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);
1266 trace_block_plug(rq->q);
1269 if (!plug->multiple_queues && last && last->q != rq->q)
1270 plug->multiple_queues = true;
1271 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1272 plug->has_elevator = true;
1274 rq_list_add(&plug->mq_list, rq);
1279 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1280 * @rq: request to insert
1281 * @at_head: insert request at head or tail of queue
1284 * Insert a fully prepared request at the back of the I/O scheduler queue
1285 * for execution. Don't wait for completion.
1288 * This function will invoke @done directly if the queue is dead.
1290 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1292 WARN_ON(irqs_disabled());
1293 WARN_ON(!blk_rq_is_passthrough(rq));
1295 blk_account_io_start(rq);
1298 * As plugging can be enabled for passthrough requests on a zoned
1299 * device, directly accessing the plug instead of using blk_mq_plug()
1300 * should not have any consequences.
1303 blk_add_rq_to_plug(current->plug, rq);
1305 blk_mq_sched_insert_request(rq, at_head, true, false);
1307 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1309 struct blk_rq_wait {
1310 struct completion done;
1314 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1316 struct blk_rq_wait *wait = rq->end_io_data;
1319 complete(&wait->done);
1320 return RQ_END_IO_NONE;
1323 bool blk_rq_is_poll(struct request *rq)
1327 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1329 if (WARN_ON_ONCE(!rq->bio))
1333 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1335 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1338 bio_poll(rq->bio, NULL, 0);
1340 } while (!completion_done(wait));
1344 * blk_execute_rq - insert a request into queue for execution
1345 * @rq: request to insert
1346 * @at_head: insert request at head or tail of queue
1349 * Insert a fully prepared request at the back of the I/O scheduler queue
1350 * for execution and wait for completion.
1351 * Return: The blk_status_t result provided to blk_mq_end_request().
1353 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1355 struct blk_rq_wait wait = {
1356 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1359 WARN_ON(irqs_disabled());
1360 WARN_ON(!blk_rq_is_passthrough(rq));
1362 rq->end_io_data = &wait;
1363 rq->end_io = blk_end_sync_rq;
1365 blk_account_io_start(rq);
1366 blk_mq_sched_insert_request(rq, at_head, true, false);
1368 if (blk_rq_is_poll(rq)) {
1369 blk_rq_poll_completion(rq, &wait.done);
1372 * Prevent hang_check timer from firing at us during very long
1375 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1378 while (!wait_for_completion_io_timeout(&wait.done,
1379 hang_check * (HZ/2)))
1382 wait_for_completion_io(&wait.done);
1387 EXPORT_SYMBOL(blk_execute_rq);
1389 static void __blk_mq_requeue_request(struct request *rq)
1391 struct request_queue *q = rq->q;
1393 blk_mq_put_driver_tag(rq);
1395 trace_block_rq_requeue(rq);
1396 rq_qos_requeue(q, rq);
1398 if (blk_mq_request_started(rq)) {
1399 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1400 rq->rq_flags &= ~RQF_TIMED_OUT;
1404 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1406 __blk_mq_requeue_request(rq);
1408 /* this request will be re-inserted to io scheduler queue */
1409 blk_mq_sched_requeue_request(rq);
1411 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1413 EXPORT_SYMBOL(blk_mq_requeue_request);
1415 static void blk_mq_requeue_work(struct work_struct *work)
1417 struct request_queue *q =
1418 container_of(work, struct request_queue, requeue_work.work);
1420 struct request *rq, *next;
1422 spin_lock_irq(&q->requeue_lock);
1423 list_splice_init(&q->requeue_list, &rq_list);
1424 spin_unlock_irq(&q->requeue_lock);
1426 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1427 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1430 rq->rq_flags &= ~RQF_SOFTBARRIER;
1431 list_del_init(&rq->queuelist);
1433 * If RQF_DONTPREP, rq has contained some driver specific
1434 * data, so insert it to hctx dispatch list to avoid any
1437 if (rq->rq_flags & RQF_DONTPREP)
1438 blk_mq_request_bypass_insert(rq, false, false);
1440 blk_mq_sched_insert_request(rq, true, false, false);
1443 while (!list_empty(&rq_list)) {
1444 rq = list_entry(rq_list.next, struct request, queuelist);
1445 list_del_init(&rq->queuelist);
1446 blk_mq_sched_insert_request(rq, false, false, false);
1449 blk_mq_run_hw_queues(q, false);
1452 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1453 bool kick_requeue_list)
1455 struct request_queue *q = rq->q;
1456 unsigned long flags;
1459 * We abuse this flag that is otherwise used by the I/O scheduler to
1460 * request head insertion from the workqueue.
1462 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1464 spin_lock_irqsave(&q->requeue_lock, flags);
1466 rq->rq_flags |= RQF_SOFTBARRIER;
1467 list_add(&rq->queuelist, &q->requeue_list);
1469 list_add_tail(&rq->queuelist, &q->requeue_list);
1471 spin_unlock_irqrestore(&q->requeue_lock, flags);
1473 if (kick_requeue_list)
1474 blk_mq_kick_requeue_list(q);
1477 void blk_mq_kick_requeue_list(struct request_queue *q)
1479 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1481 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1483 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1484 unsigned long msecs)
1486 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1487 msecs_to_jiffies(msecs));
1489 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1491 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1494 * If we find a request that isn't idle we know the queue is busy
1495 * as it's checked in the iter.
1496 * Return false to stop the iteration.
1498 if (blk_mq_request_started(rq)) {
1508 bool blk_mq_queue_inflight(struct request_queue *q)
1512 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1515 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1517 static void blk_mq_rq_timed_out(struct request *req)
1519 req->rq_flags |= RQF_TIMED_OUT;
1520 if (req->q->mq_ops->timeout) {
1521 enum blk_eh_timer_return ret;
1523 ret = req->q->mq_ops->timeout(req);
1524 if (ret == BLK_EH_DONE)
1526 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1532 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
1534 unsigned long deadline;
1536 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1538 if (rq->rq_flags & RQF_TIMED_OUT)
1541 deadline = READ_ONCE(rq->deadline);
1542 if (time_after_eq(jiffies, deadline))
1547 else if (time_after(*next, deadline))
1552 void blk_mq_put_rq_ref(struct request *rq)
1554 if (is_flush_rq(rq)) {
1555 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1556 blk_mq_free_request(rq);
1557 } else if (req_ref_put_and_test(rq)) {
1558 __blk_mq_free_request(rq);
1562 static bool blk_mq_check_expired(struct request *rq, void *priv)
1564 unsigned long *next = priv;
1567 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1568 * be reallocated underneath the timeout handler's processing, then
1569 * the expire check is reliable. If the request is not expired, then
1570 * it was completed and reallocated as a new request after returning
1571 * from blk_mq_check_expired().
1573 if (blk_mq_req_expired(rq, next))
1574 blk_mq_rq_timed_out(rq);
1578 static void blk_mq_timeout_work(struct work_struct *work)
1580 struct request_queue *q =
1581 container_of(work, struct request_queue, timeout_work);
1582 unsigned long next = 0;
1583 struct blk_mq_hw_ctx *hctx;
1586 /* A deadlock might occur if a request is stuck requiring a
1587 * timeout at the same time a queue freeze is waiting
1588 * completion, since the timeout code would not be able to
1589 * acquire the queue reference here.
1591 * That's why we don't use blk_queue_enter here; instead, we use
1592 * percpu_ref_tryget directly, because we need to be able to
1593 * obtain a reference even in the short window between the queue
1594 * starting to freeze, by dropping the first reference in
1595 * blk_freeze_queue_start, and the moment the last request is
1596 * consumed, marked by the instant q_usage_counter reaches
1599 if (!percpu_ref_tryget(&q->q_usage_counter))
1602 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
1605 mod_timer(&q->timeout, next);
1608 * Request timeouts are handled as a forward rolling timer. If
1609 * we end up here it means that no requests are pending and
1610 * also that no request has been pending for a while. Mark
1611 * each hctx as idle.
1613 queue_for_each_hw_ctx(q, hctx, i) {
1614 /* the hctx may be unmapped, so check it here */
1615 if (blk_mq_hw_queue_mapped(hctx))
1616 blk_mq_tag_idle(hctx);
1622 struct flush_busy_ctx_data {
1623 struct blk_mq_hw_ctx *hctx;
1624 struct list_head *list;
1627 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1629 struct flush_busy_ctx_data *flush_data = data;
1630 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1631 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1632 enum hctx_type type = hctx->type;
1634 spin_lock(&ctx->lock);
1635 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1636 sbitmap_clear_bit(sb, bitnr);
1637 spin_unlock(&ctx->lock);
1642 * Process software queues that have been marked busy, splicing them
1643 * to the for-dispatch
1645 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1647 struct flush_busy_ctx_data data = {
1652 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1654 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1656 struct dispatch_rq_data {
1657 struct blk_mq_hw_ctx *hctx;
1661 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1664 struct dispatch_rq_data *dispatch_data = data;
1665 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1666 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1667 enum hctx_type type = hctx->type;
1669 spin_lock(&ctx->lock);
1670 if (!list_empty(&ctx->rq_lists[type])) {
1671 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1672 list_del_init(&dispatch_data->rq->queuelist);
1673 if (list_empty(&ctx->rq_lists[type]))
1674 sbitmap_clear_bit(sb, bitnr);
1676 spin_unlock(&ctx->lock);
1678 return !dispatch_data->rq;
1681 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1682 struct blk_mq_ctx *start)
1684 unsigned off = start ? start->index_hw[hctx->type] : 0;
1685 struct dispatch_rq_data data = {
1690 __sbitmap_for_each_set(&hctx->ctx_map, off,
1691 dispatch_rq_from_ctx, &data);
1696 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1698 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1699 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1702 blk_mq_tag_busy(rq->mq_hctx);
1704 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1705 bt = &rq->mq_hctx->tags->breserved_tags;
1708 if (!hctx_may_queue(rq->mq_hctx, bt))
1712 tag = __sbitmap_queue_get(bt);
1713 if (tag == BLK_MQ_NO_TAG)
1716 rq->tag = tag + tag_offset;
1720 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1722 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1725 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1726 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1727 rq->rq_flags |= RQF_MQ_INFLIGHT;
1728 __blk_mq_inc_active_requests(hctx);
1730 hctx->tags->rqs[rq->tag] = rq;
1734 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1735 int flags, void *key)
1737 struct blk_mq_hw_ctx *hctx;
1739 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1741 spin_lock(&hctx->dispatch_wait_lock);
1742 if (!list_empty(&wait->entry)) {
1743 struct sbitmap_queue *sbq;
1745 list_del_init(&wait->entry);
1746 sbq = &hctx->tags->bitmap_tags;
1747 atomic_dec(&sbq->ws_active);
1749 spin_unlock(&hctx->dispatch_wait_lock);
1751 blk_mq_run_hw_queue(hctx, true);
1756 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1757 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1758 * restart. For both cases, take care to check the condition again after
1759 * marking us as waiting.
1761 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1764 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1765 struct wait_queue_head *wq;
1766 wait_queue_entry_t *wait;
1769 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1770 blk_mq_sched_mark_restart_hctx(hctx);
1773 * It's possible that a tag was freed in the window between the
1774 * allocation failure and adding the hardware queue to the wait
1777 * Don't clear RESTART here, someone else could have set it.
1778 * At most this will cost an extra queue run.
1780 return blk_mq_get_driver_tag(rq);
1783 wait = &hctx->dispatch_wait;
1784 if (!list_empty_careful(&wait->entry))
1787 wq = &bt_wait_ptr(sbq, hctx)->wait;
1789 spin_lock_irq(&wq->lock);
1790 spin_lock(&hctx->dispatch_wait_lock);
1791 if (!list_empty(&wait->entry)) {
1792 spin_unlock(&hctx->dispatch_wait_lock);
1793 spin_unlock_irq(&wq->lock);
1797 atomic_inc(&sbq->ws_active);
1798 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1799 __add_wait_queue(wq, wait);
1802 * It's possible that a tag was freed in the window between the
1803 * allocation failure and adding the hardware queue to the wait
1806 ret = blk_mq_get_driver_tag(rq);
1808 spin_unlock(&hctx->dispatch_wait_lock);
1809 spin_unlock_irq(&wq->lock);
1814 * We got a tag, remove ourselves from the wait queue to ensure
1815 * someone else gets the wakeup.
1817 list_del_init(&wait->entry);
1818 atomic_dec(&sbq->ws_active);
1819 spin_unlock(&hctx->dispatch_wait_lock);
1820 spin_unlock_irq(&wq->lock);
1825 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1826 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1828 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1829 * - EWMA is one simple way to compute running average value
1830 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1831 * - take 4 as factor for avoiding to get too small(0) result, and this
1832 * factor doesn't matter because EWMA decreases exponentially
1834 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1838 ewma = hctx->dispatch_busy;
1843 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1845 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1846 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1848 hctx->dispatch_busy = ewma;
1851 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1853 static void blk_mq_handle_dev_resource(struct request *rq,
1854 struct list_head *list)
1856 struct request *next =
1857 list_first_entry_or_null(list, struct request, queuelist);
1860 * If an I/O scheduler has been configured and we got a driver tag for
1861 * the next request already, free it.
1864 blk_mq_put_driver_tag(next);
1866 list_add(&rq->queuelist, list);
1867 __blk_mq_requeue_request(rq);
1870 static void blk_mq_handle_zone_resource(struct request *rq,
1871 struct list_head *zone_list)
1874 * If we end up here it is because we cannot dispatch a request to a
1875 * specific zone due to LLD level zone-write locking or other zone
1876 * related resource not being available. In this case, set the request
1877 * aside in zone_list for retrying it later.
1879 list_add(&rq->queuelist, zone_list);
1880 __blk_mq_requeue_request(rq);
1883 enum prep_dispatch {
1885 PREP_DISPATCH_NO_TAG,
1886 PREP_DISPATCH_NO_BUDGET,
1889 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1892 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1893 int budget_token = -1;
1896 budget_token = blk_mq_get_dispatch_budget(rq->q);
1897 if (budget_token < 0) {
1898 blk_mq_put_driver_tag(rq);
1899 return PREP_DISPATCH_NO_BUDGET;
1901 blk_mq_set_rq_budget_token(rq, budget_token);
1904 if (!blk_mq_get_driver_tag(rq)) {
1906 * The initial allocation attempt failed, so we need to
1907 * rerun the hardware queue when a tag is freed. The
1908 * waitqueue takes care of that. If the queue is run
1909 * before we add this entry back on the dispatch list,
1910 * we'll re-run it below.
1912 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1914 * All budgets not got from this function will be put
1915 * together during handling partial dispatch
1918 blk_mq_put_dispatch_budget(rq->q, budget_token);
1919 return PREP_DISPATCH_NO_TAG;
1923 return PREP_DISPATCH_OK;
1926 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1927 static void blk_mq_release_budgets(struct request_queue *q,
1928 struct list_head *list)
1932 list_for_each_entry(rq, list, queuelist) {
1933 int budget_token = blk_mq_get_rq_budget_token(rq);
1935 if (budget_token >= 0)
1936 blk_mq_put_dispatch_budget(q, budget_token);
1941 * Returns true if we did some work AND can potentially do more.
1943 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1944 unsigned int nr_budgets)
1946 enum prep_dispatch prep;
1947 struct request_queue *q = hctx->queue;
1948 struct request *rq, *nxt;
1950 blk_status_t ret = BLK_STS_OK;
1951 LIST_HEAD(zone_list);
1952 bool needs_resource = false;
1954 if (list_empty(list))
1958 * Now process all the entries, sending them to the driver.
1960 errors = queued = 0;
1962 struct blk_mq_queue_data bd;
1964 rq = list_first_entry(list, struct request, queuelist);
1966 WARN_ON_ONCE(hctx != rq->mq_hctx);
1967 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1968 if (prep != PREP_DISPATCH_OK)
1971 list_del_init(&rq->queuelist);
1976 * Flag last if we have no more requests, or if we have more
1977 * but can't assign a driver tag to it.
1979 if (list_empty(list))
1982 nxt = list_first_entry(list, struct request, queuelist);
1983 bd.last = !blk_mq_get_driver_tag(nxt);
1987 * once the request is queued to lld, no need to cover the
1992 ret = q->mq_ops->queue_rq(hctx, &bd);
1997 case BLK_STS_RESOURCE:
1998 needs_resource = true;
2000 case BLK_STS_DEV_RESOURCE:
2001 blk_mq_handle_dev_resource(rq, list);
2003 case BLK_STS_ZONE_RESOURCE:
2005 * Move the request to zone_list and keep going through
2006 * the dispatch list to find more requests the drive can
2009 blk_mq_handle_zone_resource(rq, &zone_list);
2010 needs_resource = true;
2014 blk_mq_end_request(rq, ret);
2016 } while (!list_empty(list));
2018 if (!list_empty(&zone_list))
2019 list_splice_tail_init(&zone_list, list);
2021 /* If we didn't flush the entire list, we could have told the driver
2022 * there was more coming, but that turned out to be a lie.
2024 if ((!list_empty(list) || errors || needs_resource ||
2025 ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued)
2026 q->mq_ops->commit_rqs(hctx);
2028 * Any items that need requeuing? Stuff them into hctx->dispatch,
2029 * that is where we will continue on next queue run.
2031 if (!list_empty(list)) {
2033 /* For non-shared tags, the RESTART check will suffice */
2034 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2035 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
2038 blk_mq_release_budgets(q, list);
2040 spin_lock(&hctx->lock);
2041 list_splice_tail_init(list, &hctx->dispatch);
2042 spin_unlock(&hctx->lock);
2045 * Order adding requests to hctx->dispatch and checking
2046 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2047 * in blk_mq_sched_restart(). Avoid restart code path to
2048 * miss the new added requests to hctx->dispatch, meantime
2049 * SCHED_RESTART is observed here.
2054 * If SCHED_RESTART was set by the caller of this function and
2055 * it is no longer set that means that it was cleared by another
2056 * thread and hence that a queue rerun is needed.
2058 * If 'no_tag' is set, that means that we failed getting
2059 * a driver tag with an I/O scheduler attached. If our dispatch
2060 * waitqueue is no longer active, ensure that we run the queue
2061 * AFTER adding our entries back to the list.
2063 * If no I/O scheduler has been configured it is possible that
2064 * the hardware queue got stopped and restarted before requests
2065 * were pushed back onto the dispatch list. Rerun the queue to
2066 * avoid starvation. Notes:
2067 * - blk_mq_run_hw_queue() checks whether or not a queue has
2068 * been stopped before rerunning a queue.
2069 * - Some but not all block drivers stop a queue before
2070 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2073 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2074 * bit is set, run queue after a delay to avoid IO stalls
2075 * that could otherwise occur if the queue is idle. We'll do
2076 * similar if we couldn't get budget or couldn't lock a zone
2077 * and SCHED_RESTART is set.
2079 needs_restart = blk_mq_sched_needs_restart(hctx);
2080 if (prep == PREP_DISPATCH_NO_BUDGET)
2081 needs_resource = true;
2082 if (!needs_restart ||
2083 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2084 blk_mq_run_hw_queue(hctx, true);
2085 else if (needs_resource)
2086 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2088 blk_mq_update_dispatch_busy(hctx, true);
2091 blk_mq_update_dispatch_busy(hctx, false);
2093 return (queued + errors) != 0;
2097 * __blk_mq_run_hw_queue - Run a hardware queue.
2098 * @hctx: Pointer to the hardware queue to run.
2100 * Send pending requests to the hardware.
2102 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
2105 * We can't run the queue inline with ints disabled. Ensure that
2106 * we catch bad users of this early.
2108 WARN_ON_ONCE(in_interrupt());
2110 blk_mq_run_dispatch_ops(hctx->queue,
2111 blk_mq_sched_dispatch_requests(hctx));
2114 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2116 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2118 if (cpu >= nr_cpu_ids)
2119 cpu = cpumask_first(hctx->cpumask);
2124 * It'd be great if the workqueue API had a way to pass
2125 * in a mask and had some smarts for more clever placement.
2126 * For now we just round-robin here, switching for every
2127 * BLK_MQ_CPU_WORK_BATCH queued items.
2129 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2132 int next_cpu = hctx->next_cpu;
2134 if (hctx->queue->nr_hw_queues == 1)
2135 return WORK_CPU_UNBOUND;
2137 if (--hctx->next_cpu_batch <= 0) {
2139 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2141 if (next_cpu >= nr_cpu_ids)
2142 next_cpu = blk_mq_first_mapped_cpu(hctx);
2143 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2147 * Do unbound schedule if we can't find a online CPU for this hctx,
2148 * and it should only happen in the path of handling CPU DEAD.
2150 if (!cpu_online(next_cpu)) {
2157 * Make sure to re-select CPU next time once after CPUs
2158 * in hctx->cpumask become online again.
2160 hctx->next_cpu = next_cpu;
2161 hctx->next_cpu_batch = 1;
2162 return WORK_CPU_UNBOUND;
2165 hctx->next_cpu = next_cpu;
2170 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2171 * @hctx: Pointer to the hardware queue to run.
2172 * @async: If we want to run the queue asynchronously.
2173 * @msecs: Milliseconds of delay to wait before running the queue.
2175 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2176 * with a delay of @msecs.
2178 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2179 unsigned long msecs)
2181 if (unlikely(blk_mq_hctx_stopped(hctx)))
2184 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2185 if (cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2186 __blk_mq_run_hw_queue(hctx);
2191 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2192 msecs_to_jiffies(msecs));
2196 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2197 * @hctx: Pointer to the hardware queue to run.
2198 * @msecs: Milliseconds of delay to wait before running the queue.
2200 * Run a hardware queue asynchronously with a delay of @msecs.
2202 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2204 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2206 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2209 * blk_mq_run_hw_queue - Start to run a hardware queue.
2210 * @hctx: Pointer to the hardware queue to run.
2211 * @async: If we want to run the queue asynchronously.
2213 * Check if the request queue is not in a quiesced state and if there are
2214 * pending requests to be sent. If this is true, run the queue to send requests
2217 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2222 * When queue is quiesced, we may be switching io scheduler, or
2223 * updating nr_hw_queues, or other things, and we can't run queue
2224 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2226 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2229 __blk_mq_run_dispatch_ops(hctx->queue, false,
2230 need_run = !blk_queue_quiesced(hctx->queue) &&
2231 blk_mq_hctx_has_pending(hctx));
2234 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2236 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2239 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2242 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2244 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2246 * If the IO scheduler does not respect hardware queues when
2247 * dispatching, we just don't bother with multiple HW queues and
2248 * dispatch from hctx for the current CPU since running multiple queues
2249 * just causes lock contention inside the scheduler and pointless cache
2252 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2254 if (!blk_mq_hctx_stopped(hctx))
2260 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2261 * @q: Pointer to the request queue to run.
2262 * @async: If we want to run the queue asynchronously.
2264 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2266 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2270 if (blk_queue_sq_sched(q))
2271 sq_hctx = blk_mq_get_sq_hctx(q);
2272 queue_for_each_hw_ctx(q, hctx, i) {
2273 if (blk_mq_hctx_stopped(hctx))
2276 * Dispatch from this hctx either if there's no hctx preferred
2277 * by IO scheduler or if it has requests that bypass the
2280 if (!sq_hctx || sq_hctx == hctx ||
2281 !list_empty_careful(&hctx->dispatch))
2282 blk_mq_run_hw_queue(hctx, async);
2285 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2288 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2289 * @q: Pointer to the request queue to run.
2290 * @msecs: Milliseconds of delay to wait before running the queues.
2292 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2294 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2298 if (blk_queue_sq_sched(q))
2299 sq_hctx = blk_mq_get_sq_hctx(q);
2300 queue_for_each_hw_ctx(q, hctx, i) {
2301 if (blk_mq_hctx_stopped(hctx))
2304 * If there is already a run_work pending, leave the
2305 * pending delay untouched. Otherwise, a hctx can stall
2306 * if another hctx is re-delaying the other's work
2307 * before the work executes.
2309 if (delayed_work_pending(&hctx->run_work))
2312 * Dispatch from this hctx either if there's no hctx preferred
2313 * by IO scheduler or if it has requests that bypass the
2316 if (!sq_hctx || sq_hctx == hctx ||
2317 !list_empty_careful(&hctx->dispatch))
2318 blk_mq_delay_run_hw_queue(hctx, msecs);
2321 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2324 * This function is often used for pausing .queue_rq() by driver when
2325 * there isn't enough resource or some conditions aren't satisfied, and
2326 * BLK_STS_RESOURCE is usually returned.
2328 * We do not guarantee that dispatch can be drained or blocked
2329 * after blk_mq_stop_hw_queue() returns. Please use
2330 * blk_mq_quiesce_queue() for that requirement.
2332 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2334 cancel_delayed_work(&hctx->run_work);
2336 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2338 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2341 * This function is often used for pausing .queue_rq() by driver when
2342 * there isn't enough resource or some conditions aren't satisfied, and
2343 * BLK_STS_RESOURCE is usually returned.
2345 * We do not guarantee that dispatch can be drained or blocked
2346 * after blk_mq_stop_hw_queues() returns. Please use
2347 * blk_mq_quiesce_queue() for that requirement.
2349 void blk_mq_stop_hw_queues(struct request_queue *q)
2351 struct blk_mq_hw_ctx *hctx;
2354 queue_for_each_hw_ctx(q, hctx, i)
2355 blk_mq_stop_hw_queue(hctx);
2357 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2359 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2361 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2363 blk_mq_run_hw_queue(hctx, false);
2365 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2367 void blk_mq_start_hw_queues(struct request_queue *q)
2369 struct blk_mq_hw_ctx *hctx;
2372 queue_for_each_hw_ctx(q, hctx, i)
2373 blk_mq_start_hw_queue(hctx);
2375 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2377 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2379 if (!blk_mq_hctx_stopped(hctx))
2382 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2383 blk_mq_run_hw_queue(hctx, async);
2385 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2387 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2389 struct blk_mq_hw_ctx *hctx;
2392 queue_for_each_hw_ctx(q, hctx, i)
2393 blk_mq_start_stopped_hw_queue(hctx, async);
2395 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2397 static void blk_mq_run_work_fn(struct work_struct *work)
2399 struct blk_mq_hw_ctx *hctx;
2401 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2404 * If we are stopped, don't run the queue.
2406 if (blk_mq_hctx_stopped(hctx))
2409 __blk_mq_run_hw_queue(hctx);
2412 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2416 struct blk_mq_ctx *ctx = rq->mq_ctx;
2417 enum hctx_type type = hctx->type;
2419 lockdep_assert_held(&ctx->lock);
2421 trace_block_rq_insert(rq);
2424 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2426 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2429 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2432 struct blk_mq_ctx *ctx = rq->mq_ctx;
2434 lockdep_assert_held(&ctx->lock);
2436 __blk_mq_insert_req_list(hctx, rq, at_head);
2437 blk_mq_hctx_mark_pending(hctx, ctx);
2441 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2442 * @rq: Pointer to request to be inserted.
2443 * @at_head: true if the request should be inserted at the head of the list.
2444 * @run_queue: If we should run the hardware queue after inserting the request.
2446 * Should only be used carefully, when the caller knows we want to
2447 * bypass a potential IO scheduler on the target device.
2449 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2452 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2454 spin_lock(&hctx->lock);
2456 list_add(&rq->queuelist, &hctx->dispatch);
2458 list_add_tail(&rq->queuelist, &hctx->dispatch);
2459 spin_unlock(&hctx->lock);
2462 blk_mq_run_hw_queue(hctx, false);
2465 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2466 struct list_head *list)
2470 enum hctx_type type = hctx->type;
2473 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2476 list_for_each_entry(rq, list, queuelist) {
2477 BUG_ON(rq->mq_ctx != ctx);
2478 trace_block_rq_insert(rq);
2481 spin_lock(&ctx->lock);
2482 list_splice_tail_init(list, &ctx->rq_lists[type]);
2483 blk_mq_hctx_mark_pending(hctx, ctx);
2484 spin_unlock(&ctx->lock);
2487 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2490 if (hctx->queue->mq_ops->commit_rqs) {
2491 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2492 hctx->queue->mq_ops->commit_rqs(hctx);
2497 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2498 unsigned int nr_segs)
2502 if (bio->bi_opf & REQ_RAHEAD)
2503 rq->cmd_flags |= REQ_FAILFAST_MASK;
2505 rq->__sector = bio->bi_iter.bi_sector;
2506 blk_rq_bio_prep(rq, bio, nr_segs);
2508 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2509 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2512 blk_account_io_start(rq);
2515 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2516 struct request *rq, bool last)
2518 struct request_queue *q = rq->q;
2519 struct blk_mq_queue_data bd = {
2526 * For OK queue, we are done. For error, caller may kill it.
2527 * Any other error (busy), just add it to our list as we
2528 * previously would have done.
2530 ret = q->mq_ops->queue_rq(hctx, &bd);
2533 blk_mq_update_dispatch_busy(hctx, false);
2535 case BLK_STS_RESOURCE:
2536 case BLK_STS_DEV_RESOURCE:
2537 blk_mq_update_dispatch_busy(hctx, true);
2538 __blk_mq_requeue_request(rq);
2541 blk_mq_update_dispatch_busy(hctx, false);
2548 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2550 bool bypass_insert, bool last)
2552 struct request_queue *q = rq->q;
2553 bool run_queue = true;
2557 * RCU or SRCU read lock is needed before checking quiesced flag.
2559 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2560 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2561 * and avoid driver to try to dispatch again.
2563 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2565 bypass_insert = false;
2569 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2572 budget_token = blk_mq_get_dispatch_budget(q);
2573 if (budget_token < 0)
2576 blk_mq_set_rq_budget_token(rq, budget_token);
2578 if (!blk_mq_get_driver_tag(rq)) {
2579 blk_mq_put_dispatch_budget(q, budget_token);
2583 return __blk_mq_issue_directly(hctx, rq, last);
2586 return BLK_STS_RESOURCE;
2588 blk_mq_sched_insert_request(rq, false, run_queue, false);
2594 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2595 * @hctx: Pointer of the associated hardware queue.
2596 * @rq: Pointer to request to be sent.
2598 * If the device has enough resources to accept a new request now, send the
2599 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2600 * we can try send it another time in the future. Requests inserted at this
2601 * queue have higher priority.
2603 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2607 __blk_mq_try_issue_directly(hctx, rq, false, true);
2609 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2610 blk_mq_request_bypass_insert(rq, false, true);
2611 else if (ret != BLK_STS_OK)
2612 blk_mq_end_request(rq, ret);
2615 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2617 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2620 static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2622 struct blk_mq_hw_ctx *hctx = NULL;
2627 while ((rq = rq_list_pop(&plug->mq_list))) {
2628 bool last = rq_list_empty(plug->mq_list);
2631 if (hctx != rq->mq_hctx) {
2633 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2637 ret = blk_mq_request_issue_directly(rq, last);
2642 case BLK_STS_RESOURCE:
2643 case BLK_STS_DEV_RESOURCE:
2644 blk_mq_request_bypass_insert(rq, false, true);
2645 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2648 blk_mq_end_request(rq, ret);
2655 * If we didn't flush the entire list, we could have told the driver
2656 * there was more coming, but that turned out to be a lie.
2659 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2662 static void __blk_mq_flush_plug_list(struct request_queue *q,
2663 struct blk_plug *plug)
2665 if (blk_queue_quiesced(q))
2667 q->mq_ops->queue_rqs(&plug->mq_list);
2670 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2672 struct blk_mq_hw_ctx *this_hctx = NULL;
2673 struct blk_mq_ctx *this_ctx = NULL;
2674 struct request *requeue_list = NULL;
2675 unsigned int depth = 0;
2679 struct request *rq = rq_list_pop(&plug->mq_list);
2682 this_hctx = rq->mq_hctx;
2683 this_ctx = rq->mq_ctx;
2684 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2685 rq_list_add(&requeue_list, rq);
2688 list_add_tail(&rq->queuelist, &list);
2690 } while (!rq_list_empty(plug->mq_list));
2692 plug->mq_list = requeue_list;
2693 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2694 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched);
2697 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2701 if (rq_list_empty(plug->mq_list))
2705 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2706 struct request_queue *q;
2708 rq = rq_list_peek(&plug->mq_list);
2712 * Peek first request and see if we have a ->queue_rqs() hook.
2713 * If we do, we can dispatch the whole plug list in one go. We
2714 * already know at this point that all requests belong to the
2715 * same queue, caller must ensure that's the case.
2717 * Since we pass off the full list to the driver at this point,
2718 * we do not increment the active request count for the queue.
2719 * Bypass shared tags for now because of that.
2721 if (q->mq_ops->queue_rqs &&
2722 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2723 blk_mq_run_dispatch_ops(q,
2724 __blk_mq_flush_plug_list(q, plug));
2725 if (rq_list_empty(plug->mq_list))
2729 blk_mq_run_dispatch_ops(q,
2730 blk_mq_plug_issue_direct(plug, false));
2731 if (rq_list_empty(plug->mq_list))
2736 blk_mq_dispatch_plug_list(plug, from_schedule);
2737 } while (!rq_list_empty(plug->mq_list));
2740 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2741 struct list_head *list)
2746 while (!list_empty(list)) {
2748 struct request *rq = list_first_entry(list, struct request,
2751 list_del_init(&rq->queuelist);
2752 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2753 if (ret != BLK_STS_OK) {
2755 if (ret == BLK_STS_RESOURCE ||
2756 ret == BLK_STS_DEV_RESOURCE) {
2757 blk_mq_request_bypass_insert(rq, false,
2761 blk_mq_end_request(rq, ret);
2767 * If we didn't flush the entire list, we could have told
2768 * the driver there was more coming, but that turned out to
2771 if ((!list_empty(list) || errors) &&
2772 hctx->queue->mq_ops->commit_rqs && queued)
2773 hctx->queue->mq_ops->commit_rqs(hctx);
2776 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2777 struct bio *bio, unsigned int nr_segs)
2779 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2780 if (blk_attempt_plug_merge(q, bio, nr_segs))
2782 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2788 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2789 struct blk_plug *plug,
2793 struct blk_mq_alloc_data data = {
2796 .cmd_flags = bio->bi_opf,
2800 if (unlikely(bio_queue_enter(bio)))
2803 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2806 rq_qos_throttle(q, bio);
2809 data.nr_tags = plug->nr_ios;
2811 data.cached_rq = &plug->cached_rq;
2814 rq = __blk_mq_alloc_requests(&data);
2817 rq_qos_cleanup(q, bio);
2818 if (bio->bi_opf & REQ_NOWAIT)
2819 bio_wouldblock_error(bio);
2825 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2826 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2832 rq = rq_list_peek(&plug->cached_rq);
2833 if (!rq || rq->q != q)
2836 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2841 if (blk_mq_get_hctx_type((*bio)->bi_opf) != rq->mq_hctx->type)
2843 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2847 * If any qos ->throttle() end up blocking, we will have flushed the
2848 * plug and hence killed the cached_rq list as well. Pop this entry
2849 * before we throttle.
2851 plug->cached_rq = rq_list_next(rq);
2852 rq_qos_throttle(q, *bio);
2854 rq->cmd_flags = (*bio)->bi_opf;
2855 INIT_LIST_HEAD(&rq->queuelist);
2859 static void bio_set_ioprio(struct bio *bio)
2861 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2862 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2863 bio->bi_ioprio = get_current_ioprio();
2864 blkcg_set_ioprio(bio);
2868 * blk_mq_submit_bio - Create and send a request to block device.
2869 * @bio: Bio pointer.
2871 * Builds up a request structure from @q and @bio and send to the device. The
2872 * request may not be queued directly to hardware if:
2873 * * This request can be merged with another one
2874 * * We want to place request at plug queue for possible future merging
2875 * * There is an IO scheduler active at this queue
2877 * It will not queue the request if there is an error with the bio, or at the
2880 void blk_mq_submit_bio(struct bio *bio)
2882 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2883 struct blk_plug *plug = blk_mq_plug(bio);
2884 const int is_sync = op_is_sync(bio->bi_opf);
2886 unsigned int nr_segs = 1;
2889 bio = blk_queue_bounce(bio, q);
2890 if (bio_may_exceed_limits(bio, &q->limits))
2891 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2893 if (!bio_integrity_prep(bio))
2896 bio_set_ioprio(bio);
2898 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2902 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2907 trace_block_getrq(bio);
2909 rq_qos_track(q, rq, bio);
2911 blk_mq_bio_to_request(rq, bio, nr_segs);
2913 ret = blk_crypto_init_request(rq);
2914 if (ret != BLK_STS_OK) {
2915 bio->bi_status = ret;
2917 blk_mq_free_request(rq);
2921 if (op_is_flush(bio->bi_opf)) {
2922 blk_insert_flush(rq);
2927 blk_add_rq_to_plug(plug, rq);
2928 else if ((rq->rq_flags & RQF_ELV) ||
2929 (rq->mq_hctx->dispatch_busy &&
2930 (q->nr_hw_queues == 1 || !is_sync)))
2931 blk_mq_sched_insert_request(rq, false, true, true);
2933 blk_mq_run_dispatch_ops(rq->q,
2934 blk_mq_try_issue_directly(rq->mq_hctx, rq));
2937 #ifdef CONFIG_BLK_MQ_STACKING
2939 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
2940 * @rq: the request being queued
2942 blk_status_t blk_insert_cloned_request(struct request *rq)
2944 struct request_queue *q = rq->q;
2945 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
2948 if (blk_rq_sectors(rq) > max_sectors) {
2950 * SCSI device does not have a good way to return if
2951 * Write Same/Zero is actually supported. If a device rejects
2952 * a non-read/write command (discard, write same,etc.) the
2953 * low-level device driver will set the relevant queue limit to
2954 * 0 to prevent blk-lib from issuing more of the offending
2955 * operations. Commands queued prior to the queue limit being
2956 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
2957 * errors being propagated to upper layers.
2959 if (max_sectors == 0)
2960 return BLK_STS_NOTSUPP;
2962 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
2963 __func__, blk_rq_sectors(rq), max_sectors);
2964 return BLK_STS_IOERR;
2968 * The queue settings related to segment counting may differ from the
2971 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
2972 if (rq->nr_phys_segments > queue_max_segments(q)) {
2973 printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n",
2974 __func__, rq->nr_phys_segments, queue_max_segments(q));
2975 return BLK_STS_IOERR;
2978 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
2979 return BLK_STS_IOERR;
2981 if (blk_crypto_insert_cloned_request(rq))
2982 return BLK_STS_IOERR;
2984 blk_account_io_start(rq);
2987 * Since we have a scheduler attached on the top device,
2988 * bypass a potential scheduler on the bottom device for
2991 blk_mq_run_dispatch_ops(q,
2992 ret = blk_mq_request_issue_directly(rq, true));
2994 blk_account_io_done(rq, ktime_get_ns());
2997 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3000 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3001 * @rq: the clone request to be cleaned up
3004 * Free all bios in @rq for a cloned request.
3006 void blk_rq_unprep_clone(struct request *rq)
3010 while ((bio = rq->bio) != NULL) {
3011 rq->bio = bio->bi_next;
3016 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3019 * blk_rq_prep_clone - Helper function to setup clone request
3020 * @rq: the request to be setup
3021 * @rq_src: original request to be cloned
3022 * @bs: bio_set that bios for clone are allocated from
3023 * @gfp_mask: memory allocation mask for bio
3024 * @bio_ctr: setup function to be called for each clone bio.
3025 * Returns %0 for success, non %0 for failure.
3026 * @data: private data to be passed to @bio_ctr
3029 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3030 * Also, pages which the original bios are pointing to are not copied
3031 * and the cloned bios just point same pages.
3032 * So cloned bios must be completed before original bios, which means
3033 * the caller must complete @rq before @rq_src.
3035 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3036 struct bio_set *bs, gfp_t gfp_mask,
3037 int (*bio_ctr)(struct bio *, struct bio *, void *),
3040 struct bio *bio, *bio_src;
3045 __rq_for_each_bio(bio_src, rq_src) {
3046 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3051 if (bio_ctr && bio_ctr(bio, bio_src, data))
3055 rq->biotail->bi_next = bio;
3058 rq->bio = rq->biotail = bio;
3063 /* Copy attributes of the original request to the clone request. */
3064 rq->__sector = blk_rq_pos(rq_src);
3065 rq->__data_len = blk_rq_bytes(rq_src);
3066 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3067 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3068 rq->special_vec = rq_src->special_vec;
3070 rq->nr_phys_segments = rq_src->nr_phys_segments;
3071 rq->ioprio = rq_src->ioprio;
3073 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3081 blk_rq_unprep_clone(rq);
3085 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3086 #endif /* CONFIG_BLK_MQ_STACKING */
3089 * Steal bios from a request and add them to a bio list.
3090 * The request must not have been partially completed before.
3092 void blk_steal_bios(struct bio_list *list, struct request *rq)
3096 list->tail->bi_next = rq->bio;
3098 list->head = rq->bio;
3099 list->tail = rq->biotail;
3107 EXPORT_SYMBOL_GPL(blk_steal_bios);
3109 static size_t order_to_size(unsigned int order)
3111 return (size_t)PAGE_SIZE << order;
3114 /* called before freeing request pool in @tags */
3115 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3116 struct blk_mq_tags *tags)
3119 unsigned long flags;
3122 * There is no need to clear mapping if driver tags is not initialized
3123 * or the mapping belongs to the driver tags.
3125 if (!drv_tags || drv_tags == tags)
3128 list_for_each_entry(page, &tags->page_list, lru) {
3129 unsigned long start = (unsigned long)page_address(page);
3130 unsigned long end = start + order_to_size(page->private);
3133 for (i = 0; i < drv_tags->nr_tags; i++) {
3134 struct request *rq = drv_tags->rqs[i];
3135 unsigned long rq_addr = (unsigned long)rq;
3137 if (rq_addr >= start && rq_addr < end) {
3138 WARN_ON_ONCE(req_ref_read(rq) != 0);
3139 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3145 * Wait until all pending iteration is done.
3147 * Request reference is cleared and it is guaranteed to be observed
3148 * after the ->lock is released.
3150 spin_lock_irqsave(&drv_tags->lock, flags);
3151 spin_unlock_irqrestore(&drv_tags->lock, flags);
3154 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3155 unsigned int hctx_idx)
3157 struct blk_mq_tags *drv_tags;
3160 if (list_empty(&tags->page_list))
3163 if (blk_mq_is_shared_tags(set->flags))
3164 drv_tags = set->shared_tags;
3166 drv_tags = set->tags[hctx_idx];
3168 if (tags->static_rqs && set->ops->exit_request) {
3171 for (i = 0; i < tags->nr_tags; i++) {
3172 struct request *rq = tags->static_rqs[i];
3176 set->ops->exit_request(set, rq, hctx_idx);
3177 tags->static_rqs[i] = NULL;
3181 blk_mq_clear_rq_mapping(drv_tags, tags);
3183 while (!list_empty(&tags->page_list)) {
3184 page = list_first_entry(&tags->page_list, struct page, lru);
3185 list_del_init(&page->lru);
3187 * Remove kmemleak object previously allocated in
3188 * blk_mq_alloc_rqs().
3190 kmemleak_free(page_address(page));
3191 __free_pages(page, page->private);
3195 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3199 kfree(tags->static_rqs);
3200 tags->static_rqs = NULL;
3202 blk_mq_free_tags(tags);
3205 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3206 unsigned int hctx_idx)
3210 for (i = 0; i < set->nr_maps; i++) {
3211 unsigned int start = set->map[i].queue_offset;
3212 unsigned int end = start + set->map[i].nr_queues;
3214 if (hctx_idx >= start && hctx_idx < end)
3218 if (i >= set->nr_maps)
3219 i = HCTX_TYPE_DEFAULT;
3224 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3225 unsigned int hctx_idx)
3227 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3229 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3232 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3233 unsigned int hctx_idx,
3234 unsigned int nr_tags,
3235 unsigned int reserved_tags)
3237 int node = blk_mq_get_hctx_node(set, hctx_idx);
3238 struct blk_mq_tags *tags;
3240 if (node == NUMA_NO_NODE)
3241 node = set->numa_node;
3243 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3244 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3248 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3249 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3252 blk_mq_free_tags(tags);
3256 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3257 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3259 if (!tags->static_rqs) {
3261 blk_mq_free_tags(tags);
3268 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3269 unsigned int hctx_idx, int node)
3273 if (set->ops->init_request) {
3274 ret = set->ops->init_request(set, rq, hctx_idx, node);
3279 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3283 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3284 struct blk_mq_tags *tags,
3285 unsigned int hctx_idx, unsigned int depth)
3287 unsigned int i, j, entries_per_page, max_order = 4;
3288 int node = blk_mq_get_hctx_node(set, hctx_idx);
3289 size_t rq_size, left;
3291 if (node == NUMA_NO_NODE)
3292 node = set->numa_node;
3294 INIT_LIST_HEAD(&tags->page_list);
3297 * rq_size is the size of the request plus driver payload, rounded
3298 * to the cacheline size
3300 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3302 left = rq_size * depth;
3304 for (i = 0; i < depth; ) {
3305 int this_order = max_order;
3310 while (this_order && left < order_to_size(this_order - 1))
3314 page = alloc_pages_node(node,
3315 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3321 if (order_to_size(this_order) < rq_size)
3328 page->private = this_order;
3329 list_add_tail(&page->lru, &tags->page_list);
3331 p = page_address(page);
3333 * Allow kmemleak to scan these pages as they contain pointers
3334 * to additional allocations like via ops->init_request().
3336 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3337 entries_per_page = order_to_size(this_order) / rq_size;
3338 to_do = min(entries_per_page, depth - i);
3339 left -= to_do * rq_size;
3340 for (j = 0; j < to_do; j++) {
3341 struct request *rq = p;
3343 tags->static_rqs[i] = rq;
3344 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3345 tags->static_rqs[i] = NULL;
3356 blk_mq_free_rqs(set, tags, hctx_idx);
3360 struct rq_iter_data {
3361 struct blk_mq_hw_ctx *hctx;
3365 static bool blk_mq_has_request(struct request *rq, void *data)
3367 struct rq_iter_data *iter_data = data;
3369 if (rq->mq_hctx != iter_data->hctx)
3371 iter_data->has_rq = true;
3375 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3377 struct blk_mq_tags *tags = hctx->sched_tags ?
3378 hctx->sched_tags : hctx->tags;
3379 struct rq_iter_data data = {
3383 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3387 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3388 struct blk_mq_hw_ctx *hctx)
3390 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3392 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3397 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3399 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3400 struct blk_mq_hw_ctx, cpuhp_online);
3402 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3403 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3407 * Prevent new request from being allocated on the current hctx.
3409 * The smp_mb__after_atomic() Pairs with the implied barrier in
3410 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3411 * seen once we return from the tag allocator.
3413 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3414 smp_mb__after_atomic();
3417 * Try to grab a reference to the queue and wait for any outstanding
3418 * requests. If we could not grab a reference the queue has been
3419 * frozen and there are no requests.
3421 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3422 while (blk_mq_hctx_has_requests(hctx))
3424 percpu_ref_put(&hctx->queue->q_usage_counter);
3430 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3432 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3433 struct blk_mq_hw_ctx, cpuhp_online);
3435 if (cpumask_test_cpu(cpu, hctx->cpumask))
3436 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3441 * 'cpu' is going away. splice any existing rq_list entries from this
3442 * software queue to the hw queue dispatch list, and ensure that it
3445 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3447 struct blk_mq_hw_ctx *hctx;
3448 struct blk_mq_ctx *ctx;
3450 enum hctx_type type;
3452 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3453 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3456 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3459 spin_lock(&ctx->lock);
3460 if (!list_empty(&ctx->rq_lists[type])) {
3461 list_splice_init(&ctx->rq_lists[type], &tmp);
3462 blk_mq_hctx_clear_pending(hctx, ctx);
3464 spin_unlock(&ctx->lock);
3466 if (list_empty(&tmp))
3469 spin_lock(&hctx->lock);
3470 list_splice_tail_init(&tmp, &hctx->dispatch);
3471 spin_unlock(&hctx->lock);
3473 blk_mq_run_hw_queue(hctx, true);
3477 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3479 if (!(hctx->flags & BLK_MQ_F_STACKING))
3480 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3481 &hctx->cpuhp_online);
3482 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3487 * Before freeing hw queue, clearing the flush request reference in
3488 * tags->rqs[] for avoiding potential UAF.
3490 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3491 unsigned int queue_depth, struct request *flush_rq)
3494 unsigned long flags;
3496 /* The hw queue may not be mapped yet */
3500 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3502 for (i = 0; i < queue_depth; i++)
3503 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3506 * Wait until all pending iteration is done.
3508 * Request reference is cleared and it is guaranteed to be observed
3509 * after the ->lock is released.
3511 spin_lock_irqsave(&tags->lock, flags);
3512 spin_unlock_irqrestore(&tags->lock, flags);
3515 /* hctx->ctxs will be freed in queue's release handler */
3516 static void blk_mq_exit_hctx(struct request_queue *q,
3517 struct blk_mq_tag_set *set,
3518 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3520 struct request *flush_rq = hctx->fq->flush_rq;
3522 if (blk_mq_hw_queue_mapped(hctx))
3523 blk_mq_tag_idle(hctx);
3525 if (blk_queue_init_done(q))
3526 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3527 set->queue_depth, flush_rq);
3528 if (set->ops->exit_request)
3529 set->ops->exit_request(set, flush_rq, hctx_idx);
3531 if (set->ops->exit_hctx)
3532 set->ops->exit_hctx(hctx, hctx_idx);
3534 blk_mq_remove_cpuhp(hctx);
3536 xa_erase(&q->hctx_table, hctx_idx);
3538 spin_lock(&q->unused_hctx_lock);
3539 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3540 spin_unlock(&q->unused_hctx_lock);
3543 static void blk_mq_exit_hw_queues(struct request_queue *q,
3544 struct blk_mq_tag_set *set, int nr_queue)
3546 struct blk_mq_hw_ctx *hctx;
3549 queue_for_each_hw_ctx(q, hctx, i) {
3552 blk_mq_exit_hctx(q, set, hctx, i);
3556 static int blk_mq_init_hctx(struct request_queue *q,
3557 struct blk_mq_tag_set *set,
3558 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3560 hctx->queue_num = hctx_idx;
3562 if (!(hctx->flags & BLK_MQ_F_STACKING))
3563 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3564 &hctx->cpuhp_online);
3565 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3567 hctx->tags = set->tags[hctx_idx];
3569 if (set->ops->init_hctx &&
3570 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3571 goto unregister_cpu_notifier;
3573 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3577 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3583 if (set->ops->exit_request)
3584 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3586 if (set->ops->exit_hctx)
3587 set->ops->exit_hctx(hctx, hctx_idx);
3588 unregister_cpu_notifier:
3589 blk_mq_remove_cpuhp(hctx);
3593 static struct blk_mq_hw_ctx *
3594 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3597 struct blk_mq_hw_ctx *hctx;
3598 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3600 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3602 goto fail_alloc_hctx;
3604 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3607 atomic_set(&hctx->nr_active, 0);
3608 if (node == NUMA_NO_NODE)
3609 node = set->numa_node;
3610 hctx->numa_node = node;
3612 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3613 spin_lock_init(&hctx->lock);
3614 INIT_LIST_HEAD(&hctx->dispatch);
3616 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3618 INIT_LIST_HEAD(&hctx->hctx_list);
3621 * Allocate space for all possible cpus to avoid allocation at
3624 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3629 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3630 gfp, node, false, false))
3634 spin_lock_init(&hctx->dispatch_wait_lock);
3635 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3636 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3638 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3642 blk_mq_hctx_kobj_init(hctx);
3647 sbitmap_free(&hctx->ctx_map);
3651 free_cpumask_var(hctx->cpumask);
3658 static void blk_mq_init_cpu_queues(struct request_queue *q,
3659 unsigned int nr_hw_queues)
3661 struct blk_mq_tag_set *set = q->tag_set;
3664 for_each_possible_cpu(i) {
3665 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3666 struct blk_mq_hw_ctx *hctx;
3670 spin_lock_init(&__ctx->lock);
3671 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3672 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3677 * Set local node, IFF we have more than one hw queue. If
3678 * not, we remain on the home node of the device
3680 for (j = 0; j < set->nr_maps; j++) {
3681 hctx = blk_mq_map_queue_type(q, j, i);
3682 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3683 hctx->numa_node = cpu_to_node(i);
3688 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3689 unsigned int hctx_idx,
3692 struct blk_mq_tags *tags;
3695 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3699 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3701 blk_mq_free_rq_map(tags);
3708 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3711 if (blk_mq_is_shared_tags(set->flags)) {
3712 set->tags[hctx_idx] = set->shared_tags;
3717 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3720 return set->tags[hctx_idx];
3723 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3724 struct blk_mq_tags *tags,
3725 unsigned int hctx_idx)
3728 blk_mq_free_rqs(set, tags, hctx_idx);
3729 blk_mq_free_rq_map(tags);
3733 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3734 unsigned int hctx_idx)
3736 if (!blk_mq_is_shared_tags(set->flags))
3737 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3739 set->tags[hctx_idx] = NULL;
3742 static void blk_mq_map_swqueue(struct request_queue *q)
3744 unsigned int j, hctx_idx;
3746 struct blk_mq_hw_ctx *hctx;
3747 struct blk_mq_ctx *ctx;
3748 struct blk_mq_tag_set *set = q->tag_set;
3750 queue_for_each_hw_ctx(q, hctx, i) {
3751 cpumask_clear(hctx->cpumask);
3753 hctx->dispatch_from = NULL;
3757 * Map software to hardware queues.
3759 * If the cpu isn't present, the cpu is mapped to first hctx.
3761 for_each_possible_cpu(i) {
3763 ctx = per_cpu_ptr(q->queue_ctx, i);
3764 for (j = 0; j < set->nr_maps; j++) {
3765 if (!set->map[j].nr_queues) {
3766 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3767 HCTX_TYPE_DEFAULT, i);
3770 hctx_idx = set->map[j].mq_map[i];
3771 /* unmapped hw queue can be remapped after CPU topo changed */
3772 if (!set->tags[hctx_idx] &&
3773 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3775 * If tags initialization fail for some hctx,
3776 * that hctx won't be brought online. In this
3777 * case, remap the current ctx to hctx[0] which
3778 * is guaranteed to always have tags allocated
3780 set->map[j].mq_map[i] = 0;
3783 hctx = blk_mq_map_queue_type(q, j, i);
3784 ctx->hctxs[j] = hctx;
3786 * If the CPU is already set in the mask, then we've
3787 * mapped this one already. This can happen if
3788 * devices share queues across queue maps.
3790 if (cpumask_test_cpu(i, hctx->cpumask))
3793 cpumask_set_cpu(i, hctx->cpumask);
3795 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3796 hctx->ctxs[hctx->nr_ctx++] = ctx;
3799 * If the nr_ctx type overflows, we have exceeded the
3800 * amount of sw queues we can support.
3802 BUG_ON(!hctx->nr_ctx);
3805 for (; j < HCTX_MAX_TYPES; j++)
3806 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3807 HCTX_TYPE_DEFAULT, i);
3810 queue_for_each_hw_ctx(q, hctx, i) {
3812 * If no software queues are mapped to this hardware queue,
3813 * disable it and free the request entries.
3815 if (!hctx->nr_ctx) {
3816 /* Never unmap queue 0. We need it as a
3817 * fallback in case of a new remap fails
3821 __blk_mq_free_map_and_rqs(set, i);
3827 hctx->tags = set->tags[i];
3828 WARN_ON(!hctx->tags);
3831 * Set the map size to the number of mapped software queues.
3832 * This is more accurate and more efficient than looping
3833 * over all possibly mapped software queues.
3835 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3838 * Initialize batch roundrobin counts
3840 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3841 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3846 * Caller needs to ensure that we're either frozen/quiesced, or that
3847 * the queue isn't live yet.
3849 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3851 struct blk_mq_hw_ctx *hctx;
3854 queue_for_each_hw_ctx(q, hctx, i) {
3856 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3858 blk_mq_tag_idle(hctx);
3859 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3864 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3867 struct request_queue *q;
3869 lockdep_assert_held(&set->tag_list_lock);
3871 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3872 blk_mq_freeze_queue(q);
3873 queue_set_hctx_shared(q, shared);
3874 blk_mq_unfreeze_queue(q);
3878 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3880 struct blk_mq_tag_set *set = q->tag_set;
3882 mutex_lock(&set->tag_list_lock);
3883 list_del(&q->tag_set_list);
3884 if (list_is_singular(&set->tag_list)) {
3885 /* just transitioned to unshared */
3886 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3887 /* update existing queue */
3888 blk_mq_update_tag_set_shared(set, false);
3890 mutex_unlock(&set->tag_list_lock);
3891 INIT_LIST_HEAD(&q->tag_set_list);
3894 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3895 struct request_queue *q)
3897 mutex_lock(&set->tag_list_lock);
3900 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3902 if (!list_empty(&set->tag_list) &&
3903 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3904 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3905 /* update existing queue */
3906 blk_mq_update_tag_set_shared(set, true);
3908 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3909 queue_set_hctx_shared(q, true);
3910 list_add_tail(&q->tag_set_list, &set->tag_list);
3912 mutex_unlock(&set->tag_list_lock);
3915 /* All allocations will be freed in release handler of q->mq_kobj */
3916 static int blk_mq_alloc_ctxs(struct request_queue *q)
3918 struct blk_mq_ctxs *ctxs;
3921 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3925 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3926 if (!ctxs->queue_ctx)
3929 for_each_possible_cpu(cpu) {
3930 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3934 q->mq_kobj = &ctxs->kobj;
3935 q->queue_ctx = ctxs->queue_ctx;
3944 * It is the actual release handler for mq, but we do it from
3945 * request queue's release handler for avoiding use-after-free
3946 * and headache because q->mq_kobj shouldn't have been introduced,
3947 * but we can't group ctx/kctx kobj without it.
3949 void blk_mq_release(struct request_queue *q)
3951 struct blk_mq_hw_ctx *hctx, *next;
3954 queue_for_each_hw_ctx(q, hctx, i)
3955 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3957 /* all hctx are in .unused_hctx_list now */
3958 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3959 list_del_init(&hctx->hctx_list);
3960 kobject_put(&hctx->kobj);
3963 xa_destroy(&q->hctx_table);
3966 * release .mq_kobj and sw queue's kobject now because
3967 * both share lifetime with request queue.
3969 blk_mq_sysfs_deinit(q);
3972 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3975 struct request_queue *q;
3978 q = blk_alloc_queue(set->numa_node, set->flags & BLK_MQ_F_BLOCKING);
3980 return ERR_PTR(-ENOMEM);
3981 q->queuedata = queuedata;
3982 ret = blk_mq_init_allocated_queue(set, q);
3985 return ERR_PTR(ret);
3990 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3992 return blk_mq_init_queue_data(set, NULL);
3994 EXPORT_SYMBOL(blk_mq_init_queue);
3997 * blk_mq_destroy_queue - shutdown a request queue
3998 * @q: request queue to shutdown
4000 * This shuts down a request queue allocated by blk_mq_init_queue() and drops
4001 * the initial reference. All future requests will failed with -ENODEV.
4003 * Context: can sleep
4005 void blk_mq_destroy_queue(struct request_queue *q)
4007 WARN_ON_ONCE(!queue_is_mq(q));
4008 WARN_ON_ONCE(blk_queue_registered(q));
4012 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4013 blk_queue_start_drain(q);
4014 blk_freeze_queue(q);
4017 blk_mq_cancel_work_sync(q);
4018 blk_mq_exit_queue(q);
4020 /* @q is and will stay empty, shutdown and put */
4023 EXPORT_SYMBOL(blk_mq_destroy_queue);
4025 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4026 struct lock_class_key *lkclass)
4028 struct request_queue *q;
4029 struct gendisk *disk;
4031 q = blk_mq_init_queue_data(set, queuedata);
4035 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4037 blk_mq_destroy_queue(q);
4038 return ERR_PTR(-ENOMEM);
4040 set_bit(GD_OWNS_QUEUE, &disk->state);
4043 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4045 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4046 struct lock_class_key *lkclass)
4048 if (!blk_get_queue(q))
4050 return __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4052 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4054 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4055 struct blk_mq_tag_set *set, struct request_queue *q,
4056 int hctx_idx, int node)
4058 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4060 /* reuse dead hctx first */
4061 spin_lock(&q->unused_hctx_lock);
4062 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4063 if (tmp->numa_node == node) {
4069 list_del_init(&hctx->hctx_list);
4070 spin_unlock(&q->unused_hctx_lock);
4073 hctx = blk_mq_alloc_hctx(q, set, node);
4077 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4083 kobject_put(&hctx->kobj);
4088 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4089 struct request_queue *q)
4091 struct blk_mq_hw_ctx *hctx;
4094 /* protect against switching io scheduler */
4095 mutex_lock(&q->sysfs_lock);
4096 for (i = 0; i < set->nr_hw_queues; i++) {
4098 int node = blk_mq_get_hctx_node(set, i);
4099 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4102 old_node = old_hctx->numa_node;
4103 blk_mq_exit_hctx(q, set, old_hctx, i);
4106 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4109 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4111 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4112 WARN_ON_ONCE(!hctx);
4116 * Increasing nr_hw_queues fails. Free the newly allocated
4117 * hctxs and keep the previous q->nr_hw_queues.
4119 if (i != set->nr_hw_queues) {
4120 j = q->nr_hw_queues;
4123 q->nr_hw_queues = set->nr_hw_queues;
4126 xa_for_each_start(&q->hctx_table, j, hctx, j)
4127 blk_mq_exit_hctx(q, set, hctx, j);
4128 mutex_unlock(&q->sysfs_lock);
4131 static void blk_mq_update_poll_flag(struct request_queue *q)
4133 struct blk_mq_tag_set *set = q->tag_set;
4135 if (set->nr_maps > HCTX_TYPE_POLL &&
4136 set->map[HCTX_TYPE_POLL].nr_queues)
4137 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4139 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4142 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4143 struct request_queue *q)
4145 WARN_ON_ONCE(blk_queue_has_srcu(q) !=
4146 !!(set->flags & BLK_MQ_F_BLOCKING));
4148 /* mark the queue as mq asap */
4149 q->mq_ops = set->ops;
4151 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
4152 blk_mq_poll_stats_bkt,
4153 BLK_MQ_POLL_STATS_BKTS, q);
4157 if (blk_mq_alloc_ctxs(q))
4160 /* init q->mq_kobj and sw queues' kobjects */
4161 blk_mq_sysfs_init(q);
4163 INIT_LIST_HEAD(&q->unused_hctx_list);
4164 spin_lock_init(&q->unused_hctx_lock);
4166 xa_init(&q->hctx_table);
4168 blk_mq_realloc_hw_ctxs(set, q);
4169 if (!q->nr_hw_queues)
4172 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4173 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4177 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4178 blk_mq_update_poll_flag(q);
4180 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4181 INIT_LIST_HEAD(&q->requeue_list);
4182 spin_lock_init(&q->requeue_lock);
4184 q->nr_requests = set->queue_depth;
4187 * Default to classic polling
4189 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
4191 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4192 blk_mq_add_queue_tag_set(set, q);
4193 blk_mq_map_swqueue(q);
4199 blk_stat_free_callback(q->poll_cb);
4205 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4207 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4208 void blk_mq_exit_queue(struct request_queue *q)
4210 struct blk_mq_tag_set *set = q->tag_set;
4212 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4213 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4214 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4215 blk_mq_del_queue_tag_set(q);
4218 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4222 if (blk_mq_is_shared_tags(set->flags)) {
4223 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4226 if (!set->shared_tags)
4230 for (i = 0; i < set->nr_hw_queues; i++) {
4231 if (!__blk_mq_alloc_map_and_rqs(set, i))
4240 __blk_mq_free_map_and_rqs(set, i);
4242 if (blk_mq_is_shared_tags(set->flags)) {
4243 blk_mq_free_map_and_rqs(set, set->shared_tags,
4244 BLK_MQ_NO_HCTX_IDX);
4251 * Allocate the request maps associated with this tag_set. Note that this
4252 * may reduce the depth asked for, if memory is tight. set->queue_depth
4253 * will be updated to reflect the allocated depth.
4255 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4260 depth = set->queue_depth;
4262 err = __blk_mq_alloc_rq_maps(set);
4266 set->queue_depth >>= 1;
4267 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4271 } while (set->queue_depth);
4273 if (!set->queue_depth || err) {
4274 pr_err("blk-mq: failed to allocate request map\n");
4278 if (depth != set->queue_depth)
4279 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4280 depth, set->queue_depth);
4285 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4288 * blk_mq_map_queues() and multiple .map_queues() implementations
4289 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4290 * number of hardware queues.
4292 if (set->nr_maps == 1)
4293 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4295 if (set->ops->map_queues && !is_kdump_kernel()) {
4299 * transport .map_queues is usually done in the following
4302 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4303 * mask = get_cpu_mask(queue)
4304 * for_each_cpu(cpu, mask)
4305 * set->map[x].mq_map[cpu] = queue;
4308 * When we need to remap, the table has to be cleared for
4309 * killing stale mapping since one CPU may not be mapped
4312 for (i = 0; i < set->nr_maps; i++)
4313 blk_mq_clear_mq_map(&set->map[i]);
4315 set->ops->map_queues(set);
4317 BUG_ON(set->nr_maps > 1);
4318 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4322 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4323 int cur_nr_hw_queues, int new_nr_hw_queues)
4325 struct blk_mq_tags **new_tags;
4327 if (cur_nr_hw_queues >= new_nr_hw_queues)
4330 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4331 GFP_KERNEL, set->numa_node);
4336 memcpy(new_tags, set->tags, cur_nr_hw_queues *
4337 sizeof(*set->tags));
4339 set->tags = new_tags;
4340 set->nr_hw_queues = new_nr_hw_queues;
4345 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
4346 int new_nr_hw_queues)
4348 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
4352 * Alloc a tag set to be associated with one or more request queues.
4353 * May fail with EINVAL for various error conditions. May adjust the
4354 * requested depth down, if it's too large. In that case, the set
4355 * value will be stored in set->queue_depth.
4357 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4361 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4363 if (!set->nr_hw_queues)
4365 if (!set->queue_depth)
4367 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4370 if (!set->ops->queue_rq)
4373 if (!set->ops->get_budget ^ !set->ops->put_budget)
4376 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4377 pr_info("blk-mq: reduced tag depth to %u\n",
4379 set->queue_depth = BLK_MQ_MAX_DEPTH;
4384 else if (set->nr_maps > HCTX_MAX_TYPES)
4388 * If a crashdump is active, then we are potentially in a very
4389 * memory constrained environment. Limit us to 1 queue and
4390 * 64 tags to prevent using too much memory.
4392 if (is_kdump_kernel()) {
4393 set->nr_hw_queues = 1;
4395 set->queue_depth = min(64U, set->queue_depth);
4398 * There is no use for more h/w queues than cpus if we just have
4401 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4402 set->nr_hw_queues = nr_cpu_ids;
4404 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
4408 for (i = 0; i < set->nr_maps; i++) {
4409 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4410 sizeof(set->map[i].mq_map[0]),
4411 GFP_KERNEL, set->numa_node);
4412 if (!set->map[i].mq_map)
4413 goto out_free_mq_map;
4414 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4417 blk_mq_update_queue_map(set);
4419 ret = blk_mq_alloc_set_map_and_rqs(set);
4421 goto out_free_mq_map;
4423 mutex_init(&set->tag_list_lock);
4424 INIT_LIST_HEAD(&set->tag_list);
4429 for (i = 0; i < set->nr_maps; i++) {
4430 kfree(set->map[i].mq_map);
4431 set->map[i].mq_map = NULL;
4437 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4439 /* allocate and initialize a tagset for a simple single-queue device */
4440 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4441 const struct blk_mq_ops *ops, unsigned int queue_depth,
4442 unsigned int set_flags)
4444 memset(set, 0, sizeof(*set));
4446 set->nr_hw_queues = 1;
4448 set->queue_depth = queue_depth;
4449 set->numa_node = NUMA_NO_NODE;
4450 set->flags = set_flags;
4451 return blk_mq_alloc_tag_set(set);
4453 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4455 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4459 for (i = 0; i < set->nr_hw_queues; i++)
4460 __blk_mq_free_map_and_rqs(set, i);
4462 if (blk_mq_is_shared_tags(set->flags)) {
4463 blk_mq_free_map_and_rqs(set, set->shared_tags,
4464 BLK_MQ_NO_HCTX_IDX);
4467 for (j = 0; j < set->nr_maps; j++) {
4468 kfree(set->map[j].mq_map);
4469 set->map[j].mq_map = NULL;
4475 EXPORT_SYMBOL(blk_mq_free_tag_set);
4477 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4479 struct blk_mq_tag_set *set = q->tag_set;
4480 struct blk_mq_hw_ctx *hctx;
4487 if (q->nr_requests == nr)
4490 blk_mq_freeze_queue(q);
4491 blk_mq_quiesce_queue(q);
4494 queue_for_each_hw_ctx(q, hctx, i) {
4498 * If we're using an MQ scheduler, just update the scheduler
4499 * queue depth. This is similar to what the old code would do.
4501 if (hctx->sched_tags) {
4502 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4505 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4510 if (q->elevator && q->elevator->type->ops.depth_updated)
4511 q->elevator->type->ops.depth_updated(hctx);
4514 q->nr_requests = nr;
4515 if (blk_mq_is_shared_tags(set->flags)) {
4517 blk_mq_tag_update_sched_shared_tags(q);
4519 blk_mq_tag_resize_shared_tags(set, nr);
4523 blk_mq_unquiesce_queue(q);
4524 blk_mq_unfreeze_queue(q);
4530 * request_queue and elevator_type pair.
4531 * It is just used by __blk_mq_update_nr_hw_queues to cache
4532 * the elevator_type associated with a request_queue.
4534 struct blk_mq_qe_pair {
4535 struct list_head node;
4536 struct request_queue *q;
4537 struct elevator_type *type;
4541 * Cache the elevator_type in qe pair list and switch the
4542 * io scheduler to 'none'
4544 static bool blk_mq_elv_switch_none(struct list_head *head,
4545 struct request_queue *q)
4547 struct blk_mq_qe_pair *qe;
4552 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4556 /* q->elevator needs protection from ->sysfs_lock */
4557 mutex_lock(&q->sysfs_lock);
4559 INIT_LIST_HEAD(&qe->node);
4561 qe->type = q->elevator->type;
4562 list_add(&qe->node, head);
4565 * After elevator_switch, the previous elevator_queue will be
4566 * released by elevator_release. The reference of the io scheduler
4567 * module get by elevator_get will also be put. So we need to get
4568 * a reference of the io scheduler module here to prevent it to be
4571 __module_get(qe->type->elevator_owner);
4572 elevator_switch(q, NULL);
4573 mutex_unlock(&q->sysfs_lock);
4578 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4579 struct request_queue *q)
4581 struct blk_mq_qe_pair *qe;
4583 list_for_each_entry(qe, head, node)
4590 static void blk_mq_elv_switch_back(struct list_head *head,
4591 struct request_queue *q)
4593 struct blk_mq_qe_pair *qe;
4594 struct elevator_type *t;
4596 qe = blk_lookup_qe_pair(head, q);
4600 list_del(&qe->node);
4603 mutex_lock(&q->sysfs_lock);
4604 elevator_switch(q, t);
4605 mutex_unlock(&q->sysfs_lock);
4608 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4611 struct request_queue *q;
4613 int prev_nr_hw_queues;
4615 lockdep_assert_held(&set->tag_list_lock);
4617 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4618 nr_hw_queues = nr_cpu_ids;
4619 if (nr_hw_queues < 1)
4621 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4624 list_for_each_entry(q, &set->tag_list, tag_set_list)
4625 blk_mq_freeze_queue(q);
4627 * Switch IO scheduler to 'none', cleaning up the data associated
4628 * with the previous scheduler. We will switch back once we are done
4629 * updating the new sw to hw queue mappings.
4631 list_for_each_entry(q, &set->tag_list, tag_set_list)
4632 if (!blk_mq_elv_switch_none(&head, q))
4635 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4636 blk_mq_debugfs_unregister_hctxs(q);
4637 blk_mq_sysfs_unregister_hctxs(q);
4640 prev_nr_hw_queues = set->nr_hw_queues;
4641 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
4645 set->nr_hw_queues = nr_hw_queues;
4647 blk_mq_update_queue_map(set);
4648 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4649 blk_mq_realloc_hw_ctxs(set, q);
4650 blk_mq_update_poll_flag(q);
4651 if (q->nr_hw_queues != set->nr_hw_queues) {
4652 int i = prev_nr_hw_queues;
4654 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4655 nr_hw_queues, prev_nr_hw_queues);
4656 for (; i < set->nr_hw_queues; i++)
4657 __blk_mq_free_map_and_rqs(set, i);
4659 set->nr_hw_queues = prev_nr_hw_queues;
4660 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4663 blk_mq_map_swqueue(q);
4667 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4668 blk_mq_sysfs_register_hctxs(q);
4669 blk_mq_debugfs_register_hctxs(q);
4673 list_for_each_entry(q, &set->tag_list, tag_set_list)
4674 blk_mq_elv_switch_back(&head, q);
4676 list_for_each_entry(q, &set->tag_list, tag_set_list)
4677 blk_mq_unfreeze_queue(q);
4680 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4682 mutex_lock(&set->tag_list_lock);
4683 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4684 mutex_unlock(&set->tag_list_lock);
4686 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4688 /* Enable polling stats and return whether they were already enabled. */
4689 static bool blk_poll_stats_enable(struct request_queue *q)
4694 return blk_stats_alloc_enable(q);
4697 static void blk_mq_poll_stats_start(struct request_queue *q)
4700 * We don't arm the callback if polling stats are not enabled or the
4701 * callback is already active.
4703 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
4706 blk_stat_activate_msecs(q->poll_cb, 100);
4709 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4711 struct request_queue *q = cb->data;
4714 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4715 if (cb->stat[bucket].nr_samples)
4716 q->poll_stat[bucket] = cb->stat[bucket];
4720 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4723 unsigned long ret = 0;
4727 * If stats collection isn't on, don't sleep but turn it on for
4730 if (!blk_poll_stats_enable(q))
4734 * As an optimistic guess, use half of the mean service time
4735 * for this type of request. We can (and should) make this smarter.
4736 * For instance, if the completion latencies are tight, we can
4737 * get closer than just half the mean. This is especially
4738 * important on devices where the completion latencies are longer
4739 * than ~10 usec. We do use the stats for the relevant IO size
4740 * if available which does lead to better estimates.
4742 bucket = blk_mq_poll_stats_bkt(rq);
4746 if (q->poll_stat[bucket].nr_samples)
4747 ret = (q->poll_stat[bucket].mean + 1) / 2;
4752 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4754 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4755 struct request *rq = blk_qc_to_rq(hctx, qc);
4756 struct hrtimer_sleeper hs;
4757 enum hrtimer_mode mode;
4762 * If a request has completed on queue that uses an I/O scheduler, we
4763 * won't get back a request from blk_qc_to_rq.
4765 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4769 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4771 * 0: use half of prev avg
4772 * >0: use this specific value
4774 if (q->poll_nsec > 0)
4775 nsecs = q->poll_nsec;
4777 nsecs = blk_mq_poll_nsecs(q, rq);
4782 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4785 * This will be replaced with the stats tracking code, using
4786 * 'avg_completion_time / 2' as the pre-sleep target.
4790 mode = HRTIMER_MODE_REL;
4791 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4792 hrtimer_set_expires(&hs.timer, kt);
4795 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4797 set_current_state(TASK_UNINTERRUPTIBLE);
4798 hrtimer_sleeper_start_expires(&hs, mode);
4801 hrtimer_cancel(&hs.timer);
4802 mode = HRTIMER_MODE_ABS;
4803 } while (hs.task && !signal_pending(current));
4805 __set_current_state(TASK_RUNNING);
4806 destroy_hrtimer_on_stack(&hs.timer);
4809 * If we sleep, have the caller restart the poll loop to reset the
4810 * state. Like for the other success return cases, the caller is
4811 * responsible for checking if the IO completed. If the IO isn't
4812 * complete, we'll get called again and will go straight to the busy
4818 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4819 struct io_comp_batch *iob, unsigned int flags)
4821 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4822 long state = get_current_state();
4826 ret = q->mq_ops->poll(hctx, iob);
4828 __set_current_state(TASK_RUNNING);
4832 if (signal_pending_state(state, current))
4833 __set_current_state(TASK_RUNNING);
4834 if (task_is_running(current))
4837 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4840 } while (!need_resched());
4842 __set_current_state(TASK_RUNNING);
4846 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4849 if (!(flags & BLK_POLL_NOSLEEP) &&
4850 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4851 if (blk_mq_poll_hybrid(q, cookie))
4854 return blk_mq_poll_classic(q, cookie, iob, flags);
4857 unsigned int blk_mq_rq_cpu(struct request *rq)
4859 return rq->mq_ctx->cpu;
4861 EXPORT_SYMBOL(blk_mq_rq_cpu);
4863 void blk_mq_cancel_work_sync(struct request_queue *q)
4865 if (queue_is_mq(q)) {
4866 struct blk_mq_hw_ctx *hctx;
4869 cancel_delayed_work_sync(&q->requeue_work);
4871 queue_for_each_hw_ctx(q, hctx, i)
4872 cancel_delayed_work_sync(&hctx->run_work);
4876 static int __init blk_mq_init(void)
4880 for_each_possible_cpu(i)
4881 init_llist_head(&per_cpu(blk_cpu_done, i));
4882 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4884 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4885 "block/softirq:dead", NULL,
4886 blk_softirq_cpu_dead);
4887 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4888 blk_mq_hctx_notify_dead);
4889 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4890 blk_mq_hctx_notify_online,
4891 blk_mq_hctx_notify_offline);
4894 subsys_initcall(blk_mq_init);