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)) ||
630 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
631 return ERR_PTR(-EINVAL);
633 if (hctx_idx >= q->nr_hw_queues)
634 return ERR_PTR(-EIO);
636 ret = blk_queue_enter(q, flags);
641 * Check if the hardware context is actually mapped to anything.
642 * If not tell the caller that it should skip this queue.
645 data.hctx = xa_load(&q->hctx_table, hctx_idx);
646 if (!blk_mq_hw_queue_mapped(data.hctx))
648 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
649 if (cpu >= nr_cpu_ids)
651 data.ctx = __blk_mq_get_ctx(q, cpu);
654 blk_mq_tag_busy(data.hctx);
656 data.rq_flags |= RQF_ELV;
658 if (flags & BLK_MQ_REQ_RESERVED)
659 data.rq_flags |= RQF_RESV;
662 tag = blk_mq_get_tag(&data);
663 if (tag == BLK_MQ_NO_TAG)
665 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag,
668 rq->__sector = (sector_t) -1;
669 rq->bio = rq->biotail = NULL;
676 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
678 static void blk_mq_finish_request(struct request *rq)
680 struct request_queue *q = rq->q;
682 if ((rq->rq_flags & RQF_ELVPRIV) &&
683 q->elevator->type->ops.finish_request) {
684 q->elevator->type->ops.finish_request(rq);
686 * For postflush request that may need to be
687 * completed twice, we should clear this flag
688 * to avoid double finish_request() on the rq.
690 rq->rq_flags &= ~RQF_ELVPRIV;
694 static void __blk_mq_free_request(struct request *rq)
696 struct request_queue *q = rq->q;
697 struct blk_mq_ctx *ctx = rq->mq_ctx;
698 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
699 const int sched_tag = rq->internal_tag;
701 blk_crypto_free_request(rq);
702 blk_pm_mark_last_busy(rq);
705 if (rq->rq_flags & RQF_MQ_INFLIGHT)
706 __blk_mq_dec_active_requests(hctx);
708 if (rq->tag != BLK_MQ_NO_TAG)
709 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
710 if (sched_tag != BLK_MQ_NO_TAG)
711 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
712 blk_mq_sched_restart(hctx);
716 void blk_mq_free_request(struct request *rq)
718 struct request_queue *q = rq->q;
720 blk_mq_finish_request(rq);
722 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
723 laptop_io_completion(q->disk->bdi);
727 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
728 if (req_ref_put_and_test(rq))
729 __blk_mq_free_request(rq);
731 EXPORT_SYMBOL_GPL(blk_mq_free_request);
733 void blk_mq_free_plug_rqs(struct blk_plug *plug)
737 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
738 blk_mq_free_request(rq);
741 void blk_dump_rq_flags(struct request *rq, char *msg)
743 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
744 rq->q->disk ? rq->q->disk->disk_name : "?",
745 (__force unsigned long long) rq->cmd_flags);
747 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
748 (unsigned long long)blk_rq_pos(rq),
749 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
750 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
751 rq->bio, rq->biotail, blk_rq_bytes(rq));
753 EXPORT_SYMBOL(blk_dump_rq_flags);
755 static void req_bio_endio(struct request *rq, struct bio *bio,
756 unsigned int nbytes, blk_status_t error)
758 if (unlikely(error)) {
759 bio->bi_status = error;
760 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
762 * Partial zone append completions cannot be supported as the
763 * BIO fragments may end up not being written sequentially.
765 if (bio->bi_iter.bi_size != nbytes)
766 bio->bi_status = BLK_STS_IOERR;
768 bio->bi_iter.bi_sector = rq->__sector;
771 bio_advance(bio, nbytes);
773 if (unlikely(rq->rq_flags & RQF_QUIET))
774 bio_set_flag(bio, BIO_QUIET);
775 /* don't actually finish bio if it's part of flush sequence */
776 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
780 static void blk_account_io_completion(struct request *req, unsigned int bytes)
782 if (req->part && blk_do_io_stat(req)) {
783 const int sgrp = op_stat_group(req_op(req));
786 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
791 static void blk_print_req_error(struct request *req, blk_status_t status)
793 printk_ratelimited(KERN_ERR
794 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
795 "phys_seg %u prio class %u\n",
796 blk_status_to_str(status),
797 req->q->disk ? req->q->disk->disk_name : "?",
798 blk_rq_pos(req), (__force u32)req_op(req),
799 blk_op_str(req_op(req)),
800 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
801 req->nr_phys_segments,
802 IOPRIO_PRIO_CLASS(req->ioprio));
806 * Fully end IO on a request. Does not support partial completions, or
809 static void blk_complete_request(struct request *req)
811 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
812 int total_bytes = blk_rq_bytes(req);
813 struct bio *bio = req->bio;
815 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
820 #ifdef CONFIG_BLK_DEV_INTEGRITY
821 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
822 req->q->integrity.profile->complete_fn(req, total_bytes);
826 * Upper layers may call blk_crypto_evict_key() anytime after the last
827 * bio_endio(). Therefore, the keyslot must be released before that.
829 blk_crypto_rq_put_keyslot(req);
831 blk_account_io_completion(req, total_bytes);
834 struct bio *next = bio->bi_next;
836 /* Completion has already been traced */
837 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
839 if (req_op(req) == REQ_OP_ZONE_APPEND)
840 bio->bi_iter.bi_sector = req->__sector;
848 * Reset counters so that the request stacking driver
849 * can find how many bytes remain in the request
859 * blk_update_request - Complete multiple bytes without completing the request
860 * @req: the request being processed
861 * @error: block status code
862 * @nr_bytes: number of bytes to complete for @req
865 * Ends I/O on a number of bytes attached to @req, but doesn't complete
866 * the request structure even if @req doesn't have leftover.
867 * If @req has leftover, sets it up for the next range of segments.
869 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
870 * %false return from this function.
873 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
874 * except in the consistency check at the end of this function.
877 * %false - this request doesn't have any more data
878 * %true - this request has more data
880 bool blk_update_request(struct request *req, blk_status_t error,
881 unsigned int nr_bytes)
885 trace_block_rq_complete(req, error, nr_bytes);
890 #ifdef CONFIG_BLK_DEV_INTEGRITY
891 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
893 req->q->integrity.profile->complete_fn(req, nr_bytes);
897 * Upper layers may call blk_crypto_evict_key() anytime after the last
898 * bio_endio(). Therefore, the keyslot must be released before that.
900 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
901 __blk_crypto_rq_put_keyslot(req);
903 if (unlikely(error && !blk_rq_is_passthrough(req) &&
904 !(req->rq_flags & RQF_QUIET)) &&
905 !test_bit(GD_DEAD, &req->q->disk->state)) {
906 blk_print_req_error(req, error);
907 trace_block_rq_error(req, error, nr_bytes);
910 blk_account_io_completion(req, nr_bytes);
914 struct bio *bio = req->bio;
915 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
917 if (bio_bytes == bio->bi_iter.bi_size)
918 req->bio = bio->bi_next;
920 /* Completion has already been traced */
921 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
922 req_bio_endio(req, bio, bio_bytes, error);
924 total_bytes += bio_bytes;
925 nr_bytes -= bio_bytes;
936 * Reset counters so that the request stacking driver
937 * can find how many bytes remain in the request
944 req->__data_len -= total_bytes;
946 /* update sector only for requests with clear definition of sector */
947 if (!blk_rq_is_passthrough(req))
948 req->__sector += total_bytes >> 9;
950 /* mixed attributes always follow the first bio */
951 if (req->rq_flags & RQF_MIXED_MERGE) {
952 req->cmd_flags &= ~REQ_FAILFAST_MASK;
953 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
956 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
958 * If total number of sectors is less than the first segment
959 * size, something has gone terribly wrong.
961 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
962 blk_dump_rq_flags(req, "request botched");
963 req->__data_len = blk_rq_cur_bytes(req);
966 /* recalculate the number of segments */
967 req->nr_phys_segments = blk_recalc_rq_segments(req);
972 EXPORT_SYMBOL_GPL(blk_update_request);
974 static void __blk_account_io_done(struct request *req, u64 now)
976 const int sgrp = op_stat_group(req_op(req));
979 update_io_ticks(req->part, jiffies, true);
980 part_stat_inc(req->part, ios[sgrp]);
981 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
985 static inline void blk_account_io_done(struct request *req, u64 now)
988 * Account IO completion. flush_rq isn't accounted as a
989 * normal IO on queueing nor completion. Accounting the
990 * containing request is enough.
992 if (blk_do_io_stat(req) && req->part &&
993 !(req->rq_flags & RQF_FLUSH_SEQ))
994 __blk_account_io_done(req, now);
997 static void __blk_account_io_start(struct request *rq)
1000 * All non-passthrough requests are created from a bio with one
1001 * exception: when a flush command that is part of a flush sequence
1002 * generated by the state machine in blk-flush.c is cloned onto the
1003 * lower device by dm-multipath we can get here without a bio.
1006 rq->part = rq->bio->bi_bdev;
1008 rq->part = rq->q->disk->part0;
1011 update_io_ticks(rq->part, jiffies, false);
1015 static inline void blk_account_io_start(struct request *req)
1017 if (blk_do_io_stat(req))
1018 __blk_account_io_start(req);
1021 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1023 if (rq->rq_flags & RQF_STATS) {
1024 blk_mq_poll_stats_start(rq->q);
1025 blk_stat_add(rq, now);
1028 blk_mq_sched_completed_request(rq, now);
1029 blk_account_io_done(rq, now);
1032 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1034 if (blk_mq_need_time_stamp(rq))
1035 __blk_mq_end_request_acct(rq, ktime_get_ns());
1037 blk_mq_finish_request(rq);
1040 rq_qos_done(rq->q, rq);
1041 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1042 blk_mq_free_request(rq);
1044 blk_mq_free_request(rq);
1047 EXPORT_SYMBOL(__blk_mq_end_request);
1049 void blk_mq_end_request(struct request *rq, blk_status_t error)
1051 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1053 __blk_mq_end_request(rq, error);
1055 EXPORT_SYMBOL(blk_mq_end_request);
1057 #define TAG_COMP_BATCH 32
1059 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1060 int *tag_array, int nr_tags)
1062 struct request_queue *q = hctx->queue;
1065 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1066 * update hctx->nr_active in batch
1068 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1069 __blk_mq_sub_active_requests(hctx, nr_tags);
1071 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1072 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1075 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1077 int tags[TAG_COMP_BATCH], nr_tags = 0;
1078 struct blk_mq_hw_ctx *cur_hctx = NULL;
1083 now = ktime_get_ns();
1085 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1087 prefetch(rq->rq_next);
1089 blk_complete_request(rq);
1091 __blk_mq_end_request_acct(rq, now);
1093 blk_mq_finish_request(rq);
1095 rq_qos_done(rq->q, rq);
1098 * If end_io handler returns NONE, then it still has
1099 * ownership of the request.
1101 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1104 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1105 if (!req_ref_put_and_test(rq))
1108 blk_crypto_free_request(rq);
1109 blk_pm_mark_last_busy(rq);
1111 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1113 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1115 cur_hctx = rq->mq_hctx;
1117 tags[nr_tags++] = rq->tag;
1121 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1123 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1125 static void blk_complete_reqs(struct llist_head *list)
1127 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1128 struct request *rq, *next;
1130 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1131 rq->q->mq_ops->complete(rq);
1134 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1136 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1139 static int blk_softirq_cpu_dead(unsigned int cpu)
1141 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1145 static void __blk_mq_complete_request_remote(void *data)
1147 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1150 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1152 int cpu = raw_smp_processor_id();
1154 if (!IS_ENABLED(CONFIG_SMP) ||
1155 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1158 * With force threaded interrupts enabled, raising softirq from an SMP
1159 * function call will always result in waking the ksoftirqd thread.
1160 * This is probably worse than completing the request on a different
1163 if (force_irqthreads())
1166 /* same CPU or cache domain? Complete locally */
1167 if (cpu == rq->mq_ctx->cpu ||
1168 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1169 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1172 /* don't try to IPI to an offline CPU */
1173 return cpu_online(rq->mq_ctx->cpu);
1176 static void blk_mq_complete_send_ipi(struct request *rq)
1178 struct llist_head *list;
1181 cpu = rq->mq_ctx->cpu;
1182 list = &per_cpu(blk_cpu_done, cpu);
1183 if (llist_add(&rq->ipi_list, list)) {
1184 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1185 smp_call_function_single_async(cpu, &rq->csd);
1189 static void blk_mq_raise_softirq(struct request *rq)
1191 struct llist_head *list;
1194 list = this_cpu_ptr(&blk_cpu_done);
1195 if (llist_add(&rq->ipi_list, list))
1196 raise_softirq(BLOCK_SOFTIRQ);
1200 bool blk_mq_complete_request_remote(struct request *rq)
1202 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1205 * For request which hctx has only one ctx mapping,
1206 * or a polled request, always complete locally,
1207 * it's pointless to redirect the completion.
1209 if (rq->mq_hctx->nr_ctx == 1 ||
1210 rq->cmd_flags & REQ_POLLED)
1213 if (blk_mq_complete_need_ipi(rq)) {
1214 blk_mq_complete_send_ipi(rq);
1218 if (rq->q->nr_hw_queues == 1) {
1219 blk_mq_raise_softirq(rq);
1224 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1227 * blk_mq_complete_request - end I/O on a request
1228 * @rq: the request being processed
1231 * Complete a request by scheduling the ->complete_rq operation.
1233 void blk_mq_complete_request(struct request *rq)
1235 if (!blk_mq_complete_request_remote(rq))
1236 rq->q->mq_ops->complete(rq);
1238 EXPORT_SYMBOL(blk_mq_complete_request);
1241 * blk_mq_start_request - Start processing a request
1242 * @rq: Pointer to request to be started
1244 * Function used by device drivers to notify the block layer that a request
1245 * is going to be processed now, so blk layer can do proper initializations
1246 * such as starting the timeout timer.
1248 void blk_mq_start_request(struct request *rq)
1250 struct request_queue *q = rq->q;
1252 trace_block_rq_issue(rq);
1254 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1255 rq->io_start_time_ns = ktime_get_ns();
1256 rq->stats_sectors = blk_rq_sectors(rq);
1257 rq->rq_flags |= RQF_STATS;
1258 rq_qos_issue(q, rq);
1261 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1264 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1266 #ifdef CONFIG_BLK_DEV_INTEGRITY
1267 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1268 q->integrity.profile->prepare_fn(rq);
1270 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1271 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1273 EXPORT_SYMBOL(blk_mq_start_request);
1276 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1277 * queues. This is important for md arrays to benefit from merging
1280 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1282 if (plug->multiple_queues)
1283 return BLK_MAX_REQUEST_COUNT * 2;
1284 return BLK_MAX_REQUEST_COUNT;
1287 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1289 struct request *last = rq_list_peek(&plug->mq_list);
1291 if (!plug->rq_count) {
1292 trace_block_plug(rq->q);
1293 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1294 (!blk_queue_nomerges(rq->q) &&
1295 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1296 blk_mq_flush_plug_list(plug, false);
1298 trace_block_plug(rq->q);
1301 if (!plug->multiple_queues && last && last->q != rq->q)
1302 plug->multiple_queues = true;
1303 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1304 plug->has_elevator = true;
1306 rq_list_add(&plug->mq_list, rq);
1311 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1312 * @rq: request to insert
1313 * @at_head: insert request at head or tail of queue
1316 * Insert a fully prepared request at the back of the I/O scheduler queue
1317 * for execution. Don't wait for completion.
1320 * This function will invoke @done directly if the queue is dead.
1322 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1324 WARN_ON(irqs_disabled());
1325 WARN_ON(!blk_rq_is_passthrough(rq));
1327 blk_account_io_start(rq);
1330 * As plugging can be enabled for passthrough requests on a zoned
1331 * device, directly accessing the plug instead of using blk_mq_plug()
1332 * should not have any consequences.
1334 if (current->plug && !at_head)
1335 blk_add_rq_to_plug(current->plug, rq);
1337 blk_mq_sched_insert_request(rq, at_head, true, false);
1339 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1341 struct blk_rq_wait {
1342 struct completion done;
1346 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1348 struct blk_rq_wait *wait = rq->end_io_data;
1351 complete(&wait->done);
1352 return RQ_END_IO_NONE;
1355 bool blk_rq_is_poll(struct request *rq)
1359 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1363 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1365 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1368 blk_mq_poll(rq->q, blk_rq_to_qc(rq), NULL, 0);
1370 } while (!completion_done(wait));
1374 * blk_execute_rq - insert a request into queue for execution
1375 * @rq: request to insert
1376 * @at_head: insert request at head or tail of queue
1379 * Insert a fully prepared request at the back of the I/O scheduler queue
1380 * for execution and wait for completion.
1381 * Return: The blk_status_t result provided to blk_mq_end_request().
1383 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1385 struct blk_rq_wait wait = {
1386 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1389 WARN_ON(irqs_disabled());
1390 WARN_ON(!blk_rq_is_passthrough(rq));
1392 rq->end_io_data = &wait;
1393 rq->end_io = blk_end_sync_rq;
1395 blk_account_io_start(rq);
1396 blk_mq_sched_insert_request(rq, at_head, true, false);
1398 if (blk_rq_is_poll(rq)) {
1399 blk_rq_poll_completion(rq, &wait.done);
1402 * Prevent hang_check timer from firing at us during very long
1405 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1408 while (!wait_for_completion_io_timeout(&wait.done,
1409 hang_check * (HZ/2)))
1412 wait_for_completion_io(&wait.done);
1417 EXPORT_SYMBOL(blk_execute_rq);
1419 static void __blk_mq_requeue_request(struct request *rq)
1421 struct request_queue *q = rq->q;
1423 blk_mq_put_driver_tag(rq);
1425 trace_block_rq_requeue(rq);
1426 rq_qos_requeue(q, rq);
1428 if (blk_mq_request_started(rq)) {
1429 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1430 rq->rq_flags &= ~RQF_TIMED_OUT;
1434 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1436 __blk_mq_requeue_request(rq);
1438 /* this request will be re-inserted to io scheduler queue */
1439 blk_mq_sched_requeue_request(rq);
1441 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1443 EXPORT_SYMBOL(blk_mq_requeue_request);
1445 static void blk_mq_requeue_work(struct work_struct *work)
1447 struct request_queue *q =
1448 container_of(work, struct request_queue, requeue_work.work);
1450 struct request *rq, *next;
1452 spin_lock_irq(&q->requeue_lock);
1453 list_splice_init(&q->requeue_list, &rq_list);
1454 spin_unlock_irq(&q->requeue_lock);
1456 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1457 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1460 rq->rq_flags &= ~RQF_SOFTBARRIER;
1461 list_del_init(&rq->queuelist);
1463 * If RQF_DONTPREP, rq has contained some driver specific
1464 * data, so insert it to hctx dispatch list to avoid any
1467 if (rq->rq_flags & RQF_DONTPREP)
1468 blk_mq_request_bypass_insert(rq, false, false);
1470 blk_mq_sched_insert_request(rq, true, false, false);
1473 while (!list_empty(&rq_list)) {
1474 rq = list_entry(rq_list.next, struct request, queuelist);
1475 list_del_init(&rq->queuelist);
1476 blk_mq_sched_insert_request(rq, false, false, false);
1479 blk_mq_run_hw_queues(q, false);
1482 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1483 bool kick_requeue_list)
1485 struct request_queue *q = rq->q;
1486 unsigned long flags;
1489 * We abuse this flag that is otherwise used by the I/O scheduler to
1490 * request head insertion from the workqueue.
1492 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1494 spin_lock_irqsave(&q->requeue_lock, flags);
1496 rq->rq_flags |= RQF_SOFTBARRIER;
1497 list_add(&rq->queuelist, &q->requeue_list);
1499 list_add_tail(&rq->queuelist, &q->requeue_list);
1501 spin_unlock_irqrestore(&q->requeue_lock, flags);
1503 if (kick_requeue_list)
1504 blk_mq_kick_requeue_list(q);
1507 void blk_mq_kick_requeue_list(struct request_queue *q)
1509 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1511 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1513 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1514 unsigned long msecs)
1516 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1517 msecs_to_jiffies(msecs));
1519 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1521 static bool blk_is_flush_data_rq(struct request *rq)
1523 return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq);
1526 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1529 * If we find a request that isn't idle we know the queue is busy
1530 * as it's checked in the iter.
1531 * Return false to stop the iteration.
1533 * In case of queue quiesce, if one flush data request is completed,
1534 * don't count it as inflight given the flush sequence is suspended,
1535 * and the original flush data request is invisible to driver, just
1536 * like other pending requests because of quiesce
1538 if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) &&
1539 blk_is_flush_data_rq(rq) &&
1540 blk_mq_request_completed(rq))) {
1550 bool blk_mq_queue_inflight(struct request_queue *q)
1554 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1557 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1559 static void blk_mq_rq_timed_out(struct request *req)
1561 req->rq_flags |= RQF_TIMED_OUT;
1562 if (req->q->mq_ops->timeout) {
1563 enum blk_eh_timer_return ret;
1565 ret = req->q->mq_ops->timeout(req);
1566 if (ret == BLK_EH_DONE)
1568 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1574 struct blk_expired_data {
1575 bool has_timedout_rq;
1577 unsigned long timeout_start;
1580 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1582 unsigned long deadline;
1584 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1586 if (rq->rq_flags & RQF_TIMED_OUT)
1589 deadline = READ_ONCE(rq->deadline);
1590 if (time_after_eq(expired->timeout_start, deadline))
1593 if (expired->next == 0)
1594 expired->next = deadline;
1595 else if (time_after(expired->next, deadline))
1596 expired->next = deadline;
1600 void blk_mq_put_rq_ref(struct request *rq)
1602 if (is_flush_rq(rq)) {
1603 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1604 blk_mq_free_request(rq);
1605 } else if (req_ref_put_and_test(rq)) {
1606 __blk_mq_free_request(rq);
1610 static bool blk_mq_check_expired(struct request *rq, void *priv)
1612 struct blk_expired_data *expired = priv;
1615 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1616 * be reallocated underneath the timeout handler's processing, then
1617 * the expire check is reliable. If the request is not expired, then
1618 * it was completed and reallocated as a new request after returning
1619 * from blk_mq_check_expired().
1621 if (blk_mq_req_expired(rq, expired)) {
1622 expired->has_timedout_rq = true;
1628 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1630 struct blk_expired_data *expired = priv;
1632 if (blk_mq_req_expired(rq, expired))
1633 blk_mq_rq_timed_out(rq);
1637 static void blk_mq_timeout_work(struct work_struct *work)
1639 struct request_queue *q =
1640 container_of(work, struct request_queue, timeout_work);
1641 struct blk_expired_data expired = {
1642 .timeout_start = jiffies,
1644 struct blk_mq_hw_ctx *hctx;
1647 /* A deadlock might occur if a request is stuck requiring a
1648 * timeout at the same time a queue freeze is waiting
1649 * completion, since the timeout code would not be able to
1650 * acquire the queue reference here.
1652 * That's why we don't use blk_queue_enter here; instead, we use
1653 * percpu_ref_tryget directly, because we need to be able to
1654 * obtain a reference even in the short window between the queue
1655 * starting to freeze, by dropping the first reference in
1656 * blk_freeze_queue_start, and the moment the last request is
1657 * consumed, marked by the instant q_usage_counter reaches
1660 if (!percpu_ref_tryget(&q->q_usage_counter))
1663 /* check if there is any timed-out request */
1664 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1665 if (expired.has_timedout_rq) {
1667 * Before walking tags, we must ensure any submit started
1668 * before the current time has finished. Since the submit
1669 * uses srcu or rcu, wait for a synchronization point to
1670 * ensure all running submits have finished
1672 blk_mq_wait_quiesce_done(q);
1675 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1678 if (expired.next != 0) {
1679 mod_timer(&q->timeout, expired.next);
1682 * Request timeouts are handled as a forward rolling timer. If
1683 * we end up here it means that no requests are pending and
1684 * also that no request has been pending for a while. Mark
1685 * each hctx as idle.
1687 queue_for_each_hw_ctx(q, hctx, i) {
1688 /* the hctx may be unmapped, so check it here */
1689 if (blk_mq_hw_queue_mapped(hctx))
1690 blk_mq_tag_idle(hctx);
1696 struct flush_busy_ctx_data {
1697 struct blk_mq_hw_ctx *hctx;
1698 struct list_head *list;
1701 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1703 struct flush_busy_ctx_data *flush_data = data;
1704 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1705 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1706 enum hctx_type type = hctx->type;
1708 spin_lock(&ctx->lock);
1709 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1710 sbitmap_clear_bit(sb, bitnr);
1711 spin_unlock(&ctx->lock);
1716 * Process software queues that have been marked busy, splicing them
1717 * to the for-dispatch
1719 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1721 struct flush_busy_ctx_data data = {
1726 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1728 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1730 struct dispatch_rq_data {
1731 struct blk_mq_hw_ctx *hctx;
1735 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1738 struct dispatch_rq_data *dispatch_data = data;
1739 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1740 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1741 enum hctx_type type = hctx->type;
1743 spin_lock(&ctx->lock);
1744 if (!list_empty(&ctx->rq_lists[type])) {
1745 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1746 list_del_init(&dispatch_data->rq->queuelist);
1747 if (list_empty(&ctx->rq_lists[type]))
1748 sbitmap_clear_bit(sb, bitnr);
1750 spin_unlock(&ctx->lock);
1752 return !dispatch_data->rq;
1755 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1756 struct blk_mq_ctx *start)
1758 unsigned off = start ? start->index_hw[hctx->type] : 0;
1759 struct dispatch_rq_data data = {
1764 __sbitmap_for_each_set(&hctx->ctx_map, off,
1765 dispatch_rq_from_ctx, &data);
1770 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1772 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1773 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1776 blk_mq_tag_busy(rq->mq_hctx);
1778 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1779 bt = &rq->mq_hctx->tags->breserved_tags;
1782 if (!hctx_may_queue(rq->mq_hctx, bt))
1786 tag = __sbitmap_queue_get(bt);
1787 if (tag == BLK_MQ_NO_TAG)
1790 rq->tag = tag + tag_offset;
1794 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1796 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1799 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1800 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1801 rq->rq_flags |= RQF_MQ_INFLIGHT;
1802 __blk_mq_inc_active_requests(hctx);
1804 hctx->tags->rqs[rq->tag] = rq;
1808 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1809 int flags, void *key)
1811 struct blk_mq_hw_ctx *hctx;
1813 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1815 spin_lock(&hctx->dispatch_wait_lock);
1816 if (!list_empty(&wait->entry)) {
1817 struct sbitmap_queue *sbq;
1819 list_del_init(&wait->entry);
1820 sbq = &hctx->tags->bitmap_tags;
1821 atomic_dec(&sbq->ws_active);
1823 spin_unlock(&hctx->dispatch_wait_lock);
1825 blk_mq_run_hw_queue(hctx, true);
1830 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1831 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1832 * restart. For both cases, take care to check the condition again after
1833 * marking us as waiting.
1835 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1838 struct sbitmap_queue *sbq;
1839 struct wait_queue_head *wq;
1840 wait_queue_entry_t *wait;
1843 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1844 !(blk_mq_is_shared_tags(hctx->flags))) {
1845 blk_mq_sched_mark_restart_hctx(hctx);
1848 * It's possible that a tag was freed in the window between the
1849 * allocation failure and adding the hardware queue to the wait
1852 * Don't clear RESTART here, someone else could have set it.
1853 * At most this will cost an extra queue run.
1855 return blk_mq_get_driver_tag(rq);
1858 wait = &hctx->dispatch_wait;
1859 if (!list_empty_careful(&wait->entry))
1862 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1863 sbq = &hctx->tags->breserved_tags;
1865 sbq = &hctx->tags->bitmap_tags;
1866 wq = &bt_wait_ptr(sbq, hctx)->wait;
1868 spin_lock_irq(&wq->lock);
1869 spin_lock(&hctx->dispatch_wait_lock);
1870 if (!list_empty(&wait->entry)) {
1871 spin_unlock(&hctx->dispatch_wait_lock);
1872 spin_unlock_irq(&wq->lock);
1876 atomic_inc(&sbq->ws_active);
1877 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1878 __add_wait_queue(wq, wait);
1881 * Add one explicit barrier since blk_mq_get_driver_tag() may
1882 * not imply barrier in case of failure.
1884 * Order adding us to wait queue and allocating driver tag.
1886 * The pair is the one implied in sbitmap_queue_wake_up() which
1887 * orders clearing sbitmap tag bits and waitqueue_active() in
1888 * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless
1890 * Otherwise, re-order of adding wait queue and getting driver tag
1891 * may cause __sbitmap_queue_wake_up() to wake up nothing because
1892 * the waitqueue_active() may not observe us in wait queue.
1897 * It's possible that a tag was freed in the window between the
1898 * allocation failure and adding the hardware queue to the wait
1901 ret = blk_mq_get_driver_tag(rq);
1903 spin_unlock(&hctx->dispatch_wait_lock);
1904 spin_unlock_irq(&wq->lock);
1909 * We got a tag, remove ourselves from the wait queue to ensure
1910 * someone else gets the wakeup.
1912 list_del_init(&wait->entry);
1913 atomic_dec(&sbq->ws_active);
1914 spin_unlock(&hctx->dispatch_wait_lock);
1915 spin_unlock_irq(&wq->lock);
1920 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1921 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1923 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1924 * - EWMA is one simple way to compute running average value
1925 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1926 * - take 4 as factor for avoiding to get too small(0) result, and this
1927 * factor doesn't matter because EWMA decreases exponentially
1929 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1933 ewma = hctx->dispatch_busy;
1938 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1940 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1941 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1943 hctx->dispatch_busy = ewma;
1946 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1948 static void blk_mq_handle_dev_resource(struct request *rq,
1949 struct list_head *list)
1951 struct request *next =
1952 list_first_entry_or_null(list, struct request, queuelist);
1955 * If an I/O scheduler has been configured and we got a driver tag for
1956 * the next request already, free it.
1959 blk_mq_put_driver_tag(next);
1961 list_add(&rq->queuelist, list);
1962 __blk_mq_requeue_request(rq);
1965 static void blk_mq_handle_zone_resource(struct request *rq,
1966 struct list_head *zone_list)
1969 * If we end up here it is because we cannot dispatch a request to a
1970 * specific zone due to LLD level zone-write locking or other zone
1971 * related resource not being available. In this case, set the request
1972 * aside in zone_list for retrying it later.
1974 list_add(&rq->queuelist, zone_list);
1975 __blk_mq_requeue_request(rq);
1978 enum prep_dispatch {
1980 PREP_DISPATCH_NO_TAG,
1981 PREP_DISPATCH_NO_BUDGET,
1984 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1987 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1988 int budget_token = -1;
1991 budget_token = blk_mq_get_dispatch_budget(rq->q);
1992 if (budget_token < 0) {
1993 blk_mq_put_driver_tag(rq);
1994 return PREP_DISPATCH_NO_BUDGET;
1996 blk_mq_set_rq_budget_token(rq, budget_token);
1999 if (!blk_mq_get_driver_tag(rq)) {
2001 * The initial allocation attempt failed, so we need to
2002 * rerun the hardware queue when a tag is freed. The
2003 * waitqueue takes care of that. If the queue is run
2004 * before we add this entry back on the dispatch list,
2005 * we'll re-run it below.
2007 if (!blk_mq_mark_tag_wait(hctx, rq)) {
2009 * All budgets not got from this function will be put
2010 * together during handling partial dispatch
2013 blk_mq_put_dispatch_budget(rq->q, budget_token);
2014 return PREP_DISPATCH_NO_TAG;
2018 return PREP_DISPATCH_OK;
2021 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
2022 static void blk_mq_release_budgets(struct request_queue *q,
2023 struct list_head *list)
2027 list_for_each_entry(rq, list, queuelist) {
2028 int budget_token = blk_mq_get_rq_budget_token(rq);
2030 if (budget_token >= 0)
2031 blk_mq_put_dispatch_budget(q, budget_token);
2036 * Returns true if we did some work AND can potentially do more.
2038 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
2039 unsigned int nr_budgets)
2041 enum prep_dispatch prep;
2042 struct request_queue *q = hctx->queue;
2043 struct request *rq, *nxt;
2045 blk_status_t ret = BLK_STS_OK;
2046 LIST_HEAD(zone_list);
2047 bool needs_resource = false;
2049 if (list_empty(list))
2053 * Now process all the entries, sending them to the driver.
2055 errors = queued = 0;
2057 struct blk_mq_queue_data bd;
2059 rq = list_first_entry(list, struct request, queuelist);
2061 WARN_ON_ONCE(hctx != rq->mq_hctx);
2062 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2063 if (prep != PREP_DISPATCH_OK)
2066 list_del_init(&rq->queuelist);
2071 * Flag last if we have no more requests, or if we have more
2072 * but can't assign a driver tag to it.
2074 if (list_empty(list))
2077 nxt = list_first_entry(list, struct request, queuelist);
2078 bd.last = !blk_mq_get_driver_tag(nxt);
2082 * once the request is queued to lld, no need to cover the
2087 ret = q->mq_ops->queue_rq(hctx, &bd);
2092 case BLK_STS_RESOURCE:
2093 needs_resource = true;
2095 case BLK_STS_DEV_RESOURCE:
2096 blk_mq_handle_dev_resource(rq, list);
2098 case BLK_STS_ZONE_RESOURCE:
2100 * Move the request to zone_list and keep going through
2101 * the dispatch list to find more requests the drive can
2104 blk_mq_handle_zone_resource(rq, &zone_list);
2105 needs_resource = true;
2109 blk_mq_end_request(rq, ret);
2111 } while (!list_empty(list));
2113 if (!list_empty(&zone_list))
2114 list_splice_tail_init(&zone_list, list);
2116 /* If we didn't flush the entire list, we could have told the driver
2117 * there was more coming, but that turned out to be a lie.
2119 if ((!list_empty(list) || errors || needs_resource ||
2120 ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued)
2121 q->mq_ops->commit_rqs(hctx);
2123 * Any items that need requeuing? Stuff them into hctx->dispatch,
2124 * that is where we will continue on next queue run.
2126 if (!list_empty(list)) {
2128 /* For non-shared tags, the RESTART check will suffice */
2129 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2130 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2131 blk_mq_is_shared_tags(hctx->flags));
2134 blk_mq_release_budgets(q, list);
2136 spin_lock(&hctx->lock);
2137 list_splice_tail_init(list, &hctx->dispatch);
2138 spin_unlock(&hctx->lock);
2141 * Order adding requests to hctx->dispatch and checking
2142 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2143 * in blk_mq_sched_restart(). Avoid restart code path to
2144 * miss the new added requests to hctx->dispatch, meantime
2145 * SCHED_RESTART is observed here.
2150 * If SCHED_RESTART was set by the caller of this function and
2151 * it is no longer set that means that it was cleared by another
2152 * thread and hence that a queue rerun is needed.
2154 * If 'no_tag' is set, that means that we failed getting
2155 * a driver tag with an I/O scheduler attached. If our dispatch
2156 * waitqueue is no longer active, ensure that we run the queue
2157 * AFTER adding our entries back to the list.
2159 * If no I/O scheduler has been configured it is possible that
2160 * the hardware queue got stopped and restarted before requests
2161 * were pushed back onto the dispatch list. Rerun the queue to
2162 * avoid starvation. Notes:
2163 * - blk_mq_run_hw_queue() checks whether or not a queue has
2164 * been stopped before rerunning a queue.
2165 * - Some but not all block drivers stop a queue before
2166 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2169 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2170 * bit is set, run queue after a delay to avoid IO stalls
2171 * that could otherwise occur if the queue is idle. We'll do
2172 * similar if we couldn't get budget or couldn't lock a zone
2173 * and SCHED_RESTART is set.
2175 needs_restart = blk_mq_sched_needs_restart(hctx);
2176 if (prep == PREP_DISPATCH_NO_BUDGET)
2177 needs_resource = true;
2178 if (!needs_restart ||
2179 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2180 blk_mq_run_hw_queue(hctx, true);
2181 else if (needs_resource)
2182 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2184 blk_mq_update_dispatch_busy(hctx, true);
2187 blk_mq_update_dispatch_busy(hctx, false);
2189 return (queued + errors) != 0;
2193 * __blk_mq_run_hw_queue - Run a hardware queue.
2194 * @hctx: Pointer to the hardware queue to run.
2196 * Send pending requests to the hardware.
2198 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
2201 * We can't run the queue inline with ints disabled. Ensure that
2202 * we catch bad users of this early.
2204 WARN_ON_ONCE(in_interrupt());
2206 blk_mq_run_dispatch_ops(hctx->queue,
2207 blk_mq_sched_dispatch_requests(hctx));
2210 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2212 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2214 if (cpu >= nr_cpu_ids)
2215 cpu = cpumask_first(hctx->cpumask);
2220 * It'd be great if the workqueue API had a way to pass
2221 * in a mask and had some smarts for more clever placement.
2222 * For now we just round-robin here, switching for every
2223 * BLK_MQ_CPU_WORK_BATCH queued items.
2225 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2228 int next_cpu = hctx->next_cpu;
2230 if (hctx->queue->nr_hw_queues == 1)
2231 return WORK_CPU_UNBOUND;
2233 if (--hctx->next_cpu_batch <= 0) {
2235 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2237 if (next_cpu >= nr_cpu_ids)
2238 next_cpu = blk_mq_first_mapped_cpu(hctx);
2239 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2243 * Do unbound schedule if we can't find a online CPU for this hctx,
2244 * and it should only happen in the path of handling CPU DEAD.
2246 if (!cpu_online(next_cpu)) {
2253 * Make sure to re-select CPU next time once after CPUs
2254 * in hctx->cpumask become online again.
2256 hctx->next_cpu = next_cpu;
2257 hctx->next_cpu_batch = 1;
2258 return WORK_CPU_UNBOUND;
2261 hctx->next_cpu = next_cpu;
2266 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2267 * @hctx: Pointer to the hardware queue to run.
2268 * @async: If we want to run the queue asynchronously.
2269 * @msecs: Milliseconds of delay to wait before running the queue.
2271 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2272 * with a delay of @msecs.
2274 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2275 unsigned long msecs)
2277 if (unlikely(blk_mq_hctx_stopped(hctx)))
2280 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2281 if (cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2282 __blk_mq_run_hw_queue(hctx);
2287 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2288 msecs_to_jiffies(msecs));
2292 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2293 * @hctx: Pointer to the hardware queue to run.
2294 * @msecs: Milliseconds of delay to wait before running the queue.
2296 * Run a hardware queue asynchronously with a delay of @msecs.
2298 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2300 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2302 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2305 * blk_mq_run_hw_queue - Start to run a hardware queue.
2306 * @hctx: Pointer to the hardware queue to run.
2307 * @async: If we want to run the queue asynchronously.
2309 * Check if the request queue is not in a quiesced state and if there are
2310 * pending requests to be sent. If this is true, run the queue to send requests
2313 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2318 * When queue is quiesced, we may be switching io scheduler, or
2319 * updating nr_hw_queues, or other things, and we can't run queue
2320 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2322 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2325 __blk_mq_run_dispatch_ops(hctx->queue, false,
2326 need_run = !blk_queue_quiesced(hctx->queue) &&
2327 blk_mq_hctx_has_pending(hctx));
2330 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2332 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2335 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2338 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2340 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2342 * If the IO scheduler does not respect hardware queues when
2343 * dispatching, we just don't bother with multiple HW queues and
2344 * dispatch from hctx for the current CPU since running multiple queues
2345 * just causes lock contention inside the scheduler and pointless cache
2348 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2350 if (!blk_mq_hctx_stopped(hctx))
2356 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2357 * @q: Pointer to the request queue to run.
2358 * @async: If we want to run the queue asynchronously.
2360 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2362 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2366 if (blk_queue_sq_sched(q))
2367 sq_hctx = blk_mq_get_sq_hctx(q);
2368 queue_for_each_hw_ctx(q, hctx, i) {
2369 if (blk_mq_hctx_stopped(hctx))
2372 * Dispatch from this hctx either if there's no hctx preferred
2373 * by IO scheduler or if it has requests that bypass the
2376 if (!sq_hctx || sq_hctx == hctx ||
2377 !list_empty_careful(&hctx->dispatch))
2378 blk_mq_run_hw_queue(hctx, async);
2381 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2384 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2385 * @q: Pointer to the request queue to run.
2386 * @msecs: Milliseconds of delay to wait before running the queues.
2388 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2390 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2394 if (blk_queue_sq_sched(q))
2395 sq_hctx = blk_mq_get_sq_hctx(q);
2396 queue_for_each_hw_ctx(q, hctx, i) {
2397 if (blk_mq_hctx_stopped(hctx))
2400 * If there is already a run_work pending, leave the
2401 * pending delay untouched. Otherwise, a hctx can stall
2402 * if another hctx is re-delaying the other's work
2403 * before the work executes.
2405 if (delayed_work_pending(&hctx->run_work))
2408 * Dispatch from this hctx either if there's no hctx preferred
2409 * by IO scheduler or if it has requests that bypass the
2412 if (!sq_hctx || sq_hctx == hctx ||
2413 !list_empty_careful(&hctx->dispatch))
2414 blk_mq_delay_run_hw_queue(hctx, msecs);
2417 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2420 * This function is often used for pausing .queue_rq() by driver when
2421 * there isn't enough resource or some conditions aren't satisfied, and
2422 * BLK_STS_RESOURCE is usually returned.
2424 * We do not guarantee that dispatch can be drained or blocked
2425 * after blk_mq_stop_hw_queue() returns. Please use
2426 * blk_mq_quiesce_queue() for that requirement.
2428 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2430 cancel_delayed_work(&hctx->run_work);
2432 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2434 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2437 * This function is often used for pausing .queue_rq() by driver when
2438 * there isn't enough resource or some conditions aren't satisfied, and
2439 * BLK_STS_RESOURCE is usually returned.
2441 * We do not guarantee that dispatch can be drained or blocked
2442 * after blk_mq_stop_hw_queues() returns. Please use
2443 * blk_mq_quiesce_queue() for that requirement.
2445 void blk_mq_stop_hw_queues(struct request_queue *q)
2447 struct blk_mq_hw_ctx *hctx;
2450 queue_for_each_hw_ctx(q, hctx, i)
2451 blk_mq_stop_hw_queue(hctx);
2453 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2455 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2457 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2459 blk_mq_run_hw_queue(hctx, false);
2461 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2463 void blk_mq_start_hw_queues(struct request_queue *q)
2465 struct blk_mq_hw_ctx *hctx;
2468 queue_for_each_hw_ctx(q, hctx, i)
2469 blk_mq_start_hw_queue(hctx);
2471 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2473 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2475 if (!blk_mq_hctx_stopped(hctx))
2478 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2479 blk_mq_run_hw_queue(hctx, async);
2481 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2483 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2485 struct blk_mq_hw_ctx *hctx;
2488 queue_for_each_hw_ctx(q, hctx, i)
2489 blk_mq_start_stopped_hw_queue(hctx, async);
2491 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2493 static void blk_mq_run_work_fn(struct work_struct *work)
2495 struct blk_mq_hw_ctx *hctx;
2497 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2500 * If we are stopped, don't run the queue.
2502 if (blk_mq_hctx_stopped(hctx))
2505 __blk_mq_run_hw_queue(hctx);
2508 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2512 struct blk_mq_ctx *ctx = rq->mq_ctx;
2513 enum hctx_type type = hctx->type;
2515 lockdep_assert_held(&ctx->lock);
2517 trace_block_rq_insert(rq);
2520 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2522 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2525 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2528 struct blk_mq_ctx *ctx = rq->mq_ctx;
2530 lockdep_assert_held(&ctx->lock);
2532 __blk_mq_insert_req_list(hctx, rq, at_head);
2533 blk_mq_hctx_mark_pending(hctx, ctx);
2537 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2538 * @rq: Pointer to request to be inserted.
2539 * @at_head: true if the request should be inserted at the head of the list.
2540 * @run_queue: If we should run the hardware queue after inserting the request.
2542 * Should only be used carefully, when the caller knows we want to
2543 * bypass a potential IO scheduler on the target device.
2545 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2548 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2550 spin_lock(&hctx->lock);
2552 list_add(&rq->queuelist, &hctx->dispatch);
2554 list_add_tail(&rq->queuelist, &hctx->dispatch);
2555 spin_unlock(&hctx->lock);
2558 blk_mq_run_hw_queue(hctx, false);
2561 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2562 struct list_head *list)
2566 enum hctx_type type = hctx->type;
2569 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2572 list_for_each_entry(rq, list, queuelist) {
2573 BUG_ON(rq->mq_ctx != ctx);
2574 trace_block_rq_insert(rq);
2577 spin_lock(&ctx->lock);
2578 list_splice_tail_init(list, &ctx->rq_lists[type]);
2579 blk_mq_hctx_mark_pending(hctx, ctx);
2580 spin_unlock(&ctx->lock);
2583 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2586 if (hctx->queue->mq_ops->commit_rqs) {
2587 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2588 hctx->queue->mq_ops->commit_rqs(hctx);
2593 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2594 unsigned int nr_segs)
2598 if (bio->bi_opf & REQ_RAHEAD)
2599 rq->cmd_flags |= REQ_FAILFAST_MASK;
2601 rq->__sector = bio->bi_iter.bi_sector;
2602 blk_rq_bio_prep(rq, bio, nr_segs);
2604 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2605 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2608 blk_account_io_start(rq);
2611 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2612 struct request *rq, bool last)
2614 struct request_queue *q = rq->q;
2615 struct blk_mq_queue_data bd = {
2622 * For OK queue, we are done. For error, caller may kill it.
2623 * Any other error (busy), just add it to our list as we
2624 * previously would have done.
2626 ret = q->mq_ops->queue_rq(hctx, &bd);
2629 blk_mq_update_dispatch_busy(hctx, false);
2631 case BLK_STS_RESOURCE:
2632 case BLK_STS_DEV_RESOURCE:
2633 blk_mq_update_dispatch_busy(hctx, true);
2634 __blk_mq_requeue_request(rq);
2637 blk_mq_update_dispatch_busy(hctx, false);
2644 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2646 bool bypass_insert, bool last)
2648 struct request_queue *q = rq->q;
2649 bool run_queue = true;
2653 * RCU or SRCU read lock is needed before checking quiesced flag.
2655 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2656 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2657 * and avoid driver to try to dispatch again.
2659 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2661 bypass_insert = false;
2665 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2668 budget_token = blk_mq_get_dispatch_budget(q);
2669 if (budget_token < 0)
2672 blk_mq_set_rq_budget_token(rq, budget_token);
2674 if (!blk_mq_get_driver_tag(rq)) {
2675 blk_mq_put_dispatch_budget(q, budget_token);
2679 return __blk_mq_issue_directly(hctx, rq, last);
2682 return BLK_STS_RESOURCE;
2684 blk_mq_sched_insert_request(rq, false, run_queue, false);
2690 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2691 * @hctx: Pointer of the associated hardware queue.
2692 * @rq: Pointer to request to be sent.
2694 * If the device has enough resources to accept a new request now, send the
2695 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2696 * we can try send it another time in the future. Requests inserted at this
2697 * queue have higher priority.
2699 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2703 __blk_mq_try_issue_directly(hctx, rq, false, true);
2705 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2706 blk_mq_request_bypass_insert(rq, false, true);
2707 else if (ret != BLK_STS_OK)
2708 blk_mq_end_request(rq, ret);
2711 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2713 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2716 static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2718 struct blk_mq_hw_ctx *hctx = NULL;
2723 while ((rq = rq_list_pop(&plug->mq_list))) {
2724 bool last = rq_list_empty(plug->mq_list);
2727 if (hctx != rq->mq_hctx) {
2729 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2733 ret = blk_mq_request_issue_directly(rq, last);
2738 case BLK_STS_RESOURCE:
2739 case BLK_STS_DEV_RESOURCE:
2740 blk_mq_request_bypass_insert(rq, false, true);
2741 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2744 blk_mq_end_request(rq, ret);
2751 * If we didn't flush the entire list, we could have told the driver
2752 * there was more coming, but that turned out to be a lie.
2755 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2758 static void __blk_mq_flush_plug_list(struct request_queue *q,
2759 struct blk_plug *plug)
2761 if (blk_queue_quiesced(q))
2763 q->mq_ops->queue_rqs(&plug->mq_list);
2766 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2768 struct blk_mq_hw_ctx *this_hctx = NULL;
2769 struct blk_mq_ctx *this_ctx = NULL;
2770 struct request *requeue_list = NULL;
2771 struct request **requeue_lastp = &requeue_list;
2772 unsigned int depth = 0;
2776 struct request *rq = rq_list_pop(&plug->mq_list);
2779 this_hctx = rq->mq_hctx;
2780 this_ctx = rq->mq_ctx;
2781 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2782 rq_list_add_tail(&requeue_lastp, rq);
2785 list_add(&rq->queuelist, &list);
2787 } while (!rq_list_empty(plug->mq_list));
2789 plug->mq_list = requeue_list;
2790 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2791 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched);
2794 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2799 * We may have been called recursively midway through handling
2800 * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2801 * To avoid mq_list changing under our feet, clear rq_count early and
2802 * bail out specifically if rq_count is 0 rather than checking
2803 * whether the mq_list is empty.
2805 if (plug->rq_count == 0)
2809 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2810 struct request_queue *q;
2812 rq = rq_list_peek(&plug->mq_list);
2816 * Peek first request and see if we have a ->queue_rqs() hook.
2817 * If we do, we can dispatch the whole plug list in one go. We
2818 * already know at this point that all requests belong to the
2819 * same queue, caller must ensure that's the case.
2821 * Since we pass off the full list to the driver at this point,
2822 * we do not increment the active request count for the queue.
2823 * Bypass shared tags for now because of that.
2825 if (q->mq_ops->queue_rqs &&
2826 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2827 blk_mq_run_dispatch_ops(q,
2828 __blk_mq_flush_plug_list(q, plug));
2829 if (rq_list_empty(plug->mq_list))
2833 blk_mq_run_dispatch_ops(q,
2834 blk_mq_plug_issue_direct(plug, false));
2835 if (rq_list_empty(plug->mq_list))
2840 blk_mq_dispatch_plug_list(plug, from_schedule);
2841 } while (!rq_list_empty(plug->mq_list));
2844 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2845 struct list_head *list)
2850 while (!list_empty(list)) {
2852 struct request *rq = list_first_entry(list, struct request,
2855 list_del_init(&rq->queuelist);
2856 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2857 if (ret != BLK_STS_OK) {
2859 if (ret == BLK_STS_RESOURCE ||
2860 ret == BLK_STS_DEV_RESOURCE) {
2861 blk_mq_request_bypass_insert(rq, false,
2865 blk_mq_end_request(rq, ret);
2871 * If we didn't flush the entire list, we could have told
2872 * the driver there was more coming, but that turned out to
2875 if ((!list_empty(list) || errors) &&
2876 hctx->queue->mq_ops->commit_rqs && queued)
2877 hctx->queue->mq_ops->commit_rqs(hctx);
2880 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2881 struct bio *bio, unsigned int nr_segs)
2883 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2884 if (blk_attempt_plug_merge(q, bio, nr_segs))
2886 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2892 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2893 struct blk_plug *plug,
2897 struct blk_mq_alloc_data data = {
2900 .cmd_flags = bio->bi_opf,
2904 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2907 rq_qos_throttle(q, bio);
2910 data.nr_tags = plug->nr_ios;
2912 data.cached_rq = &plug->cached_rq;
2915 rq = __blk_mq_alloc_requests(&data);
2918 rq_qos_cleanup(q, bio);
2919 if (bio->bi_opf & REQ_NOWAIT)
2920 bio_wouldblock_error(bio);
2924 /* return true if this @rq can be used for @bio */
2925 static bool blk_mq_can_use_cached_rq(struct request *rq, struct blk_plug *plug,
2928 enum hctx_type type = blk_mq_get_hctx_type(bio->bi_opf);
2929 enum hctx_type hctx_type = rq->mq_hctx->type;
2931 WARN_ON_ONCE(rq_list_peek(&plug->cached_rq) != rq);
2933 if (type != hctx_type &&
2934 !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2936 if (op_is_flush(rq->cmd_flags) != op_is_flush(bio->bi_opf))
2940 * If any qos ->throttle() end up blocking, we will have flushed the
2941 * plug and hence killed the cached_rq list as well. Pop this entry
2942 * before we throttle.
2944 plug->cached_rq = rq_list_next(rq);
2945 rq_qos_throttle(rq->q, bio);
2947 rq->cmd_flags = bio->bi_opf;
2948 INIT_LIST_HEAD(&rq->queuelist);
2952 static void bio_set_ioprio(struct bio *bio)
2954 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2955 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2956 bio->bi_ioprio = get_current_ioprio();
2957 blkcg_set_ioprio(bio);
2961 * blk_mq_submit_bio - Create and send a request to block device.
2962 * @bio: Bio pointer.
2964 * Builds up a request structure from @q and @bio and send to the device. The
2965 * request may not be queued directly to hardware if:
2966 * * This request can be merged with another one
2967 * * We want to place request at plug queue for possible future merging
2968 * * There is an IO scheduler active at this queue
2970 * It will not queue the request if there is an error with the bio, or at the
2973 void blk_mq_submit_bio(struct bio *bio)
2975 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2976 struct blk_plug *plug = blk_mq_plug(bio);
2977 const int is_sync = op_is_sync(bio->bi_opf);
2978 struct request *rq = NULL;
2979 unsigned int nr_segs = 1;
2982 bio = blk_queue_bounce(bio, q);
2983 bio_set_ioprio(bio);
2986 rq = rq_list_peek(&plug->cached_rq);
2987 if (rq && rq->q != q)
2991 if (unlikely(bio_may_exceed_limits(bio, &q->limits))) {
2992 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2996 if (!bio_integrity_prep(bio))
2998 if (blk_mq_attempt_bio_merge(q, bio, nr_segs))
3000 if (blk_mq_can_use_cached_rq(rq, plug, bio))
3002 percpu_ref_get(&q->q_usage_counter);
3004 if (unlikely(bio_queue_enter(bio)))
3006 if (unlikely(bio_may_exceed_limits(bio, &q->limits))) {
3007 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
3011 if (!bio_integrity_prep(bio))
3015 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
3016 if (unlikely(!rq)) {
3023 trace_block_getrq(bio);
3025 rq_qos_track(q, rq, bio);
3027 blk_mq_bio_to_request(rq, bio, nr_segs);
3029 ret = blk_crypto_rq_get_keyslot(rq);
3030 if (ret != BLK_STS_OK) {
3031 bio->bi_status = ret;
3033 blk_mq_free_request(rq);
3037 if (op_is_flush(bio->bi_opf)) {
3038 blk_insert_flush(rq);
3043 blk_add_rq_to_plug(plug, rq);
3044 else if ((rq->rq_flags & RQF_ELV) ||
3045 (rq->mq_hctx->dispatch_busy &&
3046 (q->nr_hw_queues == 1 || !is_sync)))
3047 blk_mq_sched_insert_request(rq, false, true, true);
3049 blk_mq_run_dispatch_ops(rq->q,
3050 blk_mq_try_issue_directly(rq->mq_hctx, rq));
3053 #ifdef CONFIG_BLK_MQ_STACKING
3055 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3056 * @rq: the request being queued
3058 blk_status_t blk_insert_cloned_request(struct request *rq)
3060 struct request_queue *q = rq->q;
3061 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3064 if (blk_rq_sectors(rq) > max_sectors) {
3066 * SCSI device does not have a good way to return if
3067 * Write Same/Zero is actually supported. If a device rejects
3068 * a non-read/write command (discard, write same,etc.) the
3069 * low-level device driver will set the relevant queue limit to
3070 * 0 to prevent blk-lib from issuing more of the offending
3071 * operations. Commands queued prior to the queue limit being
3072 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3073 * errors being propagated to upper layers.
3075 if (max_sectors == 0)
3076 return BLK_STS_NOTSUPP;
3078 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3079 __func__, blk_rq_sectors(rq), max_sectors);
3080 return BLK_STS_IOERR;
3084 * The queue settings related to segment counting may differ from the
3087 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3088 if (rq->nr_phys_segments > queue_max_segments(q)) {
3089 printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n",
3090 __func__, rq->nr_phys_segments, queue_max_segments(q));
3091 return BLK_STS_IOERR;
3094 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3095 return BLK_STS_IOERR;
3097 if (blk_crypto_insert_cloned_request(rq))
3098 return BLK_STS_IOERR;
3100 blk_account_io_start(rq);
3103 * Since we have a scheduler attached on the top device,
3104 * bypass a potential scheduler on the bottom device for
3107 blk_mq_run_dispatch_ops(q,
3108 ret = blk_mq_request_issue_directly(rq, true));
3110 blk_account_io_done(rq, ktime_get_ns());
3113 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3116 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3117 * @rq: the clone request to be cleaned up
3120 * Free all bios in @rq for a cloned request.
3122 void blk_rq_unprep_clone(struct request *rq)
3126 while ((bio = rq->bio) != NULL) {
3127 rq->bio = bio->bi_next;
3132 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3135 * blk_rq_prep_clone - Helper function to setup clone request
3136 * @rq: the request to be setup
3137 * @rq_src: original request to be cloned
3138 * @bs: bio_set that bios for clone are allocated from
3139 * @gfp_mask: memory allocation mask for bio
3140 * @bio_ctr: setup function to be called for each clone bio.
3141 * Returns %0 for success, non %0 for failure.
3142 * @data: private data to be passed to @bio_ctr
3145 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3146 * Also, pages which the original bios are pointing to are not copied
3147 * and the cloned bios just point same pages.
3148 * So cloned bios must be completed before original bios, which means
3149 * the caller must complete @rq before @rq_src.
3151 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3152 struct bio_set *bs, gfp_t gfp_mask,
3153 int (*bio_ctr)(struct bio *, struct bio *, void *),
3156 struct bio *bio, *bio_src;
3161 __rq_for_each_bio(bio_src, rq_src) {
3162 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3167 if (bio_ctr && bio_ctr(bio, bio_src, data))
3171 rq->biotail->bi_next = bio;
3174 rq->bio = rq->biotail = bio;
3179 /* Copy attributes of the original request to the clone request. */
3180 rq->__sector = blk_rq_pos(rq_src);
3181 rq->__data_len = blk_rq_bytes(rq_src);
3182 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3183 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3184 rq->special_vec = rq_src->special_vec;
3186 rq->nr_phys_segments = rq_src->nr_phys_segments;
3187 rq->ioprio = rq_src->ioprio;
3189 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3197 blk_rq_unprep_clone(rq);
3201 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3202 #endif /* CONFIG_BLK_MQ_STACKING */
3205 * Steal bios from a request and add them to a bio list.
3206 * The request must not have been partially completed before.
3208 void blk_steal_bios(struct bio_list *list, struct request *rq)
3212 list->tail->bi_next = rq->bio;
3214 list->head = rq->bio;
3215 list->tail = rq->biotail;
3223 EXPORT_SYMBOL_GPL(blk_steal_bios);
3225 static size_t order_to_size(unsigned int order)
3227 return (size_t)PAGE_SIZE << order;
3230 /* called before freeing request pool in @tags */
3231 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3232 struct blk_mq_tags *tags)
3235 unsigned long flags;
3238 * There is no need to clear mapping if driver tags is not initialized
3239 * or the mapping belongs to the driver tags.
3241 if (!drv_tags || drv_tags == tags)
3244 list_for_each_entry(page, &tags->page_list, lru) {
3245 unsigned long start = (unsigned long)page_address(page);
3246 unsigned long end = start + order_to_size(page->private);
3249 for (i = 0; i < drv_tags->nr_tags; i++) {
3250 struct request *rq = drv_tags->rqs[i];
3251 unsigned long rq_addr = (unsigned long)rq;
3253 if (rq_addr >= start && rq_addr < end) {
3254 WARN_ON_ONCE(req_ref_read(rq) != 0);
3255 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3261 * Wait until all pending iteration is done.
3263 * Request reference is cleared and it is guaranteed to be observed
3264 * after the ->lock is released.
3266 spin_lock_irqsave(&drv_tags->lock, flags);
3267 spin_unlock_irqrestore(&drv_tags->lock, flags);
3270 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3271 unsigned int hctx_idx)
3273 struct blk_mq_tags *drv_tags;
3276 if (list_empty(&tags->page_list))
3279 if (blk_mq_is_shared_tags(set->flags))
3280 drv_tags = set->shared_tags;
3282 drv_tags = set->tags[hctx_idx];
3284 if (tags->static_rqs && set->ops->exit_request) {
3287 for (i = 0; i < tags->nr_tags; i++) {
3288 struct request *rq = tags->static_rqs[i];
3292 set->ops->exit_request(set, rq, hctx_idx);
3293 tags->static_rqs[i] = NULL;
3297 blk_mq_clear_rq_mapping(drv_tags, tags);
3299 while (!list_empty(&tags->page_list)) {
3300 page = list_first_entry(&tags->page_list, struct page, lru);
3301 list_del_init(&page->lru);
3303 * Remove kmemleak object previously allocated in
3304 * blk_mq_alloc_rqs().
3306 kmemleak_free(page_address(page));
3307 __free_pages(page, page->private);
3311 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3315 kfree(tags->static_rqs);
3316 tags->static_rqs = NULL;
3318 blk_mq_free_tags(tags);
3321 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3322 unsigned int hctx_idx)
3326 for (i = 0; i < set->nr_maps; i++) {
3327 unsigned int start = set->map[i].queue_offset;
3328 unsigned int end = start + set->map[i].nr_queues;
3330 if (hctx_idx >= start && hctx_idx < end)
3334 if (i >= set->nr_maps)
3335 i = HCTX_TYPE_DEFAULT;
3340 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3341 unsigned int hctx_idx)
3343 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3345 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3348 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3349 unsigned int hctx_idx,
3350 unsigned int nr_tags,
3351 unsigned int reserved_tags)
3353 int node = blk_mq_get_hctx_node(set, hctx_idx);
3354 struct blk_mq_tags *tags;
3356 if (node == NUMA_NO_NODE)
3357 node = set->numa_node;
3359 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3360 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3364 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3365 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3368 blk_mq_free_tags(tags);
3372 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3373 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3375 if (!tags->static_rqs) {
3377 blk_mq_free_tags(tags);
3384 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3385 unsigned int hctx_idx, int node)
3389 if (set->ops->init_request) {
3390 ret = set->ops->init_request(set, rq, hctx_idx, node);
3395 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3399 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3400 struct blk_mq_tags *tags,
3401 unsigned int hctx_idx, unsigned int depth)
3403 unsigned int i, j, entries_per_page, max_order = 4;
3404 int node = blk_mq_get_hctx_node(set, hctx_idx);
3405 size_t rq_size, left;
3407 if (node == NUMA_NO_NODE)
3408 node = set->numa_node;
3410 INIT_LIST_HEAD(&tags->page_list);
3413 * rq_size is the size of the request plus driver payload, rounded
3414 * to the cacheline size
3416 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3418 left = rq_size * depth;
3420 for (i = 0; i < depth; ) {
3421 int this_order = max_order;
3426 while (this_order && left < order_to_size(this_order - 1))
3430 page = alloc_pages_node(node,
3431 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3437 if (order_to_size(this_order) < rq_size)
3444 page->private = this_order;
3445 list_add_tail(&page->lru, &tags->page_list);
3447 p = page_address(page);
3449 * Allow kmemleak to scan these pages as they contain pointers
3450 * to additional allocations like via ops->init_request().
3452 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3453 entries_per_page = order_to_size(this_order) / rq_size;
3454 to_do = min(entries_per_page, depth - i);
3455 left -= to_do * rq_size;
3456 for (j = 0; j < to_do; j++) {
3457 struct request *rq = p;
3459 tags->static_rqs[i] = rq;
3460 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3461 tags->static_rqs[i] = NULL;
3472 blk_mq_free_rqs(set, tags, hctx_idx);
3476 struct rq_iter_data {
3477 struct blk_mq_hw_ctx *hctx;
3481 static bool blk_mq_has_request(struct request *rq, void *data)
3483 struct rq_iter_data *iter_data = data;
3485 if (rq->mq_hctx != iter_data->hctx)
3487 iter_data->has_rq = true;
3491 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3493 struct blk_mq_tags *tags = hctx->sched_tags ?
3494 hctx->sched_tags : hctx->tags;
3495 struct rq_iter_data data = {
3499 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3503 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3504 struct blk_mq_hw_ctx *hctx)
3506 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3508 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3513 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3515 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3516 struct blk_mq_hw_ctx, cpuhp_online);
3518 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3519 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3523 * Prevent new request from being allocated on the current hctx.
3525 * The smp_mb__after_atomic() Pairs with the implied barrier in
3526 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3527 * seen once we return from the tag allocator.
3529 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3530 smp_mb__after_atomic();
3533 * Try to grab a reference to the queue and wait for any outstanding
3534 * requests. If we could not grab a reference the queue has been
3535 * frozen and there are no requests.
3537 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3538 while (blk_mq_hctx_has_requests(hctx))
3540 percpu_ref_put(&hctx->queue->q_usage_counter);
3546 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3548 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3549 struct blk_mq_hw_ctx, cpuhp_online);
3551 if (cpumask_test_cpu(cpu, hctx->cpumask))
3552 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3557 * 'cpu' is going away. splice any existing rq_list entries from this
3558 * software queue to the hw queue dispatch list, and ensure that it
3561 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3563 struct blk_mq_hw_ctx *hctx;
3564 struct blk_mq_ctx *ctx;
3566 enum hctx_type type;
3568 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3569 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3572 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3575 spin_lock(&ctx->lock);
3576 if (!list_empty(&ctx->rq_lists[type])) {
3577 list_splice_init(&ctx->rq_lists[type], &tmp);
3578 blk_mq_hctx_clear_pending(hctx, ctx);
3580 spin_unlock(&ctx->lock);
3582 if (list_empty(&tmp))
3585 spin_lock(&hctx->lock);
3586 list_splice_tail_init(&tmp, &hctx->dispatch);
3587 spin_unlock(&hctx->lock);
3589 blk_mq_run_hw_queue(hctx, true);
3593 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3595 if (!(hctx->flags & BLK_MQ_F_STACKING))
3596 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3597 &hctx->cpuhp_online);
3598 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3603 * Before freeing hw queue, clearing the flush request reference in
3604 * tags->rqs[] for avoiding potential UAF.
3606 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3607 unsigned int queue_depth, struct request *flush_rq)
3610 unsigned long flags;
3612 /* The hw queue may not be mapped yet */
3616 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3618 for (i = 0; i < queue_depth; i++)
3619 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3622 * Wait until all pending iteration is done.
3624 * Request reference is cleared and it is guaranteed to be observed
3625 * after the ->lock is released.
3627 spin_lock_irqsave(&tags->lock, flags);
3628 spin_unlock_irqrestore(&tags->lock, flags);
3631 /* hctx->ctxs will be freed in queue's release handler */
3632 static void blk_mq_exit_hctx(struct request_queue *q,
3633 struct blk_mq_tag_set *set,
3634 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3636 struct request *flush_rq = hctx->fq->flush_rq;
3638 if (blk_mq_hw_queue_mapped(hctx))
3639 blk_mq_tag_idle(hctx);
3641 if (blk_queue_init_done(q))
3642 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3643 set->queue_depth, flush_rq);
3644 if (set->ops->exit_request)
3645 set->ops->exit_request(set, flush_rq, hctx_idx);
3647 if (set->ops->exit_hctx)
3648 set->ops->exit_hctx(hctx, hctx_idx);
3650 blk_mq_remove_cpuhp(hctx);
3652 xa_erase(&q->hctx_table, hctx_idx);
3654 spin_lock(&q->unused_hctx_lock);
3655 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3656 spin_unlock(&q->unused_hctx_lock);
3659 static void blk_mq_exit_hw_queues(struct request_queue *q,
3660 struct blk_mq_tag_set *set, int nr_queue)
3662 struct blk_mq_hw_ctx *hctx;
3665 queue_for_each_hw_ctx(q, hctx, i) {
3668 blk_mq_exit_hctx(q, set, hctx, i);
3672 static int blk_mq_init_hctx(struct request_queue *q,
3673 struct blk_mq_tag_set *set,
3674 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3676 hctx->queue_num = hctx_idx;
3678 if (!(hctx->flags & BLK_MQ_F_STACKING))
3679 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3680 &hctx->cpuhp_online);
3681 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3683 hctx->tags = set->tags[hctx_idx];
3685 if (set->ops->init_hctx &&
3686 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3687 goto unregister_cpu_notifier;
3689 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3693 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3699 if (set->ops->exit_request)
3700 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3702 if (set->ops->exit_hctx)
3703 set->ops->exit_hctx(hctx, hctx_idx);
3704 unregister_cpu_notifier:
3705 blk_mq_remove_cpuhp(hctx);
3709 static struct blk_mq_hw_ctx *
3710 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3713 struct blk_mq_hw_ctx *hctx;
3714 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3716 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3718 goto fail_alloc_hctx;
3720 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3723 atomic_set(&hctx->nr_active, 0);
3724 if (node == NUMA_NO_NODE)
3725 node = set->numa_node;
3726 hctx->numa_node = node;
3728 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3729 spin_lock_init(&hctx->lock);
3730 INIT_LIST_HEAD(&hctx->dispatch);
3732 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3734 INIT_LIST_HEAD(&hctx->hctx_list);
3737 * Allocate space for all possible cpus to avoid allocation at
3740 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3745 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3746 gfp, node, false, false))
3750 spin_lock_init(&hctx->dispatch_wait_lock);
3751 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3752 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3754 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3758 blk_mq_hctx_kobj_init(hctx);
3763 sbitmap_free(&hctx->ctx_map);
3767 free_cpumask_var(hctx->cpumask);
3774 static void blk_mq_init_cpu_queues(struct request_queue *q,
3775 unsigned int nr_hw_queues)
3777 struct blk_mq_tag_set *set = q->tag_set;
3780 for_each_possible_cpu(i) {
3781 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3782 struct blk_mq_hw_ctx *hctx;
3786 spin_lock_init(&__ctx->lock);
3787 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3788 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3793 * Set local node, IFF we have more than one hw queue. If
3794 * not, we remain on the home node of the device
3796 for (j = 0; j < set->nr_maps; j++) {
3797 hctx = blk_mq_map_queue_type(q, j, i);
3798 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3799 hctx->numa_node = cpu_to_node(i);
3804 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3805 unsigned int hctx_idx,
3808 struct blk_mq_tags *tags;
3811 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3815 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3817 blk_mq_free_rq_map(tags);
3824 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3827 if (blk_mq_is_shared_tags(set->flags)) {
3828 set->tags[hctx_idx] = set->shared_tags;
3833 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3836 return set->tags[hctx_idx];
3839 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3840 struct blk_mq_tags *tags,
3841 unsigned int hctx_idx)
3844 blk_mq_free_rqs(set, tags, hctx_idx);
3845 blk_mq_free_rq_map(tags);
3849 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3850 unsigned int hctx_idx)
3852 if (!blk_mq_is_shared_tags(set->flags))
3853 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3855 set->tags[hctx_idx] = NULL;
3858 static void blk_mq_map_swqueue(struct request_queue *q)
3860 unsigned int j, hctx_idx;
3862 struct blk_mq_hw_ctx *hctx;
3863 struct blk_mq_ctx *ctx;
3864 struct blk_mq_tag_set *set = q->tag_set;
3866 queue_for_each_hw_ctx(q, hctx, i) {
3867 cpumask_clear(hctx->cpumask);
3869 hctx->dispatch_from = NULL;
3873 * Map software to hardware queues.
3875 * If the cpu isn't present, the cpu is mapped to first hctx.
3877 for_each_possible_cpu(i) {
3879 ctx = per_cpu_ptr(q->queue_ctx, i);
3880 for (j = 0; j < set->nr_maps; j++) {
3881 if (!set->map[j].nr_queues) {
3882 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3883 HCTX_TYPE_DEFAULT, i);
3886 hctx_idx = set->map[j].mq_map[i];
3887 /* unmapped hw queue can be remapped after CPU topo changed */
3888 if (!set->tags[hctx_idx] &&
3889 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3891 * If tags initialization fail for some hctx,
3892 * that hctx won't be brought online. In this
3893 * case, remap the current ctx to hctx[0] which
3894 * is guaranteed to always have tags allocated
3896 set->map[j].mq_map[i] = 0;
3899 hctx = blk_mq_map_queue_type(q, j, i);
3900 ctx->hctxs[j] = hctx;
3902 * If the CPU is already set in the mask, then we've
3903 * mapped this one already. This can happen if
3904 * devices share queues across queue maps.
3906 if (cpumask_test_cpu(i, hctx->cpumask))
3909 cpumask_set_cpu(i, hctx->cpumask);
3911 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3912 hctx->ctxs[hctx->nr_ctx++] = ctx;
3915 * If the nr_ctx type overflows, we have exceeded the
3916 * amount of sw queues we can support.
3918 BUG_ON(!hctx->nr_ctx);
3921 for (; j < HCTX_MAX_TYPES; j++)
3922 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3923 HCTX_TYPE_DEFAULT, i);
3926 queue_for_each_hw_ctx(q, hctx, i) {
3928 * If no software queues are mapped to this hardware queue,
3929 * disable it and free the request entries.
3931 if (!hctx->nr_ctx) {
3932 /* Never unmap queue 0. We need it as a
3933 * fallback in case of a new remap fails
3937 __blk_mq_free_map_and_rqs(set, i);
3943 hctx->tags = set->tags[i];
3944 WARN_ON(!hctx->tags);
3947 * Set the map size to the number of mapped software queues.
3948 * This is more accurate and more efficient than looping
3949 * over all possibly mapped software queues.
3951 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3954 * Initialize batch roundrobin counts
3956 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3957 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3962 * Caller needs to ensure that we're either frozen/quiesced, or that
3963 * the queue isn't live yet.
3965 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3967 struct blk_mq_hw_ctx *hctx;
3970 queue_for_each_hw_ctx(q, hctx, i) {
3972 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3974 blk_mq_tag_idle(hctx);
3975 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3980 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3983 struct request_queue *q;
3985 lockdep_assert_held(&set->tag_list_lock);
3987 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3988 blk_mq_freeze_queue(q);
3989 queue_set_hctx_shared(q, shared);
3990 blk_mq_unfreeze_queue(q);
3994 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3996 struct blk_mq_tag_set *set = q->tag_set;
3998 mutex_lock(&set->tag_list_lock);
3999 list_del(&q->tag_set_list);
4000 if (list_is_singular(&set->tag_list)) {
4001 /* just transitioned to unshared */
4002 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
4003 /* update existing queue */
4004 blk_mq_update_tag_set_shared(set, false);
4006 mutex_unlock(&set->tag_list_lock);
4007 INIT_LIST_HEAD(&q->tag_set_list);
4010 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
4011 struct request_queue *q)
4013 mutex_lock(&set->tag_list_lock);
4016 * Check to see if we're transitioning to shared (from 1 to 2 queues).
4018 if (!list_empty(&set->tag_list) &&
4019 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
4020 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
4021 /* update existing queue */
4022 blk_mq_update_tag_set_shared(set, true);
4024 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
4025 queue_set_hctx_shared(q, true);
4026 list_add_tail(&q->tag_set_list, &set->tag_list);
4028 mutex_unlock(&set->tag_list_lock);
4031 /* All allocations will be freed in release handler of q->mq_kobj */
4032 static int blk_mq_alloc_ctxs(struct request_queue *q)
4034 struct blk_mq_ctxs *ctxs;
4037 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
4041 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
4042 if (!ctxs->queue_ctx)
4045 for_each_possible_cpu(cpu) {
4046 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4050 q->mq_kobj = &ctxs->kobj;
4051 q->queue_ctx = ctxs->queue_ctx;
4060 * It is the actual release handler for mq, but we do it from
4061 * request queue's release handler for avoiding use-after-free
4062 * and headache because q->mq_kobj shouldn't have been introduced,
4063 * but we can't group ctx/kctx kobj without it.
4065 void blk_mq_release(struct request_queue *q)
4067 struct blk_mq_hw_ctx *hctx, *next;
4070 queue_for_each_hw_ctx(q, hctx, i)
4071 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4073 /* all hctx are in .unused_hctx_list now */
4074 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4075 list_del_init(&hctx->hctx_list);
4076 kobject_put(&hctx->kobj);
4079 xa_destroy(&q->hctx_table);
4082 * release .mq_kobj and sw queue's kobject now because
4083 * both share lifetime with request queue.
4085 blk_mq_sysfs_deinit(q);
4088 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4091 struct request_queue *q;
4094 q = blk_alloc_queue(set->numa_node, set->flags & BLK_MQ_F_BLOCKING);
4096 return ERR_PTR(-ENOMEM);
4097 q->queuedata = queuedata;
4098 ret = blk_mq_init_allocated_queue(set, q);
4101 return ERR_PTR(ret);
4106 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4108 return blk_mq_init_queue_data(set, NULL);
4110 EXPORT_SYMBOL(blk_mq_init_queue);
4113 * blk_mq_destroy_queue - shutdown a request queue
4114 * @q: request queue to shutdown
4116 * This shuts down a request queue allocated by blk_mq_init_queue() and drops
4117 * the initial reference. All future requests will failed with -ENODEV.
4119 * Context: can sleep
4121 void blk_mq_destroy_queue(struct request_queue *q)
4123 WARN_ON_ONCE(!queue_is_mq(q));
4124 WARN_ON_ONCE(blk_queue_registered(q));
4128 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4129 blk_queue_start_drain(q);
4130 blk_freeze_queue(q);
4133 blk_mq_cancel_work_sync(q);
4134 blk_mq_exit_queue(q);
4136 /* @q is and will stay empty, shutdown and put */
4139 EXPORT_SYMBOL(blk_mq_destroy_queue);
4141 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4142 struct lock_class_key *lkclass)
4144 struct request_queue *q;
4145 struct gendisk *disk;
4147 q = blk_mq_init_queue_data(set, queuedata);
4151 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4153 blk_mq_destroy_queue(q);
4154 return ERR_PTR(-ENOMEM);
4156 set_bit(GD_OWNS_QUEUE, &disk->state);
4159 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4161 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4162 struct lock_class_key *lkclass)
4164 struct gendisk *disk;
4166 if (!blk_get_queue(q))
4168 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4173 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4175 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4176 struct blk_mq_tag_set *set, struct request_queue *q,
4177 int hctx_idx, int node)
4179 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4181 /* reuse dead hctx first */
4182 spin_lock(&q->unused_hctx_lock);
4183 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4184 if (tmp->numa_node == node) {
4190 list_del_init(&hctx->hctx_list);
4191 spin_unlock(&q->unused_hctx_lock);
4194 hctx = blk_mq_alloc_hctx(q, set, node);
4198 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4204 kobject_put(&hctx->kobj);
4209 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4210 struct request_queue *q)
4212 struct blk_mq_hw_ctx *hctx;
4215 /* protect against switching io scheduler */
4216 mutex_lock(&q->sysfs_lock);
4217 for (i = 0; i < set->nr_hw_queues; i++) {
4219 int node = blk_mq_get_hctx_node(set, i);
4220 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4223 old_node = old_hctx->numa_node;
4224 blk_mq_exit_hctx(q, set, old_hctx, i);
4227 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4230 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4232 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4233 WARN_ON_ONCE(!hctx);
4237 * Increasing nr_hw_queues fails. Free the newly allocated
4238 * hctxs and keep the previous q->nr_hw_queues.
4240 if (i != set->nr_hw_queues) {
4241 j = q->nr_hw_queues;
4244 q->nr_hw_queues = set->nr_hw_queues;
4247 xa_for_each_start(&q->hctx_table, j, hctx, j)
4248 blk_mq_exit_hctx(q, set, hctx, j);
4249 mutex_unlock(&q->sysfs_lock);
4252 static void blk_mq_update_poll_flag(struct request_queue *q)
4254 struct blk_mq_tag_set *set = q->tag_set;
4256 if (set->nr_maps > HCTX_TYPE_POLL &&
4257 set->map[HCTX_TYPE_POLL].nr_queues)
4258 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4260 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4263 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4264 struct request_queue *q)
4266 WARN_ON_ONCE(blk_queue_has_srcu(q) !=
4267 !!(set->flags & BLK_MQ_F_BLOCKING));
4269 /* mark the queue as mq asap */
4270 q->mq_ops = set->ops;
4272 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
4273 blk_mq_poll_stats_bkt,
4274 BLK_MQ_POLL_STATS_BKTS, q);
4278 if (blk_mq_alloc_ctxs(q))
4281 /* init q->mq_kobj and sw queues' kobjects */
4282 blk_mq_sysfs_init(q);
4284 INIT_LIST_HEAD(&q->unused_hctx_list);
4285 spin_lock_init(&q->unused_hctx_lock);
4287 xa_init(&q->hctx_table);
4289 blk_mq_realloc_hw_ctxs(set, q);
4290 if (!q->nr_hw_queues)
4293 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4294 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4298 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4299 blk_mq_update_poll_flag(q);
4301 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4302 INIT_LIST_HEAD(&q->requeue_list);
4303 spin_lock_init(&q->requeue_lock);
4305 q->nr_requests = set->queue_depth;
4308 * Default to classic polling
4310 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
4312 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4313 blk_mq_add_queue_tag_set(set, q);
4314 blk_mq_map_swqueue(q);
4320 blk_stat_free_callback(q->poll_cb);
4326 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4328 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4329 void blk_mq_exit_queue(struct request_queue *q)
4331 struct blk_mq_tag_set *set = q->tag_set;
4333 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4334 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4335 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4336 blk_mq_del_queue_tag_set(q);
4339 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4343 if (blk_mq_is_shared_tags(set->flags)) {
4344 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4347 if (!set->shared_tags)
4351 for (i = 0; i < set->nr_hw_queues; i++) {
4352 if (!__blk_mq_alloc_map_and_rqs(set, i))
4361 __blk_mq_free_map_and_rqs(set, i);
4363 if (blk_mq_is_shared_tags(set->flags)) {
4364 blk_mq_free_map_and_rqs(set, set->shared_tags,
4365 BLK_MQ_NO_HCTX_IDX);
4372 * Allocate the request maps associated with this tag_set. Note that this
4373 * may reduce the depth asked for, if memory is tight. set->queue_depth
4374 * will be updated to reflect the allocated depth.
4376 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4381 depth = set->queue_depth;
4383 err = __blk_mq_alloc_rq_maps(set);
4387 set->queue_depth >>= 1;
4388 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4392 } while (set->queue_depth);
4394 if (!set->queue_depth || err) {
4395 pr_err("blk-mq: failed to allocate request map\n");
4399 if (depth != set->queue_depth)
4400 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4401 depth, set->queue_depth);
4406 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4409 * blk_mq_map_queues() and multiple .map_queues() implementations
4410 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4411 * number of hardware queues.
4413 if (set->nr_maps == 1)
4414 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4416 if (set->ops->map_queues && !is_kdump_kernel()) {
4420 * transport .map_queues is usually done in the following
4423 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4424 * mask = get_cpu_mask(queue)
4425 * for_each_cpu(cpu, mask)
4426 * set->map[x].mq_map[cpu] = queue;
4429 * When we need to remap, the table has to be cleared for
4430 * killing stale mapping since one CPU may not be mapped
4433 for (i = 0; i < set->nr_maps; i++)
4434 blk_mq_clear_mq_map(&set->map[i]);
4436 set->ops->map_queues(set);
4438 BUG_ON(set->nr_maps > 1);
4439 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4443 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4444 int cur_nr_hw_queues, int new_nr_hw_queues)
4446 struct blk_mq_tags **new_tags;
4448 if (cur_nr_hw_queues >= new_nr_hw_queues)
4451 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4452 GFP_KERNEL, set->numa_node);
4457 memcpy(new_tags, set->tags, cur_nr_hw_queues *
4458 sizeof(*set->tags));
4460 set->tags = new_tags;
4461 set->nr_hw_queues = new_nr_hw_queues;
4466 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
4467 int new_nr_hw_queues)
4469 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
4473 * Alloc a tag set to be associated with one or more request queues.
4474 * May fail with EINVAL for various error conditions. May adjust the
4475 * requested depth down, if it's too large. In that case, the set
4476 * value will be stored in set->queue_depth.
4478 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4482 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4484 if (!set->nr_hw_queues)
4486 if (!set->queue_depth)
4488 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4491 if (!set->ops->queue_rq)
4494 if (!set->ops->get_budget ^ !set->ops->put_budget)
4497 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4498 pr_info("blk-mq: reduced tag depth to %u\n",
4500 set->queue_depth = BLK_MQ_MAX_DEPTH;
4505 else if (set->nr_maps > HCTX_MAX_TYPES)
4509 * If a crashdump is active, then we are potentially in a very
4510 * memory constrained environment. Limit us to 1 queue and
4511 * 64 tags to prevent using too much memory.
4513 if (is_kdump_kernel()) {
4514 set->nr_hw_queues = 1;
4516 set->queue_depth = min(64U, set->queue_depth);
4519 * There is no use for more h/w queues than cpus if we just have
4522 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4523 set->nr_hw_queues = nr_cpu_ids;
4525 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
4529 for (i = 0; i < set->nr_maps; i++) {
4530 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4531 sizeof(set->map[i].mq_map[0]),
4532 GFP_KERNEL, set->numa_node);
4533 if (!set->map[i].mq_map)
4534 goto out_free_mq_map;
4535 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4538 blk_mq_update_queue_map(set);
4540 ret = blk_mq_alloc_set_map_and_rqs(set);
4542 goto out_free_mq_map;
4544 mutex_init(&set->tag_list_lock);
4545 INIT_LIST_HEAD(&set->tag_list);
4550 for (i = 0; i < set->nr_maps; i++) {
4551 kfree(set->map[i].mq_map);
4552 set->map[i].mq_map = NULL;
4558 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4560 /* allocate and initialize a tagset for a simple single-queue device */
4561 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4562 const struct blk_mq_ops *ops, unsigned int queue_depth,
4563 unsigned int set_flags)
4565 memset(set, 0, sizeof(*set));
4567 set->nr_hw_queues = 1;
4569 set->queue_depth = queue_depth;
4570 set->numa_node = NUMA_NO_NODE;
4571 set->flags = set_flags;
4572 return blk_mq_alloc_tag_set(set);
4574 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4576 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4580 for (i = 0; i < set->nr_hw_queues; i++)
4581 __blk_mq_free_map_and_rqs(set, i);
4583 if (blk_mq_is_shared_tags(set->flags)) {
4584 blk_mq_free_map_and_rqs(set, set->shared_tags,
4585 BLK_MQ_NO_HCTX_IDX);
4588 for (j = 0; j < set->nr_maps; j++) {
4589 kfree(set->map[j].mq_map);
4590 set->map[j].mq_map = NULL;
4596 EXPORT_SYMBOL(blk_mq_free_tag_set);
4598 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4600 struct blk_mq_tag_set *set = q->tag_set;
4601 struct blk_mq_hw_ctx *hctx;
4608 if (q->nr_requests == nr)
4611 blk_mq_freeze_queue(q);
4612 blk_mq_quiesce_queue(q);
4615 queue_for_each_hw_ctx(q, hctx, i) {
4619 * If we're using an MQ scheduler, just update the scheduler
4620 * queue depth. This is similar to what the old code would do.
4622 if (hctx->sched_tags) {
4623 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4626 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4631 if (q->elevator && q->elevator->type->ops.depth_updated)
4632 q->elevator->type->ops.depth_updated(hctx);
4635 q->nr_requests = nr;
4636 if (blk_mq_is_shared_tags(set->flags)) {
4638 blk_mq_tag_update_sched_shared_tags(q);
4640 blk_mq_tag_resize_shared_tags(set, nr);
4644 blk_mq_unquiesce_queue(q);
4645 blk_mq_unfreeze_queue(q);
4651 * request_queue and elevator_type pair.
4652 * It is just used by __blk_mq_update_nr_hw_queues to cache
4653 * the elevator_type associated with a request_queue.
4655 struct blk_mq_qe_pair {
4656 struct list_head node;
4657 struct request_queue *q;
4658 struct elevator_type *type;
4662 * Cache the elevator_type in qe pair list and switch the
4663 * io scheduler to 'none'
4665 static bool blk_mq_elv_switch_none(struct list_head *head,
4666 struct request_queue *q)
4668 struct blk_mq_qe_pair *qe;
4673 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4677 /* q->elevator needs protection from ->sysfs_lock */
4678 mutex_lock(&q->sysfs_lock);
4680 INIT_LIST_HEAD(&qe->node);
4682 qe->type = q->elevator->type;
4683 list_add(&qe->node, head);
4686 * After elevator_switch, the previous elevator_queue will be
4687 * released by elevator_release. The reference of the io scheduler
4688 * module get by elevator_get will also be put. So we need to get
4689 * a reference of the io scheduler module here to prevent it to be
4692 __module_get(qe->type->elevator_owner);
4693 elevator_switch(q, NULL);
4694 mutex_unlock(&q->sysfs_lock);
4699 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4700 struct request_queue *q)
4702 struct blk_mq_qe_pair *qe;
4704 list_for_each_entry(qe, head, node)
4711 static void blk_mq_elv_switch_back(struct list_head *head,
4712 struct request_queue *q)
4714 struct blk_mq_qe_pair *qe;
4715 struct elevator_type *t;
4717 qe = blk_lookup_qe_pair(head, q);
4721 list_del(&qe->node);
4724 mutex_lock(&q->sysfs_lock);
4725 elevator_switch(q, t);
4726 mutex_unlock(&q->sysfs_lock);
4729 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4732 struct request_queue *q;
4734 int prev_nr_hw_queues;
4736 lockdep_assert_held(&set->tag_list_lock);
4738 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4739 nr_hw_queues = nr_cpu_ids;
4740 if (nr_hw_queues < 1)
4742 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4745 list_for_each_entry(q, &set->tag_list, tag_set_list)
4746 blk_mq_freeze_queue(q);
4748 * Switch IO scheduler to 'none', cleaning up the data associated
4749 * with the previous scheduler. We will switch back once we are done
4750 * updating the new sw to hw queue mappings.
4752 list_for_each_entry(q, &set->tag_list, tag_set_list)
4753 if (!blk_mq_elv_switch_none(&head, q))
4756 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4757 blk_mq_debugfs_unregister_hctxs(q);
4758 blk_mq_sysfs_unregister_hctxs(q);
4761 prev_nr_hw_queues = set->nr_hw_queues;
4762 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
4766 set->nr_hw_queues = nr_hw_queues;
4768 blk_mq_update_queue_map(set);
4769 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4770 blk_mq_realloc_hw_ctxs(set, q);
4771 blk_mq_update_poll_flag(q);
4772 if (q->nr_hw_queues != set->nr_hw_queues) {
4773 int i = prev_nr_hw_queues;
4775 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4776 nr_hw_queues, prev_nr_hw_queues);
4777 for (; i < set->nr_hw_queues; i++)
4778 __blk_mq_free_map_and_rqs(set, i);
4780 set->nr_hw_queues = prev_nr_hw_queues;
4781 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4784 blk_mq_map_swqueue(q);
4788 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4789 blk_mq_sysfs_register_hctxs(q);
4790 blk_mq_debugfs_register_hctxs(q);
4794 list_for_each_entry(q, &set->tag_list, tag_set_list)
4795 blk_mq_elv_switch_back(&head, q);
4797 list_for_each_entry(q, &set->tag_list, tag_set_list)
4798 blk_mq_unfreeze_queue(q);
4801 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4803 mutex_lock(&set->tag_list_lock);
4804 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4805 mutex_unlock(&set->tag_list_lock);
4807 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4809 /* Enable polling stats and return whether they were already enabled. */
4810 static bool blk_poll_stats_enable(struct request_queue *q)
4815 return blk_stats_alloc_enable(q);
4818 static void blk_mq_poll_stats_start(struct request_queue *q)
4821 * We don't arm the callback if polling stats are not enabled or the
4822 * callback is already active.
4824 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
4827 blk_stat_activate_msecs(q->poll_cb, 100);
4830 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4832 struct request_queue *q = cb->data;
4835 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4836 if (cb->stat[bucket].nr_samples)
4837 q->poll_stat[bucket] = cb->stat[bucket];
4841 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4844 unsigned long ret = 0;
4848 * If stats collection isn't on, don't sleep but turn it on for
4851 if (!blk_poll_stats_enable(q))
4855 * As an optimistic guess, use half of the mean service time
4856 * for this type of request. We can (and should) make this smarter.
4857 * For instance, if the completion latencies are tight, we can
4858 * get closer than just half the mean. This is especially
4859 * important on devices where the completion latencies are longer
4860 * than ~10 usec. We do use the stats for the relevant IO size
4861 * if available which does lead to better estimates.
4863 bucket = blk_mq_poll_stats_bkt(rq);
4867 if (q->poll_stat[bucket].nr_samples)
4868 ret = (q->poll_stat[bucket].mean + 1) / 2;
4873 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4875 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4876 struct request *rq = blk_qc_to_rq(hctx, qc);
4877 struct hrtimer_sleeper hs;
4878 enum hrtimer_mode mode;
4883 * If a request has completed on queue that uses an I/O scheduler, we
4884 * won't get back a request from blk_qc_to_rq.
4886 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4890 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4892 * 0: use half of prev avg
4893 * >0: use this specific value
4895 if (q->poll_nsec > 0)
4896 nsecs = q->poll_nsec;
4898 nsecs = blk_mq_poll_nsecs(q, rq);
4903 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4906 * This will be replaced with the stats tracking code, using
4907 * 'avg_completion_time / 2' as the pre-sleep target.
4911 mode = HRTIMER_MODE_REL;
4912 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4913 hrtimer_set_expires(&hs.timer, kt);
4916 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4918 set_current_state(TASK_UNINTERRUPTIBLE);
4919 hrtimer_sleeper_start_expires(&hs, mode);
4922 hrtimer_cancel(&hs.timer);
4923 mode = HRTIMER_MODE_ABS;
4924 } while (hs.task && !signal_pending(current));
4926 __set_current_state(TASK_RUNNING);
4927 destroy_hrtimer_on_stack(&hs.timer);
4930 * If we sleep, have the caller restart the poll loop to reset the
4931 * state. Like for the other success return cases, the caller is
4932 * responsible for checking if the IO completed. If the IO isn't
4933 * complete, we'll get called again and will go straight to the busy
4939 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4940 struct io_comp_batch *iob, unsigned int flags)
4942 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4943 long state = get_current_state();
4947 ret = q->mq_ops->poll(hctx, iob);
4949 __set_current_state(TASK_RUNNING);
4953 if (signal_pending_state(state, current))
4954 __set_current_state(TASK_RUNNING);
4955 if (task_is_running(current))
4958 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4961 } while (!need_resched());
4963 __set_current_state(TASK_RUNNING);
4967 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4970 if (!(flags & BLK_POLL_NOSLEEP) &&
4971 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4972 if (blk_mq_poll_hybrid(q, cookie))
4975 return blk_mq_poll_classic(q, cookie, iob, flags);
4978 unsigned int blk_mq_rq_cpu(struct request *rq)
4980 return rq->mq_ctx->cpu;
4982 EXPORT_SYMBOL(blk_mq_rq_cpu);
4984 void blk_mq_cancel_work_sync(struct request_queue *q)
4986 if (queue_is_mq(q)) {
4987 struct blk_mq_hw_ctx *hctx;
4990 cancel_delayed_work_sync(&q->requeue_work);
4992 queue_for_each_hw_ctx(q, hctx, i)
4993 cancel_delayed_work_sync(&hctx->run_work);
4997 static int __init blk_mq_init(void)
5001 for_each_possible_cpu(i)
5002 init_llist_head(&per_cpu(blk_cpu_done, i));
5003 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
5005 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
5006 "block/softirq:dead", NULL,
5007 blk_softirq_cpu_dead);
5008 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
5009 blk_mq_hctx_notify_dead);
5010 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
5011 blk_mq_hctx_notify_online,
5012 blk_mq_hctx_notify_offline);
5015 subsys_initcall(blk_mq_init);