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->mq_ctx->cpu == raw_smp_processor_id()) ||
1190 rq->cmd_flags & REQ_POLLED)
1193 if (blk_mq_complete_need_ipi(rq)) {
1194 blk_mq_complete_send_ipi(rq);
1198 if (rq->q->nr_hw_queues == 1) {
1199 blk_mq_raise_softirq(rq);
1204 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1207 * blk_mq_complete_request - end I/O on a request
1208 * @rq: the request being processed
1211 * Complete a request by scheduling the ->complete_rq operation.
1213 void blk_mq_complete_request(struct request *rq)
1215 if (!blk_mq_complete_request_remote(rq))
1216 rq->q->mq_ops->complete(rq);
1218 EXPORT_SYMBOL(blk_mq_complete_request);
1221 * blk_mq_start_request - Start processing a request
1222 * @rq: Pointer to request to be started
1224 * Function used by device drivers to notify the block layer that a request
1225 * is going to be processed now, so blk layer can do proper initializations
1226 * such as starting the timeout timer.
1228 void blk_mq_start_request(struct request *rq)
1230 struct request_queue *q = rq->q;
1232 trace_block_rq_issue(rq);
1234 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1235 rq->io_start_time_ns = ktime_get_ns();
1236 rq->stats_sectors = blk_rq_sectors(rq);
1237 rq->rq_flags |= RQF_STATS;
1238 rq_qos_issue(q, rq);
1241 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1244 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1246 #ifdef CONFIG_BLK_DEV_INTEGRITY
1247 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1248 q->integrity.profile->prepare_fn(rq);
1250 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1251 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1253 EXPORT_SYMBOL(blk_mq_start_request);
1256 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1257 * queues. This is important for md arrays to benefit from merging
1260 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1262 if (plug->multiple_queues)
1263 return BLK_MAX_REQUEST_COUNT * 2;
1264 return BLK_MAX_REQUEST_COUNT;
1267 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1269 struct request *last = rq_list_peek(&plug->mq_list);
1271 if (!plug->rq_count) {
1272 trace_block_plug(rq->q);
1273 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1274 (!blk_queue_nomerges(rq->q) &&
1275 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1276 blk_mq_flush_plug_list(plug, false);
1278 trace_block_plug(rq->q);
1281 if (!plug->multiple_queues && last && last->q != rq->q)
1282 plug->multiple_queues = true;
1284 * Any request allocated from sched tags can't be issued to
1285 * ->queue_rqs() directly
1287 if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))
1288 plug->has_elevator = true;
1290 rq_list_add(&plug->mq_list, rq);
1295 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1296 * @rq: request to insert
1297 * @at_head: insert request at head or tail of queue
1300 * Insert a fully prepared request at the back of the I/O scheduler queue
1301 * for execution. Don't wait for completion.
1304 * This function will invoke @done directly if the queue is dead.
1306 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1308 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1310 WARN_ON(irqs_disabled());
1311 WARN_ON(!blk_rq_is_passthrough(rq));
1313 blk_account_io_start(rq);
1316 * As plugging can be enabled for passthrough requests on a zoned
1317 * device, directly accessing the plug instead of using blk_mq_plug()
1318 * should not have any consequences.
1320 if (current->plug && !at_head) {
1321 blk_add_rq_to_plug(current->plug, rq);
1325 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1326 blk_mq_run_hw_queue(hctx, false);
1328 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1330 struct blk_rq_wait {
1331 struct completion done;
1335 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1337 struct blk_rq_wait *wait = rq->end_io_data;
1340 complete(&wait->done);
1341 return RQ_END_IO_NONE;
1344 bool blk_rq_is_poll(struct request *rq)
1348 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1352 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1354 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1357 blk_mq_poll(rq->q, blk_rq_to_qc(rq), NULL, 0);
1359 } while (!completion_done(wait));
1363 * blk_execute_rq - insert a request into queue for execution
1364 * @rq: request to insert
1365 * @at_head: insert request at head or tail of queue
1368 * Insert a fully prepared request at the back of the I/O scheduler queue
1369 * for execution and wait for completion.
1370 * Return: The blk_status_t result provided to blk_mq_end_request().
1372 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1374 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1375 struct blk_rq_wait wait = {
1376 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1379 WARN_ON(irqs_disabled());
1380 WARN_ON(!blk_rq_is_passthrough(rq));
1382 rq->end_io_data = &wait;
1383 rq->end_io = blk_end_sync_rq;
1385 blk_account_io_start(rq);
1386 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1387 blk_mq_run_hw_queue(hctx, false);
1389 if (blk_rq_is_poll(rq)) {
1390 blk_rq_poll_completion(rq, &wait.done);
1393 * Prevent hang_check timer from firing at us during very long
1396 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1399 while (!wait_for_completion_io_timeout(&wait.done,
1400 hang_check * (HZ/2)))
1403 wait_for_completion_io(&wait.done);
1408 EXPORT_SYMBOL(blk_execute_rq);
1410 static void __blk_mq_requeue_request(struct request *rq)
1412 struct request_queue *q = rq->q;
1414 blk_mq_put_driver_tag(rq);
1416 trace_block_rq_requeue(rq);
1417 rq_qos_requeue(q, rq);
1419 if (blk_mq_request_started(rq)) {
1420 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1421 rq->rq_flags &= ~RQF_TIMED_OUT;
1425 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1427 struct request_queue *q = rq->q;
1428 unsigned long flags;
1430 __blk_mq_requeue_request(rq);
1432 /* this request will be re-inserted to io scheduler queue */
1433 blk_mq_sched_requeue_request(rq);
1435 spin_lock_irqsave(&q->requeue_lock, flags);
1436 list_add_tail(&rq->queuelist, &q->requeue_list);
1437 spin_unlock_irqrestore(&q->requeue_lock, flags);
1439 if (kick_requeue_list)
1440 blk_mq_kick_requeue_list(q);
1442 EXPORT_SYMBOL(blk_mq_requeue_request);
1444 static void blk_mq_requeue_work(struct work_struct *work)
1446 struct request_queue *q =
1447 container_of(work, struct request_queue, requeue_work.work);
1449 LIST_HEAD(flush_list);
1452 spin_lock_irq(&q->requeue_lock);
1453 list_splice_init(&q->requeue_list, &rq_list);
1454 list_splice_init(&q->flush_list, &flush_list);
1455 spin_unlock_irq(&q->requeue_lock);
1457 while (!list_empty(&rq_list)) {
1458 rq = list_entry(rq_list.next, struct request, queuelist);
1460 * If RQF_DONTPREP ist set, the request has been started by the
1461 * driver already and might have driver-specific data allocated
1462 * already. Insert it into the hctx dispatch list to avoid
1463 * block layer merges for the request.
1465 if (rq->rq_flags & RQF_DONTPREP) {
1466 list_del_init(&rq->queuelist);
1467 blk_mq_request_bypass_insert(rq, 0);
1469 list_del_init(&rq->queuelist);
1470 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1474 while (!list_empty(&flush_list)) {
1475 rq = list_entry(flush_list.next, struct request, queuelist);
1476 list_del_init(&rq->queuelist);
1477 blk_mq_insert_request(rq, 0);
1480 blk_mq_run_hw_queues(q, false);
1483 void blk_mq_kick_requeue_list(struct request_queue *q)
1485 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1487 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1489 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1490 unsigned long msecs)
1492 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1493 msecs_to_jiffies(msecs));
1495 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1497 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1500 * If we find a request that isn't idle we know the queue is busy
1501 * as it's checked in the iter.
1502 * Return false to stop the iteration.
1504 if (blk_mq_request_started(rq)) {
1514 bool blk_mq_queue_inflight(struct request_queue *q)
1518 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1521 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1523 static void blk_mq_rq_timed_out(struct request *req)
1525 req->rq_flags |= RQF_TIMED_OUT;
1526 if (req->q->mq_ops->timeout) {
1527 enum blk_eh_timer_return ret;
1529 ret = req->q->mq_ops->timeout(req);
1530 if (ret == BLK_EH_DONE)
1532 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1538 struct blk_expired_data {
1539 bool has_timedout_rq;
1541 unsigned long timeout_start;
1544 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1546 unsigned long deadline;
1548 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1550 if (rq->rq_flags & RQF_TIMED_OUT)
1553 deadline = READ_ONCE(rq->deadline);
1554 if (time_after_eq(expired->timeout_start, deadline))
1557 if (expired->next == 0)
1558 expired->next = deadline;
1559 else if (time_after(expired->next, deadline))
1560 expired->next = deadline;
1564 void blk_mq_put_rq_ref(struct request *rq)
1566 if (is_flush_rq(rq)) {
1567 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1568 blk_mq_free_request(rq);
1569 } else if (req_ref_put_and_test(rq)) {
1570 __blk_mq_free_request(rq);
1574 static bool blk_mq_check_expired(struct request *rq, void *priv)
1576 struct blk_expired_data *expired = priv;
1579 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1580 * be reallocated underneath the timeout handler's processing, then
1581 * the expire check is reliable. If the request is not expired, then
1582 * it was completed and reallocated as a new request after returning
1583 * from blk_mq_check_expired().
1585 if (blk_mq_req_expired(rq, expired)) {
1586 expired->has_timedout_rq = true;
1592 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1594 struct blk_expired_data *expired = priv;
1596 if (blk_mq_req_expired(rq, expired))
1597 blk_mq_rq_timed_out(rq);
1601 static void blk_mq_timeout_work(struct work_struct *work)
1603 struct request_queue *q =
1604 container_of(work, struct request_queue, timeout_work);
1605 struct blk_expired_data expired = {
1606 .timeout_start = jiffies,
1608 struct blk_mq_hw_ctx *hctx;
1611 /* A deadlock might occur if a request is stuck requiring a
1612 * timeout at the same time a queue freeze is waiting
1613 * completion, since the timeout code would not be able to
1614 * acquire the queue reference here.
1616 * That's why we don't use blk_queue_enter here; instead, we use
1617 * percpu_ref_tryget directly, because we need to be able to
1618 * obtain a reference even in the short window between the queue
1619 * starting to freeze, by dropping the first reference in
1620 * blk_freeze_queue_start, and the moment the last request is
1621 * consumed, marked by the instant q_usage_counter reaches
1624 if (!percpu_ref_tryget(&q->q_usage_counter))
1627 /* check if there is any timed-out request */
1628 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1629 if (expired.has_timedout_rq) {
1631 * Before walking tags, we must ensure any submit started
1632 * before the current time has finished. Since the submit
1633 * uses srcu or rcu, wait for a synchronization point to
1634 * ensure all running submits have finished
1636 blk_mq_wait_quiesce_done(q->tag_set);
1639 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1642 if (expired.next != 0) {
1643 mod_timer(&q->timeout, expired.next);
1646 * Request timeouts are handled as a forward rolling timer. If
1647 * we end up here it means that no requests are pending and
1648 * also that no request has been pending for a while. Mark
1649 * each hctx as idle.
1651 queue_for_each_hw_ctx(q, hctx, i) {
1652 /* the hctx may be unmapped, so check it here */
1653 if (blk_mq_hw_queue_mapped(hctx))
1654 blk_mq_tag_idle(hctx);
1660 struct flush_busy_ctx_data {
1661 struct blk_mq_hw_ctx *hctx;
1662 struct list_head *list;
1665 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1667 struct flush_busy_ctx_data *flush_data = data;
1668 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1669 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1670 enum hctx_type type = hctx->type;
1672 spin_lock(&ctx->lock);
1673 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1674 sbitmap_clear_bit(sb, bitnr);
1675 spin_unlock(&ctx->lock);
1680 * Process software queues that have been marked busy, splicing them
1681 * to the for-dispatch
1683 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1685 struct flush_busy_ctx_data data = {
1690 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1692 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1694 struct dispatch_rq_data {
1695 struct blk_mq_hw_ctx *hctx;
1699 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1702 struct dispatch_rq_data *dispatch_data = data;
1703 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1704 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1705 enum hctx_type type = hctx->type;
1707 spin_lock(&ctx->lock);
1708 if (!list_empty(&ctx->rq_lists[type])) {
1709 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1710 list_del_init(&dispatch_data->rq->queuelist);
1711 if (list_empty(&ctx->rq_lists[type]))
1712 sbitmap_clear_bit(sb, bitnr);
1714 spin_unlock(&ctx->lock);
1716 return !dispatch_data->rq;
1719 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1720 struct blk_mq_ctx *start)
1722 unsigned off = start ? start->index_hw[hctx->type] : 0;
1723 struct dispatch_rq_data data = {
1728 __sbitmap_for_each_set(&hctx->ctx_map, off,
1729 dispatch_rq_from_ctx, &data);
1734 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1736 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1737 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1740 blk_mq_tag_busy(rq->mq_hctx);
1742 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1743 bt = &rq->mq_hctx->tags->breserved_tags;
1746 if (!hctx_may_queue(rq->mq_hctx, bt))
1750 tag = __sbitmap_queue_get(bt);
1751 if (tag == BLK_MQ_NO_TAG)
1754 rq->tag = tag + tag_offset;
1758 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1760 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1763 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1764 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1765 rq->rq_flags |= RQF_MQ_INFLIGHT;
1766 __blk_mq_inc_active_requests(hctx);
1768 hctx->tags->rqs[rq->tag] = rq;
1772 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1773 int flags, void *key)
1775 struct blk_mq_hw_ctx *hctx;
1777 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1779 spin_lock(&hctx->dispatch_wait_lock);
1780 if (!list_empty(&wait->entry)) {
1781 struct sbitmap_queue *sbq;
1783 list_del_init(&wait->entry);
1784 sbq = &hctx->tags->bitmap_tags;
1785 atomic_dec(&sbq->ws_active);
1787 spin_unlock(&hctx->dispatch_wait_lock);
1789 blk_mq_run_hw_queue(hctx, true);
1794 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1795 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1796 * restart. For both cases, take care to check the condition again after
1797 * marking us as waiting.
1799 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1802 struct sbitmap_queue *sbq;
1803 struct wait_queue_head *wq;
1804 wait_queue_entry_t *wait;
1807 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1808 !(blk_mq_is_shared_tags(hctx->flags))) {
1809 blk_mq_sched_mark_restart_hctx(hctx);
1812 * It's possible that a tag was freed in the window between the
1813 * allocation failure and adding the hardware queue to the wait
1816 * Don't clear RESTART here, someone else could have set it.
1817 * At most this will cost an extra queue run.
1819 return blk_mq_get_driver_tag(rq);
1822 wait = &hctx->dispatch_wait;
1823 if (!list_empty_careful(&wait->entry))
1826 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1827 sbq = &hctx->tags->breserved_tags;
1829 sbq = &hctx->tags->bitmap_tags;
1830 wq = &bt_wait_ptr(sbq, hctx)->wait;
1832 spin_lock_irq(&wq->lock);
1833 spin_lock(&hctx->dispatch_wait_lock);
1834 if (!list_empty(&wait->entry)) {
1835 spin_unlock(&hctx->dispatch_wait_lock);
1836 spin_unlock_irq(&wq->lock);
1840 atomic_inc(&sbq->ws_active);
1841 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1842 __add_wait_queue(wq, wait);
1845 * It's possible that a tag was freed in the window between the
1846 * allocation failure and adding the hardware queue to the wait
1849 ret = blk_mq_get_driver_tag(rq);
1851 spin_unlock(&hctx->dispatch_wait_lock);
1852 spin_unlock_irq(&wq->lock);
1857 * We got a tag, remove ourselves from the wait queue to ensure
1858 * someone else gets the wakeup.
1860 list_del_init(&wait->entry);
1861 atomic_dec(&sbq->ws_active);
1862 spin_unlock(&hctx->dispatch_wait_lock);
1863 spin_unlock_irq(&wq->lock);
1868 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1869 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1871 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1872 * - EWMA is one simple way to compute running average value
1873 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1874 * - take 4 as factor for avoiding to get too small(0) result, and this
1875 * factor doesn't matter because EWMA decreases exponentially
1877 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1881 ewma = hctx->dispatch_busy;
1886 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1888 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1889 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1891 hctx->dispatch_busy = ewma;
1894 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1896 static void blk_mq_handle_dev_resource(struct request *rq,
1897 struct list_head *list)
1899 list_add(&rq->queuelist, list);
1900 __blk_mq_requeue_request(rq);
1903 static void blk_mq_handle_zone_resource(struct request *rq,
1904 struct list_head *zone_list)
1907 * If we end up here it is because we cannot dispatch a request to a
1908 * specific zone due to LLD level zone-write locking or other zone
1909 * related resource not being available. In this case, set the request
1910 * aside in zone_list for retrying it later.
1912 list_add(&rq->queuelist, zone_list);
1913 __blk_mq_requeue_request(rq);
1916 enum prep_dispatch {
1918 PREP_DISPATCH_NO_TAG,
1919 PREP_DISPATCH_NO_BUDGET,
1922 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1925 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1926 int budget_token = -1;
1929 budget_token = blk_mq_get_dispatch_budget(rq->q);
1930 if (budget_token < 0) {
1931 blk_mq_put_driver_tag(rq);
1932 return PREP_DISPATCH_NO_BUDGET;
1934 blk_mq_set_rq_budget_token(rq, budget_token);
1937 if (!blk_mq_get_driver_tag(rq)) {
1939 * The initial allocation attempt failed, so we need to
1940 * rerun the hardware queue when a tag is freed. The
1941 * waitqueue takes care of that. If the queue is run
1942 * before we add this entry back on the dispatch list,
1943 * we'll re-run it below.
1945 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1947 * All budgets not got from this function will be put
1948 * together during handling partial dispatch
1951 blk_mq_put_dispatch_budget(rq->q, budget_token);
1952 return PREP_DISPATCH_NO_TAG;
1956 return PREP_DISPATCH_OK;
1959 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1960 static void blk_mq_release_budgets(struct request_queue *q,
1961 struct list_head *list)
1965 list_for_each_entry(rq, list, queuelist) {
1966 int budget_token = blk_mq_get_rq_budget_token(rq);
1968 if (budget_token >= 0)
1969 blk_mq_put_dispatch_budget(q, budget_token);
1974 * blk_mq_commit_rqs will notify driver using bd->last that there is no
1975 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
1977 * Attention, we should explicitly call this in unusual cases:
1978 * 1) did not queue everything initially scheduled to queue
1979 * 2) the last attempt to queue a request failed
1981 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
1984 if (hctx->queue->mq_ops->commit_rqs && queued) {
1985 trace_block_unplug(hctx->queue, queued, !from_schedule);
1986 hctx->queue->mq_ops->commit_rqs(hctx);
1991 * Returns true if we did some work AND can potentially do more.
1993 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1994 unsigned int nr_budgets)
1996 enum prep_dispatch prep;
1997 struct request_queue *q = hctx->queue;
2000 blk_status_t ret = BLK_STS_OK;
2001 LIST_HEAD(zone_list);
2002 bool needs_resource = false;
2004 if (list_empty(list))
2008 * Now process all the entries, sending them to the driver.
2012 struct blk_mq_queue_data bd;
2014 rq = list_first_entry(list, struct request, queuelist);
2016 WARN_ON_ONCE(hctx != rq->mq_hctx);
2017 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2018 if (prep != PREP_DISPATCH_OK)
2021 list_del_init(&rq->queuelist);
2024 bd.last = list_empty(list);
2027 * once the request is queued to lld, no need to cover the
2032 ret = q->mq_ops->queue_rq(hctx, &bd);
2037 case BLK_STS_RESOURCE:
2038 needs_resource = true;
2040 case BLK_STS_DEV_RESOURCE:
2041 blk_mq_handle_dev_resource(rq, list);
2043 case BLK_STS_ZONE_RESOURCE:
2045 * Move the request to zone_list and keep going through
2046 * the dispatch list to find more requests the drive can
2049 blk_mq_handle_zone_resource(rq, &zone_list);
2050 needs_resource = true;
2053 blk_mq_end_request(rq, ret);
2055 } while (!list_empty(list));
2057 if (!list_empty(&zone_list))
2058 list_splice_tail_init(&zone_list, list);
2060 /* If we didn't flush the entire list, we could have told the driver
2061 * there was more coming, but that turned out to be a lie.
2063 if (!list_empty(list) || ret != BLK_STS_OK)
2064 blk_mq_commit_rqs(hctx, queued, false);
2067 * Any items that need requeuing? Stuff them into hctx->dispatch,
2068 * that is where we will continue on next queue run.
2070 if (!list_empty(list)) {
2072 /* For non-shared tags, the RESTART check will suffice */
2073 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2074 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2075 blk_mq_is_shared_tags(hctx->flags));
2078 blk_mq_release_budgets(q, list);
2080 spin_lock(&hctx->lock);
2081 list_splice_tail_init(list, &hctx->dispatch);
2082 spin_unlock(&hctx->lock);
2085 * Order adding requests to hctx->dispatch and checking
2086 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2087 * in blk_mq_sched_restart(). Avoid restart code path to
2088 * miss the new added requests to hctx->dispatch, meantime
2089 * SCHED_RESTART is observed here.
2094 * If SCHED_RESTART was set by the caller of this function and
2095 * it is no longer set that means that it was cleared by another
2096 * thread and hence that a queue rerun is needed.
2098 * If 'no_tag' is set, that means that we failed getting
2099 * a driver tag with an I/O scheduler attached. If our dispatch
2100 * waitqueue is no longer active, ensure that we run the queue
2101 * AFTER adding our entries back to the list.
2103 * If no I/O scheduler has been configured it is possible that
2104 * the hardware queue got stopped and restarted before requests
2105 * were pushed back onto the dispatch list. Rerun the queue to
2106 * avoid starvation. Notes:
2107 * - blk_mq_run_hw_queue() checks whether or not a queue has
2108 * been stopped before rerunning a queue.
2109 * - Some but not all block drivers stop a queue before
2110 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2113 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2114 * bit is set, run queue after a delay to avoid IO stalls
2115 * that could otherwise occur if the queue is idle. We'll do
2116 * similar if we couldn't get budget or couldn't lock a zone
2117 * and SCHED_RESTART is set.
2119 needs_restart = blk_mq_sched_needs_restart(hctx);
2120 if (prep == PREP_DISPATCH_NO_BUDGET)
2121 needs_resource = true;
2122 if (!needs_restart ||
2123 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2124 blk_mq_run_hw_queue(hctx, true);
2125 else if (needs_resource)
2126 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2128 blk_mq_update_dispatch_busy(hctx, true);
2132 blk_mq_update_dispatch_busy(hctx, false);
2136 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2138 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2140 if (cpu >= nr_cpu_ids)
2141 cpu = cpumask_first(hctx->cpumask);
2146 * It'd be great if the workqueue API had a way to pass
2147 * in a mask and had some smarts for more clever placement.
2148 * For now we just round-robin here, switching for every
2149 * BLK_MQ_CPU_WORK_BATCH queued items.
2151 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2154 int next_cpu = hctx->next_cpu;
2156 if (hctx->queue->nr_hw_queues == 1)
2157 return WORK_CPU_UNBOUND;
2159 if (--hctx->next_cpu_batch <= 0) {
2161 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2163 if (next_cpu >= nr_cpu_ids)
2164 next_cpu = blk_mq_first_mapped_cpu(hctx);
2165 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2169 * Do unbound schedule if we can't find a online CPU for this hctx,
2170 * and it should only happen in the path of handling CPU DEAD.
2172 if (!cpu_online(next_cpu)) {
2179 * Make sure to re-select CPU next time once after CPUs
2180 * in hctx->cpumask become online again.
2182 hctx->next_cpu = next_cpu;
2183 hctx->next_cpu_batch = 1;
2184 return WORK_CPU_UNBOUND;
2187 hctx->next_cpu = next_cpu;
2192 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2193 * @hctx: Pointer to the hardware queue to run.
2194 * @msecs: Milliseconds of delay to wait before running the queue.
2196 * Run a hardware queue asynchronously with a delay of @msecs.
2198 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2200 if (unlikely(blk_mq_hctx_stopped(hctx)))
2202 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2203 msecs_to_jiffies(msecs));
2205 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2208 * blk_mq_run_hw_queue - Start to run a hardware queue.
2209 * @hctx: Pointer to the hardware queue to run.
2210 * @async: If we want to run the queue asynchronously.
2212 * Check if the request queue is not in a quiesced state and if there are
2213 * pending requests to be sent. If this is true, run the queue to send requests
2216 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2221 * We can't run the queue inline with interrupts disabled.
2223 WARN_ON_ONCE(!async && in_interrupt());
2226 * When queue is quiesced, we may be switching io scheduler, or
2227 * updating nr_hw_queues, or other things, and we can't run queue
2228 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2230 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2233 __blk_mq_run_dispatch_ops(hctx->queue, false,
2234 need_run = !blk_queue_quiesced(hctx->queue) &&
2235 blk_mq_hctx_has_pending(hctx));
2240 if (async || (hctx->flags & BLK_MQ_F_BLOCKING) ||
2241 !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2242 blk_mq_delay_run_hw_queue(hctx, 0);
2246 blk_mq_run_dispatch_ops(hctx->queue,
2247 blk_mq_sched_dispatch_requests(hctx));
2249 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2252 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2255 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2257 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2259 * If the IO scheduler does not respect hardware queues when
2260 * dispatching, we just don't bother with multiple HW queues and
2261 * dispatch from hctx for the current CPU since running multiple queues
2262 * just causes lock contention inside the scheduler and pointless cache
2265 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2267 if (!blk_mq_hctx_stopped(hctx))
2273 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2274 * @q: Pointer to the request queue to run.
2275 * @async: If we want to run the queue asynchronously.
2277 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2279 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2283 if (blk_queue_sq_sched(q))
2284 sq_hctx = blk_mq_get_sq_hctx(q);
2285 queue_for_each_hw_ctx(q, hctx, i) {
2286 if (blk_mq_hctx_stopped(hctx))
2289 * Dispatch from this hctx either if there's no hctx preferred
2290 * by IO scheduler or if it has requests that bypass the
2293 if (!sq_hctx || sq_hctx == hctx ||
2294 !list_empty_careful(&hctx->dispatch))
2295 blk_mq_run_hw_queue(hctx, async);
2298 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2301 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2302 * @q: Pointer to the request queue to run.
2303 * @msecs: Milliseconds of delay to wait before running the queues.
2305 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2307 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2311 if (blk_queue_sq_sched(q))
2312 sq_hctx = blk_mq_get_sq_hctx(q);
2313 queue_for_each_hw_ctx(q, hctx, i) {
2314 if (blk_mq_hctx_stopped(hctx))
2317 * If there is already a run_work pending, leave the
2318 * pending delay untouched. Otherwise, a hctx can stall
2319 * if another hctx is re-delaying the other's work
2320 * before the work executes.
2322 if (delayed_work_pending(&hctx->run_work))
2325 * Dispatch from this hctx either if there's no hctx preferred
2326 * by IO scheduler or if it has requests that bypass the
2329 if (!sq_hctx || sq_hctx == hctx ||
2330 !list_empty_careful(&hctx->dispatch))
2331 blk_mq_delay_run_hw_queue(hctx, msecs);
2334 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2337 * This function is often used for pausing .queue_rq() by driver when
2338 * there isn't enough resource or some conditions aren't satisfied, and
2339 * BLK_STS_RESOURCE is usually returned.
2341 * We do not guarantee that dispatch can be drained or blocked
2342 * after blk_mq_stop_hw_queue() returns. Please use
2343 * blk_mq_quiesce_queue() for that requirement.
2345 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2347 cancel_delayed_work(&hctx->run_work);
2349 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2351 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2354 * This function is often used for pausing .queue_rq() by driver when
2355 * there isn't enough resource or some conditions aren't satisfied, and
2356 * BLK_STS_RESOURCE is usually returned.
2358 * We do not guarantee that dispatch can be drained or blocked
2359 * after blk_mq_stop_hw_queues() returns. Please use
2360 * blk_mq_quiesce_queue() for that requirement.
2362 void blk_mq_stop_hw_queues(struct request_queue *q)
2364 struct blk_mq_hw_ctx *hctx;
2367 queue_for_each_hw_ctx(q, hctx, i)
2368 blk_mq_stop_hw_queue(hctx);
2370 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2372 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2374 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2376 blk_mq_run_hw_queue(hctx, false);
2378 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2380 void blk_mq_start_hw_queues(struct request_queue *q)
2382 struct blk_mq_hw_ctx *hctx;
2385 queue_for_each_hw_ctx(q, hctx, i)
2386 blk_mq_start_hw_queue(hctx);
2388 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2390 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2392 if (!blk_mq_hctx_stopped(hctx))
2395 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2396 blk_mq_run_hw_queue(hctx, async);
2398 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2400 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2402 struct blk_mq_hw_ctx *hctx;
2405 queue_for_each_hw_ctx(q, hctx, i)
2406 blk_mq_start_stopped_hw_queue(hctx, async);
2408 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2410 static void blk_mq_run_work_fn(struct work_struct *work)
2412 struct blk_mq_hw_ctx *hctx =
2413 container_of(work, struct blk_mq_hw_ctx, run_work.work);
2415 blk_mq_run_dispatch_ops(hctx->queue,
2416 blk_mq_sched_dispatch_requests(hctx));
2420 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2421 * @rq: Pointer to request to be inserted.
2422 * @flags: BLK_MQ_INSERT_*
2424 * Should only be used carefully, when the caller knows we want to
2425 * bypass a potential IO scheduler on the target device.
2427 static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2429 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2431 spin_lock(&hctx->lock);
2432 if (flags & BLK_MQ_INSERT_AT_HEAD)
2433 list_add(&rq->queuelist, &hctx->dispatch);
2435 list_add_tail(&rq->queuelist, &hctx->dispatch);
2436 spin_unlock(&hctx->lock);
2439 static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2440 struct blk_mq_ctx *ctx, struct list_head *list,
2441 bool run_queue_async)
2444 enum hctx_type type = hctx->type;
2447 * Try to issue requests directly if the hw queue isn't busy to save an
2448 * extra enqueue & dequeue to the sw queue.
2450 if (!hctx->dispatch_busy && !run_queue_async) {
2451 blk_mq_run_dispatch_ops(hctx->queue,
2452 blk_mq_try_issue_list_directly(hctx, list));
2453 if (list_empty(list))
2458 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2461 list_for_each_entry(rq, list, queuelist) {
2462 BUG_ON(rq->mq_ctx != ctx);
2463 trace_block_rq_insert(rq);
2466 spin_lock(&ctx->lock);
2467 list_splice_tail_init(list, &ctx->rq_lists[type]);
2468 blk_mq_hctx_mark_pending(hctx, ctx);
2469 spin_unlock(&ctx->lock);
2471 blk_mq_run_hw_queue(hctx, run_queue_async);
2474 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2476 struct request_queue *q = rq->q;
2477 struct blk_mq_ctx *ctx = rq->mq_ctx;
2478 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2480 if (blk_rq_is_passthrough(rq)) {
2482 * Passthrough request have to be added to hctx->dispatch
2483 * directly. The device may be in a situation where it can't
2484 * handle FS request, and always returns BLK_STS_RESOURCE for
2485 * them, which gets them added to hctx->dispatch.
2487 * If a passthrough request is required to unblock the queues,
2488 * and it is added to the scheduler queue, there is no chance to
2489 * dispatch it given we prioritize requests in hctx->dispatch.
2491 blk_mq_request_bypass_insert(rq, flags);
2492 } else if (req_op(rq) == REQ_OP_FLUSH) {
2494 * Firstly normal IO request is inserted to scheduler queue or
2495 * sw queue, meantime we add flush request to dispatch queue(
2496 * hctx->dispatch) directly and there is at most one in-flight
2497 * flush request for each hw queue, so it doesn't matter to add
2498 * flush request to tail or front of the dispatch queue.
2500 * Secondly in case of NCQ, flush request belongs to non-NCQ
2501 * command, and queueing it will fail when there is any
2502 * in-flight normal IO request(NCQ command). When adding flush
2503 * rq to the front of hctx->dispatch, it is easier to introduce
2504 * extra time to flush rq's latency because of S_SCHED_RESTART
2505 * compared with adding to the tail of dispatch queue, then
2506 * chance of flush merge is increased, and less flush requests
2507 * will be issued to controller. It is observed that ~10% time
2508 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2509 * drive when adding flush rq to the front of hctx->dispatch.
2511 * Simply queue flush rq to the front of hctx->dispatch so that
2512 * intensive flush workloads can benefit in case of NCQ HW.
2514 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2515 } else if (q->elevator) {
2518 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2520 list_add(&rq->queuelist, &list);
2521 q->elevator->type->ops.insert_requests(hctx, &list, flags);
2523 trace_block_rq_insert(rq);
2525 spin_lock(&ctx->lock);
2526 if (flags & BLK_MQ_INSERT_AT_HEAD)
2527 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2529 list_add_tail(&rq->queuelist,
2530 &ctx->rq_lists[hctx->type]);
2531 blk_mq_hctx_mark_pending(hctx, ctx);
2532 spin_unlock(&ctx->lock);
2536 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2537 unsigned int nr_segs)
2541 if (bio->bi_opf & REQ_RAHEAD)
2542 rq->cmd_flags |= REQ_FAILFAST_MASK;
2544 rq->__sector = bio->bi_iter.bi_sector;
2545 blk_rq_bio_prep(rq, bio, nr_segs);
2547 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2548 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2551 blk_account_io_start(rq);
2554 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2555 struct request *rq, bool last)
2557 struct request_queue *q = rq->q;
2558 struct blk_mq_queue_data bd = {
2565 * For OK queue, we are done. For error, caller may kill it.
2566 * Any other error (busy), just add it to our list as we
2567 * previously would have done.
2569 ret = q->mq_ops->queue_rq(hctx, &bd);
2572 blk_mq_update_dispatch_busy(hctx, false);
2574 case BLK_STS_RESOURCE:
2575 case BLK_STS_DEV_RESOURCE:
2576 blk_mq_update_dispatch_busy(hctx, true);
2577 __blk_mq_requeue_request(rq);
2580 blk_mq_update_dispatch_busy(hctx, false);
2587 static bool blk_mq_get_budget_and_tag(struct request *rq)
2591 budget_token = blk_mq_get_dispatch_budget(rq->q);
2592 if (budget_token < 0)
2594 blk_mq_set_rq_budget_token(rq, budget_token);
2595 if (!blk_mq_get_driver_tag(rq)) {
2596 blk_mq_put_dispatch_budget(rq->q, budget_token);
2603 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2604 * @hctx: Pointer of the associated hardware queue.
2605 * @rq: Pointer to request to be sent.
2607 * If the device has enough resources to accept a new request now, send the
2608 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2609 * we can try send it another time in the future. Requests inserted at this
2610 * queue have higher priority.
2612 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2617 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2618 blk_mq_insert_request(rq, 0);
2622 if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
2623 blk_mq_insert_request(rq, 0);
2624 blk_mq_run_hw_queue(hctx, false);
2628 ret = __blk_mq_issue_directly(hctx, rq, true);
2632 case BLK_STS_RESOURCE:
2633 case BLK_STS_DEV_RESOURCE:
2634 blk_mq_request_bypass_insert(rq, 0);
2635 blk_mq_run_hw_queue(hctx, false);
2638 blk_mq_end_request(rq, ret);
2643 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2645 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2647 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2648 blk_mq_insert_request(rq, 0);
2652 if (!blk_mq_get_budget_and_tag(rq))
2653 return BLK_STS_RESOURCE;
2654 return __blk_mq_issue_directly(hctx, rq, last);
2657 static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2659 struct blk_mq_hw_ctx *hctx = NULL;
2662 blk_status_t ret = BLK_STS_OK;
2664 while ((rq = rq_list_pop(&plug->mq_list))) {
2665 bool last = rq_list_empty(plug->mq_list);
2667 if (hctx != rq->mq_hctx) {
2669 blk_mq_commit_rqs(hctx, queued, false);
2675 ret = blk_mq_request_issue_directly(rq, last);
2680 case BLK_STS_RESOURCE:
2681 case BLK_STS_DEV_RESOURCE:
2682 blk_mq_request_bypass_insert(rq, 0);
2683 blk_mq_run_hw_queue(hctx, false);
2686 blk_mq_end_request(rq, ret);
2692 if (ret != BLK_STS_OK)
2693 blk_mq_commit_rqs(hctx, queued, false);
2696 static void __blk_mq_flush_plug_list(struct request_queue *q,
2697 struct blk_plug *plug)
2699 if (blk_queue_quiesced(q))
2701 q->mq_ops->queue_rqs(&plug->mq_list);
2704 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2706 struct blk_mq_hw_ctx *this_hctx = NULL;
2707 struct blk_mq_ctx *this_ctx = NULL;
2708 struct request *requeue_list = NULL;
2709 struct request **requeue_lastp = &requeue_list;
2710 unsigned int depth = 0;
2711 bool is_passthrough = false;
2715 struct request *rq = rq_list_pop(&plug->mq_list);
2718 this_hctx = rq->mq_hctx;
2719 this_ctx = rq->mq_ctx;
2720 is_passthrough = blk_rq_is_passthrough(rq);
2721 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2722 is_passthrough != blk_rq_is_passthrough(rq)) {
2723 rq_list_add_tail(&requeue_lastp, rq);
2726 list_add(&rq->queuelist, &list);
2728 } while (!rq_list_empty(plug->mq_list));
2730 plug->mq_list = requeue_list;
2731 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2733 percpu_ref_get(&this_hctx->queue->q_usage_counter);
2734 /* passthrough requests should never be issued to the I/O scheduler */
2735 if (is_passthrough) {
2736 spin_lock(&this_hctx->lock);
2737 list_splice_tail_init(&list, &this_hctx->dispatch);
2738 spin_unlock(&this_hctx->lock);
2739 blk_mq_run_hw_queue(this_hctx, from_sched);
2740 } else if (this_hctx->queue->elevator) {
2741 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2743 blk_mq_run_hw_queue(this_hctx, from_sched);
2745 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2747 percpu_ref_put(&this_hctx->queue->q_usage_counter);
2750 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2754 if (rq_list_empty(plug->mq_list))
2758 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2759 struct request_queue *q;
2761 rq = rq_list_peek(&plug->mq_list);
2765 * Peek first request and see if we have a ->queue_rqs() hook.
2766 * If we do, we can dispatch the whole plug list in one go. We
2767 * already know at this point that all requests belong to the
2768 * same queue, caller must ensure that's the case.
2770 * Since we pass off the full list to the driver at this point,
2771 * we do not increment the active request count for the queue.
2772 * Bypass shared tags for now because of that.
2774 if (q->mq_ops->queue_rqs &&
2775 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2776 blk_mq_run_dispatch_ops(q,
2777 __blk_mq_flush_plug_list(q, plug));
2778 if (rq_list_empty(plug->mq_list))
2782 blk_mq_run_dispatch_ops(q,
2783 blk_mq_plug_issue_direct(plug));
2784 if (rq_list_empty(plug->mq_list))
2789 blk_mq_dispatch_plug_list(plug, from_schedule);
2790 } while (!rq_list_empty(plug->mq_list));
2793 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2794 struct list_head *list)
2797 blk_status_t ret = BLK_STS_OK;
2799 while (!list_empty(list)) {
2800 struct request *rq = list_first_entry(list, struct request,
2803 list_del_init(&rq->queuelist);
2804 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2809 case BLK_STS_RESOURCE:
2810 case BLK_STS_DEV_RESOURCE:
2811 blk_mq_request_bypass_insert(rq, 0);
2812 if (list_empty(list))
2813 blk_mq_run_hw_queue(hctx, false);
2816 blk_mq_end_request(rq, ret);
2822 if (ret != BLK_STS_OK)
2823 blk_mq_commit_rqs(hctx, queued, false);
2826 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2827 struct bio *bio, unsigned int nr_segs)
2829 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2830 if (blk_attempt_plug_merge(q, bio, nr_segs))
2832 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2838 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2839 struct blk_plug *plug,
2843 struct blk_mq_alloc_data data = {
2846 .cmd_flags = bio->bi_opf,
2850 if (unlikely(bio_queue_enter(bio)))
2853 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2856 rq_qos_throttle(q, bio);
2859 data.nr_tags = plug->nr_ios;
2861 data.cached_rq = &plug->cached_rq;
2864 rq = __blk_mq_alloc_requests(&data);
2867 rq_qos_cleanup(q, bio);
2868 if (bio->bi_opf & REQ_NOWAIT)
2869 bio_wouldblock_error(bio);
2875 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2876 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2879 enum hctx_type type, hctx_type;
2883 rq = rq_list_peek(&plug->cached_rq);
2884 if (!rq || rq->q != q)
2887 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2892 type = blk_mq_get_hctx_type((*bio)->bi_opf);
2893 hctx_type = rq->mq_hctx->type;
2894 if (type != hctx_type &&
2895 !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2897 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2901 * If any qos ->throttle() end up blocking, we will have flushed the
2902 * plug and hence killed the cached_rq list as well. Pop this entry
2903 * before we throttle.
2905 plug->cached_rq = rq_list_next(rq);
2906 rq_qos_throttle(q, *bio);
2908 rq->cmd_flags = (*bio)->bi_opf;
2909 INIT_LIST_HEAD(&rq->queuelist);
2913 static void bio_set_ioprio(struct bio *bio)
2915 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2916 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2917 bio->bi_ioprio = get_current_ioprio();
2918 blkcg_set_ioprio(bio);
2922 * blk_mq_submit_bio - Create and send a request to block device.
2923 * @bio: Bio pointer.
2925 * Builds up a request structure from @q and @bio and send to the device. The
2926 * request may not be queued directly to hardware if:
2927 * * This request can be merged with another one
2928 * * We want to place request at plug queue for possible future merging
2929 * * There is an IO scheduler active at this queue
2931 * It will not queue the request if there is an error with the bio, or at the
2934 void blk_mq_submit_bio(struct bio *bio)
2936 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2937 struct blk_plug *plug = blk_mq_plug(bio);
2938 const int is_sync = op_is_sync(bio->bi_opf);
2939 struct blk_mq_hw_ctx *hctx;
2941 unsigned int nr_segs = 1;
2944 bio = blk_queue_bounce(bio, q);
2945 if (bio_may_exceed_limits(bio, &q->limits)) {
2946 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2951 if (!bio_integrity_prep(bio))
2954 bio_set_ioprio(bio);
2956 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2960 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2965 trace_block_getrq(bio);
2967 rq_qos_track(q, rq, bio);
2969 blk_mq_bio_to_request(rq, bio, nr_segs);
2971 ret = blk_crypto_rq_get_keyslot(rq);
2972 if (ret != BLK_STS_OK) {
2973 bio->bi_status = ret;
2975 blk_mq_free_request(rq);
2979 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
2983 blk_add_rq_to_plug(plug, rq);
2988 if ((rq->rq_flags & RQF_USE_SCHED) ||
2989 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
2990 blk_mq_insert_request(rq, 0);
2991 blk_mq_run_hw_queue(hctx, true);
2993 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
2997 #ifdef CONFIG_BLK_MQ_STACKING
2999 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3000 * @rq: the request being queued
3002 blk_status_t blk_insert_cloned_request(struct request *rq)
3004 struct request_queue *q = rq->q;
3005 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3006 unsigned int max_segments = blk_rq_get_max_segments(rq);
3009 if (blk_rq_sectors(rq) > max_sectors) {
3011 * SCSI device does not have a good way to return if
3012 * Write Same/Zero is actually supported. If a device rejects
3013 * a non-read/write command (discard, write same,etc.) the
3014 * low-level device driver will set the relevant queue limit to
3015 * 0 to prevent blk-lib from issuing more of the offending
3016 * operations. Commands queued prior to the queue limit being
3017 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3018 * errors being propagated to upper layers.
3020 if (max_sectors == 0)
3021 return BLK_STS_NOTSUPP;
3023 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3024 __func__, blk_rq_sectors(rq), max_sectors);
3025 return BLK_STS_IOERR;
3029 * The queue settings related to segment counting may differ from the
3032 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3033 if (rq->nr_phys_segments > max_segments) {
3034 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3035 __func__, rq->nr_phys_segments, max_segments);
3036 return BLK_STS_IOERR;
3039 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3040 return BLK_STS_IOERR;
3042 ret = blk_crypto_rq_get_keyslot(rq);
3043 if (ret != BLK_STS_OK)
3046 blk_account_io_start(rq);
3049 * Since we have a scheduler attached on the top device,
3050 * bypass a potential scheduler on the bottom device for
3053 blk_mq_run_dispatch_ops(q,
3054 ret = blk_mq_request_issue_directly(rq, true));
3056 blk_account_io_done(rq, ktime_get_ns());
3059 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3062 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3063 * @rq: the clone request to be cleaned up
3066 * Free all bios in @rq for a cloned request.
3068 void blk_rq_unprep_clone(struct request *rq)
3072 while ((bio = rq->bio) != NULL) {
3073 rq->bio = bio->bi_next;
3078 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3081 * blk_rq_prep_clone - Helper function to setup clone request
3082 * @rq: the request to be setup
3083 * @rq_src: original request to be cloned
3084 * @bs: bio_set that bios for clone are allocated from
3085 * @gfp_mask: memory allocation mask for bio
3086 * @bio_ctr: setup function to be called for each clone bio.
3087 * Returns %0 for success, non %0 for failure.
3088 * @data: private data to be passed to @bio_ctr
3091 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3092 * Also, pages which the original bios are pointing to are not copied
3093 * and the cloned bios just point same pages.
3094 * So cloned bios must be completed before original bios, which means
3095 * the caller must complete @rq before @rq_src.
3097 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3098 struct bio_set *bs, gfp_t gfp_mask,
3099 int (*bio_ctr)(struct bio *, struct bio *, void *),
3102 struct bio *bio, *bio_src;
3107 __rq_for_each_bio(bio_src, rq_src) {
3108 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3113 if (bio_ctr && bio_ctr(bio, bio_src, data))
3117 rq->biotail->bi_next = bio;
3120 rq->bio = rq->biotail = bio;
3125 /* Copy attributes of the original request to the clone request. */
3126 rq->__sector = blk_rq_pos(rq_src);
3127 rq->__data_len = blk_rq_bytes(rq_src);
3128 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3129 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3130 rq->special_vec = rq_src->special_vec;
3132 rq->nr_phys_segments = rq_src->nr_phys_segments;
3133 rq->ioprio = rq_src->ioprio;
3135 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3143 blk_rq_unprep_clone(rq);
3147 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3148 #endif /* CONFIG_BLK_MQ_STACKING */
3151 * Steal bios from a request and add them to a bio list.
3152 * The request must not have been partially completed before.
3154 void blk_steal_bios(struct bio_list *list, struct request *rq)
3158 list->tail->bi_next = rq->bio;
3160 list->head = rq->bio;
3161 list->tail = rq->biotail;
3169 EXPORT_SYMBOL_GPL(blk_steal_bios);
3171 static size_t order_to_size(unsigned int order)
3173 return (size_t)PAGE_SIZE << order;
3176 /* called before freeing request pool in @tags */
3177 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3178 struct blk_mq_tags *tags)
3181 unsigned long flags;
3184 * There is no need to clear mapping if driver tags is not initialized
3185 * or the mapping belongs to the driver tags.
3187 if (!drv_tags || drv_tags == tags)
3190 list_for_each_entry(page, &tags->page_list, lru) {
3191 unsigned long start = (unsigned long)page_address(page);
3192 unsigned long end = start + order_to_size(page->private);
3195 for (i = 0; i < drv_tags->nr_tags; i++) {
3196 struct request *rq = drv_tags->rqs[i];
3197 unsigned long rq_addr = (unsigned long)rq;
3199 if (rq_addr >= start && rq_addr < end) {
3200 WARN_ON_ONCE(req_ref_read(rq) != 0);
3201 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3207 * Wait until all pending iteration is done.
3209 * Request reference is cleared and it is guaranteed to be observed
3210 * after the ->lock is released.
3212 spin_lock_irqsave(&drv_tags->lock, flags);
3213 spin_unlock_irqrestore(&drv_tags->lock, flags);
3216 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3217 unsigned int hctx_idx)
3219 struct blk_mq_tags *drv_tags;
3222 if (list_empty(&tags->page_list))
3225 if (blk_mq_is_shared_tags(set->flags))
3226 drv_tags = set->shared_tags;
3228 drv_tags = set->tags[hctx_idx];
3230 if (tags->static_rqs && set->ops->exit_request) {
3233 for (i = 0; i < tags->nr_tags; i++) {
3234 struct request *rq = tags->static_rqs[i];
3238 set->ops->exit_request(set, rq, hctx_idx);
3239 tags->static_rqs[i] = NULL;
3243 blk_mq_clear_rq_mapping(drv_tags, tags);
3245 while (!list_empty(&tags->page_list)) {
3246 page = list_first_entry(&tags->page_list, struct page, lru);
3247 list_del_init(&page->lru);
3249 * Remove kmemleak object previously allocated in
3250 * blk_mq_alloc_rqs().
3252 kmemleak_free(page_address(page));
3253 __free_pages(page, page->private);
3257 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3261 kfree(tags->static_rqs);
3262 tags->static_rqs = NULL;
3264 blk_mq_free_tags(tags);
3267 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3268 unsigned int hctx_idx)
3272 for (i = 0; i < set->nr_maps; i++) {
3273 unsigned int start = set->map[i].queue_offset;
3274 unsigned int end = start + set->map[i].nr_queues;
3276 if (hctx_idx >= start && hctx_idx < end)
3280 if (i >= set->nr_maps)
3281 i = HCTX_TYPE_DEFAULT;
3286 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3287 unsigned int hctx_idx)
3289 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3291 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3294 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3295 unsigned int hctx_idx,
3296 unsigned int nr_tags,
3297 unsigned int reserved_tags)
3299 int node = blk_mq_get_hctx_node(set, hctx_idx);
3300 struct blk_mq_tags *tags;
3302 if (node == NUMA_NO_NODE)
3303 node = set->numa_node;
3305 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3306 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3310 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3311 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3316 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3317 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3319 if (!tags->static_rqs)
3327 blk_mq_free_tags(tags);
3331 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3332 unsigned int hctx_idx, int node)
3336 if (set->ops->init_request) {
3337 ret = set->ops->init_request(set, rq, hctx_idx, node);
3342 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3346 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3347 struct blk_mq_tags *tags,
3348 unsigned int hctx_idx, unsigned int depth)
3350 unsigned int i, j, entries_per_page, max_order = 4;
3351 int node = blk_mq_get_hctx_node(set, hctx_idx);
3352 size_t rq_size, left;
3354 if (node == NUMA_NO_NODE)
3355 node = set->numa_node;
3357 INIT_LIST_HEAD(&tags->page_list);
3360 * rq_size is the size of the request plus driver payload, rounded
3361 * to the cacheline size
3363 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3365 left = rq_size * depth;
3367 for (i = 0; i < depth; ) {
3368 int this_order = max_order;
3373 while (this_order && left < order_to_size(this_order - 1))
3377 page = alloc_pages_node(node,
3378 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3384 if (order_to_size(this_order) < rq_size)
3391 page->private = this_order;
3392 list_add_tail(&page->lru, &tags->page_list);
3394 p = page_address(page);
3396 * Allow kmemleak to scan these pages as they contain pointers
3397 * to additional allocations like via ops->init_request().
3399 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3400 entries_per_page = order_to_size(this_order) / rq_size;
3401 to_do = min(entries_per_page, depth - i);
3402 left -= to_do * rq_size;
3403 for (j = 0; j < to_do; j++) {
3404 struct request *rq = p;
3406 tags->static_rqs[i] = rq;
3407 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3408 tags->static_rqs[i] = NULL;
3419 blk_mq_free_rqs(set, tags, hctx_idx);
3423 struct rq_iter_data {
3424 struct blk_mq_hw_ctx *hctx;
3428 static bool blk_mq_has_request(struct request *rq, void *data)
3430 struct rq_iter_data *iter_data = data;
3432 if (rq->mq_hctx != iter_data->hctx)
3434 iter_data->has_rq = true;
3438 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3440 struct blk_mq_tags *tags = hctx->sched_tags ?
3441 hctx->sched_tags : hctx->tags;
3442 struct rq_iter_data data = {
3446 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3450 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3451 struct blk_mq_hw_ctx *hctx)
3453 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3455 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3460 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3462 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3463 struct blk_mq_hw_ctx, cpuhp_online);
3465 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3466 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3470 * Prevent new request from being allocated on the current hctx.
3472 * The smp_mb__after_atomic() Pairs with the implied barrier in
3473 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3474 * seen once we return from the tag allocator.
3476 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3477 smp_mb__after_atomic();
3480 * Try to grab a reference to the queue and wait for any outstanding
3481 * requests. If we could not grab a reference the queue has been
3482 * frozen and there are no requests.
3484 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3485 while (blk_mq_hctx_has_requests(hctx))
3487 percpu_ref_put(&hctx->queue->q_usage_counter);
3493 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3495 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3496 struct blk_mq_hw_ctx, cpuhp_online);
3498 if (cpumask_test_cpu(cpu, hctx->cpumask))
3499 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3504 * 'cpu' is going away. splice any existing rq_list entries from this
3505 * software queue to the hw queue dispatch list, and ensure that it
3508 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3510 struct blk_mq_hw_ctx *hctx;
3511 struct blk_mq_ctx *ctx;
3513 enum hctx_type type;
3515 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3516 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3519 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3522 spin_lock(&ctx->lock);
3523 if (!list_empty(&ctx->rq_lists[type])) {
3524 list_splice_init(&ctx->rq_lists[type], &tmp);
3525 blk_mq_hctx_clear_pending(hctx, ctx);
3527 spin_unlock(&ctx->lock);
3529 if (list_empty(&tmp))
3532 spin_lock(&hctx->lock);
3533 list_splice_tail_init(&tmp, &hctx->dispatch);
3534 spin_unlock(&hctx->lock);
3536 blk_mq_run_hw_queue(hctx, true);
3540 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3542 if (!(hctx->flags & BLK_MQ_F_STACKING))
3543 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3544 &hctx->cpuhp_online);
3545 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3550 * Before freeing hw queue, clearing the flush request reference in
3551 * tags->rqs[] for avoiding potential UAF.
3553 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3554 unsigned int queue_depth, struct request *flush_rq)
3557 unsigned long flags;
3559 /* The hw queue may not be mapped yet */
3563 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3565 for (i = 0; i < queue_depth; i++)
3566 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3569 * Wait until all pending iteration is done.
3571 * Request reference is cleared and it is guaranteed to be observed
3572 * after the ->lock is released.
3574 spin_lock_irqsave(&tags->lock, flags);
3575 spin_unlock_irqrestore(&tags->lock, flags);
3578 /* hctx->ctxs will be freed in queue's release handler */
3579 static void blk_mq_exit_hctx(struct request_queue *q,
3580 struct blk_mq_tag_set *set,
3581 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3583 struct request *flush_rq = hctx->fq->flush_rq;
3585 if (blk_mq_hw_queue_mapped(hctx))
3586 blk_mq_tag_idle(hctx);
3588 if (blk_queue_init_done(q))
3589 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3590 set->queue_depth, flush_rq);
3591 if (set->ops->exit_request)
3592 set->ops->exit_request(set, flush_rq, hctx_idx);
3594 if (set->ops->exit_hctx)
3595 set->ops->exit_hctx(hctx, hctx_idx);
3597 blk_mq_remove_cpuhp(hctx);
3599 xa_erase(&q->hctx_table, hctx_idx);
3601 spin_lock(&q->unused_hctx_lock);
3602 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3603 spin_unlock(&q->unused_hctx_lock);
3606 static void blk_mq_exit_hw_queues(struct request_queue *q,
3607 struct blk_mq_tag_set *set, int nr_queue)
3609 struct blk_mq_hw_ctx *hctx;
3612 queue_for_each_hw_ctx(q, hctx, i) {
3615 blk_mq_exit_hctx(q, set, hctx, i);
3619 static int blk_mq_init_hctx(struct request_queue *q,
3620 struct blk_mq_tag_set *set,
3621 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3623 hctx->queue_num = hctx_idx;
3625 if (!(hctx->flags & BLK_MQ_F_STACKING))
3626 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3627 &hctx->cpuhp_online);
3628 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3630 hctx->tags = set->tags[hctx_idx];
3632 if (set->ops->init_hctx &&
3633 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3634 goto unregister_cpu_notifier;
3636 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3640 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3646 if (set->ops->exit_request)
3647 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3649 if (set->ops->exit_hctx)
3650 set->ops->exit_hctx(hctx, hctx_idx);
3651 unregister_cpu_notifier:
3652 blk_mq_remove_cpuhp(hctx);
3656 static struct blk_mq_hw_ctx *
3657 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3660 struct blk_mq_hw_ctx *hctx;
3661 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3663 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3665 goto fail_alloc_hctx;
3667 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3670 atomic_set(&hctx->nr_active, 0);
3671 if (node == NUMA_NO_NODE)
3672 node = set->numa_node;
3673 hctx->numa_node = node;
3675 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3676 spin_lock_init(&hctx->lock);
3677 INIT_LIST_HEAD(&hctx->dispatch);
3679 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3681 INIT_LIST_HEAD(&hctx->hctx_list);
3684 * Allocate space for all possible cpus to avoid allocation at
3687 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3692 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3693 gfp, node, false, false))
3697 spin_lock_init(&hctx->dispatch_wait_lock);
3698 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3699 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3701 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3705 blk_mq_hctx_kobj_init(hctx);
3710 sbitmap_free(&hctx->ctx_map);
3714 free_cpumask_var(hctx->cpumask);
3721 static void blk_mq_init_cpu_queues(struct request_queue *q,
3722 unsigned int nr_hw_queues)
3724 struct blk_mq_tag_set *set = q->tag_set;
3727 for_each_possible_cpu(i) {
3728 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3729 struct blk_mq_hw_ctx *hctx;
3733 spin_lock_init(&__ctx->lock);
3734 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3735 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3740 * Set local node, IFF we have more than one hw queue. If
3741 * not, we remain on the home node of the device
3743 for (j = 0; j < set->nr_maps; j++) {
3744 hctx = blk_mq_map_queue_type(q, j, i);
3745 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3746 hctx->numa_node = cpu_to_node(i);
3751 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3752 unsigned int hctx_idx,
3755 struct blk_mq_tags *tags;
3758 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3762 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3764 blk_mq_free_rq_map(tags);
3771 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3774 if (blk_mq_is_shared_tags(set->flags)) {
3775 set->tags[hctx_idx] = set->shared_tags;
3780 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3783 return set->tags[hctx_idx];
3786 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3787 struct blk_mq_tags *tags,
3788 unsigned int hctx_idx)
3791 blk_mq_free_rqs(set, tags, hctx_idx);
3792 blk_mq_free_rq_map(tags);
3796 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3797 unsigned int hctx_idx)
3799 if (!blk_mq_is_shared_tags(set->flags))
3800 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3802 set->tags[hctx_idx] = NULL;
3805 static void blk_mq_map_swqueue(struct request_queue *q)
3807 unsigned int j, hctx_idx;
3809 struct blk_mq_hw_ctx *hctx;
3810 struct blk_mq_ctx *ctx;
3811 struct blk_mq_tag_set *set = q->tag_set;
3813 queue_for_each_hw_ctx(q, hctx, i) {
3814 cpumask_clear(hctx->cpumask);
3816 hctx->dispatch_from = NULL;
3820 * Map software to hardware queues.
3822 * If the cpu isn't present, the cpu is mapped to first hctx.
3824 for_each_possible_cpu(i) {
3826 ctx = per_cpu_ptr(q->queue_ctx, i);
3827 for (j = 0; j < set->nr_maps; j++) {
3828 if (!set->map[j].nr_queues) {
3829 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3830 HCTX_TYPE_DEFAULT, i);
3833 hctx_idx = set->map[j].mq_map[i];
3834 /* unmapped hw queue can be remapped after CPU topo changed */
3835 if (!set->tags[hctx_idx] &&
3836 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3838 * If tags initialization fail for some hctx,
3839 * that hctx won't be brought online. In this
3840 * case, remap the current ctx to hctx[0] which
3841 * is guaranteed to always have tags allocated
3843 set->map[j].mq_map[i] = 0;
3846 hctx = blk_mq_map_queue_type(q, j, i);
3847 ctx->hctxs[j] = hctx;
3849 * If the CPU is already set in the mask, then we've
3850 * mapped this one already. This can happen if
3851 * devices share queues across queue maps.
3853 if (cpumask_test_cpu(i, hctx->cpumask))
3856 cpumask_set_cpu(i, hctx->cpumask);
3858 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3859 hctx->ctxs[hctx->nr_ctx++] = ctx;
3862 * If the nr_ctx type overflows, we have exceeded the
3863 * amount of sw queues we can support.
3865 BUG_ON(!hctx->nr_ctx);
3868 for (; j < HCTX_MAX_TYPES; j++)
3869 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3870 HCTX_TYPE_DEFAULT, i);
3873 queue_for_each_hw_ctx(q, hctx, i) {
3875 * If no software queues are mapped to this hardware queue,
3876 * disable it and free the request entries.
3878 if (!hctx->nr_ctx) {
3879 /* Never unmap queue 0. We need it as a
3880 * fallback in case of a new remap fails
3884 __blk_mq_free_map_and_rqs(set, i);
3890 hctx->tags = set->tags[i];
3891 WARN_ON(!hctx->tags);
3894 * Set the map size to the number of mapped software queues.
3895 * This is more accurate and more efficient than looping
3896 * over all possibly mapped software queues.
3898 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3901 * Initialize batch roundrobin counts
3903 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3904 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3909 * Caller needs to ensure that we're either frozen/quiesced, or that
3910 * the queue isn't live yet.
3912 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3914 struct blk_mq_hw_ctx *hctx;
3917 queue_for_each_hw_ctx(q, hctx, i) {
3919 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3921 blk_mq_tag_idle(hctx);
3922 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3927 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3930 struct request_queue *q;
3932 lockdep_assert_held(&set->tag_list_lock);
3934 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3935 blk_mq_freeze_queue(q);
3936 queue_set_hctx_shared(q, shared);
3937 blk_mq_unfreeze_queue(q);
3941 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3943 struct blk_mq_tag_set *set = q->tag_set;
3945 mutex_lock(&set->tag_list_lock);
3946 list_del(&q->tag_set_list);
3947 if (list_is_singular(&set->tag_list)) {
3948 /* just transitioned to unshared */
3949 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3950 /* update existing queue */
3951 blk_mq_update_tag_set_shared(set, false);
3953 mutex_unlock(&set->tag_list_lock);
3954 INIT_LIST_HEAD(&q->tag_set_list);
3957 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3958 struct request_queue *q)
3960 mutex_lock(&set->tag_list_lock);
3963 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3965 if (!list_empty(&set->tag_list) &&
3966 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3967 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3968 /* update existing queue */
3969 blk_mq_update_tag_set_shared(set, true);
3971 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3972 queue_set_hctx_shared(q, true);
3973 list_add_tail(&q->tag_set_list, &set->tag_list);
3975 mutex_unlock(&set->tag_list_lock);
3978 /* All allocations will be freed in release handler of q->mq_kobj */
3979 static int blk_mq_alloc_ctxs(struct request_queue *q)
3981 struct blk_mq_ctxs *ctxs;
3984 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3988 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3989 if (!ctxs->queue_ctx)
3992 for_each_possible_cpu(cpu) {
3993 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3997 q->mq_kobj = &ctxs->kobj;
3998 q->queue_ctx = ctxs->queue_ctx;
4007 * It is the actual release handler for mq, but we do it from
4008 * request queue's release handler for avoiding use-after-free
4009 * and headache because q->mq_kobj shouldn't have been introduced,
4010 * but we can't group ctx/kctx kobj without it.
4012 void blk_mq_release(struct request_queue *q)
4014 struct blk_mq_hw_ctx *hctx, *next;
4017 queue_for_each_hw_ctx(q, hctx, i)
4018 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4020 /* all hctx are in .unused_hctx_list now */
4021 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4022 list_del_init(&hctx->hctx_list);
4023 kobject_put(&hctx->kobj);
4026 xa_destroy(&q->hctx_table);
4029 * release .mq_kobj and sw queue's kobject now because
4030 * both share lifetime with request queue.
4032 blk_mq_sysfs_deinit(q);
4035 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4038 struct request_queue *q;
4041 q = blk_alloc_queue(set->numa_node);
4043 return ERR_PTR(-ENOMEM);
4044 q->queuedata = queuedata;
4045 ret = blk_mq_init_allocated_queue(set, q);
4048 return ERR_PTR(ret);
4053 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4055 return blk_mq_init_queue_data(set, NULL);
4057 EXPORT_SYMBOL(blk_mq_init_queue);
4060 * blk_mq_destroy_queue - shutdown a request queue
4061 * @q: request queue to shutdown
4063 * This shuts down a request queue allocated by blk_mq_init_queue(). All future
4064 * requests will be failed with -ENODEV. The caller is responsible for dropping
4065 * the reference from blk_mq_init_queue() by calling blk_put_queue().
4067 * Context: can sleep
4069 void blk_mq_destroy_queue(struct request_queue *q)
4071 WARN_ON_ONCE(!queue_is_mq(q));
4072 WARN_ON_ONCE(blk_queue_registered(q));
4076 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4077 blk_queue_start_drain(q);
4078 blk_mq_freeze_queue_wait(q);
4081 blk_mq_cancel_work_sync(q);
4082 blk_mq_exit_queue(q);
4084 EXPORT_SYMBOL(blk_mq_destroy_queue);
4086 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4087 struct lock_class_key *lkclass)
4089 struct request_queue *q;
4090 struct gendisk *disk;
4092 q = blk_mq_init_queue_data(set, queuedata);
4096 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4098 blk_mq_destroy_queue(q);
4100 return ERR_PTR(-ENOMEM);
4102 set_bit(GD_OWNS_QUEUE, &disk->state);
4105 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4107 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4108 struct lock_class_key *lkclass)
4110 struct gendisk *disk;
4112 if (!blk_get_queue(q))
4114 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4119 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4121 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4122 struct blk_mq_tag_set *set, struct request_queue *q,
4123 int hctx_idx, int node)
4125 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4127 /* reuse dead hctx first */
4128 spin_lock(&q->unused_hctx_lock);
4129 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4130 if (tmp->numa_node == node) {
4136 list_del_init(&hctx->hctx_list);
4137 spin_unlock(&q->unused_hctx_lock);
4140 hctx = blk_mq_alloc_hctx(q, set, node);
4144 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4150 kobject_put(&hctx->kobj);
4155 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4156 struct request_queue *q)
4158 struct blk_mq_hw_ctx *hctx;
4161 /* protect against switching io scheduler */
4162 mutex_lock(&q->sysfs_lock);
4163 for (i = 0; i < set->nr_hw_queues; i++) {
4165 int node = blk_mq_get_hctx_node(set, i);
4166 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4169 old_node = old_hctx->numa_node;
4170 blk_mq_exit_hctx(q, set, old_hctx, i);
4173 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4176 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4178 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4179 WARN_ON_ONCE(!hctx);
4183 * Increasing nr_hw_queues fails. Free the newly allocated
4184 * hctxs and keep the previous q->nr_hw_queues.
4186 if (i != set->nr_hw_queues) {
4187 j = q->nr_hw_queues;
4190 q->nr_hw_queues = set->nr_hw_queues;
4193 xa_for_each_start(&q->hctx_table, j, hctx, j)
4194 blk_mq_exit_hctx(q, set, hctx, j);
4195 mutex_unlock(&q->sysfs_lock);
4198 static void blk_mq_update_poll_flag(struct request_queue *q)
4200 struct blk_mq_tag_set *set = q->tag_set;
4202 if (set->nr_maps > HCTX_TYPE_POLL &&
4203 set->map[HCTX_TYPE_POLL].nr_queues)
4204 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4206 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4209 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4210 struct request_queue *q)
4212 /* mark the queue as mq asap */
4213 q->mq_ops = set->ops;
4215 if (blk_mq_alloc_ctxs(q))
4218 /* init q->mq_kobj and sw queues' kobjects */
4219 blk_mq_sysfs_init(q);
4221 INIT_LIST_HEAD(&q->unused_hctx_list);
4222 spin_lock_init(&q->unused_hctx_lock);
4224 xa_init(&q->hctx_table);
4226 blk_mq_realloc_hw_ctxs(set, q);
4227 if (!q->nr_hw_queues)
4230 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4231 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4235 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4236 blk_mq_update_poll_flag(q);
4238 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4239 INIT_LIST_HEAD(&q->flush_list);
4240 INIT_LIST_HEAD(&q->requeue_list);
4241 spin_lock_init(&q->requeue_lock);
4243 q->nr_requests = set->queue_depth;
4245 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4246 blk_mq_add_queue_tag_set(set, q);
4247 blk_mq_map_swqueue(q);
4256 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4258 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4259 void blk_mq_exit_queue(struct request_queue *q)
4261 struct blk_mq_tag_set *set = q->tag_set;
4263 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4264 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4265 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4266 blk_mq_del_queue_tag_set(q);
4269 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4273 if (blk_mq_is_shared_tags(set->flags)) {
4274 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4277 if (!set->shared_tags)
4281 for (i = 0; i < set->nr_hw_queues; i++) {
4282 if (!__blk_mq_alloc_map_and_rqs(set, i))
4291 __blk_mq_free_map_and_rqs(set, i);
4293 if (blk_mq_is_shared_tags(set->flags)) {
4294 blk_mq_free_map_and_rqs(set, set->shared_tags,
4295 BLK_MQ_NO_HCTX_IDX);
4302 * Allocate the request maps associated with this tag_set. Note that this
4303 * may reduce the depth asked for, if memory is tight. set->queue_depth
4304 * will be updated to reflect the allocated depth.
4306 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4311 depth = set->queue_depth;
4313 err = __blk_mq_alloc_rq_maps(set);
4317 set->queue_depth >>= 1;
4318 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4322 } while (set->queue_depth);
4324 if (!set->queue_depth || err) {
4325 pr_err("blk-mq: failed to allocate request map\n");
4329 if (depth != set->queue_depth)
4330 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4331 depth, set->queue_depth);
4336 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4339 * blk_mq_map_queues() and multiple .map_queues() implementations
4340 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4341 * number of hardware queues.
4343 if (set->nr_maps == 1)
4344 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4346 if (set->ops->map_queues && !is_kdump_kernel()) {
4350 * transport .map_queues is usually done in the following
4353 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4354 * mask = get_cpu_mask(queue)
4355 * for_each_cpu(cpu, mask)
4356 * set->map[x].mq_map[cpu] = queue;
4359 * When we need to remap, the table has to be cleared for
4360 * killing stale mapping since one CPU may not be mapped
4363 for (i = 0; i < set->nr_maps; i++)
4364 blk_mq_clear_mq_map(&set->map[i]);
4366 set->ops->map_queues(set);
4368 BUG_ON(set->nr_maps > 1);
4369 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4373 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4374 int new_nr_hw_queues)
4376 struct blk_mq_tags **new_tags;
4378 if (set->nr_hw_queues >= new_nr_hw_queues)
4381 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4382 GFP_KERNEL, set->numa_node);
4387 memcpy(new_tags, set->tags, set->nr_hw_queues *
4388 sizeof(*set->tags));
4390 set->tags = new_tags;
4392 set->nr_hw_queues = new_nr_hw_queues;
4397 * Alloc a tag set to be associated with one or more request queues.
4398 * May fail with EINVAL for various error conditions. May adjust the
4399 * requested depth down, if it's too large. In that case, the set
4400 * value will be stored in set->queue_depth.
4402 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4406 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4408 if (!set->nr_hw_queues)
4410 if (!set->queue_depth)
4412 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4415 if (!set->ops->queue_rq)
4418 if (!set->ops->get_budget ^ !set->ops->put_budget)
4421 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4422 pr_info("blk-mq: reduced tag depth to %u\n",
4424 set->queue_depth = BLK_MQ_MAX_DEPTH;
4429 else if (set->nr_maps > HCTX_MAX_TYPES)
4433 * If a crashdump is active, then we are potentially in a very
4434 * memory constrained environment. Limit us to 1 queue and
4435 * 64 tags to prevent using too much memory.
4437 if (is_kdump_kernel()) {
4438 set->nr_hw_queues = 1;
4440 set->queue_depth = min(64U, set->queue_depth);
4443 * There is no use for more h/w queues than cpus if we just have
4446 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4447 set->nr_hw_queues = nr_cpu_ids;
4449 if (set->flags & BLK_MQ_F_BLOCKING) {
4450 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4453 ret = init_srcu_struct(set->srcu);
4459 set->tags = kcalloc_node(set->nr_hw_queues,
4460 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4463 goto out_cleanup_srcu;
4465 for (i = 0; i < set->nr_maps; i++) {
4466 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4467 sizeof(set->map[i].mq_map[0]),
4468 GFP_KERNEL, set->numa_node);
4469 if (!set->map[i].mq_map)
4470 goto out_free_mq_map;
4471 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4474 blk_mq_update_queue_map(set);
4476 ret = blk_mq_alloc_set_map_and_rqs(set);
4478 goto out_free_mq_map;
4480 mutex_init(&set->tag_list_lock);
4481 INIT_LIST_HEAD(&set->tag_list);
4486 for (i = 0; i < set->nr_maps; i++) {
4487 kfree(set->map[i].mq_map);
4488 set->map[i].mq_map = NULL;
4493 if (set->flags & BLK_MQ_F_BLOCKING)
4494 cleanup_srcu_struct(set->srcu);
4496 if (set->flags & BLK_MQ_F_BLOCKING)
4500 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4502 /* allocate and initialize a tagset for a simple single-queue device */
4503 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4504 const struct blk_mq_ops *ops, unsigned int queue_depth,
4505 unsigned int set_flags)
4507 memset(set, 0, sizeof(*set));
4509 set->nr_hw_queues = 1;
4511 set->queue_depth = queue_depth;
4512 set->numa_node = NUMA_NO_NODE;
4513 set->flags = set_flags;
4514 return blk_mq_alloc_tag_set(set);
4516 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4518 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4522 for (i = 0; i < set->nr_hw_queues; i++)
4523 __blk_mq_free_map_and_rqs(set, i);
4525 if (blk_mq_is_shared_tags(set->flags)) {
4526 blk_mq_free_map_and_rqs(set, set->shared_tags,
4527 BLK_MQ_NO_HCTX_IDX);
4530 for (j = 0; j < set->nr_maps; j++) {
4531 kfree(set->map[j].mq_map);
4532 set->map[j].mq_map = NULL;
4537 if (set->flags & BLK_MQ_F_BLOCKING) {
4538 cleanup_srcu_struct(set->srcu);
4542 EXPORT_SYMBOL(blk_mq_free_tag_set);
4544 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4546 struct blk_mq_tag_set *set = q->tag_set;
4547 struct blk_mq_hw_ctx *hctx;
4554 if (q->nr_requests == nr)
4557 blk_mq_freeze_queue(q);
4558 blk_mq_quiesce_queue(q);
4561 queue_for_each_hw_ctx(q, hctx, i) {
4565 * If we're using an MQ scheduler, just update the scheduler
4566 * queue depth. This is similar to what the old code would do.
4568 if (hctx->sched_tags) {
4569 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4572 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4577 if (q->elevator && q->elevator->type->ops.depth_updated)
4578 q->elevator->type->ops.depth_updated(hctx);
4581 q->nr_requests = nr;
4582 if (blk_mq_is_shared_tags(set->flags)) {
4584 blk_mq_tag_update_sched_shared_tags(q);
4586 blk_mq_tag_resize_shared_tags(set, nr);
4590 blk_mq_unquiesce_queue(q);
4591 blk_mq_unfreeze_queue(q);
4597 * request_queue and elevator_type pair.
4598 * It is just used by __blk_mq_update_nr_hw_queues to cache
4599 * the elevator_type associated with a request_queue.
4601 struct blk_mq_qe_pair {
4602 struct list_head node;
4603 struct request_queue *q;
4604 struct elevator_type *type;
4608 * Cache the elevator_type in qe pair list and switch the
4609 * io scheduler to 'none'
4611 static bool blk_mq_elv_switch_none(struct list_head *head,
4612 struct request_queue *q)
4614 struct blk_mq_qe_pair *qe;
4616 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4620 /* q->elevator needs protection from ->sysfs_lock */
4621 mutex_lock(&q->sysfs_lock);
4623 /* the check has to be done with holding sysfs_lock */
4629 INIT_LIST_HEAD(&qe->node);
4631 qe->type = q->elevator->type;
4632 /* keep a reference to the elevator module as we'll switch back */
4633 __elevator_get(qe->type);
4634 list_add(&qe->node, head);
4635 elevator_disable(q);
4637 mutex_unlock(&q->sysfs_lock);
4642 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4643 struct request_queue *q)
4645 struct blk_mq_qe_pair *qe;
4647 list_for_each_entry(qe, head, node)
4654 static void blk_mq_elv_switch_back(struct list_head *head,
4655 struct request_queue *q)
4657 struct blk_mq_qe_pair *qe;
4658 struct elevator_type *t;
4660 qe = blk_lookup_qe_pair(head, q);
4664 list_del(&qe->node);
4667 mutex_lock(&q->sysfs_lock);
4668 elevator_switch(q, t);
4669 /* drop the reference acquired in blk_mq_elv_switch_none */
4671 mutex_unlock(&q->sysfs_lock);
4674 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4677 struct request_queue *q;
4679 int prev_nr_hw_queues;
4681 lockdep_assert_held(&set->tag_list_lock);
4683 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4684 nr_hw_queues = nr_cpu_ids;
4685 if (nr_hw_queues < 1)
4687 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4690 list_for_each_entry(q, &set->tag_list, tag_set_list)
4691 blk_mq_freeze_queue(q);
4693 * Switch IO scheduler to 'none', cleaning up the data associated
4694 * with the previous scheduler. We will switch back once we are done
4695 * updating the new sw to hw queue mappings.
4697 list_for_each_entry(q, &set->tag_list, tag_set_list)
4698 if (!blk_mq_elv_switch_none(&head, q))
4701 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4702 blk_mq_debugfs_unregister_hctxs(q);
4703 blk_mq_sysfs_unregister_hctxs(q);
4706 prev_nr_hw_queues = set->nr_hw_queues;
4707 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4711 blk_mq_update_queue_map(set);
4712 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4713 blk_mq_realloc_hw_ctxs(set, q);
4714 blk_mq_update_poll_flag(q);
4715 if (q->nr_hw_queues != set->nr_hw_queues) {
4716 int i = prev_nr_hw_queues;
4718 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4719 nr_hw_queues, prev_nr_hw_queues);
4720 for (; i < set->nr_hw_queues; i++)
4721 __blk_mq_free_map_and_rqs(set, i);
4723 set->nr_hw_queues = prev_nr_hw_queues;
4724 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4727 blk_mq_map_swqueue(q);
4731 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4732 blk_mq_sysfs_register_hctxs(q);
4733 blk_mq_debugfs_register_hctxs(q);
4737 list_for_each_entry(q, &set->tag_list, tag_set_list)
4738 blk_mq_elv_switch_back(&head, q);
4740 list_for_each_entry(q, &set->tag_list, tag_set_list)
4741 blk_mq_unfreeze_queue(q);
4744 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4746 mutex_lock(&set->tag_list_lock);
4747 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4748 mutex_unlock(&set->tag_list_lock);
4750 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4752 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4755 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4756 long state = get_current_state();
4760 ret = q->mq_ops->poll(hctx, iob);
4762 __set_current_state(TASK_RUNNING);
4766 if (signal_pending_state(state, current))
4767 __set_current_state(TASK_RUNNING);
4768 if (task_is_running(current))
4771 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4774 } while (!need_resched());
4776 __set_current_state(TASK_RUNNING);
4780 unsigned int blk_mq_rq_cpu(struct request *rq)
4782 return rq->mq_ctx->cpu;
4784 EXPORT_SYMBOL(blk_mq_rq_cpu);
4786 void blk_mq_cancel_work_sync(struct request_queue *q)
4788 struct blk_mq_hw_ctx *hctx;
4791 cancel_delayed_work_sync(&q->requeue_work);
4793 queue_for_each_hw_ctx(q, hctx, i)
4794 cancel_delayed_work_sync(&hctx->run_work);
4797 static int __init blk_mq_init(void)
4801 for_each_possible_cpu(i)
4802 init_llist_head(&per_cpu(blk_cpu_done, i));
4803 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4805 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4806 "block/softirq:dead", NULL,
4807 blk_softirq_cpu_dead);
4808 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4809 blk_mq_hctx_notify_dead);
4810 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4811 blk_mq_hctx_notify_online,
4812 blk_mq_hctx_notify_offline);
4815 subsys_initcall(blk_mq_init);