1 // SPDX-License-Identifier: GPL-2.0
3 * Block multiqueue core code
5 * Copyright (C) 2013-2014 Jens Axboe
6 * Copyright (C) 2013-2014 Christoph Hellwig
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/backing-dev.h>
11 #include <linux/bio.h>
12 #include <linux/blkdev.h>
13 #include <linux/blk-integrity.h>
14 #include <linux/kmemleak.h>
16 #include <linux/init.h>
17 #include <linux/slab.h>
18 #include <linux/workqueue.h>
19 #include <linux/smp.h>
20 #include <linux/interrupt.h>
21 #include <linux/llist.h>
22 #include <linux/cpu.h>
23 #include <linux/cache.h>
24 #include <linux/sched/sysctl.h>
25 #include <linux/sched/topology.h>
26 #include <linux/sched/signal.h>
27 #include <linux/delay.h>
28 #include <linux/crash_dump.h>
29 #include <linux/prefetch.h>
30 #include <linux/blk-crypto.h>
31 #include <linux/part_stat.h>
33 #include <trace/events/block.h>
35 #include <linux/blk-mq.h>
36 #include <linux/t10-pi.h>
39 #include "blk-mq-debugfs.h"
40 #include "blk-mq-tag.h"
43 #include "blk-mq-sched.h"
44 #include "blk-rq-qos.h"
45 #include "blk-ioprio.h"
47 static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
49 static void blk_mq_poll_stats_start(struct request_queue *q);
50 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
52 static int blk_mq_poll_stats_bkt(const struct request *rq)
54 int ddir, sectors, bucket;
56 ddir = rq_data_dir(rq);
57 sectors = blk_rq_stats_sectors(rq);
59 bucket = ddir + 2 * ilog2(sectors);
63 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
64 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
69 #define BLK_QC_T_SHIFT 16
70 #define BLK_QC_T_INTERNAL (1U << 31)
72 static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q,
75 return xa_load(&q->hctx_table,
76 (qc & ~BLK_QC_T_INTERNAL) >> BLK_QC_T_SHIFT);
79 static inline struct request *blk_qc_to_rq(struct blk_mq_hw_ctx *hctx,
82 unsigned int tag = qc & ((1U << BLK_QC_T_SHIFT) - 1);
84 if (qc & BLK_QC_T_INTERNAL)
85 return blk_mq_tag_to_rq(hctx->sched_tags, tag);
86 return blk_mq_tag_to_rq(hctx->tags, tag);
89 static inline blk_qc_t blk_rq_to_qc(struct request *rq)
91 return (rq->mq_hctx->queue_num << BLK_QC_T_SHIFT) |
93 rq->tag : (rq->internal_tag | BLK_QC_T_INTERNAL));
97 * Check if any of the ctx, dispatch list or elevator
98 * have pending work in this hardware queue.
100 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
102 return !list_empty_careful(&hctx->dispatch) ||
103 sbitmap_any_bit_set(&hctx->ctx_map) ||
104 blk_mq_sched_has_work(hctx);
108 * Mark this ctx as having pending work in this hardware queue
110 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
111 struct blk_mq_ctx *ctx)
113 const int bit = ctx->index_hw[hctx->type];
115 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
116 sbitmap_set_bit(&hctx->ctx_map, bit);
119 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
120 struct blk_mq_ctx *ctx)
122 const int bit = ctx->index_hw[hctx->type];
124 sbitmap_clear_bit(&hctx->ctx_map, bit);
128 struct block_device *part;
129 unsigned int inflight[2];
132 static bool blk_mq_check_inflight(struct request *rq, void *priv)
134 struct mq_inflight *mi = priv;
136 if (rq->part && blk_do_io_stat(rq) &&
137 (!mi->part->bd_partno || rq->part == mi->part) &&
138 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
139 mi->inflight[rq_data_dir(rq)]++;
144 unsigned int blk_mq_in_flight(struct request_queue *q,
145 struct block_device *part)
147 struct mq_inflight mi = { .part = part };
149 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
151 return mi.inflight[0] + mi.inflight[1];
154 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
155 unsigned int inflight[2])
157 struct mq_inflight mi = { .part = part };
159 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
160 inflight[0] = mi.inflight[0];
161 inflight[1] = mi.inflight[1];
164 void blk_freeze_queue_start(struct request_queue *q)
166 mutex_lock(&q->mq_freeze_lock);
167 if (++q->mq_freeze_depth == 1) {
168 percpu_ref_kill(&q->q_usage_counter);
169 mutex_unlock(&q->mq_freeze_lock);
171 blk_mq_run_hw_queues(q, false);
173 mutex_unlock(&q->mq_freeze_lock);
176 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
178 void blk_mq_freeze_queue_wait(struct request_queue *q)
180 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
182 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
184 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
185 unsigned long timeout)
187 return wait_event_timeout(q->mq_freeze_wq,
188 percpu_ref_is_zero(&q->q_usage_counter),
191 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
194 * Guarantee no request is in use, so we can change any data structure of
195 * the queue afterward.
197 void blk_freeze_queue(struct request_queue *q)
200 * In the !blk_mq case we are only calling this to kill the
201 * q_usage_counter, otherwise this increases the freeze depth
202 * and waits for it to return to zero. For this reason there is
203 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
204 * exported to drivers as the only user for unfreeze is blk_mq.
206 blk_freeze_queue_start(q);
207 blk_mq_freeze_queue_wait(q);
210 void blk_mq_freeze_queue(struct request_queue *q)
213 * ...just an alias to keep freeze and unfreeze actions balanced
214 * in the blk_mq_* namespace
218 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
220 void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
222 mutex_lock(&q->mq_freeze_lock);
224 q->q_usage_counter.data->force_atomic = true;
225 q->mq_freeze_depth--;
226 WARN_ON_ONCE(q->mq_freeze_depth < 0);
227 if (!q->mq_freeze_depth) {
228 percpu_ref_resurrect(&q->q_usage_counter);
229 wake_up_all(&q->mq_freeze_wq);
231 mutex_unlock(&q->mq_freeze_lock);
234 void blk_mq_unfreeze_queue(struct request_queue *q)
236 __blk_mq_unfreeze_queue(q, false);
238 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
241 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
242 * mpt3sas driver such that this function can be removed.
244 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
248 spin_lock_irqsave(&q->queue_lock, flags);
249 if (!q->quiesce_depth++)
250 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
251 spin_unlock_irqrestore(&q->queue_lock, flags);
253 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
256 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
259 * Note: it is driver's responsibility for making sure that quiesce has
262 void blk_mq_wait_quiesce_done(struct request_queue *q)
264 if (blk_queue_has_srcu(q))
265 synchronize_srcu(q->srcu);
269 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
272 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
275 * Note: this function does not prevent that the struct request end_io()
276 * callback function is invoked. Once this function is returned, we make
277 * sure no dispatch can happen until the queue is unquiesced via
278 * blk_mq_unquiesce_queue().
280 void blk_mq_quiesce_queue(struct request_queue *q)
282 blk_mq_quiesce_queue_nowait(q);
283 blk_mq_wait_quiesce_done(q);
285 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
288 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
291 * This function recovers queue into the state before quiescing
292 * which is done by blk_mq_quiesce_queue.
294 void blk_mq_unquiesce_queue(struct request_queue *q)
297 bool run_queue = false;
299 spin_lock_irqsave(&q->queue_lock, flags);
300 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
302 } else if (!--q->quiesce_depth) {
303 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
306 spin_unlock_irqrestore(&q->queue_lock, flags);
308 /* dispatch requests which are inserted during quiescing */
310 blk_mq_run_hw_queues(q, true);
312 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
314 void blk_mq_wake_waiters(struct request_queue *q)
316 struct blk_mq_hw_ctx *hctx;
319 queue_for_each_hw_ctx(q, hctx, i)
320 if (blk_mq_hw_queue_mapped(hctx))
321 blk_mq_tag_wakeup_all(hctx->tags, true);
324 void blk_rq_init(struct request_queue *q, struct request *rq)
326 memset(rq, 0, sizeof(*rq));
328 INIT_LIST_HEAD(&rq->queuelist);
330 rq->__sector = (sector_t) -1;
331 INIT_HLIST_NODE(&rq->hash);
332 RB_CLEAR_NODE(&rq->rb_node);
333 rq->tag = BLK_MQ_NO_TAG;
334 rq->internal_tag = BLK_MQ_NO_TAG;
335 rq->start_time_ns = ktime_get_ns();
337 blk_crypto_rq_set_defaults(rq);
339 EXPORT_SYMBOL(blk_rq_init);
341 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
342 struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns)
344 struct blk_mq_ctx *ctx = data->ctx;
345 struct blk_mq_hw_ctx *hctx = data->hctx;
346 struct request_queue *q = data->q;
347 struct request *rq = tags->static_rqs[tag];
352 rq->cmd_flags = data->cmd_flags;
354 if (data->flags & BLK_MQ_REQ_PM)
355 data->rq_flags |= RQF_PM;
356 if (blk_queue_io_stat(q))
357 data->rq_flags |= RQF_IO_STAT;
358 rq->rq_flags = data->rq_flags;
360 if (!(data->rq_flags & RQF_ELV)) {
362 rq->internal_tag = BLK_MQ_NO_TAG;
364 rq->tag = BLK_MQ_NO_TAG;
365 rq->internal_tag = tag;
369 if (blk_mq_need_time_stamp(rq))
370 rq->start_time_ns = ktime_get_ns();
372 rq->start_time_ns = 0;
374 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
375 rq->alloc_time_ns = alloc_time_ns;
377 rq->io_start_time_ns = 0;
378 rq->stats_sectors = 0;
379 rq->nr_phys_segments = 0;
380 #if defined(CONFIG_BLK_DEV_INTEGRITY)
381 rq->nr_integrity_segments = 0;
384 rq->end_io_data = NULL;
386 blk_crypto_rq_set_defaults(rq);
387 INIT_LIST_HEAD(&rq->queuelist);
388 /* tag was already set */
389 WRITE_ONCE(rq->deadline, 0);
392 if (rq->rq_flags & RQF_ELV) {
393 struct elevator_queue *e = data->q->elevator;
395 INIT_HLIST_NODE(&rq->hash);
396 RB_CLEAR_NODE(&rq->rb_node);
398 if (!op_is_flush(data->cmd_flags) &&
399 e->type->ops.prepare_request) {
400 e->type->ops.prepare_request(rq);
401 rq->rq_flags |= RQF_ELVPRIV;
408 static inline struct request *
409 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data,
412 unsigned int tag, tag_offset;
413 struct blk_mq_tags *tags;
415 unsigned long tag_mask;
418 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
419 if (unlikely(!tag_mask))
422 tags = blk_mq_tags_from_data(data);
423 for (i = 0; tag_mask; i++) {
424 if (!(tag_mask & (1UL << i)))
426 tag = tag_offset + i;
427 prefetch(tags->static_rqs[tag]);
428 tag_mask &= ~(1UL << i);
429 rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns);
430 rq_list_add(data->cached_rq, rq);
433 /* caller already holds a reference, add for remainder */
434 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
437 return rq_list_pop(data->cached_rq);
440 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
442 struct request_queue *q = data->q;
443 u64 alloc_time_ns = 0;
447 /* alloc_time includes depth and tag waits */
448 if (blk_queue_rq_alloc_time(q))
449 alloc_time_ns = ktime_get_ns();
451 if (data->cmd_flags & REQ_NOWAIT)
452 data->flags |= BLK_MQ_REQ_NOWAIT;
455 struct elevator_queue *e = q->elevator;
457 data->rq_flags |= RQF_ELV;
460 * Flush/passthrough requests are special and go directly to the
461 * dispatch list. Don't include reserved tags in the
462 * limiting, as it isn't useful.
464 if (!op_is_flush(data->cmd_flags) &&
465 !blk_op_is_passthrough(data->cmd_flags) &&
466 e->type->ops.limit_depth &&
467 !(data->flags & BLK_MQ_REQ_RESERVED))
468 e->type->ops.limit_depth(data->cmd_flags, data);
472 data->ctx = blk_mq_get_ctx(q);
473 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
474 if (!(data->rq_flags & RQF_ELV))
475 blk_mq_tag_busy(data->hctx);
477 if (data->flags & BLK_MQ_REQ_RESERVED)
478 data->rq_flags |= RQF_RESV;
481 * Try batched alloc if we want more than 1 tag.
483 if (data->nr_tags > 1) {
484 rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns);
491 * Waiting allocations only fail because of an inactive hctx. In that
492 * case just retry the hctx assignment and tag allocation as CPU hotplug
493 * should have migrated us to an online CPU by now.
495 tag = blk_mq_get_tag(data);
496 if (tag == BLK_MQ_NO_TAG) {
497 if (data->flags & BLK_MQ_REQ_NOWAIT)
500 * Give up the CPU and sleep for a random short time to
501 * ensure that thread using a realtime scheduling class
502 * are migrated off the CPU, and thus off the hctx that
509 return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag,
513 static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
514 struct blk_plug *plug,
516 blk_mq_req_flags_t flags)
518 struct blk_mq_alloc_data data = {
522 .nr_tags = plug->nr_ios,
523 .cached_rq = &plug->cached_rq,
527 if (blk_queue_enter(q, flags))
532 rq = __blk_mq_alloc_requests(&data);
538 static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
540 blk_mq_req_flags_t flags)
542 struct blk_plug *plug = current->plug;
547 if (rq_list_empty(plug->cached_rq)) {
548 if (plug->nr_ios == 1)
550 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
555 rq = rq_list_peek(&plug->cached_rq);
556 if (!rq || rq->q != q)
559 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
561 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
564 plug->cached_rq = rq_list_next(rq);
567 INIT_LIST_HEAD(&rq->queuelist);
571 struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
572 blk_mq_req_flags_t flags)
576 rq = blk_mq_alloc_cached_request(q, opf, flags);
578 struct blk_mq_alloc_data data = {
586 ret = blk_queue_enter(q, flags);
590 rq = __blk_mq_alloc_requests(&data);
595 rq->__sector = (sector_t) -1;
596 rq->bio = rq->biotail = NULL;
600 return ERR_PTR(-EWOULDBLOCK);
602 EXPORT_SYMBOL(blk_mq_alloc_request);
604 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
605 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
607 struct blk_mq_alloc_data data = {
613 u64 alloc_time_ns = 0;
619 /* alloc_time includes depth and tag waits */
620 if (blk_queue_rq_alloc_time(q))
621 alloc_time_ns = ktime_get_ns();
624 * If the tag allocator sleeps we could get an allocation for a
625 * different hardware context. No need to complicate the low level
626 * allocator for this for the rare use case of a command tied to
629 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
630 return ERR_PTR(-EINVAL);
632 if (hctx_idx >= q->nr_hw_queues)
633 return ERR_PTR(-EIO);
635 ret = blk_queue_enter(q, flags);
640 * Check if the hardware context is actually mapped to anything.
641 * If not tell the caller that it should skip this queue.
644 data.hctx = xa_load(&q->hctx_table, hctx_idx);
645 if (!blk_mq_hw_queue_mapped(data.hctx))
647 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
648 if (cpu >= nr_cpu_ids)
650 data.ctx = __blk_mq_get_ctx(q, cpu);
653 blk_mq_tag_busy(data.hctx);
655 data.rq_flags |= RQF_ELV;
657 if (flags & BLK_MQ_REQ_RESERVED)
658 data.rq_flags |= RQF_RESV;
661 tag = blk_mq_get_tag(&data);
662 if (tag == BLK_MQ_NO_TAG)
664 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag,
667 rq->__sector = (sector_t) -1;
668 rq->bio = rq->biotail = NULL;
675 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
677 static void __blk_mq_free_request(struct request *rq)
679 struct request_queue *q = rq->q;
680 struct blk_mq_ctx *ctx = rq->mq_ctx;
681 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
682 const int sched_tag = rq->internal_tag;
684 blk_crypto_free_request(rq);
685 blk_pm_mark_last_busy(rq);
687 if (rq->tag != BLK_MQ_NO_TAG)
688 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
689 if (sched_tag != BLK_MQ_NO_TAG)
690 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
691 blk_mq_sched_restart(hctx);
695 void blk_mq_free_request(struct request *rq)
697 struct request_queue *q = rq->q;
698 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
700 if ((rq->rq_flags & RQF_ELVPRIV) &&
701 q->elevator->type->ops.finish_request)
702 q->elevator->type->ops.finish_request(rq);
704 if (rq->rq_flags & RQF_MQ_INFLIGHT)
705 __blk_mq_dec_active_requests(hctx);
707 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
708 laptop_io_completion(q->disk->bdi);
712 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
713 if (req_ref_put_and_test(rq))
714 __blk_mq_free_request(rq);
716 EXPORT_SYMBOL_GPL(blk_mq_free_request);
718 void blk_mq_free_plug_rqs(struct blk_plug *plug)
722 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
723 blk_mq_free_request(rq);
726 void blk_dump_rq_flags(struct request *rq, char *msg)
728 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
729 rq->q->disk ? rq->q->disk->disk_name : "?",
730 (__force unsigned long long) rq->cmd_flags);
732 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
733 (unsigned long long)blk_rq_pos(rq),
734 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
735 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
736 rq->bio, rq->biotail, blk_rq_bytes(rq));
738 EXPORT_SYMBOL(blk_dump_rq_flags);
740 static void req_bio_endio(struct request *rq, struct bio *bio,
741 unsigned int nbytes, blk_status_t error)
743 if (unlikely(error)) {
744 bio->bi_status = error;
745 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
747 * Partial zone append completions cannot be supported as the
748 * BIO fragments may end up not being written sequentially.
750 if (bio->bi_iter.bi_size != nbytes)
751 bio->bi_status = BLK_STS_IOERR;
753 bio->bi_iter.bi_sector = rq->__sector;
756 bio_advance(bio, nbytes);
758 if (unlikely(rq->rq_flags & RQF_QUIET))
759 bio_set_flag(bio, BIO_QUIET);
760 /* don't actually finish bio if it's part of flush sequence */
761 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
765 static void blk_account_io_completion(struct request *req, unsigned int bytes)
767 if (req->part && blk_do_io_stat(req)) {
768 const int sgrp = op_stat_group(req_op(req));
771 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
776 static void blk_print_req_error(struct request *req, blk_status_t status)
778 printk_ratelimited(KERN_ERR
779 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
780 "phys_seg %u prio class %u\n",
781 blk_status_to_str(status),
782 req->q->disk ? req->q->disk->disk_name : "?",
783 blk_rq_pos(req), (__force u32)req_op(req),
784 blk_op_str(req_op(req)),
785 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
786 req->nr_phys_segments,
787 IOPRIO_PRIO_CLASS(req->ioprio));
791 * Fully end IO on a request. Does not support partial completions, or
794 static void blk_complete_request(struct request *req)
796 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
797 int total_bytes = blk_rq_bytes(req);
798 struct bio *bio = req->bio;
800 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
805 #ifdef CONFIG_BLK_DEV_INTEGRITY
806 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
807 req->q->integrity.profile->complete_fn(req, total_bytes);
810 blk_account_io_completion(req, total_bytes);
813 struct bio *next = bio->bi_next;
815 /* Completion has already been traced */
816 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
818 if (req_op(req) == REQ_OP_ZONE_APPEND)
819 bio->bi_iter.bi_sector = req->__sector;
827 * Reset counters so that the request stacking driver
828 * can find how many bytes remain in the request
838 * blk_update_request - Complete multiple bytes without completing the request
839 * @req: the request being processed
840 * @error: block status code
841 * @nr_bytes: number of bytes to complete for @req
844 * Ends I/O on a number of bytes attached to @req, but doesn't complete
845 * the request structure even if @req doesn't have leftover.
846 * If @req has leftover, sets it up for the next range of segments.
848 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
849 * %false return from this function.
852 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
853 * except in the consistency check at the end of this function.
856 * %false - this request doesn't have any more data
857 * %true - this request has more data
859 bool blk_update_request(struct request *req, blk_status_t error,
860 unsigned int nr_bytes)
864 trace_block_rq_complete(req, error, nr_bytes);
869 #ifdef CONFIG_BLK_DEV_INTEGRITY
870 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
872 req->q->integrity.profile->complete_fn(req, nr_bytes);
875 if (unlikely(error && !blk_rq_is_passthrough(req) &&
876 !(req->rq_flags & RQF_QUIET)) &&
877 !test_bit(GD_DEAD, &req->q->disk->state)) {
878 blk_print_req_error(req, error);
879 trace_block_rq_error(req, error, nr_bytes);
882 blk_account_io_completion(req, nr_bytes);
886 struct bio *bio = req->bio;
887 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
889 if (bio_bytes == bio->bi_iter.bi_size)
890 req->bio = bio->bi_next;
892 /* Completion has already been traced */
893 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
894 req_bio_endio(req, bio, bio_bytes, error);
896 total_bytes += bio_bytes;
897 nr_bytes -= bio_bytes;
908 * Reset counters so that the request stacking driver
909 * can find how many bytes remain in the request
916 req->__data_len -= total_bytes;
918 /* update sector only for requests with clear definition of sector */
919 if (!blk_rq_is_passthrough(req))
920 req->__sector += total_bytes >> 9;
922 /* mixed attributes always follow the first bio */
923 if (req->rq_flags & RQF_MIXED_MERGE) {
924 req->cmd_flags &= ~REQ_FAILFAST_MASK;
925 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
928 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
930 * If total number of sectors is less than the first segment
931 * size, something has gone terribly wrong.
933 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
934 blk_dump_rq_flags(req, "request botched");
935 req->__data_len = blk_rq_cur_bytes(req);
938 /* recalculate the number of segments */
939 req->nr_phys_segments = blk_recalc_rq_segments(req);
944 EXPORT_SYMBOL_GPL(blk_update_request);
946 static void __blk_account_io_done(struct request *req, u64 now)
948 const int sgrp = op_stat_group(req_op(req));
951 update_io_ticks(req->part, jiffies, true);
952 part_stat_inc(req->part, ios[sgrp]);
953 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
957 static inline void blk_account_io_done(struct request *req, u64 now)
960 * Account IO completion. flush_rq isn't accounted as a
961 * normal IO on queueing nor completion. Accounting the
962 * containing request is enough.
964 if (blk_do_io_stat(req) && req->part &&
965 !(req->rq_flags & RQF_FLUSH_SEQ))
966 __blk_account_io_done(req, now);
969 static void __blk_account_io_start(struct request *rq)
972 * All non-passthrough requests are created from a bio with one
973 * exception: when a flush command that is part of a flush sequence
974 * generated by the state machine in blk-flush.c is cloned onto the
975 * lower device by dm-multipath we can get here without a bio.
978 rq->part = rq->bio->bi_bdev;
980 rq->part = rq->q->disk->part0;
983 update_io_ticks(rq->part, jiffies, false);
987 static inline void blk_account_io_start(struct request *req)
989 if (blk_do_io_stat(req))
990 __blk_account_io_start(req);
993 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
995 if (rq->rq_flags & RQF_STATS) {
996 blk_mq_poll_stats_start(rq->q);
997 blk_stat_add(rq, now);
1000 blk_mq_sched_completed_request(rq, now);
1001 blk_account_io_done(rq, now);
1004 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1006 if (blk_mq_need_time_stamp(rq))
1007 __blk_mq_end_request_acct(rq, ktime_get_ns());
1010 rq_qos_done(rq->q, rq);
1011 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1012 blk_mq_free_request(rq);
1014 blk_mq_free_request(rq);
1017 EXPORT_SYMBOL(__blk_mq_end_request);
1019 void blk_mq_end_request(struct request *rq, blk_status_t error)
1021 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1023 __blk_mq_end_request(rq, error);
1025 EXPORT_SYMBOL(blk_mq_end_request);
1027 #define TAG_COMP_BATCH 32
1029 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1030 int *tag_array, int nr_tags)
1032 struct request_queue *q = hctx->queue;
1035 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1036 * update hctx->nr_active in batch
1038 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1039 __blk_mq_sub_active_requests(hctx, nr_tags);
1041 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1042 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1045 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1047 int tags[TAG_COMP_BATCH], nr_tags = 0;
1048 struct blk_mq_hw_ctx *cur_hctx = NULL;
1053 now = ktime_get_ns();
1055 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1057 prefetch(rq->rq_next);
1059 blk_complete_request(rq);
1061 __blk_mq_end_request_acct(rq, now);
1063 rq_qos_done(rq->q, rq);
1066 * If end_io handler returns NONE, then it still has
1067 * ownership of the request.
1069 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1072 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1073 if (!req_ref_put_and_test(rq))
1076 blk_crypto_free_request(rq);
1077 blk_pm_mark_last_busy(rq);
1079 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1081 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1083 cur_hctx = rq->mq_hctx;
1085 tags[nr_tags++] = rq->tag;
1089 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1091 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1093 static void blk_complete_reqs(struct llist_head *list)
1095 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1096 struct request *rq, *next;
1098 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1099 rq->q->mq_ops->complete(rq);
1102 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1104 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1107 static int blk_softirq_cpu_dead(unsigned int cpu)
1109 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1113 static void __blk_mq_complete_request_remote(void *data)
1115 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1118 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1120 int cpu = raw_smp_processor_id();
1122 if (!IS_ENABLED(CONFIG_SMP) ||
1123 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1126 * With force threaded interrupts enabled, raising softirq from an SMP
1127 * function call will always result in waking the ksoftirqd thread.
1128 * This is probably worse than completing the request on a different
1131 if (force_irqthreads())
1134 /* same CPU or cache domain? Complete locally */
1135 if (cpu == rq->mq_ctx->cpu ||
1136 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1137 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1140 /* don't try to IPI to an offline CPU */
1141 return cpu_online(rq->mq_ctx->cpu);
1144 static void blk_mq_complete_send_ipi(struct request *rq)
1146 struct llist_head *list;
1149 cpu = rq->mq_ctx->cpu;
1150 list = &per_cpu(blk_cpu_done, cpu);
1151 if (llist_add(&rq->ipi_list, list)) {
1152 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1153 smp_call_function_single_async(cpu, &rq->csd);
1157 static void blk_mq_raise_softirq(struct request *rq)
1159 struct llist_head *list;
1162 list = this_cpu_ptr(&blk_cpu_done);
1163 if (llist_add(&rq->ipi_list, list))
1164 raise_softirq(BLOCK_SOFTIRQ);
1168 bool blk_mq_complete_request_remote(struct request *rq)
1170 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1173 * For request which hctx has only one ctx mapping,
1174 * or a polled request, always complete locally,
1175 * it's pointless to redirect the completion.
1177 if (rq->mq_hctx->nr_ctx == 1 ||
1178 rq->cmd_flags & REQ_POLLED)
1181 if (blk_mq_complete_need_ipi(rq)) {
1182 blk_mq_complete_send_ipi(rq);
1186 if (rq->q->nr_hw_queues == 1) {
1187 blk_mq_raise_softirq(rq);
1192 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1195 * blk_mq_complete_request - end I/O on a request
1196 * @rq: the request being processed
1199 * Complete a request by scheduling the ->complete_rq operation.
1201 void blk_mq_complete_request(struct request *rq)
1203 if (!blk_mq_complete_request_remote(rq))
1204 rq->q->mq_ops->complete(rq);
1206 EXPORT_SYMBOL(blk_mq_complete_request);
1209 * blk_mq_start_request - Start processing a request
1210 * @rq: Pointer to request to be started
1212 * Function used by device drivers to notify the block layer that a request
1213 * is going to be processed now, so blk layer can do proper initializations
1214 * such as starting the timeout timer.
1216 void blk_mq_start_request(struct request *rq)
1218 struct request_queue *q = rq->q;
1220 trace_block_rq_issue(rq);
1222 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1223 rq->io_start_time_ns = ktime_get_ns();
1224 rq->stats_sectors = blk_rq_sectors(rq);
1225 rq->rq_flags |= RQF_STATS;
1226 rq_qos_issue(q, rq);
1229 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1232 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1234 #ifdef CONFIG_BLK_DEV_INTEGRITY
1235 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1236 q->integrity.profile->prepare_fn(rq);
1238 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1239 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1241 EXPORT_SYMBOL(blk_mq_start_request);
1244 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1245 * queues. This is important for md arrays to benefit from merging
1248 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1250 if (plug->multiple_queues)
1251 return BLK_MAX_REQUEST_COUNT * 2;
1252 return BLK_MAX_REQUEST_COUNT;
1255 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1257 struct request *last = rq_list_peek(&plug->mq_list);
1259 if (!plug->rq_count) {
1260 trace_block_plug(rq->q);
1261 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1262 (!blk_queue_nomerges(rq->q) &&
1263 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1264 blk_mq_flush_plug_list(plug, false);
1265 trace_block_plug(rq->q);
1268 if (!plug->multiple_queues && last && last->q != rq->q)
1269 plug->multiple_queues = true;
1270 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1271 plug->has_elevator = true;
1273 rq_list_add(&plug->mq_list, rq);
1278 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1279 * @rq: request to insert
1280 * @at_head: insert request at head or tail of queue
1283 * Insert a fully prepared request at the back of the I/O scheduler queue
1284 * for execution. Don't wait for completion.
1287 * This function will invoke @done directly if the queue is dead.
1289 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1291 WARN_ON(irqs_disabled());
1292 WARN_ON(!blk_rq_is_passthrough(rq));
1294 blk_account_io_start(rq);
1297 * As plugging can be enabled for passthrough requests on a zoned
1298 * device, directly accessing the plug instead of using blk_mq_plug()
1299 * should not have any consequences.
1302 blk_add_rq_to_plug(current->plug, rq);
1304 blk_mq_sched_insert_request(rq, at_head, true, false);
1306 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1308 struct blk_rq_wait {
1309 struct completion done;
1313 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1315 struct blk_rq_wait *wait = rq->end_io_data;
1318 complete(&wait->done);
1319 return RQ_END_IO_NONE;
1322 bool blk_rq_is_poll(struct request *rq)
1326 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1328 if (WARN_ON_ONCE(!rq->bio))
1332 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1334 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1337 bio_poll(rq->bio, NULL, 0);
1339 } while (!completion_done(wait));
1343 * blk_execute_rq - insert a request into queue for execution
1344 * @rq: request to insert
1345 * @at_head: insert request at head or tail of queue
1348 * Insert a fully prepared request at the back of the I/O scheduler queue
1349 * for execution and wait for completion.
1350 * Return: The blk_status_t result provided to blk_mq_end_request().
1352 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1354 struct blk_rq_wait wait = {
1355 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1358 WARN_ON(irqs_disabled());
1359 WARN_ON(!blk_rq_is_passthrough(rq));
1361 rq->end_io_data = &wait;
1362 rq->end_io = blk_end_sync_rq;
1364 blk_account_io_start(rq);
1365 blk_mq_sched_insert_request(rq, at_head, true, false);
1367 if (blk_rq_is_poll(rq)) {
1368 blk_rq_poll_completion(rq, &wait.done);
1371 * Prevent hang_check timer from firing at us during very long
1374 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1377 while (!wait_for_completion_io_timeout(&wait.done,
1378 hang_check * (HZ/2)))
1381 wait_for_completion_io(&wait.done);
1386 EXPORT_SYMBOL(blk_execute_rq);
1388 static void __blk_mq_requeue_request(struct request *rq)
1390 struct request_queue *q = rq->q;
1392 blk_mq_put_driver_tag(rq);
1394 trace_block_rq_requeue(rq);
1395 rq_qos_requeue(q, rq);
1397 if (blk_mq_request_started(rq)) {
1398 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1399 rq->rq_flags &= ~RQF_TIMED_OUT;
1403 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1405 __blk_mq_requeue_request(rq);
1407 /* this request will be re-inserted to io scheduler queue */
1408 blk_mq_sched_requeue_request(rq);
1410 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1412 EXPORT_SYMBOL(blk_mq_requeue_request);
1414 static void blk_mq_requeue_work(struct work_struct *work)
1416 struct request_queue *q =
1417 container_of(work, struct request_queue, requeue_work.work);
1419 struct request *rq, *next;
1421 spin_lock_irq(&q->requeue_lock);
1422 list_splice_init(&q->requeue_list, &rq_list);
1423 spin_unlock_irq(&q->requeue_lock);
1425 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1426 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1429 rq->rq_flags &= ~RQF_SOFTBARRIER;
1430 list_del_init(&rq->queuelist);
1432 * If RQF_DONTPREP, rq has contained some driver specific
1433 * data, so insert it to hctx dispatch list to avoid any
1436 if (rq->rq_flags & RQF_DONTPREP)
1437 blk_mq_request_bypass_insert(rq, false, false);
1439 blk_mq_sched_insert_request(rq, true, false, false);
1442 while (!list_empty(&rq_list)) {
1443 rq = list_entry(rq_list.next, struct request, queuelist);
1444 list_del_init(&rq->queuelist);
1445 blk_mq_sched_insert_request(rq, false, false, false);
1448 blk_mq_run_hw_queues(q, false);
1451 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1452 bool kick_requeue_list)
1454 struct request_queue *q = rq->q;
1455 unsigned long flags;
1458 * We abuse this flag that is otherwise used by the I/O scheduler to
1459 * request head insertion from the workqueue.
1461 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1463 spin_lock_irqsave(&q->requeue_lock, flags);
1465 rq->rq_flags |= RQF_SOFTBARRIER;
1466 list_add(&rq->queuelist, &q->requeue_list);
1468 list_add_tail(&rq->queuelist, &q->requeue_list);
1470 spin_unlock_irqrestore(&q->requeue_lock, flags);
1472 if (kick_requeue_list)
1473 blk_mq_kick_requeue_list(q);
1476 void blk_mq_kick_requeue_list(struct request_queue *q)
1478 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1480 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1482 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1483 unsigned long msecs)
1485 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1486 msecs_to_jiffies(msecs));
1488 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1490 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1493 * If we find a request that isn't idle we know the queue is busy
1494 * as it's checked in the iter.
1495 * Return false to stop the iteration.
1497 if (blk_mq_request_started(rq)) {
1507 bool blk_mq_queue_inflight(struct request_queue *q)
1511 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1514 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1516 static void blk_mq_rq_timed_out(struct request *req)
1518 req->rq_flags |= RQF_TIMED_OUT;
1519 if (req->q->mq_ops->timeout) {
1520 enum blk_eh_timer_return ret;
1522 ret = req->q->mq_ops->timeout(req);
1523 if (ret == BLK_EH_DONE)
1525 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1531 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
1533 unsigned long deadline;
1535 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1537 if (rq->rq_flags & RQF_TIMED_OUT)
1540 deadline = READ_ONCE(rq->deadline);
1541 if (time_after_eq(jiffies, deadline))
1546 else if (time_after(*next, deadline))
1551 void blk_mq_put_rq_ref(struct request *rq)
1553 if (is_flush_rq(rq)) {
1554 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1555 blk_mq_free_request(rq);
1556 } else if (req_ref_put_and_test(rq)) {
1557 __blk_mq_free_request(rq);
1561 static bool blk_mq_check_expired(struct request *rq, void *priv)
1563 unsigned long *next = priv;
1566 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1567 * be reallocated underneath the timeout handler's processing, then
1568 * the expire check is reliable. If the request is not expired, then
1569 * it was completed and reallocated as a new request after returning
1570 * from blk_mq_check_expired().
1572 if (blk_mq_req_expired(rq, next))
1573 blk_mq_rq_timed_out(rq);
1577 static void blk_mq_timeout_work(struct work_struct *work)
1579 struct request_queue *q =
1580 container_of(work, struct request_queue, timeout_work);
1581 unsigned long next = 0;
1582 struct blk_mq_hw_ctx *hctx;
1585 /* A deadlock might occur if a request is stuck requiring a
1586 * timeout at the same time a queue freeze is waiting
1587 * completion, since the timeout code would not be able to
1588 * acquire the queue reference here.
1590 * That's why we don't use blk_queue_enter here; instead, we use
1591 * percpu_ref_tryget directly, because we need to be able to
1592 * obtain a reference even in the short window between the queue
1593 * starting to freeze, by dropping the first reference in
1594 * blk_freeze_queue_start, and the moment the last request is
1595 * consumed, marked by the instant q_usage_counter reaches
1598 if (!percpu_ref_tryget(&q->q_usage_counter))
1601 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
1604 mod_timer(&q->timeout, next);
1607 * Request timeouts are handled as a forward rolling timer. If
1608 * we end up here it means that no requests are pending and
1609 * also that no request has been pending for a while. Mark
1610 * each hctx as idle.
1612 queue_for_each_hw_ctx(q, hctx, i) {
1613 /* the hctx may be unmapped, so check it here */
1614 if (blk_mq_hw_queue_mapped(hctx))
1615 blk_mq_tag_idle(hctx);
1621 struct flush_busy_ctx_data {
1622 struct blk_mq_hw_ctx *hctx;
1623 struct list_head *list;
1626 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1628 struct flush_busy_ctx_data *flush_data = data;
1629 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1630 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1631 enum hctx_type type = hctx->type;
1633 spin_lock(&ctx->lock);
1634 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1635 sbitmap_clear_bit(sb, bitnr);
1636 spin_unlock(&ctx->lock);
1641 * Process software queues that have been marked busy, splicing them
1642 * to the for-dispatch
1644 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1646 struct flush_busy_ctx_data data = {
1651 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1653 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1655 struct dispatch_rq_data {
1656 struct blk_mq_hw_ctx *hctx;
1660 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1663 struct dispatch_rq_data *dispatch_data = data;
1664 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1665 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1666 enum hctx_type type = hctx->type;
1668 spin_lock(&ctx->lock);
1669 if (!list_empty(&ctx->rq_lists[type])) {
1670 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1671 list_del_init(&dispatch_data->rq->queuelist);
1672 if (list_empty(&ctx->rq_lists[type]))
1673 sbitmap_clear_bit(sb, bitnr);
1675 spin_unlock(&ctx->lock);
1677 return !dispatch_data->rq;
1680 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1681 struct blk_mq_ctx *start)
1683 unsigned off = start ? start->index_hw[hctx->type] : 0;
1684 struct dispatch_rq_data data = {
1689 __sbitmap_for_each_set(&hctx->ctx_map, off,
1690 dispatch_rq_from_ctx, &data);
1695 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1697 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1698 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1701 blk_mq_tag_busy(rq->mq_hctx);
1703 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1704 bt = &rq->mq_hctx->tags->breserved_tags;
1707 if (!hctx_may_queue(rq->mq_hctx, bt))
1711 tag = __sbitmap_queue_get(bt);
1712 if (tag == BLK_MQ_NO_TAG)
1715 rq->tag = tag + tag_offset;
1719 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1721 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1724 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1725 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1726 rq->rq_flags |= RQF_MQ_INFLIGHT;
1727 __blk_mq_inc_active_requests(hctx);
1729 hctx->tags->rqs[rq->tag] = rq;
1733 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1734 int flags, void *key)
1736 struct blk_mq_hw_ctx *hctx;
1738 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1740 spin_lock(&hctx->dispatch_wait_lock);
1741 if (!list_empty(&wait->entry)) {
1742 struct sbitmap_queue *sbq;
1744 list_del_init(&wait->entry);
1745 sbq = &hctx->tags->bitmap_tags;
1746 atomic_dec(&sbq->ws_active);
1748 spin_unlock(&hctx->dispatch_wait_lock);
1750 blk_mq_run_hw_queue(hctx, true);
1755 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1756 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1757 * restart. For both cases, take care to check the condition again after
1758 * marking us as waiting.
1760 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1763 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1764 struct wait_queue_head *wq;
1765 wait_queue_entry_t *wait;
1768 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1769 blk_mq_sched_mark_restart_hctx(hctx);
1772 * It's possible that a tag was freed in the window between the
1773 * allocation failure and adding the hardware queue to the wait
1776 * Don't clear RESTART here, someone else could have set it.
1777 * At most this will cost an extra queue run.
1779 return blk_mq_get_driver_tag(rq);
1782 wait = &hctx->dispatch_wait;
1783 if (!list_empty_careful(&wait->entry))
1786 wq = &bt_wait_ptr(sbq, hctx)->wait;
1788 spin_lock_irq(&wq->lock);
1789 spin_lock(&hctx->dispatch_wait_lock);
1790 if (!list_empty(&wait->entry)) {
1791 spin_unlock(&hctx->dispatch_wait_lock);
1792 spin_unlock_irq(&wq->lock);
1796 atomic_inc(&sbq->ws_active);
1797 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1798 __add_wait_queue(wq, wait);
1801 * It's possible that a tag was freed in the window between the
1802 * allocation failure and adding the hardware queue to the wait
1805 ret = blk_mq_get_driver_tag(rq);
1807 spin_unlock(&hctx->dispatch_wait_lock);
1808 spin_unlock_irq(&wq->lock);
1813 * We got a tag, remove ourselves from the wait queue to ensure
1814 * someone else gets the wakeup.
1816 list_del_init(&wait->entry);
1817 atomic_dec(&sbq->ws_active);
1818 spin_unlock(&hctx->dispatch_wait_lock);
1819 spin_unlock_irq(&wq->lock);
1824 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1825 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1827 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1828 * - EWMA is one simple way to compute running average value
1829 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1830 * - take 4 as factor for avoiding to get too small(0) result, and this
1831 * factor doesn't matter because EWMA decreases exponentially
1833 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1837 ewma = hctx->dispatch_busy;
1842 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1844 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1845 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1847 hctx->dispatch_busy = ewma;
1850 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1852 static void blk_mq_handle_dev_resource(struct request *rq,
1853 struct list_head *list)
1855 struct request *next =
1856 list_first_entry_or_null(list, struct request, queuelist);
1859 * If an I/O scheduler has been configured and we got a driver tag for
1860 * the next request already, free it.
1863 blk_mq_put_driver_tag(next);
1865 list_add(&rq->queuelist, list);
1866 __blk_mq_requeue_request(rq);
1869 static void blk_mq_handle_zone_resource(struct request *rq,
1870 struct list_head *zone_list)
1873 * If we end up here it is because we cannot dispatch a request to a
1874 * specific zone due to LLD level zone-write locking or other zone
1875 * related resource not being available. In this case, set the request
1876 * aside in zone_list for retrying it later.
1878 list_add(&rq->queuelist, zone_list);
1879 __blk_mq_requeue_request(rq);
1882 enum prep_dispatch {
1884 PREP_DISPATCH_NO_TAG,
1885 PREP_DISPATCH_NO_BUDGET,
1888 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1891 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1892 int budget_token = -1;
1895 budget_token = blk_mq_get_dispatch_budget(rq->q);
1896 if (budget_token < 0) {
1897 blk_mq_put_driver_tag(rq);
1898 return PREP_DISPATCH_NO_BUDGET;
1900 blk_mq_set_rq_budget_token(rq, budget_token);
1903 if (!blk_mq_get_driver_tag(rq)) {
1905 * The initial allocation attempt failed, so we need to
1906 * rerun the hardware queue when a tag is freed. The
1907 * waitqueue takes care of that. If the queue is run
1908 * before we add this entry back on the dispatch list,
1909 * we'll re-run it below.
1911 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1913 * All budgets not got from this function will be put
1914 * together during handling partial dispatch
1917 blk_mq_put_dispatch_budget(rq->q, budget_token);
1918 return PREP_DISPATCH_NO_TAG;
1922 return PREP_DISPATCH_OK;
1925 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1926 static void blk_mq_release_budgets(struct request_queue *q,
1927 struct list_head *list)
1931 list_for_each_entry(rq, list, queuelist) {
1932 int budget_token = blk_mq_get_rq_budget_token(rq);
1934 if (budget_token >= 0)
1935 blk_mq_put_dispatch_budget(q, budget_token);
1940 * Returns true if we did some work AND can potentially do more.
1942 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1943 unsigned int nr_budgets)
1945 enum prep_dispatch prep;
1946 struct request_queue *q = hctx->queue;
1947 struct request *rq, *nxt;
1949 blk_status_t ret = BLK_STS_OK;
1950 LIST_HEAD(zone_list);
1951 bool needs_resource = false;
1953 if (list_empty(list))
1957 * Now process all the entries, sending them to the driver.
1959 errors = queued = 0;
1961 struct blk_mq_queue_data bd;
1963 rq = list_first_entry(list, struct request, queuelist);
1965 WARN_ON_ONCE(hctx != rq->mq_hctx);
1966 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1967 if (prep != PREP_DISPATCH_OK)
1970 list_del_init(&rq->queuelist);
1975 * Flag last if we have no more requests, or if we have more
1976 * but can't assign a driver tag to it.
1978 if (list_empty(list))
1981 nxt = list_first_entry(list, struct request, queuelist);
1982 bd.last = !blk_mq_get_driver_tag(nxt);
1986 * once the request is queued to lld, no need to cover the
1991 ret = q->mq_ops->queue_rq(hctx, &bd);
1996 case BLK_STS_RESOURCE:
1997 needs_resource = true;
1999 case BLK_STS_DEV_RESOURCE:
2000 blk_mq_handle_dev_resource(rq, list);
2002 case BLK_STS_ZONE_RESOURCE:
2004 * Move the request to zone_list and keep going through
2005 * the dispatch list to find more requests the drive can
2008 blk_mq_handle_zone_resource(rq, &zone_list);
2009 needs_resource = true;
2013 blk_mq_end_request(rq, ret);
2015 } while (!list_empty(list));
2017 if (!list_empty(&zone_list))
2018 list_splice_tail_init(&zone_list, list);
2020 /* If we didn't flush the entire list, we could have told the driver
2021 * there was more coming, but that turned out to be a lie.
2023 if ((!list_empty(list) || errors || needs_resource ||
2024 ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued)
2025 q->mq_ops->commit_rqs(hctx);
2027 * Any items that need requeuing? Stuff them into hctx->dispatch,
2028 * that is where we will continue on next queue run.
2030 if (!list_empty(list)) {
2032 /* For non-shared tags, the RESTART check will suffice */
2033 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2034 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
2037 blk_mq_release_budgets(q, list);
2039 spin_lock(&hctx->lock);
2040 list_splice_tail_init(list, &hctx->dispatch);
2041 spin_unlock(&hctx->lock);
2044 * Order adding requests to hctx->dispatch and checking
2045 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2046 * in blk_mq_sched_restart(). Avoid restart code path to
2047 * miss the new added requests to hctx->dispatch, meantime
2048 * SCHED_RESTART is observed here.
2053 * If SCHED_RESTART was set by the caller of this function and
2054 * it is no longer set that means that it was cleared by another
2055 * thread and hence that a queue rerun is needed.
2057 * If 'no_tag' is set, that means that we failed getting
2058 * a driver tag with an I/O scheduler attached. If our dispatch
2059 * waitqueue is no longer active, ensure that we run the queue
2060 * AFTER adding our entries back to the list.
2062 * If no I/O scheduler has been configured it is possible that
2063 * the hardware queue got stopped and restarted before requests
2064 * were pushed back onto the dispatch list. Rerun the queue to
2065 * avoid starvation. Notes:
2066 * - blk_mq_run_hw_queue() checks whether or not a queue has
2067 * been stopped before rerunning a queue.
2068 * - Some but not all block drivers stop a queue before
2069 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2072 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2073 * bit is set, run queue after a delay to avoid IO stalls
2074 * that could otherwise occur if the queue is idle. We'll do
2075 * similar if we couldn't get budget or couldn't lock a zone
2076 * and SCHED_RESTART is set.
2078 needs_restart = blk_mq_sched_needs_restart(hctx);
2079 if (prep == PREP_DISPATCH_NO_BUDGET)
2080 needs_resource = true;
2081 if (!needs_restart ||
2082 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2083 blk_mq_run_hw_queue(hctx, true);
2084 else if (needs_resource)
2085 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2087 blk_mq_update_dispatch_busy(hctx, true);
2090 blk_mq_update_dispatch_busy(hctx, false);
2092 return (queued + errors) != 0;
2096 * __blk_mq_run_hw_queue - Run a hardware queue.
2097 * @hctx: Pointer to the hardware queue to run.
2099 * Send pending requests to the hardware.
2101 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
2104 * We can't run the queue inline with ints disabled. Ensure that
2105 * we catch bad users of this early.
2107 WARN_ON_ONCE(in_interrupt());
2109 blk_mq_run_dispatch_ops(hctx->queue,
2110 blk_mq_sched_dispatch_requests(hctx));
2113 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2115 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2117 if (cpu >= nr_cpu_ids)
2118 cpu = cpumask_first(hctx->cpumask);
2123 * It'd be great if the workqueue API had a way to pass
2124 * in a mask and had some smarts for more clever placement.
2125 * For now we just round-robin here, switching for every
2126 * BLK_MQ_CPU_WORK_BATCH queued items.
2128 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2131 int next_cpu = hctx->next_cpu;
2133 if (hctx->queue->nr_hw_queues == 1)
2134 return WORK_CPU_UNBOUND;
2136 if (--hctx->next_cpu_batch <= 0) {
2138 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2140 if (next_cpu >= nr_cpu_ids)
2141 next_cpu = blk_mq_first_mapped_cpu(hctx);
2142 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2146 * Do unbound schedule if we can't find a online CPU for this hctx,
2147 * and it should only happen in the path of handling CPU DEAD.
2149 if (!cpu_online(next_cpu)) {
2156 * Make sure to re-select CPU next time once after CPUs
2157 * in hctx->cpumask become online again.
2159 hctx->next_cpu = next_cpu;
2160 hctx->next_cpu_batch = 1;
2161 return WORK_CPU_UNBOUND;
2164 hctx->next_cpu = next_cpu;
2169 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2170 * @hctx: Pointer to the hardware queue to run.
2171 * @async: If we want to run the queue asynchronously.
2172 * @msecs: Milliseconds of delay to wait before running the queue.
2174 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2175 * with a delay of @msecs.
2177 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2178 unsigned long msecs)
2180 if (unlikely(blk_mq_hctx_stopped(hctx)))
2183 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2184 if (cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2185 __blk_mq_run_hw_queue(hctx);
2190 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2191 msecs_to_jiffies(msecs));
2195 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2196 * @hctx: Pointer to the hardware queue to run.
2197 * @msecs: Milliseconds of delay to wait before running the queue.
2199 * Run a hardware queue asynchronously with a delay of @msecs.
2201 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2203 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2205 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2208 * blk_mq_run_hw_queue - Start to run a hardware queue.
2209 * @hctx: Pointer to the hardware queue to run.
2210 * @async: If we want to run the queue asynchronously.
2212 * Check if the request queue is not in a quiesced state and if there are
2213 * pending requests to be sent. If this is true, run the queue to send requests
2216 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2221 * When queue is quiesced, we may be switching io scheduler, or
2222 * updating nr_hw_queues, or other things, and we can't run queue
2223 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2225 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2228 __blk_mq_run_dispatch_ops(hctx->queue, false,
2229 need_run = !blk_queue_quiesced(hctx->queue) &&
2230 blk_mq_hctx_has_pending(hctx));
2233 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2235 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2238 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2241 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2243 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2245 * If the IO scheduler does not respect hardware queues when
2246 * dispatching, we just don't bother with multiple HW queues and
2247 * dispatch from hctx for the current CPU since running multiple queues
2248 * just causes lock contention inside the scheduler and pointless cache
2251 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2253 if (!blk_mq_hctx_stopped(hctx))
2259 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2260 * @q: Pointer to the request queue to run.
2261 * @async: If we want to run the queue asynchronously.
2263 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2265 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2269 if (blk_queue_sq_sched(q))
2270 sq_hctx = blk_mq_get_sq_hctx(q);
2271 queue_for_each_hw_ctx(q, hctx, i) {
2272 if (blk_mq_hctx_stopped(hctx))
2275 * Dispatch from this hctx either if there's no hctx preferred
2276 * by IO scheduler or if it has requests that bypass the
2279 if (!sq_hctx || sq_hctx == hctx ||
2280 !list_empty_careful(&hctx->dispatch))
2281 blk_mq_run_hw_queue(hctx, async);
2284 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2287 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2288 * @q: Pointer to the request queue to run.
2289 * @msecs: Milliseconds of delay to wait before running the queues.
2291 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2293 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2297 if (blk_queue_sq_sched(q))
2298 sq_hctx = blk_mq_get_sq_hctx(q);
2299 queue_for_each_hw_ctx(q, hctx, i) {
2300 if (blk_mq_hctx_stopped(hctx))
2303 * If there is already a run_work pending, leave the
2304 * pending delay untouched. Otherwise, a hctx can stall
2305 * if another hctx is re-delaying the other's work
2306 * before the work executes.
2308 if (delayed_work_pending(&hctx->run_work))
2311 * Dispatch from this hctx either if there's no hctx preferred
2312 * by IO scheduler or if it has requests that bypass the
2315 if (!sq_hctx || sq_hctx == hctx ||
2316 !list_empty_careful(&hctx->dispatch))
2317 blk_mq_delay_run_hw_queue(hctx, msecs);
2320 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2323 * This function is often used for pausing .queue_rq() by driver when
2324 * there isn't enough resource or some conditions aren't satisfied, and
2325 * BLK_STS_RESOURCE is usually returned.
2327 * We do not guarantee that dispatch can be drained or blocked
2328 * after blk_mq_stop_hw_queue() returns. Please use
2329 * blk_mq_quiesce_queue() for that requirement.
2331 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2333 cancel_delayed_work(&hctx->run_work);
2335 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2337 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2340 * This function is often used for pausing .queue_rq() by driver when
2341 * there isn't enough resource or some conditions aren't satisfied, and
2342 * BLK_STS_RESOURCE is usually returned.
2344 * We do not guarantee that dispatch can be drained or blocked
2345 * after blk_mq_stop_hw_queues() returns. Please use
2346 * blk_mq_quiesce_queue() for that requirement.
2348 void blk_mq_stop_hw_queues(struct request_queue *q)
2350 struct blk_mq_hw_ctx *hctx;
2353 queue_for_each_hw_ctx(q, hctx, i)
2354 blk_mq_stop_hw_queue(hctx);
2356 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2358 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2360 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2362 blk_mq_run_hw_queue(hctx, false);
2364 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2366 void blk_mq_start_hw_queues(struct request_queue *q)
2368 struct blk_mq_hw_ctx *hctx;
2371 queue_for_each_hw_ctx(q, hctx, i)
2372 blk_mq_start_hw_queue(hctx);
2374 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2376 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2378 if (!blk_mq_hctx_stopped(hctx))
2381 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2382 blk_mq_run_hw_queue(hctx, async);
2384 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2386 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2388 struct blk_mq_hw_ctx *hctx;
2391 queue_for_each_hw_ctx(q, hctx, i)
2392 blk_mq_start_stopped_hw_queue(hctx, async);
2394 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2396 static void blk_mq_run_work_fn(struct work_struct *work)
2398 struct blk_mq_hw_ctx *hctx;
2400 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2403 * If we are stopped, don't run the queue.
2405 if (blk_mq_hctx_stopped(hctx))
2408 __blk_mq_run_hw_queue(hctx);
2411 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2415 struct blk_mq_ctx *ctx = rq->mq_ctx;
2416 enum hctx_type type = hctx->type;
2418 lockdep_assert_held(&ctx->lock);
2420 trace_block_rq_insert(rq);
2423 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2425 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2428 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2431 struct blk_mq_ctx *ctx = rq->mq_ctx;
2433 lockdep_assert_held(&ctx->lock);
2435 __blk_mq_insert_req_list(hctx, rq, at_head);
2436 blk_mq_hctx_mark_pending(hctx, ctx);
2440 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2441 * @rq: Pointer to request to be inserted.
2442 * @at_head: true if the request should be inserted at the head of the list.
2443 * @run_queue: If we should run the hardware queue after inserting the request.
2445 * Should only be used carefully, when the caller knows we want to
2446 * bypass a potential IO scheduler on the target device.
2448 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2451 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2453 spin_lock(&hctx->lock);
2455 list_add(&rq->queuelist, &hctx->dispatch);
2457 list_add_tail(&rq->queuelist, &hctx->dispatch);
2458 spin_unlock(&hctx->lock);
2461 blk_mq_run_hw_queue(hctx, false);
2464 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2465 struct list_head *list)
2469 enum hctx_type type = hctx->type;
2472 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2475 list_for_each_entry(rq, list, queuelist) {
2476 BUG_ON(rq->mq_ctx != ctx);
2477 trace_block_rq_insert(rq);
2480 spin_lock(&ctx->lock);
2481 list_splice_tail_init(list, &ctx->rq_lists[type]);
2482 blk_mq_hctx_mark_pending(hctx, ctx);
2483 spin_unlock(&ctx->lock);
2486 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2489 if (hctx->queue->mq_ops->commit_rqs) {
2490 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2491 hctx->queue->mq_ops->commit_rqs(hctx);
2496 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2497 unsigned int nr_segs)
2501 if (bio->bi_opf & REQ_RAHEAD)
2502 rq->cmd_flags |= REQ_FAILFAST_MASK;
2504 rq->__sector = bio->bi_iter.bi_sector;
2505 blk_rq_bio_prep(rq, bio, nr_segs);
2507 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2508 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2511 blk_account_io_start(rq);
2514 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2515 struct request *rq, bool last)
2517 struct request_queue *q = rq->q;
2518 struct blk_mq_queue_data bd = {
2525 * For OK queue, we are done. For error, caller may kill it.
2526 * Any other error (busy), just add it to our list as we
2527 * previously would have done.
2529 ret = q->mq_ops->queue_rq(hctx, &bd);
2532 blk_mq_update_dispatch_busy(hctx, false);
2534 case BLK_STS_RESOURCE:
2535 case BLK_STS_DEV_RESOURCE:
2536 blk_mq_update_dispatch_busy(hctx, true);
2537 __blk_mq_requeue_request(rq);
2540 blk_mq_update_dispatch_busy(hctx, false);
2547 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2549 bool bypass_insert, bool last)
2551 struct request_queue *q = rq->q;
2552 bool run_queue = true;
2556 * RCU or SRCU read lock is needed before checking quiesced flag.
2558 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2559 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2560 * and avoid driver to try to dispatch again.
2562 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2564 bypass_insert = false;
2568 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2571 budget_token = blk_mq_get_dispatch_budget(q);
2572 if (budget_token < 0)
2575 blk_mq_set_rq_budget_token(rq, budget_token);
2577 if (!blk_mq_get_driver_tag(rq)) {
2578 blk_mq_put_dispatch_budget(q, budget_token);
2582 return __blk_mq_issue_directly(hctx, rq, last);
2585 return BLK_STS_RESOURCE;
2587 blk_mq_sched_insert_request(rq, false, run_queue, false);
2593 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2594 * @hctx: Pointer of the associated hardware queue.
2595 * @rq: Pointer to request to be sent.
2597 * If the device has enough resources to accept a new request now, send the
2598 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2599 * we can try send it another time in the future. Requests inserted at this
2600 * queue have higher priority.
2602 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2606 __blk_mq_try_issue_directly(hctx, rq, false, true);
2608 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2609 blk_mq_request_bypass_insert(rq, false, true);
2610 else if (ret != BLK_STS_OK)
2611 blk_mq_end_request(rq, ret);
2614 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2616 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2619 static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2621 struct blk_mq_hw_ctx *hctx = NULL;
2626 while ((rq = rq_list_pop(&plug->mq_list))) {
2627 bool last = rq_list_empty(plug->mq_list);
2630 if (hctx != rq->mq_hctx) {
2632 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2636 ret = blk_mq_request_issue_directly(rq, last);
2641 case BLK_STS_RESOURCE:
2642 case BLK_STS_DEV_RESOURCE:
2643 blk_mq_request_bypass_insert(rq, false, true);
2644 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2647 blk_mq_end_request(rq, ret);
2654 * If we didn't flush the entire list, we could have told the driver
2655 * there was more coming, but that turned out to be a lie.
2658 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2661 static void __blk_mq_flush_plug_list(struct request_queue *q,
2662 struct blk_plug *plug)
2664 if (blk_queue_quiesced(q))
2666 q->mq_ops->queue_rqs(&plug->mq_list);
2669 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2671 struct blk_mq_hw_ctx *this_hctx = NULL;
2672 struct blk_mq_ctx *this_ctx = NULL;
2673 struct request *requeue_list = NULL;
2674 unsigned int depth = 0;
2678 struct request *rq = rq_list_pop(&plug->mq_list);
2681 this_hctx = rq->mq_hctx;
2682 this_ctx = rq->mq_ctx;
2683 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2684 rq_list_add(&requeue_list, rq);
2687 list_add_tail(&rq->queuelist, &list);
2689 } while (!rq_list_empty(plug->mq_list));
2691 plug->mq_list = requeue_list;
2692 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2693 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched);
2696 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2700 if (rq_list_empty(plug->mq_list))
2704 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2705 struct request_queue *q;
2707 rq = rq_list_peek(&plug->mq_list);
2711 * Peek first request and see if we have a ->queue_rqs() hook.
2712 * If we do, we can dispatch the whole plug list in one go. We
2713 * already know at this point that all requests belong to the
2714 * same queue, caller must ensure that's the case.
2716 * Since we pass off the full list to the driver at this point,
2717 * we do not increment the active request count for the queue.
2718 * Bypass shared tags for now because of that.
2720 if (q->mq_ops->queue_rqs &&
2721 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2722 blk_mq_run_dispatch_ops(q,
2723 __blk_mq_flush_plug_list(q, plug));
2724 if (rq_list_empty(plug->mq_list))
2728 blk_mq_run_dispatch_ops(q,
2729 blk_mq_plug_issue_direct(plug, false));
2730 if (rq_list_empty(plug->mq_list))
2735 blk_mq_dispatch_plug_list(plug, from_schedule);
2736 } while (!rq_list_empty(plug->mq_list));
2739 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2740 struct list_head *list)
2745 while (!list_empty(list)) {
2747 struct request *rq = list_first_entry(list, struct request,
2750 list_del_init(&rq->queuelist);
2751 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2752 if (ret != BLK_STS_OK) {
2754 if (ret == BLK_STS_RESOURCE ||
2755 ret == BLK_STS_DEV_RESOURCE) {
2756 blk_mq_request_bypass_insert(rq, false,
2760 blk_mq_end_request(rq, ret);
2766 * If we didn't flush the entire list, we could have told
2767 * the driver there was more coming, but that turned out to
2770 if ((!list_empty(list) || errors) &&
2771 hctx->queue->mq_ops->commit_rqs && queued)
2772 hctx->queue->mq_ops->commit_rqs(hctx);
2775 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2776 struct bio *bio, unsigned int nr_segs)
2778 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2779 if (blk_attempt_plug_merge(q, bio, nr_segs))
2781 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2787 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2788 struct blk_plug *plug,
2792 struct blk_mq_alloc_data data = {
2795 .cmd_flags = bio->bi_opf,
2799 if (unlikely(bio_queue_enter(bio)))
2802 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2805 rq_qos_throttle(q, bio);
2808 data.nr_tags = plug->nr_ios;
2810 data.cached_rq = &plug->cached_rq;
2813 rq = __blk_mq_alloc_requests(&data);
2816 rq_qos_cleanup(q, bio);
2817 if (bio->bi_opf & REQ_NOWAIT)
2818 bio_wouldblock_error(bio);
2824 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2825 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2831 rq = rq_list_peek(&plug->cached_rq);
2832 if (!rq || rq->q != q)
2835 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2840 if (blk_mq_get_hctx_type((*bio)->bi_opf) != rq->mq_hctx->type)
2842 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2846 * If any qos ->throttle() end up blocking, we will have flushed the
2847 * plug and hence killed the cached_rq list as well. Pop this entry
2848 * before we throttle.
2850 plug->cached_rq = rq_list_next(rq);
2851 rq_qos_throttle(q, *bio);
2853 rq->cmd_flags = (*bio)->bi_opf;
2854 INIT_LIST_HEAD(&rq->queuelist);
2858 static void bio_set_ioprio(struct bio *bio)
2860 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2861 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2862 bio->bi_ioprio = get_current_ioprio();
2863 blkcg_set_ioprio(bio);
2867 * blk_mq_submit_bio - Create and send a request to block device.
2868 * @bio: Bio pointer.
2870 * Builds up a request structure from @q and @bio and send to the device. The
2871 * request may not be queued directly to hardware if:
2872 * * This request can be merged with another one
2873 * * We want to place request at plug queue for possible future merging
2874 * * There is an IO scheduler active at this queue
2876 * It will not queue the request if there is an error with the bio, or at the
2879 void blk_mq_submit_bio(struct bio *bio)
2881 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2882 struct blk_plug *plug = blk_mq_plug(bio);
2883 const int is_sync = op_is_sync(bio->bi_opf);
2885 unsigned int nr_segs = 1;
2888 bio = blk_queue_bounce(bio, q);
2889 if (bio_may_exceed_limits(bio, &q->limits))
2890 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2892 if (!bio_integrity_prep(bio))
2895 bio_set_ioprio(bio);
2897 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2901 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2906 trace_block_getrq(bio);
2908 rq_qos_track(q, rq, bio);
2910 blk_mq_bio_to_request(rq, bio, nr_segs);
2912 ret = blk_crypto_init_request(rq);
2913 if (ret != BLK_STS_OK) {
2914 bio->bi_status = ret;
2916 blk_mq_free_request(rq);
2920 if (op_is_flush(bio->bi_opf)) {
2921 blk_insert_flush(rq);
2926 blk_add_rq_to_plug(plug, rq);
2927 else if ((rq->rq_flags & RQF_ELV) ||
2928 (rq->mq_hctx->dispatch_busy &&
2929 (q->nr_hw_queues == 1 || !is_sync)))
2930 blk_mq_sched_insert_request(rq, false, true, true);
2932 blk_mq_run_dispatch_ops(rq->q,
2933 blk_mq_try_issue_directly(rq->mq_hctx, rq));
2936 #ifdef CONFIG_BLK_MQ_STACKING
2938 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
2939 * @rq: the request being queued
2941 blk_status_t blk_insert_cloned_request(struct request *rq)
2943 struct request_queue *q = rq->q;
2944 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
2947 if (blk_rq_sectors(rq) > max_sectors) {
2949 * SCSI device does not have a good way to return if
2950 * Write Same/Zero is actually supported. If a device rejects
2951 * a non-read/write command (discard, write same,etc.) the
2952 * low-level device driver will set the relevant queue limit to
2953 * 0 to prevent blk-lib from issuing more of the offending
2954 * operations. Commands queued prior to the queue limit being
2955 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
2956 * errors being propagated to upper layers.
2958 if (max_sectors == 0)
2959 return BLK_STS_NOTSUPP;
2961 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
2962 __func__, blk_rq_sectors(rq), max_sectors);
2963 return BLK_STS_IOERR;
2967 * The queue settings related to segment counting may differ from the
2970 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
2971 if (rq->nr_phys_segments > queue_max_segments(q)) {
2972 printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n",
2973 __func__, rq->nr_phys_segments, queue_max_segments(q));
2974 return BLK_STS_IOERR;
2977 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
2978 return BLK_STS_IOERR;
2980 if (blk_crypto_insert_cloned_request(rq))
2981 return BLK_STS_IOERR;
2983 blk_account_io_start(rq);
2986 * Since we have a scheduler attached on the top device,
2987 * bypass a potential scheduler on the bottom device for
2990 blk_mq_run_dispatch_ops(q,
2991 ret = blk_mq_request_issue_directly(rq, true));
2993 blk_account_io_done(rq, ktime_get_ns());
2996 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
2999 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3000 * @rq: the clone request to be cleaned up
3003 * Free all bios in @rq for a cloned request.
3005 void blk_rq_unprep_clone(struct request *rq)
3009 while ((bio = rq->bio) != NULL) {
3010 rq->bio = bio->bi_next;
3015 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3018 * blk_rq_prep_clone - Helper function to setup clone request
3019 * @rq: the request to be setup
3020 * @rq_src: original request to be cloned
3021 * @bs: bio_set that bios for clone are allocated from
3022 * @gfp_mask: memory allocation mask for bio
3023 * @bio_ctr: setup function to be called for each clone bio.
3024 * Returns %0 for success, non %0 for failure.
3025 * @data: private data to be passed to @bio_ctr
3028 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3029 * Also, pages which the original bios are pointing to are not copied
3030 * and the cloned bios just point same pages.
3031 * So cloned bios must be completed before original bios, which means
3032 * the caller must complete @rq before @rq_src.
3034 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3035 struct bio_set *bs, gfp_t gfp_mask,
3036 int (*bio_ctr)(struct bio *, struct bio *, void *),
3039 struct bio *bio, *bio_src;
3044 __rq_for_each_bio(bio_src, rq_src) {
3045 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3050 if (bio_ctr && bio_ctr(bio, bio_src, data))
3054 rq->biotail->bi_next = bio;
3057 rq->bio = rq->biotail = bio;
3062 /* Copy attributes of the original request to the clone request. */
3063 rq->__sector = blk_rq_pos(rq_src);
3064 rq->__data_len = blk_rq_bytes(rq_src);
3065 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3066 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3067 rq->special_vec = rq_src->special_vec;
3069 rq->nr_phys_segments = rq_src->nr_phys_segments;
3070 rq->ioprio = rq_src->ioprio;
3072 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3080 blk_rq_unprep_clone(rq);
3084 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3085 #endif /* CONFIG_BLK_MQ_STACKING */
3088 * Steal bios from a request and add them to a bio list.
3089 * The request must not have been partially completed before.
3091 void blk_steal_bios(struct bio_list *list, struct request *rq)
3095 list->tail->bi_next = rq->bio;
3097 list->head = rq->bio;
3098 list->tail = rq->biotail;
3106 EXPORT_SYMBOL_GPL(blk_steal_bios);
3108 static size_t order_to_size(unsigned int order)
3110 return (size_t)PAGE_SIZE << order;
3113 /* called before freeing request pool in @tags */
3114 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3115 struct blk_mq_tags *tags)
3118 unsigned long flags;
3121 * There is no need to clear mapping if driver tags is not initialized
3122 * or the mapping belongs to the driver tags.
3124 if (!drv_tags || drv_tags == tags)
3127 list_for_each_entry(page, &tags->page_list, lru) {
3128 unsigned long start = (unsigned long)page_address(page);
3129 unsigned long end = start + order_to_size(page->private);
3132 for (i = 0; i < drv_tags->nr_tags; i++) {
3133 struct request *rq = drv_tags->rqs[i];
3134 unsigned long rq_addr = (unsigned long)rq;
3136 if (rq_addr >= start && rq_addr < end) {
3137 WARN_ON_ONCE(req_ref_read(rq) != 0);
3138 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3144 * Wait until all pending iteration is done.
3146 * Request reference is cleared and it is guaranteed to be observed
3147 * after the ->lock is released.
3149 spin_lock_irqsave(&drv_tags->lock, flags);
3150 spin_unlock_irqrestore(&drv_tags->lock, flags);
3153 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3154 unsigned int hctx_idx)
3156 struct blk_mq_tags *drv_tags;
3159 if (list_empty(&tags->page_list))
3162 if (blk_mq_is_shared_tags(set->flags))
3163 drv_tags = set->shared_tags;
3165 drv_tags = set->tags[hctx_idx];
3167 if (tags->static_rqs && set->ops->exit_request) {
3170 for (i = 0; i < tags->nr_tags; i++) {
3171 struct request *rq = tags->static_rqs[i];
3175 set->ops->exit_request(set, rq, hctx_idx);
3176 tags->static_rqs[i] = NULL;
3180 blk_mq_clear_rq_mapping(drv_tags, tags);
3182 while (!list_empty(&tags->page_list)) {
3183 page = list_first_entry(&tags->page_list, struct page, lru);
3184 list_del_init(&page->lru);
3186 * Remove kmemleak object previously allocated in
3187 * blk_mq_alloc_rqs().
3189 kmemleak_free(page_address(page));
3190 __free_pages(page, page->private);
3194 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3198 kfree(tags->static_rqs);
3199 tags->static_rqs = NULL;
3201 blk_mq_free_tags(tags);
3204 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3205 unsigned int hctx_idx)
3209 for (i = 0; i < set->nr_maps; i++) {
3210 unsigned int start = set->map[i].queue_offset;
3211 unsigned int end = start + set->map[i].nr_queues;
3213 if (hctx_idx >= start && hctx_idx < end)
3217 if (i >= set->nr_maps)
3218 i = HCTX_TYPE_DEFAULT;
3223 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3224 unsigned int hctx_idx)
3226 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3228 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3231 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3232 unsigned int hctx_idx,
3233 unsigned int nr_tags,
3234 unsigned int reserved_tags)
3236 int node = blk_mq_get_hctx_node(set, hctx_idx);
3237 struct blk_mq_tags *tags;
3239 if (node == NUMA_NO_NODE)
3240 node = set->numa_node;
3242 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3243 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3247 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3248 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3251 blk_mq_free_tags(tags);
3255 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3256 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3258 if (!tags->static_rqs) {
3260 blk_mq_free_tags(tags);
3267 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3268 unsigned int hctx_idx, int node)
3272 if (set->ops->init_request) {
3273 ret = set->ops->init_request(set, rq, hctx_idx, node);
3278 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3282 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3283 struct blk_mq_tags *tags,
3284 unsigned int hctx_idx, unsigned int depth)
3286 unsigned int i, j, entries_per_page, max_order = 4;
3287 int node = blk_mq_get_hctx_node(set, hctx_idx);
3288 size_t rq_size, left;
3290 if (node == NUMA_NO_NODE)
3291 node = set->numa_node;
3293 INIT_LIST_HEAD(&tags->page_list);
3296 * rq_size is the size of the request plus driver payload, rounded
3297 * to the cacheline size
3299 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3301 left = rq_size * depth;
3303 for (i = 0; i < depth; ) {
3304 int this_order = max_order;
3309 while (this_order && left < order_to_size(this_order - 1))
3313 page = alloc_pages_node(node,
3314 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3320 if (order_to_size(this_order) < rq_size)
3327 page->private = this_order;
3328 list_add_tail(&page->lru, &tags->page_list);
3330 p = page_address(page);
3332 * Allow kmemleak to scan these pages as they contain pointers
3333 * to additional allocations like via ops->init_request().
3335 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3336 entries_per_page = order_to_size(this_order) / rq_size;
3337 to_do = min(entries_per_page, depth - i);
3338 left -= to_do * rq_size;
3339 for (j = 0; j < to_do; j++) {
3340 struct request *rq = p;
3342 tags->static_rqs[i] = rq;
3343 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3344 tags->static_rqs[i] = NULL;
3355 blk_mq_free_rqs(set, tags, hctx_idx);
3359 struct rq_iter_data {
3360 struct blk_mq_hw_ctx *hctx;
3364 static bool blk_mq_has_request(struct request *rq, void *data)
3366 struct rq_iter_data *iter_data = data;
3368 if (rq->mq_hctx != iter_data->hctx)
3370 iter_data->has_rq = true;
3374 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3376 struct blk_mq_tags *tags = hctx->sched_tags ?
3377 hctx->sched_tags : hctx->tags;
3378 struct rq_iter_data data = {
3382 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3386 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3387 struct blk_mq_hw_ctx *hctx)
3389 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3391 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3396 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3398 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3399 struct blk_mq_hw_ctx, cpuhp_online);
3401 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3402 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3406 * Prevent new request from being allocated on the current hctx.
3408 * The smp_mb__after_atomic() Pairs with the implied barrier in
3409 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3410 * seen once we return from the tag allocator.
3412 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3413 smp_mb__after_atomic();
3416 * Try to grab a reference to the queue and wait for any outstanding
3417 * requests. If we could not grab a reference the queue has been
3418 * frozen and there are no requests.
3420 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3421 while (blk_mq_hctx_has_requests(hctx))
3423 percpu_ref_put(&hctx->queue->q_usage_counter);
3429 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3431 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3432 struct blk_mq_hw_ctx, cpuhp_online);
3434 if (cpumask_test_cpu(cpu, hctx->cpumask))
3435 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3440 * 'cpu' is going away. splice any existing rq_list entries from this
3441 * software queue to the hw queue dispatch list, and ensure that it
3444 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3446 struct blk_mq_hw_ctx *hctx;
3447 struct blk_mq_ctx *ctx;
3449 enum hctx_type type;
3451 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3452 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3455 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3458 spin_lock(&ctx->lock);
3459 if (!list_empty(&ctx->rq_lists[type])) {
3460 list_splice_init(&ctx->rq_lists[type], &tmp);
3461 blk_mq_hctx_clear_pending(hctx, ctx);
3463 spin_unlock(&ctx->lock);
3465 if (list_empty(&tmp))
3468 spin_lock(&hctx->lock);
3469 list_splice_tail_init(&tmp, &hctx->dispatch);
3470 spin_unlock(&hctx->lock);
3472 blk_mq_run_hw_queue(hctx, true);
3476 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3478 if (!(hctx->flags & BLK_MQ_F_STACKING))
3479 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3480 &hctx->cpuhp_online);
3481 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3486 * Before freeing hw queue, clearing the flush request reference in
3487 * tags->rqs[] for avoiding potential UAF.
3489 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3490 unsigned int queue_depth, struct request *flush_rq)
3493 unsigned long flags;
3495 /* The hw queue may not be mapped yet */
3499 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3501 for (i = 0; i < queue_depth; i++)
3502 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3505 * Wait until all pending iteration is done.
3507 * Request reference is cleared and it is guaranteed to be observed
3508 * after the ->lock is released.
3510 spin_lock_irqsave(&tags->lock, flags);
3511 spin_unlock_irqrestore(&tags->lock, flags);
3514 /* hctx->ctxs will be freed in queue's release handler */
3515 static void blk_mq_exit_hctx(struct request_queue *q,
3516 struct blk_mq_tag_set *set,
3517 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3519 struct request *flush_rq = hctx->fq->flush_rq;
3521 if (blk_mq_hw_queue_mapped(hctx))
3522 blk_mq_tag_idle(hctx);
3524 if (blk_queue_init_done(q))
3525 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3526 set->queue_depth, flush_rq);
3527 if (set->ops->exit_request)
3528 set->ops->exit_request(set, flush_rq, hctx_idx);
3530 if (set->ops->exit_hctx)
3531 set->ops->exit_hctx(hctx, hctx_idx);
3533 blk_mq_remove_cpuhp(hctx);
3535 xa_erase(&q->hctx_table, hctx_idx);
3537 spin_lock(&q->unused_hctx_lock);
3538 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3539 spin_unlock(&q->unused_hctx_lock);
3542 static void blk_mq_exit_hw_queues(struct request_queue *q,
3543 struct blk_mq_tag_set *set, int nr_queue)
3545 struct blk_mq_hw_ctx *hctx;
3548 queue_for_each_hw_ctx(q, hctx, i) {
3551 blk_mq_exit_hctx(q, set, hctx, i);
3555 static int blk_mq_init_hctx(struct request_queue *q,
3556 struct blk_mq_tag_set *set,
3557 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3559 hctx->queue_num = hctx_idx;
3561 if (!(hctx->flags & BLK_MQ_F_STACKING))
3562 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3563 &hctx->cpuhp_online);
3564 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3566 hctx->tags = set->tags[hctx_idx];
3568 if (set->ops->init_hctx &&
3569 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3570 goto unregister_cpu_notifier;
3572 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3576 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3582 if (set->ops->exit_request)
3583 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3585 if (set->ops->exit_hctx)
3586 set->ops->exit_hctx(hctx, hctx_idx);
3587 unregister_cpu_notifier:
3588 blk_mq_remove_cpuhp(hctx);
3592 static struct blk_mq_hw_ctx *
3593 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3596 struct blk_mq_hw_ctx *hctx;
3597 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3599 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3601 goto fail_alloc_hctx;
3603 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3606 atomic_set(&hctx->nr_active, 0);
3607 if (node == NUMA_NO_NODE)
3608 node = set->numa_node;
3609 hctx->numa_node = node;
3611 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3612 spin_lock_init(&hctx->lock);
3613 INIT_LIST_HEAD(&hctx->dispatch);
3615 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3617 INIT_LIST_HEAD(&hctx->hctx_list);
3620 * Allocate space for all possible cpus to avoid allocation at
3623 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3628 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3629 gfp, node, false, false))
3633 spin_lock_init(&hctx->dispatch_wait_lock);
3634 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3635 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3637 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3641 blk_mq_hctx_kobj_init(hctx);
3646 sbitmap_free(&hctx->ctx_map);
3650 free_cpumask_var(hctx->cpumask);
3657 static void blk_mq_init_cpu_queues(struct request_queue *q,
3658 unsigned int nr_hw_queues)
3660 struct blk_mq_tag_set *set = q->tag_set;
3663 for_each_possible_cpu(i) {
3664 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3665 struct blk_mq_hw_ctx *hctx;
3669 spin_lock_init(&__ctx->lock);
3670 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3671 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3676 * Set local node, IFF we have more than one hw queue. If
3677 * not, we remain on the home node of the device
3679 for (j = 0; j < set->nr_maps; j++) {
3680 hctx = blk_mq_map_queue_type(q, j, i);
3681 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3682 hctx->numa_node = cpu_to_node(i);
3687 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3688 unsigned int hctx_idx,
3691 struct blk_mq_tags *tags;
3694 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3698 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3700 blk_mq_free_rq_map(tags);
3707 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3710 if (blk_mq_is_shared_tags(set->flags)) {
3711 set->tags[hctx_idx] = set->shared_tags;
3716 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3719 return set->tags[hctx_idx];
3722 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3723 struct blk_mq_tags *tags,
3724 unsigned int hctx_idx)
3727 blk_mq_free_rqs(set, tags, hctx_idx);
3728 blk_mq_free_rq_map(tags);
3732 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3733 unsigned int hctx_idx)
3735 if (!blk_mq_is_shared_tags(set->flags))
3736 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3738 set->tags[hctx_idx] = NULL;
3741 static void blk_mq_map_swqueue(struct request_queue *q)
3743 unsigned int j, hctx_idx;
3745 struct blk_mq_hw_ctx *hctx;
3746 struct blk_mq_ctx *ctx;
3747 struct blk_mq_tag_set *set = q->tag_set;
3749 queue_for_each_hw_ctx(q, hctx, i) {
3750 cpumask_clear(hctx->cpumask);
3752 hctx->dispatch_from = NULL;
3756 * Map software to hardware queues.
3758 * If the cpu isn't present, the cpu is mapped to first hctx.
3760 for_each_possible_cpu(i) {
3762 ctx = per_cpu_ptr(q->queue_ctx, i);
3763 for (j = 0; j < set->nr_maps; j++) {
3764 if (!set->map[j].nr_queues) {
3765 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3766 HCTX_TYPE_DEFAULT, i);
3769 hctx_idx = set->map[j].mq_map[i];
3770 /* unmapped hw queue can be remapped after CPU topo changed */
3771 if (!set->tags[hctx_idx] &&
3772 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3774 * If tags initialization fail for some hctx,
3775 * that hctx won't be brought online. In this
3776 * case, remap the current ctx to hctx[0] which
3777 * is guaranteed to always have tags allocated
3779 set->map[j].mq_map[i] = 0;
3782 hctx = blk_mq_map_queue_type(q, j, i);
3783 ctx->hctxs[j] = hctx;
3785 * If the CPU is already set in the mask, then we've
3786 * mapped this one already. This can happen if
3787 * devices share queues across queue maps.
3789 if (cpumask_test_cpu(i, hctx->cpumask))
3792 cpumask_set_cpu(i, hctx->cpumask);
3794 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3795 hctx->ctxs[hctx->nr_ctx++] = ctx;
3798 * If the nr_ctx type overflows, we have exceeded the
3799 * amount of sw queues we can support.
3801 BUG_ON(!hctx->nr_ctx);
3804 for (; j < HCTX_MAX_TYPES; j++)
3805 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3806 HCTX_TYPE_DEFAULT, i);
3809 queue_for_each_hw_ctx(q, hctx, i) {
3811 * If no software queues are mapped to this hardware queue,
3812 * disable it and free the request entries.
3814 if (!hctx->nr_ctx) {
3815 /* Never unmap queue 0. We need it as a
3816 * fallback in case of a new remap fails
3820 __blk_mq_free_map_and_rqs(set, i);
3826 hctx->tags = set->tags[i];
3827 WARN_ON(!hctx->tags);
3830 * Set the map size to the number of mapped software queues.
3831 * This is more accurate and more efficient than looping
3832 * over all possibly mapped software queues.
3834 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3837 * Initialize batch roundrobin counts
3839 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3840 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3845 * Caller needs to ensure that we're either frozen/quiesced, or that
3846 * the queue isn't live yet.
3848 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3850 struct blk_mq_hw_ctx *hctx;
3853 queue_for_each_hw_ctx(q, hctx, i) {
3855 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3857 blk_mq_tag_idle(hctx);
3858 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3863 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3866 struct request_queue *q;
3868 lockdep_assert_held(&set->tag_list_lock);
3870 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3871 blk_mq_freeze_queue(q);
3872 queue_set_hctx_shared(q, shared);
3873 blk_mq_unfreeze_queue(q);
3877 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3879 struct blk_mq_tag_set *set = q->tag_set;
3881 mutex_lock(&set->tag_list_lock);
3882 list_del(&q->tag_set_list);
3883 if (list_is_singular(&set->tag_list)) {
3884 /* just transitioned to unshared */
3885 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3886 /* update existing queue */
3887 blk_mq_update_tag_set_shared(set, false);
3889 mutex_unlock(&set->tag_list_lock);
3890 INIT_LIST_HEAD(&q->tag_set_list);
3893 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3894 struct request_queue *q)
3896 mutex_lock(&set->tag_list_lock);
3899 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3901 if (!list_empty(&set->tag_list) &&
3902 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3903 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3904 /* update existing queue */
3905 blk_mq_update_tag_set_shared(set, true);
3907 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3908 queue_set_hctx_shared(q, true);
3909 list_add_tail(&q->tag_set_list, &set->tag_list);
3911 mutex_unlock(&set->tag_list_lock);
3914 /* All allocations will be freed in release handler of q->mq_kobj */
3915 static int blk_mq_alloc_ctxs(struct request_queue *q)
3917 struct blk_mq_ctxs *ctxs;
3920 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3924 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3925 if (!ctxs->queue_ctx)
3928 for_each_possible_cpu(cpu) {
3929 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3933 q->mq_kobj = &ctxs->kobj;
3934 q->queue_ctx = ctxs->queue_ctx;
3943 * It is the actual release handler for mq, but we do it from
3944 * request queue's release handler for avoiding use-after-free
3945 * and headache because q->mq_kobj shouldn't have been introduced,
3946 * but we can't group ctx/kctx kobj without it.
3948 void blk_mq_release(struct request_queue *q)
3950 struct blk_mq_hw_ctx *hctx, *next;
3953 queue_for_each_hw_ctx(q, hctx, i)
3954 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3956 /* all hctx are in .unused_hctx_list now */
3957 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3958 list_del_init(&hctx->hctx_list);
3959 kobject_put(&hctx->kobj);
3962 xa_destroy(&q->hctx_table);
3965 * release .mq_kobj and sw queue's kobject now because
3966 * both share lifetime with request queue.
3968 blk_mq_sysfs_deinit(q);
3971 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3974 struct request_queue *q;
3977 q = blk_alloc_queue(set->numa_node, set->flags & BLK_MQ_F_BLOCKING);
3979 return ERR_PTR(-ENOMEM);
3980 q->queuedata = queuedata;
3981 ret = blk_mq_init_allocated_queue(set, q);
3984 return ERR_PTR(ret);
3989 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3991 return blk_mq_init_queue_data(set, NULL);
3993 EXPORT_SYMBOL(blk_mq_init_queue);
3996 * blk_mq_destroy_queue - shutdown a request queue
3997 * @q: request queue to shutdown
3999 * This shuts down a request queue allocated by blk_mq_init_queue() and drops
4000 * the initial reference. All future requests will failed with -ENODEV.
4002 * Context: can sleep
4004 void blk_mq_destroy_queue(struct request_queue *q)
4006 WARN_ON_ONCE(!queue_is_mq(q));
4007 WARN_ON_ONCE(blk_queue_registered(q));
4011 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4012 blk_queue_start_drain(q);
4013 blk_freeze_queue(q);
4016 blk_mq_cancel_work_sync(q);
4017 blk_mq_exit_queue(q);
4019 /* @q is and will stay empty, shutdown and put */
4022 EXPORT_SYMBOL(blk_mq_destroy_queue);
4024 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4025 struct lock_class_key *lkclass)
4027 struct request_queue *q;
4028 struct gendisk *disk;
4030 q = blk_mq_init_queue_data(set, queuedata);
4034 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4036 blk_mq_destroy_queue(q);
4037 return ERR_PTR(-ENOMEM);
4039 set_bit(GD_OWNS_QUEUE, &disk->state);
4042 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4044 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4045 struct lock_class_key *lkclass)
4047 if (!blk_get_queue(q))
4049 return __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4051 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4053 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4054 struct blk_mq_tag_set *set, struct request_queue *q,
4055 int hctx_idx, int node)
4057 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4059 /* reuse dead hctx first */
4060 spin_lock(&q->unused_hctx_lock);
4061 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4062 if (tmp->numa_node == node) {
4068 list_del_init(&hctx->hctx_list);
4069 spin_unlock(&q->unused_hctx_lock);
4072 hctx = blk_mq_alloc_hctx(q, set, node);
4076 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4082 kobject_put(&hctx->kobj);
4087 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4088 struct request_queue *q)
4090 struct blk_mq_hw_ctx *hctx;
4093 /* protect against switching io scheduler */
4094 mutex_lock(&q->sysfs_lock);
4095 for (i = 0; i < set->nr_hw_queues; i++) {
4097 int node = blk_mq_get_hctx_node(set, i);
4098 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4101 old_node = old_hctx->numa_node;
4102 blk_mq_exit_hctx(q, set, old_hctx, i);
4105 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4108 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4110 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4111 WARN_ON_ONCE(!hctx);
4115 * Increasing nr_hw_queues fails. Free the newly allocated
4116 * hctxs and keep the previous q->nr_hw_queues.
4118 if (i != set->nr_hw_queues) {
4119 j = q->nr_hw_queues;
4122 q->nr_hw_queues = set->nr_hw_queues;
4125 xa_for_each_start(&q->hctx_table, j, hctx, j)
4126 blk_mq_exit_hctx(q, set, hctx, j);
4127 mutex_unlock(&q->sysfs_lock);
4130 static void blk_mq_update_poll_flag(struct request_queue *q)
4132 struct blk_mq_tag_set *set = q->tag_set;
4134 if (set->nr_maps > HCTX_TYPE_POLL &&
4135 set->map[HCTX_TYPE_POLL].nr_queues)
4136 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4138 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4141 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4142 struct request_queue *q)
4144 WARN_ON_ONCE(blk_queue_has_srcu(q) !=
4145 !!(set->flags & BLK_MQ_F_BLOCKING));
4147 /* mark the queue as mq asap */
4148 q->mq_ops = set->ops;
4150 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
4151 blk_mq_poll_stats_bkt,
4152 BLK_MQ_POLL_STATS_BKTS, q);
4156 if (blk_mq_alloc_ctxs(q))
4159 /* init q->mq_kobj and sw queues' kobjects */
4160 blk_mq_sysfs_init(q);
4162 INIT_LIST_HEAD(&q->unused_hctx_list);
4163 spin_lock_init(&q->unused_hctx_lock);
4165 xa_init(&q->hctx_table);
4167 blk_mq_realloc_hw_ctxs(set, q);
4168 if (!q->nr_hw_queues)
4171 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4172 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4176 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4177 blk_mq_update_poll_flag(q);
4179 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4180 INIT_LIST_HEAD(&q->requeue_list);
4181 spin_lock_init(&q->requeue_lock);
4183 q->nr_requests = set->queue_depth;
4186 * Default to classic polling
4188 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
4190 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4191 blk_mq_add_queue_tag_set(set, q);
4192 blk_mq_map_swqueue(q);
4196 xa_destroy(&q->hctx_table);
4197 q->nr_hw_queues = 0;
4198 blk_mq_sysfs_deinit(q);
4200 blk_stat_free_callback(q->poll_cb);
4206 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4208 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4209 void blk_mq_exit_queue(struct request_queue *q)
4211 struct blk_mq_tag_set *set = q->tag_set;
4213 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4214 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4215 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4216 blk_mq_del_queue_tag_set(q);
4219 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4223 if (blk_mq_is_shared_tags(set->flags)) {
4224 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4227 if (!set->shared_tags)
4231 for (i = 0; i < set->nr_hw_queues; i++) {
4232 if (!__blk_mq_alloc_map_and_rqs(set, i))
4241 __blk_mq_free_map_and_rqs(set, i);
4243 if (blk_mq_is_shared_tags(set->flags)) {
4244 blk_mq_free_map_and_rqs(set, set->shared_tags,
4245 BLK_MQ_NO_HCTX_IDX);
4252 * Allocate the request maps associated with this tag_set. Note that this
4253 * may reduce the depth asked for, if memory is tight. set->queue_depth
4254 * will be updated to reflect the allocated depth.
4256 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4261 depth = set->queue_depth;
4263 err = __blk_mq_alloc_rq_maps(set);
4267 set->queue_depth >>= 1;
4268 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4272 } while (set->queue_depth);
4274 if (!set->queue_depth || err) {
4275 pr_err("blk-mq: failed to allocate request map\n");
4279 if (depth != set->queue_depth)
4280 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4281 depth, set->queue_depth);
4286 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4289 * blk_mq_map_queues() and multiple .map_queues() implementations
4290 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4291 * number of hardware queues.
4293 if (set->nr_maps == 1)
4294 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4296 if (set->ops->map_queues && !is_kdump_kernel()) {
4300 * transport .map_queues is usually done in the following
4303 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4304 * mask = get_cpu_mask(queue)
4305 * for_each_cpu(cpu, mask)
4306 * set->map[x].mq_map[cpu] = queue;
4309 * When we need to remap, the table has to be cleared for
4310 * killing stale mapping since one CPU may not be mapped
4313 for (i = 0; i < set->nr_maps; i++)
4314 blk_mq_clear_mq_map(&set->map[i]);
4316 set->ops->map_queues(set);
4318 BUG_ON(set->nr_maps > 1);
4319 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4323 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4324 int cur_nr_hw_queues, int new_nr_hw_queues)
4326 struct blk_mq_tags **new_tags;
4328 if (cur_nr_hw_queues >= new_nr_hw_queues)
4331 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4332 GFP_KERNEL, set->numa_node);
4337 memcpy(new_tags, set->tags, cur_nr_hw_queues *
4338 sizeof(*set->tags));
4340 set->tags = new_tags;
4341 set->nr_hw_queues = new_nr_hw_queues;
4346 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
4347 int new_nr_hw_queues)
4349 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
4353 * Alloc a tag set to be associated with one or more request queues.
4354 * May fail with EINVAL for various error conditions. May adjust the
4355 * requested depth down, if it's too large. In that case, the set
4356 * value will be stored in set->queue_depth.
4358 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4362 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4364 if (!set->nr_hw_queues)
4366 if (!set->queue_depth)
4368 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4371 if (!set->ops->queue_rq)
4374 if (!set->ops->get_budget ^ !set->ops->put_budget)
4377 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4378 pr_info("blk-mq: reduced tag depth to %u\n",
4380 set->queue_depth = BLK_MQ_MAX_DEPTH;
4385 else if (set->nr_maps > HCTX_MAX_TYPES)
4389 * If a crashdump is active, then we are potentially in a very
4390 * memory constrained environment. Limit us to 1 queue and
4391 * 64 tags to prevent using too much memory.
4393 if (is_kdump_kernel()) {
4394 set->nr_hw_queues = 1;
4396 set->queue_depth = min(64U, set->queue_depth);
4399 * There is no use for more h/w queues than cpus if we just have
4402 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4403 set->nr_hw_queues = nr_cpu_ids;
4405 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
4409 for (i = 0; i < set->nr_maps; i++) {
4410 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4411 sizeof(set->map[i].mq_map[0]),
4412 GFP_KERNEL, set->numa_node);
4413 if (!set->map[i].mq_map)
4414 goto out_free_mq_map;
4415 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4418 blk_mq_update_queue_map(set);
4420 ret = blk_mq_alloc_set_map_and_rqs(set);
4422 goto out_free_mq_map;
4424 mutex_init(&set->tag_list_lock);
4425 INIT_LIST_HEAD(&set->tag_list);
4430 for (i = 0; i < set->nr_maps; i++) {
4431 kfree(set->map[i].mq_map);
4432 set->map[i].mq_map = NULL;
4438 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4440 /* allocate and initialize a tagset for a simple single-queue device */
4441 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4442 const struct blk_mq_ops *ops, unsigned int queue_depth,
4443 unsigned int set_flags)
4445 memset(set, 0, sizeof(*set));
4447 set->nr_hw_queues = 1;
4449 set->queue_depth = queue_depth;
4450 set->numa_node = NUMA_NO_NODE;
4451 set->flags = set_flags;
4452 return blk_mq_alloc_tag_set(set);
4454 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4456 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4460 for (i = 0; i < set->nr_hw_queues; i++)
4461 __blk_mq_free_map_and_rqs(set, i);
4463 if (blk_mq_is_shared_tags(set->flags)) {
4464 blk_mq_free_map_and_rqs(set, set->shared_tags,
4465 BLK_MQ_NO_HCTX_IDX);
4468 for (j = 0; j < set->nr_maps; j++) {
4469 kfree(set->map[j].mq_map);
4470 set->map[j].mq_map = NULL;
4476 EXPORT_SYMBOL(blk_mq_free_tag_set);
4478 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4480 struct blk_mq_tag_set *set = q->tag_set;
4481 struct blk_mq_hw_ctx *hctx;
4488 if (q->nr_requests == nr)
4491 blk_mq_freeze_queue(q);
4492 blk_mq_quiesce_queue(q);
4495 queue_for_each_hw_ctx(q, hctx, i) {
4499 * If we're using an MQ scheduler, just update the scheduler
4500 * queue depth. This is similar to what the old code would do.
4502 if (hctx->sched_tags) {
4503 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4506 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4511 if (q->elevator && q->elevator->type->ops.depth_updated)
4512 q->elevator->type->ops.depth_updated(hctx);
4515 q->nr_requests = nr;
4516 if (blk_mq_is_shared_tags(set->flags)) {
4518 blk_mq_tag_update_sched_shared_tags(q);
4520 blk_mq_tag_resize_shared_tags(set, nr);
4524 blk_mq_unquiesce_queue(q);
4525 blk_mq_unfreeze_queue(q);
4531 * request_queue and elevator_type pair.
4532 * It is just used by __blk_mq_update_nr_hw_queues to cache
4533 * the elevator_type associated with a request_queue.
4535 struct blk_mq_qe_pair {
4536 struct list_head node;
4537 struct request_queue *q;
4538 struct elevator_type *type;
4542 * Cache the elevator_type in qe pair list and switch the
4543 * io scheduler to 'none'
4545 static bool blk_mq_elv_switch_none(struct list_head *head,
4546 struct request_queue *q)
4548 struct blk_mq_qe_pair *qe;
4553 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4557 /* q->elevator needs protection from ->sysfs_lock */
4558 mutex_lock(&q->sysfs_lock);
4560 INIT_LIST_HEAD(&qe->node);
4562 qe->type = q->elevator->type;
4563 list_add(&qe->node, head);
4566 * After elevator_switch, the previous elevator_queue will be
4567 * released by elevator_release. The reference of the io scheduler
4568 * module get by elevator_get will also be put. So we need to get
4569 * a reference of the io scheduler module here to prevent it to be
4572 __module_get(qe->type->elevator_owner);
4573 elevator_switch(q, NULL);
4574 mutex_unlock(&q->sysfs_lock);
4579 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4580 struct request_queue *q)
4582 struct blk_mq_qe_pair *qe;
4584 list_for_each_entry(qe, head, node)
4591 static void blk_mq_elv_switch_back(struct list_head *head,
4592 struct request_queue *q)
4594 struct blk_mq_qe_pair *qe;
4595 struct elevator_type *t;
4597 qe = blk_lookup_qe_pair(head, q);
4601 list_del(&qe->node);
4604 mutex_lock(&q->sysfs_lock);
4605 elevator_switch(q, t);
4606 mutex_unlock(&q->sysfs_lock);
4609 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4612 struct request_queue *q;
4614 int prev_nr_hw_queues;
4616 lockdep_assert_held(&set->tag_list_lock);
4618 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4619 nr_hw_queues = nr_cpu_ids;
4620 if (nr_hw_queues < 1)
4622 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4625 list_for_each_entry(q, &set->tag_list, tag_set_list)
4626 blk_mq_freeze_queue(q);
4628 * Switch IO scheduler to 'none', cleaning up the data associated
4629 * with the previous scheduler. We will switch back once we are done
4630 * updating the new sw to hw queue mappings.
4632 list_for_each_entry(q, &set->tag_list, tag_set_list)
4633 if (!blk_mq_elv_switch_none(&head, q))
4636 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4637 blk_mq_debugfs_unregister_hctxs(q);
4638 blk_mq_sysfs_unregister_hctxs(q);
4641 prev_nr_hw_queues = set->nr_hw_queues;
4642 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
4646 set->nr_hw_queues = nr_hw_queues;
4648 blk_mq_update_queue_map(set);
4649 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4650 blk_mq_realloc_hw_ctxs(set, q);
4651 blk_mq_update_poll_flag(q);
4652 if (q->nr_hw_queues != set->nr_hw_queues) {
4653 int i = prev_nr_hw_queues;
4655 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4656 nr_hw_queues, prev_nr_hw_queues);
4657 for (; i < set->nr_hw_queues; i++)
4658 __blk_mq_free_map_and_rqs(set, i);
4660 set->nr_hw_queues = prev_nr_hw_queues;
4661 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4664 blk_mq_map_swqueue(q);
4668 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4669 blk_mq_sysfs_register_hctxs(q);
4670 blk_mq_debugfs_register_hctxs(q);
4674 list_for_each_entry(q, &set->tag_list, tag_set_list)
4675 blk_mq_elv_switch_back(&head, q);
4677 list_for_each_entry(q, &set->tag_list, tag_set_list)
4678 blk_mq_unfreeze_queue(q);
4681 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4683 mutex_lock(&set->tag_list_lock);
4684 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4685 mutex_unlock(&set->tag_list_lock);
4687 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4689 /* Enable polling stats and return whether they were already enabled. */
4690 static bool blk_poll_stats_enable(struct request_queue *q)
4695 return blk_stats_alloc_enable(q);
4698 static void blk_mq_poll_stats_start(struct request_queue *q)
4701 * We don't arm the callback if polling stats are not enabled or the
4702 * callback is already active.
4704 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
4707 blk_stat_activate_msecs(q->poll_cb, 100);
4710 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4712 struct request_queue *q = cb->data;
4715 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4716 if (cb->stat[bucket].nr_samples)
4717 q->poll_stat[bucket] = cb->stat[bucket];
4721 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4724 unsigned long ret = 0;
4728 * If stats collection isn't on, don't sleep but turn it on for
4731 if (!blk_poll_stats_enable(q))
4735 * As an optimistic guess, use half of the mean service time
4736 * for this type of request. We can (and should) make this smarter.
4737 * For instance, if the completion latencies are tight, we can
4738 * get closer than just half the mean. This is especially
4739 * important on devices where the completion latencies are longer
4740 * than ~10 usec. We do use the stats for the relevant IO size
4741 * if available which does lead to better estimates.
4743 bucket = blk_mq_poll_stats_bkt(rq);
4747 if (q->poll_stat[bucket].nr_samples)
4748 ret = (q->poll_stat[bucket].mean + 1) / 2;
4753 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4755 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4756 struct request *rq = blk_qc_to_rq(hctx, qc);
4757 struct hrtimer_sleeper hs;
4758 enum hrtimer_mode mode;
4763 * If a request has completed on queue that uses an I/O scheduler, we
4764 * won't get back a request from blk_qc_to_rq.
4766 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4770 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4772 * 0: use half of prev avg
4773 * >0: use this specific value
4775 if (q->poll_nsec > 0)
4776 nsecs = q->poll_nsec;
4778 nsecs = blk_mq_poll_nsecs(q, rq);
4783 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4786 * This will be replaced with the stats tracking code, using
4787 * 'avg_completion_time / 2' as the pre-sleep target.
4791 mode = HRTIMER_MODE_REL;
4792 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4793 hrtimer_set_expires(&hs.timer, kt);
4796 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4798 set_current_state(TASK_UNINTERRUPTIBLE);
4799 hrtimer_sleeper_start_expires(&hs, mode);
4802 hrtimer_cancel(&hs.timer);
4803 mode = HRTIMER_MODE_ABS;
4804 } while (hs.task && !signal_pending(current));
4806 __set_current_state(TASK_RUNNING);
4807 destroy_hrtimer_on_stack(&hs.timer);
4810 * If we sleep, have the caller restart the poll loop to reset the
4811 * state. Like for the other success return cases, the caller is
4812 * responsible for checking if the IO completed. If the IO isn't
4813 * complete, we'll get called again and will go straight to the busy
4819 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4820 struct io_comp_batch *iob, unsigned int flags)
4822 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4823 long state = get_current_state();
4827 ret = q->mq_ops->poll(hctx, iob);
4829 __set_current_state(TASK_RUNNING);
4833 if (signal_pending_state(state, current))
4834 __set_current_state(TASK_RUNNING);
4835 if (task_is_running(current))
4838 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4841 } while (!need_resched());
4843 __set_current_state(TASK_RUNNING);
4847 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4850 if (!(flags & BLK_POLL_NOSLEEP) &&
4851 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4852 if (blk_mq_poll_hybrid(q, cookie))
4855 return blk_mq_poll_classic(q, cookie, iob, flags);
4858 unsigned int blk_mq_rq_cpu(struct request *rq)
4860 return rq->mq_ctx->cpu;
4862 EXPORT_SYMBOL(blk_mq_rq_cpu);
4864 void blk_mq_cancel_work_sync(struct request_queue *q)
4866 if (queue_is_mq(q)) {
4867 struct blk_mq_hw_ctx *hctx;
4870 cancel_delayed_work_sync(&q->requeue_work);
4872 queue_for_each_hw_ctx(q, hctx, i)
4873 cancel_delayed_work_sync(&hctx->run_work);
4877 static int __init blk_mq_init(void)
4881 for_each_possible_cpu(i)
4882 init_llist_head(&per_cpu(blk_cpu_done, i));
4883 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4885 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4886 "block/softirq:dead", NULL,
4887 blk_softirq_cpu_dead);
4888 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4889 blk_mq_hctx_notify_dead);
4890 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4891 blk_mq_hctx_notify_online,
4892 blk_mq_hctx_notify_offline);
4895 subsys_initcall(blk_mq_init);