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
257 * @set: tag_set to wait on
259 * Note: it is driver's responsibility for making sure that quiesce has
260 * been started on or more of the request_queues of the tag_set. This
261 * function only waits for the quiesce on those request_queues that had
262 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
264 void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
266 if (set->flags & BLK_MQ_F_BLOCKING)
267 synchronize_srcu(set->srcu);
271 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
274 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
277 * Note: this function does not prevent that the struct request end_io()
278 * callback function is invoked. Once this function is returned, we make
279 * sure no dispatch can happen until the queue is unquiesced via
280 * blk_mq_unquiesce_queue().
282 void blk_mq_quiesce_queue(struct request_queue *q)
284 blk_mq_quiesce_queue_nowait(q);
285 /* nothing to wait for non-mq queues */
287 blk_mq_wait_quiesce_done(q->tag_set);
289 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
292 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
295 * This function recovers queue into the state before quiescing
296 * which is done by blk_mq_quiesce_queue.
298 void blk_mq_unquiesce_queue(struct request_queue *q)
301 bool run_queue = false;
303 spin_lock_irqsave(&q->queue_lock, flags);
304 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
306 } else if (!--q->quiesce_depth) {
307 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
310 spin_unlock_irqrestore(&q->queue_lock, flags);
312 /* dispatch requests which are inserted during quiescing */
314 blk_mq_run_hw_queues(q, true);
316 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
318 void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
320 struct request_queue *q;
322 mutex_lock(&set->tag_list_lock);
323 list_for_each_entry(q, &set->tag_list, tag_set_list) {
324 if (!blk_queue_skip_tagset_quiesce(q))
325 blk_mq_quiesce_queue_nowait(q);
327 blk_mq_wait_quiesce_done(set);
328 mutex_unlock(&set->tag_list_lock);
330 EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
332 void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
334 struct request_queue *q;
336 mutex_lock(&set->tag_list_lock);
337 list_for_each_entry(q, &set->tag_list, tag_set_list) {
338 if (!blk_queue_skip_tagset_quiesce(q))
339 blk_mq_unquiesce_queue(q);
341 mutex_unlock(&set->tag_list_lock);
343 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
345 void blk_mq_wake_waiters(struct request_queue *q)
347 struct blk_mq_hw_ctx *hctx;
350 queue_for_each_hw_ctx(q, hctx, i)
351 if (blk_mq_hw_queue_mapped(hctx))
352 blk_mq_tag_wakeup_all(hctx->tags, true);
355 void blk_rq_init(struct request_queue *q, struct request *rq)
357 memset(rq, 0, sizeof(*rq));
359 INIT_LIST_HEAD(&rq->queuelist);
361 rq->__sector = (sector_t) -1;
362 INIT_HLIST_NODE(&rq->hash);
363 RB_CLEAR_NODE(&rq->rb_node);
364 rq->tag = BLK_MQ_NO_TAG;
365 rq->internal_tag = BLK_MQ_NO_TAG;
366 rq->start_time_ns = ktime_get_ns();
368 blk_crypto_rq_set_defaults(rq);
370 EXPORT_SYMBOL(blk_rq_init);
372 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
373 struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns)
375 struct blk_mq_ctx *ctx = data->ctx;
376 struct blk_mq_hw_ctx *hctx = data->hctx;
377 struct request_queue *q = data->q;
378 struct request *rq = tags->static_rqs[tag];
383 rq->cmd_flags = data->cmd_flags;
385 if (data->flags & BLK_MQ_REQ_PM)
386 data->rq_flags |= RQF_PM;
387 if (blk_queue_io_stat(q))
388 data->rq_flags |= RQF_IO_STAT;
389 rq->rq_flags = data->rq_flags;
391 if (!(data->rq_flags & RQF_ELV)) {
393 rq->internal_tag = BLK_MQ_NO_TAG;
395 rq->tag = BLK_MQ_NO_TAG;
396 rq->internal_tag = tag;
400 if (blk_mq_need_time_stamp(rq))
401 rq->start_time_ns = ktime_get_ns();
403 rq->start_time_ns = 0;
405 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
406 rq->alloc_time_ns = alloc_time_ns;
408 rq->io_start_time_ns = 0;
409 rq->stats_sectors = 0;
410 rq->nr_phys_segments = 0;
411 #if defined(CONFIG_BLK_DEV_INTEGRITY)
412 rq->nr_integrity_segments = 0;
415 rq->end_io_data = NULL;
417 blk_crypto_rq_set_defaults(rq);
418 INIT_LIST_HEAD(&rq->queuelist);
419 /* tag was already set */
420 WRITE_ONCE(rq->deadline, 0);
423 if (rq->rq_flags & RQF_ELV) {
424 struct elevator_queue *e = data->q->elevator;
426 INIT_HLIST_NODE(&rq->hash);
427 RB_CLEAR_NODE(&rq->rb_node);
429 if (!op_is_flush(data->cmd_flags) &&
430 e->type->ops.prepare_request) {
431 e->type->ops.prepare_request(rq);
432 rq->rq_flags |= RQF_ELVPRIV;
439 static inline struct request *
440 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data,
443 unsigned int tag, tag_offset;
444 struct blk_mq_tags *tags;
446 unsigned long tag_mask;
449 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
450 if (unlikely(!tag_mask))
453 tags = blk_mq_tags_from_data(data);
454 for (i = 0; tag_mask; i++) {
455 if (!(tag_mask & (1UL << i)))
457 tag = tag_offset + i;
458 prefetch(tags->static_rqs[tag]);
459 tag_mask &= ~(1UL << i);
460 rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns);
461 rq_list_add(data->cached_rq, rq);
464 /* caller already holds a reference, add for remainder */
465 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
468 return rq_list_pop(data->cached_rq);
471 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
473 struct request_queue *q = data->q;
474 u64 alloc_time_ns = 0;
478 /* alloc_time includes depth and tag waits */
479 if (blk_queue_rq_alloc_time(q))
480 alloc_time_ns = ktime_get_ns();
482 if (data->cmd_flags & REQ_NOWAIT)
483 data->flags |= BLK_MQ_REQ_NOWAIT;
486 struct elevator_queue *e = q->elevator;
488 data->rq_flags |= RQF_ELV;
491 * Flush/passthrough requests are special and go directly to the
492 * dispatch list. Don't include reserved tags in the
493 * limiting, as it isn't useful.
495 if (!op_is_flush(data->cmd_flags) &&
496 !blk_op_is_passthrough(data->cmd_flags) &&
497 e->type->ops.limit_depth &&
498 !(data->flags & BLK_MQ_REQ_RESERVED))
499 e->type->ops.limit_depth(data->cmd_flags, data);
503 data->ctx = blk_mq_get_ctx(q);
504 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
505 if (!(data->rq_flags & RQF_ELV))
506 blk_mq_tag_busy(data->hctx);
508 if (data->flags & BLK_MQ_REQ_RESERVED)
509 data->rq_flags |= RQF_RESV;
512 * Try batched alloc if we want more than 1 tag.
514 if (data->nr_tags > 1) {
515 rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns);
522 * Waiting allocations only fail because of an inactive hctx. In that
523 * case just retry the hctx assignment and tag allocation as CPU hotplug
524 * should have migrated us to an online CPU by now.
526 tag = blk_mq_get_tag(data);
527 if (tag == BLK_MQ_NO_TAG) {
528 if (data->flags & BLK_MQ_REQ_NOWAIT)
531 * Give up the CPU and sleep for a random short time to
532 * ensure that thread using a realtime scheduling class
533 * are migrated off the CPU, and thus off the hctx that
540 return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag,
544 static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
545 struct blk_plug *plug,
547 blk_mq_req_flags_t flags)
549 struct blk_mq_alloc_data data = {
553 .nr_tags = plug->nr_ios,
554 .cached_rq = &plug->cached_rq,
558 if (blk_queue_enter(q, flags))
563 rq = __blk_mq_alloc_requests(&data);
569 static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
571 blk_mq_req_flags_t flags)
573 struct blk_plug *plug = current->plug;
579 if (rq_list_empty(plug->cached_rq)) {
580 if (plug->nr_ios == 1)
582 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
586 rq = rq_list_peek(&plug->cached_rq);
587 if (!rq || rq->q != q)
590 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
592 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
595 plug->cached_rq = rq_list_next(rq);
599 INIT_LIST_HEAD(&rq->queuelist);
603 struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
604 blk_mq_req_flags_t flags)
608 rq = blk_mq_alloc_cached_request(q, opf, flags);
610 struct blk_mq_alloc_data data = {
618 ret = blk_queue_enter(q, flags);
622 rq = __blk_mq_alloc_requests(&data);
627 rq->__sector = (sector_t) -1;
628 rq->bio = rq->biotail = NULL;
632 return ERR_PTR(-EWOULDBLOCK);
634 EXPORT_SYMBOL(blk_mq_alloc_request);
636 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
637 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
639 struct blk_mq_alloc_data data = {
645 u64 alloc_time_ns = 0;
651 /* alloc_time includes depth and tag waits */
652 if (blk_queue_rq_alloc_time(q))
653 alloc_time_ns = ktime_get_ns();
656 * If the tag allocator sleeps we could get an allocation for a
657 * different hardware context. No need to complicate the low level
658 * allocator for this for the rare use case of a command tied to
661 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
662 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
663 return ERR_PTR(-EINVAL);
665 if (hctx_idx >= q->nr_hw_queues)
666 return ERR_PTR(-EIO);
668 ret = blk_queue_enter(q, flags);
673 * Check if the hardware context is actually mapped to anything.
674 * If not tell the caller that it should skip this queue.
677 data.hctx = xa_load(&q->hctx_table, hctx_idx);
678 if (!blk_mq_hw_queue_mapped(data.hctx))
680 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
681 if (cpu >= nr_cpu_ids)
683 data.ctx = __blk_mq_get_ctx(q, cpu);
686 blk_mq_tag_busy(data.hctx);
688 data.rq_flags |= RQF_ELV;
690 if (flags & BLK_MQ_REQ_RESERVED)
691 data.rq_flags |= RQF_RESV;
694 tag = blk_mq_get_tag(&data);
695 if (tag == BLK_MQ_NO_TAG)
697 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag,
700 rq->__sector = (sector_t) -1;
701 rq->bio = rq->biotail = NULL;
708 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
710 static void __blk_mq_free_request(struct request *rq)
712 struct request_queue *q = rq->q;
713 struct blk_mq_ctx *ctx = rq->mq_ctx;
714 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
715 const int sched_tag = rq->internal_tag;
717 blk_crypto_free_request(rq);
718 blk_pm_mark_last_busy(rq);
720 if (rq->tag != BLK_MQ_NO_TAG)
721 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
722 if (sched_tag != BLK_MQ_NO_TAG)
723 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
724 blk_mq_sched_restart(hctx);
728 void blk_mq_free_request(struct request *rq)
730 struct request_queue *q = rq->q;
731 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
733 if ((rq->rq_flags & RQF_ELVPRIV) &&
734 q->elevator->type->ops.finish_request)
735 q->elevator->type->ops.finish_request(rq);
737 if (rq->rq_flags & RQF_MQ_INFLIGHT)
738 __blk_mq_dec_active_requests(hctx);
740 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
741 laptop_io_completion(q->disk->bdi);
745 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
746 if (req_ref_put_and_test(rq))
747 __blk_mq_free_request(rq);
749 EXPORT_SYMBOL_GPL(blk_mq_free_request);
751 void blk_mq_free_plug_rqs(struct blk_plug *plug)
755 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
756 blk_mq_free_request(rq);
759 void blk_dump_rq_flags(struct request *rq, char *msg)
761 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
762 rq->q->disk ? rq->q->disk->disk_name : "?",
763 (__force unsigned long long) rq->cmd_flags);
765 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
766 (unsigned long long)blk_rq_pos(rq),
767 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
768 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
769 rq->bio, rq->biotail, blk_rq_bytes(rq));
771 EXPORT_SYMBOL(blk_dump_rq_flags);
773 static void req_bio_endio(struct request *rq, struct bio *bio,
774 unsigned int nbytes, blk_status_t error)
776 if (unlikely(error)) {
777 bio->bi_status = error;
778 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
780 * Partial zone append completions cannot be supported as the
781 * BIO fragments may end up not being written sequentially.
783 if (bio->bi_iter.bi_size != nbytes)
784 bio->bi_status = BLK_STS_IOERR;
786 bio->bi_iter.bi_sector = rq->__sector;
789 bio_advance(bio, nbytes);
791 if (unlikely(rq->rq_flags & RQF_QUIET))
792 bio_set_flag(bio, BIO_QUIET);
793 /* don't actually finish bio if it's part of flush sequence */
794 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
798 static void blk_account_io_completion(struct request *req, unsigned int bytes)
800 if (req->part && blk_do_io_stat(req)) {
801 const int sgrp = op_stat_group(req_op(req));
804 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
809 static void blk_print_req_error(struct request *req, blk_status_t status)
811 printk_ratelimited(KERN_ERR
812 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
813 "phys_seg %u prio class %u\n",
814 blk_status_to_str(status),
815 req->q->disk ? req->q->disk->disk_name : "?",
816 blk_rq_pos(req), (__force u32)req_op(req),
817 blk_op_str(req_op(req)),
818 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
819 req->nr_phys_segments,
820 IOPRIO_PRIO_CLASS(req->ioprio));
824 * Fully end IO on a request. Does not support partial completions, or
827 static void blk_complete_request(struct request *req)
829 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
830 int total_bytes = blk_rq_bytes(req);
831 struct bio *bio = req->bio;
833 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
838 #ifdef CONFIG_BLK_DEV_INTEGRITY
839 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
840 req->q->integrity.profile->complete_fn(req, total_bytes);
844 * Upper layers may call blk_crypto_evict_key() anytime after the last
845 * bio_endio(). Therefore, the keyslot must be released before that.
847 blk_crypto_rq_put_keyslot(req);
849 blk_account_io_completion(req, total_bytes);
852 struct bio *next = bio->bi_next;
854 /* Completion has already been traced */
855 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
857 if (req_op(req) == REQ_OP_ZONE_APPEND)
858 bio->bi_iter.bi_sector = req->__sector;
866 * Reset counters so that the request stacking driver
867 * can find how many bytes remain in the request
877 * blk_update_request - Complete multiple bytes without completing the request
878 * @req: the request being processed
879 * @error: block status code
880 * @nr_bytes: number of bytes to complete for @req
883 * Ends I/O on a number of bytes attached to @req, but doesn't complete
884 * the request structure even if @req doesn't have leftover.
885 * If @req has leftover, sets it up for the next range of segments.
887 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
888 * %false return from this function.
891 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
892 * except in the consistency check at the end of this function.
895 * %false - this request doesn't have any more data
896 * %true - this request has more data
898 bool blk_update_request(struct request *req, blk_status_t error,
899 unsigned int nr_bytes)
903 trace_block_rq_complete(req, error, nr_bytes);
908 #ifdef CONFIG_BLK_DEV_INTEGRITY
909 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
911 req->q->integrity.profile->complete_fn(req, nr_bytes);
915 * Upper layers may call blk_crypto_evict_key() anytime after the last
916 * bio_endio(). Therefore, the keyslot must be released before that.
918 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
919 __blk_crypto_rq_put_keyslot(req);
921 if (unlikely(error && !blk_rq_is_passthrough(req) &&
922 !(req->rq_flags & RQF_QUIET)) &&
923 !test_bit(GD_DEAD, &req->q->disk->state)) {
924 blk_print_req_error(req, error);
925 trace_block_rq_error(req, error, nr_bytes);
928 blk_account_io_completion(req, nr_bytes);
932 struct bio *bio = req->bio;
933 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
935 if (bio_bytes == bio->bi_iter.bi_size)
936 req->bio = bio->bi_next;
938 /* Completion has already been traced */
939 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
940 req_bio_endio(req, bio, bio_bytes, error);
942 total_bytes += bio_bytes;
943 nr_bytes -= bio_bytes;
954 * Reset counters so that the request stacking driver
955 * can find how many bytes remain in the request
962 req->__data_len -= total_bytes;
964 /* update sector only for requests with clear definition of sector */
965 if (!blk_rq_is_passthrough(req))
966 req->__sector += total_bytes >> 9;
968 /* mixed attributes always follow the first bio */
969 if (req->rq_flags & RQF_MIXED_MERGE) {
970 req->cmd_flags &= ~REQ_FAILFAST_MASK;
971 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
974 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
976 * If total number of sectors is less than the first segment
977 * size, something has gone terribly wrong.
979 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
980 blk_dump_rq_flags(req, "request botched");
981 req->__data_len = blk_rq_cur_bytes(req);
984 /* recalculate the number of segments */
985 req->nr_phys_segments = blk_recalc_rq_segments(req);
990 EXPORT_SYMBOL_GPL(blk_update_request);
992 static void __blk_account_io_done(struct request *req, u64 now)
994 const int sgrp = op_stat_group(req_op(req));
997 update_io_ticks(req->part, jiffies, true);
998 part_stat_inc(req->part, ios[sgrp]);
999 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
1003 static inline void blk_account_io_done(struct request *req, u64 now)
1006 * Account IO completion. flush_rq isn't accounted as a
1007 * normal IO on queueing nor completion. Accounting the
1008 * containing request is enough.
1010 if (blk_do_io_stat(req) && req->part &&
1011 !(req->rq_flags & RQF_FLUSH_SEQ))
1012 __blk_account_io_done(req, now);
1015 static void __blk_account_io_start(struct request *rq)
1018 * All non-passthrough requests are created from a bio with one
1019 * exception: when a flush command that is part of a flush sequence
1020 * generated by the state machine in blk-flush.c is cloned onto the
1021 * lower device by dm-multipath we can get here without a bio.
1024 rq->part = rq->bio->bi_bdev;
1026 rq->part = rq->q->disk->part0;
1029 update_io_ticks(rq->part, jiffies, false);
1033 static inline void blk_account_io_start(struct request *req)
1035 if (blk_do_io_stat(req))
1036 __blk_account_io_start(req);
1039 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1041 if (rq->rq_flags & RQF_STATS) {
1042 blk_mq_poll_stats_start(rq->q);
1043 blk_stat_add(rq, now);
1046 blk_mq_sched_completed_request(rq, now);
1047 blk_account_io_done(rq, now);
1050 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1052 if (blk_mq_need_time_stamp(rq))
1053 __blk_mq_end_request_acct(rq, ktime_get_ns());
1056 rq_qos_done(rq->q, rq);
1057 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1058 blk_mq_free_request(rq);
1060 blk_mq_free_request(rq);
1063 EXPORT_SYMBOL(__blk_mq_end_request);
1065 void blk_mq_end_request(struct request *rq, blk_status_t error)
1067 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1069 __blk_mq_end_request(rq, error);
1071 EXPORT_SYMBOL(blk_mq_end_request);
1073 #define TAG_COMP_BATCH 32
1075 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1076 int *tag_array, int nr_tags)
1078 struct request_queue *q = hctx->queue;
1081 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1082 * update hctx->nr_active in batch
1084 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1085 __blk_mq_sub_active_requests(hctx, nr_tags);
1087 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1088 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1091 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1093 int tags[TAG_COMP_BATCH], nr_tags = 0;
1094 struct blk_mq_hw_ctx *cur_hctx = NULL;
1099 now = ktime_get_ns();
1101 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1103 prefetch(rq->rq_next);
1105 blk_complete_request(rq);
1107 __blk_mq_end_request_acct(rq, now);
1109 rq_qos_done(rq->q, rq);
1112 * If end_io handler returns NONE, then it still has
1113 * ownership of the request.
1115 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1118 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1119 if (!req_ref_put_and_test(rq))
1122 blk_crypto_free_request(rq);
1123 blk_pm_mark_last_busy(rq);
1125 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1127 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1129 cur_hctx = rq->mq_hctx;
1131 tags[nr_tags++] = rq->tag;
1135 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1137 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1139 static void blk_complete_reqs(struct llist_head *list)
1141 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1142 struct request *rq, *next;
1144 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1145 rq->q->mq_ops->complete(rq);
1148 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1150 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1153 static int blk_softirq_cpu_dead(unsigned int cpu)
1155 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1159 static void __blk_mq_complete_request_remote(void *data)
1161 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1164 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1166 int cpu = raw_smp_processor_id();
1168 if (!IS_ENABLED(CONFIG_SMP) ||
1169 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1172 * With force threaded interrupts enabled, raising softirq from an SMP
1173 * function call will always result in waking the ksoftirqd thread.
1174 * This is probably worse than completing the request on a different
1177 if (force_irqthreads())
1180 /* same CPU or cache domain? Complete locally */
1181 if (cpu == rq->mq_ctx->cpu ||
1182 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1183 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1186 /* don't try to IPI to an offline CPU */
1187 return cpu_online(rq->mq_ctx->cpu);
1190 static void blk_mq_complete_send_ipi(struct request *rq)
1192 struct llist_head *list;
1195 cpu = rq->mq_ctx->cpu;
1196 list = &per_cpu(blk_cpu_done, cpu);
1197 if (llist_add(&rq->ipi_list, list)) {
1198 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1199 smp_call_function_single_async(cpu, &rq->csd);
1203 static void blk_mq_raise_softirq(struct request *rq)
1205 struct llist_head *list;
1208 list = this_cpu_ptr(&blk_cpu_done);
1209 if (llist_add(&rq->ipi_list, list))
1210 raise_softirq(BLOCK_SOFTIRQ);
1214 bool blk_mq_complete_request_remote(struct request *rq)
1216 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1219 * For request which hctx has only one ctx mapping,
1220 * or a polled request, always complete locally,
1221 * it's pointless to redirect the completion.
1223 if (rq->mq_hctx->nr_ctx == 1 ||
1224 rq->cmd_flags & REQ_POLLED)
1227 if (blk_mq_complete_need_ipi(rq)) {
1228 blk_mq_complete_send_ipi(rq);
1232 if (rq->q->nr_hw_queues == 1) {
1233 blk_mq_raise_softirq(rq);
1238 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1241 * blk_mq_complete_request - end I/O on a request
1242 * @rq: the request being processed
1245 * Complete a request by scheduling the ->complete_rq operation.
1247 void blk_mq_complete_request(struct request *rq)
1249 if (!blk_mq_complete_request_remote(rq))
1250 rq->q->mq_ops->complete(rq);
1252 EXPORT_SYMBOL(blk_mq_complete_request);
1255 * blk_mq_start_request - Start processing a request
1256 * @rq: Pointer to request to be started
1258 * Function used by device drivers to notify the block layer that a request
1259 * is going to be processed now, so blk layer can do proper initializations
1260 * such as starting the timeout timer.
1262 void blk_mq_start_request(struct request *rq)
1264 struct request_queue *q = rq->q;
1266 trace_block_rq_issue(rq);
1268 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1269 rq->io_start_time_ns = ktime_get_ns();
1270 rq->stats_sectors = blk_rq_sectors(rq);
1271 rq->rq_flags |= RQF_STATS;
1272 rq_qos_issue(q, rq);
1275 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1278 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1280 #ifdef CONFIG_BLK_DEV_INTEGRITY
1281 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1282 q->integrity.profile->prepare_fn(rq);
1284 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1285 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1287 EXPORT_SYMBOL(blk_mq_start_request);
1290 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1291 * queues. This is important for md arrays to benefit from merging
1294 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1296 if (plug->multiple_queues)
1297 return BLK_MAX_REQUEST_COUNT * 2;
1298 return BLK_MAX_REQUEST_COUNT;
1301 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1303 struct request *last = rq_list_peek(&plug->mq_list);
1305 if (!plug->rq_count) {
1306 trace_block_plug(rq->q);
1307 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1308 (!blk_queue_nomerges(rq->q) &&
1309 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1310 blk_mq_flush_plug_list(plug, false);
1312 trace_block_plug(rq->q);
1315 if (!plug->multiple_queues && last && last->q != rq->q)
1316 plug->multiple_queues = true;
1317 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1318 plug->has_elevator = true;
1320 rq_list_add(&plug->mq_list, rq);
1325 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1326 * @rq: request to insert
1327 * @at_head: insert request at head or tail of queue
1330 * Insert a fully prepared request at the back of the I/O scheduler queue
1331 * for execution. Don't wait for completion.
1334 * This function will invoke @done directly if the queue is dead.
1336 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1338 WARN_ON(irqs_disabled());
1339 WARN_ON(!blk_rq_is_passthrough(rq));
1341 blk_account_io_start(rq);
1344 * As plugging can be enabled for passthrough requests on a zoned
1345 * device, directly accessing the plug instead of using blk_mq_plug()
1346 * should not have any consequences.
1349 blk_add_rq_to_plug(current->plug, rq);
1351 blk_mq_sched_insert_request(rq, at_head, true, false);
1353 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1355 struct blk_rq_wait {
1356 struct completion done;
1360 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1362 struct blk_rq_wait *wait = rq->end_io_data;
1365 complete(&wait->done);
1366 return RQ_END_IO_NONE;
1369 bool blk_rq_is_poll(struct request *rq)
1373 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1375 if (WARN_ON_ONCE(!rq->bio))
1379 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1381 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1384 bio_poll(rq->bio, NULL, 0);
1386 } while (!completion_done(wait));
1390 * blk_execute_rq - insert a request into queue for execution
1391 * @rq: request to insert
1392 * @at_head: insert request at head or tail of queue
1395 * Insert a fully prepared request at the back of the I/O scheduler queue
1396 * for execution and wait for completion.
1397 * Return: The blk_status_t result provided to blk_mq_end_request().
1399 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1401 struct blk_rq_wait wait = {
1402 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1405 WARN_ON(irqs_disabled());
1406 WARN_ON(!blk_rq_is_passthrough(rq));
1408 rq->end_io_data = &wait;
1409 rq->end_io = blk_end_sync_rq;
1411 blk_account_io_start(rq);
1412 blk_mq_sched_insert_request(rq, at_head, true, false);
1414 if (blk_rq_is_poll(rq)) {
1415 blk_rq_poll_completion(rq, &wait.done);
1418 * Prevent hang_check timer from firing at us during very long
1421 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1424 while (!wait_for_completion_io_timeout(&wait.done,
1425 hang_check * (HZ/2)))
1428 wait_for_completion_io(&wait.done);
1433 EXPORT_SYMBOL(blk_execute_rq);
1435 static void __blk_mq_requeue_request(struct request *rq)
1437 struct request_queue *q = rq->q;
1439 blk_mq_put_driver_tag(rq);
1441 trace_block_rq_requeue(rq);
1442 rq_qos_requeue(q, rq);
1444 if (blk_mq_request_started(rq)) {
1445 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1446 rq->rq_flags &= ~RQF_TIMED_OUT;
1450 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1452 __blk_mq_requeue_request(rq);
1454 /* this request will be re-inserted to io scheduler queue */
1455 blk_mq_sched_requeue_request(rq);
1457 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1459 EXPORT_SYMBOL(blk_mq_requeue_request);
1461 static void blk_mq_requeue_work(struct work_struct *work)
1463 struct request_queue *q =
1464 container_of(work, struct request_queue, requeue_work.work);
1466 struct request *rq, *next;
1468 spin_lock_irq(&q->requeue_lock);
1469 list_splice_init(&q->requeue_list, &rq_list);
1470 spin_unlock_irq(&q->requeue_lock);
1472 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1473 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1476 rq->rq_flags &= ~RQF_SOFTBARRIER;
1477 list_del_init(&rq->queuelist);
1479 * If RQF_DONTPREP, rq has contained some driver specific
1480 * data, so insert it to hctx dispatch list to avoid any
1483 if (rq->rq_flags & RQF_DONTPREP)
1484 blk_mq_request_bypass_insert(rq, false, false);
1486 blk_mq_sched_insert_request(rq, true, false, false);
1489 while (!list_empty(&rq_list)) {
1490 rq = list_entry(rq_list.next, struct request, queuelist);
1491 list_del_init(&rq->queuelist);
1492 blk_mq_sched_insert_request(rq, false, false, false);
1495 blk_mq_run_hw_queues(q, false);
1498 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1499 bool kick_requeue_list)
1501 struct request_queue *q = rq->q;
1502 unsigned long flags;
1505 * We abuse this flag that is otherwise used by the I/O scheduler to
1506 * request head insertion from the workqueue.
1508 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1510 spin_lock_irqsave(&q->requeue_lock, flags);
1512 rq->rq_flags |= RQF_SOFTBARRIER;
1513 list_add(&rq->queuelist, &q->requeue_list);
1515 list_add_tail(&rq->queuelist, &q->requeue_list);
1517 spin_unlock_irqrestore(&q->requeue_lock, flags);
1519 if (kick_requeue_list)
1520 blk_mq_kick_requeue_list(q);
1523 void blk_mq_kick_requeue_list(struct request_queue *q)
1525 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1527 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1529 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1530 unsigned long msecs)
1532 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1533 msecs_to_jiffies(msecs));
1535 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1537 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1540 * If we find a request that isn't idle we know the queue is busy
1541 * as it's checked in the iter.
1542 * Return false to stop the iteration.
1544 if (blk_mq_request_started(rq)) {
1554 bool blk_mq_queue_inflight(struct request_queue *q)
1558 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1561 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1563 static void blk_mq_rq_timed_out(struct request *req)
1565 req->rq_flags |= RQF_TIMED_OUT;
1566 if (req->q->mq_ops->timeout) {
1567 enum blk_eh_timer_return ret;
1569 ret = req->q->mq_ops->timeout(req);
1570 if (ret == BLK_EH_DONE)
1572 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1578 struct blk_expired_data {
1579 bool has_timedout_rq;
1581 unsigned long timeout_start;
1584 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1586 unsigned long deadline;
1588 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1590 if (rq->rq_flags & RQF_TIMED_OUT)
1593 deadline = READ_ONCE(rq->deadline);
1594 if (time_after_eq(expired->timeout_start, deadline))
1597 if (expired->next == 0)
1598 expired->next = deadline;
1599 else if (time_after(expired->next, deadline))
1600 expired->next = deadline;
1604 void blk_mq_put_rq_ref(struct request *rq)
1606 if (is_flush_rq(rq)) {
1607 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1608 blk_mq_free_request(rq);
1609 } else if (req_ref_put_and_test(rq)) {
1610 __blk_mq_free_request(rq);
1614 static bool blk_mq_check_expired(struct request *rq, void *priv)
1616 struct blk_expired_data *expired = priv;
1619 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1620 * be reallocated underneath the timeout handler's processing, then
1621 * the expire check is reliable. If the request is not expired, then
1622 * it was completed and reallocated as a new request after returning
1623 * from blk_mq_check_expired().
1625 if (blk_mq_req_expired(rq, expired)) {
1626 expired->has_timedout_rq = true;
1632 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1634 struct blk_expired_data *expired = priv;
1636 if (blk_mq_req_expired(rq, expired))
1637 blk_mq_rq_timed_out(rq);
1641 static void blk_mq_timeout_work(struct work_struct *work)
1643 struct request_queue *q =
1644 container_of(work, struct request_queue, timeout_work);
1645 struct blk_expired_data expired = {
1646 .timeout_start = jiffies,
1648 struct blk_mq_hw_ctx *hctx;
1651 /* A deadlock might occur if a request is stuck requiring a
1652 * timeout at the same time a queue freeze is waiting
1653 * completion, since the timeout code would not be able to
1654 * acquire the queue reference here.
1656 * That's why we don't use blk_queue_enter here; instead, we use
1657 * percpu_ref_tryget directly, because we need to be able to
1658 * obtain a reference even in the short window between the queue
1659 * starting to freeze, by dropping the first reference in
1660 * blk_freeze_queue_start, and the moment the last request is
1661 * consumed, marked by the instant q_usage_counter reaches
1664 if (!percpu_ref_tryget(&q->q_usage_counter))
1667 /* check if there is any timed-out request */
1668 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1669 if (expired.has_timedout_rq) {
1671 * Before walking tags, we must ensure any submit started
1672 * before the current time has finished. Since the submit
1673 * uses srcu or rcu, wait for a synchronization point to
1674 * ensure all running submits have finished
1676 blk_mq_wait_quiesce_done(q->tag_set);
1679 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1682 if (expired.next != 0) {
1683 mod_timer(&q->timeout, expired.next);
1686 * Request timeouts are handled as a forward rolling timer. If
1687 * we end up here it means that no requests are pending and
1688 * also that no request has been pending for a while. Mark
1689 * each hctx as idle.
1691 queue_for_each_hw_ctx(q, hctx, i) {
1692 /* the hctx may be unmapped, so check it here */
1693 if (blk_mq_hw_queue_mapped(hctx))
1694 blk_mq_tag_idle(hctx);
1700 struct flush_busy_ctx_data {
1701 struct blk_mq_hw_ctx *hctx;
1702 struct list_head *list;
1705 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1707 struct flush_busy_ctx_data *flush_data = data;
1708 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1709 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1710 enum hctx_type type = hctx->type;
1712 spin_lock(&ctx->lock);
1713 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1714 sbitmap_clear_bit(sb, bitnr);
1715 spin_unlock(&ctx->lock);
1720 * Process software queues that have been marked busy, splicing them
1721 * to the for-dispatch
1723 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1725 struct flush_busy_ctx_data data = {
1730 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1732 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1734 struct dispatch_rq_data {
1735 struct blk_mq_hw_ctx *hctx;
1739 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1742 struct dispatch_rq_data *dispatch_data = data;
1743 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1744 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1745 enum hctx_type type = hctx->type;
1747 spin_lock(&ctx->lock);
1748 if (!list_empty(&ctx->rq_lists[type])) {
1749 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1750 list_del_init(&dispatch_data->rq->queuelist);
1751 if (list_empty(&ctx->rq_lists[type]))
1752 sbitmap_clear_bit(sb, bitnr);
1754 spin_unlock(&ctx->lock);
1756 return !dispatch_data->rq;
1759 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1760 struct blk_mq_ctx *start)
1762 unsigned off = start ? start->index_hw[hctx->type] : 0;
1763 struct dispatch_rq_data data = {
1768 __sbitmap_for_each_set(&hctx->ctx_map, off,
1769 dispatch_rq_from_ctx, &data);
1774 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1776 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1777 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1780 blk_mq_tag_busy(rq->mq_hctx);
1782 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1783 bt = &rq->mq_hctx->tags->breserved_tags;
1786 if (!hctx_may_queue(rq->mq_hctx, bt))
1790 tag = __sbitmap_queue_get(bt);
1791 if (tag == BLK_MQ_NO_TAG)
1794 rq->tag = tag + tag_offset;
1798 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1800 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1803 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1804 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1805 rq->rq_flags |= RQF_MQ_INFLIGHT;
1806 __blk_mq_inc_active_requests(hctx);
1808 hctx->tags->rqs[rq->tag] = rq;
1812 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1813 int flags, void *key)
1815 struct blk_mq_hw_ctx *hctx;
1817 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1819 spin_lock(&hctx->dispatch_wait_lock);
1820 if (!list_empty(&wait->entry)) {
1821 struct sbitmap_queue *sbq;
1823 list_del_init(&wait->entry);
1824 sbq = &hctx->tags->bitmap_tags;
1825 atomic_dec(&sbq->ws_active);
1827 spin_unlock(&hctx->dispatch_wait_lock);
1829 blk_mq_run_hw_queue(hctx, true);
1834 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1835 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1836 * restart. For both cases, take care to check the condition again after
1837 * marking us as waiting.
1839 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1842 struct sbitmap_queue *sbq;
1843 struct wait_queue_head *wq;
1844 wait_queue_entry_t *wait;
1847 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1848 !(blk_mq_is_shared_tags(hctx->flags))) {
1849 blk_mq_sched_mark_restart_hctx(hctx);
1852 * It's possible that a tag was freed in the window between the
1853 * allocation failure and adding the hardware queue to the wait
1856 * Don't clear RESTART here, someone else could have set it.
1857 * At most this will cost an extra queue run.
1859 return blk_mq_get_driver_tag(rq);
1862 wait = &hctx->dispatch_wait;
1863 if (!list_empty_careful(&wait->entry))
1866 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1867 sbq = &hctx->tags->breserved_tags;
1869 sbq = &hctx->tags->bitmap_tags;
1870 wq = &bt_wait_ptr(sbq, hctx)->wait;
1872 spin_lock_irq(&wq->lock);
1873 spin_lock(&hctx->dispatch_wait_lock);
1874 if (!list_empty(&wait->entry)) {
1875 spin_unlock(&hctx->dispatch_wait_lock);
1876 spin_unlock_irq(&wq->lock);
1880 atomic_inc(&sbq->ws_active);
1881 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1882 __add_wait_queue(wq, wait);
1885 * It's possible that a tag was freed in the window between the
1886 * allocation failure and adding the hardware queue to the wait
1889 ret = blk_mq_get_driver_tag(rq);
1891 spin_unlock(&hctx->dispatch_wait_lock);
1892 spin_unlock_irq(&wq->lock);
1897 * We got a tag, remove ourselves from the wait queue to ensure
1898 * someone else gets the wakeup.
1900 list_del_init(&wait->entry);
1901 atomic_dec(&sbq->ws_active);
1902 spin_unlock(&hctx->dispatch_wait_lock);
1903 spin_unlock_irq(&wq->lock);
1908 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1909 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1911 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1912 * - EWMA is one simple way to compute running average value
1913 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1914 * - take 4 as factor for avoiding to get too small(0) result, and this
1915 * factor doesn't matter because EWMA decreases exponentially
1917 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1921 ewma = hctx->dispatch_busy;
1926 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1928 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1929 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1931 hctx->dispatch_busy = ewma;
1934 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1936 static void blk_mq_handle_dev_resource(struct request *rq,
1937 struct list_head *list)
1939 list_add(&rq->queuelist, list);
1940 __blk_mq_requeue_request(rq);
1943 static void blk_mq_handle_zone_resource(struct request *rq,
1944 struct list_head *zone_list)
1947 * If we end up here it is because we cannot dispatch a request to a
1948 * specific zone due to LLD level zone-write locking or other zone
1949 * related resource not being available. In this case, set the request
1950 * aside in zone_list for retrying it later.
1952 list_add(&rq->queuelist, zone_list);
1953 __blk_mq_requeue_request(rq);
1956 enum prep_dispatch {
1958 PREP_DISPATCH_NO_TAG,
1959 PREP_DISPATCH_NO_BUDGET,
1962 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1965 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1966 int budget_token = -1;
1969 budget_token = blk_mq_get_dispatch_budget(rq->q);
1970 if (budget_token < 0) {
1971 blk_mq_put_driver_tag(rq);
1972 return PREP_DISPATCH_NO_BUDGET;
1974 blk_mq_set_rq_budget_token(rq, budget_token);
1977 if (!blk_mq_get_driver_tag(rq)) {
1979 * The initial allocation attempt failed, so we need to
1980 * rerun the hardware queue when a tag is freed. The
1981 * waitqueue takes care of that. If the queue is run
1982 * before we add this entry back on the dispatch list,
1983 * we'll re-run it below.
1985 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1987 * All budgets not got from this function will be put
1988 * together during handling partial dispatch
1991 blk_mq_put_dispatch_budget(rq->q, budget_token);
1992 return PREP_DISPATCH_NO_TAG;
1996 return PREP_DISPATCH_OK;
1999 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
2000 static void blk_mq_release_budgets(struct request_queue *q,
2001 struct list_head *list)
2005 list_for_each_entry(rq, list, queuelist) {
2006 int budget_token = blk_mq_get_rq_budget_token(rq);
2008 if (budget_token >= 0)
2009 blk_mq_put_dispatch_budget(q, budget_token);
2014 * blk_mq_commit_rqs will notify driver using bd->last that there is no
2015 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
2017 * Attention, we should explicitly call this in unusual cases:
2018 * 1) did not queue everything initially scheduled to queue
2019 * 2) the last attempt to queue a request failed
2021 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
2024 if (hctx->queue->mq_ops->commit_rqs && queued) {
2025 trace_block_unplug(hctx->queue, queued, !from_schedule);
2026 hctx->queue->mq_ops->commit_rqs(hctx);
2031 * Returns true if we did some work AND can potentially do more.
2033 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
2034 unsigned int nr_budgets)
2036 enum prep_dispatch prep;
2037 struct request_queue *q = hctx->queue;
2040 blk_status_t ret = BLK_STS_OK;
2041 LIST_HEAD(zone_list);
2042 bool needs_resource = false;
2044 if (list_empty(list))
2048 * Now process all the entries, sending them to the driver.
2052 struct blk_mq_queue_data bd;
2054 rq = list_first_entry(list, struct request, queuelist);
2056 WARN_ON_ONCE(hctx != rq->mq_hctx);
2057 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2058 if (prep != PREP_DISPATCH_OK)
2061 list_del_init(&rq->queuelist);
2064 bd.last = list_empty(list);
2067 * once the request is queued to lld, no need to cover the
2072 ret = q->mq_ops->queue_rq(hctx, &bd);
2077 case BLK_STS_RESOURCE:
2078 needs_resource = true;
2080 case BLK_STS_DEV_RESOURCE:
2081 blk_mq_handle_dev_resource(rq, list);
2083 case BLK_STS_ZONE_RESOURCE:
2085 * Move the request to zone_list and keep going through
2086 * the dispatch list to find more requests the drive can
2089 blk_mq_handle_zone_resource(rq, &zone_list);
2090 needs_resource = true;
2093 blk_mq_end_request(rq, ret);
2095 } while (!list_empty(list));
2097 if (!list_empty(&zone_list))
2098 list_splice_tail_init(&zone_list, list);
2100 /* If we didn't flush the entire list, we could have told the driver
2101 * there was more coming, but that turned out to be a lie.
2103 if (!list_empty(list) || ret != BLK_STS_OK)
2104 blk_mq_commit_rqs(hctx, queued, false);
2107 * Any items that need requeuing? Stuff them into hctx->dispatch,
2108 * that is where we will continue on next queue run.
2110 if (!list_empty(list)) {
2112 /* For non-shared tags, the RESTART check will suffice */
2113 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2114 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2115 blk_mq_is_shared_tags(hctx->flags));
2118 blk_mq_release_budgets(q, list);
2120 spin_lock(&hctx->lock);
2121 list_splice_tail_init(list, &hctx->dispatch);
2122 spin_unlock(&hctx->lock);
2125 * Order adding requests to hctx->dispatch and checking
2126 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2127 * in blk_mq_sched_restart(). Avoid restart code path to
2128 * miss the new added requests to hctx->dispatch, meantime
2129 * SCHED_RESTART is observed here.
2134 * If SCHED_RESTART was set by the caller of this function and
2135 * it is no longer set that means that it was cleared by another
2136 * thread and hence that a queue rerun is needed.
2138 * If 'no_tag' is set, that means that we failed getting
2139 * a driver tag with an I/O scheduler attached. If our dispatch
2140 * waitqueue is no longer active, ensure that we run the queue
2141 * AFTER adding our entries back to the list.
2143 * If no I/O scheduler has been configured it is possible that
2144 * the hardware queue got stopped and restarted before requests
2145 * were pushed back onto the dispatch list. Rerun the queue to
2146 * avoid starvation. Notes:
2147 * - blk_mq_run_hw_queue() checks whether or not a queue has
2148 * been stopped before rerunning a queue.
2149 * - Some but not all block drivers stop a queue before
2150 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2153 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2154 * bit is set, run queue after a delay to avoid IO stalls
2155 * that could otherwise occur if the queue is idle. We'll do
2156 * similar if we couldn't get budget or couldn't lock a zone
2157 * and SCHED_RESTART is set.
2159 needs_restart = blk_mq_sched_needs_restart(hctx);
2160 if (prep == PREP_DISPATCH_NO_BUDGET)
2161 needs_resource = true;
2162 if (!needs_restart ||
2163 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2164 blk_mq_run_hw_queue(hctx, true);
2165 else if (needs_resource)
2166 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2168 blk_mq_update_dispatch_busy(hctx, true);
2172 blk_mq_update_dispatch_busy(hctx, false);
2177 * __blk_mq_run_hw_queue - Run a hardware queue.
2178 * @hctx: Pointer to the hardware queue to run.
2180 * Send pending requests to the hardware.
2182 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
2185 * We can't run the queue inline with ints disabled. Ensure that
2186 * we catch bad users of this early.
2188 WARN_ON_ONCE(in_interrupt());
2190 blk_mq_run_dispatch_ops(hctx->queue,
2191 blk_mq_sched_dispatch_requests(hctx));
2194 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2196 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2198 if (cpu >= nr_cpu_ids)
2199 cpu = cpumask_first(hctx->cpumask);
2204 * It'd be great if the workqueue API had a way to pass
2205 * in a mask and had some smarts for more clever placement.
2206 * For now we just round-robin here, switching for every
2207 * BLK_MQ_CPU_WORK_BATCH queued items.
2209 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2212 int next_cpu = hctx->next_cpu;
2214 if (hctx->queue->nr_hw_queues == 1)
2215 return WORK_CPU_UNBOUND;
2217 if (--hctx->next_cpu_batch <= 0) {
2219 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2221 if (next_cpu >= nr_cpu_ids)
2222 next_cpu = blk_mq_first_mapped_cpu(hctx);
2223 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2227 * Do unbound schedule if we can't find a online CPU for this hctx,
2228 * and it should only happen in the path of handling CPU DEAD.
2230 if (!cpu_online(next_cpu)) {
2237 * Make sure to re-select CPU next time once after CPUs
2238 * in hctx->cpumask become online again.
2240 hctx->next_cpu = next_cpu;
2241 hctx->next_cpu_batch = 1;
2242 return WORK_CPU_UNBOUND;
2245 hctx->next_cpu = next_cpu;
2250 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2251 * @hctx: Pointer to the hardware queue to run.
2252 * @async: If we want to run the queue asynchronously.
2253 * @msecs: Milliseconds of delay to wait before running the queue.
2255 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2256 * with a delay of @msecs.
2258 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2259 unsigned long msecs)
2261 if (unlikely(blk_mq_hctx_stopped(hctx)))
2264 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2265 if (cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2266 __blk_mq_run_hw_queue(hctx);
2271 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2272 msecs_to_jiffies(msecs));
2276 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2277 * @hctx: Pointer to the hardware queue to run.
2278 * @msecs: Milliseconds of delay to wait before running the queue.
2280 * Run a hardware queue asynchronously with a delay of @msecs.
2282 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2284 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2286 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2289 * blk_mq_run_hw_queue - Start to run a hardware queue.
2290 * @hctx: Pointer to the hardware queue to run.
2291 * @async: If we want to run the queue asynchronously.
2293 * Check if the request queue is not in a quiesced state and if there are
2294 * pending requests to be sent. If this is true, run the queue to send requests
2297 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2302 * When queue is quiesced, we may be switching io scheduler, or
2303 * updating nr_hw_queues, or other things, and we can't run queue
2304 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2306 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2309 __blk_mq_run_dispatch_ops(hctx->queue, false,
2310 need_run = !blk_queue_quiesced(hctx->queue) &&
2311 blk_mq_hctx_has_pending(hctx));
2314 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2316 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2319 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2322 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2324 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2326 * If the IO scheduler does not respect hardware queues when
2327 * dispatching, we just don't bother with multiple HW queues and
2328 * dispatch from hctx for the current CPU since running multiple queues
2329 * just causes lock contention inside the scheduler and pointless cache
2332 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2334 if (!blk_mq_hctx_stopped(hctx))
2340 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2341 * @q: Pointer to the request queue to run.
2342 * @async: If we want to run the queue asynchronously.
2344 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2346 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2350 if (blk_queue_sq_sched(q))
2351 sq_hctx = blk_mq_get_sq_hctx(q);
2352 queue_for_each_hw_ctx(q, hctx, i) {
2353 if (blk_mq_hctx_stopped(hctx))
2356 * Dispatch from this hctx either if there's no hctx preferred
2357 * by IO scheduler or if it has requests that bypass the
2360 if (!sq_hctx || sq_hctx == hctx ||
2361 !list_empty_careful(&hctx->dispatch))
2362 blk_mq_run_hw_queue(hctx, async);
2365 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2368 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2369 * @q: Pointer to the request queue to run.
2370 * @msecs: Milliseconds of delay to wait before running the queues.
2372 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2374 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2378 if (blk_queue_sq_sched(q))
2379 sq_hctx = blk_mq_get_sq_hctx(q);
2380 queue_for_each_hw_ctx(q, hctx, i) {
2381 if (blk_mq_hctx_stopped(hctx))
2384 * If there is already a run_work pending, leave the
2385 * pending delay untouched. Otherwise, a hctx can stall
2386 * if another hctx is re-delaying the other's work
2387 * before the work executes.
2389 if (delayed_work_pending(&hctx->run_work))
2392 * Dispatch from this hctx either if there's no hctx preferred
2393 * by IO scheduler or if it has requests that bypass the
2396 if (!sq_hctx || sq_hctx == hctx ||
2397 !list_empty_careful(&hctx->dispatch))
2398 blk_mq_delay_run_hw_queue(hctx, msecs);
2401 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2404 * This function is often used for pausing .queue_rq() by driver when
2405 * there isn't enough resource or some conditions aren't satisfied, and
2406 * BLK_STS_RESOURCE is usually returned.
2408 * We do not guarantee that dispatch can be drained or blocked
2409 * after blk_mq_stop_hw_queue() returns. Please use
2410 * blk_mq_quiesce_queue() for that requirement.
2412 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2414 cancel_delayed_work(&hctx->run_work);
2416 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2418 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2421 * This function is often used for pausing .queue_rq() by driver when
2422 * there isn't enough resource or some conditions aren't satisfied, and
2423 * BLK_STS_RESOURCE is usually returned.
2425 * We do not guarantee that dispatch can be drained or blocked
2426 * after blk_mq_stop_hw_queues() returns. Please use
2427 * blk_mq_quiesce_queue() for that requirement.
2429 void blk_mq_stop_hw_queues(struct request_queue *q)
2431 struct blk_mq_hw_ctx *hctx;
2434 queue_for_each_hw_ctx(q, hctx, i)
2435 blk_mq_stop_hw_queue(hctx);
2437 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2439 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2441 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2443 blk_mq_run_hw_queue(hctx, false);
2445 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2447 void blk_mq_start_hw_queues(struct request_queue *q)
2449 struct blk_mq_hw_ctx *hctx;
2452 queue_for_each_hw_ctx(q, hctx, i)
2453 blk_mq_start_hw_queue(hctx);
2455 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2457 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2459 if (!blk_mq_hctx_stopped(hctx))
2462 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2463 blk_mq_run_hw_queue(hctx, async);
2465 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2467 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2469 struct blk_mq_hw_ctx *hctx;
2472 queue_for_each_hw_ctx(q, hctx, i)
2473 blk_mq_start_stopped_hw_queue(hctx, async);
2475 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2477 static void blk_mq_run_work_fn(struct work_struct *work)
2479 struct blk_mq_hw_ctx *hctx;
2481 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2484 * If we are stopped, don't run the queue.
2486 if (blk_mq_hctx_stopped(hctx))
2489 __blk_mq_run_hw_queue(hctx);
2492 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2496 struct blk_mq_ctx *ctx = rq->mq_ctx;
2497 enum hctx_type type = hctx->type;
2499 lockdep_assert_held(&ctx->lock);
2501 trace_block_rq_insert(rq);
2504 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2506 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2509 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2512 struct blk_mq_ctx *ctx = rq->mq_ctx;
2514 lockdep_assert_held(&ctx->lock);
2516 __blk_mq_insert_req_list(hctx, rq, at_head);
2517 blk_mq_hctx_mark_pending(hctx, ctx);
2521 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2522 * @rq: Pointer to request to be inserted.
2523 * @at_head: true if the request should be inserted at the head of the list.
2524 * @run_queue: If we should run the hardware queue after inserting the request.
2526 * Should only be used carefully, when the caller knows we want to
2527 * bypass a potential IO scheduler on the target device.
2529 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2532 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2534 spin_lock(&hctx->lock);
2536 list_add(&rq->queuelist, &hctx->dispatch);
2538 list_add_tail(&rq->queuelist, &hctx->dispatch);
2539 spin_unlock(&hctx->lock);
2542 blk_mq_run_hw_queue(hctx, false);
2545 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2546 struct list_head *list)
2550 enum hctx_type type = hctx->type;
2553 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2556 list_for_each_entry(rq, list, queuelist) {
2557 BUG_ON(rq->mq_ctx != ctx);
2558 trace_block_rq_insert(rq);
2561 spin_lock(&ctx->lock);
2562 list_splice_tail_init(list, &ctx->rq_lists[type]);
2563 blk_mq_hctx_mark_pending(hctx, ctx);
2564 spin_unlock(&ctx->lock);
2567 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2568 unsigned int nr_segs)
2572 if (bio->bi_opf & REQ_RAHEAD)
2573 rq->cmd_flags |= REQ_FAILFAST_MASK;
2575 rq->__sector = bio->bi_iter.bi_sector;
2576 blk_rq_bio_prep(rq, bio, nr_segs);
2578 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2579 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2582 blk_account_io_start(rq);
2585 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2586 struct request *rq, bool last)
2588 struct request_queue *q = rq->q;
2589 struct blk_mq_queue_data bd = {
2596 * For OK queue, we are done. For error, caller may kill it.
2597 * Any other error (busy), just add it to our list as we
2598 * previously would have done.
2600 ret = q->mq_ops->queue_rq(hctx, &bd);
2603 blk_mq_update_dispatch_busy(hctx, false);
2605 case BLK_STS_RESOURCE:
2606 case BLK_STS_DEV_RESOURCE:
2607 blk_mq_update_dispatch_busy(hctx, true);
2608 __blk_mq_requeue_request(rq);
2611 blk_mq_update_dispatch_busy(hctx, false);
2618 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2620 bool bypass_insert, bool last)
2622 struct request_queue *q = rq->q;
2623 bool run_queue = true;
2627 * RCU or SRCU read lock is needed before checking quiesced flag.
2629 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2630 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2631 * and avoid driver to try to dispatch again.
2633 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2635 bypass_insert = false;
2639 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2642 budget_token = blk_mq_get_dispatch_budget(q);
2643 if (budget_token < 0)
2646 blk_mq_set_rq_budget_token(rq, budget_token);
2648 if (!blk_mq_get_driver_tag(rq)) {
2649 blk_mq_put_dispatch_budget(q, budget_token);
2653 return __blk_mq_issue_directly(hctx, rq, last);
2656 return BLK_STS_RESOURCE;
2658 blk_mq_sched_insert_request(rq, false, run_queue, false);
2664 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2665 * @hctx: Pointer of the associated hardware queue.
2666 * @rq: Pointer to request to be sent.
2668 * If the device has enough resources to accept a new request now, send the
2669 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2670 * we can try send it another time in the future. Requests inserted at this
2671 * queue have higher priority.
2673 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2677 __blk_mq_try_issue_directly(hctx, rq, false, true);
2679 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2680 blk_mq_request_bypass_insert(rq, false, true);
2681 else if (ret != BLK_STS_OK)
2682 blk_mq_end_request(rq, ret);
2685 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2687 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2690 static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2692 struct blk_mq_hw_ctx *hctx = NULL;
2695 blk_status_t ret = BLK_STS_OK;
2697 while ((rq = rq_list_pop(&plug->mq_list))) {
2698 bool last = rq_list_empty(plug->mq_list);
2700 if (hctx != rq->mq_hctx) {
2702 blk_mq_commit_rqs(hctx, queued, false);
2708 ret = blk_mq_request_issue_directly(rq, last);
2713 case BLK_STS_RESOURCE:
2714 case BLK_STS_DEV_RESOURCE:
2715 blk_mq_request_bypass_insert(rq, false, true);
2718 blk_mq_end_request(rq, ret);
2724 if (ret != BLK_STS_OK)
2725 blk_mq_commit_rqs(hctx, queued, false);
2728 static void __blk_mq_flush_plug_list(struct request_queue *q,
2729 struct blk_plug *plug)
2731 if (blk_queue_quiesced(q))
2733 q->mq_ops->queue_rqs(&plug->mq_list);
2736 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2738 struct blk_mq_hw_ctx *this_hctx = NULL;
2739 struct blk_mq_ctx *this_ctx = NULL;
2740 struct request *requeue_list = NULL;
2741 unsigned int depth = 0;
2745 struct request *rq = rq_list_pop(&plug->mq_list);
2748 this_hctx = rq->mq_hctx;
2749 this_ctx = rq->mq_ctx;
2750 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2751 rq_list_add(&requeue_list, rq);
2754 list_add_tail(&rq->queuelist, &list);
2756 } while (!rq_list_empty(plug->mq_list));
2758 plug->mq_list = requeue_list;
2759 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2760 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched);
2763 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2767 if (rq_list_empty(plug->mq_list))
2771 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2772 struct request_queue *q;
2774 rq = rq_list_peek(&plug->mq_list);
2778 * Peek first request and see if we have a ->queue_rqs() hook.
2779 * If we do, we can dispatch the whole plug list in one go. We
2780 * already know at this point that all requests belong to the
2781 * same queue, caller must ensure that's the case.
2783 * Since we pass off the full list to the driver at this point,
2784 * we do not increment the active request count for the queue.
2785 * Bypass shared tags for now because of that.
2787 if (q->mq_ops->queue_rqs &&
2788 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2789 blk_mq_run_dispatch_ops(q,
2790 __blk_mq_flush_plug_list(q, plug));
2791 if (rq_list_empty(plug->mq_list))
2795 blk_mq_run_dispatch_ops(q,
2796 blk_mq_plug_issue_direct(plug));
2797 if (rq_list_empty(plug->mq_list))
2802 blk_mq_dispatch_plug_list(plug, from_schedule);
2803 } while (!rq_list_empty(plug->mq_list));
2806 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2807 struct list_head *list)
2810 blk_status_t ret = BLK_STS_OK;
2812 while (!list_empty(list)) {
2813 struct request *rq = list_first_entry(list, struct request,
2816 list_del_init(&rq->queuelist);
2817 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2822 case BLK_STS_RESOURCE:
2823 case BLK_STS_DEV_RESOURCE:
2824 blk_mq_request_bypass_insert(rq, false,
2828 blk_mq_end_request(rq, ret);
2834 if (ret != BLK_STS_OK)
2835 blk_mq_commit_rqs(hctx, queued, false);
2838 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2839 struct bio *bio, unsigned int nr_segs)
2841 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2842 if (blk_attempt_plug_merge(q, bio, nr_segs))
2844 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2850 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2851 struct blk_plug *plug,
2855 struct blk_mq_alloc_data data = {
2858 .cmd_flags = bio->bi_opf,
2862 if (unlikely(bio_queue_enter(bio)))
2865 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2868 rq_qos_throttle(q, bio);
2871 data.nr_tags = plug->nr_ios;
2873 data.cached_rq = &plug->cached_rq;
2876 rq = __blk_mq_alloc_requests(&data);
2879 rq_qos_cleanup(q, bio);
2880 if (bio->bi_opf & REQ_NOWAIT)
2881 bio_wouldblock_error(bio);
2887 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2888 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2891 enum hctx_type type, hctx_type;
2896 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2901 rq = rq_list_peek(&plug->cached_rq);
2902 if (!rq || rq->q != q)
2905 type = blk_mq_get_hctx_type((*bio)->bi_opf);
2906 hctx_type = rq->mq_hctx->type;
2907 if (type != hctx_type &&
2908 !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2910 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2914 * If any qos ->throttle() end up blocking, we will have flushed the
2915 * plug and hence killed the cached_rq list as well. Pop this entry
2916 * before we throttle.
2918 plug->cached_rq = rq_list_next(rq);
2919 rq_qos_throttle(q, *bio);
2921 rq->cmd_flags = (*bio)->bi_opf;
2922 INIT_LIST_HEAD(&rq->queuelist);
2926 static void bio_set_ioprio(struct bio *bio)
2928 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2929 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2930 bio->bi_ioprio = get_current_ioprio();
2931 blkcg_set_ioprio(bio);
2935 * blk_mq_submit_bio - Create and send a request to block device.
2936 * @bio: Bio pointer.
2938 * Builds up a request structure from @q and @bio and send to the device. The
2939 * request may not be queued directly to hardware if:
2940 * * This request can be merged with another one
2941 * * We want to place request at plug queue for possible future merging
2942 * * There is an IO scheduler active at this queue
2944 * It will not queue the request if there is an error with the bio, or at the
2947 void blk_mq_submit_bio(struct bio *bio)
2949 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2950 struct blk_plug *plug = blk_mq_plug(bio);
2951 const int is_sync = op_is_sync(bio->bi_opf);
2953 unsigned int nr_segs = 1;
2956 bio = blk_queue_bounce(bio, q);
2957 if (bio_may_exceed_limits(bio, &q->limits)) {
2958 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2963 if (!bio_integrity_prep(bio))
2966 bio_set_ioprio(bio);
2968 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2972 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2977 trace_block_getrq(bio);
2979 rq_qos_track(q, rq, bio);
2981 blk_mq_bio_to_request(rq, bio, nr_segs);
2983 ret = blk_crypto_rq_get_keyslot(rq);
2984 if (ret != BLK_STS_OK) {
2985 bio->bi_status = ret;
2987 blk_mq_free_request(rq);
2991 if (op_is_flush(bio->bi_opf)) {
2992 blk_insert_flush(rq);
2997 blk_add_rq_to_plug(plug, rq);
2998 else if ((rq->rq_flags & RQF_ELV) ||
2999 (rq->mq_hctx->dispatch_busy &&
3000 (q->nr_hw_queues == 1 || !is_sync)))
3001 blk_mq_sched_insert_request(rq, false, true, true);
3003 blk_mq_run_dispatch_ops(rq->q,
3004 blk_mq_try_issue_directly(rq->mq_hctx, rq));
3007 #ifdef CONFIG_BLK_MQ_STACKING
3009 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3010 * @rq: the request being queued
3012 blk_status_t blk_insert_cloned_request(struct request *rq)
3014 struct request_queue *q = rq->q;
3015 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3016 unsigned int max_segments = blk_rq_get_max_segments(rq);
3019 if (blk_rq_sectors(rq) > max_sectors) {
3021 * SCSI device does not have a good way to return if
3022 * Write Same/Zero is actually supported. If a device rejects
3023 * a non-read/write command (discard, write same,etc.) the
3024 * low-level device driver will set the relevant queue limit to
3025 * 0 to prevent blk-lib from issuing more of the offending
3026 * operations. Commands queued prior to the queue limit being
3027 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3028 * errors being propagated to upper layers.
3030 if (max_sectors == 0)
3031 return BLK_STS_NOTSUPP;
3033 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3034 __func__, blk_rq_sectors(rq), max_sectors);
3035 return BLK_STS_IOERR;
3039 * The queue settings related to segment counting may differ from the
3042 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3043 if (rq->nr_phys_segments > max_segments) {
3044 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3045 __func__, rq->nr_phys_segments, max_segments);
3046 return BLK_STS_IOERR;
3049 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3050 return BLK_STS_IOERR;
3052 ret = blk_crypto_rq_get_keyslot(rq);
3053 if (ret != BLK_STS_OK)
3056 blk_account_io_start(rq);
3059 * Since we have a scheduler attached on the top device,
3060 * bypass a potential scheduler on the bottom device for
3063 blk_mq_run_dispatch_ops(q,
3064 ret = blk_mq_request_issue_directly(rq, true));
3066 blk_account_io_done(rq, ktime_get_ns());
3069 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3072 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3073 * @rq: the clone request to be cleaned up
3076 * Free all bios in @rq for a cloned request.
3078 void blk_rq_unprep_clone(struct request *rq)
3082 while ((bio = rq->bio) != NULL) {
3083 rq->bio = bio->bi_next;
3088 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3091 * blk_rq_prep_clone - Helper function to setup clone request
3092 * @rq: the request to be setup
3093 * @rq_src: original request to be cloned
3094 * @bs: bio_set that bios for clone are allocated from
3095 * @gfp_mask: memory allocation mask for bio
3096 * @bio_ctr: setup function to be called for each clone bio.
3097 * Returns %0 for success, non %0 for failure.
3098 * @data: private data to be passed to @bio_ctr
3101 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3102 * Also, pages which the original bios are pointing to are not copied
3103 * and the cloned bios just point same pages.
3104 * So cloned bios must be completed before original bios, which means
3105 * the caller must complete @rq before @rq_src.
3107 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3108 struct bio_set *bs, gfp_t gfp_mask,
3109 int (*bio_ctr)(struct bio *, struct bio *, void *),
3112 struct bio *bio, *bio_src;
3117 __rq_for_each_bio(bio_src, rq_src) {
3118 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3123 if (bio_ctr && bio_ctr(bio, bio_src, data))
3127 rq->biotail->bi_next = bio;
3130 rq->bio = rq->biotail = bio;
3135 /* Copy attributes of the original request to the clone request. */
3136 rq->__sector = blk_rq_pos(rq_src);
3137 rq->__data_len = blk_rq_bytes(rq_src);
3138 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3139 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3140 rq->special_vec = rq_src->special_vec;
3142 rq->nr_phys_segments = rq_src->nr_phys_segments;
3143 rq->ioprio = rq_src->ioprio;
3145 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3153 blk_rq_unprep_clone(rq);
3157 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3158 #endif /* CONFIG_BLK_MQ_STACKING */
3161 * Steal bios from a request and add them to a bio list.
3162 * The request must not have been partially completed before.
3164 void blk_steal_bios(struct bio_list *list, struct request *rq)
3168 list->tail->bi_next = rq->bio;
3170 list->head = rq->bio;
3171 list->tail = rq->biotail;
3179 EXPORT_SYMBOL_GPL(blk_steal_bios);
3181 static size_t order_to_size(unsigned int order)
3183 return (size_t)PAGE_SIZE << order;
3186 /* called before freeing request pool in @tags */
3187 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3188 struct blk_mq_tags *tags)
3191 unsigned long flags;
3194 * There is no need to clear mapping if driver tags is not initialized
3195 * or the mapping belongs to the driver tags.
3197 if (!drv_tags || drv_tags == tags)
3200 list_for_each_entry(page, &tags->page_list, lru) {
3201 unsigned long start = (unsigned long)page_address(page);
3202 unsigned long end = start + order_to_size(page->private);
3205 for (i = 0; i < drv_tags->nr_tags; i++) {
3206 struct request *rq = drv_tags->rqs[i];
3207 unsigned long rq_addr = (unsigned long)rq;
3209 if (rq_addr >= start && rq_addr < end) {
3210 WARN_ON_ONCE(req_ref_read(rq) != 0);
3211 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3217 * Wait until all pending iteration is done.
3219 * Request reference is cleared and it is guaranteed to be observed
3220 * after the ->lock is released.
3222 spin_lock_irqsave(&drv_tags->lock, flags);
3223 spin_unlock_irqrestore(&drv_tags->lock, flags);
3226 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3227 unsigned int hctx_idx)
3229 struct blk_mq_tags *drv_tags;
3232 if (list_empty(&tags->page_list))
3235 if (blk_mq_is_shared_tags(set->flags))
3236 drv_tags = set->shared_tags;
3238 drv_tags = set->tags[hctx_idx];
3240 if (tags->static_rqs && set->ops->exit_request) {
3243 for (i = 0; i < tags->nr_tags; i++) {
3244 struct request *rq = tags->static_rqs[i];
3248 set->ops->exit_request(set, rq, hctx_idx);
3249 tags->static_rqs[i] = NULL;
3253 blk_mq_clear_rq_mapping(drv_tags, tags);
3255 while (!list_empty(&tags->page_list)) {
3256 page = list_first_entry(&tags->page_list, struct page, lru);
3257 list_del_init(&page->lru);
3259 * Remove kmemleak object previously allocated in
3260 * blk_mq_alloc_rqs().
3262 kmemleak_free(page_address(page));
3263 __free_pages(page, page->private);
3267 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3271 kfree(tags->static_rqs);
3272 tags->static_rqs = NULL;
3274 blk_mq_free_tags(tags);
3277 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3278 unsigned int hctx_idx)
3282 for (i = 0; i < set->nr_maps; i++) {
3283 unsigned int start = set->map[i].queue_offset;
3284 unsigned int end = start + set->map[i].nr_queues;
3286 if (hctx_idx >= start && hctx_idx < end)
3290 if (i >= set->nr_maps)
3291 i = HCTX_TYPE_DEFAULT;
3296 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3297 unsigned int hctx_idx)
3299 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3301 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3304 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3305 unsigned int hctx_idx,
3306 unsigned int nr_tags,
3307 unsigned int reserved_tags)
3309 int node = blk_mq_get_hctx_node(set, hctx_idx);
3310 struct blk_mq_tags *tags;
3312 if (node == NUMA_NO_NODE)
3313 node = set->numa_node;
3315 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3316 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3320 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3321 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3326 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3327 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3329 if (!tags->static_rqs)
3337 blk_mq_free_tags(tags);
3341 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3342 unsigned int hctx_idx, int node)
3346 if (set->ops->init_request) {
3347 ret = set->ops->init_request(set, rq, hctx_idx, node);
3352 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3356 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3357 struct blk_mq_tags *tags,
3358 unsigned int hctx_idx, unsigned int depth)
3360 unsigned int i, j, entries_per_page, max_order = 4;
3361 int node = blk_mq_get_hctx_node(set, hctx_idx);
3362 size_t rq_size, left;
3364 if (node == NUMA_NO_NODE)
3365 node = set->numa_node;
3367 INIT_LIST_HEAD(&tags->page_list);
3370 * rq_size is the size of the request plus driver payload, rounded
3371 * to the cacheline size
3373 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3375 left = rq_size * depth;
3377 for (i = 0; i < depth; ) {
3378 int this_order = max_order;
3383 while (this_order && left < order_to_size(this_order - 1))
3387 page = alloc_pages_node(node,
3388 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3394 if (order_to_size(this_order) < rq_size)
3401 page->private = this_order;
3402 list_add_tail(&page->lru, &tags->page_list);
3404 p = page_address(page);
3406 * Allow kmemleak to scan these pages as they contain pointers
3407 * to additional allocations like via ops->init_request().
3409 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3410 entries_per_page = order_to_size(this_order) / rq_size;
3411 to_do = min(entries_per_page, depth - i);
3412 left -= to_do * rq_size;
3413 for (j = 0; j < to_do; j++) {
3414 struct request *rq = p;
3416 tags->static_rqs[i] = rq;
3417 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3418 tags->static_rqs[i] = NULL;
3429 blk_mq_free_rqs(set, tags, hctx_idx);
3433 struct rq_iter_data {
3434 struct blk_mq_hw_ctx *hctx;
3438 static bool blk_mq_has_request(struct request *rq, void *data)
3440 struct rq_iter_data *iter_data = data;
3442 if (rq->mq_hctx != iter_data->hctx)
3444 iter_data->has_rq = true;
3448 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3450 struct blk_mq_tags *tags = hctx->sched_tags ?
3451 hctx->sched_tags : hctx->tags;
3452 struct rq_iter_data data = {
3456 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3460 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3461 struct blk_mq_hw_ctx *hctx)
3463 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3465 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3470 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3472 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3473 struct blk_mq_hw_ctx, cpuhp_online);
3475 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3476 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3480 * Prevent new request from being allocated on the current hctx.
3482 * The smp_mb__after_atomic() Pairs with the implied barrier in
3483 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3484 * seen once we return from the tag allocator.
3486 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3487 smp_mb__after_atomic();
3490 * Try to grab a reference to the queue and wait for any outstanding
3491 * requests. If we could not grab a reference the queue has been
3492 * frozen and there are no requests.
3494 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3495 while (blk_mq_hctx_has_requests(hctx))
3497 percpu_ref_put(&hctx->queue->q_usage_counter);
3503 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3505 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3506 struct blk_mq_hw_ctx, cpuhp_online);
3508 if (cpumask_test_cpu(cpu, hctx->cpumask))
3509 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3514 * 'cpu' is going away. splice any existing rq_list entries from this
3515 * software queue to the hw queue dispatch list, and ensure that it
3518 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3520 struct blk_mq_hw_ctx *hctx;
3521 struct blk_mq_ctx *ctx;
3523 enum hctx_type type;
3525 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3526 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3529 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3532 spin_lock(&ctx->lock);
3533 if (!list_empty(&ctx->rq_lists[type])) {
3534 list_splice_init(&ctx->rq_lists[type], &tmp);
3535 blk_mq_hctx_clear_pending(hctx, ctx);
3537 spin_unlock(&ctx->lock);
3539 if (list_empty(&tmp))
3542 spin_lock(&hctx->lock);
3543 list_splice_tail_init(&tmp, &hctx->dispatch);
3544 spin_unlock(&hctx->lock);
3546 blk_mq_run_hw_queue(hctx, true);
3550 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3552 if (!(hctx->flags & BLK_MQ_F_STACKING))
3553 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3554 &hctx->cpuhp_online);
3555 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3560 * Before freeing hw queue, clearing the flush request reference in
3561 * tags->rqs[] for avoiding potential UAF.
3563 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3564 unsigned int queue_depth, struct request *flush_rq)
3567 unsigned long flags;
3569 /* The hw queue may not be mapped yet */
3573 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3575 for (i = 0; i < queue_depth; i++)
3576 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3579 * Wait until all pending iteration is done.
3581 * Request reference is cleared and it is guaranteed to be observed
3582 * after the ->lock is released.
3584 spin_lock_irqsave(&tags->lock, flags);
3585 spin_unlock_irqrestore(&tags->lock, flags);
3588 /* hctx->ctxs will be freed in queue's release handler */
3589 static void blk_mq_exit_hctx(struct request_queue *q,
3590 struct blk_mq_tag_set *set,
3591 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3593 struct request *flush_rq = hctx->fq->flush_rq;
3595 if (blk_mq_hw_queue_mapped(hctx))
3596 blk_mq_tag_idle(hctx);
3598 if (blk_queue_init_done(q))
3599 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3600 set->queue_depth, flush_rq);
3601 if (set->ops->exit_request)
3602 set->ops->exit_request(set, flush_rq, hctx_idx);
3604 if (set->ops->exit_hctx)
3605 set->ops->exit_hctx(hctx, hctx_idx);
3607 blk_mq_remove_cpuhp(hctx);
3609 xa_erase(&q->hctx_table, hctx_idx);
3611 spin_lock(&q->unused_hctx_lock);
3612 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3613 spin_unlock(&q->unused_hctx_lock);
3616 static void blk_mq_exit_hw_queues(struct request_queue *q,
3617 struct blk_mq_tag_set *set, int nr_queue)
3619 struct blk_mq_hw_ctx *hctx;
3622 queue_for_each_hw_ctx(q, hctx, i) {
3625 blk_mq_exit_hctx(q, set, hctx, i);
3629 static int blk_mq_init_hctx(struct request_queue *q,
3630 struct blk_mq_tag_set *set,
3631 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3633 hctx->queue_num = hctx_idx;
3635 if (!(hctx->flags & BLK_MQ_F_STACKING))
3636 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3637 &hctx->cpuhp_online);
3638 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3640 hctx->tags = set->tags[hctx_idx];
3642 if (set->ops->init_hctx &&
3643 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3644 goto unregister_cpu_notifier;
3646 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3650 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3656 if (set->ops->exit_request)
3657 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3659 if (set->ops->exit_hctx)
3660 set->ops->exit_hctx(hctx, hctx_idx);
3661 unregister_cpu_notifier:
3662 blk_mq_remove_cpuhp(hctx);
3666 static struct blk_mq_hw_ctx *
3667 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3670 struct blk_mq_hw_ctx *hctx;
3671 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3673 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3675 goto fail_alloc_hctx;
3677 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3680 atomic_set(&hctx->nr_active, 0);
3681 if (node == NUMA_NO_NODE)
3682 node = set->numa_node;
3683 hctx->numa_node = node;
3685 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3686 spin_lock_init(&hctx->lock);
3687 INIT_LIST_HEAD(&hctx->dispatch);
3689 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3691 INIT_LIST_HEAD(&hctx->hctx_list);
3694 * Allocate space for all possible cpus to avoid allocation at
3697 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3702 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3703 gfp, node, false, false))
3707 spin_lock_init(&hctx->dispatch_wait_lock);
3708 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3709 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3711 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3715 blk_mq_hctx_kobj_init(hctx);
3720 sbitmap_free(&hctx->ctx_map);
3724 free_cpumask_var(hctx->cpumask);
3731 static void blk_mq_init_cpu_queues(struct request_queue *q,
3732 unsigned int nr_hw_queues)
3734 struct blk_mq_tag_set *set = q->tag_set;
3737 for_each_possible_cpu(i) {
3738 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3739 struct blk_mq_hw_ctx *hctx;
3743 spin_lock_init(&__ctx->lock);
3744 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3745 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3750 * Set local node, IFF we have more than one hw queue. If
3751 * not, we remain on the home node of the device
3753 for (j = 0; j < set->nr_maps; j++) {
3754 hctx = blk_mq_map_queue_type(q, j, i);
3755 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3756 hctx->numa_node = cpu_to_node(i);
3761 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3762 unsigned int hctx_idx,
3765 struct blk_mq_tags *tags;
3768 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3772 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3774 blk_mq_free_rq_map(tags);
3781 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3784 if (blk_mq_is_shared_tags(set->flags)) {
3785 set->tags[hctx_idx] = set->shared_tags;
3790 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3793 return set->tags[hctx_idx];
3796 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3797 struct blk_mq_tags *tags,
3798 unsigned int hctx_idx)
3801 blk_mq_free_rqs(set, tags, hctx_idx);
3802 blk_mq_free_rq_map(tags);
3806 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3807 unsigned int hctx_idx)
3809 if (!blk_mq_is_shared_tags(set->flags))
3810 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3812 set->tags[hctx_idx] = NULL;
3815 static void blk_mq_map_swqueue(struct request_queue *q)
3817 unsigned int j, hctx_idx;
3819 struct blk_mq_hw_ctx *hctx;
3820 struct blk_mq_ctx *ctx;
3821 struct blk_mq_tag_set *set = q->tag_set;
3823 queue_for_each_hw_ctx(q, hctx, i) {
3824 cpumask_clear(hctx->cpumask);
3826 hctx->dispatch_from = NULL;
3830 * Map software to hardware queues.
3832 * If the cpu isn't present, the cpu is mapped to first hctx.
3834 for_each_possible_cpu(i) {
3836 ctx = per_cpu_ptr(q->queue_ctx, i);
3837 for (j = 0; j < set->nr_maps; j++) {
3838 if (!set->map[j].nr_queues) {
3839 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3840 HCTX_TYPE_DEFAULT, i);
3843 hctx_idx = set->map[j].mq_map[i];
3844 /* unmapped hw queue can be remapped after CPU topo changed */
3845 if (!set->tags[hctx_idx] &&
3846 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3848 * If tags initialization fail for some hctx,
3849 * that hctx won't be brought online. In this
3850 * case, remap the current ctx to hctx[0] which
3851 * is guaranteed to always have tags allocated
3853 set->map[j].mq_map[i] = 0;
3856 hctx = blk_mq_map_queue_type(q, j, i);
3857 ctx->hctxs[j] = hctx;
3859 * If the CPU is already set in the mask, then we've
3860 * mapped this one already. This can happen if
3861 * devices share queues across queue maps.
3863 if (cpumask_test_cpu(i, hctx->cpumask))
3866 cpumask_set_cpu(i, hctx->cpumask);
3868 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3869 hctx->ctxs[hctx->nr_ctx++] = ctx;
3872 * If the nr_ctx type overflows, we have exceeded the
3873 * amount of sw queues we can support.
3875 BUG_ON(!hctx->nr_ctx);
3878 for (; j < HCTX_MAX_TYPES; j++)
3879 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3880 HCTX_TYPE_DEFAULT, i);
3883 queue_for_each_hw_ctx(q, hctx, i) {
3885 * If no software queues are mapped to this hardware queue,
3886 * disable it and free the request entries.
3888 if (!hctx->nr_ctx) {
3889 /* Never unmap queue 0. We need it as a
3890 * fallback in case of a new remap fails
3894 __blk_mq_free_map_and_rqs(set, i);
3900 hctx->tags = set->tags[i];
3901 WARN_ON(!hctx->tags);
3904 * Set the map size to the number of mapped software queues.
3905 * This is more accurate and more efficient than looping
3906 * over all possibly mapped software queues.
3908 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3911 * Initialize batch roundrobin counts
3913 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3914 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3919 * Caller needs to ensure that we're either frozen/quiesced, or that
3920 * the queue isn't live yet.
3922 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3924 struct blk_mq_hw_ctx *hctx;
3927 queue_for_each_hw_ctx(q, hctx, i) {
3929 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3931 blk_mq_tag_idle(hctx);
3932 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3937 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3940 struct request_queue *q;
3942 lockdep_assert_held(&set->tag_list_lock);
3944 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3945 blk_mq_freeze_queue(q);
3946 queue_set_hctx_shared(q, shared);
3947 blk_mq_unfreeze_queue(q);
3951 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3953 struct blk_mq_tag_set *set = q->tag_set;
3955 mutex_lock(&set->tag_list_lock);
3956 list_del(&q->tag_set_list);
3957 if (list_is_singular(&set->tag_list)) {
3958 /* just transitioned to unshared */
3959 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3960 /* update existing queue */
3961 blk_mq_update_tag_set_shared(set, false);
3963 mutex_unlock(&set->tag_list_lock);
3964 INIT_LIST_HEAD(&q->tag_set_list);
3967 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3968 struct request_queue *q)
3970 mutex_lock(&set->tag_list_lock);
3973 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3975 if (!list_empty(&set->tag_list) &&
3976 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3977 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3978 /* update existing queue */
3979 blk_mq_update_tag_set_shared(set, true);
3981 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3982 queue_set_hctx_shared(q, true);
3983 list_add_tail(&q->tag_set_list, &set->tag_list);
3985 mutex_unlock(&set->tag_list_lock);
3988 /* All allocations will be freed in release handler of q->mq_kobj */
3989 static int blk_mq_alloc_ctxs(struct request_queue *q)
3991 struct blk_mq_ctxs *ctxs;
3994 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3998 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3999 if (!ctxs->queue_ctx)
4002 for_each_possible_cpu(cpu) {
4003 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4007 q->mq_kobj = &ctxs->kobj;
4008 q->queue_ctx = ctxs->queue_ctx;
4017 * It is the actual release handler for mq, but we do it from
4018 * request queue's release handler for avoiding use-after-free
4019 * and headache because q->mq_kobj shouldn't have been introduced,
4020 * but we can't group ctx/kctx kobj without it.
4022 void blk_mq_release(struct request_queue *q)
4024 struct blk_mq_hw_ctx *hctx, *next;
4027 queue_for_each_hw_ctx(q, hctx, i)
4028 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4030 /* all hctx are in .unused_hctx_list now */
4031 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4032 list_del_init(&hctx->hctx_list);
4033 kobject_put(&hctx->kobj);
4036 xa_destroy(&q->hctx_table);
4039 * release .mq_kobj and sw queue's kobject now because
4040 * both share lifetime with request queue.
4042 blk_mq_sysfs_deinit(q);
4045 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4048 struct request_queue *q;
4051 q = blk_alloc_queue(set->numa_node);
4053 return ERR_PTR(-ENOMEM);
4054 q->queuedata = queuedata;
4055 ret = blk_mq_init_allocated_queue(set, q);
4058 return ERR_PTR(ret);
4063 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4065 return blk_mq_init_queue_data(set, NULL);
4067 EXPORT_SYMBOL(blk_mq_init_queue);
4070 * blk_mq_destroy_queue - shutdown a request queue
4071 * @q: request queue to shutdown
4073 * This shuts down a request queue allocated by blk_mq_init_queue(). All future
4074 * requests will be failed with -ENODEV. The caller is responsible for dropping
4075 * the reference from blk_mq_init_queue() by calling blk_put_queue().
4077 * Context: can sleep
4079 void blk_mq_destroy_queue(struct request_queue *q)
4081 WARN_ON_ONCE(!queue_is_mq(q));
4082 WARN_ON_ONCE(blk_queue_registered(q));
4086 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4087 blk_queue_start_drain(q);
4088 blk_mq_freeze_queue_wait(q);
4091 blk_mq_cancel_work_sync(q);
4092 blk_mq_exit_queue(q);
4094 EXPORT_SYMBOL(blk_mq_destroy_queue);
4096 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4097 struct lock_class_key *lkclass)
4099 struct request_queue *q;
4100 struct gendisk *disk;
4102 q = blk_mq_init_queue_data(set, queuedata);
4106 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4108 blk_mq_destroy_queue(q);
4110 return ERR_PTR(-ENOMEM);
4112 set_bit(GD_OWNS_QUEUE, &disk->state);
4115 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4117 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4118 struct lock_class_key *lkclass)
4120 struct gendisk *disk;
4122 if (!blk_get_queue(q))
4124 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4129 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4131 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4132 struct blk_mq_tag_set *set, struct request_queue *q,
4133 int hctx_idx, int node)
4135 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4137 /* reuse dead hctx first */
4138 spin_lock(&q->unused_hctx_lock);
4139 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4140 if (tmp->numa_node == node) {
4146 list_del_init(&hctx->hctx_list);
4147 spin_unlock(&q->unused_hctx_lock);
4150 hctx = blk_mq_alloc_hctx(q, set, node);
4154 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4160 kobject_put(&hctx->kobj);
4165 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4166 struct request_queue *q)
4168 struct blk_mq_hw_ctx *hctx;
4171 /* protect against switching io scheduler */
4172 mutex_lock(&q->sysfs_lock);
4173 for (i = 0; i < set->nr_hw_queues; i++) {
4175 int node = blk_mq_get_hctx_node(set, i);
4176 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4179 old_node = old_hctx->numa_node;
4180 blk_mq_exit_hctx(q, set, old_hctx, i);
4183 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4186 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4188 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4189 WARN_ON_ONCE(!hctx);
4193 * Increasing nr_hw_queues fails. Free the newly allocated
4194 * hctxs and keep the previous q->nr_hw_queues.
4196 if (i != set->nr_hw_queues) {
4197 j = q->nr_hw_queues;
4200 q->nr_hw_queues = set->nr_hw_queues;
4203 xa_for_each_start(&q->hctx_table, j, hctx, j)
4204 blk_mq_exit_hctx(q, set, hctx, j);
4205 mutex_unlock(&q->sysfs_lock);
4208 static void blk_mq_update_poll_flag(struct request_queue *q)
4210 struct blk_mq_tag_set *set = q->tag_set;
4212 if (set->nr_maps > HCTX_TYPE_POLL &&
4213 set->map[HCTX_TYPE_POLL].nr_queues)
4214 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4216 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4219 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4220 struct request_queue *q)
4222 /* mark the queue as mq asap */
4223 q->mq_ops = set->ops;
4225 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
4226 blk_mq_poll_stats_bkt,
4227 BLK_MQ_POLL_STATS_BKTS, q);
4231 if (blk_mq_alloc_ctxs(q))
4234 /* init q->mq_kobj and sw queues' kobjects */
4235 blk_mq_sysfs_init(q);
4237 INIT_LIST_HEAD(&q->unused_hctx_list);
4238 spin_lock_init(&q->unused_hctx_lock);
4240 xa_init(&q->hctx_table);
4242 blk_mq_realloc_hw_ctxs(set, q);
4243 if (!q->nr_hw_queues)
4246 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4247 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4251 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4252 blk_mq_update_poll_flag(q);
4254 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4255 INIT_LIST_HEAD(&q->requeue_list);
4256 spin_lock_init(&q->requeue_lock);
4258 q->nr_requests = set->queue_depth;
4261 * Default to classic polling
4263 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
4265 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4266 blk_mq_add_queue_tag_set(set, q);
4267 blk_mq_map_swqueue(q);
4273 blk_stat_free_callback(q->poll_cb);
4279 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4281 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4282 void blk_mq_exit_queue(struct request_queue *q)
4284 struct blk_mq_tag_set *set = q->tag_set;
4286 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4287 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4288 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4289 blk_mq_del_queue_tag_set(q);
4292 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4296 if (blk_mq_is_shared_tags(set->flags)) {
4297 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4300 if (!set->shared_tags)
4304 for (i = 0; i < set->nr_hw_queues; i++) {
4305 if (!__blk_mq_alloc_map_and_rqs(set, i))
4314 __blk_mq_free_map_and_rqs(set, i);
4316 if (blk_mq_is_shared_tags(set->flags)) {
4317 blk_mq_free_map_and_rqs(set, set->shared_tags,
4318 BLK_MQ_NO_HCTX_IDX);
4325 * Allocate the request maps associated with this tag_set. Note that this
4326 * may reduce the depth asked for, if memory is tight. set->queue_depth
4327 * will be updated to reflect the allocated depth.
4329 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4334 depth = set->queue_depth;
4336 err = __blk_mq_alloc_rq_maps(set);
4340 set->queue_depth >>= 1;
4341 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4345 } while (set->queue_depth);
4347 if (!set->queue_depth || err) {
4348 pr_err("blk-mq: failed to allocate request map\n");
4352 if (depth != set->queue_depth)
4353 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4354 depth, set->queue_depth);
4359 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4362 * blk_mq_map_queues() and multiple .map_queues() implementations
4363 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4364 * number of hardware queues.
4366 if (set->nr_maps == 1)
4367 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4369 if (set->ops->map_queues && !is_kdump_kernel()) {
4373 * transport .map_queues is usually done in the following
4376 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4377 * mask = get_cpu_mask(queue)
4378 * for_each_cpu(cpu, mask)
4379 * set->map[x].mq_map[cpu] = queue;
4382 * When we need to remap, the table has to be cleared for
4383 * killing stale mapping since one CPU may not be mapped
4386 for (i = 0; i < set->nr_maps; i++)
4387 blk_mq_clear_mq_map(&set->map[i]);
4389 set->ops->map_queues(set);
4391 BUG_ON(set->nr_maps > 1);
4392 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4396 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4397 int new_nr_hw_queues)
4399 struct blk_mq_tags **new_tags;
4401 if (set->nr_hw_queues >= new_nr_hw_queues)
4404 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4405 GFP_KERNEL, set->numa_node);
4410 memcpy(new_tags, set->tags, set->nr_hw_queues *
4411 sizeof(*set->tags));
4413 set->tags = new_tags;
4415 set->nr_hw_queues = new_nr_hw_queues;
4420 * Alloc a tag set to be associated with one or more request queues.
4421 * May fail with EINVAL for various error conditions. May adjust the
4422 * requested depth down, if it's too large. In that case, the set
4423 * value will be stored in set->queue_depth.
4425 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4429 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4431 if (!set->nr_hw_queues)
4433 if (!set->queue_depth)
4435 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4438 if (!set->ops->queue_rq)
4441 if (!set->ops->get_budget ^ !set->ops->put_budget)
4444 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4445 pr_info("blk-mq: reduced tag depth to %u\n",
4447 set->queue_depth = BLK_MQ_MAX_DEPTH;
4452 else if (set->nr_maps > HCTX_MAX_TYPES)
4456 * If a crashdump is active, then we are potentially in a very
4457 * memory constrained environment. Limit us to 1 queue and
4458 * 64 tags to prevent using too much memory.
4460 if (is_kdump_kernel()) {
4461 set->nr_hw_queues = 1;
4463 set->queue_depth = min(64U, set->queue_depth);
4466 * There is no use for more h/w queues than cpus if we just have
4469 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4470 set->nr_hw_queues = nr_cpu_ids;
4472 if (set->flags & BLK_MQ_F_BLOCKING) {
4473 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4476 ret = init_srcu_struct(set->srcu);
4482 set->tags = kcalloc_node(set->nr_hw_queues,
4483 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4486 goto out_cleanup_srcu;
4488 for (i = 0; i < set->nr_maps; i++) {
4489 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4490 sizeof(set->map[i].mq_map[0]),
4491 GFP_KERNEL, set->numa_node);
4492 if (!set->map[i].mq_map)
4493 goto out_free_mq_map;
4494 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4497 blk_mq_update_queue_map(set);
4499 ret = blk_mq_alloc_set_map_and_rqs(set);
4501 goto out_free_mq_map;
4503 mutex_init(&set->tag_list_lock);
4504 INIT_LIST_HEAD(&set->tag_list);
4509 for (i = 0; i < set->nr_maps; i++) {
4510 kfree(set->map[i].mq_map);
4511 set->map[i].mq_map = NULL;
4516 if (set->flags & BLK_MQ_F_BLOCKING)
4517 cleanup_srcu_struct(set->srcu);
4519 if (set->flags & BLK_MQ_F_BLOCKING)
4523 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4525 /* allocate and initialize a tagset for a simple single-queue device */
4526 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4527 const struct blk_mq_ops *ops, unsigned int queue_depth,
4528 unsigned int set_flags)
4530 memset(set, 0, sizeof(*set));
4532 set->nr_hw_queues = 1;
4534 set->queue_depth = queue_depth;
4535 set->numa_node = NUMA_NO_NODE;
4536 set->flags = set_flags;
4537 return blk_mq_alloc_tag_set(set);
4539 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4541 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4545 for (i = 0; i < set->nr_hw_queues; i++)
4546 __blk_mq_free_map_and_rqs(set, i);
4548 if (blk_mq_is_shared_tags(set->flags)) {
4549 blk_mq_free_map_and_rqs(set, set->shared_tags,
4550 BLK_MQ_NO_HCTX_IDX);
4553 for (j = 0; j < set->nr_maps; j++) {
4554 kfree(set->map[j].mq_map);
4555 set->map[j].mq_map = NULL;
4560 if (set->flags & BLK_MQ_F_BLOCKING) {
4561 cleanup_srcu_struct(set->srcu);
4565 EXPORT_SYMBOL(blk_mq_free_tag_set);
4567 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4569 struct blk_mq_tag_set *set = q->tag_set;
4570 struct blk_mq_hw_ctx *hctx;
4577 if (q->nr_requests == nr)
4580 blk_mq_freeze_queue(q);
4581 blk_mq_quiesce_queue(q);
4584 queue_for_each_hw_ctx(q, hctx, i) {
4588 * If we're using an MQ scheduler, just update the scheduler
4589 * queue depth. This is similar to what the old code would do.
4591 if (hctx->sched_tags) {
4592 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4595 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4600 if (q->elevator && q->elevator->type->ops.depth_updated)
4601 q->elevator->type->ops.depth_updated(hctx);
4604 q->nr_requests = nr;
4605 if (blk_mq_is_shared_tags(set->flags)) {
4607 blk_mq_tag_update_sched_shared_tags(q);
4609 blk_mq_tag_resize_shared_tags(set, nr);
4613 blk_mq_unquiesce_queue(q);
4614 blk_mq_unfreeze_queue(q);
4620 * request_queue and elevator_type pair.
4621 * It is just used by __blk_mq_update_nr_hw_queues to cache
4622 * the elevator_type associated with a request_queue.
4624 struct blk_mq_qe_pair {
4625 struct list_head node;
4626 struct request_queue *q;
4627 struct elevator_type *type;
4631 * Cache the elevator_type in qe pair list and switch the
4632 * io scheduler to 'none'
4634 static bool blk_mq_elv_switch_none(struct list_head *head,
4635 struct request_queue *q)
4637 struct blk_mq_qe_pair *qe;
4642 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4646 /* q->elevator needs protection from ->sysfs_lock */
4647 mutex_lock(&q->sysfs_lock);
4649 INIT_LIST_HEAD(&qe->node);
4651 qe->type = q->elevator->type;
4652 /* keep a reference to the elevator module as we'll switch back */
4653 __elevator_get(qe->type);
4654 list_add(&qe->node, head);
4655 elevator_disable(q);
4656 mutex_unlock(&q->sysfs_lock);
4661 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4662 struct request_queue *q)
4664 struct blk_mq_qe_pair *qe;
4666 list_for_each_entry(qe, head, node)
4673 static void blk_mq_elv_switch_back(struct list_head *head,
4674 struct request_queue *q)
4676 struct blk_mq_qe_pair *qe;
4677 struct elevator_type *t;
4679 qe = blk_lookup_qe_pair(head, q);
4683 list_del(&qe->node);
4686 mutex_lock(&q->sysfs_lock);
4687 elevator_switch(q, t);
4688 /* drop the reference acquired in blk_mq_elv_switch_none */
4690 mutex_unlock(&q->sysfs_lock);
4693 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4696 struct request_queue *q;
4698 int prev_nr_hw_queues;
4700 lockdep_assert_held(&set->tag_list_lock);
4702 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4703 nr_hw_queues = nr_cpu_ids;
4704 if (nr_hw_queues < 1)
4706 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4709 list_for_each_entry(q, &set->tag_list, tag_set_list)
4710 blk_mq_freeze_queue(q);
4712 * Switch IO scheduler to 'none', cleaning up the data associated
4713 * with the previous scheduler. We will switch back once we are done
4714 * updating the new sw to hw queue mappings.
4716 list_for_each_entry(q, &set->tag_list, tag_set_list)
4717 if (!blk_mq_elv_switch_none(&head, q))
4720 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4721 blk_mq_debugfs_unregister_hctxs(q);
4722 blk_mq_sysfs_unregister_hctxs(q);
4725 prev_nr_hw_queues = set->nr_hw_queues;
4726 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4730 blk_mq_update_queue_map(set);
4731 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4732 blk_mq_realloc_hw_ctxs(set, q);
4733 blk_mq_update_poll_flag(q);
4734 if (q->nr_hw_queues != set->nr_hw_queues) {
4735 int i = prev_nr_hw_queues;
4737 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4738 nr_hw_queues, prev_nr_hw_queues);
4739 for (; i < set->nr_hw_queues; i++)
4740 __blk_mq_free_map_and_rqs(set, i);
4742 set->nr_hw_queues = prev_nr_hw_queues;
4743 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4746 blk_mq_map_swqueue(q);
4750 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4751 blk_mq_sysfs_register_hctxs(q);
4752 blk_mq_debugfs_register_hctxs(q);
4756 list_for_each_entry(q, &set->tag_list, tag_set_list)
4757 blk_mq_elv_switch_back(&head, q);
4759 list_for_each_entry(q, &set->tag_list, tag_set_list)
4760 blk_mq_unfreeze_queue(q);
4763 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4765 mutex_lock(&set->tag_list_lock);
4766 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4767 mutex_unlock(&set->tag_list_lock);
4769 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4771 /* Enable polling stats and return whether they were already enabled. */
4772 static bool blk_poll_stats_enable(struct request_queue *q)
4777 return blk_stats_alloc_enable(q);
4780 static void blk_mq_poll_stats_start(struct request_queue *q)
4783 * We don't arm the callback if polling stats are not enabled or the
4784 * callback is already active.
4786 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
4789 blk_stat_activate_msecs(q->poll_cb, 100);
4792 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4794 struct request_queue *q = cb->data;
4797 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4798 if (cb->stat[bucket].nr_samples)
4799 q->poll_stat[bucket] = cb->stat[bucket];
4803 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4806 unsigned long ret = 0;
4810 * If stats collection isn't on, don't sleep but turn it on for
4813 if (!blk_poll_stats_enable(q))
4817 * As an optimistic guess, use half of the mean service time
4818 * for this type of request. We can (and should) make this smarter.
4819 * For instance, if the completion latencies are tight, we can
4820 * get closer than just half the mean. This is especially
4821 * important on devices where the completion latencies are longer
4822 * than ~10 usec. We do use the stats for the relevant IO size
4823 * if available which does lead to better estimates.
4825 bucket = blk_mq_poll_stats_bkt(rq);
4829 if (q->poll_stat[bucket].nr_samples)
4830 ret = (q->poll_stat[bucket].mean + 1) / 2;
4835 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4837 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4838 struct request *rq = blk_qc_to_rq(hctx, qc);
4839 struct hrtimer_sleeper hs;
4840 enum hrtimer_mode mode;
4845 * If a request has completed on queue that uses an I/O scheduler, we
4846 * won't get back a request from blk_qc_to_rq.
4848 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4852 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4854 * 0: use half of prev avg
4855 * >0: use this specific value
4857 if (q->poll_nsec > 0)
4858 nsecs = q->poll_nsec;
4860 nsecs = blk_mq_poll_nsecs(q, rq);
4865 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4868 * This will be replaced with the stats tracking code, using
4869 * 'avg_completion_time / 2' as the pre-sleep target.
4873 mode = HRTIMER_MODE_REL;
4874 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4875 hrtimer_set_expires(&hs.timer, kt);
4878 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4880 set_current_state(TASK_UNINTERRUPTIBLE);
4881 hrtimer_sleeper_start_expires(&hs, mode);
4884 hrtimer_cancel(&hs.timer);
4885 mode = HRTIMER_MODE_ABS;
4886 } while (hs.task && !signal_pending(current));
4888 __set_current_state(TASK_RUNNING);
4889 destroy_hrtimer_on_stack(&hs.timer);
4892 * If we sleep, have the caller restart the poll loop to reset the
4893 * state. Like for the other success return cases, the caller is
4894 * responsible for checking if the IO completed. If the IO isn't
4895 * complete, we'll get called again and will go straight to the busy
4901 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4902 struct io_comp_batch *iob, unsigned int flags)
4904 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4905 long state = get_current_state();
4909 ret = q->mq_ops->poll(hctx, iob);
4911 __set_current_state(TASK_RUNNING);
4915 if (signal_pending_state(state, current))
4916 __set_current_state(TASK_RUNNING);
4917 if (task_is_running(current))
4920 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4923 } while (!need_resched());
4925 __set_current_state(TASK_RUNNING);
4929 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4932 if (!(flags & BLK_POLL_NOSLEEP) &&
4933 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4934 if (blk_mq_poll_hybrid(q, cookie))
4937 return blk_mq_poll_classic(q, cookie, iob, flags);
4940 unsigned int blk_mq_rq_cpu(struct request *rq)
4942 return rq->mq_ctx->cpu;
4944 EXPORT_SYMBOL(blk_mq_rq_cpu);
4946 void blk_mq_cancel_work_sync(struct request_queue *q)
4948 struct blk_mq_hw_ctx *hctx;
4951 cancel_delayed_work_sync(&q->requeue_work);
4953 queue_for_each_hw_ctx(q, hctx, i)
4954 cancel_delayed_work_sync(&hctx->run_work);
4957 static int __init blk_mq_init(void)
4961 for_each_possible_cpu(i)
4962 init_llist_head(&per_cpu(blk_cpu_done, i));
4963 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4965 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4966 "block/softirq:dead", NULL,
4967 blk_softirq_cpu_dead);
4968 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4969 blk_mq_hctx_notify_dead);
4970 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4971 blk_mq_hctx_notify_online,
4972 blk_mq_hctx_notify_offline);
4975 subsys_initcall(blk_mq_init);