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>
32 #include <trace/events/block.h>
34 #include <linux/blk-mq.h>
35 #include <linux/t10-pi.h>
38 #include "blk-mq-debugfs.h"
39 #include "blk-mq-tag.h"
42 #include "blk-mq-sched.h"
43 #include "blk-rq-qos.h"
45 static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
47 static void blk_mq_poll_stats_start(struct request_queue *q);
48 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
50 static int blk_mq_poll_stats_bkt(const struct request *rq)
52 int ddir, sectors, bucket;
54 ddir = rq_data_dir(rq);
55 sectors = blk_rq_stats_sectors(rq);
57 bucket = ddir + 2 * ilog2(sectors);
61 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
62 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
67 #define BLK_QC_T_SHIFT 16
68 #define BLK_QC_T_INTERNAL (1U << 31)
70 static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q,
73 return q->queue_hw_ctx[(qc & ~BLK_QC_T_INTERNAL) >> BLK_QC_T_SHIFT];
76 static inline struct request *blk_qc_to_rq(struct blk_mq_hw_ctx *hctx,
79 unsigned int tag = qc & ((1U << BLK_QC_T_SHIFT) - 1);
81 if (qc & BLK_QC_T_INTERNAL)
82 return blk_mq_tag_to_rq(hctx->sched_tags, tag);
83 return blk_mq_tag_to_rq(hctx->tags, tag);
86 static inline blk_qc_t blk_rq_to_qc(struct request *rq)
88 return (rq->mq_hctx->queue_num << BLK_QC_T_SHIFT) |
90 rq->tag : (rq->internal_tag | BLK_QC_T_INTERNAL));
94 * Check if any of the ctx, dispatch list or elevator
95 * have pending work in this hardware queue.
97 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
99 return !list_empty_careful(&hctx->dispatch) ||
100 sbitmap_any_bit_set(&hctx->ctx_map) ||
101 blk_mq_sched_has_work(hctx);
105 * Mark this ctx as having pending work in this hardware queue
107 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
108 struct blk_mq_ctx *ctx)
110 const int bit = ctx->index_hw[hctx->type];
112 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
113 sbitmap_set_bit(&hctx->ctx_map, bit);
116 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
117 struct blk_mq_ctx *ctx)
119 const int bit = ctx->index_hw[hctx->type];
121 sbitmap_clear_bit(&hctx->ctx_map, bit);
125 struct block_device *part;
126 unsigned int inflight[2];
129 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
130 struct request *rq, void *priv,
133 struct mq_inflight *mi = priv;
135 if ((!mi->part->bd_partno || rq->part == mi->part) &&
136 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
137 mi->inflight[rq_data_dir(rq)]++;
142 unsigned int blk_mq_in_flight(struct request_queue *q,
143 struct block_device *part)
145 struct mq_inflight mi = { .part = part };
147 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
149 return mi.inflight[0] + mi.inflight[1];
152 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
153 unsigned int inflight[2])
155 struct mq_inflight mi = { .part = part };
157 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
158 inflight[0] = mi.inflight[0];
159 inflight[1] = mi.inflight[1];
162 void blk_freeze_queue_start(struct request_queue *q)
164 mutex_lock(&q->mq_freeze_lock);
165 if (++q->mq_freeze_depth == 1) {
166 percpu_ref_kill(&q->q_usage_counter);
167 mutex_unlock(&q->mq_freeze_lock);
169 blk_mq_run_hw_queues(q, false);
171 mutex_unlock(&q->mq_freeze_lock);
174 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
176 void blk_mq_freeze_queue_wait(struct request_queue *q)
178 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
180 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
182 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
183 unsigned long timeout)
185 return wait_event_timeout(q->mq_freeze_wq,
186 percpu_ref_is_zero(&q->q_usage_counter),
189 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
192 * Guarantee no request is in use, so we can change any data structure of
193 * the queue afterward.
195 void blk_freeze_queue(struct request_queue *q)
198 * In the !blk_mq case we are only calling this to kill the
199 * q_usage_counter, otherwise this increases the freeze depth
200 * and waits for it to return to zero. For this reason there is
201 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
202 * exported to drivers as the only user for unfreeze is blk_mq.
204 blk_freeze_queue_start(q);
205 blk_mq_freeze_queue_wait(q);
208 void blk_mq_freeze_queue(struct request_queue *q)
211 * ...just an alias to keep freeze and unfreeze actions balanced
212 * in the blk_mq_* namespace
216 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
218 void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
220 mutex_lock(&q->mq_freeze_lock);
222 q->q_usage_counter.data->force_atomic = true;
223 q->mq_freeze_depth--;
224 WARN_ON_ONCE(q->mq_freeze_depth < 0);
225 if (!q->mq_freeze_depth) {
226 percpu_ref_resurrect(&q->q_usage_counter);
227 wake_up_all(&q->mq_freeze_wq);
229 mutex_unlock(&q->mq_freeze_lock);
232 void blk_mq_unfreeze_queue(struct request_queue *q)
234 __blk_mq_unfreeze_queue(q, false);
236 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
239 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
240 * mpt3sas driver such that this function can be removed.
242 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
246 spin_lock_irqsave(&q->queue_lock, flags);
247 if (!q->quiesce_depth++)
248 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
249 spin_unlock_irqrestore(&q->queue_lock, flags);
251 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
254 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
257 * Note: this function does not prevent that the struct request end_io()
258 * callback function is invoked. Once this function is returned, we make
259 * sure no dispatch can happen until the queue is unquiesced via
260 * blk_mq_unquiesce_queue().
262 void blk_mq_quiesce_queue(struct request_queue *q)
264 struct blk_mq_hw_ctx *hctx;
268 blk_mq_quiesce_queue_nowait(q);
270 queue_for_each_hw_ctx(q, hctx, i) {
271 if (hctx->flags & BLK_MQ_F_BLOCKING)
272 synchronize_srcu(hctx->srcu);
279 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
282 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
285 * This function recovers queue into the state before quiescing
286 * which is done by blk_mq_quiesce_queue.
288 void blk_mq_unquiesce_queue(struct request_queue *q)
291 bool run_queue = false;
293 spin_lock_irqsave(&q->queue_lock, flags);
294 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
296 } else if (!--q->quiesce_depth) {
297 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
300 spin_unlock_irqrestore(&q->queue_lock, flags);
302 /* dispatch requests which are inserted during quiescing */
304 blk_mq_run_hw_queues(q, true);
306 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
308 void blk_mq_wake_waiters(struct request_queue *q)
310 struct blk_mq_hw_ctx *hctx;
313 queue_for_each_hw_ctx(q, hctx, i)
314 if (blk_mq_hw_queue_mapped(hctx))
315 blk_mq_tag_wakeup_all(hctx->tags, true);
318 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
319 unsigned int tag, u64 alloc_time_ns)
321 struct blk_mq_ctx *ctx = data->ctx;
322 struct blk_mq_hw_ctx *hctx = data->hctx;
323 struct request_queue *q = data->q;
324 struct elevator_queue *e = q->elevator;
325 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
326 struct request *rq = tags->static_rqs[tag];
327 unsigned int rq_flags = 0;
331 rq->tag = BLK_MQ_NO_TAG;
332 rq->internal_tag = tag;
335 rq->internal_tag = BLK_MQ_NO_TAG;
338 if (data->flags & BLK_MQ_REQ_PM)
340 if (blk_queue_io_stat(q))
341 rq_flags |= RQF_IO_STAT;
342 rq->rq_flags = rq_flags;
344 if (blk_mq_need_time_stamp(rq))
345 rq->start_time_ns = ktime_get_ns();
347 rq->start_time_ns = 0;
348 /* csd/requeue_work/fifo_time is initialized before use */
352 rq->cmd_flags = data->cmd_flags;
355 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
356 rq->alloc_time_ns = alloc_time_ns;
358 rq->io_start_time_ns = 0;
359 rq->stats_sectors = 0;
360 rq->nr_phys_segments = 0;
361 #if defined(CONFIG_BLK_DEV_INTEGRITY)
362 rq->nr_integrity_segments = 0;
366 rq->end_io_data = NULL;
368 blk_crypto_rq_set_defaults(rq);
369 INIT_LIST_HEAD(&rq->queuelist);
370 /* tag was already set */
371 WRITE_ONCE(rq->deadline, 0);
372 refcount_set(&rq->ref, 1);
374 if (rq->rq_flags & RQF_ELV) {
375 struct elevator_queue *e = data->q->elevator;
378 INIT_HLIST_NODE(&rq->hash);
379 RB_CLEAR_NODE(&rq->rb_node);
381 if (!op_is_flush(data->cmd_flags) &&
382 e->type->ops.prepare_request) {
383 if (e->type->icq_cache)
384 blk_mq_sched_assign_ioc(rq);
386 e->type->ops.prepare_request(rq);
387 rq->rq_flags |= RQF_ELVPRIV;
394 static inline struct request *
395 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data,
398 unsigned int tag, tag_offset;
403 tags = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
407 for (i = 0; tags; i++) {
408 if (!(tags & (1UL << i)))
410 tag = tag_offset + i;
412 rq = blk_mq_rq_ctx_init(data, tag, alloc_time_ns);
413 rq_list_add(data->cached_rq, rq);
417 return rq_list_pop(data->cached_rq);
420 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
422 struct request_queue *q = data->q;
423 struct elevator_queue *e = q->elevator;
424 u64 alloc_time_ns = 0;
428 /* alloc_time includes depth and tag waits */
429 if (blk_queue_rq_alloc_time(q))
430 alloc_time_ns = ktime_get_ns();
432 if (data->cmd_flags & REQ_NOWAIT)
433 data->flags |= BLK_MQ_REQ_NOWAIT;
437 * Flush/passthrough requests are special and go directly to the
438 * dispatch list. Don't include reserved tags in the
439 * limiting, as it isn't useful.
441 if (!op_is_flush(data->cmd_flags) &&
442 !blk_op_is_passthrough(data->cmd_flags) &&
443 e->type->ops.limit_depth &&
444 !(data->flags & BLK_MQ_REQ_RESERVED))
445 e->type->ops.limit_depth(data->cmd_flags, data);
449 data->ctx = blk_mq_get_ctx(q);
450 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
452 blk_mq_tag_busy(data->hctx);
455 * Try batched alloc if we want more than 1 tag.
457 if (data->nr_tags > 1) {
458 rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns);
465 * Waiting allocations only fail because of an inactive hctx. In that
466 * case just retry the hctx assignment and tag allocation as CPU hotplug
467 * should have migrated us to an online CPU by now.
469 tag = blk_mq_get_tag(data);
470 if (tag == BLK_MQ_NO_TAG) {
471 if (data->flags & BLK_MQ_REQ_NOWAIT)
474 * Give up the CPU and sleep for a random short time to
475 * ensure that thread using a realtime scheduling class
476 * are migrated off the CPU, and thus off the hctx that
483 return blk_mq_rq_ctx_init(data, tag, alloc_time_ns);
486 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
487 blk_mq_req_flags_t flags)
489 struct blk_mq_alloc_data data = {
498 ret = blk_queue_enter(q, flags);
502 rq = __blk_mq_alloc_requests(&data);
506 rq->__sector = (sector_t) -1;
507 rq->bio = rq->biotail = NULL;
511 return ERR_PTR(-EWOULDBLOCK);
513 EXPORT_SYMBOL(blk_mq_alloc_request);
515 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
516 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
518 struct blk_mq_alloc_data data = {
524 u64 alloc_time_ns = 0;
529 /* alloc_time includes depth and tag waits */
530 if (blk_queue_rq_alloc_time(q))
531 alloc_time_ns = ktime_get_ns();
534 * If the tag allocator sleeps we could get an allocation for a
535 * different hardware context. No need to complicate the low level
536 * allocator for this for the rare use case of a command tied to
539 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
540 return ERR_PTR(-EINVAL);
542 if (hctx_idx >= q->nr_hw_queues)
543 return ERR_PTR(-EIO);
545 ret = blk_queue_enter(q, flags);
550 * Check if the hardware context is actually mapped to anything.
551 * If not tell the caller that it should skip this queue.
554 data.hctx = q->queue_hw_ctx[hctx_idx];
555 if (!blk_mq_hw_queue_mapped(data.hctx))
557 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
558 data.ctx = __blk_mq_get_ctx(q, cpu);
561 blk_mq_tag_busy(data.hctx);
564 tag = blk_mq_get_tag(&data);
565 if (tag == BLK_MQ_NO_TAG)
567 return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns);
573 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
575 static void __blk_mq_free_request(struct request *rq)
577 struct request_queue *q = rq->q;
578 struct blk_mq_ctx *ctx = rq->mq_ctx;
579 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
580 const int sched_tag = rq->internal_tag;
582 blk_crypto_free_request(rq);
583 blk_pm_mark_last_busy(rq);
585 if (rq->tag != BLK_MQ_NO_TAG)
586 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
587 if (sched_tag != BLK_MQ_NO_TAG)
588 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
589 blk_mq_sched_restart(hctx);
593 void blk_mq_free_request(struct request *rq)
595 struct request_queue *q = rq->q;
596 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
598 if (rq->rq_flags & RQF_ELVPRIV) {
599 struct elevator_queue *e = q->elevator;
601 if (e->type->ops.finish_request)
602 e->type->ops.finish_request(rq);
604 put_io_context(rq->elv.icq->ioc);
609 if (rq->rq_flags & RQF_MQ_INFLIGHT)
610 __blk_mq_dec_active_requests(hctx);
612 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
613 laptop_io_completion(q->disk->bdi);
617 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
618 if (refcount_dec_and_test(&rq->ref))
619 __blk_mq_free_request(rq);
621 EXPORT_SYMBOL_GPL(blk_mq_free_request);
623 void blk_mq_free_plug_rqs(struct blk_plug *plug)
627 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL) {
628 percpu_ref_get(&rq->q->q_usage_counter);
629 blk_mq_free_request(rq);
633 static void req_bio_endio(struct request *rq, struct bio *bio,
634 unsigned int nbytes, blk_status_t error)
636 if (unlikely(error)) {
637 bio->bi_status = error;
638 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
640 * Partial zone append completions cannot be supported as the
641 * BIO fragments may end up not being written sequentially.
643 if (bio->bi_iter.bi_size != nbytes)
644 bio->bi_status = BLK_STS_IOERR;
646 bio->bi_iter.bi_sector = rq->__sector;
649 bio_advance(bio, nbytes);
651 if (unlikely(rq->rq_flags & RQF_QUIET))
652 bio_set_flag(bio, BIO_QUIET);
653 /* don't actually finish bio if it's part of flush sequence */
654 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
658 static void blk_account_io_completion(struct request *req, unsigned int bytes)
660 if (req->part && blk_do_io_stat(req)) {
661 const int sgrp = op_stat_group(req_op(req));
664 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
670 * blk_update_request - Complete multiple bytes without completing the request
671 * @req: the request being processed
672 * @error: block status code
673 * @nr_bytes: number of bytes to complete for @req
676 * Ends I/O on a number of bytes attached to @req, but doesn't complete
677 * the request structure even if @req doesn't have leftover.
678 * If @req has leftover, sets it up for the next range of segments.
680 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
681 * %false return from this function.
684 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
685 * except in the consistency check at the end of this function.
688 * %false - this request doesn't have any more data
689 * %true - this request has more data
691 bool blk_update_request(struct request *req, blk_status_t error,
692 unsigned int nr_bytes)
696 trace_block_rq_complete(req, error, nr_bytes);
701 #ifdef CONFIG_BLK_DEV_INTEGRITY
702 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
704 req->q->integrity.profile->complete_fn(req, nr_bytes);
707 if (unlikely(error && !blk_rq_is_passthrough(req) &&
708 !(req->rq_flags & RQF_QUIET)))
709 blk_print_req_error(req, error);
711 blk_account_io_completion(req, nr_bytes);
715 struct bio *bio = req->bio;
716 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
718 if (bio_bytes == bio->bi_iter.bi_size)
719 req->bio = bio->bi_next;
721 /* Completion has already been traced */
722 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
723 req_bio_endio(req, bio, bio_bytes, error);
725 total_bytes += bio_bytes;
726 nr_bytes -= bio_bytes;
737 * Reset counters so that the request stacking driver
738 * can find how many bytes remain in the request
745 req->__data_len -= total_bytes;
747 /* update sector only for requests with clear definition of sector */
748 if (!blk_rq_is_passthrough(req))
749 req->__sector += total_bytes >> 9;
751 /* mixed attributes always follow the first bio */
752 if (req->rq_flags & RQF_MIXED_MERGE) {
753 req->cmd_flags &= ~REQ_FAILFAST_MASK;
754 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
757 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
759 * If total number of sectors is less than the first segment
760 * size, something has gone terribly wrong.
762 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
763 blk_dump_rq_flags(req, "request botched");
764 req->__data_len = blk_rq_cur_bytes(req);
767 /* recalculate the number of segments */
768 req->nr_phys_segments = blk_recalc_rq_segments(req);
773 EXPORT_SYMBOL_GPL(blk_update_request);
775 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
777 if (rq->rq_flags & RQF_STATS) {
778 blk_mq_poll_stats_start(rq->q);
779 blk_stat_add(rq, now);
782 blk_mq_sched_completed_request(rq, now);
783 blk_account_io_done(rq, now);
786 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
788 if (blk_mq_need_time_stamp(rq))
789 __blk_mq_end_request_acct(rq, ktime_get_ns());
792 rq_qos_done(rq->q, rq);
793 rq->end_io(rq, error);
795 blk_mq_free_request(rq);
798 EXPORT_SYMBOL(__blk_mq_end_request);
800 void blk_mq_end_request(struct request *rq, blk_status_t error)
802 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
804 __blk_mq_end_request(rq, error);
806 EXPORT_SYMBOL(blk_mq_end_request);
808 #define TAG_COMP_BATCH 32
810 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
811 int *tag_array, int nr_tags)
813 struct request_queue *q = hctx->queue;
815 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
816 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
819 void blk_mq_end_request_batch(struct io_comp_batch *iob)
821 int tags[TAG_COMP_BATCH], nr_tags = 0;
822 struct blk_mq_hw_ctx *last_hctx = NULL;
827 now = ktime_get_ns();
829 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
831 prefetch(rq->rq_next);
833 blk_update_request(rq, BLK_STS_OK, blk_rq_bytes(rq));
835 __blk_mq_end_request_acct(rq, now);
837 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
838 if (!refcount_dec_and_test(&rq->ref))
841 blk_crypto_free_request(rq);
842 blk_pm_mark_last_busy(rq);
843 rq_qos_done(rq->q, rq);
845 if (nr_tags == TAG_COMP_BATCH ||
846 (last_hctx && last_hctx != rq->mq_hctx)) {
847 blk_mq_flush_tag_batch(last_hctx, tags, nr_tags);
850 tags[nr_tags++] = rq->tag;
851 last_hctx = rq->mq_hctx;
855 blk_mq_flush_tag_batch(last_hctx, tags, nr_tags);
857 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
859 static void blk_complete_reqs(struct llist_head *list)
861 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
862 struct request *rq, *next;
864 llist_for_each_entry_safe(rq, next, entry, ipi_list)
865 rq->q->mq_ops->complete(rq);
868 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
870 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
873 static int blk_softirq_cpu_dead(unsigned int cpu)
875 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
879 static void __blk_mq_complete_request_remote(void *data)
881 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
884 static inline bool blk_mq_complete_need_ipi(struct request *rq)
886 int cpu = raw_smp_processor_id();
888 if (!IS_ENABLED(CONFIG_SMP) ||
889 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
892 * With force threaded interrupts enabled, raising softirq from an SMP
893 * function call will always result in waking the ksoftirqd thread.
894 * This is probably worse than completing the request on a different
897 if (force_irqthreads())
900 /* same CPU or cache domain? Complete locally */
901 if (cpu == rq->mq_ctx->cpu ||
902 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
903 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
906 /* don't try to IPI to an offline CPU */
907 return cpu_online(rq->mq_ctx->cpu);
910 static void blk_mq_complete_send_ipi(struct request *rq)
912 struct llist_head *list;
915 cpu = rq->mq_ctx->cpu;
916 list = &per_cpu(blk_cpu_done, cpu);
917 if (llist_add(&rq->ipi_list, list)) {
918 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
919 smp_call_function_single_async(cpu, &rq->csd);
923 static void blk_mq_raise_softirq(struct request *rq)
925 struct llist_head *list;
928 list = this_cpu_ptr(&blk_cpu_done);
929 if (llist_add(&rq->ipi_list, list))
930 raise_softirq(BLOCK_SOFTIRQ);
934 bool blk_mq_complete_request_remote(struct request *rq)
936 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
939 * For a polled request, always complete locallly, it's pointless
940 * to redirect the completion.
942 if (rq->cmd_flags & REQ_POLLED)
945 if (blk_mq_complete_need_ipi(rq)) {
946 blk_mq_complete_send_ipi(rq);
950 if (rq->q->nr_hw_queues == 1) {
951 blk_mq_raise_softirq(rq);
956 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
959 * blk_mq_complete_request - end I/O on a request
960 * @rq: the request being processed
963 * Complete a request by scheduling the ->complete_rq operation.
965 void blk_mq_complete_request(struct request *rq)
967 if (!blk_mq_complete_request_remote(rq))
968 rq->q->mq_ops->complete(rq);
970 EXPORT_SYMBOL(blk_mq_complete_request);
972 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
973 __releases(hctx->srcu)
975 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
978 srcu_read_unlock(hctx->srcu, srcu_idx);
981 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
982 __acquires(hctx->srcu)
984 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
985 /* shut up gcc false positive */
989 *srcu_idx = srcu_read_lock(hctx->srcu);
993 * blk_mq_start_request - Start processing a request
994 * @rq: Pointer to request to be started
996 * Function used by device drivers to notify the block layer that a request
997 * is going to be processed now, so blk layer can do proper initializations
998 * such as starting the timeout timer.
1000 void blk_mq_start_request(struct request *rq)
1002 struct request_queue *q = rq->q;
1004 trace_block_rq_issue(rq);
1006 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1008 #ifdef CONFIG_BLK_CGROUP
1010 start_time = bio_issue_time(&rq->bio->bi_issue);
1013 start_time = ktime_get_ns();
1014 rq->io_start_time_ns = start_time;
1015 rq->stats_sectors = blk_rq_sectors(rq);
1016 rq->rq_flags |= RQF_STATS;
1017 rq_qos_issue(q, rq);
1020 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1023 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1025 #ifdef CONFIG_BLK_DEV_INTEGRITY
1026 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1027 q->integrity.profile->prepare_fn(rq);
1029 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1030 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1032 EXPORT_SYMBOL(blk_mq_start_request);
1034 static void __blk_mq_requeue_request(struct request *rq)
1036 struct request_queue *q = rq->q;
1038 blk_mq_put_driver_tag(rq);
1040 trace_block_rq_requeue(rq);
1041 rq_qos_requeue(q, rq);
1043 if (blk_mq_request_started(rq)) {
1044 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1045 rq->rq_flags &= ~RQF_TIMED_OUT;
1049 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1051 __blk_mq_requeue_request(rq);
1053 /* this request will be re-inserted to io scheduler queue */
1054 blk_mq_sched_requeue_request(rq);
1056 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1058 EXPORT_SYMBOL(blk_mq_requeue_request);
1060 static void blk_mq_requeue_work(struct work_struct *work)
1062 struct request_queue *q =
1063 container_of(work, struct request_queue, requeue_work.work);
1065 struct request *rq, *next;
1067 spin_lock_irq(&q->requeue_lock);
1068 list_splice_init(&q->requeue_list, &rq_list);
1069 spin_unlock_irq(&q->requeue_lock);
1071 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1072 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1075 rq->rq_flags &= ~RQF_SOFTBARRIER;
1076 list_del_init(&rq->queuelist);
1078 * If RQF_DONTPREP, rq has contained some driver specific
1079 * data, so insert it to hctx dispatch list to avoid any
1082 if (rq->rq_flags & RQF_DONTPREP)
1083 blk_mq_request_bypass_insert(rq, false, false);
1085 blk_mq_sched_insert_request(rq, true, false, false);
1088 while (!list_empty(&rq_list)) {
1089 rq = list_entry(rq_list.next, struct request, queuelist);
1090 list_del_init(&rq->queuelist);
1091 blk_mq_sched_insert_request(rq, false, false, false);
1094 blk_mq_run_hw_queues(q, false);
1097 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1098 bool kick_requeue_list)
1100 struct request_queue *q = rq->q;
1101 unsigned long flags;
1104 * We abuse this flag that is otherwise used by the I/O scheduler to
1105 * request head insertion from the workqueue.
1107 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1109 spin_lock_irqsave(&q->requeue_lock, flags);
1111 rq->rq_flags |= RQF_SOFTBARRIER;
1112 list_add(&rq->queuelist, &q->requeue_list);
1114 list_add_tail(&rq->queuelist, &q->requeue_list);
1116 spin_unlock_irqrestore(&q->requeue_lock, flags);
1118 if (kick_requeue_list)
1119 blk_mq_kick_requeue_list(q);
1122 void blk_mq_kick_requeue_list(struct request_queue *q)
1124 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1126 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1128 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1129 unsigned long msecs)
1131 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1132 msecs_to_jiffies(msecs));
1134 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1136 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
1137 void *priv, bool reserved)
1140 * If we find a request that isn't idle and the queue matches,
1141 * we know the queue is busy. Return false to stop the iteration.
1143 if (blk_mq_request_started(rq) && rq->q == hctx->queue) {
1153 bool blk_mq_queue_inflight(struct request_queue *q)
1157 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1160 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1162 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
1164 req->rq_flags |= RQF_TIMED_OUT;
1165 if (req->q->mq_ops->timeout) {
1166 enum blk_eh_timer_return ret;
1168 ret = req->q->mq_ops->timeout(req, reserved);
1169 if (ret == BLK_EH_DONE)
1171 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1177 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
1179 unsigned long deadline;
1181 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1183 if (rq->rq_flags & RQF_TIMED_OUT)
1186 deadline = READ_ONCE(rq->deadline);
1187 if (time_after_eq(jiffies, deadline))
1192 else if (time_after(*next, deadline))
1197 void blk_mq_put_rq_ref(struct request *rq)
1199 if (is_flush_rq(rq))
1201 else if (refcount_dec_and_test(&rq->ref))
1202 __blk_mq_free_request(rq);
1205 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
1206 struct request *rq, void *priv, bool reserved)
1208 unsigned long *next = priv;
1211 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1212 * be reallocated underneath the timeout handler's processing, then
1213 * the expire check is reliable. If the request is not expired, then
1214 * it was completed and reallocated as a new request after returning
1215 * from blk_mq_check_expired().
1217 if (blk_mq_req_expired(rq, next))
1218 blk_mq_rq_timed_out(rq, reserved);
1222 static void blk_mq_timeout_work(struct work_struct *work)
1224 struct request_queue *q =
1225 container_of(work, struct request_queue, timeout_work);
1226 unsigned long next = 0;
1227 struct blk_mq_hw_ctx *hctx;
1230 /* A deadlock might occur if a request is stuck requiring a
1231 * timeout at the same time a queue freeze is waiting
1232 * completion, since the timeout code would not be able to
1233 * acquire the queue reference here.
1235 * That's why we don't use blk_queue_enter here; instead, we use
1236 * percpu_ref_tryget directly, because we need to be able to
1237 * obtain a reference even in the short window between the queue
1238 * starting to freeze, by dropping the first reference in
1239 * blk_freeze_queue_start, and the moment the last request is
1240 * consumed, marked by the instant q_usage_counter reaches
1243 if (!percpu_ref_tryget(&q->q_usage_counter))
1246 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
1249 mod_timer(&q->timeout, next);
1252 * Request timeouts are handled as a forward rolling timer. If
1253 * we end up here it means that no requests are pending and
1254 * also that no request has been pending for a while. Mark
1255 * each hctx as idle.
1257 queue_for_each_hw_ctx(q, hctx, i) {
1258 /* the hctx may be unmapped, so check it here */
1259 if (blk_mq_hw_queue_mapped(hctx))
1260 blk_mq_tag_idle(hctx);
1266 struct flush_busy_ctx_data {
1267 struct blk_mq_hw_ctx *hctx;
1268 struct list_head *list;
1271 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1273 struct flush_busy_ctx_data *flush_data = data;
1274 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1275 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1276 enum hctx_type type = hctx->type;
1278 spin_lock(&ctx->lock);
1279 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1280 sbitmap_clear_bit(sb, bitnr);
1281 spin_unlock(&ctx->lock);
1286 * Process software queues that have been marked busy, splicing them
1287 * to the for-dispatch
1289 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1291 struct flush_busy_ctx_data data = {
1296 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1298 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1300 struct dispatch_rq_data {
1301 struct blk_mq_hw_ctx *hctx;
1305 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1308 struct dispatch_rq_data *dispatch_data = data;
1309 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1310 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1311 enum hctx_type type = hctx->type;
1313 spin_lock(&ctx->lock);
1314 if (!list_empty(&ctx->rq_lists[type])) {
1315 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1316 list_del_init(&dispatch_data->rq->queuelist);
1317 if (list_empty(&ctx->rq_lists[type]))
1318 sbitmap_clear_bit(sb, bitnr);
1320 spin_unlock(&ctx->lock);
1322 return !dispatch_data->rq;
1325 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1326 struct blk_mq_ctx *start)
1328 unsigned off = start ? start->index_hw[hctx->type] : 0;
1329 struct dispatch_rq_data data = {
1334 __sbitmap_for_each_set(&hctx->ctx_map, off,
1335 dispatch_rq_from_ctx, &data);
1340 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1342 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1343 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1346 blk_mq_tag_busy(rq->mq_hctx);
1348 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1349 bt = &rq->mq_hctx->tags->breserved_tags;
1352 if (!hctx_may_queue(rq->mq_hctx, bt))
1356 tag = __sbitmap_queue_get(bt);
1357 if (tag == BLK_MQ_NO_TAG)
1360 rq->tag = tag + tag_offset;
1364 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1366 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1369 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1370 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1371 rq->rq_flags |= RQF_MQ_INFLIGHT;
1372 __blk_mq_inc_active_requests(hctx);
1374 hctx->tags->rqs[rq->tag] = rq;
1378 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1379 int flags, void *key)
1381 struct blk_mq_hw_ctx *hctx;
1383 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1385 spin_lock(&hctx->dispatch_wait_lock);
1386 if (!list_empty(&wait->entry)) {
1387 struct sbitmap_queue *sbq;
1389 list_del_init(&wait->entry);
1390 sbq = &hctx->tags->bitmap_tags;
1391 atomic_dec(&sbq->ws_active);
1393 spin_unlock(&hctx->dispatch_wait_lock);
1395 blk_mq_run_hw_queue(hctx, true);
1400 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1401 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1402 * restart. For both cases, take care to check the condition again after
1403 * marking us as waiting.
1405 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1408 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1409 struct wait_queue_head *wq;
1410 wait_queue_entry_t *wait;
1413 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1414 blk_mq_sched_mark_restart_hctx(hctx);
1417 * It's possible that a tag was freed in the window between the
1418 * allocation failure and adding the hardware queue to the wait
1421 * Don't clear RESTART here, someone else could have set it.
1422 * At most this will cost an extra queue run.
1424 return blk_mq_get_driver_tag(rq);
1427 wait = &hctx->dispatch_wait;
1428 if (!list_empty_careful(&wait->entry))
1431 wq = &bt_wait_ptr(sbq, hctx)->wait;
1433 spin_lock_irq(&wq->lock);
1434 spin_lock(&hctx->dispatch_wait_lock);
1435 if (!list_empty(&wait->entry)) {
1436 spin_unlock(&hctx->dispatch_wait_lock);
1437 spin_unlock_irq(&wq->lock);
1441 atomic_inc(&sbq->ws_active);
1442 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1443 __add_wait_queue(wq, wait);
1446 * It's possible that a tag was freed in the window between the
1447 * allocation failure and adding the hardware queue to the wait
1450 ret = blk_mq_get_driver_tag(rq);
1452 spin_unlock(&hctx->dispatch_wait_lock);
1453 spin_unlock_irq(&wq->lock);
1458 * We got a tag, remove ourselves from the wait queue to ensure
1459 * someone else gets the wakeup.
1461 list_del_init(&wait->entry);
1462 atomic_dec(&sbq->ws_active);
1463 spin_unlock(&hctx->dispatch_wait_lock);
1464 spin_unlock_irq(&wq->lock);
1469 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1470 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1472 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1473 * - EWMA is one simple way to compute running average value
1474 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1475 * - take 4 as factor for avoiding to get too small(0) result, and this
1476 * factor doesn't matter because EWMA decreases exponentially
1478 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1482 ewma = hctx->dispatch_busy;
1487 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1489 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1490 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1492 hctx->dispatch_busy = ewma;
1495 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1497 static void blk_mq_handle_dev_resource(struct request *rq,
1498 struct list_head *list)
1500 struct request *next =
1501 list_first_entry_or_null(list, struct request, queuelist);
1504 * If an I/O scheduler has been configured and we got a driver tag for
1505 * the next request already, free it.
1508 blk_mq_put_driver_tag(next);
1510 list_add(&rq->queuelist, list);
1511 __blk_mq_requeue_request(rq);
1514 static void blk_mq_handle_zone_resource(struct request *rq,
1515 struct list_head *zone_list)
1518 * If we end up here it is because we cannot dispatch a request to a
1519 * specific zone due to LLD level zone-write locking or other zone
1520 * related resource not being available. In this case, set the request
1521 * aside in zone_list for retrying it later.
1523 list_add(&rq->queuelist, zone_list);
1524 __blk_mq_requeue_request(rq);
1527 enum prep_dispatch {
1529 PREP_DISPATCH_NO_TAG,
1530 PREP_DISPATCH_NO_BUDGET,
1533 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1536 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1537 int budget_token = -1;
1540 budget_token = blk_mq_get_dispatch_budget(rq->q);
1541 if (budget_token < 0) {
1542 blk_mq_put_driver_tag(rq);
1543 return PREP_DISPATCH_NO_BUDGET;
1545 blk_mq_set_rq_budget_token(rq, budget_token);
1548 if (!blk_mq_get_driver_tag(rq)) {
1550 * The initial allocation attempt failed, so we need to
1551 * rerun the hardware queue when a tag is freed. The
1552 * waitqueue takes care of that. If the queue is run
1553 * before we add this entry back on the dispatch list,
1554 * we'll re-run it below.
1556 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1558 * All budgets not got from this function will be put
1559 * together during handling partial dispatch
1562 blk_mq_put_dispatch_budget(rq->q, budget_token);
1563 return PREP_DISPATCH_NO_TAG;
1567 return PREP_DISPATCH_OK;
1570 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1571 static void blk_mq_release_budgets(struct request_queue *q,
1572 struct list_head *list)
1576 list_for_each_entry(rq, list, queuelist) {
1577 int budget_token = blk_mq_get_rq_budget_token(rq);
1579 if (budget_token >= 0)
1580 blk_mq_put_dispatch_budget(q, budget_token);
1585 * Returns true if we did some work AND can potentially do more.
1587 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1588 unsigned int nr_budgets)
1590 enum prep_dispatch prep;
1591 struct request_queue *q = hctx->queue;
1592 struct request *rq, *nxt;
1594 blk_status_t ret = BLK_STS_OK;
1595 LIST_HEAD(zone_list);
1597 if (list_empty(list))
1601 * Now process all the entries, sending them to the driver.
1603 errors = queued = 0;
1605 struct blk_mq_queue_data bd;
1607 rq = list_first_entry(list, struct request, queuelist);
1609 WARN_ON_ONCE(hctx != rq->mq_hctx);
1610 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1611 if (prep != PREP_DISPATCH_OK)
1614 list_del_init(&rq->queuelist);
1619 * Flag last if we have no more requests, or if we have more
1620 * but can't assign a driver tag to it.
1622 if (list_empty(list))
1625 nxt = list_first_entry(list, struct request, queuelist);
1626 bd.last = !blk_mq_get_driver_tag(nxt);
1630 * once the request is queued to lld, no need to cover the
1635 ret = q->mq_ops->queue_rq(hctx, &bd);
1640 case BLK_STS_RESOURCE:
1641 case BLK_STS_DEV_RESOURCE:
1642 blk_mq_handle_dev_resource(rq, list);
1644 case BLK_STS_ZONE_RESOURCE:
1646 * Move the request to zone_list and keep going through
1647 * the dispatch list to find more requests the drive can
1650 blk_mq_handle_zone_resource(rq, &zone_list);
1654 blk_mq_end_request(rq, ret);
1656 } while (!list_empty(list));
1658 if (!list_empty(&zone_list))
1659 list_splice_tail_init(&zone_list, list);
1661 /* If we didn't flush the entire list, we could have told the driver
1662 * there was more coming, but that turned out to be a lie.
1664 if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued)
1665 q->mq_ops->commit_rqs(hctx);
1667 * Any items that need requeuing? Stuff them into hctx->dispatch,
1668 * that is where we will continue on next queue run.
1670 if (!list_empty(list)) {
1672 /* For non-shared tags, the RESTART check will suffice */
1673 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1674 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1675 bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET;
1678 blk_mq_release_budgets(q, list);
1680 spin_lock(&hctx->lock);
1681 list_splice_tail_init(list, &hctx->dispatch);
1682 spin_unlock(&hctx->lock);
1685 * Order adding requests to hctx->dispatch and checking
1686 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1687 * in blk_mq_sched_restart(). Avoid restart code path to
1688 * miss the new added requests to hctx->dispatch, meantime
1689 * SCHED_RESTART is observed here.
1694 * If SCHED_RESTART was set by the caller of this function and
1695 * it is no longer set that means that it was cleared by another
1696 * thread and hence that a queue rerun is needed.
1698 * If 'no_tag' is set, that means that we failed getting
1699 * a driver tag with an I/O scheduler attached. If our dispatch
1700 * waitqueue is no longer active, ensure that we run the queue
1701 * AFTER adding our entries back to the list.
1703 * If no I/O scheduler has been configured it is possible that
1704 * the hardware queue got stopped and restarted before requests
1705 * were pushed back onto the dispatch list. Rerun the queue to
1706 * avoid starvation. Notes:
1707 * - blk_mq_run_hw_queue() checks whether or not a queue has
1708 * been stopped before rerunning a queue.
1709 * - Some but not all block drivers stop a queue before
1710 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1713 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1714 * bit is set, run queue after a delay to avoid IO stalls
1715 * that could otherwise occur if the queue is idle. We'll do
1716 * similar if we couldn't get budget and SCHED_RESTART is set.
1718 needs_restart = blk_mq_sched_needs_restart(hctx);
1719 if (!needs_restart ||
1720 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1721 blk_mq_run_hw_queue(hctx, true);
1722 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1724 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1726 blk_mq_update_dispatch_busy(hctx, true);
1729 blk_mq_update_dispatch_busy(hctx, false);
1731 return (queued + errors) != 0;
1735 * __blk_mq_run_hw_queue - Run a hardware queue.
1736 * @hctx: Pointer to the hardware queue to run.
1738 * Send pending requests to the hardware.
1740 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1745 * We can't run the queue inline with ints disabled. Ensure that
1746 * we catch bad users of this early.
1748 WARN_ON_ONCE(in_interrupt());
1750 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1752 hctx_lock(hctx, &srcu_idx);
1753 blk_mq_sched_dispatch_requests(hctx);
1754 hctx_unlock(hctx, srcu_idx);
1757 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1759 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1761 if (cpu >= nr_cpu_ids)
1762 cpu = cpumask_first(hctx->cpumask);
1767 * It'd be great if the workqueue API had a way to pass
1768 * in a mask and had some smarts for more clever placement.
1769 * For now we just round-robin here, switching for every
1770 * BLK_MQ_CPU_WORK_BATCH queued items.
1772 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1775 int next_cpu = hctx->next_cpu;
1777 if (hctx->queue->nr_hw_queues == 1)
1778 return WORK_CPU_UNBOUND;
1780 if (--hctx->next_cpu_batch <= 0) {
1782 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1784 if (next_cpu >= nr_cpu_ids)
1785 next_cpu = blk_mq_first_mapped_cpu(hctx);
1786 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1790 * Do unbound schedule if we can't find a online CPU for this hctx,
1791 * and it should only happen in the path of handling CPU DEAD.
1793 if (!cpu_online(next_cpu)) {
1800 * Make sure to re-select CPU next time once after CPUs
1801 * in hctx->cpumask become online again.
1803 hctx->next_cpu = next_cpu;
1804 hctx->next_cpu_batch = 1;
1805 return WORK_CPU_UNBOUND;
1808 hctx->next_cpu = next_cpu;
1813 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1814 * @hctx: Pointer to the hardware queue to run.
1815 * @async: If we want to run the queue asynchronously.
1816 * @msecs: Milliseconds of delay to wait before running the queue.
1818 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1819 * with a delay of @msecs.
1821 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1822 unsigned long msecs)
1824 if (unlikely(blk_mq_hctx_stopped(hctx)))
1827 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1828 int cpu = get_cpu();
1829 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1830 __blk_mq_run_hw_queue(hctx);
1838 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1839 msecs_to_jiffies(msecs));
1843 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1844 * @hctx: Pointer to the hardware queue to run.
1845 * @msecs: Milliseconds of delay to wait before running the queue.
1847 * Run a hardware queue asynchronously with a delay of @msecs.
1849 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1851 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1853 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1856 * blk_mq_run_hw_queue - Start to run a hardware queue.
1857 * @hctx: Pointer to the hardware queue to run.
1858 * @async: If we want to run the queue asynchronously.
1860 * Check if the request queue is not in a quiesced state and if there are
1861 * pending requests to be sent. If this is true, run the queue to send requests
1864 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1870 * When queue is quiesced, we may be switching io scheduler, or
1871 * updating nr_hw_queues, or other things, and we can't run queue
1872 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1874 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1877 hctx_lock(hctx, &srcu_idx);
1878 need_run = !blk_queue_quiesced(hctx->queue) &&
1879 blk_mq_hctx_has_pending(hctx);
1880 hctx_unlock(hctx, srcu_idx);
1883 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1885 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1888 * Is the request queue handled by an IO scheduler that does not respect
1889 * hardware queues when dispatching?
1891 static bool blk_mq_has_sqsched(struct request_queue *q)
1893 struct elevator_queue *e = q->elevator;
1895 if (e && e->type->ops.dispatch_request &&
1896 !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE))
1902 * Return prefered queue to dispatch from (if any) for non-mq aware IO
1905 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
1907 struct blk_mq_hw_ctx *hctx;
1910 * If the IO scheduler does not respect hardware queues when
1911 * dispatching, we just don't bother with multiple HW queues and
1912 * dispatch from hctx for the current CPU since running multiple queues
1913 * just causes lock contention inside the scheduler and pointless cache
1916 hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT,
1917 raw_smp_processor_id());
1918 if (!blk_mq_hctx_stopped(hctx))
1924 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
1925 * @q: Pointer to the request queue to run.
1926 * @async: If we want to run the queue asynchronously.
1928 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1930 struct blk_mq_hw_ctx *hctx, *sq_hctx;
1934 if (blk_mq_has_sqsched(q))
1935 sq_hctx = blk_mq_get_sq_hctx(q);
1936 queue_for_each_hw_ctx(q, hctx, i) {
1937 if (blk_mq_hctx_stopped(hctx))
1940 * Dispatch from this hctx either if there's no hctx preferred
1941 * by IO scheduler or if it has requests that bypass the
1944 if (!sq_hctx || sq_hctx == hctx ||
1945 !list_empty_careful(&hctx->dispatch))
1946 blk_mq_run_hw_queue(hctx, async);
1949 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1952 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1953 * @q: Pointer to the request queue to run.
1954 * @msecs: Milliseconds of delay to wait before running the queues.
1956 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1958 struct blk_mq_hw_ctx *hctx, *sq_hctx;
1962 if (blk_mq_has_sqsched(q))
1963 sq_hctx = blk_mq_get_sq_hctx(q);
1964 queue_for_each_hw_ctx(q, hctx, i) {
1965 if (blk_mq_hctx_stopped(hctx))
1968 * Dispatch from this hctx either if there's no hctx preferred
1969 * by IO scheduler or if it has requests that bypass the
1972 if (!sq_hctx || sq_hctx == hctx ||
1973 !list_empty_careful(&hctx->dispatch))
1974 blk_mq_delay_run_hw_queue(hctx, msecs);
1977 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
1980 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1981 * @q: request queue.
1983 * The caller is responsible for serializing this function against
1984 * blk_mq_{start,stop}_hw_queue().
1986 bool blk_mq_queue_stopped(struct request_queue *q)
1988 struct blk_mq_hw_ctx *hctx;
1991 queue_for_each_hw_ctx(q, hctx, i)
1992 if (blk_mq_hctx_stopped(hctx))
1997 EXPORT_SYMBOL(blk_mq_queue_stopped);
2000 * This function is often used for pausing .queue_rq() by driver when
2001 * there isn't enough resource or some conditions aren't satisfied, and
2002 * BLK_STS_RESOURCE is usually returned.
2004 * We do not guarantee that dispatch can be drained or blocked
2005 * after blk_mq_stop_hw_queue() returns. Please use
2006 * blk_mq_quiesce_queue() for that requirement.
2008 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2010 cancel_delayed_work(&hctx->run_work);
2012 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2014 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2017 * This function is often used for pausing .queue_rq() by driver when
2018 * there isn't enough resource or some conditions aren't satisfied, and
2019 * BLK_STS_RESOURCE is usually returned.
2021 * We do not guarantee that dispatch can be drained or blocked
2022 * after blk_mq_stop_hw_queues() returns. Please use
2023 * blk_mq_quiesce_queue() for that requirement.
2025 void blk_mq_stop_hw_queues(struct request_queue *q)
2027 struct blk_mq_hw_ctx *hctx;
2030 queue_for_each_hw_ctx(q, hctx, i)
2031 blk_mq_stop_hw_queue(hctx);
2033 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2035 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2037 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2039 blk_mq_run_hw_queue(hctx, false);
2041 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2043 void blk_mq_start_hw_queues(struct request_queue *q)
2045 struct blk_mq_hw_ctx *hctx;
2048 queue_for_each_hw_ctx(q, hctx, i)
2049 blk_mq_start_hw_queue(hctx);
2051 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2053 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2055 if (!blk_mq_hctx_stopped(hctx))
2058 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2059 blk_mq_run_hw_queue(hctx, async);
2061 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2063 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2065 struct blk_mq_hw_ctx *hctx;
2068 queue_for_each_hw_ctx(q, hctx, i)
2069 blk_mq_start_stopped_hw_queue(hctx, async);
2071 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2073 static void blk_mq_run_work_fn(struct work_struct *work)
2075 struct blk_mq_hw_ctx *hctx;
2077 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2080 * If we are stopped, don't run the queue.
2082 if (blk_mq_hctx_stopped(hctx))
2085 __blk_mq_run_hw_queue(hctx);
2088 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2092 struct blk_mq_ctx *ctx = rq->mq_ctx;
2093 enum hctx_type type = hctx->type;
2095 lockdep_assert_held(&ctx->lock);
2097 trace_block_rq_insert(rq);
2100 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2102 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2105 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2108 struct blk_mq_ctx *ctx = rq->mq_ctx;
2110 lockdep_assert_held(&ctx->lock);
2112 __blk_mq_insert_req_list(hctx, rq, at_head);
2113 blk_mq_hctx_mark_pending(hctx, ctx);
2117 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2118 * @rq: Pointer to request to be inserted.
2119 * @at_head: true if the request should be inserted at the head of the list.
2120 * @run_queue: If we should run the hardware queue after inserting the request.
2122 * Should only be used carefully, when the caller knows we want to
2123 * bypass a potential IO scheduler on the target device.
2125 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2128 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2130 spin_lock(&hctx->lock);
2132 list_add(&rq->queuelist, &hctx->dispatch);
2134 list_add_tail(&rq->queuelist, &hctx->dispatch);
2135 spin_unlock(&hctx->lock);
2138 blk_mq_run_hw_queue(hctx, false);
2141 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2142 struct list_head *list)
2146 enum hctx_type type = hctx->type;
2149 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2152 list_for_each_entry(rq, list, queuelist) {
2153 BUG_ON(rq->mq_ctx != ctx);
2154 trace_block_rq_insert(rq);
2157 spin_lock(&ctx->lock);
2158 list_splice_tail_init(list, &ctx->rq_lists[type]);
2159 blk_mq_hctx_mark_pending(hctx, ctx);
2160 spin_unlock(&ctx->lock);
2163 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2166 if (hctx->queue->mq_ops->commit_rqs) {
2167 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2168 hctx->queue->mq_ops->commit_rqs(hctx);
2173 static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2175 struct blk_mq_hw_ctx *hctx = NULL;
2180 while ((rq = rq_list_pop(&plug->mq_list))) {
2181 bool last = rq_list_empty(plug->mq_list);
2184 if (hctx != rq->mq_hctx) {
2186 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2190 ret = blk_mq_request_issue_directly(rq, last);
2195 case BLK_STS_RESOURCE:
2196 case BLK_STS_DEV_RESOURCE:
2197 blk_mq_request_bypass_insert(rq, false, last);
2198 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2201 blk_mq_end_request(rq, ret);
2208 * If we didn't flush the entire list, we could have told the driver
2209 * there was more coming, but that turned out to be a lie.
2212 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2215 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2217 struct blk_mq_hw_ctx *this_hctx;
2218 struct blk_mq_ctx *this_ctx;
2222 if (rq_list_empty(plug->mq_list))
2226 if (!plug->multiple_queues && !plug->has_elevator) {
2227 blk_mq_plug_issue_direct(plug, from_schedule);
2228 if (rq_list_empty(plug->mq_list))
2238 rq = rq_list_pop(&plug->mq_list);
2241 this_hctx = rq->mq_hctx;
2242 this_ctx = rq->mq_ctx;
2243 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2244 trace_block_unplug(this_hctx->queue, depth,
2246 blk_mq_sched_insert_requests(this_hctx, this_ctx,
2247 &list, from_schedule);
2249 this_hctx = rq->mq_hctx;
2250 this_ctx = rq->mq_ctx;
2254 list_add(&rq->queuelist, &list);
2256 } while (!rq_list_empty(plug->mq_list));
2258 if (!list_empty(&list)) {
2259 trace_block_unplug(this_hctx->queue, depth, !from_schedule);
2260 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list,
2265 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2266 unsigned int nr_segs)
2270 if (bio->bi_opf & REQ_RAHEAD)
2271 rq->cmd_flags |= REQ_FAILFAST_MASK;
2273 rq->__sector = bio->bi_iter.bi_sector;
2274 rq->write_hint = bio->bi_write_hint;
2275 blk_rq_bio_prep(rq, bio, nr_segs);
2277 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2278 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2281 blk_account_io_start(rq);
2284 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2285 struct request *rq, bool last)
2287 struct request_queue *q = rq->q;
2288 struct blk_mq_queue_data bd = {
2295 * For OK queue, we are done. For error, caller may kill it.
2296 * Any other error (busy), just add it to our list as we
2297 * previously would have done.
2299 ret = q->mq_ops->queue_rq(hctx, &bd);
2302 blk_mq_update_dispatch_busy(hctx, false);
2304 case BLK_STS_RESOURCE:
2305 case BLK_STS_DEV_RESOURCE:
2306 blk_mq_update_dispatch_busy(hctx, true);
2307 __blk_mq_requeue_request(rq);
2310 blk_mq_update_dispatch_busy(hctx, false);
2317 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2319 bool bypass_insert, bool last)
2321 struct request_queue *q = rq->q;
2322 bool run_queue = true;
2326 * RCU or SRCU read lock is needed before checking quiesced flag.
2328 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2329 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2330 * and avoid driver to try to dispatch again.
2332 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2334 bypass_insert = false;
2338 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2341 budget_token = blk_mq_get_dispatch_budget(q);
2342 if (budget_token < 0)
2345 blk_mq_set_rq_budget_token(rq, budget_token);
2347 if (!blk_mq_get_driver_tag(rq)) {
2348 blk_mq_put_dispatch_budget(q, budget_token);
2352 return __blk_mq_issue_directly(hctx, rq, last);
2355 return BLK_STS_RESOURCE;
2357 blk_mq_sched_insert_request(rq, false, run_queue, false);
2363 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2364 * @hctx: Pointer of the associated hardware queue.
2365 * @rq: Pointer to request to be sent.
2367 * If the device has enough resources to accept a new request now, send the
2368 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2369 * we can try send it another time in the future. Requests inserted at this
2370 * queue have higher priority.
2372 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2378 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2380 hctx_lock(hctx, &srcu_idx);
2382 ret = __blk_mq_try_issue_directly(hctx, rq, false, true);
2383 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2384 blk_mq_request_bypass_insert(rq, false, true);
2385 else if (ret != BLK_STS_OK)
2386 blk_mq_end_request(rq, ret);
2388 hctx_unlock(hctx, srcu_idx);
2391 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2395 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2397 hctx_lock(hctx, &srcu_idx);
2398 ret = __blk_mq_try_issue_directly(hctx, rq, true, last);
2399 hctx_unlock(hctx, srcu_idx);
2404 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2405 struct list_head *list)
2410 while (!list_empty(list)) {
2412 struct request *rq = list_first_entry(list, struct request,
2415 list_del_init(&rq->queuelist);
2416 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2417 if (ret != BLK_STS_OK) {
2418 if (ret == BLK_STS_RESOURCE ||
2419 ret == BLK_STS_DEV_RESOURCE) {
2420 blk_mq_request_bypass_insert(rq, false,
2424 blk_mq_end_request(rq, ret);
2431 * If we didn't flush the entire list, we could have told
2432 * the driver there was more coming, but that turned out to
2435 if ((!list_empty(list) || errors) &&
2436 hctx->queue->mq_ops->commit_rqs && queued)
2437 hctx->queue->mq_ops->commit_rqs(hctx);
2440 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2442 if (!plug->multiple_queues) {
2443 struct request *nxt = rq_list_peek(&plug->mq_list);
2445 if (nxt && nxt->q != rq->q)
2446 plug->multiple_queues = true;
2448 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
2449 plug->has_elevator = true;
2451 rq_list_add(&plug->mq_list, rq);
2456 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
2457 * queues. This is important for md arrays to benefit from merging
2460 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
2462 if (plug->multiple_queues)
2463 return BLK_MAX_REQUEST_COUNT * 2;
2464 return BLK_MAX_REQUEST_COUNT;
2468 * blk_mq_submit_bio - Create and send a request to block device.
2469 * @bio: Bio pointer.
2471 * Builds up a request structure from @q and @bio and send to the device. The
2472 * request may not be queued directly to hardware if:
2473 * * This request can be merged with another one
2474 * * We want to place request at plug queue for possible future merging
2475 * * There is an IO scheduler active at this queue
2477 * It will not queue the request if there is an error with the bio, or at the
2480 void blk_mq_submit_bio(struct bio *bio)
2482 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2483 const int is_sync = op_is_sync(bio->bi_opf);
2485 struct blk_plug *plug;
2486 bool same_queue_rq = false;
2487 unsigned int nr_segs = 1;
2490 blk_queue_bounce(q, &bio);
2491 if (blk_may_split(q, bio))
2492 __blk_queue_split(q, &bio, &nr_segs);
2494 if (!bio_integrity_prep(bio))
2497 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2498 if (blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
2500 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2504 rq_qos_throttle(q, bio);
2506 plug = blk_mq_plug(q, bio);
2507 if (plug && plug->cached_rq) {
2508 rq = rq_list_pop(&plug->cached_rq);
2509 INIT_LIST_HEAD(&rq->queuelist);
2511 struct blk_mq_alloc_data data = {
2514 .cmd_flags = bio->bi_opf,
2518 data.nr_tags = plug->nr_ios;
2520 data.cached_rq = &plug->cached_rq;
2522 rq = __blk_mq_alloc_requests(&data);
2523 if (unlikely(!rq)) {
2524 rq_qos_cleanup(q, bio);
2525 if (bio->bi_opf & REQ_NOWAIT)
2526 bio_wouldblock_error(bio);
2531 trace_block_getrq(bio);
2533 rq_qos_track(q, rq, bio);
2535 blk_mq_bio_to_request(rq, bio, nr_segs);
2537 ret = blk_crypto_init_request(rq);
2538 if (ret != BLK_STS_OK) {
2539 bio->bi_status = ret;
2541 blk_mq_free_request(rq);
2545 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
2548 if (plug && (q->nr_hw_queues == 1 ||
2549 blk_mq_is_shared_tags(rq->mq_hctx->flags) ||
2550 q->mq_ops->commit_rqs || !blk_queue_nonrot(q))) {
2552 * Use plugging if we have a ->commit_rqs() hook as well, as
2553 * we know the driver uses bd->last in a smart fashion.
2555 * Use normal plugging if this disk is slow HDD, as sequential
2556 * IO may benefit a lot from plug merging.
2558 unsigned int request_count = plug->rq_count;
2559 struct request *last = NULL;
2561 if (!request_count) {
2562 trace_block_plug(q);
2563 } else if (!blk_queue_nomerges(q)) {
2564 last = rq_list_peek(&plug->mq_list);
2565 if (blk_rq_bytes(last) < BLK_PLUG_FLUSH_SIZE)
2569 if (request_count >= blk_plug_max_rq_count(plug) || last) {
2570 blk_mq_flush_plug_list(plug, false);
2571 trace_block_plug(q);
2574 blk_add_rq_to_plug(plug, rq);
2575 } else if (rq->rq_flags & RQF_ELV) {
2576 /* Insert the request at the IO scheduler queue */
2577 blk_mq_sched_insert_request(rq, false, true, true);
2578 } else if (plug && !blk_queue_nomerges(q)) {
2579 struct request *next_rq = NULL;
2582 * We do limited plugging. If the bio can be merged, do that.
2583 * Otherwise the existing request in the plug list will be
2584 * issued. So the plug list will have one request at most
2585 * The plug list might get flushed before this. If that happens,
2586 * the plug list is empty, and same_queue_rq is invalid.
2588 if (same_queue_rq) {
2589 next_rq = rq_list_pop(&plug->mq_list);
2592 blk_add_rq_to_plug(plug, rq);
2593 trace_block_plug(q);
2596 trace_block_unplug(q, 1, true);
2597 blk_mq_try_issue_directly(next_rq->mq_hctx, next_rq);
2599 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2600 !rq->mq_hctx->dispatch_busy) {
2602 * There is no scheduler and we can try to send directly
2605 blk_mq_try_issue_directly(rq->mq_hctx, rq);
2608 blk_mq_sched_insert_request(rq, false, true, true);
2616 static size_t order_to_size(unsigned int order)
2618 return (size_t)PAGE_SIZE << order;
2621 /* called before freeing request pool in @tags */
2622 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
2623 struct blk_mq_tags *tags)
2626 unsigned long flags;
2628 /* There is no need to clear a driver tags own mapping */
2629 if (drv_tags == tags)
2632 list_for_each_entry(page, &tags->page_list, lru) {
2633 unsigned long start = (unsigned long)page_address(page);
2634 unsigned long end = start + order_to_size(page->private);
2637 for (i = 0; i < drv_tags->nr_tags; i++) {
2638 struct request *rq = drv_tags->rqs[i];
2639 unsigned long rq_addr = (unsigned long)rq;
2641 if (rq_addr >= start && rq_addr < end) {
2642 WARN_ON_ONCE(refcount_read(&rq->ref) != 0);
2643 cmpxchg(&drv_tags->rqs[i], rq, NULL);
2649 * Wait until all pending iteration is done.
2651 * Request reference is cleared and it is guaranteed to be observed
2652 * after the ->lock is released.
2654 spin_lock_irqsave(&drv_tags->lock, flags);
2655 spin_unlock_irqrestore(&drv_tags->lock, flags);
2658 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2659 unsigned int hctx_idx)
2661 struct blk_mq_tags *drv_tags;
2664 if (blk_mq_is_shared_tags(set->flags))
2665 drv_tags = set->shared_tags;
2667 drv_tags = set->tags[hctx_idx];
2669 if (tags->static_rqs && set->ops->exit_request) {
2672 for (i = 0; i < tags->nr_tags; i++) {
2673 struct request *rq = tags->static_rqs[i];
2677 set->ops->exit_request(set, rq, hctx_idx);
2678 tags->static_rqs[i] = NULL;
2682 blk_mq_clear_rq_mapping(drv_tags, tags);
2684 while (!list_empty(&tags->page_list)) {
2685 page = list_first_entry(&tags->page_list, struct page, lru);
2686 list_del_init(&page->lru);
2688 * Remove kmemleak object previously allocated in
2689 * blk_mq_alloc_rqs().
2691 kmemleak_free(page_address(page));
2692 __free_pages(page, page->private);
2696 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2700 kfree(tags->static_rqs);
2701 tags->static_rqs = NULL;
2703 blk_mq_free_tags(tags);
2706 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2707 unsigned int hctx_idx,
2708 unsigned int nr_tags,
2709 unsigned int reserved_tags)
2711 struct blk_mq_tags *tags;
2714 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2715 if (node == NUMA_NO_NODE)
2716 node = set->numa_node;
2718 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2719 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2723 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2724 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2727 blk_mq_free_tags(tags);
2731 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2732 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2734 if (!tags->static_rqs) {
2736 blk_mq_free_tags(tags);
2743 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2744 unsigned int hctx_idx, int node)
2748 if (set->ops->init_request) {
2749 ret = set->ops->init_request(set, rq, hctx_idx, node);
2754 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2758 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
2759 struct blk_mq_tags *tags,
2760 unsigned int hctx_idx, unsigned int depth)
2762 unsigned int i, j, entries_per_page, max_order = 4;
2763 size_t rq_size, left;
2766 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2767 if (node == NUMA_NO_NODE)
2768 node = set->numa_node;
2770 INIT_LIST_HEAD(&tags->page_list);
2773 * rq_size is the size of the request plus driver payload, rounded
2774 * to the cacheline size
2776 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2778 left = rq_size * depth;
2780 for (i = 0; i < depth; ) {
2781 int this_order = max_order;
2786 while (this_order && left < order_to_size(this_order - 1))
2790 page = alloc_pages_node(node,
2791 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2797 if (order_to_size(this_order) < rq_size)
2804 page->private = this_order;
2805 list_add_tail(&page->lru, &tags->page_list);
2807 p = page_address(page);
2809 * Allow kmemleak to scan these pages as they contain pointers
2810 * to additional allocations like via ops->init_request().
2812 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2813 entries_per_page = order_to_size(this_order) / rq_size;
2814 to_do = min(entries_per_page, depth - i);
2815 left -= to_do * rq_size;
2816 for (j = 0; j < to_do; j++) {
2817 struct request *rq = p;
2819 tags->static_rqs[i] = rq;
2820 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2821 tags->static_rqs[i] = NULL;
2832 blk_mq_free_rqs(set, tags, hctx_idx);
2836 struct rq_iter_data {
2837 struct blk_mq_hw_ctx *hctx;
2841 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2843 struct rq_iter_data *iter_data = data;
2845 if (rq->mq_hctx != iter_data->hctx)
2847 iter_data->has_rq = true;
2851 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2853 struct blk_mq_tags *tags = hctx->sched_tags ?
2854 hctx->sched_tags : hctx->tags;
2855 struct rq_iter_data data = {
2859 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2863 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2864 struct blk_mq_hw_ctx *hctx)
2866 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2868 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2873 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2875 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2876 struct blk_mq_hw_ctx, cpuhp_online);
2878 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2879 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2883 * Prevent new request from being allocated on the current hctx.
2885 * The smp_mb__after_atomic() Pairs with the implied barrier in
2886 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2887 * seen once we return from the tag allocator.
2889 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2890 smp_mb__after_atomic();
2893 * Try to grab a reference to the queue and wait for any outstanding
2894 * requests. If we could not grab a reference the queue has been
2895 * frozen and there are no requests.
2897 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
2898 while (blk_mq_hctx_has_requests(hctx))
2900 percpu_ref_put(&hctx->queue->q_usage_counter);
2906 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
2908 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2909 struct blk_mq_hw_ctx, cpuhp_online);
2911 if (cpumask_test_cpu(cpu, hctx->cpumask))
2912 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2917 * 'cpu' is going away. splice any existing rq_list entries from this
2918 * software queue to the hw queue dispatch list, and ensure that it
2921 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2923 struct blk_mq_hw_ctx *hctx;
2924 struct blk_mq_ctx *ctx;
2926 enum hctx_type type;
2928 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2929 if (!cpumask_test_cpu(cpu, hctx->cpumask))
2932 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2935 spin_lock(&ctx->lock);
2936 if (!list_empty(&ctx->rq_lists[type])) {
2937 list_splice_init(&ctx->rq_lists[type], &tmp);
2938 blk_mq_hctx_clear_pending(hctx, ctx);
2940 spin_unlock(&ctx->lock);
2942 if (list_empty(&tmp))
2945 spin_lock(&hctx->lock);
2946 list_splice_tail_init(&tmp, &hctx->dispatch);
2947 spin_unlock(&hctx->lock);
2949 blk_mq_run_hw_queue(hctx, true);
2953 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2955 if (!(hctx->flags & BLK_MQ_F_STACKING))
2956 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
2957 &hctx->cpuhp_online);
2958 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2963 * Before freeing hw queue, clearing the flush request reference in
2964 * tags->rqs[] for avoiding potential UAF.
2966 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
2967 unsigned int queue_depth, struct request *flush_rq)
2970 unsigned long flags;
2972 /* The hw queue may not be mapped yet */
2976 WARN_ON_ONCE(refcount_read(&flush_rq->ref) != 0);
2978 for (i = 0; i < queue_depth; i++)
2979 cmpxchg(&tags->rqs[i], flush_rq, NULL);
2982 * Wait until all pending iteration is done.
2984 * Request reference is cleared and it is guaranteed to be observed
2985 * after the ->lock is released.
2987 spin_lock_irqsave(&tags->lock, flags);
2988 spin_unlock_irqrestore(&tags->lock, flags);
2991 /* hctx->ctxs will be freed in queue's release handler */
2992 static void blk_mq_exit_hctx(struct request_queue *q,
2993 struct blk_mq_tag_set *set,
2994 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2996 struct request *flush_rq = hctx->fq->flush_rq;
2998 if (blk_mq_hw_queue_mapped(hctx))
2999 blk_mq_tag_idle(hctx);
3001 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3002 set->queue_depth, flush_rq);
3003 if (set->ops->exit_request)
3004 set->ops->exit_request(set, flush_rq, hctx_idx);
3006 if (set->ops->exit_hctx)
3007 set->ops->exit_hctx(hctx, hctx_idx);
3009 blk_mq_remove_cpuhp(hctx);
3011 spin_lock(&q->unused_hctx_lock);
3012 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3013 spin_unlock(&q->unused_hctx_lock);
3016 static void blk_mq_exit_hw_queues(struct request_queue *q,
3017 struct blk_mq_tag_set *set, int nr_queue)
3019 struct blk_mq_hw_ctx *hctx;
3022 queue_for_each_hw_ctx(q, hctx, i) {
3025 blk_mq_debugfs_unregister_hctx(hctx);
3026 blk_mq_exit_hctx(q, set, hctx, i);
3030 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
3032 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
3034 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
3035 __alignof__(struct blk_mq_hw_ctx)) !=
3036 sizeof(struct blk_mq_hw_ctx));
3038 if (tag_set->flags & BLK_MQ_F_BLOCKING)
3039 hw_ctx_size += sizeof(struct srcu_struct);
3044 static int blk_mq_init_hctx(struct request_queue *q,
3045 struct blk_mq_tag_set *set,
3046 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3048 hctx->queue_num = hctx_idx;
3050 if (!(hctx->flags & BLK_MQ_F_STACKING))
3051 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3052 &hctx->cpuhp_online);
3053 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3055 hctx->tags = set->tags[hctx_idx];
3057 if (set->ops->init_hctx &&
3058 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3059 goto unregister_cpu_notifier;
3061 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3067 if (set->ops->exit_hctx)
3068 set->ops->exit_hctx(hctx, hctx_idx);
3069 unregister_cpu_notifier:
3070 blk_mq_remove_cpuhp(hctx);
3074 static struct blk_mq_hw_ctx *
3075 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3078 struct blk_mq_hw_ctx *hctx;
3079 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3081 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
3083 goto fail_alloc_hctx;
3085 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3088 atomic_set(&hctx->nr_active, 0);
3089 if (node == NUMA_NO_NODE)
3090 node = set->numa_node;
3091 hctx->numa_node = node;
3093 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3094 spin_lock_init(&hctx->lock);
3095 INIT_LIST_HEAD(&hctx->dispatch);
3097 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3099 INIT_LIST_HEAD(&hctx->hctx_list);
3102 * Allocate space for all possible cpus to avoid allocation at
3105 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3110 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3111 gfp, node, false, false))
3115 spin_lock_init(&hctx->dispatch_wait_lock);
3116 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3117 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3119 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3123 if (hctx->flags & BLK_MQ_F_BLOCKING)
3124 init_srcu_struct(hctx->srcu);
3125 blk_mq_hctx_kobj_init(hctx);
3130 sbitmap_free(&hctx->ctx_map);
3134 free_cpumask_var(hctx->cpumask);
3141 static void blk_mq_init_cpu_queues(struct request_queue *q,
3142 unsigned int nr_hw_queues)
3144 struct blk_mq_tag_set *set = q->tag_set;
3147 for_each_possible_cpu(i) {
3148 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3149 struct blk_mq_hw_ctx *hctx;
3153 spin_lock_init(&__ctx->lock);
3154 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3155 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3160 * Set local node, IFF we have more than one hw queue. If
3161 * not, we remain on the home node of the device
3163 for (j = 0; j < set->nr_maps; j++) {
3164 hctx = blk_mq_map_queue_type(q, j, i);
3165 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3166 hctx->numa_node = cpu_to_node(i);
3171 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3172 unsigned int hctx_idx,
3175 struct blk_mq_tags *tags;
3178 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3182 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3184 blk_mq_free_rq_map(tags);
3191 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3194 if (blk_mq_is_shared_tags(set->flags)) {
3195 set->tags[hctx_idx] = set->shared_tags;
3200 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3203 return set->tags[hctx_idx];
3206 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3207 struct blk_mq_tags *tags,
3208 unsigned int hctx_idx)
3211 blk_mq_free_rqs(set, tags, hctx_idx);
3212 blk_mq_free_rq_map(tags);
3216 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3217 unsigned int hctx_idx)
3219 if (!blk_mq_is_shared_tags(set->flags))
3220 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3222 set->tags[hctx_idx] = NULL;
3225 static void blk_mq_map_swqueue(struct request_queue *q)
3227 unsigned int i, j, hctx_idx;
3228 struct blk_mq_hw_ctx *hctx;
3229 struct blk_mq_ctx *ctx;
3230 struct blk_mq_tag_set *set = q->tag_set;
3232 queue_for_each_hw_ctx(q, hctx, i) {
3233 cpumask_clear(hctx->cpumask);
3235 hctx->dispatch_from = NULL;
3239 * Map software to hardware queues.
3241 * If the cpu isn't present, the cpu is mapped to first hctx.
3243 for_each_possible_cpu(i) {
3245 ctx = per_cpu_ptr(q->queue_ctx, i);
3246 for (j = 0; j < set->nr_maps; j++) {
3247 if (!set->map[j].nr_queues) {
3248 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3249 HCTX_TYPE_DEFAULT, i);
3252 hctx_idx = set->map[j].mq_map[i];
3253 /* unmapped hw queue can be remapped after CPU topo changed */
3254 if (!set->tags[hctx_idx] &&
3255 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3257 * If tags initialization fail for some hctx,
3258 * that hctx won't be brought online. In this
3259 * case, remap the current ctx to hctx[0] which
3260 * is guaranteed to always have tags allocated
3262 set->map[j].mq_map[i] = 0;
3265 hctx = blk_mq_map_queue_type(q, j, i);
3266 ctx->hctxs[j] = hctx;
3268 * If the CPU is already set in the mask, then we've
3269 * mapped this one already. This can happen if
3270 * devices share queues across queue maps.
3272 if (cpumask_test_cpu(i, hctx->cpumask))
3275 cpumask_set_cpu(i, hctx->cpumask);
3277 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3278 hctx->ctxs[hctx->nr_ctx++] = ctx;
3281 * If the nr_ctx type overflows, we have exceeded the
3282 * amount of sw queues we can support.
3284 BUG_ON(!hctx->nr_ctx);
3287 for (; j < HCTX_MAX_TYPES; j++)
3288 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3289 HCTX_TYPE_DEFAULT, i);
3292 queue_for_each_hw_ctx(q, hctx, i) {
3294 * If no software queues are mapped to this hardware queue,
3295 * disable it and free the request entries.
3297 if (!hctx->nr_ctx) {
3298 /* Never unmap queue 0. We need it as a
3299 * fallback in case of a new remap fails
3303 __blk_mq_free_map_and_rqs(set, i);
3309 hctx->tags = set->tags[i];
3310 WARN_ON(!hctx->tags);
3313 * Set the map size to the number of mapped software queues.
3314 * This is more accurate and more efficient than looping
3315 * over all possibly mapped software queues.
3317 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3320 * Initialize batch roundrobin counts
3322 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3323 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3328 * Caller needs to ensure that we're either frozen/quiesced, or that
3329 * the queue isn't live yet.
3331 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3333 struct blk_mq_hw_ctx *hctx;
3336 queue_for_each_hw_ctx(q, hctx, i) {
3338 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3340 blk_mq_tag_idle(hctx);
3341 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3346 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3349 struct request_queue *q;
3351 lockdep_assert_held(&set->tag_list_lock);
3353 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3354 blk_mq_freeze_queue(q);
3355 queue_set_hctx_shared(q, shared);
3356 blk_mq_unfreeze_queue(q);
3360 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3362 struct blk_mq_tag_set *set = q->tag_set;
3364 mutex_lock(&set->tag_list_lock);
3365 list_del(&q->tag_set_list);
3366 if (list_is_singular(&set->tag_list)) {
3367 /* just transitioned to unshared */
3368 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3369 /* update existing queue */
3370 blk_mq_update_tag_set_shared(set, false);
3372 mutex_unlock(&set->tag_list_lock);
3373 INIT_LIST_HEAD(&q->tag_set_list);
3376 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3377 struct request_queue *q)
3379 mutex_lock(&set->tag_list_lock);
3382 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3384 if (!list_empty(&set->tag_list) &&
3385 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3386 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3387 /* update existing queue */
3388 blk_mq_update_tag_set_shared(set, true);
3390 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3391 queue_set_hctx_shared(q, true);
3392 list_add_tail(&q->tag_set_list, &set->tag_list);
3394 mutex_unlock(&set->tag_list_lock);
3397 /* All allocations will be freed in release handler of q->mq_kobj */
3398 static int blk_mq_alloc_ctxs(struct request_queue *q)
3400 struct blk_mq_ctxs *ctxs;
3403 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3407 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3408 if (!ctxs->queue_ctx)
3411 for_each_possible_cpu(cpu) {
3412 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3416 q->mq_kobj = &ctxs->kobj;
3417 q->queue_ctx = ctxs->queue_ctx;
3426 * It is the actual release handler for mq, but we do it from
3427 * request queue's release handler for avoiding use-after-free
3428 * and headache because q->mq_kobj shouldn't have been introduced,
3429 * but we can't group ctx/kctx kobj without it.
3431 void blk_mq_release(struct request_queue *q)
3433 struct blk_mq_hw_ctx *hctx, *next;
3436 queue_for_each_hw_ctx(q, hctx, i)
3437 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3439 /* all hctx are in .unused_hctx_list now */
3440 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3441 list_del_init(&hctx->hctx_list);
3442 kobject_put(&hctx->kobj);
3445 kfree(q->queue_hw_ctx);
3448 * release .mq_kobj and sw queue's kobject now because
3449 * both share lifetime with request queue.
3451 blk_mq_sysfs_deinit(q);
3454 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3457 struct request_queue *q;
3460 q = blk_alloc_queue(set->numa_node);
3462 return ERR_PTR(-ENOMEM);
3463 q->queuedata = queuedata;
3464 ret = blk_mq_init_allocated_queue(set, q);
3466 blk_cleanup_queue(q);
3467 return ERR_PTR(ret);
3472 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3474 return blk_mq_init_queue_data(set, NULL);
3476 EXPORT_SYMBOL(blk_mq_init_queue);
3478 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
3479 struct lock_class_key *lkclass)
3481 struct request_queue *q;
3482 struct gendisk *disk;
3484 q = blk_mq_init_queue_data(set, queuedata);
3488 disk = __alloc_disk_node(q, set->numa_node, lkclass);
3490 blk_cleanup_queue(q);
3491 return ERR_PTR(-ENOMEM);
3495 EXPORT_SYMBOL(__blk_mq_alloc_disk);
3497 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3498 struct blk_mq_tag_set *set, struct request_queue *q,
3499 int hctx_idx, int node)
3501 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3503 /* reuse dead hctx first */
3504 spin_lock(&q->unused_hctx_lock);
3505 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3506 if (tmp->numa_node == node) {
3512 list_del_init(&hctx->hctx_list);
3513 spin_unlock(&q->unused_hctx_lock);
3516 hctx = blk_mq_alloc_hctx(q, set, node);
3520 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3526 kobject_put(&hctx->kobj);
3531 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3532 struct request_queue *q)
3535 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3537 if (q->nr_hw_queues < set->nr_hw_queues) {
3538 struct blk_mq_hw_ctx **new_hctxs;
3540 new_hctxs = kcalloc_node(set->nr_hw_queues,
3541 sizeof(*new_hctxs), GFP_KERNEL,
3546 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3548 q->queue_hw_ctx = new_hctxs;
3553 /* protect against switching io scheduler */
3554 mutex_lock(&q->sysfs_lock);
3555 for (i = 0; i < set->nr_hw_queues; i++) {
3557 struct blk_mq_hw_ctx *hctx;
3559 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3561 * If the hw queue has been mapped to another numa node,
3562 * we need to realloc the hctx. If allocation fails, fallback
3563 * to use the previous one.
3565 if (hctxs[i] && (hctxs[i]->numa_node == node))
3568 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3571 blk_mq_exit_hctx(q, set, hctxs[i], i);
3575 pr_warn("Allocate new hctx on node %d fails,\
3576 fallback to previous one on node %d\n",
3577 node, hctxs[i]->numa_node);
3583 * Increasing nr_hw_queues fails. Free the newly allocated
3584 * hctxs and keep the previous q->nr_hw_queues.
3586 if (i != set->nr_hw_queues) {
3587 j = q->nr_hw_queues;
3591 end = q->nr_hw_queues;
3592 q->nr_hw_queues = set->nr_hw_queues;
3595 for (; j < end; j++) {
3596 struct blk_mq_hw_ctx *hctx = hctxs[j];
3599 __blk_mq_free_map_and_rqs(set, j);
3600 blk_mq_exit_hctx(q, set, hctx, j);
3604 mutex_unlock(&q->sysfs_lock);
3607 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3608 struct request_queue *q)
3610 /* mark the queue as mq asap */
3611 q->mq_ops = set->ops;
3613 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3614 blk_mq_poll_stats_bkt,
3615 BLK_MQ_POLL_STATS_BKTS, q);
3619 if (blk_mq_alloc_ctxs(q))
3622 /* init q->mq_kobj and sw queues' kobjects */
3623 blk_mq_sysfs_init(q);
3625 INIT_LIST_HEAD(&q->unused_hctx_list);
3626 spin_lock_init(&q->unused_hctx_lock);
3628 blk_mq_realloc_hw_ctxs(set, q);
3629 if (!q->nr_hw_queues)
3632 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3633 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3637 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3638 if (set->nr_maps > HCTX_TYPE_POLL &&
3639 set->map[HCTX_TYPE_POLL].nr_queues)
3640 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3642 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3643 INIT_LIST_HEAD(&q->requeue_list);
3644 spin_lock_init(&q->requeue_lock);
3646 q->nr_requests = set->queue_depth;
3649 * Default to classic polling
3651 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3653 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3654 blk_mq_add_queue_tag_set(set, q);
3655 blk_mq_map_swqueue(q);
3659 kfree(q->queue_hw_ctx);
3660 q->nr_hw_queues = 0;
3661 blk_mq_sysfs_deinit(q);
3663 blk_stat_free_callback(q->poll_cb);
3669 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3671 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3672 void blk_mq_exit_queue(struct request_queue *q)
3674 struct blk_mq_tag_set *set = q->tag_set;
3676 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
3677 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3678 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
3679 blk_mq_del_queue_tag_set(q);
3682 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3686 if (blk_mq_is_shared_tags(set->flags)) {
3687 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
3690 if (!set->shared_tags)
3694 for (i = 0; i < set->nr_hw_queues; i++) {
3695 if (!__blk_mq_alloc_map_and_rqs(set, i))
3704 __blk_mq_free_map_and_rqs(set, i);
3706 if (blk_mq_is_shared_tags(set->flags)) {
3707 blk_mq_free_map_and_rqs(set, set->shared_tags,
3708 BLK_MQ_NO_HCTX_IDX);
3715 * Allocate the request maps associated with this tag_set. Note that this
3716 * may reduce the depth asked for, if memory is tight. set->queue_depth
3717 * will be updated to reflect the allocated depth.
3719 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
3724 depth = set->queue_depth;
3726 err = __blk_mq_alloc_rq_maps(set);
3730 set->queue_depth >>= 1;
3731 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3735 } while (set->queue_depth);
3737 if (!set->queue_depth || err) {
3738 pr_err("blk-mq: failed to allocate request map\n");
3742 if (depth != set->queue_depth)
3743 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3744 depth, set->queue_depth);
3749 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3752 * blk_mq_map_queues() and multiple .map_queues() implementations
3753 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3754 * number of hardware queues.
3756 if (set->nr_maps == 1)
3757 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3759 if (set->ops->map_queues && !is_kdump_kernel()) {
3763 * transport .map_queues is usually done in the following
3766 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3767 * mask = get_cpu_mask(queue)
3768 * for_each_cpu(cpu, mask)
3769 * set->map[x].mq_map[cpu] = queue;
3772 * When we need to remap, the table has to be cleared for
3773 * killing stale mapping since one CPU may not be mapped
3776 for (i = 0; i < set->nr_maps; i++)
3777 blk_mq_clear_mq_map(&set->map[i]);
3779 return set->ops->map_queues(set);
3781 BUG_ON(set->nr_maps > 1);
3782 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3786 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3787 int cur_nr_hw_queues, int new_nr_hw_queues)
3789 struct blk_mq_tags **new_tags;
3791 if (cur_nr_hw_queues >= new_nr_hw_queues)
3794 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3795 GFP_KERNEL, set->numa_node);
3800 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3801 sizeof(*set->tags));
3803 set->tags = new_tags;
3804 set->nr_hw_queues = new_nr_hw_queues;
3809 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
3810 int new_nr_hw_queues)
3812 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
3816 * Alloc a tag set to be associated with one or more request queues.
3817 * May fail with EINVAL for various error conditions. May adjust the
3818 * requested depth down, if it's too large. In that case, the set
3819 * value will be stored in set->queue_depth.
3821 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3825 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3827 if (!set->nr_hw_queues)
3829 if (!set->queue_depth)
3831 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3834 if (!set->ops->queue_rq)
3837 if (!set->ops->get_budget ^ !set->ops->put_budget)
3840 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3841 pr_info("blk-mq: reduced tag depth to %u\n",
3843 set->queue_depth = BLK_MQ_MAX_DEPTH;
3848 else if (set->nr_maps > HCTX_MAX_TYPES)
3852 * If a crashdump is active, then we are potentially in a very
3853 * memory constrained environment. Limit us to 1 queue and
3854 * 64 tags to prevent using too much memory.
3856 if (is_kdump_kernel()) {
3857 set->nr_hw_queues = 1;
3859 set->queue_depth = min(64U, set->queue_depth);
3862 * There is no use for more h/w queues than cpus if we just have
3865 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3866 set->nr_hw_queues = nr_cpu_ids;
3868 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
3872 for (i = 0; i < set->nr_maps; i++) {
3873 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3874 sizeof(set->map[i].mq_map[0]),
3875 GFP_KERNEL, set->numa_node);
3876 if (!set->map[i].mq_map)
3877 goto out_free_mq_map;
3878 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3881 ret = blk_mq_update_queue_map(set);
3883 goto out_free_mq_map;
3885 ret = blk_mq_alloc_set_map_and_rqs(set);
3887 goto out_free_mq_map;
3889 mutex_init(&set->tag_list_lock);
3890 INIT_LIST_HEAD(&set->tag_list);
3895 for (i = 0; i < set->nr_maps; i++) {
3896 kfree(set->map[i].mq_map);
3897 set->map[i].mq_map = NULL;
3903 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3905 /* allocate and initialize a tagset for a simple single-queue device */
3906 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
3907 const struct blk_mq_ops *ops, unsigned int queue_depth,
3908 unsigned int set_flags)
3910 memset(set, 0, sizeof(*set));
3912 set->nr_hw_queues = 1;
3914 set->queue_depth = queue_depth;
3915 set->numa_node = NUMA_NO_NODE;
3916 set->flags = set_flags;
3917 return blk_mq_alloc_tag_set(set);
3919 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
3921 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3925 for (i = 0; i < set->nr_hw_queues; i++)
3926 __blk_mq_free_map_and_rqs(set, i);
3928 if (blk_mq_is_shared_tags(set->flags)) {
3929 blk_mq_free_map_and_rqs(set, set->shared_tags,
3930 BLK_MQ_NO_HCTX_IDX);
3933 for (j = 0; j < set->nr_maps; j++) {
3934 kfree(set->map[j].mq_map);
3935 set->map[j].mq_map = NULL;
3941 EXPORT_SYMBOL(blk_mq_free_tag_set);
3943 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3945 struct blk_mq_tag_set *set = q->tag_set;
3946 struct blk_mq_hw_ctx *hctx;
3952 if (q->nr_requests == nr)
3955 blk_mq_freeze_queue(q);
3956 blk_mq_quiesce_queue(q);
3959 queue_for_each_hw_ctx(q, hctx, i) {
3963 * If we're using an MQ scheduler, just update the scheduler
3964 * queue depth. This is similar to what the old code would do.
3966 if (hctx->sched_tags) {
3967 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3970 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3975 if (q->elevator && q->elevator->type->ops.depth_updated)
3976 q->elevator->type->ops.depth_updated(hctx);
3979 q->nr_requests = nr;
3980 if (blk_mq_is_shared_tags(set->flags)) {
3982 blk_mq_tag_update_sched_shared_tags(q);
3984 blk_mq_tag_resize_shared_tags(set, nr);
3988 blk_mq_unquiesce_queue(q);
3989 blk_mq_unfreeze_queue(q);
3995 * request_queue and elevator_type pair.
3996 * It is just used by __blk_mq_update_nr_hw_queues to cache
3997 * the elevator_type associated with a request_queue.
3999 struct blk_mq_qe_pair {
4000 struct list_head node;
4001 struct request_queue *q;
4002 struct elevator_type *type;
4006 * Cache the elevator_type in qe pair list and switch the
4007 * io scheduler to 'none'
4009 static bool blk_mq_elv_switch_none(struct list_head *head,
4010 struct request_queue *q)
4012 struct blk_mq_qe_pair *qe;
4017 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4021 INIT_LIST_HEAD(&qe->node);
4023 qe->type = q->elevator->type;
4024 list_add(&qe->node, head);
4026 mutex_lock(&q->sysfs_lock);
4028 * After elevator_switch_mq, the previous elevator_queue will be
4029 * released by elevator_release. The reference of the io scheduler
4030 * module get by elevator_get will also be put. So we need to get
4031 * a reference of the io scheduler module here to prevent it to be
4034 __module_get(qe->type->elevator_owner);
4035 elevator_switch_mq(q, NULL);
4036 mutex_unlock(&q->sysfs_lock);
4041 static void blk_mq_elv_switch_back(struct list_head *head,
4042 struct request_queue *q)
4044 struct blk_mq_qe_pair *qe;
4045 struct elevator_type *t = NULL;
4047 list_for_each_entry(qe, head, node)
4056 list_del(&qe->node);
4059 mutex_lock(&q->sysfs_lock);
4060 elevator_switch_mq(q, t);
4061 mutex_unlock(&q->sysfs_lock);
4064 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4067 struct request_queue *q;
4069 int prev_nr_hw_queues;
4071 lockdep_assert_held(&set->tag_list_lock);
4073 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4074 nr_hw_queues = nr_cpu_ids;
4075 if (nr_hw_queues < 1)
4077 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4080 list_for_each_entry(q, &set->tag_list, tag_set_list)
4081 blk_mq_freeze_queue(q);
4083 * Switch IO scheduler to 'none', cleaning up the data associated
4084 * with the previous scheduler. We will switch back once we are done
4085 * updating the new sw to hw queue mappings.
4087 list_for_each_entry(q, &set->tag_list, tag_set_list)
4088 if (!blk_mq_elv_switch_none(&head, q))
4091 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4092 blk_mq_debugfs_unregister_hctxs(q);
4093 blk_mq_sysfs_unregister(q);
4096 prev_nr_hw_queues = set->nr_hw_queues;
4097 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
4101 set->nr_hw_queues = nr_hw_queues;
4103 blk_mq_update_queue_map(set);
4104 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4105 blk_mq_realloc_hw_ctxs(set, q);
4106 if (q->nr_hw_queues != set->nr_hw_queues) {
4107 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4108 nr_hw_queues, prev_nr_hw_queues);
4109 set->nr_hw_queues = prev_nr_hw_queues;
4110 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4113 blk_mq_map_swqueue(q);
4117 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4118 blk_mq_sysfs_register(q);
4119 blk_mq_debugfs_register_hctxs(q);
4123 list_for_each_entry(q, &set->tag_list, tag_set_list)
4124 blk_mq_elv_switch_back(&head, q);
4126 list_for_each_entry(q, &set->tag_list, tag_set_list)
4127 blk_mq_unfreeze_queue(q);
4130 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4132 mutex_lock(&set->tag_list_lock);
4133 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4134 mutex_unlock(&set->tag_list_lock);
4136 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4138 /* Enable polling stats and return whether they were already enabled. */
4139 static bool blk_poll_stats_enable(struct request_queue *q)
4141 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
4142 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
4144 blk_stat_add_callback(q, q->poll_cb);
4148 static void blk_mq_poll_stats_start(struct request_queue *q)
4151 * We don't arm the callback if polling stats are not enabled or the
4152 * callback is already active.
4154 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
4155 blk_stat_is_active(q->poll_cb))
4158 blk_stat_activate_msecs(q->poll_cb, 100);
4161 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4163 struct request_queue *q = cb->data;
4166 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4167 if (cb->stat[bucket].nr_samples)
4168 q->poll_stat[bucket] = cb->stat[bucket];
4172 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4175 unsigned long ret = 0;
4179 * If stats collection isn't on, don't sleep but turn it on for
4182 if (!blk_poll_stats_enable(q))
4186 * As an optimistic guess, use half of the mean service time
4187 * for this type of request. We can (and should) make this smarter.
4188 * For instance, if the completion latencies are tight, we can
4189 * get closer than just half the mean. This is especially
4190 * important on devices where the completion latencies are longer
4191 * than ~10 usec. We do use the stats for the relevant IO size
4192 * if available which does lead to better estimates.
4194 bucket = blk_mq_poll_stats_bkt(rq);
4198 if (q->poll_stat[bucket].nr_samples)
4199 ret = (q->poll_stat[bucket].mean + 1) / 2;
4204 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4206 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4207 struct request *rq = blk_qc_to_rq(hctx, qc);
4208 struct hrtimer_sleeper hs;
4209 enum hrtimer_mode mode;
4214 * If a request has completed on queue that uses an I/O scheduler, we
4215 * won't get back a request from blk_qc_to_rq.
4217 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4221 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4223 * 0: use half of prev avg
4224 * >0: use this specific value
4226 if (q->poll_nsec > 0)
4227 nsecs = q->poll_nsec;
4229 nsecs = blk_mq_poll_nsecs(q, rq);
4234 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4237 * This will be replaced with the stats tracking code, using
4238 * 'avg_completion_time / 2' as the pre-sleep target.
4242 mode = HRTIMER_MODE_REL;
4243 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4244 hrtimer_set_expires(&hs.timer, kt);
4247 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4249 set_current_state(TASK_UNINTERRUPTIBLE);
4250 hrtimer_sleeper_start_expires(&hs, mode);
4253 hrtimer_cancel(&hs.timer);
4254 mode = HRTIMER_MODE_ABS;
4255 } while (hs.task && !signal_pending(current));
4257 __set_current_state(TASK_RUNNING);
4258 destroy_hrtimer_on_stack(&hs.timer);
4261 * If we sleep, have the caller restart the poll loop to reset the
4262 * state. Like for the other success return cases, the caller is
4263 * responsible for checking if the IO completed. If the IO isn't
4264 * complete, we'll get called again and will go straight to the busy
4270 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4271 struct io_comp_batch *iob, unsigned int flags)
4273 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4274 long state = get_current_state();
4278 ret = q->mq_ops->poll(hctx, iob);
4280 __set_current_state(TASK_RUNNING);
4284 if (signal_pending_state(state, current))
4285 __set_current_state(TASK_RUNNING);
4286 if (task_is_running(current))
4289 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4292 } while (!need_resched());
4294 __set_current_state(TASK_RUNNING);
4298 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4301 if (!(flags & BLK_POLL_NOSLEEP) &&
4302 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4303 if (blk_mq_poll_hybrid(q, cookie))
4306 return blk_mq_poll_classic(q, cookie, iob, flags);
4309 unsigned int blk_mq_rq_cpu(struct request *rq)
4311 return rq->mq_ctx->cpu;
4313 EXPORT_SYMBOL(blk_mq_rq_cpu);
4315 static int __init blk_mq_init(void)
4319 for_each_possible_cpu(i)
4320 init_llist_head(&per_cpu(blk_cpu_done, i));
4321 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4323 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4324 "block/softirq:dead", NULL,
4325 blk_softirq_cpu_dead);
4326 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4327 blk_mq_hctx_notify_dead);
4328 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4329 blk_mq_hctx_notify_online,
4330 blk_mq_hctx_notify_offline);
4333 subsys_initcall(blk_mq_init);