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_wait_quiesce_done() - wait until in-progress quiesce is done
257 * Note: it is driver's responsibility for making sure that quiesce has
260 void blk_mq_wait_quiesce_done(struct request_queue *q)
262 struct blk_mq_hw_ctx *hctx;
266 queue_for_each_hw_ctx(q, hctx, i) {
267 if (hctx->flags & BLK_MQ_F_BLOCKING)
268 synchronize_srcu(hctx->srcu);
275 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
278 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
281 * Note: this function does not prevent that the struct request end_io()
282 * callback function is invoked. Once this function is returned, we make
283 * sure no dispatch can happen until the queue is unquiesced via
284 * blk_mq_unquiesce_queue().
286 void blk_mq_quiesce_queue(struct request_queue *q)
288 blk_mq_quiesce_queue_nowait(q);
289 blk_mq_wait_quiesce_done(q);
291 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
294 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
297 * This function recovers queue into the state before quiescing
298 * which is done by blk_mq_quiesce_queue.
300 void blk_mq_unquiesce_queue(struct request_queue *q)
303 bool run_queue = false;
305 spin_lock_irqsave(&q->queue_lock, flags);
306 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
308 } else if (!--q->quiesce_depth) {
309 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
312 spin_unlock_irqrestore(&q->queue_lock, flags);
314 /* dispatch requests which are inserted during quiescing */
316 blk_mq_run_hw_queues(q, true);
318 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
320 void blk_mq_wake_waiters(struct request_queue *q)
322 struct blk_mq_hw_ctx *hctx;
325 queue_for_each_hw_ctx(q, hctx, i)
326 if (blk_mq_hw_queue_mapped(hctx))
327 blk_mq_tag_wakeup_all(hctx->tags, true);
330 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
331 struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns)
333 struct blk_mq_ctx *ctx = data->ctx;
334 struct blk_mq_hw_ctx *hctx = data->hctx;
335 struct request_queue *q = data->q;
336 struct request *rq = tags->static_rqs[tag];
341 rq->cmd_flags = data->cmd_flags;
343 if (data->flags & BLK_MQ_REQ_PM)
344 data->rq_flags |= RQF_PM;
345 if (blk_queue_io_stat(q))
346 data->rq_flags |= RQF_IO_STAT;
347 rq->rq_flags = data->rq_flags;
349 if (!(data->rq_flags & RQF_ELV)) {
351 rq->internal_tag = BLK_MQ_NO_TAG;
353 rq->tag = BLK_MQ_NO_TAG;
354 rq->internal_tag = tag;
358 if (blk_mq_need_time_stamp(rq))
359 rq->start_time_ns = ktime_get_ns();
361 rq->start_time_ns = 0;
364 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
365 rq->alloc_time_ns = alloc_time_ns;
367 rq->io_start_time_ns = 0;
368 rq->stats_sectors = 0;
369 rq->nr_phys_segments = 0;
370 #if defined(CONFIG_BLK_DEV_INTEGRITY)
371 rq->nr_integrity_segments = 0;
374 rq->end_io_data = NULL;
376 blk_crypto_rq_set_defaults(rq);
377 INIT_LIST_HEAD(&rq->queuelist);
378 /* tag was already set */
379 WRITE_ONCE(rq->deadline, 0);
380 refcount_set(&rq->ref, 1);
382 if (rq->rq_flags & RQF_ELV) {
383 struct elevator_queue *e = data->q->elevator;
386 INIT_HLIST_NODE(&rq->hash);
387 RB_CLEAR_NODE(&rq->rb_node);
389 if (!op_is_flush(data->cmd_flags) &&
390 e->type->ops.prepare_request) {
391 if (e->type->icq_cache)
392 blk_mq_sched_assign_ioc(rq);
394 e->type->ops.prepare_request(rq);
395 rq->rq_flags |= RQF_ELVPRIV;
402 static inline struct request *
403 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data,
406 unsigned int tag, tag_offset;
407 struct blk_mq_tags *tags;
409 unsigned long tag_mask;
412 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
413 if (unlikely(!tag_mask))
416 tags = blk_mq_tags_from_data(data);
417 for (i = 0; tag_mask; i++) {
418 if (!(tag_mask & (1UL << i)))
420 tag = tag_offset + i;
421 prefetch(tags->static_rqs[tag]);
422 tag_mask &= ~(1UL << i);
423 rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns);
424 rq_list_add(data->cached_rq, rq);
427 /* caller already holds a reference, add for remainder */
428 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
431 return rq_list_pop(data->cached_rq);
434 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
436 struct request_queue *q = data->q;
437 u64 alloc_time_ns = 0;
441 /* alloc_time includes depth and tag waits */
442 if (blk_queue_rq_alloc_time(q))
443 alloc_time_ns = ktime_get_ns();
445 if (data->cmd_flags & REQ_NOWAIT)
446 data->flags |= BLK_MQ_REQ_NOWAIT;
449 struct elevator_queue *e = q->elevator;
451 data->rq_flags |= RQF_ELV;
454 * Flush/passthrough requests are special and go directly to the
455 * dispatch list. Don't include reserved tags in the
456 * limiting, as it isn't useful.
458 if (!op_is_flush(data->cmd_flags) &&
459 !blk_op_is_passthrough(data->cmd_flags) &&
460 e->type->ops.limit_depth &&
461 !(data->flags & BLK_MQ_REQ_RESERVED))
462 e->type->ops.limit_depth(data->cmd_flags, data);
466 data->ctx = blk_mq_get_ctx(q);
467 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
468 if (!(data->rq_flags & RQF_ELV))
469 blk_mq_tag_busy(data->hctx);
472 * Try batched alloc if we want more than 1 tag.
474 if (data->nr_tags > 1) {
475 rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns);
482 * Waiting allocations only fail because of an inactive hctx. In that
483 * case just retry the hctx assignment and tag allocation as CPU hotplug
484 * should have migrated us to an online CPU by now.
486 tag = blk_mq_get_tag(data);
487 if (tag == BLK_MQ_NO_TAG) {
488 if (data->flags & BLK_MQ_REQ_NOWAIT)
491 * Give up the CPU and sleep for a random short time to
492 * ensure that thread using a realtime scheduling class
493 * are migrated off the CPU, and thus off the hctx that
500 return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag,
504 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
505 blk_mq_req_flags_t flags)
507 struct blk_mq_alloc_data data = {
516 ret = blk_queue_enter(q, flags);
520 rq = __blk_mq_alloc_requests(&data);
524 rq->__sector = (sector_t) -1;
525 rq->bio = rq->biotail = NULL;
529 return ERR_PTR(-EWOULDBLOCK);
531 EXPORT_SYMBOL(blk_mq_alloc_request);
533 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
534 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
536 struct blk_mq_alloc_data data = {
542 u64 alloc_time_ns = 0;
547 /* alloc_time includes depth and tag waits */
548 if (blk_queue_rq_alloc_time(q))
549 alloc_time_ns = ktime_get_ns();
552 * If the tag allocator sleeps we could get an allocation for a
553 * different hardware context. No need to complicate the low level
554 * allocator for this for the rare use case of a command tied to
557 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
558 return ERR_PTR(-EINVAL);
560 if (hctx_idx >= q->nr_hw_queues)
561 return ERR_PTR(-EIO);
563 ret = blk_queue_enter(q, flags);
568 * Check if the hardware context is actually mapped to anything.
569 * If not tell the caller that it should skip this queue.
572 data.hctx = q->queue_hw_ctx[hctx_idx];
573 if (!blk_mq_hw_queue_mapped(data.hctx))
575 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
576 data.ctx = __blk_mq_get_ctx(q, cpu);
579 blk_mq_tag_busy(data.hctx);
581 data.rq_flags |= RQF_ELV;
584 tag = blk_mq_get_tag(&data);
585 if (tag == BLK_MQ_NO_TAG)
587 return blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag,
594 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
596 static void __blk_mq_free_request(struct request *rq)
598 struct request_queue *q = rq->q;
599 struct blk_mq_ctx *ctx = rq->mq_ctx;
600 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
601 const int sched_tag = rq->internal_tag;
603 blk_crypto_free_request(rq);
604 blk_pm_mark_last_busy(rq);
606 if (rq->tag != BLK_MQ_NO_TAG)
607 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
608 if (sched_tag != BLK_MQ_NO_TAG)
609 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
610 blk_mq_sched_restart(hctx);
614 void blk_mq_free_request(struct request *rq)
616 struct request_queue *q = rq->q;
617 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
619 if (rq->rq_flags & RQF_ELVPRIV) {
620 struct elevator_queue *e = q->elevator;
622 if (e->type->ops.finish_request)
623 e->type->ops.finish_request(rq);
625 put_io_context(rq->elv.icq->ioc);
630 if (rq->rq_flags & RQF_MQ_INFLIGHT)
631 __blk_mq_dec_active_requests(hctx);
633 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
634 laptop_io_completion(q->disk->bdi);
638 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
639 if (refcount_dec_and_test(&rq->ref))
640 __blk_mq_free_request(rq);
642 EXPORT_SYMBOL_GPL(blk_mq_free_request);
644 void blk_mq_free_plug_rqs(struct blk_plug *plug)
648 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
649 blk_mq_free_request(rq);
652 static void req_bio_endio(struct request *rq, struct bio *bio,
653 unsigned int nbytes, blk_status_t error)
655 if (unlikely(error)) {
656 bio->bi_status = error;
657 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
659 * Partial zone append completions cannot be supported as the
660 * BIO fragments may end up not being written sequentially.
662 if (bio->bi_iter.bi_size != nbytes)
663 bio->bi_status = BLK_STS_IOERR;
665 bio->bi_iter.bi_sector = rq->__sector;
668 bio_advance(bio, nbytes);
670 if (unlikely(rq->rq_flags & RQF_QUIET))
671 bio_set_flag(bio, BIO_QUIET);
672 /* don't actually finish bio if it's part of flush sequence */
673 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
677 static void blk_account_io_completion(struct request *req, unsigned int bytes)
679 if (req->part && blk_do_io_stat(req)) {
680 const int sgrp = op_stat_group(req_op(req));
683 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
689 * blk_update_request - Complete multiple bytes without completing the request
690 * @req: the request being processed
691 * @error: block status code
692 * @nr_bytes: number of bytes to complete for @req
695 * Ends I/O on a number of bytes attached to @req, but doesn't complete
696 * the request structure even if @req doesn't have leftover.
697 * If @req has leftover, sets it up for the next range of segments.
699 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
700 * %false return from this function.
703 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
704 * except in the consistency check at the end of this function.
707 * %false - this request doesn't have any more data
708 * %true - this request has more data
710 bool blk_update_request(struct request *req, blk_status_t error,
711 unsigned int nr_bytes)
715 trace_block_rq_complete(req, error, nr_bytes);
720 #ifdef CONFIG_BLK_DEV_INTEGRITY
721 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
723 req->q->integrity.profile->complete_fn(req, nr_bytes);
726 if (unlikely(error && !blk_rq_is_passthrough(req) &&
727 !(req->rq_flags & RQF_QUIET)))
728 blk_print_req_error(req, error);
730 blk_account_io_completion(req, nr_bytes);
734 struct bio *bio = req->bio;
735 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
737 if (bio_bytes == bio->bi_iter.bi_size)
738 req->bio = bio->bi_next;
740 /* Completion has already been traced */
741 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
742 req_bio_endio(req, bio, bio_bytes, error);
744 total_bytes += bio_bytes;
745 nr_bytes -= bio_bytes;
756 * Reset counters so that the request stacking driver
757 * can find how many bytes remain in the request
764 req->__data_len -= total_bytes;
766 /* update sector only for requests with clear definition of sector */
767 if (!blk_rq_is_passthrough(req))
768 req->__sector += total_bytes >> 9;
770 /* mixed attributes always follow the first bio */
771 if (req->rq_flags & RQF_MIXED_MERGE) {
772 req->cmd_flags &= ~REQ_FAILFAST_MASK;
773 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
776 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
778 * If total number of sectors is less than the first segment
779 * size, something has gone terribly wrong.
781 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
782 blk_dump_rq_flags(req, "request botched");
783 req->__data_len = blk_rq_cur_bytes(req);
786 /* recalculate the number of segments */
787 req->nr_phys_segments = blk_recalc_rq_segments(req);
792 EXPORT_SYMBOL_GPL(blk_update_request);
794 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
796 if (rq->rq_flags & RQF_STATS) {
797 blk_mq_poll_stats_start(rq->q);
798 blk_stat_add(rq, now);
801 blk_mq_sched_completed_request(rq, now);
802 blk_account_io_done(rq, now);
805 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
807 if (blk_mq_need_time_stamp(rq))
808 __blk_mq_end_request_acct(rq, ktime_get_ns());
811 rq_qos_done(rq->q, rq);
812 rq->end_io(rq, error);
814 blk_mq_free_request(rq);
817 EXPORT_SYMBOL(__blk_mq_end_request);
819 void blk_mq_end_request(struct request *rq, blk_status_t error)
821 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
823 __blk_mq_end_request(rq, error);
825 EXPORT_SYMBOL(blk_mq_end_request);
827 #define TAG_COMP_BATCH 32
829 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
830 int *tag_array, int nr_tags)
832 struct request_queue *q = hctx->queue;
835 * All requests should have been marked as RQF_MQ_INFLIGHT, so
836 * update hctx->nr_active in batch
838 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
839 __blk_mq_sub_active_requests(hctx, nr_tags);
841 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
842 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
845 void blk_mq_end_request_batch(struct io_comp_batch *iob)
847 int tags[TAG_COMP_BATCH], nr_tags = 0;
848 struct blk_mq_hw_ctx *cur_hctx = NULL;
853 now = ktime_get_ns();
855 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
857 prefetch(rq->rq_next);
859 blk_update_request(rq, BLK_STS_OK, blk_rq_bytes(rq));
861 __blk_mq_end_request_acct(rq, now);
863 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
864 if (!refcount_dec_and_test(&rq->ref))
867 blk_crypto_free_request(rq);
868 blk_pm_mark_last_busy(rq);
869 rq_qos_done(rq->q, rq);
871 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
873 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
875 cur_hctx = rq->mq_hctx;
877 tags[nr_tags++] = rq->tag;
881 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
883 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
885 static void blk_complete_reqs(struct llist_head *list)
887 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
888 struct request *rq, *next;
890 llist_for_each_entry_safe(rq, next, entry, ipi_list)
891 rq->q->mq_ops->complete(rq);
894 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
896 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
899 static int blk_softirq_cpu_dead(unsigned int cpu)
901 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
905 static void __blk_mq_complete_request_remote(void *data)
907 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
910 static inline bool blk_mq_complete_need_ipi(struct request *rq)
912 int cpu = raw_smp_processor_id();
914 if (!IS_ENABLED(CONFIG_SMP) ||
915 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
918 * With force threaded interrupts enabled, raising softirq from an SMP
919 * function call will always result in waking the ksoftirqd thread.
920 * This is probably worse than completing the request on a different
923 if (force_irqthreads())
926 /* same CPU or cache domain? Complete locally */
927 if (cpu == rq->mq_ctx->cpu ||
928 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
929 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
932 /* don't try to IPI to an offline CPU */
933 return cpu_online(rq->mq_ctx->cpu);
936 static void blk_mq_complete_send_ipi(struct request *rq)
938 struct llist_head *list;
941 cpu = rq->mq_ctx->cpu;
942 list = &per_cpu(blk_cpu_done, cpu);
943 if (llist_add(&rq->ipi_list, list)) {
944 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
945 smp_call_function_single_async(cpu, &rq->csd);
949 static void blk_mq_raise_softirq(struct request *rq)
951 struct llist_head *list;
954 list = this_cpu_ptr(&blk_cpu_done);
955 if (llist_add(&rq->ipi_list, list))
956 raise_softirq(BLOCK_SOFTIRQ);
960 bool blk_mq_complete_request_remote(struct request *rq)
962 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
965 * For a polled request, always complete locallly, it's pointless
966 * to redirect the completion.
968 if (rq->cmd_flags & REQ_POLLED)
971 if (blk_mq_complete_need_ipi(rq)) {
972 blk_mq_complete_send_ipi(rq);
976 if (rq->q->nr_hw_queues == 1) {
977 blk_mq_raise_softirq(rq);
982 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
985 * blk_mq_complete_request - end I/O on a request
986 * @rq: the request being processed
989 * Complete a request by scheduling the ->complete_rq operation.
991 void blk_mq_complete_request(struct request *rq)
993 if (!blk_mq_complete_request_remote(rq))
994 rq->q->mq_ops->complete(rq);
996 EXPORT_SYMBOL(blk_mq_complete_request);
998 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
999 __releases(hctx->srcu)
1001 if (!(hctx->flags & BLK_MQ_F_BLOCKING))
1004 srcu_read_unlock(hctx->srcu, srcu_idx);
1007 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
1008 __acquires(hctx->srcu)
1010 if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1011 /* shut up gcc false positive */
1015 *srcu_idx = srcu_read_lock(hctx->srcu);
1019 * blk_mq_start_request - Start processing a request
1020 * @rq: Pointer to request to be started
1022 * Function used by device drivers to notify the block layer that a request
1023 * is going to be processed now, so blk layer can do proper initializations
1024 * such as starting the timeout timer.
1026 void blk_mq_start_request(struct request *rq)
1028 struct request_queue *q = rq->q;
1030 trace_block_rq_issue(rq);
1032 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1034 #ifdef CONFIG_BLK_CGROUP
1036 start_time = bio_issue_time(&rq->bio->bi_issue);
1039 start_time = ktime_get_ns();
1040 rq->io_start_time_ns = start_time;
1041 rq->stats_sectors = blk_rq_sectors(rq);
1042 rq->rq_flags |= RQF_STATS;
1043 rq_qos_issue(q, rq);
1046 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1049 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1051 #ifdef CONFIG_BLK_DEV_INTEGRITY
1052 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1053 q->integrity.profile->prepare_fn(rq);
1055 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1056 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1058 EXPORT_SYMBOL(blk_mq_start_request);
1060 static void __blk_mq_requeue_request(struct request *rq)
1062 struct request_queue *q = rq->q;
1064 blk_mq_put_driver_tag(rq);
1066 trace_block_rq_requeue(rq);
1067 rq_qos_requeue(q, rq);
1069 if (blk_mq_request_started(rq)) {
1070 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1071 rq->rq_flags &= ~RQF_TIMED_OUT;
1075 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1077 __blk_mq_requeue_request(rq);
1079 /* this request will be re-inserted to io scheduler queue */
1080 blk_mq_sched_requeue_request(rq);
1082 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1084 EXPORT_SYMBOL(blk_mq_requeue_request);
1086 static void blk_mq_requeue_work(struct work_struct *work)
1088 struct request_queue *q =
1089 container_of(work, struct request_queue, requeue_work.work);
1091 struct request *rq, *next;
1093 spin_lock_irq(&q->requeue_lock);
1094 list_splice_init(&q->requeue_list, &rq_list);
1095 spin_unlock_irq(&q->requeue_lock);
1097 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1098 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1101 rq->rq_flags &= ~RQF_SOFTBARRIER;
1102 list_del_init(&rq->queuelist);
1104 * If RQF_DONTPREP, rq has contained some driver specific
1105 * data, so insert it to hctx dispatch list to avoid any
1108 if (rq->rq_flags & RQF_DONTPREP)
1109 blk_mq_request_bypass_insert(rq, false, false);
1111 blk_mq_sched_insert_request(rq, true, false, false);
1114 while (!list_empty(&rq_list)) {
1115 rq = list_entry(rq_list.next, struct request, queuelist);
1116 list_del_init(&rq->queuelist);
1117 blk_mq_sched_insert_request(rq, false, false, false);
1120 blk_mq_run_hw_queues(q, false);
1123 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1124 bool kick_requeue_list)
1126 struct request_queue *q = rq->q;
1127 unsigned long flags;
1130 * We abuse this flag that is otherwise used by the I/O scheduler to
1131 * request head insertion from the workqueue.
1133 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1135 spin_lock_irqsave(&q->requeue_lock, flags);
1137 rq->rq_flags |= RQF_SOFTBARRIER;
1138 list_add(&rq->queuelist, &q->requeue_list);
1140 list_add_tail(&rq->queuelist, &q->requeue_list);
1142 spin_unlock_irqrestore(&q->requeue_lock, flags);
1144 if (kick_requeue_list)
1145 blk_mq_kick_requeue_list(q);
1148 void blk_mq_kick_requeue_list(struct request_queue *q)
1150 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1152 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1154 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1155 unsigned long msecs)
1157 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1158 msecs_to_jiffies(msecs));
1160 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1162 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
1163 void *priv, bool reserved)
1166 * If we find a request that isn't idle and the queue matches,
1167 * we know the queue is busy. Return false to stop the iteration.
1169 if (blk_mq_request_started(rq) && rq->q == hctx->queue) {
1179 bool blk_mq_queue_inflight(struct request_queue *q)
1183 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1186 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1188 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
1190 req->rq_flags |= RQF_TIMED_OUT;
1191 if (req->q->mq_ops->timeout) {
1192 enum blk_eh_timer_return ret;
1194 ret = req->q->mq_ops->timeout(req, reserved);
1195 if (ret == BLK_EH_DONE)
1197 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1203 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
1205 unsigned long deadline;
1207 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1209 if (rq->rq_flags & RQF_TIMED_OUT)
1212 deadline = READ_ONCE(rq->deadline);
1213 if (time_after_eq(jiffies, deadline))
1218 else if (time_after(*next, deadline))
1223 void blk_mq_put_rq_ref(struct request *rq)
1225 if (is_flush_rq(rq))
1227 else if (refcount_dec_and_test(&rq->ref))
1228 __blk_mq_free_request(rq);
1231 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
1232 struct request *rq, void *priv, bool reserved)
1234 unsigned long *next = priv;
1237 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1238 * be reallocated underneath the timeout handler's processing, then
1239 * the expire check is reliable. If the request is not expired, then
1240 * it was completed and reallocated as a new request after returning
1241 * from blk_mq_check_expired().
1243 if (blk_mq_req_expired(rq, next))
1244 blk_mq_rq_timed_out(rq, reserved);
1248 static void blk_mq_timeout_work(struct work_struct *work)
1250 struct request_queue *q =
1251 container_of(work, struct request_queue, timeout_work);
1252 unsigned long next = 0;
1253 struct blk_mq_hw_ctx *hctx;
1256 /* A deadlock might occur if a request is stuck requiring a
1257 * timeout at the same time a queue freeze is waiting
1258 * completion, since the timeout code would not be able to
1259 * acquire the queue reference here.
1261 * That's why we don't use blk_queue_enter here; instead, we use
1262 * percpu_ref_tryget directly, because we need to be able to
1263 * obtain a reference even in the short window between the queue
1264 * starting to freeze, by dropping the first reference in
1265 * blk_freeze_queue_start, and the moment the last request is
1266 * consumed, marked by the instant q_usage_counter reaches
1269 if (!percpu_ref_tryget(&q->q_usage_counter))
1272 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
1275 mod_timer(&q->timeout, next);
1278 * Request timeouts are handled as a forward rolling timer. If
1279 * we end up here it means that no requests are pending and
1280 * also that no request has been pending for a while. Mark
1281 * each hctx as idle.
1283 queue_for_each_hw_ctx(q, hctx, i) {
1284 /* the hctx may be unmapped, so check it here */
1285 if (blk_mq_hw_queue_mapped(hctx))
1286 blk_mq_tag_idle(hctx);
1292 struct flush_busy_ctx_data {
1293 struct blk_mq_hw_ctx *hctx;
1294 struct list_head *list;
1297 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1299 struct flush_busy_ctx_data *flush_data = data;
1300 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1301 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1302 enum hctx_type type = hctx->type;
1304 spin_lock(&ctx->lock);
1305 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1306 sbitmap_clear_bit(sb, bitnr);
1307 spin_unlock(&ctx->lock);
1312 * Process software queues that have been marked busy, splicing them
1313 * to the for-dispatch
1315 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1317 struct flush_busy_ctx_data data = {
1322 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1324 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1326 struct dispatch_rq_data {
1327 struct blk_mq_hw_ctx *hctx;
1331 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1334 struct dispatch_rq_data *dispatch_data = data;
1335 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1336 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1337 enum hctx_type type = hctx->type;
1339 spin_lock(&ctx->lock);
1340 if (!list_empty(&ctx->rq_lists[type])) {
1341 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1342 list_del_init(&dispatch_data->rq->queuelist);
1343 if (list_empty(&ctx->rq_lists[type]))
1344 sbitmap_clear_bit(sb, bitnr);
1346 spin_unlock(&ctx->lock);
1348 return !dispatch_data->rq;
1351 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1352 struct blk_mq_ctx *start)
1354 unsigned off = start ? start->index_hw[hctx->type] : 0;
1355 struct dispatch_rq_data data = {
1360 __sbitmap_for_each_set(&hctx->ctx_map, off,
1361 dispatch_rq_from_ctx, &data);
1366 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1368 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1369 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1372 blk_mq_tag_busy(rq->mq_hctx);
1374 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1375 bt = &rq->mq_hctx->tags->breserved_tags;
1378 if (!hctx_may_queue(rq->mq_hctx, bt))
1382 tag = __sbitmap_queue_get(bt);
1383 if (tag == BLK_MQ_NO_TAG)
1386 rq->tag = tag + tag_offset;
1390 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1392 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1395 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1396 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1397 rq->rq_flags |= RQF_MQ_INFLIGHT;
1398 __blk_mq_inc_active_requests(hctx);
1400 hctx->tags->rqs[rq->tag] = rq;
1404 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1405 int flags, void *key)
1407 struct blk_mq_hw_ctx *hctx;
1409 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1411 spin_lock(&hctx->dispatch_wait_lock);
1412 if (!list_empty(&wait->entry)) {
1413 struct sbitmap_queue *sbq;
1415 list_del_init(&wait->entry);
1416 sbq = &hctx->tags->bitmap_tags;
1417 atomic_dec(&sbq->ws_active);
1419 spin_unlock(&hctx->dispatch_wait_lock);
1421 blk_mq_run_hw_queue(hctx, true);
1426 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1427 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1428 * restart. For both cases, take care to check the condition again after
1429 * marking us as waiting.
1431 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1434 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1435 struct wait_queue_head *wq;
1436 wait_queue_entry_t *wait;
1439 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1440 blk_mq_sched_mark_restart_hctx(hctx);
1443 * It's possible that a tag was freed in the window between the
1444 * allocation failure and adding the hardware queue to the wait
1447 * Don't clear RESTART here, someone else could have set it.
1448 * At most this will cost an extra queue run.
1450 return blk_mq_get_driver_tag(rq);
1453 wait = &hctx->dispatch_wait;
1454 if (!list_empty_careful(&wait->entry))
1457 wq = &bt_wait_ptr(sbq, hctx)->wait;
1459 spin_lock_irq(&wq->lock);
1460 spin_lock(&hctx->dispatch_wait_lock);
1461 if (!list_empty(&wait->entry)) {
1462 spin_unlock(&hctx->dispatch_wait_lock);
1463 spin_unlock_irq(&wq->lock);
1467 atomic_inc(&sbq->ws_active);
1468 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1469 __add_wait_queue(wq, wait);
1472 * It's possible that a tag was freed in the window between the
1473 * allocation failure and adding the hardware queue to the wait
1476 ret = blk_mq_get_driver_tag(rq);
1478 spin_unlock(&hctx->dispatch_wait_lock);
1479 spin_unlock_irq(&wq->lock);
1484 * We got a tag, remove ourselves from the wait queue to ensure
1485 * someone else gets the wakeup.
1487 list_del_init(&wait->entry);
1488 atomic_dec(&sbq->ws_active);
1489 spin_unlock(&hctx->dispatch_wait_lock);
1490 spin_unlock_irq(&wq->lock);
1495 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1496 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1498 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1499 * - EWMA is one simple way to compute running average value
1500 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1501 * - take 4 as factor for avoiding to get too small(0) result, and this
1502 * factor doesn't matter because EWMA decreases exponentially
1504 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1508 ewma = hctx->dispatch_busy;
1513 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1515 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1516 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1518 hctx->dispatch_busy = ewma;
1521 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1523 static void blk_mq_handle_dev_resource(struct request *rq,
1524 struct list_head *list)
1526 struct request *next =
1527 list_first_entry_or_null(list, struct request, queuelist);
1530 * If an I/O scheduler has been configured and we got a driver tag for
1531 * the next request already, free it.
1534 blk_mq_put_driver_tag(next);
1536 list_add(&rq->queuelist, list);
1537 __blk_mq_requeue_request(rq);
1540 static void blk_mq_handle_zone_resource(struct request *rq,
1541 struct list_head *zone_list)
1544 * If we end up here it is because we cannot dispatch a request to a
1545 * specific zone due to LLD level zone-write locking or other zone
1546 * related resource not being available. In this case, set the request
1547 * aside in zone_list for retrying it later.
1549 list_add(&rq->queuelist, zone_list);
1550 __blk_mq_requeue_request(rq);
1553 enum prep_dispatch {
1555 PREP_DISPATCH_NO_TAG,
1556 PREP_DISPATCH_NO_BUDGET,
1559 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1562 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1563 int budget_token = -1;
1566 budget_token = blk_mq_get_dispatch_budget(rq->q);
1567 if (budget_token < 0) {
1568 blk_mq_put_driver_tag(rq);
1569 return PREP_DISPATCH_NO_BUDGET;
1571 blk_mq_set_rq_budget_token(rq, budget_token);
1574 if (!blk_mq_get_driver_tag(rq)) {
1576 * The initial allocation attempt failed, so we need to
1577 * rerun the hardware queue when a tag is freed. The
1578 * waitqueue takes care of that. If the queue is run
1579 * before we add this entry back on the dispatch list,
1580 * we'll re-run it below.
1582 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1584 * All budgets not got from this function will be put
1585 * together during handling partial dispatch
1588 blk_mq_put_dispatch_budget(rq->q, budget_token);
1589 return PREP_DISPATCH_NO_TAG;
1593 return PREP_DISPATCH_OK;
1596 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1597 static void blk_mq_release_budgets(struct request_queue *q,
1598 struct list_head *list)
1602 list_for_each_entry(rq, list, queuelist) {
1603 int budget_token = blk_mq_get_rq_budget_token(rq);
1605 if (budget_token >= 0)
1606 blk_mq_put_dispatch_budget(q, budget_token);
1611 * Returns true if we did some work AND can potentially do more.
1613 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1614 unsigned int nr_budgets)
1616 enum prep_dispatch prep;
1617 struct request_queue *q = hctx->queue;
1618 struct request *rq, *nxt;
1620 blk_status_t ret = BLK_STS_OK;
1621 LIST_HEAD(zone_list);
1622 bool needs_resource = false;
1624 if (list_empty(list))
1628 * Now process all the entries, sending them to the driver.
1630 errors = queued = 0;
1632 struct blk_mq_queue_data bd;
1634 rq = list_first_entry(list, struct request, queuelist);
1636 WARN_ON_ONCE(hctx != rq->mq_hctx);
1637 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1638 if (prep != PREP_DISPATCH_OK)
1641 list_del_init(&rq->queuelist);
1646 * Flag last if we have no more requests, or if we have more
1647 * but can't assign a driver tag to it.
1649 if (list_empty(list))
1652 nxt = list_first_entry(list, struct request, queuelist);
1653 bd.last = !blk_mq_get_driver_tag(nxt);
1657 * once the request is queued to lld, no need to cover the
1662 ret = q->mq_ops->queue_rq(hctx, &bd);
1667 case BLK_STS_RESOURCE:
1668 needs_resource = true;
1670 case BLK_STS_DEV_RESOURCE:
1671 blk_mq_handle_dev_resource(rq, list);
1673 case BLK_STS_ZONE_RESOURCE:
1675 * Move the request to zone_list and keep going through
1676 * the dispatch list to find more requests the drive can
1679 blk_mq_handle_zone_resource(rq, &zone_list);
1680 needs_resource = true;
1684 blk_mq_end_request(rq, ret);
1686 } while (!list_empty(list));
1688 if (!list_empty(&zone_list))
1689 list_splice_tail_init(&zone_list, list);
1691 /* If we didn't flush the entire list, we could have told the driver
1692 * there was more coming, but that turned out to be a lie.
1694 if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued)
1695 q->mq_ops->commit_rqs(hctx);
1697 * Any items that need requeuing? Stuff them into hctx->dispatch,
1698 * that is where we will continue on next queue run.
1700 if (!list_empty(list)) {
1702 /* For non-shared tags, the RESTART check will suffice */
1703 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1704 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1707 blk_mq_release_budgets(q, list);
1709 spin_lock(&hctx->lock);
1710 list_splice_tail_init(list, &hctx->dispatch);
1711 spin_unlock(&hctx->lock);
1714 * Order adding requests to hctx->dispatch and checking
1715 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1716 * in blk_mq_sched_restart(). Avoid restart code path to
1717 * miss the new added requests to hctx->dispatch, meantime
1718 * SCHED_RESTART is observed here.
1723 * If SCHED_RESTART was set by the caller of this function and
1724 * it is no longer set that means that it was cleared by another
1725 * thread and hence that a queue rerun is needed.
1727 * If 'no_tag' is set, that means that we failed getting
1728 * a driver tag with an I/O scheduler attached. If our dispatch
1729 * waitqueue is no longer active, ensure that we run the queue
1730 * AFTER adding our entries back to the list.
1732 * If no I/O scheduler has been configured it is possible that
1733 * the hardware queue got stopped and restarted before requests
1734 * were pushed back onto the dispatch list. Rerun the queue to
1735 * avoid starvation. Notes:
1736 * - blk_mq_run_hw_queue() checks whether or not a queue has
1737 * been stopped before rerunning a queue.
1738 * - Some but not all block drivers stop a queue before
1739 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1742 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1743 * bit is set, run queue after a delay to avoid IO stalls
1744 * that could otherwise occur if the queue is idle. We'll do
1745 * similar if we couldn't get budget or couldn't lock a zone
1746 * and SCHED_RESTART is set.
1748 needs_restart = blk_mq_sched_needs_restart(hctx);
1749 if (prep == PREP_DISPATCH_NO_BUDGET)
1750 needs_resource = true;
1751 if (!needs_restart ||
1752 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1753 blk_mq_run_hw_queue(hctx, true);
1754 else if (needs_restart && needs_resource)
1755 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1757 blk_mq_update_dispatch_busy(hctx, true);
1760 blk_mq_update_dispatch_busy(hctx, false);
1762 return (queued + errors) != 0;
1766 * __blk_mq_run_hw_queue - Run a hardware queue.
1767 * @hctx: Pointer to the hardware queue to run.
1769 * Send pending requests to the hardware.
1771 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1776 * We can't run the queue inline with ints disabled. Ensure that
1777 * we catch bad users of this early.
1779 WARN_ON_ONCE(in_interrupt());
1781 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1783 hctx_lock(hctx, &srcu_idx);
1784 blk_mq_sched_dispatch_requests(hctx);
1785 hctx_unlock(hctx, srcu_idx);
1788 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1790 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1792 if (cpu >= nr_cpu_ids)
1793 cpu = cpumask_first(hctx->cpumask);
1798 * It'd be great if the workqueue API had a way to pass
1799 * in a mask and had some smarts for more clever placement.
1800 * For now we just round-robin here, switching for every
1801 * BLK_MQ_CPU_WORK_BATCH queued items.
1803 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1806 int next_cpu = hctx->next_cpu;
1808 if (hctx->queue->nr_hw_queues == 1)
1809 return WORK_CPU_UNBOUND;
1811 if (--hctx->next_cpu_batch <= 0) {
1813 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1815 if (next_cpu >= nr_cpu_ids)
1816 next_cpu = blk_mq_first_mapped_cpu(hctx);
1817 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1821 * Do unbound schedule if we can't find a online CPU for this hctx,
1822 * and it should only happen in the path of handling CPU DEAD.
1824 if (!cpu_online(next_cpu)) {
1831 * Make sure to re-select CPU next time once after CPUs
1832 * in hctx->cpumask become online again.
1834 hctx->next_cpu = next_cpu;
1835 hctx->next_cpu_batch = 1;
1836 return WORK_CPU_UNBOUND;
1839 hctx->next_cpu = next_cpu;
1844 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
1845 * @hctx: Pointer to the hardware queue to run.
1846 * @async: If we want to run the queue asynchronously.
1847 * @msecs: Milliseconds of delay to wait before running the queue.
1849 * If !@async, try to run the queue now. Else, run the queue asynchronously and
1850 * with a delay of @msecs.
1852 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1853 unsigned long msecs)
1855 if (unlikely(blk_mq_hctx_stopped(hctx)))
1858 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1859 int cpu = get_cpu();
1860 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1861 __blk_mq_run_hw_queue(hctx);
1869 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1870 msecs_to_jiffies(msecs));
1874 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
1875 * @hctx: Pointer to the hardware queue to run.
1876 * @msecs: Milliseconds of delay to wait before running the queue.
1878 * Run a hardware queue asynchronously with a delay of @msecs.
1880 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1882 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1884 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1887 * blk_mq_run_hw_queue - Start to run a hardware queue.
1888 * @hctx: Pointer to the hardware queue to run.
1889 * @async: If we want to run the queue asynchronously.
1891 * Check if the request queue is not in a quiesced state and if there are
1892 * pending requests to be sent. If this is true, run the queue to send requests
1895 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1901 * When queue is quiesced, we may be switching io scheduler, or
1902 * updating nr_hw_queues, or other things, and we can't run queue
1903 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1905 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1908 hctx_lock(hctx, &srcu_idx);
1909 need_run = !blk_queue_quiesced(hctx->queue) &&
1910 blk_mq_hctx_has_pending(hctx);
1911 hctx_unlock(hctx, srcu_idx);
1914 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1916 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1919 * Is the request queue handled by an IO scheduler that does not respect
1920 * hardware queues when dispatching?
1922 static bool blk_mq_has_sqsched(struct request_queue *q)
1924 struct elevator_queue *e = q->elevator;
1926 if (e && e->type->ops.dispatch_request &&
1927 !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE))
1933 * Return prefered queue to dispatch from (if any) for non-mq aware IO
1936 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
1938 struct blk_mq_hw_ctx *hctx;
1941 * If the IO scheduler does not respect hardware queues when
1942 * dispatching, we just don't bother with multiple HW queues and
1943 * dispatch from hctx for the current CPU since running multiple queues
1944 * just causes lock contention inside the scheduler and pointless cache
1947 hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT,
1948 raw_smp_processor_id());
1949 if (!blk_mq_hctx_stopped(hctx))
1955 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
1956 * @q: Pointer to the request queue to run.
1957 * @async: If we want to run the queue asynchronously.
1959 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1961 struct blk_mq_hw_ctx *hctx, *sq_hctx;
1965 if (blk_mq_has_sqsched(q))
1966 sq_hctx = blk_mq_get_sq_hctx(q);
1967 queue_for_each_hw_ctx(q, hctx, i) {
1968 if (blk_mq_hctx_stopped(hctx))
1971 * Dispatch from this hctx either if there's no hctx preferred
1972 * by IO scheduler or if it has requests that bypass the
1975 if (!sq_hctx || sq_hctx == hctx ||
1976 !list_empty_careful(&hctx->dispatch))
1977 blk_mq_run_hw_queue(hctx, async);
1980 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1983 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
1984 * @q: Pointer to the request queue to run.
1985 * @msecs: Milliseconds of delay to wait before running the queues.
1987 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
1989 struct blk_mq_hw_ctx *hctx, *sq_hctx;
1993 if (blk_mq_has_sqsched(q))
1994 sq_hctx = blk_mq_get_sq_hctx(q);
1995 queue_for_each_hw_ctx(q, hctx, i) {
1996 if (blk_mq_hctx_stopped(hctx))
1999 * Dispatch from this hctx either if there's no hctx preferred
2000 * by IO scheduler or if it has requests that bypass the
2003 if (!sq_hctx || sq_hctx == hctx ||
2004 !list_empty_careful(&hctx->dispatch))
2005 blk_mq_delay_run_hw_queue(hctx, msecs);
2008 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2011 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
2012 * @q: request queue.
2014 * The caller is responsible for serializing this function against
2015 * blk_mq_{start,stop}_hw_queue().
2017 bool blk_mq_queue_stopped(struct request_queue *q)
2019 struct blk_mq_hw_ctx *hctx;
2022 queue_for_each_hw_ctx(q, hctx, i)
2023 if (blk_mq_hctx_stopped(hctx))
2028 EXPORT_SYMBOL(blk_mq_queue_stopped);
2031 * This function is often used for pausing .queue_rq() by driver when
2032 * there isn't enough resource or some conditions aren't satisfied, and
2033 * BLK_STS_RESOURCE is usually returned.
2035 * We do not guarantee that dispatch can be drained or blocked
2036 * after blk_mq_stop_hw_queue() returns. Please use
2037 * blk_mq_quiesce_queue() for that requirement.
2039 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2041 cancel_delayed_work(&hctx->run_work);
2043 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2045 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2048 * This function is often used for pausing .queue_rq() by driver when
2049 * there isn't enough resource or some conditions aren't satisfied, and
2050 * BLK_STS_RESOURCE is usually returned.
2052 * We do not guarantee that dispatch can be drained or blocked
2053 * after blk_mq_stop_hw_queues() returns. Please use
2054 * blk_mq_quiesce_queue() for that requirement.
2056 void blk_mq_stop_hw_queues(struct request_queue *q)
2058 struct blk_mq_hw_ctx *hctx;
2061 queue_for_each_hw_ctx(q, hctx, i)
2062 blk_mq_stop_hw_queue(hctx);
2064 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2066 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2068 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2070 blk_mq_run_hw_queue(hctx, false);
2072 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2074 void blk_mq_start_hw_queues(struct request_queue *q)
2076 struct blk_mq_hw_ctx *hctx;
2079 queue_for_each_hw_ctx(q, hctx, i)
2080 blk_mq_start_hw_queue(hctx);
2082 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2084 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2086 if (!blk_mq_hctx_stopped(hctx))
2089 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2090 blk_mq_run_hw_queue(hctx, async);
2092 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2094 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2096 struct blk_mq_hw_ctx *hctx;
2099 queue_for_each_hw_ctx(q, hctx, i)
2100 blk_mq_start_stopped_hw_queue(hctx, async);
2102 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2104 static void blk_mq_run_work_fn(struct work_struct *work)
2106 struct blk_mq_hw_ctx *hctx;
2108 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2111 * If we are stopped, don't run the queue.
2113 if (blk_mq_hctx_stopped(hctx))
2116 __blk_mq_run_hw_queue(hctx);
2119 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2123 struct blk_mq_ctx *ctx = rq->mq_ctx;
2124 enum hctx_type type = hctx->type;
2126 lockdep_assert_held(&ctx->lock);
2128 trace_block_rq_insert(rq);
2131 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2133 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2136 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2139 struct blk_mq_ctx *ctx = rq->mq_ctx;
2141 lockdep_assert_held(&ctx->lock);
2143 __blk_mq_insert_req_list(hctx, rq, at_head);
2144 blk_mq_hctx_mark_pending(hctx, ctx);
2148 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2149 * @rq: Pointer to request to be inserted.
2150 * @at_head: true if the request should be inserted at the head of the list.
2151 * @run_queue: If we should run the hardware queue after inserting the request.
2153 * Should only be used carefully, when the caller knows we want to
2154 * bypass a potential IO scheduler on the target device.
2156 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2159 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2161 spin_lock(&hctx->lock);
2163 list_add(&rq->queuelist, &hctx->dispatch);
2165 list_add_tail(&rq->queuelist, &hctx->dispatch);
2166 spin_unlock(&hctx->lock);
2169 blk_mq_run_hw_queue(hctx, false);
2172 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2173 struct list_head *list)
2177 enum hctx_type type = hctx->type;
2180 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2183 list_for_each_entry(rq, list, queuelist) {
2184 BUG_ON(rq->mq_ctx != ctx);
2185 trace_block_rq_insert(rq);
2188 spin_lock(&ctx->lock);
2189 list_splice_tail_init(list, &ctx->rq_lists[type]);
2190 blk_mq_hctx_mark_pending(hctx, ctx);
2191 spin_unlock(&ctx->lock);
2194 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2197 if (hctx->queue->mq_ops->commit_rqs) {
2198 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2199 hctx->queue->mq_ops->commit_rqs(hctx);
2204 static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2206 struct blk_mq_hw_ctx *hctx = NULL;
2211 while ((rq = rq_list_pop(&plug->mq_list))) {
2212 bool last = rq_list_empty(plug->mq_list);
2215 if (hctx != rq->mq_hctx) {
2217 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2221 ret = blk_mq_request_issue_directly(rq, last);
2226 case BLK_STS_RESOURCE:
2227 case BLK_STS_DEV_RESOURCE:
2228 blk_mq_request_bypass_insert(rq, false, last);
2229 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2232 blk_mq_end_request(rq, ret);
2239 * If we didn't flush the entire list, we could have told the driver
2240 * there was more coming, but that turned out to be a lie.
2243 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2246 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2248 struct blk_mq_hw_ctx *this_hctx;
2249 struct blk_mq_ctx *this_ctx;
2253 if (rq_list_empty(plug->mq_list))
2257 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2258 blk_mq_plug_issue_direct(plug, false);
2259 if (rq_list_empty(plug->mq_list))
2269 rq = rq_list_pop(&plug->mq_list);
2272 this_hctx = rq->mq_hctx;
2273 this_ctx = rq->mq_ctx;
2274 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2275 trace_block_unplug(this_hctx->queue, depth,
2277 blk_mq_sched_insert_requests(this_hctx, this_ctx,
2278 &list, from_schedule);
2280 this_hctx = rq->mq_hctx;
2281 this_ctx = rq->mq_ctx;
2285 list_add(&rq->queuelist, &list);
2287 } while (!rq_list_empty(plug->mq_list));
2289 if (!list_empty(&list)) {
2290 trace_block_unplug(this_hctx->queue, depth, !from_schedule);
2291 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list,
2296 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2297 unsigned int nr_segs)
2301 if (bio->bi_opf & REQ_RAHEAD)
2302 rq->cmd_flags |= REQ_FAILFAST_MASK;
2304 rq->__sector = bio->bi_iter.bi_sector;
2305 rq->write_hint = bio->bi_write_hint;
2306 blk_rq_bio_prep(rq, bio, nr_segs);
2308 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2309 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2312 blk_account_io_start(rq);
2315 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2316 struct request *rq, bool last)
2318 struct request_queue *q = rq->q;
2319 struct blk_mq_queue_data bd = {
2326 * For OK queue, we are done. For error, caller may kill it.
2327 * Any other error (busy), just add it to our list as we
2328 * previously would have done.
2330 ret = q->mq_ops->queue_rq(hctx, &bd);
2333 blk_mq_update_dispatch_busy(hctx, false);
2335 case BLK_STS_RESOURCE:
2336 case BLK_STS_DEV_RESOURCE:
2337 blk_mq_update_dispatch_busy(hctx, true);
2338 __blk_mq_requeue_request(rq);
2341 blk_mq_update_dispatch_busy(hctx, false);
2348 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2350 bool bypass_insert, bool last)
2352 struct request_queue *q = rq->q;
2353 bool run_queue = true;
2357 * RCU or SRCU read lock is needed before checking quiesced flag.
2359 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2360 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2361 * and avoid driver to try to dispatch again.
2363 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2365 bypass_insert = false;
2369 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2372 budget_token = blk_mq_get_dispatch_budget(q);
2373 if (budget_token < 0)
2376 blk_mq_set_rq_budget_token(rq, budget_token);
2378 if (!blk_mq_get_driver_tag(rq)) {
2379 blk_mq_put_dispatch_budget(q, budget_token);
2383 return __blk_mq_issue_directly(hctx, rq, last);
2386 return BLK_STS_RESOURCE;
2388 blk_mq_sched_insert_request(rq, false, run_queue, false);
2394 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2395 * @hctx: Pointer of the associated hardware queue.
2396 * @rq: Pointer to request to be sent.
2398 * If the device has enough resources to accept a new request now, send the
2399 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2400 * we can try send it another time in the future. Requests inserted at this
2401 * queue have higher priority.
2403 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2409 might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
2411 hctx_lock(hctx, &srcu_idx);
2413 ret = __blk_mq_try_issue_directly(hctx, rq, false, true);
2414 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2415 blk_mq_request_bypass_insert(rq, false, true);
2416 else if (ret != BLK_STS_OK)
2417 blk_mq_end_request(rq, ret);
2419 hctx_unlock(hctx, srcu_idx);
2422 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2426 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2428 hctx_lock(hctx, &srcu_idx);
2429 ret = __blk_mq_try_issue_directly(hctx, rq, true, last);
2430 hctx_unlock(hctx, srcu_idx);
2435 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2436 struct list_head *list)
2441 while (!list_empty(list)) {
2443 struct request *rq = list_first_entry(list, struct request,
2446 list_del_init(&rq->queuelist);
2447 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2448 if (ret != BLK_STS_OK) {
2449 if (ret == BLK_STS_RESOURCE ||
2450 ret == BLK_STS_DEV_RESOURCE) {
2451 blk_mq_request_bypass_insert(rq, false,
2455 blk_mq_end_request(rq, ret);
2462 * If we didn't flush the entire list, we could have told
2463 * the driver there was more coming, but that turned out to
2466 if ((!list_empty(list) || errors) &&
2467 hctx->queue->mq_ops->commit_rqs && queued)
2468 hctx->queue->mq_ops->commit_rqs(hctx);
2471 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2473 if (!plug->multiple_queues) {
2474 struct request *nxt = rq_list_peek(&plug->mq_list);
2476 if (nxt && nxt->q != rq->q)
2477 plug->multiple_queues = true;
2479 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
2480 plug->has_elevator = true;
2482 rq_list_add(&plug->mq_list, rq);
2487 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
2488 * queues. This is important for md arrays to benefit from merging
2491 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
2493 if (plug->multiple_queues)
2494 return BLK_MAX_REQUEST_COUNT * 2;
2495 return BLK_MAX_REQUEST_COUNT;
2498 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2499 struct bio *bio, unsigned int nr_segs,
2500 bool *same_queue_rq)
2502 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2503 if (blk_attempt_plug_merge(q, bio, nr_segs, same_queue_rq))
2505 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2511 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2512 struct blk_plug *plug,
2515 bool *same_queue_rq)
2517 struct blk_mq_alloc_data data = {
2520 .cmd_flags = bio->bi_opf,
2524 if (blk_mq_attempt_bio_merge(q, bio, nsegs, same_queue_rq))
2527 rq_qos_throttle(q, bio);
2530 data.nr_tags = plug->nr_ios;
2532 data.cached_rq = &plug->cached_rq;
2535 rq = __blk_mq_alloc_requests(&data);
2539 rq_qos_cleanup(q, bio);
2540 if (bio->bi_opf & REQ_NOWAIT)
2541 bio_wouldblock_error(bio);
2546 static inline bool blk_mq_can_use_cached_rq(struct request *rq, struct bio *bio)
2548 if (blk_mq_get_hctx_type(bio->bi_opf) != rq->mq_hctx->type)
2551 if (op_is_flush(rq->cmd_flags) != op_is_flush(bio->bi_opf))
2557 static inline struct request *blk_mq_get_request(struct request_queue *q,
2558 struct blk_plug *plug,
2561 bool *same_queue_rq)
2564 bool checked = false;
2567 rq = rq_list_peek(&plug->cached_rq);
2568 if (rq && rq->q == q) {
2569 if (unlikely(!submit_bio_checks(bio)))
2571 if (blk_mq_attempt_bio_merge(q, bio, nsegs,
2575 if (!blk_mq_can_use_cached_rq(rq, bio))
2577 rq->cmd_flags = bio->bi_opf;
2578 plug->cached_rq = rq_list_next(rq);
2579 INIT_LIST_HEAD(&rq->queuelist);
2580 rq_qos_throttle(q, bio);
2586 if (unlikely(bio_queue_enter(bio)))
2588 if (unlikely(!checked && !submit_bio_checks(bio)))
2590 rq = blk_mq_get_new_requests(q, plug, bio, nsegs, same_queue_rq);
2599 * blk_mq_submit_bio - Create and send a request to block device.
2600 * @bio: Bio pointer.
2602 * Builds up a request structure from @q and @bio and send to the device. The
2603 * request may not be queued directly to hardware if:
2604 * * This request can be merged with another one
2605 * * We want to place request at plug queue for possible future merging
2606 * * There is an IO scheduler active at this queue
2608 * It will not queue the request if there is an error with the bio, or at the
2611 void blk_mq_submit_bio(struct bio *bio)
2613 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2614 const int is_sync = op_is_sync(bio->bi_opf);
2616 struct blk_plug *plug;
2617 bool same_queue_rq = false;
2618 unsigned int nr_segs = 1;
2621 if (unlikely(!blk_crypto_bio_prep(&bio)))
2624 blk_queue_bounce(q, &bio);
2625 if (blk_may_split(q, bio))
2626 __blk_queue_split(q, &bio, &nr_segs);
2628 if (!bio_integrity_prep(bio))
2631 plug = blk_mq_plug(q, bio);
2632 rq = blk_mq_get_request(q, plug, bio, nr_segs, &same_queue_rq);
2636 trace_block_getrq(bio);
2638 rq_qos_track(q, rq, bio);
2640 blk_mq_bio_to_request(rq, bio, nr_segs);
2642 ret = blk_crypto_init_request(rq);
2643 if (ret != BLK_STS_OK) {
2644 bio->bi_status = ret;
2646 blk_mq_free_request(rq);
2650 if (op_is_flush(bio->bi_opf)) {
2651 blk_insert_flush(rq);
2655 if (plug && (q->nr_hw_queues == 1 ||
2656 blk_mq_is_shared_tags(rq->mq_hctx->flags) ||
2657 q->mq_ops->commit_rqs || !blk_queue_nonrot(q))) {
2659 * Use plugging if we have a ->commit_rqs() hook as well, as
2660 * we know the driver uses bd->last in a smart fashion.
2662 * Use normal plugging if this disk is slow HDD, as sequential
2663 * IO may benefit a lot from plug merging.
2665 unsigned int request_count = plug->rq_count;
2666 struct request *last = NULL;
2668 if (!request_count) {
2669 trace_block_plug(q);
2670 } else if (!blk_queue_nomerges(q)) {
2671 last = rq_list_peek(&plug->mq_list);
2672 if (blk_rq_bytes(last) < BLK_PLUG_FLUSH_SIZE)
2676 if (request_count >= blk_plug_max_rq_count(plug) || last) {
2677 blk_mq_flush_plug_list(plug, false);
2678 trace_block_plug(q);
2681 blk_add_rq_to_plug(plug, rq);
2682 } else if (rq->rq_flags & RQF_ELV) {
2683 /* Insert the request at the IO scheduler queue */
2684 blk_mq_sched_insert_request(rq, false, true, true);
2685 } else if (plug && !blk_queue_nomerges(q)) {
2686 struct request *next_rq = NULL;
2689 * We do limited plugging. If the bio can be merged, do that.
2690 * Otherwise the existing request in the plug list will be
2691 * issued. So the plug list will have one request at most
2692 * The plug list might get flushed before this. If that happens,
2693 * the plug list is empty, and same_queue_rq is invalid.
2695 if (same_queue_rq) {
2696 next_rq = rq_list_pop(&plug->mq_list);
2699 blk_add_rq_to_plug(plug, rq);
2700 trace_block_plug(q);
2703 trace_block_unplug(q, 1, true);
2704 blk_mq_try_issue_directly(next_rq->mq_hctx, next_rq);
2706 } else if ((q->nr_hw_queues > 1 && is_sync) ||
2707 !rq->mq_hctx->dispatch_busy) {
2709 * There is no scheduler and we can try to send directly
2712 blk_mq_try_issue_directly(rq->mq_hctx, rq);
2715 blk_mq_sched_insert_request(rq, false, true, true);
2719 static size_t order_to_size(unsigned int order)
2721 return (size_t)PAGE_SIZE << order;
2724 /* called before freeing request pool in @tags */
2725 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
2726 struct blk_mq_tags *tags)
2729 unsigned long flags;
2731 /* There is no need to clear a driver tags own mapping */
2732 if (drv_tags == tags)
2735 list_for_each_entry(page, &tags->page_list, lru) {
2736 unsigned long start = (unsigned long)page_address(page);
2737 unsigned long end = start + order_to_size(page->private);
2740 for (i = 0; i < drv_tags->nr_tags; i++) {
2741 struct request *rq = drv_tags->rqs[i];
2742 unsigned long rq_addr = (unsigned long)rq;
2744 if (rq_addr >= start && rq_addr < end) {
2745 WARN_ON_ONCE(refcount_read(&rq->ref) != 0);
2746 cmpxchg(&drv_tags->rqs[i], rq, NULL);
2752 * Wait until all pending iteration is done.
2754 * Request reference is cleared and it is guaranteed to be observed
2755 * after the ->lock is released.
2757 spin_lock_irqsave(&drv_tags->lock, flags);
2758 spin_unlock_irqrestore(&drv_tags->lock, flags);
2761 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2762 unsigned int hctx_idx)
2764 struct blk_mq_tags *drv_tags;
2767 if (blk_mq_is_shared_tags(set->flags))
2768 drv_tags = set->shared_tags;
2770 drv_tags = set->tags[hctx_idx];
2772 if (tags->static_rqs && set->ops->exit_request) {
2775 for (i = 0; i < tags->nr_tags; i++) {
2776 struct request *rq = tags->static_rqs[i];
2780 set->ops->exit_request(set, rq, hctx_idx);
2781 tags->static_rqs[i] = NULL;
2785 blk_mq_clear_rq_mapping(drv_tags, tags);
2787 while (!list_empty(&tags->page_list)) {
2788 page = list_first_entry(&tags->page_list, struct page, lru);
2789 list_del_init(&page->lru);
2791 * Remove kmemleak object previously allocated in
2792 * blk_mq_alloc_rqs().
2794 kmemleak_free(page_address(page));
2795 __free_pages(page, page->private);
2799 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2803 kfree(tags->static_rqs);
2804 tags->static_rqs = NULL;
2806 blk_mq_free_tags(tags);
2809 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2810 unsigned int hctx_idx,
2811 unsigned int nr_tags,
2812 unsigned int reserved_tags)
2814 struct blk_mq_tags *tags;
2817 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2818 if (node == NUMA_NO_NODE)
2819 node = set->numa_node;
2821 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2822 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2826 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2827 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2830 blk_mq_free_tags(tags);
2834 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2835 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2837 if (!tags->static_rqs) {
2839 blk_mq_free_tags(tags);
2846 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2847 unsigned int hctx_idx, int node)
2851 if (set->ops->init_request) {
2852 ret = set->ops->init_request(set, rq, hctx_idx, node);
2857 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2861 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
2862 struct blk_mq_tags *tags,
2863 unsigned int hctx_idx, unsigned int depth)
2865 unsigned int i, j, entries_per_page, max_order = 4;
2866 size_t rq_size, left;
2869 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2870 if (node == NUMA_NO_NODE)
2871 node = set->numa_node;
2873 INIT_LIST_HEAD(&tags->page_list);
2876 * rq_size is the size of the request plus driver payload, rounded
2877 * to the cacheline size
2879 rq_size = round_up(sizeof(struct request) + set->cmd_size,
2881 left = rq_size * depth;
2883 for (i = 0; i < depth; ) {
2884 int this_order = max_order;
2889 while (this_order && left < order_to_size(this_order - 1))
2893 page = alloc_pages_node(node,
2894 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2900 if (order_to_size(this_order) < rq_size)
2907 page->private = this_order;
2908 list_add_tail(&page->lru, &tags->page_list);
2910 p = page_address(page);
2912 * Allow kmemleak to scan these pages as they contain pointers
2913 * to additional allocations like via ops->init_request().
2915 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2916 entries_per_page = order_to_size(this_order) / rq_size;
2917 to_do = min(entries_per_page, depth - i);
2918 left -= to_do * rq_size;
2919 for (j = 0; j < to_do; j++) {
2920 struct request *rq = p;
2922 tags->static_rqs[i] = rq;
2923 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2924 tags->static_rqs[i] = NULL;
2935 blk_mq_free_rqs(set, tags, hctx_idx);
2939 struct rq_iter_data {
2940 struct blk_mq_hw_ctx *hctx;
2944 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
2946 struct rq_iter_data *iter_data = data;
2948 if (rq->mq_hctx != iter_data->hctx)
2950 iter_data->has_rq = true;
2954 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
2956 struct blk_mq_tags *tags = hctx->sched_tags ?
2957 hctx->sched_tags : hctx->tags;
2958 struct rq_iter_data data = {
2962 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
2966 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
2967 struct blk_mq_hw_ctx *hctx)
2969 if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu)
2971 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
2976 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
2978 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
2979 struct blk_mq_hw_ctx, cpuhp_online);
2981 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
2982 !blk_mq_last_cpu_in_hctx(cpu, hctx))
2986 * Prevent new request from being allocated on the current hctx.
2988 * The smp_mb__after_atomic() Pairs with the implied barrier in
2989 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
2990 * seen once we return from the tag allocator.
2992 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
2993 smp_mb__after_atomic();
2996 * Try to grab a reference to the queue and wait for any outstanding
2997 * requests. If we could not grab a reference the queue has been
2998 * frozen and there are no requests.
3000 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3001 while (blk_mq_hctx_has_requests(hctx))
3003 percpu_ref_put(&hctx->queue->q_usage_counter);
3009 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3011 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3012 struct blk_mq_hw_ctx, cpuhp_online);
3014 if (cpumask_test_cpu(cpu, hctx->cpumask))
3015 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3020 * 'cpu' is going away. splice any existing rq_list entries from this
3021 * software queue to the hw queue dispatch list, and ensure that it
3024 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3026 struct blk_mq_hw_ctx *hctx;
3027 struct blk_mq_ctx *ctx;
3029 enum hctx_type type;
3031 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3032 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3035 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3038 spin_lock(&ctx->lock);
3039 if (!list_empty(&ctx->rq_lists[type])) {
3040 list_splice_init(&ctx->rq_lists[type], &tmp);
3041 blk_mq_hctx_clear_pending(hctx, ctx);
3043 spin_unlock(&ctx->lock);
3045 if (list_empty(&tmp))
3048 spin_lock(&hctx->lock);
3049 list_splice_tail_init(&tmp, &hctx->dispatch);
3050 spin_unlock(&hctx->lock);
3052 blk_mq_run_hw_queue(hctx, true);
3056 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3058 if (!(hctx->flags & BLK_MQ_F_STACKING))
3059 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3060 &hctx->cpuhp_online);
3061 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3066 * Before freeing hw queue, clearing the flush request reference in
3067 * tags->rqs[] for avoiding potential UAF.
3069 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3070 unsigned int queue_depth, struct request *flush_rq)
3073 unsigned long flags;
3075 /* The hw queue may not be mapped yet */
3079 WARN_ON_ONCE(refcount_read(&flush_rq->ref) != 0);
3081 for (i = 0; i < queue_depth; i++)
3082 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3085 * Wait until all pending iteration is done.
3087 * Request reference is cleared and it is guaranteed to be observed
3088 * after the ->lock is released.
3090 spin_lock_irqsave(&tags->lock, flags);
3091 spin_unlock_irqrestore(&tags->lock, flags);
3094 /* hctx->ctxs will be freed in queue's release handler */
3095 static void blk_mq_exit_hctx(struct request_queue *q,
3096 struct blk_mq_tag_set *set,
3097 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3099 struct request *flush_rq = hctx->fq->flush_rq;
3101 if (blk_mq_hw_queue_mapped(hctx))
3102 blk_mq_tag_idle(hctx);
3104 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3105 set->queue_depth, flush_rq);
3106 if (set->ops->exit_request)
3107 set->ops->exit_request(set, flush_rq, hctx_idx);
3109 if (set->ops->exit_hctx)
3110 set->ops->exit_hctx(hctx, hctx_idx);
3112 blk_mq_remove_cpuhp(hctx);
3114 spin_lock(&q->unused_hctx_lock);
3115 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3116 spin_unlock(&q->unused_hctx_lock);
3119 static void blk_mq_exit_hw_queues(struct request_queue *q,
3120 struct blk_mq_tag_set *set, int nr_queue)
3122 struct blk_mq_hw_ctx *hctx;
3125 queue_for_each_hw_ctx(q, hctx, i) {
3128 blk_mq_debugfs_unregister_hctx(hctx);
3129 blk_mq_exit_hctx(q, set, hctx, i);
3133 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
3135 int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
3137 BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
3138 __alignof__(struct blk_mq_hw_ctx)) !=
3139 sizeof(struct blk_mq_hw_ctx));
3141 if (tag_set->flags & BLK_MQ_F_BLOCKING)
3142 hw_ctx_size += sizeof(struct srcu_struct);
3147 static int blk_mq_init_hctx(struct request_queue *q,
3148 struct blk_mq_tag_set *set,
3149 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3151 hctx->queue_num = hctx_idx;
3153 if (!(hctx->flags & BLK_MQ_F_STACKING))
3154 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3155 &hctx->cpuhp_online);
3156 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3158 hctx->tags = set->tags[hctx_idx];
3160 if (set->ops->init_hctx &&
3161 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3162 goto unregister_cpu_notifier;
3164 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3170 if (set->ops->exit_hctx)
3171 set->ops->exit_hctx(hctx, hctx_idx);
3172 unregister_cpu_notifier:
3173 blk_mq_remove_cpuhp(hctx);
3177 static struct blk_mq_hw_ctx *
3178 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3181 struct blk_mq_hw_ctx *hctx;
3182 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3184 hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
3186 goto fail_alloc_hctx;
3188 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3191 atomic_set(&hctx->nr_active, 0);
3192 if (node == NUMA_NO_NODE)
3193 node = set->numa_node;
3194 hctx->numa_node = node;
3196 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3197 spin_lock_init(&hctx->lock);
3198 INIT_LIST_HEAD(&hctx->dispatch);
3200 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3202 INIT_LIST_HEAD(&hctx->hctx_list);
3205 * Allocate space for all possible cpus to avoid allocation at
3208 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3213 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3214 gfp, node, false, false))
3218 spin_lock_init(&hctx->dispatch_wait_lock);
3219 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3220 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3222 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3226 if (hctx->flags & BLK_MQ_F_BLOCKING)
3227 init_srcu_struct(hctx->srcu);
3228 blk_mq_hctx_kobj_init(hctx);
3233 sbitmap_free(&hctx->ctx_map);
3237 free_cpumask_var(hctx->cpumask);
3244 static void blk_mq_init_cpu_queues(struct request_queue *q,
3245 unsigned int nr_hw_queues)
3247 struct blk_mq_tag_set *set = q->tag_set;
3250 for_each_possible_cpu(i) {
3251 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3252 struct blk_mq_hw_ctx *hctx;
3256 spin_lock_init(&__ctx->lock);
3257 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3258 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3263 * Set local node, IFF we have more than one hw queue. If
3264 * not, we remain on the home node of the device
3266 for (j = 0; j < set->nr_maps; j++) {
3267 hctx = blk_mq_map_queue_type(q, j, i);
3268 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3269 hctx->numa_node = cpu_to_node(i);
3274 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3275 unsigned int hctx_idx,
3278 struct blk_mq_tags *tags;
3281 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3285 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3287 blk_mq_free_rq_map(tags);
3294 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3297 if (blk_mq_is_shared_tags(set->flags)) {
3298 set->tags[hctx_idx] = set->shared_tags;
3303 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3306 return set->tags[hctx_idx];
3309 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3310 struct blk_mq_tags *tags,
3311 unsigned int hctx_idx)
3314 blk_mq_free_rqs(set, tags, hctx_idx);
3315 blk_mq_free_rq_map(tags);
3319 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3320 unsigned int hctx_idx)
3322 if (!blk_mq_is_shared_tags(set->flags))
3323 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3325 set->tags[hctx_idx] = NULL;
3328 static void blk_mq_map_swqueue(struct request_queue *q)
3330 unsigned int i, j, hctx_idx;
3331 struct blk_mq_hw_ctx *hctx;
3332 struct blk_mq_ctx *ctx;
3333 struct blk_mq_tag_set *set = q->tag_set;
3335 queue_for_each_hw_ctx(q, hctx, i) {
3336 cpumask_clear(hctx->cpumask);
3338 hctx->dispatch_from = NULL;
3342 * Map software to hardware queues.
3344 * If the cpu isn't present, the cpu is mapped to first hctx.
3346 for_each_possible_cpu(i) {
3348 ctx = per_cpu_ptr(q->queue_ctx, i);
3349 for (j = 0; j < set->nr_maps; j++) {
3350 if (!set->map[j].nr_queues) {
3351 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3352 HCTX_TYPE_DEFAULT, i);
3355 hctx_idx = set->map[j].mq_map[i];
3356 /* unmapped hw queue can be remapped after CPU topo changed */
3357 if (!set->tags[hctx_idx] &&
3358 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3360 * If tags initialization fail for some hctx,
3361 * that hctx won't be brought online. In this
3362 * case, remap the current ctx to hctx[0] which
3363 * is guaranteed to always have tags allocated
3365 set->map[j].mq_map[i] = 0;
3368 hctx = blk_mq_map_queue_type(q, j, i);
3369 ctx->hctxs[j] = hctx;
3371 * If the CPU is already set in the mask, then we've
3372 * mapped this one already. This can happen if
3373 * devices share queues across queue maps.
3375 if (cpumask_test_cpu(i, hctx->cpumask))
3378 cpumask_set_cpu(i, hctx->cpumask);
3380 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3381 hctx->ctxs[hctx->nr_ctx++] = ctx;
3384 * If the nr_ctx type overflows, we have exceeded the
3385 * amount of sw queues we can support.
3387 BUG_ON(!hctx->nr_ctx);
3390 for (; j < HCTX_MAX_TYPES; j++)
3391 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3392 HCTX_TYPE_DEFAULT, i);
3395 queue_for_each_hw_ctx(q, hctx, i) {
3397 * If no software queues are mapped to this hardware queue,
3398 * disable it and free the request entries.
3400 if (!hctx->nr_ctx) {
3401 /* Never unmap queue 0. We need it as a
3402 * fallback in case of a new remap fails
3406 __blk_mq_free_map_and_rqs(set, i);
3412 hctx->tags = set->tags[i];
3413 WARN_ON(!hctx->tags);
3416 * Set the map size to the number of mapped software queues.
3417 * This is more accurate and more efficient than looping
3418 * over all possibly mapped software queues.
3420 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3423 * Initialize batch roundrobin counts
3425 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3426 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3431 * Caller needs to ensure that we're either frozen/quiesced, or that
3432 * the queue isn't live yet.
3434 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3436 struct blk_mq_hw_ctx *hctx;
3439 queue_for_each_hw_ctx(q, hctx, i) {
3441 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3443 blk_mq_tag_idle(hctx);
3444 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3449 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3452 struct request_queue *q;
3454 lockdep_assert_held(&set->tag_list_lock);
3456 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3457 blk_mq_freeze_queue(q);
3458 queue_set_hctx_shared(q, shared);
3459 blk_mq_unfreeze_queue(q);
3463 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3465 struct blk_mq_tag_set *set = q->tag_set;
3467 mutex_lock(&set->tag_list_lock);
3468 list_del(&q->tag_set_list);
3469 if (list_is_singular(&set->tag_list)) {
3470 /* just transitioned to unshared */
3471 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3472 /* update existing queue */
3473 blk_mq_update_tag_set_shared(set, false);
3475 mutex_unlock(&set->tag_list_lock);
3476 INIT_LIST_HEAD(&q->tag_set_list);
3479 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3480 struct request_queue *q)
3482 mutex_lock(&set->tag_list_lock);
3485 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3487 if (!list_empty(&set->tag_list) &&
3488 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3489 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3490 /* update existing queue */
3491 blk_mq_update_tag_set_shared(set, true);
3493 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3494 queue_set_hctx_shared(q, true);
3495 list_add_tail(&q->tag_set_list, &set->tag_list);
3497 mutex_unlock(&set->tag_list_lock);
3500 /* All allocations will be freed in release handler of q->mq_kobj */
3501 static int blk_mq_alloc_ctxs(struct request_queue *q)
3503 struct blk_mq_ctxs *ctxs;
3506 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3510 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3511 if (!ctxs->queue_ctx)
3514 for_each_possible_cpu(cpu) {
3515 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3519 q->mq_kobj = &ctxs->kobj;
3520 q->queue_ctx = ctxs->queue_ctx;
3529 * It is the actual release handler for mq, but we do it from
3530 * request queue's release handler for avoiding use-after-free
3531 * and headache because q->mq_kobj shouldn't have been introduced,
3532 * but we can't group ctx/kctx kobj without it.
3534 void blk_mq_release(struct request_queue *q)
3536 struct blk_mq_hw_ctx *hctx, *next;
3539 queue_for_each_hw_ctx(q, hctx, i)
3540 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3542 /* all hctx are in .unused_hctx_list now */
3543 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3544 list_del_init(&hctx->hctx_list);
3545 kobject_put(&hctx->kobj);
3548 kfree(q->queue_hw_ctx);
3551 * release .mq_kobj and sw queue's kobject now because
3552 * both share lifetime with request queue.
3554 blk_mq_sysfs_deinit(q);
3557 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3560 struct request_queue *q;
3563 q = blk_alloc_queue(set->numa_node);
3565 return ERR_PTR(-ENOMEM);
3566 q->queuedata = queuedata;
3567 ret = blk_mq_init_allocated_queue(set, q);
3569 blk_cleanup_queue(q);
3570 return ERR_PTR(ret);
3575 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3577 return blk_mq_init_queue_data(set, NULL);
3579 EXPORT_SYMBOL(blk_mq_init_queue);
3581 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
3582 struct lock_class_key *lkclass)
3584 struct request_queue *q;
3585 struct gendisk *disk;
3587 q = blk_mq_init_queue_data(set, queuedata);
3591 disk = __alloc_disk_node(q, set->numa_node, lkclass);
3593 blk_cleanup_queue(q);
3594 return ERR_PTR(-ENOMEM);
3598 EXPORT_SYMBOL(__blk_mq_alloc_disk);
3600 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3601 struct blk_mq_tag_set *set, struct request_queue *q,
3602 int hctx_idx, int node)
3604 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3606 /* reuse dead hctx first */
3607 spin_lock(&q->unused_hctx_lock);
3608 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3609 if (tmp->numa_node == node) {
3615 list_del_init(&hctx->hctx_list);
3616 spin_unlock(&q->unused_hctx_lock);
3619 hctx = blk_mq_alloc_hctx(q, set, node);
3623 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3629 kobject_put(&hctx->kobj);
3634 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3635 struct request_queue *q)
3638 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3640 if (q->nr_hw_queues < set->nr_hw_queues) {
3641 struct blk_mq_hw_ctx **new_hctxs;
3643 new_hctxs = kcalloc_node(set->nr_hw_queues,
3644 sizeof(*new_hctxs), GFP_KERNEL,
3649 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3651 q->queue_hw_ctx = new_hctxs;
3656 /* protect against switching io scheduler */
3657 mutex_lock(&q->sysfs_lock);
3658 for (i = 0; i < set->nr_hw_queues; i++) {
3660 struct blk_mq_hw_ctx *hctx;
3662 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3664 * If the hw queue has been mapped to another numa node,
3665 * we need to realloc the hctx. If allocation fails, fallback
3666 * to use the previous one.
3668 if (hctxs[i] && (hctxs[i]->numa_node == node))
3671 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3674 blk_mq_exit_hctx(q, set, hctxs[i], i);
3678 pr_warn("Allocate new hctx on node %d fails,\
3679 fallback to previous one on node %d\n",
3680 node, hctxs[i]->numa_node);
3686 * Increasing nr_hw_queues fails. Free the newly allocated
3687 * hctxs and keep the previous q->nr_hw_queues.
3689 if (i != set->nr_hw_queues) {
3690 j = q->nr_hw_queues;
3694 end = q->nr_hw_queues;
3695 q->nr_hw_queues = set->nr_hw_queues;
3698 for (; j < end; j++) {
3699 struct blk_mq_hw_ctx *hctx = hctxs[j];
3702 blk_mq_exit_hctx(q, set, hctx, j);
3706 mutex_unlock(&q->sysfs_lock);
3709 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
3710 struct request_queue *q)
3712 /* mark the queue as mq asap */
3713 q->mq_ops = set->ops;
3715 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
3716 blk_mq_poll_stats_bkt,
3717 BLK_MQ_POLL_STATS_BKTS, q);
3721 if (blk_mq_alloc_ctxs(q))
3724 /* init q->mq_kobj and sw queues' kobjects */
3725 blk_mq_sysfs_init(q);
3727 INIT_LIST_HEAD(&q->unused_hctx_list);
3728 spin_lock_init(&q->unused_hctx_lock);
3730 blk_mq_realloc_hw_ctxs(set, q);
3731 if (!q->nr_hw_queues)
3734 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
3735 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
3739 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
3740 if (set->nr_maps > HCTX_TYPE_POLL &&
3741 set->map[HCTX_TYPE_POLL].nr_queues)
3742 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
3744 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
3745 INIT_LIST_HEAD(&q->requeue_list);
3746 spin_lock_init(&q->requeue_lock);
3748 q->nr_requests = set->queue_depth;
3751 * Default to classic polling
3753 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
3755 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
3756 blk_mq_add_queue_tag_set(set, q);
3757 blk_mq_map_swqueue(q);
3761 kfree(q->queue_hw_ctx);
3762 q->nr_hw_queues = 0;
3763 blk_mq_sysfs_deinit(q);
3765 blk_stat_free_callback(q->poll_cb);
3771 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
3773 /* tags can _not_ be used after returning from blk_mq_exit_queue */
3774 void blk_mq_exit_queue(struct request_queue *q)
3776 struct blk_mq_tag_set *set = q->tag_set;
3778 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
3779 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
3780 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
3781 blk_mq_del_queue_tag_set(q);
3784 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3788 if (blk_mq_is_shared_tags(set->flags)) {
3789 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
3792 if (!set->shared_tags)
3796 for (i = 0; i < set->nr_hw_queues; i++) {
3797 if (!__blk_mq_alloc_map_and_rqs(set, i))
3806 __blk_mq_free_map_and_rqs(set, i);
3808 if (blk_mq_is_shared_tags(set->flags)) {
3809 blk_mq_free_map_and_rqs(set, set->shared_tags,
3810 BLK_MQ_NO_HCTX_IDX);
3817 * Allocate the request maps associated with this tag_set. Note that this
3818 * may reduce the depth asked for, if memory is tight. set->queue_depth
3819 * will be updated to reflect the allocated depth.
3821 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
3826 depth = set->queue_depth;
3828 err = __blk_mq_alloc_rq_maps(set);
3832 set->queue_depth >>= 1;
3833 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3837 } while (set->queue_depth);
3839 if (!set->queue_depth || err) {
3840 pr_err("blk-mq: failed to allocate request map\n");
3844 if (depth != set->queue_depth)
3845 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3846 depth, set->queue_depth);
3851 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3854 * blk_mq_map_queues() and multiple .map_queues() implementations
3855 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3856 * number of hardware queues.
3858 if (set->nr_maps == 1)
3859 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3861 if (set->ops->map_queues && !is_kdump_kernel()) {
3865 * transport .map_queues is usually done in the following
3868 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3869 * mask = get_cpu_mask(queue)
3870 * for_each_cpu(cpu, mask)
3871 * set->map[x].mq_map[cpu] = queue;
3874 * When we need to remap, the table has to be cleared for
3875 * killing stale mapping since one CPU may not be mapped
3878 for (i = 0; i < set->nr_maps; i++)
3879 blk_mq_clear_mq_map(&set->map[i]);
3881 return set->ops->map_queues(set);
3883 BUG_ON(set->nr_maps > 1);
3884 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3888 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
3889 int cur_nr_hw_queues, int new_nr_hw_queues)
3891 struct blk_mq_tags **new_tags;
3893 if (cur_nr_hw_queues >= new_nr_hw_queues)
3896 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3897 GFP_KERNEL, set->numa_node);
3902 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3903 sizeof(*set->tags));
3905 set->tags = new_tags;
3906 set->nr_hw_queues = new_nr_hw_queues;
3911 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
3912 int new_nr_hw_queues)
3914 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
3918 * Alloc a tag set to be associated with one or more request queues.
3919 * May fail with EINVAL for various error conditions. May adjust the
3920 * requested depth down, if it's too large. In that case, the set
3921 * value will be stored in set->queue_depth.
3923 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3927 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3929 if (!set->nr_hw_queues)
3931 if (!set->queue_depth)
3933 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3936 if (!set->ops->queue_rq)
3939 if (!set->ops->get_budget ^ !set->ops->put_budget)
3942 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3943 pr_info("blk-mq: reduced tag depth to %u\n",
3945 set->queue_depth = BLK_MQ_MAX_DEPTH;
3950 else if (set->nr_maps > HCTX_MAX_TYPES)
3954 * If a crashdump is active, then we are potentially in a very
3955 * memory constrained environment. Limit us to 1 queue and
3956 * 64 tags to prevent using too much memory.
3958 if (is_kdump_kernel()) {
3959 set->nr_hw_queues = 1;
3961 set->queue_depth = min(64U, set->queue_depth);
3964 * There is no use for more h/w queues than cpus if we just have
3967 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3968 set->nr_hw_queues = nr_cpu_ids;
3970 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
3974 for (i = 0; i < set->nr_maps; i++) {
3975 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3976 sizeof(set->map[i].mq_map[0]),
3977 GFP_KERNEL, set->numa_node);
3978 if (!set->map[i].mq_map)
3979 goto out_free_mq_map;
3980 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3983 ret = blk_mq_update_queue_map(set);
3985 goto out_free_mq_map;
3987 ret = blk_mq_alloc_set_map_and_rqs(set);
3989 goto out_free_mq_map;
3991 mutex_init(&set->tag_list_lock);
3992 INIT_LIST_HEAD(&set->tag_list);
3997 for (i = 0; i < set->nr_maps; i++) {
3998 kfree(set->map[i].mq_map);
3999 set->map[i].mq_map = NULL;
4005 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4007 /* allocate and initialize a tagset for a simple single-queue device */
4008 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4009 const struct blk_mq_ops *ops, unsigned int queue_depth,
4010 unsigned int set_flags)
4012 memset(set, 0, sizeof(*set));
4014 set->nr_hw_queues = 1;
4016 set->queue_depth = queue_depth;
4017 set->numa_node = NUMA_NO_NODE;
4018 set->flags = set_flags;
4019 return blk_mq_alloc_tag_set(set);
4021 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4023 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4027 for (i = 0; i < set->nr_hw_queues; i++)
4028 __blk_mq_free_map_and_rqs(set, i);
4030 if (blk_mq_is_shared_tags(set->flags)) {
4031 blk_mq_free_map_and_rqs(set, set->shared_tags,
4032 BLK_MQ_NO_HCTX_IDX);
4035 for (j = 0; j < set->nr_maps; j++) {
4036 kfree(set->map[j].mq_map);
4037 set->map[j].mq_map = NULL;
4043 EXPORT_SYMBOL(blk_mq_free_tag_set);
4045 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4047 struct blk_mq_tag_set *set = q->tag_set;
4048 struct blk_mq_hw_ctx *hctx;
4054 if (q->nr_requests == nr)
4057 blk_mq_freeze_queue(q);
4058 blk_mq_quiesce_queue(q);
4061 queue_for_each_hw_ctx(q, hctx, i) {
4065 * If we're using an MQ scheduler, just update the scheduler
4066 * queue depth. This is similar to what the old code would do.
4068 if (hctx->sched_tags) {
4069 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4072 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4077 if (q->elevator && q->elevator->type->ops.depth_updated)
4078 q->elevator->type->ops.depth_updated(hctx);
4081 q->nr_requests = nr;
4082 if (blk_mq_is_shared_tags(set->flags)) {
4084 blk_mq_tag_update_sched_shared_tags(q);
4086 blk_mq_tag_resize_shared_tags(set, nr);
4090 blk_mq_unquiesce_queue(q);
4091 blk_mq_unfreeze_queue(q);
4097 * request_queue and elevator_type pair.
4098 * It is just used by __blk_mq_update_nr_hw_queues to cache
4099 * the elevator_type associated with a request_queue.
4101 struct blk_mq_qe_pair {
4102 struct list_head node;
4103 struct request_queue *q;
4104 struct elevator_type *type;
4108 * Cache the elevator_type in qe pair list and switch the
4109 * io scheduler to 'none'
4111 static bool blk_mq_elv_switch_none(struct list_head *head,
4112 struct request_queue *q)
4114 struct blk_mq_qe_pair *qe;
4119 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4123 INIT_LIST_HEAD(&qe->node);
4125 qe->type = q->elevator->type;
4126 list_add(&qe->node, head);
4128 mutex_lock(&q->sysfs_lock);
4130 * After elevator_switch_mq, the previous elevator_queue will be
4131 * released by elevator_release. The reference of the io scheduler
4132 * module get by elevator_get will also be put. So we need to get
4133 * a reference of the io scheduler module here to prevent it to be
4136 __module_get(qe->type->elevator_owner);
4137 elevator_switch_mq(q, NULL);
4138 mutex_unlock(&q->sysfs_lock);
4143 static void blk_mq_elv_switch_back(struct list_head *head,
4144 struct request_queue *q)
4146 struct blk_mq_qe_pair *qe;
4147 struct elevator_type *t = NULL;
4149 list_for_each_entry(qe, head, node)
4158 list_del(&qe->node);
4161 mutex_lock(&q->sysfs_lock);
4162 elevator_switch_mq(q, t);
4163 mutex_unlock(&q->sysfs_lock);
4166 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4169 struct request_queue *q;
4171 int prev_nr_hw_queues;
4173 lockdep_assert_held(&set->tag_list_lock);
4175 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4176 nr_hw_queues = nr_cpu_ids;
4177 if (nr_hw_queues < 1)
4179 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4182 list_for_each_entry(q, &set->tag_list, tag_set_list)
4183 blk_mq_freeze_queue(q);
4185 * Switch IO scheduler to 'none', cleaning up the data associated
4186 * with the previous scheduler. We will switch back once we are done
4187 * updating the new sw to hw queue mappings.
4189 list_for_each_entry(q, &set->tag_list, tag_set_list)
4190 if (!blk_mq_elv_switch_none(&head, q))
4193 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4194 blk_mq_debugfs_unregister_hctxs(q);
4195 blk_mq_sysfs_unregister(q);
4198 prev_nr_hw_queues = set->nr_hw_queues;
4199 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
4203 set->nr_hw_queues = nr_hw_queues;
4205 blk_mq_update_queue_map(set);
4206 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4207 blk_mq_realloc_hw_ctxs(set, q);
4208 if (q->nr_hw_queues != set->nr_hw_queues) {
4209 int i = prev_nr_hw_queues;
4211 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4212 nr_hw_queues, prev_nr_hw_queues);
4213 for (; i < set->nr_hw_queues; i++)
4214 __blk_mq_free_map_and_rqs(set, i);
4216 set->nr_hw_queues = prev_nr_hw_queues;
4217 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4220 blk_mq_map_swqueue(q);
4224 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4225 blk_mq_sysfs_register(q);
4226 blk_mq_debugfs_register_hctxs(q);
4230 list_for_each_entry(q, &set->tag_list, tag_set_list)
4231 blk_mq_elv_switch_back(&head, q);
4233 list_for_each_entry(q, &set->tag_list, tag_set_list)
4234 blk_mq_unfreeze_queue(q);
4237 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4239 mutex_lock(&set->tag_list_lock);
4240 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4241 mutex_unlock(&set->tag_list_lock);
4243 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4245 /* Enable polling stats and return whether they were already enabled. */
4246 static bool blk_poll_stats_enable(struct request_queue *q)
4248 if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
4249 blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
4251 blk_stat_add_callback(q, q->poll_cb);
4255 static void blk_mq_poll_stats_start(struct request_queue *q)
4258 * We don't arm the callback if polling stats are not enabled or the
4259 * callback is already active.
4261 if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
4262 blk_stat_is_active(q->poll_cb))
4265 blk_stat_activate_msecs(q->poll_cb, 100);
4268 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4270 struct request_queue *q = cb->data;
4273 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4274 if (cb->stat[bucket].nr_samples)
4275 q->poll_stat[bucket] = cb->stat[bucket];
4279 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4282 unsigned long ret = 0;
4286 * If stats collection isn't on, don't sleep but turn it on for
4289 if (!blk_poll_stats_enable(q))
4293 * As an optimistic guess, use half of the mean service time
4294 * for this type of request. We can (and should) make this smarter.
4295 * For instance, if the completion latencies are tight, we can
4296 * get closer than just half the mean. This is especially
4297 * important on devices where the completion latencies are longer
4298 * than ~10 usec. We do use the stats for the relevant IO size
4299 * if available which does lead to better estimates.
4301 bucket = blk_mq_poll_stats_bkt(rq);
4305 if (q->poll_stat[bucket].nr_samples)
4306 ret = (q->poll_stat[bucket].mean + 1) / 2;
4311 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4313 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4314 struct request *rq = blk_qc_to_rq(hctx, qc);
4315 struct hrtimer_sleeper hs;
4316 enum hrtimer_mode mode;
4321 * If a request has completed on queue that uses an I/O scheduler, we
4322 * won't get back a request from blk_qc_to_rq.
4324 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4328 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4330 * 0: use half of prev avg
4331 * >0: use this specific value
4333 if (q->poll_nsec > 0)
4334 nsecs = q->poll_nsec;
4336 nsecs = blk_mq_poll_nsecs(q, rq);
4341 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4344 * This will be replaced with the stats tracking code, using
4345 * 'avg_completion_time / 2' as the pre-sleep target.
4349 mode = HRTIMER_MODE_REL;
4350 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4351 hrtimer_set_expires(&hs.timer, kt);
4354 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4356 set_current_state(TASK_UNINTERRUPTIBLE);
4357 hrtimer_sleeper_start_expires(&hs, mode);
4360 hrtimer_cancel(&hs.timer);
4361 mode = HRTIMER_MODE_ABS;
4362 } while (hs.task && !signal_pending(current));
4364 __set_current_state(TASK_RUNNING);
4365 destroy_hrtimer_on_stack(&hs.timer);
4368 * If we sleep, have the caller restart the poll loop to reset the
4369 * state. Like for the other success return cases, the caller is
4370 * responsible for checking if the IO completed. If the IO isn't
4371 * complete, we'll get called again and will go straight to the busy
4377 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4378 struct io_comp_batch *iob, unsigned int flags)
4380 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4381 long state = get_current_state();
4385 ret = q->mq_ops->poll(hctx, iob);
4387 __set_current_state(TASK_RUNNING);
4391 if (signal_pending_state(state, current))
4392 __set_current_state(TASK_RUNNING);
4393 if (task_is_running(current))
4396 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4399 } while (!need_resched());
4401 __set_current_state(TASK_RUNNING);
4405 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4408 if (!(flags & BLK_POLL_NOSLEEP) &&
4409 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4410 if (blk_mq_poll_hybrid(q, cookie))
4413 return blk_mq_poll_classic(q, cookie, iob, flags);
4416 unsigned int blk_mq_rq_cpu(struct request *rq)
4418 return rq->mq_ctx->cpu;
4420 EXPORT_SYMBOL(blk_mq_rq_cpu);
4422 void blk_mq_cancel_work_sync(struct request_queue *q)
4424 if (queue_is_mq(q)) {
4425 struct blk_mq_hw_ctx *hctx;
4428 cancel_delayed_work_sync(&q->requeue_work);
4430 queue_for_each_hw_ctx(q, hctx, i)
4431 cancel_delayed_work_sync(&hctx->run_work);
4435 static int __init blk_mq_init(void)
4439 for_each_possible_cpu(i)
4440 init_llist_head(&per_cpu(blk_cpu_done, i));
4441 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4443 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4444 "block/softirq:dead", NULL,
4445 blk_softirq_cpu_dead);
4446 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4447 blk_mq_hctx_notify_dead);
4448 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4449 blk_mq_hctx_notify_online,
4450 blk_mq_hctx_notify_offline);
4453 subsys_initcall(blk_mq_init);