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_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
49 struct list_head *list);
51 static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q,
54 return xa_load(&q->hctx_table, qc);
57 static inline blk_qc_t blk_rq_to_qc(struct request *rq)
59 return rq->mq_hctx->queue_num;
63 * Check if any of the ctx, dispatch list or elevator
64 * have pending work in this hardware queue.
66 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
68 return !list_empty_careful(&hctx->dispatch) ||
69 sbitmap_any_bit_set(&hctx->ctx_map) ||
70 blk_mq_sched_has_work(hctx);
74 * Mark this ctx as having pending work in this hardware queue
76 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
77 struct blk_mq_ctx *ctx)
79 const int bit = ctx->index_hw[hctx->type];
81 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
82 sbitmap_set_bit(&hctx->ctx_map, bit);
85 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
86 struct blk_mq_ctx *ctx)
88 const int bit = ctx->index_hw[hctx->type];
90 sbitmap_clear_bit(&hctx->ctx_map, bit);
94 struct block_device *part;
95 unsigned int inflight[2];
98 static bool blk_mq_check_inflight(struct request *rq, void *priv)
100 struct mq_inflight *mi = priv;
102 if (rq->part && blk_do_io_stat(rq) &&
103 (!mi->part->bd_partno || rq->part == mi->part) &&
104 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
105 mi->inflight[rq_data_dir(rq)]++;
110 unsigned int blk_mq_in_flight(struct request_queue *q,
111 struct block_device *part)
113 struct mq_inflight mi = { .part = part };
115 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
117 return mi.inflight[0] + mi.inflight[1];
120 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
121 unsigned int inflight[2])
123 struct mq_inflight mi = { .part = part };
125 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
126 inflight[0] = mi.inflight[0];
127 inflight[1] = mi.inflight[1];
130 void blk_freeze_queue_start(struct request_queue *q)
132 mutex_lock(&q->mq_freeze_lock);
133 if (++q->mq_freeze_depth == 1) {
134 percpu_ref_kill(&q->q_usage_counter);
135 mutex_unlock(&q->mq_freeze_lock);
137 blk_mq_run_hw_queues(q, false);
139 mutex_unlock(&q->mq_freeze_lock);
142 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
144 void blk_mq_freeze_queue_wait(struct request_queue *q)
146 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
148 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
150 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
151 unsigned long timeout)
153 return wait_event_timeout(q->mq_freeze_wq,
154 percpu_ref_is_zero(&q->q_usage_counter),
157 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
160 * Guarantee no request is in use, so we can change any data structure of
161 * the queue afterward.
163 void blk_freeze_queue(struct request_queue *q)
166 * In the !blk_mq case we are only calling this to kill the
167 * q_usage_counter, otherwise this increases the freeze depth
168 * and waits for it to return to zero. For this reason there is
169 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
170 * exported to drivers as the only user for unfreeze is blk_mq.
172 blk_freeze_queue_start(q);
173 blk_mq_freeze_queue_wait(q);
176 void blk_mq_freeze_queue(struct request_queue *q)
179 * ...just an alias to keep freeze and unfreeze actions balanced
180 * in the blk_mq_* namespace
184 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
186 void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
188 mutex_lock(&q->mq_freeze_lock);
190 q->q_usage_counter.data->force_atomic = true;
191 q->mq_freeze_depth--;
192 WARN_ON_ONCE(q->mq_freeze_depth < 0);
193 if (!q->mq_freeze_depth) {
194 percpu_ref_resurrect(&q->q_usage_counter);
195 wake_up_all(&q->mq_freeze_wq);
197 mutex_unlock(&q->mq_freeze_lock);
200 void blk_mq_unfreeze_queue(struct request_queue *q)
202 __blk_mq_unfreeze_queue(q, false);
204 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
207 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
208 * mpt3sas driver such that this function can be removed.
210 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
214 spin_lock_irqsave(&q->queue_lock, flags);
215 if (!q->quiesce_depth++)
216 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
217 spin_unlock_irqrestore(&q->queue_lock, flags);
219 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
222 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
223 * @set: tag_set to wait on
225 * Note: it is driver's responsibility for making sure that quiesce has
226 * been started on or more of the request_queues of the tag_set. This
227 * function only waits for the quiesce on those request_queues that had
228 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
230 void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
232 if (set->flags & BLK_MQ_F_BLOCKING)
233 synchronize_srcu(set->srcu);
237 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
240 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
243 * Note: this function does not prevent that the struct request end_io()
244 * callback function is invoked. Once this function is returned, we make
245 * sure no dispatch can happen until the queue is unquiesced via
246 * blk_mq_unquiesce_queue().
248 void blk_mq_quiesce_queue(struct request_queue *q)
250 blk_mq_quiesce_queue_nowait(q);
251 /* nothing to wait for non-mq queues */
253 blk_mq_wait_quiesce_done(q->tag_set);
255 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
258 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
261 * This function recovers queue into the state before quiescing
262 * which is done by blk_mq_quiesce_queue.
264 void blk_mq_unquiesce_queue(struct request_queue *q)
267 bool run_queue = false;
269 spin_lock_irqsave(&q->queue_lock, flags);
270 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
272 } else if (!--q->quiesce_depth) {
273 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
276 spin_unlock_irqrestore(&q->queue_lock, flags);
278 /* dispatch requests which are inserted during quiescing */
280 blk_mq_run_hw_queues(q, true);
282 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
284 void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
286 struct request_queue *q;
288 mutex_lock(&set->tag_list_lock);
289 list_for_each_entry(q, &set->tag_list, tag_set_list) {
290 if (!blk_queue_skip_tagset_quiesce(q))
291 blk_mq_quiesce_queue_nowait(q);
293 blk_mq_wait_quiesce_done(set);
294 mutex_unlock(&set->tag_list_lock);
296 EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
298 void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
300 struct request_queue *q;
302 mutex_lock(&set->tag_list_lock);
303 list_for_each_entry(q, &set->tag_list, tag_set_list) {
304 if (!blk_queue_skip_tagset_quiesce(q))
305 blk_mq_unquiesce_queue(q);
307 mutex_unlock(&set->tag_list_lock);
309 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
311 void blk_mq_wake_waiters(struct request_queue *q)
313 struct blk_mq_hw_ctx *hctx;
316 queue_for_each_hw_ctx(q, hctx, i)
317 if (blk_mq_hw_queue_mapped(hctx))
318 blk_mq_tag_wakeup_all(hctx->tags, true);
321 void blk_rq_init(struct request_queue *q, struct request *rq)
323 memset(rq, 0, sizeof(*rq));
325 INIT_LIST_HEAD(&rq->queuelist);
327 rq->__sector = (sector_t) -1;
328 INIT_HLIST_NODE(&rq->hash);
329 RB_CLEAR_NODE(&rq->rb_node);
330 rq->tag = BLK_MQ_NO_TAG;
331 rq->internal_tag = BLK_MQ_NO_TAG;
332 rq->start_time_ns = ktime_get_ns();
334 blk_crypto_rq_set_defaults(rq);
336 EXPORT_SYMBOL(blk_rq_init);
338 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
339 struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns)
341 struct blk_mq_ctx *ctx = data->ctx;
342 struct blk_mq_hw_ctx *hctx = data->hctx;
343 struct request_queue *q = data->q;
344 struct request *rq = tags->static_rqs[tag];
349 rq->cmd_flags = data->cmd_flags;
351 if (data->flags & BLK_MQ_REQ_PM)
352 data->rq_flags |= RQF_PM;
353 if (blk_queue_io_stat(q))
354 data->rq_flags |= RQF_IO_STAT;
355 rq->rq_flags = data->rq_flags;
357 if (!(data->rq_flags & RQF_ELV)) {
359 rq->internal_tag = BLK_MQ_NO_TAG;
361 rq->tag = BLK_MQ_NO_TAG;
362 rq->internal_tag = tag;
366 if (blk_mq_need_time_stamp(rq))
367 rq->start_time_ns = ktime_get_ns();
369 rq->start_time_ns = 0;
371 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
372 rq->alloc_time_ns = alloc_time_ns;
374 rq->io_start_time_ns = 0;
375 rq->stats_sectors = 0;
376 rq->nr_phys_segments = 0;
377 #if defined(CONFIG_BLK_DEV_INTEGRITY)
378 rq->nr_integrity_segments = 0;
381 rq->end_io_data = NULL;
383 blk_crypto_rq_set_defaults(rq);
384 INIT_LIST_HEAD(&rq->queuelist);
385 /* tag was already set */
386 WRITE_ONCE(rq->deadline, 0);
389 if (rq->rq_flags & RQF_ELV) {
390 struct elevator_queue *e = data->q->elevator;
392 INIT_HLIST_NODE(&rq->hash);
393 RB_CLEAR_NODE(&rq->rb_node);
395 if (!op_is_flush(data->cmd_flags) &&
396 e->type->ops.prepare_request) {
397 e->type->ops.prepare_request(rq);
398 rq->rq_flags |= RQF_ELVPRIV;
405 static inline struct request *
406 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data,
409 unsigned int tag, tag_offset;
410 struct blk_mq_tags *tags;
412 unsigned long tag_mask;
415 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
416 if (unlikely(!tag_mask))
419 tags = blk_mq_tags_from_data(data);
420 for (i = 0; tag_mask; i++) {
421 if (!(tag_mask & (1UL << i)))
423 tag = tag_offset + i;
424 prefetch(tags->static_rqs[tag]);
425 tag_mask &= ~(1UL << i);
426 rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns);
427 rq_list_add(data->cached_rq, rq);
430 /* caller already holds a reference, add for remainder */
431 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
434 return rq_list_pop(data->cached_rq);
437 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
439 struct request_queue *q = data->q;
440 u64 alloc_time_ns = 0;
444 /* alloc_time includes depth and tag waits */
445 if (blk_queue_rq_alloc_time(q))
446 alloc_time_ns = ktime_get_ns();
448 if (data->cmd_flags & REQ_NOWAIT)
449 data->flags |= BLK_MQ_REQ_NOWAIT;
452 struct elevator_queue *e = q->elevator;
454 data->rq_flags |= RQF_ELV;
457 * Flush/passthrough requests are special and go directly to the
458 * dispatch list. Don't include reserved tags in the
459 * limiting, as it isn't useful.
461 if (!op_is_flush(data->cmd_flags) &&
462 !blk_op_is_passthrough(data->cmd_flags) &&
463 e->type->ops.limit_depth &&
464 !(data->flags & BLK_MQ_REQ_RESERVED))
465 e->type->ops.limit_depth(data->cmd_flags, data);
469 data->ctx = blk_mq_get_ctx(q);
470 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
471 if (!(data->rq_flags & RQF_ELV))
472 blk_mq_tag_busy(data->hctx);
474 if (data->flags & BLK_MQ_REQ_RESERVED)
475 data->rq_flags |= RQF_RESV;
478 * Try batched alloc if we want more than 1 tag.
480 if (data->nr_tags > 1) {
481 rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns);
488 * Waiting allocations only fail because of an inactive hctx. In that
489 * case just retry the hctx assignment and tag allocation as CPU hotplug
490 * should have migrated us to an online CPU by now.
492 tag = blk_mq_get_tag(data);
493 if (tag == BLK_MQ_NO_TAG) {
494 if (data->flags & BLK_MQ_REQ_NOWAIT)
497 * Give up the CPU and sleep for a random short time to
498 * ensure that thread using a realtime scheduling class
499 * are migrated off the CPU, and thus off the hctx that
506 return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag,
510 static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
511 struct blk_plug *plug,
513 blk_mq_req_flags_t flags)
515 struct blk_mq_alloc_data data = {
519 .nr_tags = plug->nr_ios,
520 .cached_rq = &plug->cached_rq,
524 if (blk_queue_enter(q, flags))
529 rq = __blk_mq_alloc_requests(&data);
535 static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
537 blk_mq_req_flags_t flags)
539 struct blk_plug *plug = current->plug;
545 if (rq_list_empty(plug->cached_rq)) {
546 if (plug->nr_ios == 1)
548 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
552 rq = rq_list_peek(&plug->cached_rq);
553 if (!rq || rq->q != q)
556 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
558 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
561 plug->cached_rq = rq_list_next(rq);
565 INIT_LIST_HEAD(&rq->queuelist);
569 struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
570 blk_mq_req_flags_t flags)
574 rq = blk_mq_alloc_cached_request(q, opf, flags);
576 struct blk_mq_alloc_data data = {
584 ret = blk_queue_enter(q, flags);
588 rq = __blk_mq_alloc_requests(&data);
593 rq->__sector = (sector_t) -1;
594 rq->bio = rq->biotail = NULL;
598 return ERR_PTR(-EWOULDBLOCK);
600 EXPORT_SYMBOL(blk_mq_alloc_request);
602 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
603 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
605 struct blk_mq_alloc_data data = {
611 u64 alloc_time_ns = 0;
617 /* alloc_time includes depth and tag waits */
618 if (blk_queue_rq_alloc_time(q))
619 alloc_time_ns = ktime_get_ns();
622 * If the tag allocator sleeps we could get an allocation for a
623 * different hardware context. No need to complicate the low level
624 * allocator for this for the rare use case of a command tied to
627 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
628 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
629 return ERR_PTR(-EINVAL);
631 if (hctx_idx >= q->nr_hw_queues)
632 return ERR_PTR(-EIO);
634 ret = blk_queue_enter(q, flags);
639 * Check if the hardware context is actually mapped to anything.
640 * If not tell the caller that it should skip this queue.
643 data.hctx = xa_load(&q->hctx_table, hctx_idx);
644 if (!blk_mq_hw_queue_mapped(data.hctx))
646 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
647 if (cpu >= nr_cpu_ids)
649 data.ctx = __blk_mq_get_ctx(q, cpu);
652 blk_mq_tag_busy(data.hctx);
654 data.rq_flags |= RQF_ELV;
656 if (flags & BLK_MQ_REQ_RESERVED)
657 data.rq_flags |= RQF_RESV;
660 tag = blk_mq_get_tag(&data);
661 if (tag == BLK_MQ_NO_TAG)
663 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag,
666 rq->__sector = (sector_t) -1;
667 rq->bio = rq->biotail = NULL;
674 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
676 static void __blk_mq_free_request(struct request *rq)
678 struct request_queue *q = rq->q;
679 struct blk_mq_ctx *ctx = rq->mq_ctx;
680 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
681 const int sched_tag = rq->internal_tag;
683 blk_crypto_free_request(rq);
684 blk_pm_mark_last_busy(rq);
686 if (rq->tag != BLK_MQ_NO_TAG)
687 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
688 if (sched_tag != BLK_MQ_NO_TAG)
689 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
690 blk_mq_sched_restart(hctx);
694 void blk_mq_free_request(struct request *rq)
696 struct request_queue *q = rq->q;
697 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
699 if ((rq->rq_flags & RQF_ELVPRIV) &&
700 q->elevator->type->ops.finish_request)
701 q->elevator->type->ops.finish_request(rq);
703 if (rq->rq_flags & RQF_MQ_INFLIGHT)
704 __blk_mq_dec_active_requests(hctx);
706 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
707 laptop_io_completion(q->disk->bdi);
711 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
712 if (req_ref_put_and_test(rq))
713 __blk_mq_free_request(rq);
715 EXPORT_SYMBOL_GPL(blk_mq_free_request);
717 void blk_mq_free_plug_rqs(struct blk_plug *plug)
721 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
722 blk_mq_free_request(rq);
725 void blk_dump_rq_flags(struct request *rq, char *msg)
727 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
728 rq->q->disk ? rq->q->disk->disk_name : "?",
729 (__force unsigned long long) rq->cmd_flags);
731 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
732 (unsigned long long)blk_rq_pos(rq),
733 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
734 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
735 rq->bio, rq->biotail, blk_rq_bytes(rq));
737 EXPORT_SYMBOL(blk_dump_rq_flags);
739 static void req_bio_endio(struct request *rq, struct bio *bio,
740 unsigned int nbytes, blk_status_t error)
742 if (unlikely(error)) {
743 bio->bi_status = error;
744 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
746 * Partial zone append completions cannot be supported as the
747 * BIO fragments may end up not being written sequentially.
749 if (bio->bi_iter.bi_size != nbytes)
750 bio->bi_status = BLK_STS_IOERR;
752 bio->bi_iter.bi_sector = rq->__sector;
755 bio_advance(bio, nbytes);
757 if (unlikely(rq->rq_flags & RQF_QUIET))
758 bio_set_flag(bio, BIO_QUIET);
759 /* don't actually finish bio if it's part of flush sequence */
760 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
764 static void blk_account_io_completion(struct request *req, unsigned int bytes)
766 if (req->part && blk_do_io_stat(req)) {
767 const int sgrp = op_stat_group(req_op(req));
770 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
775 static void blk_print_req_error(struct request *req, blk_status_t status)
777 printk_ratelimited(KERN_ERR
778 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
779 "phys_seg %u prio class %u\n",
780 blk_status_to_str(status),
781 req->q->disk ? req->q->disk->disk_name : "?",
782 blk_rq_pos(req), (__force u32)req_op(req),
783 blk_op_str(req_op(req)),
784 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
785 req->nr_phys_segments,
786 IOPRIO_PRIO_CLASS(req->ioprio));
790 * Fully end IO on a request. Does not support partial completions, or
793 static void blk_complete_request(struct request *req)
795 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
796 int total_bytes = blk_rq_bytes(req);
797 struct bio *bio = req->bio;
799 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
804 #ifdef CONFIG_BLK_DEV_INTEGRITY
805 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
806 req->q->integrity.profile->complete_fn(req, total_bytes);
810 * Upper layers may call blk_crypto_evict_key() anytime after the last
811 * bio_endio(). Therefore, the keyslot must be released before that.
813 blk_crypto_rq_put_keyslot(req);
815 blk_account_io_completion(req, total_bytes);
818 struct bio *next = bio->bi_next;
820 /* Completion has already been traced */
821 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
823 if (req_op(req) == REQ_OP_ZONE_APPEND)
824 bio->bi_iter.bi_sector = req->__sector;
832 * Reset counters so that the request stacking driver
833 * can find how many bytes remain in the request
843 * blk_update_request - Complete multiple bytes without completing the request
844 * @req: the request being processed
845 * @error: block status code
846 * @nr_bytes: number of bytes to complete for @req
849 * Ends I/O on a number of bytes attached to @req, but doesn't complete
850 * the request structure even if @req doesn't have leftover.
851 * If @req has leftover, sets it up for the next range of segments.
853 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
854 * %false return from this function.
857 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
858 * except in the consistency check at the end of this function.
861 * %false - this request doesn't have any more data
862 * %true - this request has more data
864 bool blk_update_request(struct request *req, blk_status_t error,
865 unsigned int nr_bytes)
869 trace_block_rq_complete(req, error, nr_bytes);
874 #ifdef CONFIG_BLK_DEV_INTEGRITY
875 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
877 req->q->integrity.profile->complete_fn(req, nr_bytes);
881 * Upper layers may call blk_crypto_evict_key() anytime after the last
882 * bio_endio(). Therefore, the keyslot must be released before that.
884 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
885 __blk_crypto_rq_put_keyslot(req);
887 if (unlikely(error && !blk_rq_is_passthrough(req) &&
888 !(req->rq_flags & RQF_QUIET)) &&
889 !test_bit(GD_DEAD, &req->q->disk->state)) {
890 blk_print_req_error(req, error);
891 trace_block_rq_error(req, error, nr_bytes);
894 blk_account_io_completion(req, nr_bytes);
898 struct bio *bio = req->bio;
899 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
901 if (bio_bytes == bio->bi_iter.bi_size)
902 req->bio = bio->bi_next;
904 /* Completion has already been traced */
905 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
906 req_bio_endio(req, bio, bio_bytes, error);
908 total_bytes += bio_bytes;
909 nr_bytes -= bio_bytes;
920 * Reset counters so that the request stacking driver
921 * can find how many bytes remain in the request
928 req->__data_len -= total_bytes;
930 /* update sector only for requests with clear definition of sector */
931 if (!blk_rq_is_passthrough(req))
932 req->__sector += total_bytes >> 9;
934 /* mixed attributes always follow the first bio */
935 if (req->rq_flags & RQF_MIXED_MERGE) {
936 req->cmd_flags &= ~REQ_FAILFAST_MASK;
937 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
940 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
942 * If total number of sectors is less than the first segment
943 * size, something has gone terribly wrong.
945 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
946 blk_dump_rq_flags(req, "request botched");
947 req->__data_len = blk_rq_cur_bytes(req);
950 /* recalculate the number of segments */
951 req->nr_phys_segments = blk_recalc_rq_segments(req);
956 EXPORT_SYMBOL_GPL(blk_update_request);
958 static inline void blk_account_io_done(struct request *req, u64 now)
961 * Account IO completion. flush_rq isn't accounted as a
962 * normal IO on queueing nor completion. Accounting the
963 * containing request is enough.
965 if (blk_do_io_stat(req) && req->part &&
966 !(req->rq_flags & RQF_FLUSH_SEQ)) {
967 const int sgrp = op_stat_group(req_op(req));
970 update_io_ticks(req->part, jiffies, true);
971 part_stat_inc(req->part, ios[sgrp]);
972 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
977 static inline void blk_account_io_start(struct request *req)
979 if (blk_do_io_stat(req)) {
981 * All non-passthrough requests are created from a bio with one
982 * exception: when a flush command that is part of a flush sequence
983 * generated by the state machine in blk-flush.c is cloned onto the
984 * lower device by dm-multipath we can get here without a bio.
987 req->part = req->bio->bi_bdev;
989 req->part = req->q->disk->part0;
992 update_io_ticks(req->part, jiffies, false);
997 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
999 if (rq->rq_flags & RQF_STATS)
1000 blk_stat_add(rq, now);
1002 blk_mq_sched_completed_request(rq, now);
1003 blk_account_io_done(rq, now);
1006 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1008 if (blk_mq_need_time_stamp(rq))
1009 __blk_mq_end_request_acct(rq, ktime_get_ns());
1012 rq_qos_done(rq->q, rq);
1013 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1014 blk_mq_free_request(rq);
1016 blk_mq_free_request(rq);
1019 EXPORT_SYMBOL(__blk_mq_end_request);
1021 void blk_mq_end_request(struct request *rq, blk_status_t error)
1023 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1025 __blk_mq_end_request(rq, error);
1027 EXPORT_SYMBOL(blk_mq_end_request);
1029 #define TAG_COMP_BATCH 32
1031 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1032 int *tag_array, int nr_tags)
1034 struct request_queue *q = hctx->queue;
1037 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1038 * update hctx->nr_active in batch
1040 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1041 __blk_mq_sub_active_requests(hctx, nr_tags);
1043 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1044 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1047 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1049 int tags[TAG_COMP_BATCH], nr_tags = 0;
1050 struct blk_mq_hw_ctx *cur_hctx = NULL;
1055 now = ktime_get_ns();
1057 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1059 prefetch(rq->rq_next);
1061 blk_complete_request(rq);
1063 __blk_mq_end_request_acct(rq, now);
1065 rq_qos_done(rq->q, rq);
1068 * If end_io handler returns NONE, then it still has
1069 * ownership of the request.
1071 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1074 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1075 if (!req_ref_put_and_test(rq))
1078 blk_crypto_free_request(rq);
1079 blk_pm_mark_last_busy(rq);
1081 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1083 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1085 cur_hctx = rq->mq_hctx;
1087 tags[nr_tags++] = rq->tag;
1091 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1093 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1095 static void blk_complete_reqs(struct llist_head *list)
1097 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1098 struct request *rq, *next;
1100 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1101 rq->q->mq_ops->complete(rq);
1104 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1106 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1109 static int blk_softirq_cpu_dead(unsigned int cpu)
1111 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1115 static void __blk_mq_complete_request_remote(void *data)
1117 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1120 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1122 int cpu = raw_smp_processor_id();
1124 if (!IS_ENABLED(CONFIG_SMP) ||
1125 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1128 * With force threaded interrupts enabled, raising softirq from an SMP
1129 * function call will always result in waking the ksoftirqd thread.
1130 * This is probably worse than completing the request on a different
1133 if (force_irqthreads())
1136 /* same CPU or cache domain? Complete locally */
1137 if (cpu == rq->mq_ctx->cpu ||
1138 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1139 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1142 /* don't try to IPI to an offline CPU */
1143 return cpu_online(rq->mq_ctx->cpu);
1146 static void blk_mq_complete_send_ipi(struct request *rq)
1148 struct llist_head *list;
1151 cpu = rq->mq_ctx->cpu;
1152 list = &per_cpu(blk_cpu_done, cpu);
1153 if (llist_add(&rq->ipi_list, list)) {
1154 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1155 smp_call_function_single_async(cpu, &rq->csd);
1159 static void blk_mq_raise_softirq(struct request *rq)
1161 struct llist_head *list;
1164 list = this_cpu_ptr(&blk_cpu_done);
1165 if (llist_add(&rq->ipi_list, list))
1166 raise_softirq(BLOCK_SOFTIRQ);
1170 bool blk_mq_complete_request_remote(struct request *rq)
1172 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1175 * For request which hctx has only one ctx mapping,
1176 * or a polled request, always complete locally,
1177 * it's pointless to redirect the completion.
1179 if (rq->mq_hctx->nr_ctx == 1 ||
1180 rq->cmd_flags & REQ_POLLED)
1183 if (blk_mq_complete_need_ipi(rq)) {
1184 blk_mq_complete_send_ipi(rq);
1188 if (rq->q->nr_hw_queues == 1) {
1189 blk_mq_raise_softirq(rq);
1194 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1197 * blk_mq_complete_request - end I/O on a request
1198 * @rq: the request being processed
1201 * Complete a request by scheduling the ->complete_rq operation.
1203 void blk_mq_complete_request(struct request *rq)
1205 if (!blk_mq_complete_request_remote(rq))
1206 rq->q->mq_ops->complete(rq);
1208 EXPORT_SYMBOL(blk_mq_complete_request);
1211 * blk_mq_start_request - Start processing a request
1212 * @rq: Pointer to request to be started
1214 * Function used by device drivers to notify the block layer that a request
1215 * is going to be processed now, so blk layer can do proper initializations
1216 * such as starting the timeout timer.
1218 void blk_mq_start_request(struct request *rq)
1220 struct request_queue *q = rq->q;
1222 trace_block_rq_issue(rq);
1224 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1225 rq->io_start_time_ns = ktime_get_ns();
1226 rq->stats_sectors = blk_rq_sectors(rq);
1227 rq->rq_flags |= RQF_STATS;
1228 rq_qos_issue(q, rq);
1231 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1234 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1236 #ifdef CONFIG_BLK_DEV_INTEGRITY
1237 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1238 q->integrity.profile->prepare_fn(rq);
1240 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1241 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1243 EXPORT_SYMBOL(blk_mq_start_request);
1246 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1247 * queues. This is important for md arrays to benefit from merging
1250 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1252 if (plug->multiple_queues)
1253 return BLK_MAX_REQUEST_COUNT * 2;
1254 return BLK_MAX_REQUEST_COUNT;
1257 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1259 struct request *last = rq_list_peek(&plug->mq_list);
1261 if (!plug->rq_count) {
1262 trace_block_plug(rq->q);
1263 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1264 (!blk_queue_nomerges(rq->q) &&
1265 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1266 blk_mq_flush_plug_list(plug, false);
1268 trace_block_plug(rq->q);
1271 if (!plug->multiple_queues && last && last->q != rq->q)
1272 plug->multiple_queues = true;
1273 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1274 plug->has_elevator = true;
1276 rq_list_add(&plug->mq_list, rq);
1281 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1282 * @rq: request to insert
1283 * @at_head: insert request at head or tail of queue
1286 * Insert a fully prepared request at the back of the I/O scheduler queue
1287 * for execution. Don't wait for completion.
1290 * This function will invoke @done directly if the queue is dead.
1292 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1294 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1296 WARN_ON(irqs_disabled());
1297 WARN_ON(!blk_rq_is_passthrough(rq));
1299 blk_account_io_start(rq);
1302 * As plugging can be enabled for passthrough requests on a zoned
1303 * device, directly accessing the plug instead of using blk_mq_plug()
1304 * should not have any consequences.
1306 if (current->plug && !at_head) {
1307 blk_add_rq_to_plug(current->plug, rq);
1311 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1312 blk_mq_run_hw_queue(hctx, false);
1314 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1316 struct blk_rq_wait {
1317 struct completion done;
1321 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1323 struct blk_rq_wait *wait = rq->end_io_data;
1326 complete(&wait->done);
1327 return RQ_END_IO_NONE;
1330 bool blk_rq_is_poll(struct request *rq)
1334 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1338 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1340 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1343 blk_mq_poll(rq->q, blk_rq_to_qc(rq), NULL, 0);
1345 } while (!completion_done(wait));
1349 * blk_execute_rq - insert a request into queue for execution
1350 * @rq: request to insert
1351 * @at_head: insert request at head or tail of queue
1354 * Insert a fully prepared request at the back of the I/O scheduler queue
1355 * for execution and wait for completion.
1356 * Return: The blk_status_t result provided to blk_mq_end_request().
1358 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1360 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1361 struct blk_rq_wait wait = {
1362 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1365 WARN_ON(irqs_disabled());
1366 WARN_ON(!blk_rq_is_passthrough(rq));
1368 rq->end_io_data = &wait;
1369 rq->end_io = blk_end_sync_rq;
1371 blk_account_io_start(rq);
1372 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1373 blk_mq_run_hw_queue(hctx, false);
1375 if (blk_rq_is_poll(rq)) {
1376 blk_rq_poll_completion(rq, &wait.done);
1379 * Prevent hang_check timer from firing at us during very long
1382 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1385 while (!wait_for_completion_io_timeout(&wait.done,
1386 hang_check * (HZ/2)))
1389 wait_for_completion_io(&wait.done);
1394 EXPORT_SYMBOL(blk_execute_rq);
1396 static void __blk_mq_requeue_request(struct request *rq)
1398 struct request_queue *q = rq->q;
1400 blk_mq_put_driver_tag(rq);
1402 trace_block_rq_requeue(rq);
1403 rq_qos_requeue(q, rq);
1405 if (blk_mq_request_started(rq)) {
1406 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1407 rq->rq_flags &= ~RQF_TIMED_OUT;
1411 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1413 struct request_queue *q = rq->q;
1415 __blk_mq_requeue_request(rq);
1417 /* this request will be re-inserted to io scheduler queue */
1418 blk_mq_sched_requeue_request(rq);
1420 blk_mq_add_to_requeue_list(rq, BLK_MQ_INSERT_AT_HEAD);
1422 if (kick_requeue_list)
1423 blk_mq_kick_requeue_list(q);
1425 EXPORT_SYMBOL(blk_mq_requeue_request);
1427 static void blk_mq_requeue_work(struct work_struct *work)
1429 struct request_queue *q =
1430 container_of(work, struct request_queue, requeue_work.work);
1432 struct request *rq, *next;
1434 spin_lock_irq(&q->requeue_lock);
1435 list_splice_init(&q->requeue_list, &rq_list);
1436 spin_unlock_irq(&q->requeue_lock);
1438 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1440 * If RQF_DONTPREP ist set, the request has been started by the
1441 * driver already and might have driver-specific data allocated
1442 * already. Insert it into the hctx dispatch list to avoid
1443 * block layer merges for the request.
1445 if (rq->rq_flags & RQF_DONTPREP) {
1446 rq->rq_flags &= ~RQF_SOFTBARRIER;
1447 list_del_init(&rq->queuelist);
1448 blk_mq_request_bypass_insert(rq, 0);
1449 } else if (rq->rq_flags & RQF_SOFTBARRIER) {
1450 rq->rq_flags &= ~RQF_SOFTBARRIER;
1451 list_del_init(&rq->queuelist);
1452 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1456 while (!list_empty(&rq_list)) {
1457 rq = list_entry(rq_list.next, struct request, queuelist);
1458 list_del_init(&rq->queuelist);
1459 blk_mq_insert_request(rq, 0);
1462 blk_mq_run_hw_queues(q, false);
1465 void blk_mq_add_to_requeue_list(struct request *rq, blk_insert_t insert_flags)
1467 struct request_queue *q = rq->q;
1468 unsigned long flags;
1471 * We abuse this flag that is otherwise used by the I/O scheduler to
1472 * request head insertion from the workqueue.
1474 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1476 spin_lock_irqsave(&q->requeue_lock, flags);
1477 if (insert_flags & BLK_MQ_INSERT_AT_HEAD) {
1478 rq->rq_flags |= RQF_SOFTBARRIER;
1479 list_add(&rq->queuelist, &q->requeue_list);
1481 list_add_tail(&rq->queuelist, &q->requeue_list);
1483 spin_unlock_irqrestore(&q->requeue_lock, flags);
1486 void blk_mq_kick_requeue_list(struct request_queue *q)
1488 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1490 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1492 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1493 unsigned long msecs)
1495 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1496 msecs_to_jiffies(msecs));
1498 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1500 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1503 * If we find a request that isn't idle we know the queue is busy
1504 * as it's checked in the iter.
1505 * Return false to stop the iteration.
1507 if (blk_mq_request_started(rq)) {
1517 bool blk_mq_queue_inflight(struct request_queue *q)
1521 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1524 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1526 static void blk_mq_rq_timed_out(struct request *req)
1528 req->rq_flags |= RQF_TIMED_OUT;
1529 if (req->q->mq_ops->timeout) {
1530 enum blk_eh_timer_return ret;
1532 ret = req->q->mq_ops->timeout(req);
1533 if (ret == BLK_EH_DONE)
1535 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1541 struct blk_expired_data {
1542 bool has_timedout_rq;
1544 unsigned long timeout_start;
1547 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1549 unsigned long deadline;
1551 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1553 if (rq->rq_flags & RQF_TIMED_OUT)
1556 deadline = READ_ONCE(rq->deadline);
1557 if (time_after_eq(expired->timeout_start, deadline))
1560 if (expired->next == 0)
1561 expired->next = deadline;
1562 else if (time_after(expired->next, deadline))
1563 expired->next = deadline;
1567 void blk_mq_put_rq_ref(struct request *rq)
1569 if (is_flush_rq(rq)) {
1570 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1571 blk_mq_free_request(rq);
1572 } else if (req_ref_put_and_test(rq)) {
1573 __blk_mq_free_request(rq);
1577 static bool blk_mq_check_expired(struct request *rq, void *priv)
1579 struct blk_expired_data *expired = priv;
1582 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1583 * be reallocated underneath the timeout handler's processing, then
1584 * the expire check is reliable. If the request is not expired, then
1585 * it was completed and reallocated as a new request after returning
1586 * from blk_mq_check_expired().
1588 if (blk_mq_req_expired(rq, expired)) {
1589 expired->has_timedout_rq = true;
1595 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1597 struct blk_expired_data *expired = priv;
1599 if (blk_mq_req_expired(rq, expired))
1600 blk_mq_rq_timed_out(rq);
1604 static void blk_mq_timeout_work(struct work_struct *work)
1606 struct request_queue *q =
1607 container_of(work, struct request_queue, timeout_work);
1608 struct blk_expired_data expired = {
1609 .timeout_start = jiffies,
1611 struct blk_mq_hw_ctx *hctx;
1614 /* A deadlock might occur if a request is stuck requiring a
1615 * timeout at the same time a queue freeze is waiting
1616 * completion, since the timeout code would not be able to
1617 * acquire the queue reference here.
1619 * That's why we don't use blk_queue_enter here; instead, we use
1620 * percpu_ref_tryget directly, because we need to be able to
1621 * obtain a reference even in the short window between the queue
1622 * starting to freeze, by dropping the first reference in
1623 * blk_freeze_queue_start, and the moment the last request is
1624 * consumed, marked by the instant q_usage_counter reaches
1627 if (!percpu_ref_tryget(&q->q_usage_counter))
1630 /* check if there is any timed-out request */
1631 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1632 if (expired.has_timedout_rq) {
1634 * Before walking tags, we must ensure any submit started
1635 * before the current time has finished. Since the submit
1636 * uses srcu or rcu, wait for a synchronization point to
1637 * ensure all running submits have finished
1639 blk_mq_wait_quiesce_done(q->tag_set);
1642 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1645 if (expired.next != 0) {
1646 mod_timer(&q->timeout, expired.next);
1649 * Request timeouts are handled as a forward rolling timer. If
1650 * we end up here it means that no requests are pending and
1651 * also that no request has been pending for a while. Mark
1652 * each hctx as idle.
1654 queue_for_each_hw_ctx(q, hctx, i) {
1655 /* the hctx may be unmapped, so check it here */
1656 if (blk_mq_hw_queue_mapped(hctx))
1657 blk_mq_tag_idle(hctx);
1663 struct flush_busy_ctx_data {
1664 struct blk_mq_hw_ctx *hctx;
1665 struct list_head *list;
1668 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1670 struct flush_busy_ctx_data *flush_data = data;
1671 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1672 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1673 enum hctx_type type = hctx->type;
1675 spin_lock(&ctx->lock);
1676 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1677 sbitmap_clear_bit(sb, bitnr);
1678 spin_unlock(&ctx->lock);
1683 * Process software queues that have been marked busy, splicing them
1684 * to the for-dispatch
1686 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1688 struct flush_busy_ctx_data data = {
1693 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1695 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1697 struct dispatch_rq_data {
1698 struct blk_mq_hw_ctx *hctx;
1702 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1705 struct dispatch_rq_data *dispatch_data = data;
1706 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1707 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1708 enum hctx_type type = hctx->type;
1710 spin_lock(&ctx->lock);
1711 if (!list_empty(&ctx->rq_lists[type])) {
1712 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1713 list_del_init(&dispatch_data->rq->queuelist);
1714 if (list_empty(&ctx->rq_lists[type]))
1715 sbitmap_clear_bit(sb, bitnr);
1717 spin_unlock(&ctx->lock);
1719 return !dispatch_data->rq;
1722 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1723 struct blk_mq_ctx *start)
1725 unsigned off = start ? start->index_hw[hctx->type] : 0;
1726 struct dispatch_rq_data data = {
1731 __sbitmap_for_each_set(&hctx->ctx_map, off,
1732 dispatch_rq_from_ctx, &data);
1737 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1739 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1740 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1743 blk_mq_tag_busy(rq->mq_hctx);
1745 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1746 bt = &rq->mq_hctx->tags->breserved_tags;
1749 if (!hctx_may_queue(rq->mq_hctx, bt))
1753 tag = __sbitmap_queue_get(bt);
1754 if (tag == BLK_MQ_NO_TAG)
1757 rq->tag = tag + tag_offset;
1761 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1763 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1766 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1767 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1768 rq->rq_flags |= RQF_MQ_INFLIGHT;
1769 __blk_mq_inc_active_requests(hctx);
1771 hctx->tags->rqs[rq->tag] = rq;
1775 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1776 int flags, void *key)
1778 struct blk_mq_hw_ctx *hctx;
1780 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1782 spin_lock(&hctx->dispatch_wait_lock);
1783 if (!list_empty(&wait->entry)) {
1784 struct sbitmap_queue *sbq;
1786 list_del_init(&wait->entry);
1787 sbq = &hctx->tags->bitmap_tags;
1788 atomic_dec(&sbq->ws_active);
1790 spin_unlock(&hctx->dispatch_wait_lock);
1792 blk_mq_run_hw_queue(hctx, true);
1797 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1798 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1799 * restart. For both cases, take care to check the condition again after
1800 * marking us as waiting.
1802 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1805 struct sbitmap_queue *sbq;
1806 struct wait_queue_head *wq;
1807 wait_queue_entry_t *wait;
1810 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1811 !(blk_mq_is_shared_tags(hctx->flags))) {
1812 blk_mq_sched_mark_restart_hctx(hctx);
1815 * It's possible that a tag was freed in the window between the
1816 * allocation failure and adding the hardware queue to the wait
1819 * Don't clear RESTART here, someone else could have set it.
1820 * At most this will cost an extra queue run.
1822 return blk_mq_get_driver_tag(rq);
1825 wait = &hctx->dispatch_wait;
1826 if (!list_empty_careful(&wait->entry))
1829 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1830 sbq = &hctx->tags->breserved_tags;
1832 sbq = &hctx->tags->bitmap_tags;
1833 wq = &bt_wait_ptr(sbq, hctx)->wait;
1835 spin_lock_irq(&wq->lock);
1836 spin_lock(&hctx->dispatch_wait_lock);
1837 if (!list_empty(&wait->entry)) {
1838 spin_unlock(&hctx->dispatch_wait_lock);
1839 spin_unlock_irq(&wq->lock);
1843 atomic_inc(&sbq->ws_active);
1844 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1845 __add_wait_queue(wq, wait);
1848 * It's possible that a tag was freed in the window between the
1849 * allocation failure and adding the hardware queue to the wait
1852 ret = blk_mq_get_driver_tag(rq);
1854 spin_unlock(&hctx->dispatch_wait_lock);
1855 spin_unlock_irq(&wq->lock);
1860 * We got a tag, remove ourselves from the wait queue to ensure
1861 * someone else gets the wakeup.
1863 list_del_init(&wait->entry);
1864 atomic_dec(&sbq->ws_active);
1865 spin_unlock(&hctx->dispatch_wait_lock);
1866 spin_unlock_irq(&wq->lock);
1871 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1872 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1874 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1875 * - EWMA is one simple way to compute running average value
1876 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1877 * - take 4 as factor for avoiding to get too small(0) result, and this
1878 * factor doesn't matter because EWMA decreases exponentially
1880 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1884 ewma = hctx->dispatch_busy;
1889 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1891 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1892 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1894 hctx->dispatch_busy = ewma;
1897 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1899 static void blk_mq_handle_dev_resource(struct request *rq,
1900 struct list_head *list)
1902 list_add(&rq->queuelist, list);
1903 __blk_mq_requeue_request(rq);
1906 static void blk_mq_handle_zone_resource(struct request *rq,
1907 struct list_head *zone_list)
1910 * If we end up here it is because we cannot dispatch a request to a
1911 * specific zone due to LLD level zone-write locking or other zone
1912 * related resource not being available. In this case, set the request
1913 * aside in zone_list for retrying it later.
1915 list_add(&rq->queuelist, zone_list);
1916 __blk_mq_requeue_request(rq);
1919 enum prep_dispatch {
1921 PREP_DISPATCH_NO_TAG,
1922 PREP_DISPATCH_NO_BUDGET,
1925 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1928 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1929 int budget_token = -1;
1932 budget_token = blk_mq_get_dispatch_budget(rq->q);
1933 if (budget_token < 0) {
1934 blk_mq_put_driver_tag(rq);
1935 return PREP_DISPATCH_NO_BUDGET;
1937 blk_mq_set_rq_budget_token(rq, budget_token);
1940 if (!blk_mq_get_driver_tag(rq)) {
1942 * The initial allocation attempt failed, so we need to
1943 * rerun the hardware queue when a tag is freed. The
1944 * waitqueue takes care of that. If the queue is run
1945 * before we add this entry back on the dispatch list,
1946 * we'll re-run it below.
1948 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1950 * All budgets not got from this function will be put
1951 * together during handling partial dispatch
1954 blk_mq_put_dispatch_budget(rq->q, budget_token);
1955 return PREP_DISPATCH_NO_TAG;
1959 return PREP_DISPATCH_OK;
1962 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1963 static void blk_mq_release_budgets(struct request_queue *q,
1964 struct list_head *list)
1968 list_for_each_entry(rq, list, queuelist) {
1969 int budget_token = blk_mq_get_rq_budget_token(rq);
1971 if (budget_token >= 0)
1972 blk_mq_put_dispatch_budget(q, budget_token);
1977 * blk_mq_commit_rqs will notify driver using bd->last that there is no
1978 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
1980 * Attention, we should explicitly call this in unusual cases:
1981 * 1) did not queue everything initially scheduled to queue
1982 * 2) the last attempt to queue a request failed
1984 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
1987 if (hctx->queue->mq_ops->commit_rqs && queued) {
1988 trace_block_unplug(hctx->queue, queued, !from_schedule);
1989 hctx->queue->mq_ops->commit_rqs(hctx);
1994 * Returns true if we did some work AND can potentially do more.
1996 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1997 unsigned int nr_budgets)
1999 enum prep_dispatch prep;
2000 struct request_queue *q = hctx->queue;
2003 blk_status_t ret = BLK_STS_OK;
2004 LIST_HEAD(zone_list);
2005 bool needs_resource = false;
2007 if (list_empty(list))
2011 * Now process all the entries, sending them to the driver.
2015 struct blk_mq_queue_data bd;
2017 rq = list_first_entry(list, struct request, queuelist);
2019 WARN_ON_ONCE(hctx != rq->mq_hctx);
2020 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2021 if (prep != PREP_DISPATCH_OK)
2024 list_del_init(&rq->queuelist);
2027 bd.last = list_empty(list);
2030 * once the request is queued to lld, no need to cover the
2035 ret = q->mq_ops->queue_rq(hctx, &bd);
2040 case BLK_STS_RESOURCE:
2041 needs_resource = true;
2043 case BLK_STS_DEV_RESOURCE:
2044 blk_mq_handle_dev_resource(rq, list);
2046 case BLK_STS_ZONE_RESOURCE:
2048 * Move the request to zone_list and keep going through
2049 * the dispatch list to find more requests the drive can
2052 blk_mq_handle_zone_resource(rq, &zone_list);
2053 needs_resource = true;
2056 blk_mq_end_request(rq, ret);
2058 } while (!list_empty(list));
2060 if (!list_empty(&zone_list))
2061 list_splice_tail_init(&zone_list, list);
2063 /* If we didn't flush the entire list, we could have told the driver
2064 * there was more coming, but that turned out to be a lie.
2066 if (!list_empty(list) || ret != BLK_STS_OK)
2067 blk_mq_commit_rqs(hctx, queued, false);
2070 * Any items that need requeuing? Stuff them into hctx->dispatch,
2071 * that is where we will continue on next queue run.
2073 if (!list_empty(list)) {
2075 /* For non-shared tags, the RESTART check will suffice */
2076 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2077 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2078 blk_mq_is_shared_tags(hctx->flags));
2081 blk_mq_release_budgets(q, list);
2083 spin_lock(&hctx->lock);
2084 list_splice_tail_init(list, &hctx->dispatch);
2085 spin_unlock(&hctx->lock);
2088 * Order adding requests to hctx->dispatch and checking
2089 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2090 * in blk_mq_sched_restart(). Avoid restart code path to
2091 * miss the new added requests to hctx->dispatch, meantime
2092 * SCHED_RESTART is observed here.
2097 * If SCHED_RESTART was set by the caller of this function and
2098 * it is no longer set that means that it was cleared by another
2099 * thread and hence that a queue rerun is needed.
2101 * If 'no_tag' is set, that means that we failed getting
2102 * a driver tag with an I/O scheduler attached. If our dispatch
2103 * waitqueue is no longer active, ensure that we run the queue
2104 * AFTER adding our entries back to the list.
2106 * If no I/O scheduler has been configured it is possible that
2107 * the hardware queue got stopped and restarted before requests
2108 * were pushed back onto the dispatch list. Rerun the queue to
2109 * avoid starvation. Notes:
2110 * - blk_mq_run_hw_queue() checks whether or not a queue has
2111 * been stopped before rerunning a queue.
2112 * - Some but not all block drivers stop a queue before
2113 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2116 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2117 * bit is set, run queue after a delay to avoid IO stalls
2118 * that could otherwise occur if the queue is idle. We'll do
2119 * similar if we couldn't get budget or couldn't lock a zone
2120 * and SCHED_RESTART is set.
2122 needs_restart = blk_mq_sched_needs_restart(hctx);
2123 if (prep == PREP_DISPATCH_NO_BUDGET)
2124 needs_resource = true;
2125 if (!needs_restart ||
2126 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2127 blk_mq_run_hw_queue(hctx, true);
2128 else if (needs_resource)
2129 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2131 blk_mq_update_dispatch_busy(hctx, true);
2135 blk_mq_update_dispatch_busy(hctx, false);
2139 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2141 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2143 if (cpu >= nr_cpu_ids)
2144 cpu = cpumask_first(hctx->cpumask);
2149 * It'd be great if the workqueue API had a way to pass
2150 * in a mask and had some smarts for more clever placement.
2151 * For now we just round-robin here, switching for every
2152 * BLK_MQ_CPU_WORK_BATCH queued items.
2154 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2157 int next_cpu = hctx->next_cpu;
2159 if (hctx->queue->nr_hw_queues == 1)
2160 return WORK_CPU_UNBOUND;
2162 if (--hctx->next_cpu_batch <= 0) {
2164 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2166 if (next_cpu >= nr_cpu_ids)
2167 next_cpu = blk_mq_first_mapped_cpu(hctx);
2168 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2172 * Do unbound schedule if we can't find a online CPU for this hctx,
2173 * and it should only happen in the path of handling CPU DEAD.
2175 if (!cpu_online(next_cpu)) {
2182 * Make sure to re-select CPU next time once after CPUs
2183 * in hctx->cpumask become online again.
2185 hctx->next_cpu = next_cpu;
2186 hctx->next_cpu_batch = 1;
2187 return WORK_CPU_UNBOUND;
2190 hctx->next_cpu = next_cpu;
2195 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2196 * @hctx: Pointer to the hardware queue to run.
2197 * @msecs: Milliseconds of delay to wait before running the queue.
2199 * Run a hardware queue asynchronously with a delay of @msecs.
2201 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2203 if (unlikely(blk_mq_hctx_stopped(hctx)))
2205 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2206 msecs_to_jiffies(msecs));
2208 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2211 * blk_mq_run_hw_queue - Start to run a hardware queue.
2212 * @hctx: Pointer to the hardware queue to run.
2213 * @async: If we want to run the queue asynchronously.
2215 * Check if the request queue is not in a quiesced state and if there are
2216 * pending requests to be sent. If this is true, run the queue to send requests
2219 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2224 * We can't run the queue inline with interrupts disabled.
2226 WARN_ON_ONCE(!async && in_interrupt());
2229 * When queue is quiesced, we may be switching io scheduler, or
2230 * updating nr_hw_queues, or other things, and we can't run queue
2231 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2233 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2236 __blk_mq_run_dispatch_ops(hctx->queue, false,
2237 need_run = !blk_queue_quiesced(hctx->queue) &&
2238 blk_mq_hctx_has_pending(hctx));
2243 if (async || (hctx->flags & BLK_MQ_F_BLOCKING) ||
2244 !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2245 blk_mq_delay_run_hw_queue(hctx, 0);
2249 blk_mq_run_dispatch_ops(hctx->queue,
2250 blk_mq_sched_dispatch_requests(hctx));
2252 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2255 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2258 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2260 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2262 * If the IO scheduler does not respect hardware queues when
2263 * dispatching, we just don't bother with multiple HW queues and
2264 * dispatch from hctx for the current CPU since running multiple queues
2265 * just causes lock contention inside the scheduler and pointless cache
2268 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2270 if (!blk_mq_hctx_stopped(hctx))
2276 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2277 * @q: Pointer to the request queue to run.
2278 * @async: If we want to run the queue asynchronously.
2280 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2282 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2286 if (blk_queue_sq_sched(q))
2287 sq_hctx = blk_mq_get_sq_hctx(q);
2288 queue_for_each_hw_ctx(q, hctx, i) {
2289 if (blk_mq_hctx_stopped(hctx))
2292 * Dispatch from this hctx either if there's no hctx preferred
2293 * by IO scheduler or if it has requests that bypass the
2296 if (!sq_hctx || sq_hctx == hctx ||
2297 !list_empty_careful(&hctx->dispatch))
2298 blk_mq_run_hw_queue(hctx, async);
2301 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2304 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2305 * @q: Pointer to the request queue to run.
2306 * @msecs: Milliseconds of delay to wait before running the queues.
2308 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2310 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2314 if (blk_queue_sq_sched(q))
2315 sq_hctx = blk_mq_get_sq_hctx(q);
2316 queue_for_each_hw_ctx(q, hctx, i) {
2317 if (blk_mq_hctx_stopped(hctx))
2320 * If there is already a run_work pending, leave the
2321 * pending delay untouched. Otherwise, a hctx can stall
2322 * if another hctx is re-delaying the other's work
2323 * before the work executes.
2325 if (delayed_work_pending(&hctx->run_work))
2328 * Dispatch from this hctx either if there's no hctx preferred
2329 * by IO scheduler or if it has requests that bypass the
2332 if (!sq_hctx || sq_hctx == hctx ||
2333 !list_empty_careful(&hctx->dispatch))
2334 blk_mq_delay_run_hw_queue(hctx, msecs);
2337 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2340 * This function is often used for pausing .queue_rq() by driver when
2341 * there isn't enough resource or some conditions aren't satisfied, and
2342 * BLK_STS_RESOURCE is usually returned.
2344 * We do not guarantee that dispatch can be drained or blocked
2345 * after blk_mq_stop_hw_queue() returns. Please use
2346 * blk_mq_quiesce_queue() for that requirement.
2348 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2350 cancel_delayed_work(&hctx->run_work);
2352 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2354 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2357 * This function is often used for pausing .queue_rq() by driver when
2358 * there isn't enough resource or some conditions aren't satisfied, and
2359 * BLK_STS_RESOURCE is usually returned.
2361 * We do not guarantee that dispatch can be drained or blocked
2362 * after blk_mq_stop_hw_queues() returns. Please use
2363 * blk_mq_quiesce_queue() for that requirement.
2365 void blk_mq_stop_hw_queues(struct request_queue *q)
2367 struct blk_mq_hw_ctx *hctx;
2370 queue_for_each_hw_ctx(q, hctx, i)
2371 blk_mq_stop_hw_queue(hctx);
2373 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2375 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2377 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2379 blk_mq_run_hw_queue(hctx, false);
2381 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2383 void blk_mq_start_hw_queues(struct request_queue *q)
2385 struct blk_mq_hw_ctx *hctx;
2388 queue_for_each_hw_ctx(q, hctx, i)
2389 blk_mq_start_hw_queue(hctx);
2391 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2393 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2395 if (!blk_mq_hctx_stopped(hctx))
2398 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2399 blk_mq_run_hw_queue(hctx, async);
2401 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2403 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2405 struct blk_mq_hw_ctx *hctx;
2408 queue_for_each_hw_ctx(q, hctx, i)
2409 blk_mq_start_stopped_hw_queue(hctx, async);
2411 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2413 static void blk_mq_run_work_fn(struct work_struct *work)
2415 struct blk_mq_hw_ctx *hctx =
2416 container_of(work, struct blk_mq_hw_ctx, run_work.work);
2418 blk_mq_run_dispatch_ops(hctx->queue,
2419 blk_mq_sched_dispatch_requests(hctx));
2423 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2424 * @rq: Pointer to request to be inserted.
2425 * @flags: BLK_MQ_INSERT_*
2427 * Should only be used carefully, when the caller knows we want to
2428 * bypass a potential IO scheduler on the target device.
2430 void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2432 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2434 spin_lock(&hctx->lock);
2435 if (flags & BLK_MQ_INSERT_AT_HEAD)
2436 list_add(&rq->queuelist, &hctx->dispatch);
2438 list_add_tail(&rq->queuelist, &hctx->dispatch);
2439 spin_unlock(&hctx->lock);
2442 static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2443 struct blk_mq_ctx *ctx, struct list_head *list,
2444 bool run_queue_async)
2447 enum hctx_type type = hctx->type;
2450 * Try to issue requests directly if the hw queue isn't busy to save an
2451 * extra enqueue & dequeue to the sw queue.
2453 if (!hctx->dispatch_busy && !run_queue_async) {
2454 blk_mq_run_dispatch_ops(hctx->queue,
2455 blk_mq_try_issue_list_directly(hctx, list));
2456 if (list_empty(list))
2461 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2464 list_for_each_entry(rq, list, queuelist) {
2465 BUG_ON(rq->mq_ctx != ctx);
2466 trace_block_rq_insert(rq);
2469 spin_lock(&ctx->lock);
2470 list_splice_tail_init(list, &ctx->rq_lists[type]);
2471 blk_mq_hctx_mark_pending(hctx, ctx);
2472 spin_unlock(&ctx->lock);
2474 blk_mq_run_hw_queue(hctx, run_queue_async);
2477 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2479 struct request_queue *q = rq->q;
2480 struct blk_mq_ctx *ctx = rq->mq_ctx;
2481 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2483 if (blk_rq_is_passthrough(rq)) {
2485 * Passthrough request have to be added to hctx->dispatch
2486 * directly. The device may be in a situation where it can't
2487 * handle FS request, and always returns BLK_STS_RESOURCE for
2488 * them, which gets them added to hctx->dispatch.
2490 * If a passthrough request is required to unblock the queues,
2491 * and it is added to the scheduler queue, there is no chance to
2492 * dispatch it given we prioritize requests in hctx->dispatch.
2494 blk_mq_request_bypass_insert(rq, flags);
2495 } else if (rq->rq_flags & RQF_FLUSH_SEQ) {
2497 * Firstly normal IO request is inserted to scheduler queue or
2498 * sw queue, meantime we add flush request to dispatch queue(
2499 * hctx->dispatch) directly and there is at most one in-flight
2500 * flush request for each hw queue, so it doesn't matter to add
2501 * flush request to tail or front of the dispatch queue.
2503 * Secondly in case of NCQ, flush request belongs to non-NCQ
2504 * command, and queueing it will fail when there is any
2505 * in-flight normal IO request(NCQ command). When adding flush
2506 * rq to the front of hctx->dispatch, it is easier to introduce
2507 * extra time to flush rq's latency because of S_SCHED_RESTART
2508 * compared with adding to the tail of dispatch queue, then
2509 * chance of flush merge is increased, and less flush requests
2510 * will be issued to controller. It is observed that ~10% time
2511 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2512 * drive when adding flush rq to the front of hctx->dispatch.
2514 * Simply queue flush rq to the front of hctx->dispatch so that
2515 * intensive flush workloads can benefit in case of NCQ HW.
2517 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2518 } else if (q->elevator) {
2521 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2523 list_add(&rq->queuelist, &list);
2524 q->elevator->type->ops.insert_requests(hctx, &list, flags);
2526 trace_block_rq_insert(rq);
2528 spin_lock(&ctx->lock);
2529 if (flags & BLK_MQ_INSERT_AT_HEAD)
2530 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2532 list_add_tail(&rq->queuelist,
2533 &ctx->rq_lists[hctx->type]);
2534 blk_mq_hctx_mark_pending(hctx, ctx);
2535 spin_unlock(&ctx->lock);
2539 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2540 unsigned int nr_segs)
2544 if (bio->bi_opf & REQ_RAHEAD)
2545 rq->cmd_flags |= REQ_FAILFAST_MASK;
2547 rq->__sector = bio->bi_iter.bi_sector;
2548 blk_rq_bio_prep(rq, bio, nr_segs);
2550 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2551 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2554 blk_account_io_start(rq);
2557 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2558 struct request *rq, bool last)
2560 struct request_queue *q = rq->q;
2561 struct blk_mq_queue_data bd = {
2568 * For OK queue, we are done. For error, caller may kill it.
2569 * Any other error (busy), just add it to our list as we
2570 * previously would have done.
2572 ret = q->mq_ops->queue_rq(hctx, &bd);
2575 blk_mq_update_dispatch_busy(hctx, false);
2577 case BLK_STS_RESOURCE:
2578 case BLK_STS_DEV_RESOURCE:
2579 blk_mq_update_dispatch_busy(hctx, true);
2580 __blk_mq_requeue_request(rq);
2583 blk_mq_update_dispatch_busy(hctx, false);
2590 static bool blk_mq_get_budget_and_tag(struct request *rq)
2594 budget_token = blk_mq_get_dispatch_budget(rq->q);
2595 if (budget_token < 0)
2597 blk_mq_set_rq_budget_token(rq, budget_token);
2598 if (!blk_mq_get_driver_tag(rq)) {
2599 blk_mq_put_dispatch_budget(rq->q, budget_token);
2606 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2607 * @hctx: Pointer of the associated hardware queue.
2608 * @rq: Pointer to request to be sent.
2610 * If the device has enough resources to accept a new request now, send the
2611 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2612 * we can try send it another time in the future. Requests inserted at this
2613 * queue have higher priority.
2615 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2620 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2621 blk_mq_insert_request(rq, 0);
2625 if ((rq->rq_flags & RQF_ELV) || !blk_mq_get_budget_and_tag(rq)) {
2626 blk_mq_insert_request(rq, 0);
2627 blk_mq_run_hw_queue(hctx, false);
2631 ret = __blk_mq_issue_directly(hctx, rq, true);
2635 case BLK_STS_RESOURCE:
2636 case BLK_STS_DEV_RESOURCE:
2637 blk_mq_request_bypass_insert(rq, 0);
2638 blk_mq_run_hw_queue(hctx, false);
2641 blk_mq_end_request(rq, ret);
2646 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2648 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2650 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2651 blk_mq_insert_request(rq, 0);
2655 if (!blk_mq_get_budget_and_tag(rq))
2656 return BLK_STS_RESOURCE;
2657 return __blk_mq_issue_directly(hctx, rq, last);
2660 static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2662 struct blk_mq_hw_ctx *hctx = NULL;
2665 blk_status_t ret = BLK_STS_OK;
2667 while ((rq = rq_list_pop(&plug->mq_list))) {
2668 bool last = rq_list_empty(plug->mq_list);
2670 if (hctx != rq->mq_hctx) {
2672 blk_mq_commit_rqs(hctx, queued, false);
2678 ret = blk_mq_request_issue_directly(rq, last);
2683 case BLK_STS_RESOURCE:
2684 case BLK_STS_DEV_RESOURCE:
2685 blk_mq_request_bypass_insert(rq, 0);
2686 blk_mq_run_hw_queue(hctx, false);
2689 blk_mq_end_request(rq, ret);
2695 if (ret != BLK_STS_OK)
2696 blk_mq_commit_rqs(hctx, queued, false);
2699 static void __blk_mq_flush_plug_list(struct request_queue *q,
2700 struct blk_plug *plug)
2702 if (blk_queue_quiesced(q))
2704 q->mq_ops->queue_rqs(&plug->mq_list);
2707 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2709 struct blk_mq_hw_ctx *this_hctx = NULL;
2710 struct blk_mq_ctx *this_ctx = NULL;
2711 struct request *requeue_list = NULL;
2712 struct request **requeue_lastp = &requeue_list;
2713 unsigned int depth = 0;
2714 bool is_passthrough = false;
2718 struct request *rq = rq_list_pop(&plug->mq_list);
2721 this_hctx = rq->mq_hctx;
2722 this_ctx = rq->mq_ctx;
2723 is_passthrough = blk_rq_is_passthrough(rq);
2724 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2725 is_passthrough != blk_rq_is_passthrough(rq)) {
2726 rq_list_add_tail(&requeue_lastp, rq);
2729 list_add(&rq->queuelist, &list);
2731 } while (!rq_list_empty(plug->mq_list));
2733 plug->mq_list = requeue_list;
2734 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2736 percpu_ref_get(&this_hctx->queue->q_usage_counter);
2737 /* passthrough requests should never be issued to the I/O scheduler */
2738 if (this_hctx->queue->elevator && !is_passthrough) {
2739 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2741 blk_mq_run_hw_queue(this_hctx, from_sched);
2743 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2745 percpu_ref_put(&this_hctx->queue->q_usage_counter);
2748 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2752 if (rq_list_empty(plug->mq_list))
2756 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2757 struct request_queue *q;
2759 rq = rq_list_peek(&plug->mq_list);
2763 * Peek first request and see if we have a ->queue_rqs() hook.
2764 * If we do, we can dispatch the whole plug list in one go. We
2765 * already know at this point that all requests belong to the
2766 * same queue, caller must ensure that's the case.
2768 * Since we pass off the full list to the driver at this point,
2769 * we do not increment the active request count for the queue.
2770 * Bypass shared tags for now because of that.
2772 if (q->mq_ops->queue_rqs &&
2773 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2774 blk_mq_run_dispatch_ops(q,
2775 __blk_mq_flush_plug_list(q, plug));
2776 if (rq_list_empty(plug->mq_list))
2780 blk_mq_run_dispatch_ops(q,
2781 blk_mq_plug_issue_direct(plug));
2782 if (rq_list_empty(plug->mq_list))
2787 blk_mq_dispatch_plug_list(plug, from_schedule);
2788 } while (!rq_list_empty(plug->mq_list));
2791 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2792 struct list_head *list)
2795 blk_status_t ret = BLK_STS_OK;
2797 while (!list_empty(list)) {
2798 struct request *rq = list_first_entry(list, struct request,
2801 list_del_init(&rq->queuelist);
2802 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2807 case BLK_STS_RESOURCE:
2808 case BLK_STS_DEV_RESOURCE:
2809 blk_mq_request_bypass_insert(rq, 0);
2810 if (list_empty(list))
2811 blk_mq_run_hw_queue(hctx, false);
2814 blk_mq_end_request(rq, ret);
2820 if (ret != BLK_STS_OK)
2821 blk_mq_commit_rqs(hctx, queued, false);
2824 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2825 struct bio *bio, unsigned int nr_segs)
2827 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2828 if (blk_attempt_plug_merge(q, bio, nr_segs))
2830 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2836 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2837 struct blk_plug *plug,
2841 struct blk_mq_alloc_data data = {
2844 .cmd_flags = bio->bi_opf,
2848 if (unlikely(bio_queue_enter(bio)))
2851 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2854 rq_qos_throttle(q, bio);
2857 data.nr_tags = plug->nr_ios;
2859 data.cached_rq = &plug->cached_rq;
2862 rq = __blk_mq_alloc_requests(&data);
2865 rq_qos_cleanup(q, bio);
2866 if (bio->bi_opf & REQ_NOWAIT)
2867 bio_wouldblock_error(bio);
2873 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2874 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2877 enum hctx_type type, hctx_type;
2881 rq = rq_list_peek(&plug->cached_rq);
2882 if (!rq || rq->q != q)
2885 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2890 type = blk_mq_get_hctx_type((*bio)->bi_opf);
2891 hctx_type = rq->mq_hctx->type;
2892 if (type != hctx_type &&
2893 !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2895 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2899 * If any qos ->throttle() end up blocking, we will have flushed the
2900 * plug and hence killed the cached_rq list as well. Pop this entry
2901 * before we throttle.
2903 plug->cached_rq = rq_list_next(rq);
2904 rq_qos_throttle(q, *bio);
2906 rq->cmd_flags = (*bio)->bi_opf;
2907 INIT_LIST_HEAD(&rq->queuelist);
2911 static void bio_set_ioprio(struct bio *bio)
2913 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2914 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2915 bio->bi_ioprio = get_current_ioprio();
2916 blkcg_set_ioprio(bio);
2920 * blk_mq_submit_bio - Create and send a request to block device.
2921 * @bio: Bio pointer.
2923 * Builds up a request structure from @q and @bio and send to the device. The
2924 * request may not be queued directly to hardware if:
2925 * * This request can be merged with another one
2926 * * We want to place request at plug queue for possible future merging
2927 * * There is an IO scheduler active at this queue
2929 * It will not queue the request if there is an error with the bio, or at the
2932 void blk_mq_submit_bio(struct bio *bio)
2934 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2935 struct blk_plug *plug = blk_mq_plug(bio);
2936 const int is_sync = op_is_sync(bio->bi_opf);
2937 struct blk_mq_hw_ctx *hctx;
2939 unsigned int nr_segs = 1;
2942 bio = blk_queue_bounce(bio, q);
2943 if (bio_may_exceed_limits(bio, &q->limits)) {
2944 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2949 if (!bio_integrity_prep(bio))
2952 bio_set_ioprio(bio);
2954 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2958 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2963 trace_block_getrq(bio);
2965 rq_qos_track(q, rq, bio);
2967 blk_mq_bio_to_request(rq, bio, nr_segs);
2969 ret = blk_crypto_rq_get_keyslot(rq);
2970 if (ret != BLK_STS_OK) {
2971 bio->bi_status = ret;
2973 blk_mq_free_request(rq);
2977 if (op_is_flush(bio->bi_opf)) {
2978 blk_insert_flush(rq);
2983 blk_add_rq_to_plug(plug, rq);
2988 if ((rq->rq_flags & RQF_ELV) ||
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->requeue_list);
4240 spin_lock_init(&q->requeue_lock);
4242 q->nr_requests = set->queue_depth;
4244 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4245 blk_mq_add_queue_tag_set(set, q);
4246 blk_mq_map_swqueue(q);
4255 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4257 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4258 void blk_mq_exit_queue(struct request_queue *q)
4260 struct blk_mq_tag_set *set = q->tag_set;
4262 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4263 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4264 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4265 blk_mq_del_queue_tag_set(q);
4268 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4272 if (blk_mq_is_shared_tags(set->flags)) {
4273 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4276 if (!set->shared_tags)
4280 for (i = 0; i < set->nr_hw_queues; i++) {
4281 if (!__blk_mq_alloc_map_and_rqs(set, i))
4290 __blk_mq_free_map_and_rqs(set, i);
4292 if (blk_mq_is_shared_tags(set->flags)) {
4293 blk_mq_free_map_and_rqs(set, set->shared_tags,
4294 BLK_MQ_NO_HCTX_IDX);
4301 * Allocate the request maps associated with this tag_set. Note that this
4302 * may reduce the depth asked for, if memory is tight. set->queue_depth
4303 * will be updated to reflect the allocated depth.
4305 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4310 depth = set->queue_depth;
4312 err = __blk_mq_alloc_rq_maps(set);
4316 set->queue_depth >>= 1;
4317 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4321 } while (set->queue_depth);
4323 if (!set->queue_depth || err) {
4324 pr_err("blk-mq: failed to allocate request map\n");
4328 if (depth != set->queue_depth)
4329 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4330 depth, set->queue_depth);
4335 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4338 * blk_mq_map_queues() and multiple .map_queues() implementations
4339 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4340 * number of hardware queues.
4342 if (set->nr_maps == 1)
4343 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4345 if (set->ops->map_queues && !is_kdump_kernel()) {
4349 * transport .map_queues is usually done in the following
4352 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4353 * mask = get_cpu_mask(queue)
4354 * for_each_cpu(cpu, mask)
4355 * set->map[x].mq_map[cpu] = queue;
4358 * When we need to remap, the table has to be cleared for
4359 * killing stale mapping since one CPU may not be mapped
4362 for (i = 0; i < set->nr_maps; i++)
4363 blk_mq_clear_mq_map(&set->map[i]);
4365 set->ops->map_queues(set);
4367 BUG_ON(set->nr_maps > 1);
4368 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4372 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4373 int new_nr_hw_queues)
4375 struct blk_mq_tags **new_tags;
4377 if (set->nr_hw_queues >= new_nr_hw_queues)
4380 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4381 GFP_KERNEL, set->numa_node);
4386 memcpy(new_tags, set->tags, set->nr_hw_queues *
4387 sizeof(*set->tags));
4389 set->tags = new_tags;
4391 set->nr_hw_queues = new_nr_hw_queues;
4396 * Alloc a tag set to be associated with one or more request queues.
4397 * May fail with EINVAL for various error conditions. May adjust the
4398 * requested depth down, if it's too large. In that case, the set
4399 * value will be stored in set->queue_depth.
4401 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4405 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4407 if (!set->nr_hw_queues)
4409 if (!set->queue_depth)
4411 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4414 if (!set->ops->queue_rq)
4417 if (!set->ops->get_budget ^ !set->ops->put_budget)
4420 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4421 pr_info("blk-mq: reduced tag depth to %u\n",
4423 set->queue_depth = BLK_MQ_MAX_DEPTH;
4428 else if (set->nr_maps > HCTX_MAX_TYPES)
4432 * If a crashdump is active, then we are potentially in a very
4433 * memory constrained environment. Limit us to 1 queue and
4434 * 64 tags to prevent using too much memory.
4436 if (is_kdump_kernel()) {
4437 set->nr_hw_queues = 1;
4439 set->queue_depth = min(64U, set->queue_depth);
4442 * There is no use for more h/w queues than cpus if we just have
4445 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4446 set->nr_hw_queues = nr_cpu_ids;
4448 if (set->flags & BLK_MQ_F_BLOCKING) {
4449 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4452 ret = init_srcu_struct(set->srcu);
4458 set->tags = kcalloc_node(set->nr_hw_queues,
4459 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4462 goto out_cleanup_srcu;
4464 for (i = 0; i < set->nr_maps; i++) {
4465 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4466 sizeof(set->map[i].mq_map[0]),
4467 GFP_KERNEL, set->numa_node);
4468 if (!set->map[i].mq_map)
4469 goto out_free_mq_map;
4470 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4473 blk_mq_update_queue_map(set);
4475 ret = blk_mq_alloc_set_map_and_rqs(set);
4477 goto out_free_mq_map;
4479 mutex_init(&set->tag_list_lock);
4480 INIT_LIST_HEAD(&set->tag_list);
4485 for (i = 0; i < set->nr_maps; i++) {
4486 kfree(set->map[i].mq_map);
4487 set->map[i].mq_map = NULL;
4492 if (set->flags & BLK_MQ_F_BLOCKING)
4493 cleanup_srcu_struct(set->srcu);
4495 if (set->flags & BLK_MQ_F_BLOCKING)
4499 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4501 /* allocate and initialize a tagset for a simple single-queue device */
4502 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4503 const struct blk_mq_ops *ops, unsigned int queue_depth,
4504 unsigned int set_flags)
4506 memset(set, 0, sizeof(*set));
4508 set->nr_hw_queues = 1;
4510 set->queue_depth = queue_depth;
4511 set->numa_node = NUMA_NO_NODE;
4512 set->flags = set_flags;
4513 return blk_mq_alloc_tag_set(set);
4515 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4517 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4521 for (i = 0; i < set->nr_hw_queues; i++)
4522 __blk_mq_free_map_and_rqs(set, i);
4524 if (blk_mq_is_shared_tags(set->flags)) {
4525 blk_mq_free_map_and_rqs(set, set->shared_tags,
4526 BLK_MQ_NO_HCTX_IDX);
4529 for (j = 0; j < set->nr_maps; j++) {
4530 kfree(set->map[j].mq_map);
4531 set->map[j].mq_map = NULL;
4536 if (set->flags & BLK_MQ_F_BLOCKING) {
4537 cleanup_srcu_struct(set->srcu);
4541 EXPORT_SYMBOL(blk_mq_free_tag_set);
4543 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4545 struct blk_mq_tag_set *set = q->tag_set;
4546 struct blk_mq_hw_ctx *hctx;
4553 if (q->nr_requests == nr)
4556 blk_mq_freeze_queue(q);
4557 blk_mq_quiesce_queue(q);
4560 queue_for_each_hw_ctx(q, hctx, i) {
4564 * If we're using an MQ scheduler, just update the scheduler
4565 * queue depth. This is similar to what the old code would do.
4567 if (hctx->sched_tags) {
4568 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4571 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4576 if (q->elevator && q->elevator->type->ops.depth_updated)
4577 q->elevator->type->ops.depth_updated(hctx);
4580 q->nr_requests = nr;
4581 if (blk_mq_is_shared_tags(set->flags)) {
4583 blk_mq_tag_update_sched_shared_tags(q);
4585 blk_mq_tag_resize_shared_tags(set, nr);
4589 blk_mq_unquiesce_queue(q);
4590 blk_mq_unfreeze_queue(q);
4596 * request_queue and elevator_type pair.
4597 * It is just used by __blk_mq_update_nr_hw_queues to cache
4598 * the elevator_type associated with a request_queue.
4600 struct blk_mq_qe_pair {
4601 struct list_head node;
4602 struct request_queue *q;
4603 struct elevator_type *type;
4607 * Cache the elevator_type in qe pair list and switch the
4608 * io scheduler to 'none'
4610 static bool blk_mq_elv_switch_none(struct list_head *head,
4611 struct request_queue *q)
4613 struct blk_mq_qe_pair *qe;
4618 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4622 /* q->elevator needs protection from ->sysfs_lock */
4623 mutex_lock(&q->sysfs_lock);
4625 INIT_LIST_HEAD(&qe->node);
4627 qe->type = q->elevator->type;
4628 /* keep a reference to the elevator module as we'll switch back */
4629 __elevator_get(qe->type);
4630 list_add(&qe->node, head);
4631 elevator_disable(q);
4632 mutex_unlock(&q->sysfs_lock);
4637 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4638 struct request_queue *q)
4640 struct blk_mq_qe_pair *qe;
4642 list_for_each_entry(qe, head, node)
4649 static void blk_mq_elv_switch_back(struct list_head *head,
4650 struct request_queue *q)
4652 struct blk_mq_qe_pair *qe;
4653 struct elevator_type *t;
4655 qe = blk_lookup_qe_pair(head, q);
4659 list_del(&qe->node);
4662 mutex_lock(&q->sysfs_lock);
4663 elevator_switch(q, t);
4664 /* drop the reference acquired in blk_mq_elv_switch_none */
4666 mutex_unlock(&q->sysfs_lock);
4669 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4672 struct request_queue *q;
4674 int prev_nr_hw_queues;
4676 lockdep_assert_held(&set->tag_list_lock);
4678 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4679 nr_hw_queues = nr_cpu_ids;
4680 if (nr_hw_queues < 1)
4682 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4685 list_for_each_entry(q, &set->tag_list, tag_set_list)
4686 blk_mq_freeze_queue(q);
4688 * Switch IO scheduler to 'none', cleaning up the data associated
4689 * with the previous scheduler. We will switch back once we are done
4690 * updating the new sw to hw queue mappings.
4692 list_for_each_entry(q, &set->tag_list, tag_set_list)
4693 if (!blk_mq_elv_switch_none(&head, q))
4696 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4697 blk_mq_debugfs_unregister_hctxs(q);
4698 blk_mq_sysfs_unregister_hctxs(q);
4701 prev_nr_hw_queues = set->nr_hw_queues;
4702 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4706 blk_mq_update_queue_map(set);
4707 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4708 blk_mq_realloc_hw_ctxs(set, q);
4709 blk_mq_update_poll_flag(q);
4710 if (q->nr_hw_queues != set->nr_hw_queues) {
4711 int i = prev_nr_hw_queues;
4713 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4714 nr_hw_queues, prev_nr_hw_queues);
4715 for (; i < set->nr_hw_queues; i++)
4716 __blk_mq_free_map_and_rqs(set, i);
4718 set->nr_hw_queues = prev_nr_hw_queues;
4719 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4722 blk_mq_map_swqueue(q);
4726 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4727 blk_mq_sysfs_register_hctxs(q);
4728 blk_mq_debugfs_register_hctxs(q);
4732 list_for_each_entry(q, &set->tag_list, tag_set_list)
4733 blk_mq_elv_switch_back(&head, q);
4735 list_for_each_entry(q, &set->tag_list, tag_set_list)
4736 blk_mq_unfreeze_queue(q);
4739 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4741 mutex_lock(&set->tag_list_lock);
4742 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4743 mutex_unlock(&set->tag_list_lock);
4745 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4747 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4750 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4751 long state = get_current_state();
4755 ret = q->mq_ops->poll(hctx, iob);
4757 __set_current_state(TASK_RUNNING);
4761 if (signal_pending_state(state, current))
4762 __set_current_state(TASK_RUNNING);
4763 if (task_is_running(current))
4766 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4769 } while (!need_resched());
4771 __set_current_state(TASK_RUNNING);
4775 unsigned int blk_mq_rq_cpu(struct request *rq)
4777 return rq->mq_ctx->cpu;
4779 EXPORT_SYMBOL(blk_mq_rq_cpu);
4781 void blk_mq_cancel_work_sync(struct request_queue *q)
4783 struct blk_mq_hw_ctx *hctx;
4786 cancel_delayed_work_sync(&q->requeue_work);
4788 queue_for_each_hw_ctx(q, hctx, i)
4789 cancel_delayed_work_sync(&hctx->run_work);
4792 static int __init blk_mq_init(void)
4796 for_each_possible_cpu(i)
4797 init_llist_head(&per_cpu(blk_cpu_done, i));
4798 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4800 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4801 "block/softirq:dead", NULL,
4802 blk_softirq_cpu_dead);
4803 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4804 blk_mq_hctx_notify_dead);
4805 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4806 blk_mq_hctx_notify_online,
4807 blk_mq_hctx_notify_offline);
4810 subsys_initcall(blk_mq_init);