1711 return !dispatch_data->rq;
1714 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1715 struct blk_mq_ctx *start)
1717 unsigned off = start ? start->index_hw[hctx->type] : 0;
1718 struct dispatch_rq_data data = {
1723 __sbitmap_for_each_set(&hctx->ctx_map, off,
1724 dispatch_rq_from_ctx, &data);
1729 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1731 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1732 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1735 blk_mq_tag_busy(rq->mq_hctx);
1737 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1738 bt = &rq->mq_hctx->tags->breserved_tags;
1741 if (!hctx_may_queue(rq->mq_hctx, bt))
1745 tag = __sbitmap_queue_get(bt);
1746 if (tag == BLK_MQ_NO_TAG)
1749 rq->tag = tag + tag_offset;
1753 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1755 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1758 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1759 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1760 rq->rq_flags |= RQF_MQ_INFLIGHT;
1761 __blk_mq_inc_active_requests(hctx);
1763 hctx->tags->rqs[rq->tag] = rq;
1767 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1768 int flags, void *key)
1770 struct blk_mq_hw_ctx *hctx;
1772 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1774 spin_lock(&hctx->dispatch_wait_lock);
1775 if (!list_empty(&wait->entry)) {
1776 struct sbitmap_queue *sbq;
1778 list_del_init(&wait->entry);
1779 sbq = &hctx->tags->bitmap_tags;
1780 atomic_dec(&sbq->ws_active);
1782 spin_unlock(&hctx->dispatch_wait_lock);
1784 blk_mq_run_hw_queue(hctx, true);
1789 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1790 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1791 * restart. For both cases, take care to check the condition again after
1792 * marking us as waiting.
1794 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1797 struct sbitmap_queue *sbq;
1798 struct wait_queue_head *wq;
1799 wait_queue_entry_t *wait;
1802 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1803 !(blk_mq_is_shared_tags(hctx->flags))) {
1804 blk_mq_sched_mark_restart_hctx(hctx);
1807 * It's possible that a tag was freed in the window between the
1808 * allocation failure and adding the hardware queue to the wait
1811 * Don't clear RESTART here, someone else could have set it.
1812 * At most this will cost an extra queue run.
1814 return blk_mq_get_driver_tag(rq);
1817 wait = &hctx->dispatch_wait;
1818 if (!list_empty_careful(&wait->entry))
1821 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1822 sbq = &hctx->tags->breserved_tags;
1824 sbq = &hctx->tags->bitmap_tags;
1825 wq = &bt_wait_ptr(sbq, hctx)->wait;
1827 spin_lock_irq(&wq->lock);
1828 spin_lock(&hctx->dispatch_wait_lock);
1829 if (!list_empty(&wait->entry)) {
1830 spin_unlock(&hctx->dispatch_wait_lock);
1831 spin_unlock_irq(&wq->lock);
1835 atomic_inc(&sbq->ws_active);
1836 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1837 __add_wait_queue(wq, wait);
1840 * It's possible that a tag was freed in the window between the
1841 * allocation failure and adding the hardware queue to the wait
1844 ret = blk_mq_get_driver_tag(rq);
1846 spin_unlock(&hctx->dispatch_wait_lock);
1847 spin_unlock_irq(&wq->lock);
1852 * We got a tag, remove ourselves from the wait queue to ensure
1853 * someone else gets the wakeup.
1855 list_del_init(&wait->entry);
1856 atomic_dec(&sbq->ws_active);
1857 spin_unlock(&hctx->dispatch_wait_lock);
1858 spin_unlock_irq(&wq->lock);
1863 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1864 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1866 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1867 * - EWMA is one simple way to compute running average value
1868 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1869 * - take 4 as factor for avoiding to get too small(0) result, and this
1870 * factor doesn't matter because EWMA decreases exponentially
1872 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1876 ewma = hctx->dispatch_busy;
1881 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1883 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1884 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1886 hctx->dispatch_busy = ewma;
1889 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1891 static void blk_mq_handle_dev_resource(struct request *rq,
1892 struct list_head *list)
1894 list_add(&rq->queuelist, list);
1895 __blk_mq_requeue_request(rq);
1898 static void blk_mq_handle_zone_resource(struct request *rq,
1899 struct list_head *zone_list)
1902 * If we end up here it is because we cannot dispatch a request to a
1903 * specific zone due to LLD level zone-write locking or other zone
1904 * related resource not being available. In this case, set the request
1905 * aside in zone_list for retrying it later.
1907 list_add(&rq->queuelist, zone_list);
1908 __blk_mq_requeue_request(rq);
1911 enum prep_dispatch {
1913 PREP_DISPATCH_NO_TAG,
1914 PREP_DISPATCH_NO_BUDGET,
1917 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1920 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1921 int budget_token = -1;
1924 budget_token = blk_mq_get_dispatch_budget(rq->q);
1925 if (budget_token < 0) {
1926 blk_mq_put_driver_tag(rq);
1927 return PREP_DISPATCH_NO_BUDGET;
1929 blk_mq_set_rq_budget_token(rq, budget_token);
1932 if (!blk_mq_get_driver_tag(rq)) {
1934 * The initial allocation attempt failed, so we need to
1935 * rerun the hardware queue when a tag is freed. The
1936 * waitqueue takes care of that. If the queue is run
1937 * before we add this entry back on the dispatch list,
1938 * we'll re-run it below.
1940 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1942 * All budgets not got from this function will be put
1943 * together during handling partial dispatch
1946 blk_mq_put_dispatch_budget(rq->q, budget_token);
1947 return PREP_DISPATCH_NO_TAG;
1951 return PREP_DISPATCH_OK;
1954 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1955 static void blk_mq_release_budgets(struct request_queue *q,
1956 struct list_head *list)
1960 list_for_each_entry(rq, list, queuelist) {
1961 int budget_token = blk_mq_get_rq_budget_token(rq);
1963 if (budget_token >= 0)
1964 blk_mq_put_dispatch_budget(q, budget_token);
1969 * blk_mq_commit_rqs will notify driver using bd->last that there is no
1970 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
1972 * Attention, we should explicitly call this in unusual cases:
1973 * 1) did not queue everything initially scheduled to queue
1974 * 2) the last attempt to queue a request failed
1976 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
1979 if (hctx->queue->mq_ops->commit_rqs && queued) {
1980 trace_block_unplug(hctx->queue, queued, !from_schedule);
1981 hctx->queue->mq_ops->commit_rqs(hctx);
1986 * Returns true if we did some work AND can potentially do more.
1988 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1989 unsigned int nr_budgets)
1991 enum prep_dispatch prep;
1992 struct request_queue *q = hctx->queue;
1995 blk_status_t ret = BLK_STS_OK;
1996 LIST_HEAD(zone_list);
1997 bool needs_resource = false;
1999 if (list_empty(list))
2003 * Now process all the entries, sending them to the driver.
2007 struct blk_mq_queue_data bd;
2009 rq = list_first_entry(list, struct request, queuelist);
2011 WARN_ON_ONCE(hctx != rq->mq_hctx);
2012 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2013 if (prep != PREP_DISPATCH_OK)
2016 list_del_init(&rq->queuelist);
2019 bd.last = list_empty(list);
2022 * once the request is queued to lld, no need to cover the
2027 ret = q->mq_ops->queue_rq(hctx, &bd);
2032 case BLK_STS_RESOURCE:
2033 needs_resource = true;
2035 case BLK_STS_DEV_RESOURCE:
2036 blk_mq_handle_dev_resource(rq, list);
2038 case BLK_STS_ZONE_RESOURCE:
2040 * Move the request to zone_list and keep going through
2041 * the dispatch list to find more requests the drive can
2044 blk_mq_handle_zone_resource(rq, &zone_list);
2045 needs_resource = true;
2048 blk_mq_end_request(rq, ret);
2050 } while (!list_empty(list));
2052 if (!list_empty(&zone_list))
2053 list_splice_tail_init(&zone_list, list);
2055 /* If we didn't flush the entire list, we could have told the driver
2056 * there was more coming, but that turned out to be a lie.
2058 if (!list_empty(list) || ret != BLK_STS_OK)
2059 blk_mq_commit_rqs(hctx, queued, false);
2062 * Any items that need requeuing? Stuff them into hctx->dispatch,
2063 * that is where we will continue on next queue run.
2065 if (!list_empty(list)) {
2067 /* For non-shared tags, the RESTART check will suffice */
2068 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2069 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2070 blk_mq_is_shared_tags(hctx->flags));
2073 blk_mq_release_budgets(q, list);
2075 spin_lock(&hctx->lock);
2076 list_splice_tail_init(list, &hctx->dispatch);
2077 spin_unlock(&hctx->lock);
2080 * Order adding requests to hctx->dispatch and checking
2081 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2082 * in blk_mq_sched_restart(). Avoid restart code path to
2083 * miss the new added requests to hctx->dispatch, meantime
2084 * SCHED_RESTART is observed here.
2089 * If SCHED_RESTART was set by the caller of this function and
2090 * it is no longer set that means that it was cleared by another
2091 * thread and hence that a queue rerun is needed.
2093 * If 'no_tag' is set, that means that we failed getting
2094 * a driver tag with an I/O scheduler attached. If our dispatch
2095 * waitqueue is no longer active, ensure that we run the queue
2096 * AFTER adding our entries back to the list.
2098 * If no I/O scheduler has been configured it is possible that
2099 * the hardware queue got stopped and restarted before requests
2100 * were pushed back onto the dispatch list. Rerun the queue to
2101 * avoid starvation. Notes:
2102 * - blk_mq_run_hw_queue() checks whether or not a queue has
2103 * been stopped before rerunning a queue.
2104 * - Some but not all block drivers stop a queue before
2105 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2108 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2109 * bit is set, run queue after a delay to avoid IO stalls
2110 * that could otherwise occur if the queue is idle. We'll do
2111 * similar if we couldn't get budget or couldn't lock a zone
2112 * and SCHED_RESTART is set.
2114 needs_restart = blk_mq_sched_needs_restart(hctx);
2115 if (prep == PREP_DISPATCH_NO_BUDGET)
2116 needs_resource = true;
2117 if (!needs_restart ||
2118 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2119 blk_mq_run_hw_queue(hctx, true);
2120 else if (needs_resource)
2121 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2123 blk_mq_update_dispatch_busy(hctx, true);
2127 blk_mq_update_dispatch_busy(hctx, false);
2131 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2133 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2135 if (cpu >= nr_cpu_ids)
2136 cpu = cpumask_first(hctx->cpumask);
2141 * It'd be great if the workqueue API had a way to pass
2142 * in a mask and had some smarts for more clever placement.
2143 * For now we just round-robin here, switching for every
2144 * BLK_MQ_CPU_WORK_BATCH queued items.
2146 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2149 int next_cpu = hctx->next_cpu;
2151 if (hctx->queue->nr_hw_queues == 1)
2152 return WORK_CPU_UNBOUND;
2154 if (--hctx->next_cpu_batch <= 0) {
2156 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2158 if (next_cpu >= nr_cpu_ids)
2159 next_cpu = blk_mq_first_mapped_cpu(hctx);
2160 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2164 * Do unbound schedule if we can't find a online CPU for this hctx,
2165 * and it should only happen in the path of handling CPU DEAD.
2167 if (!cpu_online(next_cpu)) {
2174 * Make sure to re-select CPU next time once after CPUs
2175 * in hctx->cpumask become online again.
2177 hctx->next_cpu = next_cpu;
2178 hctx->next_cpu_batch = 1;
2179 return WORK_CPU_UNBOUND;
2182 hctx->next_cpu = next_cpu;
2187 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2188 * @hctx: Pointer to the hardware queue to run.
2189 * @msecs: Milliseconds of delay to wait before running the queue.
2191 * Run a hardware queue asynchronously with a delay of @msecs.
2193 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2195 if (unlikely(blk_mq_hctx_stopped(hctx)))
2197 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2198 msecs_to_jiffies(msecs));
2200 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2203 * blk_mq_run_hw_queue - Start to run a hardware queue.
2204 * @hctx: Pointer to the hardware queue to run.
2205 * @async: If we want to run the queue asynchronously.
2207 * Check if the request queue is not in a quiesced state and if there are
2208 * pending requests to be sent. If this is true, run the queue to send requests
2211 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2216 * We can't run the queue inline with interrupts disabled.
2218 WARN_ON_ONCE(!async && in_interrupt());
2221 * When queue is quiesced, we may be switching io scheduler, or
2222 * updating nr_hw_queues, or other things, and we can't run queue
2223 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2225 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2228 __blk_mq_run_dispatch_ops(hctx->queue, false,
2229 need_run = !blk_queue_quiesced(hctx->queue) &&
2230 blk_mq_hctx_has_pending(hctx));
2235 if (async || (hctx->flags & BLK_MQ_F_BLOCKING) ||
2236 !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2237 blk_mq_delay_run_hw_queue(hctx, 0);
2241 blk_mq_run_dispatch_ops(hctx->queue,
2242 blk_mq_sched_dispatch_requests(hctx));
2244 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2247 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2250 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2252 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2254 * If the IO scheduler does not respect hardware queues when
2255 * dispatching, we just don't bother with multiple HW queues and
2256 * dispatch from hctx for the current CPU since running multiple queues
2257 * just causes lock contention inside the scheduler and pointless cache
2260 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2262 if (!blk_mq_hctx_stopped(hctx))
2268 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2269 * @q: Pointer to the request queue to run.
2270 * @async: If we want to run the queue asynchronously.
2272 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2274 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2278 if (blk_queue_sq_sched(q))
2279 sq_hctx = blk_mq_get_sq_hctx(q);
2280 queue_for_each_hw_ctx(q, hctx, i) {
2281 if (blk_mq_hctx_stopped(hctx))
2284 * Dispatch from this hctx either if there's no hctx preferred
2285 * by IO scheduler or if it has requests that bypass the
2288 if (!sq_hctx || sq_hctx == hctx ||
2289 !list_empty_careful(&hctx->dispatch))
2290 blk_mq_run_hw_queue(hctx, async);
2293 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2296 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2297 * @q: Pointer to the request queue to run.
2298 * @msecs: Milliseconds of delay to wait before running the queues.
2300 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2302 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2306 if (blk_queue_sq_sched(q))
2307 sq_hctx = blk_mq_get_sq_hctx(q);
2308 queue_for_each_hw_ctx(q, hctx, i) {
2309 if (blk_mq_hctx_stopped(hctx))
2312 * If there is already a run_work pending, leave the
2313 * pending delay untouched. Otherwise, a hctx can stall
2314 * if another hctx is re-delaying the other's work
2315 * before the work executes.
2317 if (delayed_work_pending(&hctx->run_work))
2320 * Dispatch from this hctx either if there's no hctx preferred
2321 * by IO scheduler or if it has requests that bypass the
2324 if (!sq_hctx || sq_hctx == hctx ||
2325 !list_empty_careful(&hctx->dispatch))
2326 blk_mq_delay_run_hw_queue(hctx, msecs);
2329 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2332 * This function is often used for pausing .queue_rq() by driver when
2333 * there isn't enough resource or some conditions aren't satisfied, and
2334 * BLK_STS_RESOURCE is usually returned.
2336 * We do not guarantee that dispatch can be drained or blocked
2337 * after blk_mq_stop_hw_queue() returns. Please use
2338 * blk_mq_quiesce_queue() for that requirement.
2340 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2342 cancel_delayed_work(&hctx->run_work);
2344 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2346 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2349 * This function is often used for pausing .queue_rq() by driver when
2350 * there isn't enough resource or some conditions aren't satisfied, and
2351 * BLK_STS_RESOURCE is usually returned.
2353 * We do not guarantee that dispatch can be drained or blocked
2354 * after blk_mq_stop_hw_queues() returns. Please use
2355 * blk_mq_quiesce_queue() for that requirement.
2357 void blk_mq_stop_hw_queues(struct request_queue *q)
2359 struct blk_mq_hw_ctx *hctx;
2362 queue_for_each_hw_ctx(q, hctx, i)
2363 blk_mq_stop_hw_queue(hctx);
2365 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2367 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2369 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2371 blk_mq_run_hw_queue(hctx, false);
2373 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2375 void blk_mq_start_hw_queues(struct request_queue *q)
2377 struct blk_mq_hw_ctx *hctx;
2380 queue_for_each_hw_ctx(q, hctx, i)
2381 blk_mq_start_hw_queue(hctx);
2383 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2385 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2387 if (!blk_mq_hctx_stopped(hctx))
2390 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2391 blk_mq_run_hw_queue(hctx, async);
2393 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2395 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2397 struct blk_mq_hw_ctx *hctx;
2400 queue_for_each_hw_ctx(q, hctx, i)
2401 blk_mq_start_stopped_hw_queue(hctx, async);
2403 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2405 static void blk_mq_run_work_fn(struct work_struct *work)
2407 struct blk_mq_hw_ctx *hctx =
2408 container_of(work, struct blk_mq_hw_ctx, run_work.work);
2410 blk_mq_run_dispatch_ops(hctx->queue,
2411 blk_mq_sched_dispatch_requests(hctx));
2415 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2416 * @rq: Pointer to request to be inserted.
2417 * @flags: BLK_MQ_INSERT_*
2419 * Should only be used carefully, when the caller knows we want to
2420 * bypass a potential IO scheduler on the target device.
2422 static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2424 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2426 spin_lock(&hctx->lock);
2427 if (flags & BLK_MQ_INSERT_AT_HEAD)
2428 list_add(&rq->queuelist, &hctx->dispatch);
2430 list_add_tail(&rq->queuelist, &hctx->dispatch);
2431 spin_unlock(&hctx->lock);
2434 static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2435 struct blk_mq_ctx *ctx, struct list_head *list,
2436 bool run_queue_async)
2439 enum hctx_type type = hctx->type;
2442 * Try to issue requests directly if the hw queue isn't busy to save an
2443 * extra enqueue & dequeue to the sw queue.
2445 if (!hctx->dispatch_busy && !run_queue_async) {
2446 blk_mq_run_dispatch_ops(hctx->queue,
2447 blk_mq_try_issue_list_directly(hctx, list));
2448 if (list_empty(list))
2453 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2456 list_for_each_entry(rq, list, queuelist) {
2457 BUG_ON(rq->mq_ctx != ctx);
2458 trace_block_rq_insert(rq);
2461 spin_lock(&ctx->lock);
2462 list_splice_tail_init(list, &ctx->rq_lists[type]);
2463 blk_mq_hctx_mark_pending(hctx, ctx);
2464 spin_unlock(&ctx->lock);
2466 blk_mq_run_hw_queue(hctx, run_queue_async);
2469 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2471 struct request_queue *q = rq->q;
2472 struct blk_mq_ctx *ctx = rq->mq_ctx;
2473 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2475 if (blk_rq_is_passthrough(rq)) {
2477 * Passthrough request have to be added to hctx->dispatch
2478 * directly. The device may be in a situation where it can't
2479 * handle FS request, and always returns BLK_STS_RESOURCE for
2480 * them, which gets them added to hctx->dispatch.
2482 * If a passthrough request is required to unblock the queues,
2483 * and it is added to the scheduler queue, there is no chance to
2484 * dispatch it given we prioritize requests in hctx->dispatch.
2486 blk_mq_request_bypass_insert(rq, flags);
2487 } else if (req_op(rq) == REQ_OP_FLUSH) {
2489 * Firstly normal IO request is inserted to scheduler queue or
2490 * sw queue, meantime we add flush request to dispatch queue(
2491 * hctx->dispatch) directly and there is at most one in-flight
2492 * flush request for each hw queue, so it doesn't matter to add
2493 * flush request to tail or front of the dispatch queue.
2495 * Secondly in case of NCQ, flush request belongs to non-NCQ
2496 * command, and queueing it will fail when there is any
2497 * in-flight normal IO request(NCQ command). When adding flush
2498 * rq to the front of hctx->dispatch, it is easier to introduce
2499 * extra time to flush rq's latency because of S_SCHED_RESTART
2500 * compared with adding to the tail of dispatch queue, then
2501 * chance of flush merge is increased, and less flush requests
2502 * will be issued to controller. It is observed that ~10% time
2503 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2504 * drive when adding flush rq to the front of hctx->dispatch.
2506 * Simply queue flush rq to the front of hctx->dispatch so that
2507 * intensive flush workloads can benefit in case of NCQ HW.
2509 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2510 } else if (q->elevator) {
2513 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2515 list_add(&rq->queuelist, &list);
2516 q->elevator->type->ops.insert_requests(hctx, &list, flags);
2518 trace_block_rq_insert(rq);
2520 spin_lock(&ctx->lock);
2521 if (flags & BLK_MQ_INSERT_AT_HEAD)
2522 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2524 list_add_tail(&rq->queuelist,
2525 &ctx->rq_lists[hctx->type]);
2526 blk_mq_hctx_mark_pending(hctx, ctx);
2527 spin_unlock(&ctx->lock);
2531 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2532 unsigned int nr_segs)
2536 if (bio->bi_opf & REQ_RAHEAD)
2537 rq->cmd_flags |= REQ_FAILFAST_MASK;
2539 rq->__sector = bio->bi_iter.bi_sector;
2540 blk_rq_bio_prep(rq, bio, nr_segs);
2542 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2543 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2546 blk_account_io_start(rq);
2549 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2550 struct request *rq, bool last)
2552 struct request_queue *q = rq->q;
2553 struct blk_mq_queue_data bd = {
2560 * For OK queue, we are done. For error, caller may kill it.
2561 * Any other error (busy), just add it to our list as we
2562 * previously would have done.
2564 ret = q->mq_ops->queue_rq(hctx, &bd);
2567 blk_mq_update_dispatch_busy(hctx, false);
2569 case BLK_STS_RESOURCE:
2570 case BLK_STS_DEV_RESOURCE:
2571 blk_mq_update_dispatch_busy(hctx, true);
2572 __blk_mq_requeue_request(rq);
2575 blk_mq_update_dispatch_busy(hctx, false);
2582 static bool blk_mq_get_budget_and_tag(struct request *rq)
2586 budget_token = blk_mq_get_dispatch_budget(rq->q);
2587 if (budget_token < 0)
2589 blk_mq_set_rq_budget_token(rq, budget_token);
2590 if (!blk_mq_get_driver_tag(rq)) {
2591 blk_mq_put_dispatch_budget(rq->q, budget_token);
2598 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2599 * @hctx: Pointer of the associated hardware queue.
2600 * @rq: Pointer to request to be sent.
2602 * If the device has enough resources to accept a new request now, send the
2603 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2604 * we can try send it another time in the future. Requests inserted at this
2605 * queue have higher priority.
2607 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2612 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2613 blk_mq_insert_request(rq, 0);
2617 if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
2618 blk_mq_insert_request(rq, 0);
2619 blk_mq_run_hw_queue(hctx, false);
2623 ret = __blk_mq_issue_directly(hctx, rq, true);
2627 case BLK_STS_RESOURCE:
2628 case BLK_STS_DEV_RESOURCE:
2629 blk_mq_request_bypass_insert(rq, 0);
2630 blk_mq_run_hw_queue(hctx, false);
2633 blk_mq_end_request(rq, ret);
2638 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2640 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2642 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2643 blk_mq_insert_request(rq, 0);
2647 if (!blk_mq_get_budget_and_tag(rq))
2648 return BLK_STS_RESOURCE;
2649 return __blk_mq_issue_directly(hctx, rq, last);
2652 static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2654 struct blk_mq_hw_ctx *hctx = NULL;
2657 blk_status_t ret = BLK_STS_OK;
2659 while ((rq = rq_list_pop(&plug->mq_list))) {
2660 bool last = rq_list_empty(plug->mq_list);
2662 if (hctx != rq->mq_hctx) {
2664 blk_mq_commit_rqs(hctx, queued, false);
2670 ret = blk_mq_request_issue_directly(rq, last);
2675 case BLK_STS_RESOURCE:
2676 case BLK_STS_DEV_RESOURCE:
2677 blk_mq_request_bypass_insert(rq, 0);
2678 blk_mq_run_hw_queue(hctx, false);
2681 blk_mq_end_request(rq, ret);
2687 if (ret != BLK_STS_OK)
2688 blk_mq_commit_rqs(hctx, queued, false);
2691 static void __blk_mq_flush_plug_list(struct request_queue *q,
2692 struct blk_plug *plug)
2694 if (blk_queue_quiesced(q))
2696 q->mq_ops->queue_rqs(&plug->mq_list);
2699 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2701 struct blk_mq_hw_ctx *this_hctx = NULL;
2702 struct blk_mq_ctx *this_ctx = NULL;
2703 struct request *requeue_list = NULL;
2704 struct request **requeue_lastp = &requeue_list;
2705 unsigned int depth = 0;
2706 bool is_passthrough = false;
2710 struct request *rq = rq_list_pop(&plug->mq_list);
2713 this_hctx = rq->mq_hctx;
2714 this_ctx = rq->mq_ctx;
2715 is_passthrough = blk_rq_is_passthrough(rq);
2716 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2717 is_passthrough != blk_rq_is_passthrough(rq)) {
2718 rq_list_add_tail(&requeue_lastp, rq);
2721 list_add(&rq->queuelist, &list);
2723 } while (!rq_list_empty(plug->mq_list));
2725 plug->mq_list = requeue_list;
2726 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2728 percpu_ref_get(&this_hctx->queue->q_usage_counter);
2729 /* passthrough requests should never be issued to the I/O scheduler */
2730 if (this_hctx->queue->elevator && !is_passthrough) {
2731 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2733 blk_mq_run_hw_queue(this_hctx, from_sched);
2735 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2737 percpu_ref_put(&this_hctx->queue->q_usage_counter);
2740 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2744 if (rq_list_empty(plug->mq_list))
2748 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2749 struct request_queue *q;
2751 rq = rq_list_peek(&plug->mq_list);
2755 * Peek first request and see if we have a ->queue_rqs() hook.
2756 * If we do, we can dispatch the whole plug list in one go. We
2757 * already know at this point that all requests belong to the
2758 * same queue, caller must ensure that's the case.
2760 * Since we pass off the full list to the driver at this point,
2761 * we do not increment the active request count for the queue.
2762 * Bypass shared tags for now because of that.
2764 if (q->mq_ops->queue_rqs &&
2765 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2766 blk_mq_run_dispatch_ops(q,
2767 __blk_mq_flush_plug_list(q, plug));
2768 if (rq_list_empty(plug->mq_list))
2772 blk_mq_run_dispatch_ops(q,
2773 blk_mq_plug_issue_direct(plug));
2774 if (rq_list_empty(plug->mq_list))
2779 blk_mq_dispatch_plug_list(plug, from_schedule);
2780 } while (!rq_list_empty(plug->mq_list));
2783 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2784 struct list_head *list)
2787 blk_status_t ret = BLK_STS_OK;
2789 while (!list_empty(list)) {
2790 struct request *rq = list_first_entry(list, struct request,
2793 list_del_init(&rq->queuelist);
2794 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2799 case BLK_STS_RESOURCE:
2800 case BLK_STS_DEV_RESOURCE:
2801 blk_mq_request_bypass_insert(rq, 0);
2802 if (list_empty(list))
2803 blk_mq_run_hw_queue(hctx, false);
2806 blk_mq_end_request(rq, ret);
2812 if (ret != BLK_STS_OK)
2813 blk_mq_commit_rqs(hctx, queued, false);
2816 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2817 struct bio *bio, unsigned int nr_segs)
2819 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2820 if (blk_attempt_plug_merge(q, bio, nr_segs))
2822 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2828 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2829 struct blk_plug *plug,
2833 struct blk_mq_alloc_data data = {
2836 .cmd_flags = bio->bi_opf,
2840 if (unlikely(bio_queue_enter(bio)))
2843 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2846 rq_qos_throttle(q, bio);
2849 data.nr_tags = plug->nr_ios;
2851 data.cached_rq = &plug->cached_rq;
2854 rq = __blk_mq_alloc_requests(&data);
2857 rq_qos_cleanup(q, bio);
2858 if (bio->bi_opf & REQ_NOWAIT)
2859 bio_wouldblock_error(bio);
2865 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2866 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2869 enum hctx_type type, hctx_type;
2873 rq = rq_list_peek(&plug->cached_rq);
2874 if (!rq || rq->q != q)
2877 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2882 type = blk_mq_get_hctx_type((*bio)->bi_opf);
2883 hctx_type = rq->mq_hctx->type;
2884 if (type != hctx_type &&
2885 !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2887 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2891 * If any qos ->throttle() end up blocking, we will have flushed the
2892 * plug and hence killed the cached_rq list as well. Pop this entry
2893 * before we throttle.
2895 plug->cached_rq = rq_list_next(rq);
2896 rq_qos_throttle(q, *bio);
2898 rq->cmd_flags = (*bio)->bi_opf;
2899 INIT_LIST_HEAD(&rq->queuelist);
2903 static void bio_set_ioprio(struct bio *bio)
2905 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2906 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2907 bio->bi_ioprio = get_current_ioprio();
2908 blkcg_set_ioprio(bio);
2912 * blk_mq_submit_bio - Create and send a request to block device.
2913 * @bio: Bio pointer.
2915 * Builds up a request structure from @q and @bio and send to the device. The
2916 * request may not be queued directly to hardware if:
2917 * * This request can be merged with another one
2918 * * We want to place request at plug queue for possible future merging
2919 * * There is an IO scheduler active at this queue
2921 * It will not queue the request if there is an error with the bio, or at the
2924 void blk_mq_submit_bio(struct bio *bio)
2926 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2927 struct blk_plug *plug = blk_mq_plug(bio);
2928 const int is_sync = op_is_sync(bio->bi_opf);
2929 struct blk_mq_hw_ctx *hctx;
2931 unsigned int nr_segs = 1;
2934 bio = blk_queue_bounce(bio, q);
2935 if (bio_may_exceed_limits(bio, &q->limits)) {
2936 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2941 if (!bio_integrity_prep(bio))
2944 bio_set_ioprio(bio);
2946 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2950 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2955 trace_block_getrq(bio);
2957 rq_qos_track(q, rq, bio);
2959 blk_mq_bio_to_request(rq, bio, nr_segs);
2961 ret = blk_crypto_rq_get_keyslot(rq);
2962 if (ret != BLK_STS_OK) {
2963 bio->bi_status = ret;
2965 blk_mq_free_request(rq);
2969 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
2973 blk_add_rq_to_plug(plug, rq);
2978 if ((rq->rq_flags & RQF_USE_SCHED) ||
2979 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
2980 blk_mq_insert_request(rq, 0);
2981 blk_mq_run_hw_queue(hctx, true);
2983 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
2987 #ifdef CONFIG_BLK_MQ_STACKING
2989 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
2990 * @rq: the request being queued
2992 blk_status_t blk_insert_cloned_request(struct request *rq)
2994 struct request_queue *q = rq->q;
2995 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
2996 unsigned int max_segments = blk_rq_get_max_segments(rq);
2999 if (blk_rq_sectors(rq) > max_sectors) {
3001 * SCSI device does not have a good way to return if
3002 * Write Same/Zero is actually supported. If a device rejects
3003 * a non-read/write command (discard, write same,etc.) the
3004 * low-level device driver will set the relevant queue limit to
3005 * 0 to prevent blk-lib from issuing more of the offending
3006 * operations. Commands queued prior to the queue limit being
3007 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3008 * errors being propagated to upper layers.
3010 if (max_sectors == 0)
3011 return BLK_STS_NOTSUPP;
3013 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3014 __func__, blk_rq_sectors(rq), max_sectors);
3015 return BLK_STS_IOERR;
3019 * The queue settings related to segment counting may differ from the
3022 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3023 if (rq->nr_phys_segments > max_segments) {
3024 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3025 __func__, rq->nr_phys_segments, max_segments);
3026 return BLK_STS_IOERR;
3029 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3030 return BLK_STS_IOERR;
3032 ret = blk_crypto_rq_get_keyslot(rq);
3033 if (ret != BLK_STS_OK)
3036 blk_account_io_start(rq);
3039 * Since we have a scheduler attached on the top device,
3040 * bypass a potential scheduler on the bottom device for
3043 blk_mq_run_dispatch_ops(q,
3044 ret = blk_mq_request_issue_directly(rq, true));
3046 blk_account_io_done(rq, ktime_get_ns());
3049 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3052 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3053 * @rq: the clone request to be cleaned up
3056 * Free all bios in @rq for a cloned request.
3058 void blk_rq_unprep_clone(struct request *rq)
3062 while ((bio = rq->bio) != NULL) {
3063 rq->bio = bio->bi_next;
3068 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3071 * blk_rq_prep_clone - Helper function to setup clone request
3072 * @rq: the request to be setup
3073 * @rq_src: original request to be cloned
3074 * @bs: bio_set that bios for clone are allocated from
3075 * @gfp_mask: memory allocation mask for bio
3076 * @bio_ctr: setup function to be called for each clone bio.
3077 * Returns %0 for success, non %0 for failure.
3078 * @data: private data to be passed to @bio_ctr
3081 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3082 * Also, pages which the original bios are pointing to are not copied
3083 * and the cloned bios just point same pages.
3084 * So cloned bios must be completed before original bios, which means
3085 * the caller must complete @rq before @rq_src.
3087 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3088 struct bio_set *bs, gfp_t gfp_mask,
3089 int (*bio_ctr)(struct bio *, struct bio *, void *),
3092 struct bio *bio, *bio_src;
3097 __rq_for_each_bio(bio_src, rq_src) {
3098 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3103 if (bio_ctr && bio_ctr(bio, bio_src, data))
3107 rq->biotail->bi_next = bio;
3110 rq->bio = rq->biotail = bio;
3115 /* Copy attributes of the original request to the clone request. */
3116 rq->__sector = blk_rq_pos(rq_src);
3117 rq->__data_len = blk_rq_bytes(rq_src);
3118 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3119 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3120 rq->special_vec = rq_src->special_vec;
3122 rq->nr_phys_segments = rq_src->nr_phys_segments;
3123 rq->ioprio = rq_src->ioprio;
3125 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3133 blk_rq_unprep_clone(rq);
3137 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3138 #endif /* CONFIG_BLK_MQ_STACKING */
3141 * Steal bios from a request and add them to a bio list.
3142 * The request must not have been partially completed before.
3144 void blk_steal_bios(struct bio_list *list, struct request *rq)
3148 list->tail->bi_next = rq->bio;
3150 list->head = rq->bio;
3151 list->tail = rq->biotail;
3159 EXPORT_SYMBOL_GPL(blk_steal_bios);
3161 static size_t order_to_size(unsigned int order)
3163 return (size_t)PAGE_SIZE << order;
3166 /* called before freeing request pool in @tags */
3167 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3168 struct blk_mq_tags *tags)
3171 unsigned long flags;
3174 * There is no need to clear mapping if driver tags is not initialized
3175 * or the mapping belongs to the driver tags.
3177 if (!drv_tags || drv_tags == tags)
3180 list_for_each_entry(page, &tags->page_list, lru) {
3181 unsigned long start = (unsigned long)page_address(page);
3182 unsigned long end = start + order_to_size(page->private);
3185 for (i = 0; i < drv_tags->nr_tags; i++) {
3186 struct request *rq = drv_tags->rqs[i];
3187 unsigned long rq_addr = (unsigned long)rq;
3189 if (rq_addr >= start && rq_addr < end) {
3190 WARN_ON_ONCE(req_ref_read(rq) != 0);
3191 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3197 * Wait until all pending iteration is done.
3199 * Request reference is cleared and it is guaranteed to be observed
3200 * after the ->lock is released.
3202 spin_lock_irqsave(&drv_tags->lock, flags);
3203 spin_unlock_irqrestore(&drv_tags->lock, flags);
3206 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3207 unsigned int hctx_idx)
3209 struct blk_mq_tags *drv_tags;
3212 if (list_empty(&tags->page_list))
3215 if (blk_mq_is_shared_tags(set->flags))
3216 drv_tags = set->shared_tags;
3218 drv_tags = set->tags[hctx_idx];
3220 if (tags->static_rqs && set->ops->exit_request) {
3223 for (i = 0; i < tags->nr_tags; i++) {
3224 struct request *rq = tags->static_rqs[i];
3228 set->ops->exit_request(set, rq, hctx_idx);
3229 tags->static_rqs[i] = NULL;
3233 blk_mq_clear_rq_mapping(drv_tags, tags);
3235 while (!list_empty(&tags->page_list)) {
3236 page = list_first_entry(&tags->page_list, struct page, lru);
3237 list_del_init(&page->lru);
3239 * Remove kmemleak object previously allocated in
3240 * blk_mq_alloc_rqs().
3242 kmemleak_free(page_address(page));
3243 __free_pages(page, page->private);
3247 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3251 kfree(tags->static_rqs);
3252 tags->static_rqs = NULL;
3254 blk_mq_free_tags(tags);
3257 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3258 unsigned int hctx_idx)
3262 for (i = 0; i < set->nr_maps; i++) {
3263 unsigned int start = set->map[i].queue_offset;
3264 unsigned int end = start + set->map[i].nr_queues;
3266 if (hctx_idx >= start && hctx_idx < end)
3270 if (i >= set->nr_maps)
3271 i = HCTX_TYPE_DEFAULT;
3276 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3277 unsigned int hctx_idx)
3279 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3281 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3284 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3285 unsigned int hctx_idx,
3286 unsigned int nr_tags,
3287 unsigned int reserved_tags)
3289 int node = blk_mq_get_hctx_node(set, hctx_idx);
3290 struct blk_mq_tags *tags;
3292 if (node == NUMA_NO_NODE)
3293 node = set->numa_node;
3295 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3296 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3300 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3301 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3306 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3307 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3309 if (!tags->static_rqs)
3317 blk_mq_free_tags(tags);
3321 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3322 unsigned int hctx_idx, int node)
3326 if (set->ops->init_request) {
3327 ret = set->ops->init_request(set, rq, hctx_idx, node);
3332 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3336 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3337 struct blk_mq_tags *tags,
3338 unsigned int hctx_idx, unsigned int depth)
3340 unsigned int i, j, entries_per_page, max_order = 4;
3341 int node = blk_mq_get_hctx_node(set, hctx_idx);
3342 size_t rq_size, left;
3344 if (node == NUMA_NO_NODE)
3345 node = set->numa_node;
3347 INIT_LIST_HEAD(&tags->page_list);
3350 * rq_size is the size of the request plus driver payload, rounded
3351 * to the cacheline size
3353 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3355 left = rq_size * depth;
3357 for (i = 0; i < depth; ) {
3358 int this_order = max_order;
3363 while (this_order && left < order_to_size(this_order - 1))
3367 page = alloc_pages_node(node,
3368 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3374 if (order_to_size(this_order) < rq_size)
3381 page->private = this_order;
3382 list_add_tail(&page->lru, &tags->page_list);
3384 p = page_address(page);
3386 * Allow kmemleak to scan these pages as they contain pointers
3387 * to additional allocations like via ops->init_request().
3389 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3390 entries_per_page = order_to_size(this_order) / rq_size;
3391 to_do = min(entries_per_page, depth - i);
3392 left -= to_do * rq_size;
3393 for (j = 0; j < to_do; j++) {
3394 struct request *rq = p;
3396 tags->static_rqs[i] = rq;
3397 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3398 tags->static_rqs[i] = NULL;
3409 blk_mq_free_rqs(set, tags, hctx_idx);
3413 struct rq_iter_data {
3414 struct blk_mq_hw_ctx *hctx;
3418 static bool blk_mq_has_request(struct request *rq, void *data)
3420 struct rq_iter_data *iter_data = data;
3422 if (rq->mq_hctx != iter_data->hctx)
3424 iter_data->has_rq = true;
3428 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3430 struct blk_mq_tags *tags = hctx->sched_tags ?
3431 hctx->sched_tags : hctx->tags;
3432 struct rq_iter_data data = {
3436 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3440 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3441 struct blk_mq_hw_ctx *hctx)
3443 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3445 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3450 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3452 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3453 struct blk_mq_hw_ctx, cpuhp_online);
3455 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3456 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3460 * Prevent new request from being allocated on the current hctx.
3462 * The smp_mb__after_atomic() Pairs with the implied barrier in
3463 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3464 * seen once we return from the tag allocator.
3466 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3467 smp_mb__after_atomic();
3470 * Try to grab a reference to the queue and wait for any outstanding
3471 * requests. If we could not grab a reference the queue has been
3472 * frozen and there are no requests.
3474 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3475 while (blk_mq_hctx_has_requests(hctx))
3477 percpu_ref_put(&hctx->queue->q_usage_counter);
3483 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3485 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3486 struct blk_mq_hw_ctx, cpuhp_online);
3488 if (cpumask_test_cpu(cpu, hctx->cpumask))
3489 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3494 * 'cpu' is going away. splice any existing rq_list entries from this
3495 * software queue to the hw queue dispatch list, and ensure that it
3498 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3500 struct blk_mq_hw_ctx *hctx;
3501 struct blk_mq_ctx *ctx;
3503 enum hctx_type type;
3505 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3506 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3509 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3512 spin_lock(&ctx->lock);
3513 if (!list_empty(&ctx->rq_lists[type])) {
3514 list_splice_init(&ctx->rq_lists[type], &tmp);
3515 blk_mq_hctx_clear_pending(hctx, ctx);
3517 spin_unlock(&ctx->lock);
3519 if (list_empty(&tmp))
3522 spin_lock(&hctx->lock);
3523 list_splice_tail_init(&tmp, &hctx->dispatch);
3524 spin_unlock(&hctx->lock);
3526 blk_mq_run_hw_queue(hctx, true);
3530 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3532 if (!(hctx->flags & BLK_MQ_F_STACKING))
3533 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3534 &hctx->cpuhp_online);
3535 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3540 * Before freeing hw queue, clearing the flush request reference in
3541 * tags->rqs[] for avoiding potential UAF.
3543 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3544 unsigned int queue_depth, struct request *flush_rq)
3547 unsigned long flags;
3549 /* The hw queue may not be mapped yet */
3553 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3555 for (i = 0; i < queue_depth; i++)
3556 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3559 * Wait until all pending iteration is done.
3561 * Request reference is cleared and it is guaranteed to be observed
3562 * after the ->lock is released.
3564 spin_lock_irqsave(&tags->lock, flags);
3565 spin_unlock_irqrestore(&tags->lock, flags);
3568 /* hctx->ctxs will be freed in queue's release handler */
3569 static void blk_mq_exit_hctx(struct request_queue *q,
3570 struct blk_mq_tag_set *set,
3571 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3573 struct request *flush_rq = hctx->fq->flush_rq;
3575 if (blk_mq_hw_queue_mapped(hctx))
3576 blk_mq_tag_idle(hctx);
3578 if (blk_queue_init_done(q))
3579 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3580 set->queue_depth, flush_rq);
3581 if (set->ops->exit_request)
3582 set->ops->exit_request(set, flush_rq, hctx_idx);
3584 if (set->ops->exit_hctx)
3585 set->ops->exit_hctx(hctx, hctx_idx);
3587 blk_mq_remove_cpuhp(hctx);
3589 xa_erase(&q->hctx_table, hctx_idx);
3591 spin_lock(&q->unused_hctx_lock);
3592 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3593 spin_unlock(&q->unused_hctx_lock);
3596 static void blk_mq_exit_hw_queues(struct request_queue *q,
3597 struct blk_mq_tag_set *set, int nr_queue)
3599 struct blk_mq_hw_ctx *hctx;
3602 queue_for_each_hw_ctx(q, hctx, i) {
3605 blk_mq_exit_hctx(q, set, hctx, i);
3609 static int blk_mq_init_hctx(struct request_queue *q,
3610 struct blk_mq_tag_set *set,
3611 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3613 hctx->queue_num = hctx_idx;
3615 if (!(hctx->flags & BLK_MQ_F_STACKING))
3616 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3617 &hctx->cpuhp_online);
3618 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3620 hctx->tags = set->tags[hctx_idx];
3622 if (set->ops->init_hctx &&
3623 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3624 goto unregister_cpu_notifier;
3626 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3630 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3636 if (set->ops->exit_request)
3637 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3639 if (set->ops->exit_hctx)
3640 set->ops->exit_hctx(hctx, hctx_idx);
3641 unregister_cpu_notifier:
3642 blk_mq_remove_cpuhp(hctx);
3646 static struct blk_mq_hw_ctx *
3647 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3650 struct blk_mq_hw_ctx *hctx;
3651 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3653 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3655 goto fail_alloc_hctx;
3657 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3660 atomic_set(&hctx->nr_active, 0);
3661 if (node == NUMA_NO_NODE)
3662 node = set->numa_node;
3663 hctx->numa_node = node;
3665 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3666 spin_lock_init(&hctx->lock);
3667 INIT_LIST_HEAD(&hctx->dispatch);
3669 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3671 INIT_LIST_HEAD(&hctx->hctx_list);
3674 * Allocate space for all possible cpus to avoid allocation at
3677 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3682 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3683 gfp, node, false, false))
3687 spin_lock_init(&hctx->dispatch_wait_lock);
3688 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3689 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3691 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3695 blk_mq_hctx_kobj_init(hctx);
3700 sbitmap_free(&hctx->ctx_map);
3704 free_cpumask_var(hctx->cpumask);
3711 static void blk_mq_init_cpu_queues(struct request_queue *q,
3712 unsigned int nr_hw_queues)
3714 struct blk_mq_tag_set *set = q->tag_set;
3717 for_each_possible_cpu(i) {
3718 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3719 struct blk_mq_hw_ctx *hctx;
3723 spin_lock_init(&__ctx->lock);
3724 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3725 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3730 * Set local node, IFF we have more than one hw queue. If
3731 * not, we remain on the home node of the device
3733 for (j = 0; j < set->nr_maps; j++) {
3734 hctx = blk_mq_map_queue_type(q, j, i);
3735 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3736 hctx->numa_node = cpu_to_node(i);
3741 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3742 unsigned int hctx_idx,
3745 struct blk_mq_tags *tags;
3748 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3752 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3754 blk_mq_free_rq_map(tags);
3761 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3764 if (blk_mq_is_shared_tags(set->flags)) {
3765 set->tags[hctx_idx] = set->shared_tags;
3770 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3773 return set->tags[hctx_idx];
3776 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3777 struct blk_mq_tags *tags,
3778 unsigned int hctx_idx)
3781 blk_mq_free_rqs(set, tags, hctx_idx);
3782 blk_mq_free_rq_map(tags);
3786 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3787 unsigned int hctx_idx)
3789 if (!blk_mq_is_shared_tags(set->flags))
3790 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3792 set->tags[hctx_idx] = NULL;
3795 static void blk_mq_map_swqueue(struct request_queue *q)
3797 unsigned int j, hctx_idx;
3799 struct blk_mq_hw_ctx *hctx;
3800 struct blk_mq_ctx *ctx;
3801 struct blk_mq_tag_set *set = q->tag_set;
3803 queue_for_each_hw_ctx(q, hctx, i) {
3804 cpumask_clear(hctx->cpumask);
3806 hctx->dispatch_from = NULL;
3810 * Map software to hardware queues.
3812 * If the cpu isn't present, the cpu is mapped to first hctx.
3814 for_each_possible_cpu(i) {
3816 ctx = per_cpu_ptr(q->queue_ctx, i);
3817 for (j = 0; j < set->nr_maps; j++) {
3818 if (!set->map[j].nr_queues) {
3819 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3820 HCTX_TYPE_DEFAULT, i);
3823 hctx_idx = set->map[j].mq_map[i];
3824 /* unmapped hw queue can be remapped after CPU topo changed */
3825 if (!set->tags[hctx_idx] &&
3826 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3828 * If tags initialization fail for some hctx,
3829 * that hctx won't be brought online. In this
3830 * case, remap the current ctx to hctx[0] which
3831 * is guaranteed to always have tags allocated
3833 set->map[j].mq_map[i] = 0;
3836 hctx = blk_mq_map_queue_type(q, j, i);
3837 ctx->hctxs[j] = hctx;
3839 * If the CPU is already set in the mask, then we've
3840 * mapped this one already. This can happen if
3841 * devices share queues across queue maps.
3843 if (cpumask_test_cpu(i, hctx->cpumask))
3846 cpumask_set_cpu(i, hctx->cpumask);
3848 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3849 hctx->ctxs[hctx->nr_ctx++] = ctx;
3852 * If the nr_ctx type overflows, we have exceeded the
3853 * amount of sw queues we can support.
3855 BUG_ON(!hctx->nr_ctx);
3858 for (; j < HCTX_MAX_TYPES; j++)
3859 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3860 HCTX_TYPE_DEFAULT, i);
3863 queue_for_each_hw_ctx(q, hctx, i) {
3865 * If no software queues are mapped to this hardware queue,
3866 * disable it and free the request entries.
3868 if (!hctx->nr_ctx) {
3869 /* Never unmap queue 0. We need it as a
3870 * fallback in case of a new remap fails
3874 __blk_mq_free_map_and_rqs(set, i);
3880 hctx->tags = set->tags[i];
3881 WARN_ON(!hctx->tags);
3884 * Set the map size to the number of mapped software queues.
3885 * This is more accurate and more efficient than looping
3886 * over all possibly mapped software queues.
3888 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3891 * Initialize batch roundrobin counts
3893 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3894 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3899 * Caller needs to ensure that we're either frozen/quiesced, or that
3900 * the queue isn't live yet.
3902 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3904 struct blk_mq_hw_ctx *hctx;
3907 queue_for_each_hw_ctx(q, hctx, i) {
3909 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3911 blk_mq_tag_idle(hctx);
3912 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3917 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3920 struct request_queue *q;
3922 lockdep_assert_held(&set->tag_list_lock);
3924 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3925 blk_mq_freeze_queue(q);
3926 queue_set_hctx_shared(q, shared);
3927 blk_mq_unfreeze_queue(q);
3931 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3933 struct blk_mq_tag_set *set = q->tag_set;
3935 mutex_lock(&set->tag_list_lock);
3936 list_del(&q->tag_set_list);
3937 if (list_is_singular(&set->tag_list)) {
3938 /* just transitioned to unshared */
3939 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3940 /* update existing queue */
3941 blk_mq_update_tag_set_shared(set, false);
3943 mutex_unlock(&set->tag_list_lock);
3944 INIT_LIST_HEAD(&q->tag_set_list);
3947 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3948 struct request_queue *q)
3950 mutex_lock(&set->tag_list_lock);
3953 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3955 if (!list_empty(&set->tag_list) &&
3956 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3957 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3958 /* update existing queue */
3959 blk_mq_update_tag_set_shared(set, true);
3961 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3962 queue_set_hctx_shared(q, true);
3963 list_add_tail(&q->tag_set_list, &set->tag_list);
3965 mutex_unlock(&set->tag_list_lock);
3968 /* All allocations will be freed in release handler of q->mq_kobj */
3969 static int blk_mq_alloc_ctxs(struct request_queue *q)
3971 struct blk_mq_ctxs *ctxs;
3974 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3978 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3979 if (!ctxs->queue_ctx)
3982 for_each_possible_cpu(cpu) {
3983 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3987 q->mq_kobj = &ctxs->kobj;
3988 q->queue_ctx = ctxs->queue_ctx;
3997 * It is the actual release handler for mq, but we do it from
3998 * request queue's release handler for avoiding use-after-free
3999 * and headache because q->mq_kobj shouldn't have been introduced,
4000 * but we can't group ctx/kctx kobj without it.
4002 void blk_mq_release(struct request_queue *q)
4004 struct blk_mq_hw_ctx *hctx, *next;
4007 queue_for_each_hw_ctx(q, hctx, i)
4008 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4010 /* all hctx are in .unused_hctx_list now */
4011 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4012 list_del_init(&hctx->hctx_list);
4013 kobject_put(&hctx->kobj);
4016 xa_destroy(&q->hctx_table);
4019 * release .mq_kobj and sw queue's kobject now because
4020 * both share lifetime with request queue.
4022 blk_mq_sysfs_deinit(q);
4025 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4028 struct request_queue *q;
4031 q = blk_alloc_queue(set->numa_node);
4033 return ERR_PTR(-ENOMEM);
4034 q->queuedata = queuedata;
4035 ret = blk_mq_init_allocated_queue(set, q);
4038 return ERR_PTR(ret);
4043 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4045 return blk_mq_init_queue_data(set, NULL);
4047 EXPORT_SYMBOL(blk_mq_init_queue);
4050 * blk_mq_destroy_queue - shutdown a request queue
4051 * @q: request queue to shutdown
4053 * This shuts down a request queue allocated by blk_mq_init_queue(). All future
4054 * requests will be failed with -ENODEV. The caller is responsible for dropping
4055 * the reference from blk_mq_init_queue() by calling blk_put_queue().
4057 * Context: can sleep
4059 void blk_mq_destroy_queue(struct request_queue *q)
4061 WARN_ON_ONCE(!queue_is_mq(q));
4062 WARN_ON_ONCE(blk_queue_registered(q));
4066 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4067 blk_queue_start_drain(q);
4068 blk_mq_freeze_queue_wait(q);
4071 blk_mq_cancel_work_sync(q);
4072 blk_mq_exit_queue(q);
4074 EXPORT_SYMBOL(blk_mq_destroy_queue);
4076 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4077 struct lock_class_key *lkclass)
4079 struct request_queue *q;
4080 struct gendisk *disk;
4082 q = blk_mq_init_queue_data(set, queuedata);
4086 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4088 blk_mq_destroy_queue(q);
4090 return ERR_PTR(-ENOMEM);
4092 set_bit(GD_OWNS_QUEUE, &disk->state);
4095 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4097 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4098 struct lock_class_key *lkclass)
4100 struct gendisk *disk;
4102 if (!blk_get_queue(q))
4104 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4109 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4111 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4112 struct blk_mq_tag_set *set, struct request_queue *q,
4113 int hctx_idx, int node)
4115 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4117 /* reuse dead hctx first */
4118 spin_lock(&q->unused_hctx_lock);
4119 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4120 if (tmp->numa_node == node) {
4126 list_del_init(&hctx->hctx_list);
4127 spin_unlock(&q->unused_hctx_lock);
4130 hctx = blk_mq_alloc_hctx(q, set, node);
4134 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4140 kobject_put(&hctx->kobj);
4145 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4146 struct request_queue *q)
4148 struct blk_mq_hw_ctx *hctx;
4151 /* protect against switching io scheduler */
4152 mutex_lock(&q->sysfs_lock);
4153 for (i = 0; i < set->nr_hw_queues; i++) {
4155 int node = blk_mq_get_hctx_node(set, i);
4156 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4159 old_node = old_hctx->numa_node;
4160 blk_mq_exit_hctx(q, set, old_hctx, i);
4163 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4166 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4168 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4169 WARN_ON_ONCE(!hctx);
4173 * Increasing nr_hw_queues fails. Free the newly allocated
4174 * hctxs and keep the previous q->nr_hw_queues.
4176 if (i != set->nr_hw_queues) {
4177 j = q->nr_hw_queues;
4180 q->nr_hw_queues = set->nr_hw_queues;
4183 xa_for_each_start(&q->hctx_table, j, hctx, j)
4184 blk_mq_exit_hctx(q, set, hctx, j);
4185 mutex_unlock(&q->sysfs_lock);
4188 static void blk_mq_update_poll_flag(struct request_queue *q)
4190 struct blk_mq_tag_set *set = q->tag_set;
4192 if (set->nr_maps > HCTX_TYPE_POLL &&
4193 set->map[HCTX_TYPE_POLL].nr_queues)
4194 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4196 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4199 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4200 struct request_queue *q)
4202 /* mark the queue as mq asap */
4203 q->mq_ops = set->ops;
4205 if (blk_mq_alloc_ctxs(q))
4208 /* init q->mq_kobj and sw queues' kobjects */
4209 blk_mq_sysfs_init(q);
4211 INIT_LIST_HEAD(&q->unused_hctx_list);
4212 spin_lock_init(&q->unused_hctx_lock);
4214 xa_init(&q->hctx_table);
4216 blk_mq_realloc_hw_ctxs(set, q);
4217 if (!q->nr_hw_queues)
4220 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4221 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4225 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4226 blk_mq_update_poll_flag(q);
4228 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4229 INIT_LIST_HEAD(&q->flush_list);
4230 INIT_LIST_HEAD(&q->requeue_list);
4231 spin_lock_init(&q->requeue_lock);
4233 q->nr_requests = set->queue_depth;
4235 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4236 blk_mq_add_queue_tag_set(set, q);
4237 blk_mq_map_swqueue(q);
4246 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4248 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4249 void blk_mq_exit_queue(struct request_queue *q)
4251 struct blk_mq_tag_set *set = q->tag_set;
4253 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4254 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4255 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4256 blk_mq_del_queue_tag_set(q);
4259 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4263 if (blk_mq_is_shared_tags(set->flags)) {
4264 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4267 if (!set->shared_tags)
4271 for (i = 0; i < set->nr_hw_queues; i++) {
4272 if (!__blk_mq_alloc_map_and_rqs(set, i))
4281 __blk_mq_free_map_and_rqs(set, i);
4283 if (blk_mq_is_shared_tags(set->flags)) {
4284 blk_mq_free_map_and_rqs(set, set->shared_tags,
4285 BLK_MQ_NO_HCTX_IDX);
4292 * Allocate the request maps associated with this tag_set. Note that this
4293 * may reduce the depth asked for, if memory is tight. set->queue_depth
4294 * will be updated to reflect the allocated depth.
4296 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4301 depth = set->queue_depth;
4303 err = __blk_mq_alloc_rq_maps(set);
4307 set->queue_depth >>= 1;
4308 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4312 } while (set->queue_depth);
4314 if (!set->queue_depth || err) {
4315 pr_err("blk-mq: failed to allocate request map\n");
4319 if (depth != set->queue_depth)
4320 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4321 depth, set->queue_depth);
4326 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4329 * blk_mq_map_queues() and multiple .map_queues() implementations
4330 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4331 * number of hardware queues.
4333 if (set->nr_maps == 1)
4334 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4336 if (set->ops->map_queues && !is_kdump_kernel()) {
4340 * transport .map_queues is usually done in the following
4343 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4344 * mask = get_cpu_mask(queue)
4345 * for_each_cpu(cpu, mask)
4346 * set->map[x].mq_map[cpu] = queue;
4349 * When we need to remap, the table has to be cleared for
4350 * killing stale mapping since one CPU may not be mapped
4353 for (i = 0; i < set->nr_maps; i++)
4354 blk_mq_clear_mq_map(&set->map[i]);
4356 set->ops->map_queues(set);
4358 BUG_ON(set->nr_maps > 1);
4359 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4363 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4364 int new_nr_hw_queues)
4366 struct blk_mq_tags **new_tags;
4368 if (set->nr_hw_queues >= new_nr_hw_queues)
4371 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4372 GFP_KERNEL, set->numa_node);
4377 memcpy(new_tags, set->tags, set->nr_hw_queues *
4378 sizeof(*set->tags));
4380 set->tags = new_tags;
4382 set->nr_hw_queues = new_nr_hw_queues;
4387 * Alloc a tag set to be associated with one or more request queues.
4388 * May fail with EINVAL for various error conditions. May adjust the
4389 * requested depth down, if it's too large. In that case, the set
4390 * value will be stored in set->queue_depth.
4392 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4396 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4398 if (!set->nr_hw_queues)
4400 if (!set->queue_depth)
4402 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4405 if (!set->ops->queue_rq)
4408 if (!set->ops->get_budget ^ !set->ops->put_budget)
4411 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4412 pr_info("blk-mq: reduced tag depth to %u\n",
4414 set->queue_depth = BLK_MQ_MAX_DEPTH;
4419 else if (set->nr_maps > HCTX_MAX_TYPES)
4423 * If a crashdump is active, then we are potentially in a very
4424 * memory constrained environment. Limit us to 1 queue and
4425 * 64 tags to prevent using too much memory.
4427 if (is_kdump_kernel()) {
4428 set->nr_hw_queues = 1;
4430 set->queue_depth = min(64U, set->queue_depth);
4433 * There is no use for more h/w queues than cpus if we just have
4436 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4437 set->nr_hw_queues = nr_cpu_ids;
4439 if (set->flags & BLK_MQ_F_BLOCKING) {
4440 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4443 ret = init_srcu_struct(set->srcu);
4449 set->tags = kcalloc_node(set->nr_hw_queues,
4450 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4453 goto out_cleanup_srcu;
4455 for (i = 0; i < set->nr_maps; i++) {
4456 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4457 sizeof(set->map[i].mq_map[0]),
4458 GFP_KERNEL, set->numa_node);
4459 if (!set->map[i].mq_map)
4460 goto out_free_mq_map;
4461 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4464 blk_mq_update_queue_map(set);
4466 ret = blk_mq_alloc_set_map_and_rqs(set);
4468 goto out_free_mq_map;
4470 mutex_init(&set->tag_list_lock);
4471 INIT_LIST_HEAD(&set->tag_list);
4476 for (i = 0; i < set->nr_maps; i++) {
4477 kfree(set->map[i].mq_map);
4478 set->map[i].mq_map = NULL;
4483 if (set->flags & BLK_MQ_F_BLOCKING)
4484 cleanup_srcu_struct(set->srcu);
4486 if (set->flags & BLK_MQ_F_BLOCKING)
4490 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4492 /* allocate and initialize a tagset for a simple single-queue device */
4493 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4494 const struct blk_mq_ops *ops, unsigned int queue_depth,
4495 unsigned int set_flags)
4497 memset(set, 0, sizeof(*set));
4499 set->nr_hw_queues = 1;
4501 set->queue_depth = queue_depth;
4502 set->numa_node = NUMA_NO_NODE;
4503 set->flags = set_flags;
4504 return blk_mq_alloc_tag_set(set);
4506 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4508 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4512 for (i = 0; i < set->nr_hw_queues; i++)
4513 __blk_mq_free_map_and_rqs(set, i);
4515 if (blk_mq_is_shared_tags(set->flags)) {
4516 blk_mq_free_map_and_rqs(set, set->shared_tags,
4517 BLK_MQ_NO_HCTX_IDX);
4520 for (j = 0; j < set->nr_maps; j++) {
4521 kfree(set->map[j].mq_map);
4522 set->map[j].mq_map = NULL;
4527 if (set->flags & BLK_MQ_F_BLOCKING) {
4528 cleanup_srcu_struct(set->srcu);
4532 EXPORT_SYMBOL(blk_mq_free_tag_set);
4534 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4536 struct blk_mq_tag_set *set = q->tag_set;
4537 struct blk_mq_hw_ctx *hctx;
4544 if (q->nr_requests == nr)
4547 blk_mq_freeze_queue(q);
4548 blk_mq_quiesce_queue(q);
4551 queue_for_each_hw_ctx(q, hctx, i) {
4555 * If we're using an MQ scheduler, just update the scheduler
4556 * queue depth. This is similar to what the old code would do.
4558 if (hctx->sched_tags) {
4559 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4562 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4567 if (q->elevator && q->elevator->type->ops.depth_updated)
4568 q->elevator->type->ops.depth_updated(hctx);
4571 q->nr_requests = nr;
4572 if (blk_mq_is_shared_tags(set->flags)) {
4574 blk_mq_tag_update_sched_shared_tags(q);
4576 blk_mq_tag_resize_shared_tags(set, nr);
4580 blk_mq_unquiesce_queue(q);
4581 blk_mq_unfreeze_queue(q);
4587 * request_queue and elevator_type pair.
4588 * It is just used by __blk_mq_update_nr_hw_queues to cache
4589 * the elevator_type associated with a request_queue.
4591 struct blk_mq_qe_pair {
4592 struct list_head node;
4593 struct request_queue *q;
4594 struct elevator_type *type;
4598 * Cache the elevator_type in qe pair list and switch the
4599 * io scheduler to 'none'
4601 static bool blk_mq_elv_switch_none(struct list_head *head,
4602 struct request_queue *q)
4604 struct blk_mq_qe_pair *qe;
4609 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4613 /* q->elevator needs protection from ->sysfs_lock */
4614 mutex_lock(&q->sysfs_lock);
4616 INIT_LIST_HEAD(&qe->node);
4618 qe->type = q->elevator->type;
4619 /* keep a reference to the elevator module as we'll switch back */
4620 __elevator_get(qe->type);
4621 list_add(&qe->node, head);
4622 elevator_disable(q);
4623 mutex_unlock(&q->sysfs_lock);
4628 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4629 struct request_queue *q)
4631 struct blk_mq_qe_pair *qe;
4633 list_for_each_entry(qe, head, node)
4640 static void blk_mq_elv_switch_back(struct list_head *head,
4641 struct request_queue *q)
4643 struct blk_mq_qe_pair *qe;
4644 struct elevator_type *t;
4646 qe = blk_lookup_qe_pair(head, q);
4650 list_del(&qe->node);
4653 mutex_lock(&q->sysfs_lock);
4654 elevator_switch(q, t);
4655 /* drop the reference acquired in blk_mq_elv_switch_none */
4657 mutex_unlock(&q->sysfs_lock);
4660 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4663 struct request_queue *q;
4665 int prev_nr_hw_queues;
4667 lockdep_assert_held(&set->tag_list_lock);
4669 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4670 nr_hw_queues = nr_cpu_ids;
4671 if (nr_hw_queues < 1)
4673 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4676 list_for_each_entry(q, &set->tag_list, tag_set_list)
4677 blk_mq_freeze_queue(q);
4679 * Switch IO scheduler to 'none', cleaning up the data associated
4680 * with the previous scheduler. We will switch back once we are done
4681 * updating the new sw to hw queue mappings.
4683 list_for_each_entry(q, &set->tag_list, tag_set_list)
4684 if (!blk_mq_elv_switch_none(&head, q))
4687 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4688 blk_mq_debugfs_unregister_hctxs(q);
4689 blk_mq_sysfs_unregister_hctxs(q);
4692 prev_nr_hw_queues = set->nr_hw_queues;
4693 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4697 blk_mq_update_queue_map(set);
4698 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4699 blk_mq_realloc_hw_ctxs(set, q);
4700 blk_mq_update_poll_flag(q);
4701 if (q->nr_hw_queues != set->nr_hw_queues) {
4702 int i = prev_nr_hw_queues;
4704 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4705 nr_hw_queues, prev_nr_hw_queues);
4706 for (; i < set->nr_hw_queues; i++)
4707 __blk_mq_free_map_and_rqs(set, i);
4709 set->nr_hw_queues = prev_nr_hw_queues;
4710 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4713 blk_mq_map_swqueue(q);
4717 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4718 blk_mq_sysfs_register_hctxs(q);
4719 blk_mq_debugfs_register_hctxs(q);
4723 list_for_each_entry(q, &set->tag_list, tag_set_list)
4724 blk_mq_elv_switch_back(&head, q);
4726 list_for_each_entry(q, &set->tag_list, tag_set_list)
4727 blk_mq_unfreeze_queue(q);
4730 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4732 mutex_lock(&set->tag_list_lock);
4733 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4734 mutex_unlock(&set->tag_list_lock);
4736 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4738 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4741 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4742 long state = get_current_state();
4746 ret = q->mq_ops->poll(hctx, iob);
4748 __set_current_state(TASK_RUNNING);
4752 if (signal_pending_state(state, current))
4753 __set_current_state(TASK_RUNNING);
4754 if (task_is_running(current))
4757 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4760 } while (!need_resched());
4762 __set_current_state(TASK_RUNNING);
4766 unsigned int blk_mq_rq_cpu(struct request *rq)
4768 return rq->mq_ctx->cpu;
4770 EXPORT_SYMBOL(blk_mq_rq_cpu);
4772 void blk_mq_cancel_work_sync(struct request_queue *q)
4774 struct blk_mq_hw_ctx *hctx;
4777 cancel_delayed_work_sync(&q->requeue_work);
4779 queue_for_each_hw_ctx(q, hctx, i)
4780 cancel_delayed_work_sync(&hctx->run_work);
4783 static int __init blk_mq_init(void)
4787 for_each_possible_cpu(i)
4788 init_llist_head(&per_cpu(blk_cpu_done, i));
4789 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4791 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4792 "block/softirq:dead", NULL,
4793 blk_softirq_cpu_dead);
4794 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4795 blk_mq_hctx_notify_dead);
4796 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4797 blk_mq_hctx_notify_online,
4798 blk_mq_hctx_notify_offline);
4801 subsys_initcall(blk_mq_init);