1 // SPDX-License-Identifier: GPL-2.0
3 * Block multiqueue core code
5 * Copyright (C) 2013-2014 Jens Axboe
6 * Copyright (C) 2013-2014 Christoph Hellwig
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/backing-dev.h>
11 #include <linux/bio.h>
12 #include <linux/blkdev.h>
13 #include <linux/blk-integrity.h>
14 #include <linux/kmemleak.h>
16 #include <linux/init.h>
17 #include <linux/slab.h>
18 #include <linux/workqueue.h>
19 #include <linux/smp.h>
20 #include <linux/interrupt.h>
21 #include <linux/llist.h>
22 #include <linux/cpu.h>
23 #include <linux/cache.h>
24 #include <linux/sched/sysctl.h>
25 #include <linux/sched/topology.h>
26 #include <linux/sched/signal.h>
27 #include <linux/delay.h>
28 #include <linux/crash_dump.h>
29 #include <linux/prefetch.h>
30 #include <linux/blk-crypto.h>
31 #include <linux/part_stat.h>
33 #include <trace/events/block.h>
35 #include <linux/blk-mq.h>
36 #include <linux/t10-pi.h>
39 #include "blk-mq-debugfs.h"
40 #include "blk-mq-tag.h"
43 #include "blk-mq-sched.h"
44 #include "blk-rq-qos.h"
45 #include "blk-ioprio.h"
47 static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
49 static void blk_mq_poll_stats_start(struct request_queue *q);
50 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
52 static int blk_mq_poll_stats_bkt(const struct request *rq)
54 int ddir, sectors, bucket;
56 ddir = rq_data_dir(rq);
57 sectors = blk_rq_stats_sectors(rq);
59 bucket = ddir + 2 * ilog2(sectors);
63 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
64 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
69 #define BLK_QC_T_SHIFT 16
70 #define BLK_QC_T_INTERNAL (1U << 31)
72 static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q,
75 return xa_load(&q->hctx_table,
76 (qc & ~BLK_QC_T_INTERNAL) >> BLK_QC_T_SHIFT);
79 static inline struct request *blk_qc_to_rq(struct blk_mq_hw_ctx *hctx,
82 unsigned int tag = qc & ((1U << BLK_QC_T_SHIFT) - 1);
84 if (qc & BLK_QC_T_INTERNAL)
85 return blk_mq_tag_to_rq(hctx->sched_tags, tag);
86 return blk_mq_tag_to_rq(hctx->tags, tag);
89 static inline blk_qc_t blk_rq_to_qc(struct request *rq)
91 return (rq->mq_hctx->queue_num << BLK_QC_T_SHIFT) |
93 rq->tag : (rq->internal_tag | BLK_QC_T_INTERNAL));
97 * Check if any of the ctx, dispatch list or elevator
98 * have pending work in this hardware queue.
100 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
102 return !list_empty_careful(&hctx->dispatch) ||
103 sbitmap_any_bit_set(&hctx->ctx_map) ||
104 blk_mq_sched_has_work(hctx);
108 * Mark this ctx as having pending work in this hardware queue
110 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
111 struct blk_mq_ctx *ctx)
113 const int bit = ctx->index_hw[hctx->type];
115 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
116 sbitmap_set_bit(&hctx->ctx_map, bit);
119 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
120 struct blk_mq_ctx *ctx)
122 const int bit = ctx->index_hw[hctx->type];
124 sbitmap_clear_bit(&hctx->ctx_map, bit);
128 struct block_device *part;
129 unsigned int inflight[2];
132 static bool blk_mq_check_inflight(struct request *rq, void *priv)
134 struct mq_inflight *mi = priv;
136 if (rq->part && blk_do_io_stat(rq) &&
137 (!mi->part->bd_partno || rq->part == mi->part) &&
138 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
139 mi->inflight[rq_data_dir(rq)]++;
144 unsigned int blk_mq_in_flight(struct request_queue *q,
145 struct block_device *part)
147 struct mq_inflight mi = { .part = part };
149 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
151 return mi.inflight[0] + mi.inflight[1];
154 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
155 unsigned int inflight[2])
157 struct mq_inflight mi = { .part = part };
159 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
160 inflight[0] = mi.inflight[0];
161 inflight[1] = mi.inflight[1];
164 void blk_freeze_queue_start(struct request_queue *q)
166 mutex_lock(&q->mq_freeze_lock);
167 if (++q->mq_freeze_depth == 1) {
168 percpu_ref_kill(&q->q_usage_counter);
169 mutex_unlock(&q->mq_freeze_lock);
171 blk_mq_run_hw_queues(q, false);
173 mutex_unlock(&q->mq_freeze_lock);
176 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
178 void blk_mq_freeze_queue_wait(struct request_queue *q)
180 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
182 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
184 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
185 unsigned long timeout)
187 return wait_event_timeout(q->mq_freeze_wq,
188 percpu_ref_is_zero(&q->q_usage_counter),
191 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
194 * Guarantee no request is in use, so we can change any data structure of
195 * the queue afterward.
197 void blk_freeze_queue(struct request_queue *q)
200 * In the !blk_mq case we are only calling this to kill the
201 * q_usage_counter, otherwise this increases the freeze depth
202 * and waits for it to return to zero. For this reason there is
203 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
204 * exported to drivers as the only user for unfreeze is blk_mq.
206 blk_freeze_queue_start(q);
207 blk_mq_freeze_queue_wait(q);
210 void blk_mq_freeze_queue(struct request_queue *q)
213 * ...just an alias to keep freeze and unfreeze actions balanced
214 * in the blk_mq_* namespace
218 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
220 void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
222 mutex_lock(&q->mq_freeze_lock);
224 q->q_usage_counter.data->force_atomic = true;
225 q->mq_freeze_depth--;
226 WARN_ON_ONCE(q->mq_freeze_depth < 0);
227 if (!q->mq_freeze_depth) {
228 percpu_ref_resurrect(&q->q_usage_counter);
229 wake_up_all(&q->mq_freeze_wq);
231 mutex_unlock(&q->mq_freeze_lock);
234 void blk_mq_unfreeze_queue(struct request_queue *q)
236 __blk_mq_unfreeze_queue(q, false);
238 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
241 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
242 * mpt3sas driver such that this function can be removed.
244 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
248 spin_lock_irqsave(&q->queue_lock, flags);
249 if (!q->quiesce_depth++)
250 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
251 spin_unlock_irqrestore(&q->queue_lock, flags);
253 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
256 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
259 * Note: it is driver's responsibility for making sure that quiesce has
262 void blk_mq_wait_quiesce_done(struct request_queue *q)
264 if (blk_queue_has_srcu(q))
265 synchronize_srcu(q->srcu);
269 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
272 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
275 * Note: this function does not prevent that the struct request end_io()
276 * callback function is invoked. Once this function is returned, we make
277 * sure no dispatch can happen until the queue is unquiesced via
278 * blk_mq_unquiesce_queue().
280 void blk_mq_quiesce_queue(struct request_queue *q)
282 blk_mq_quiesce_queue_nowait(q);
283 blk_mq_wait_quiesce_done(q);
285 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
288 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
291 * This function recovers queue into the state before quiescing
292 * which is done by blk_mq_quiesce_queue.
294 void blk_mq_unquiesce_queue(struct request_queue *q)
297 bool run_queue = false;
299 spin_lock_irqsave(&q->queue_lock, flags);
300 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
302 } else if (!--q->quiesce_depth) {
303 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
306 spin_unlock_irqrestore(&q->queue_lock, flags);
308 /* dispatch requests which are inserted during quiescing */
310 blk_mq_run_hw_queues(q, true);
312 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
314 void blk_mq_wake_waiters(struct request_queue *q)
316 struct blk_mq_hw_ctx *hctx;
319 queue_for_each_hw_ctx(q, hctx, i)
320 if (blk_mq_hw_queue_mapped(hctx))
321 blk_mq_tag_wakeup_all(hctx->tags, true);
324 void blk_rq_init(struct request_queue *q, struct request *rq)
326 memset(rq, 0, sizeof(*rq));
328 INIT_LIST_HEAD(&rq->queuelist);
330 rq->__sector = (sector_t) -1;
331 INIT_HLIST_NODE(&rq->hash);
332 RB_CLEAR_NODE(&rq->rb_node);
333 rq->tag = BLK_MQ_NO_TAG;
334 rq->internal_tag = BLK_MQ_NO_TAG;
335 rq->start_time_ns = ktime_get_ns();
337 blk_crypto_rq_set_defaults(rq);
339 EXPORT_SYMBOL(blk_rq_init);
341 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
342 struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns)
344 struct blk_mq_ctx *ctx = data->ctx;
345 struct blk_mq_hw_ctx *hctx = data->hctx;
346 struct request_queue *q = data->q;
347 struct request *rq = tags->static_rqs[tag];
352 rq->cmd_flags = data->cmd_flags;
354 if (data->flags & BLK_MQ_REQ_PM)
355 data->rq_flags |= RQF_PM;
356 if (blk_queue_io_stat(q))
357 data->rq_flags |= RQF_IO_STAT;
358 rq->rq_flags = data->rq_flags;
360 if (!(data->rq_flags & RQF_ELV)) {
362 rq->internal_tag = BLK_MQ_NO_TAG;
364 rq->tag = BLK_MQ_NO_TAG;
365 rq->internal_tag = tag;
369 if (blk_mq_need_time_stamp(rq))
370 rq->start_time_ns = ktime_get_ns();
372 rq->start_time_ns = 0;
374 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
375 rq->alloc_time_ns = alloc_time_ns;
377 rq->io_start_time_ns = 0;
378 rq->stats_sectors = 0;
379 rq->nr_phys_segments = 0;
380 #if defined(CONFIG_BLK_DEV_INTEGRITY)
381 rq->nr_integrity_segments = 0;
384 rq->end_io_data = NULL;
386 blk_crypto_rq_set_defaults(rq);
387 INIT_LIST_HEAD(&rq->queuelist);
388 /* tag was already set */
389 WRITE_ONCE(rq->deadline, 0);
392 if (rq->rq_flags & RQF_ELV) {
393 struct elevator_queue *e = data->q->elevator;
395 INIT_HLIST_NODE(&rq->hash);
396 RB_CLEAR_NODE(&rq->rb_node);
398 if (!op_is_flush(data->cmd_flags) &&
399 e->type->ops.prepare_request) {
400 e->type->ops.prepare_request(rq);
401 rq->rq_flags |= RQF_ELVPRIV;
408 static inline struct request *
409 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data,
412 unsigned int tag, tag_offset;
413 struct blk_mq_tags *tags;
415 unsigned long tag_mask;
418 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
419 if (unlikely(!tag_mask))
422 tags = blk_mq_tags_from_data(data);
423 for (i = 0; tag_mask; i++) {
424 if (!(tag_mask & (1UL << i)))
426 tag = tag_offset + i;
427 prefetch(tags->static_rqs[tag]);
428 tag_mask &= ~(1UL << i);
429 rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns);
430 rq_list_add(data->cached_rq, rq);
433 /* caller already holds a reference, add for remainder */
434 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
437 return rq_list_pop(data->cached_rq);
440 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
442 struct request_queue *q = data->q;
443 u64 alloc_time_ns = 0;
447 /* alloc_time includes depth and tag waits */
448 if (blk_queue_rq_alloc_time(q))
449 alloc_time_ns = ktime_get_ns();
451 if (data->cmd_flags & REQ_NOWAIT)
452 data->flags |= BLK_MQ_REQ_NOWAIT;
455 struct elevator_queue *e = q->elevator;
457 data->rq_flags |= RQF_ELV;
460 * Flush/passthrough requests are special and go directly to the
461 * dispatch list. Don't include reserved tags in the
462 * limiting, as it isn't useful.
464 if (!op_is_flush(data->cmd_flags) &&
465 !blk_op_is_passthrough(data->cmd_flags) &&
466 e->type->ops.limit_depth &&
467 !(data->flags & BLK_MQ_REQ_RESERVED))
468 e->type->ops.limit_depth(data->cmd_flags, data);
472 data->ctx = blk_mq_get_ctx(q);
473 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
474 if (!(data->rq_flags & RQF_ELV))
475 blk_mq_tag_busy(data->hctx);
477 if (data->flags & BLK_MQ_REQ_RESERVED)
478 data->rq_flags |= RQF_RESV;
481 * Try batched alloc if we want more than 1 tag.
483 if (data->nr_tags > 1) {
484 rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns);
491 * Waiting allocations only fail because of an inactive hctx. In that
492 * case just retry the hctx assignment and tag allocation as CPU hotplug
493 * should have migrated us to an online CPU by now.
495 tag = blk_mq_get_tag(data);
496 if (tag == BLK_MQ_NO_TAG) {
497 if (data->flags & BLK_MQ_REQ_NOWAIT)
500 * Give up the CPU and sleep for a random short time to
501 * ensure that thread using a realtime scheduling class
502 * are migrated off the CPU, and thus off the hctx that
509 return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag,
513 static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
514 struct blk_plug *plug,
516 blk_mq_req_flags_t flags)
518 struct blk_mq_alloc_data data = {
522 .nr_tags = plug->nr_ios,
523 .cached_rq = &plug->cached_rq,
527 if (blk_queue_enter(q, flags))
532 rq = __blk_mq_alloc_requests(&data);
538 static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
540 blk_mq_req_flags_t flags)
542 struct blk_plug *plug = current->plug;
547 if (rq_list_empty(plug->cached_rq)) {
548 if (plug->nr_ios == 1)
550 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
555 rq = rq_list_peek(&plug->cached_rq);
556 if (!rq || rq->q != q)
559 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
561 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
564 plug->cached_rq = rq_list_next(rq);
567 INIT_LIST_HEAD(&rq->queuelist);
571 struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
572 blk_mq_req_flags_t flags)
576 rq = blk_mq_alloc_cached_request(q, opf, flags);
578 struct blk_mq_alloc_data data = {
586 ret = blk_queue_enter(q, flags);
590 rq = __blk_mq_alloc_requests(&data);
595 rq->__sector = (sector_t) -1;
596 rq->bio = rq->biotail = NULL;
600 return ERR_PTR(-EWOULDBLOCK);
602 EXPORT_SYMBOL(blk_mq_alloc_request);
604 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
605 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
607 struct blk_mq_alloc_data data = {
613 u64 alloc_time_ns = 0;
618 /* alloc_time includes depth and tag waits */
619 if (blk_queue_rq_alloc_time(q))
620 alloc_time_ns = ktime_get_ns();
623 * If the tag allocator sleeps we could get an allocation for a
624 * different hardware context. No need to complicate the low level
625 * allocator for this for the rare use case of a command tied to
628 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
629 return ERR_PTR(-EINVAL);
631 if (hctx_idx >= q->nr_hw_queues)
632 return ERR_PTR(-EIO);
634 ret = blk_queue_enter(q, flags);
639 * Check if the hardware context is actually mapped to anything.
640 * If not tell the caller that it should skip this queue.
643 data.hctx = xa_load(&q->hctx_table, hctx_idx);
644 if (!blk_mq_hw_queue_mapped(data.hctx))
646 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
647 if (cpu >= nr_cpu_ids)
649 data.ctx = __blk_mq_get_ctx(q, cpu);
652 blk_mq_tag_busy(data.hctx);
654 data.rq_flags |= RQF_ELV;
656 if (flags & BLK_MQ_REQ_RESERVED)
657 data.rq_flags |= RQF_RESV;
660 tag = blk_mq_get_tag(&data);
661 if (tag == BLK_MQ_NO_TAG)
663 return blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag,
670 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
672 static void __blk_mq_free_request(struct request *rq)
674 struct request_queue *q = rq->q;
675 struct blk_mq_ctx *ctx = rq->mq_ctx;
676 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
677 const int sched_tag = rq->internal_tag;
679 blk_crypto_free_request(rq);
680 blk_pm_mark_last_busy(rq);
682 if (rq->tag != BLK_MQ_NO_TAG)
683 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
684 if (sched_tag != BLK_MQ_NO_TAG)
685 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
686 blk_mq_sched_restart(hctx);
690 void blk_mq_free_request(struct request *rq)
692 struct request_queue *q = rq->q;
693 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
695 if ((rq->rq_flags & RQF_ELVPRIV) &&
696 q->elevator->type->ops.finish_request)
697 q->elevator->type->ops.finish_request(rq);
699 if (rq->rq_flags & RQF_MQ_INFLIGHT)
700 __blk_mq_dec_active_requests(hctx);
702 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
703 laptop_io_completion(q->disk->bdi);
707 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
708 if (req_ref_put_and_test(rq))
709 __blk_mq_free_request(rq);
711 EXPORT_SYMBOL_GPL(blk_mq_free_request);
713 void blk_mq_free_plug_rqs(struct blk_plug *plug)
717 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
718 blk_mq_free_request(rq);
721 void blk_dump_rq_flags(struct request *rq, char *msg)
723 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
724 rq->q->disk ? rq->q->disk->disk_name : "?",
725 (__force unsigned long long) rq->cmd_flags);
727 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
728 (unsigned long long)blk_rq_pos(rq),
729 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
730 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
731 rq->bio, rq->biotail, blk_rq_bytes(rq));
733 EXPORT_SYMBOL(blk_dump_rq_flags);
735 static void req_bio_endio(struct request *rq, struct bio *bio,
736 unsigned int nbytes, blk_status_t error)
738 if (unlikely(error)) {
739 bio->bi_status = error;
740 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
742 * Partial zone append completions cannot be supported as the
743 * BIO fragments may end up not being written sequentially.
745 if (bio->bi_iter.bi_size != nbytes)
746 bio->bi_status = BLK_STS_IOERR;
748 bio->bi_iter.bi_sector = rq->__sector;
751 bio_advance(bio, nbytes);
753 if (unlikely(rq->rq_flags & RQF_QUIET))
754 bio_set_flag(bio, BIO_QUIET);
755 /* don't actually finish bio if it's part of flush sequence */
756 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
760 static void blk_account_io_completion(struct request *req, unsigned int bytes)
762 if (req->part && blk_do_io_stat(req)) {
763 const int sgrp = op_stat_group(req_op(req));
766 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
771 static void blk_print_req_error(struct request *req, blk_status_t status)
773 printk_ratelimited(KERN_ERR
774 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
775 "phys_seg %u prio class %u\n",
776 blk_status_to_str(status),
777 req->q->disk ? req->q->disk->disk_name : "?",
778 blk_rq_pos(req), (__force u32)req_op(req),
779 blk_op_str(req_op(req)),
780 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
781 req->nr_phys_segments,
782 IOPRIO_PRIO_CLASS(req->ioprio));
786 * Fully end IO on a request. Does not support partial completions, or
789 static void blk_complete_request(struct request *req)
791 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
792 int total_bytes = blk_rq_bytes(req);
793 struct bio *bio = req->bio;
795 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
800 #ifdef CONFIG_BLK_DEV_INTEGRITY
801 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
802 req->q->integrity.profile->complete_fn(req, total_bytes);
805 blk_account_io_completion(req, total_bytes);
808 struct bio *next = bio->bi_next;
810 /* Completion has already been traced */
811 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
813 if (req_op(req) == REQ_OP_ZONE_APPEND)
814 bio->bi_iter.bi_sector = req->__sector;
822 * Reset counters so that the request stacking driver
823 * can find how many bytes remain in the request
833 * blk_update_request - Complete multiple bytes without completing the request
834 * @req: the request being processed
835 * @error: block status code
836 * @nr_bytes: number of bytes to complete for @req
839 * Ends I/O on a number of bytes attached to @req, but doesn't complete
840 * the request structure even if @req doesn't have leftover.
841 * If @req has leftover, sets it up for the next range of segments.
843 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
844 * %false return from this function.
847 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
848 * except in the consistency check at the end of this function.
851 * %false - this request doesn't have any more data
852 * %true - this request has more data
854 bool blk_update_request(struct request *req, blk_status_t error,
855 unsigned int nr_bytes)
859 trace_block_rq_complete(req, error, nr_bytes);
864 #ifdef CONFIG_BLK_DEV_INTEGRITY
865 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
867 req->q->integrity.profile->complete_fn(req, nr_bytes);
870 if (unlikely(error && !blk_rq_is_passthrough(req) &&
871 !(req->rq_flags & RQF_QUIET)) &&
872 !test_bit(GD_DEAD, &req->q->disk->state)) {
873 blk_print_req_error(req, error);
874 trace_block_rq_error(req, error, nr_bytes);
877 blk_account_io_completion(req, nr_bytes);
881 struct bio *bio = req->bio;
882 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
884 if (bio_bytes == bio->bi_iter.bi_size)
885 req->bio = bio->bi_next;
887 /* Completion has already been traced */
888 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
889 req_bio_endio(req, bio, bio_bytes, error);
891 total_bytes += bio_bytes;
892 nr_bytes -= bio_bytes;
903 * Reset counters so that the request stacking driver
904 * can find how many bytes remain in the request
911 req->__data_len -= total_bytes;
913 /* update sector only for requests with clear definition of sector */
914 if (!blk_rq_is_passthrough(req))
915 req->__sector += total_bytes >> 9;
917 /* mixed attributes always follow the first bio */
918 if (req->rq_flags & RQF_MIXED_MERGE) {
919 req->cmd_flags &= ~REQ_FAILFAST_MASK;
920 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
923 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
925 * If total number of sectors is less than the first segment
926 * size, something has gone terribly wrong.
928 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
929 blk_dump_rq_flags(req, "request botched");
930 req->__data_len = blk_rq_cur_bytes(req);
933 /* recalculate the number of segments */
934 req->nr_phys_segments = blk_recalc_rq_segments(req);
939 EXPORT_SYMBOL_GPL(blk_update_request);
941 static void __blk_account_io_done(struct request *req, u64 now)
943 const int sgrp = op_stat_group(req_op(req));
946 update_io_ticks(req->part, jiffies, true);
947 part_stat_inc(req->part, ios[sgrp]);
948 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
952 static inline void blk_account_io_done(struct request *req, u64 now)
955 * Account IO completion. flush_rq isn't accounted as a
956 * normal IO on queueing nor completion. Accounting the
957 * containing request is enough.
959 if (blk_do_io_stat(req) && req->part &&
960 !(req->rq_flags & RQF_FLUSH_SEQ))
961 __blk_account_io_done(req, now);
964 static void __blk_account_io_start(struct request *rq)
967 * All non-passthrough requests are created from a bio with one
968 * exception: when a flush command that is part of a flush sequence
969 * generated by the state machine in blk-flush.c is cloned onto the
970 * lower device by dm-multipath we can get here without a bio.
973 rq->part = rq->bio->bi_bdev;
975 rq->part = rq->q->disk->part0;
978 update_io_ticks(rq->part, jiffies, false);
982 static inline void blk_account_io_start(struct request *req)
984 if (blk_do_io_stat(req))
985 __blk_account_io_start(req);
988 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
990 if (rq->rq_flags & RQF_STATS) {
991 blk_mq_poll_stats_start(rq->q);
992 blk_stat_add(rq, now);
995 blk_mq_sched_completed_request(rq, now);
996 blk_account_io_done(rq, now);
999 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1001 if (blk_mq_need_time_stamp(rq))
1002 __blk_mq_end_request_acct(rq, ktime_get_ns());
1005 rq_qos_done(rq->q, rq);
1006 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1007 blk_mq_free_request(rq);
1009 blk_mq_free_request(rq);
1012 EXPORT_SYMBOL(__blk_mq_end_request);
1014 void blk_mq_end_request(struct request *rq, blk_status_t error)
1016 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1018 __blk_mq_end_request(rq, error);
1020 EXPORT_SYMBOL(blk_mq_end_request);
1022 #define TAG_COMP_BATCH 32
1024 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1025 int *tag_array, int nr_tags)
1027 struct request_queue *q = hctx->queue;
1030 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1031 * update hctx->nr_active in batch
1033 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1034 __blk_mq_sub_active_requests(hctx, nr_tags);
1036 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1037 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1040 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1042 int tags[TAG_COMP_BATCH], nr_tags = 0;
1043 struct blk_mq_hw_ctx *cur_hctx = NULL;
1048 now = ktime_get_ns();
1050 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1052 prefetch(rq->rq_next);
1054 blk_complete_request(rq);
1056 __blk_mq_end_request_acct(rq, now);
1058 rq_qos_done(rq->q, rq);
1061 * If end_io handler returns NONE, then it still has
1062 * ownership of the request.
1064 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1067 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1068 if (!req_ref_put_and_test(rq))
1071 blk_crypto_free_request(rq);
1072 blk_pm_mark_last_busy(rq);
1074 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1076 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1078 cur_hctx = rq->mq_hctx;
1080 tags[nr_tags++] = rq->tag;
1084 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1086 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1088 static void blk_complete_reqs(struct llist_head *list)
1090 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1091 struct request *rq, *next;
1093 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1094 rq->q->mq_ops->complete(rq);
1097 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1099 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1102 static int blk_softirq_cpu_dead(unsigned int cpu)
1104 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1108 static void __blk_mq_complete_request_remote(void *data)
1110 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1113 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1115 int cpu = raw_smp_processor_id();
1117 if (!IS_ENABLED(CONFIG_SMP) ||
1118 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1121 * With force threaded interrupts enabled, raising softirq from an SMP
1122 * function call will always result in waking the ksoftirqd thread.
1123 * This is probably worse than completing the request on a different
1126 if (force_irqthreads())
1129 /* same CPU or cache domain? Complete locally */
1130 if (cpu == rq->mq_ctx->cpu ||
1131 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1132 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1135 /* don't try to IPI to an offline CPU */
1136 return cpu_online(rq->mq_ctx->cpu);
1139 static void blk_mq_complete_send_ipi(struct request *rq)
1141 struct llist_head *list;
1144 cpu = rq->mq_ctx->cpu;
1145 list = &per_cpu(blk_cpu_done, cpu);
1146 if (llist_add(&rq->ipi_list, list)) {
1147 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1148 smp_call_function_single_async(cpu, &rq->csd);
1152 static void blk_mq_raise_softirq(struct request *rq)
1154 struct llist_head *list;
1157 list = this_cpu_ptr(&blk_cpu_done);
1158 if (llist_add(&rq->ipi_list, list))
1159 raise_softirq(BLOCK_SOFTIRQ);
1163 bool blk_mq_complete_request_remote(struct request *rq)
1165 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1168 * For request which hctx has only one ctx mapping,
1169 * or a polled request, always complete locally,
1170 * it's pointless to redirect the completion.
1172 if (rq->mq_hctx->nr_ctx == 1 ||
1173 rq->cmd_flags & REQ_POLLED)
1176 if (blk_mq_complete_need_ipi(rq)) {
1177 blk_mq_complete_send_ipi(rq);
1181 if (rq->q->nr_hw_queues == 1) {
1182 blk_mq_raise_softirq(rq);
1187 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1190 * blk_mq_complete_request - end I/O on a request
1191 * @rq: the request being processed
1194 * Complete a request by scheduling the ->complete_rq operation.
1196 void blk_mq_complete_request(struct request *rq)
1198 if (!blk_mq_complete_request_remote(rq))
1199 rq->q->mq_ops->complete(rq);
1201 EXPORT_SYMBOL(blk_mq_complete_request);
1204 * blk_mq_start_request - Start processing a request
1205 * @rq: Pointer to request to be started
1207 * Function used by device drivers to notify the block layer that a request
1208 * is going to be processed now, so blk layer can do proper initializations
1209 * such as starting the timeout timer.
1211 void blk_mq_start_request(struct request *rq)
1213 struct request_queue *q = rq->q;
1215 trace_block_rq_issue(rq);
1217 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1218 rq->io_start_time_ns = ktime_get_ns();
1219 rq->stats_sectors = blk_rq_sectors(rq);
1220 rq->rq_flags |= RQF_STATS;
1221 rq_qos_issue(q, rq);
1224 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1227 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1229 #ifdef CONFIG_BLK_DEV_INTEGRITY
1230 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1231 q->integrity.profile->prepare_fn(rq);
1233 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1234 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1236 EXPORT_SYMBOL(blk_mq_start_request);
1239 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1240 * queues. This is important for md arrays to benefit from merging
1243 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1245 if (plug->multiple_queues)
1246 return BLK_MAX_REQUEST_COUNT * 2;
1247 return BLK_MAX_REQUEST_COUNT;
1250 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1252 struct request *last = rq_list_peek(&plug->mq_list);
1254 if (!plug->rq_count) {
1255 trace_block_plug(rq->q);
1256 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1257 (!blk_queue_nomerges(rq->q) &&
1258 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1259 blk_mq_flush_plug_list(plug, false);
1260 trace_block_plug(rq->q);
1263 if (!plug->multiple_queues && last && last->q != rq->q)
1264 plug->multiple_queues = true;
1265 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1266 plug->has_elevator = true;
1268 rq_list_add(&plug->mq_list, rq);
1273 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1274 * @rq: request to insert
1275 * @at_head: insert request at head or tail of queue
1278 * Insert a fully prepared request at the back of the I/O scheduler queue
1279 * for execution. Don't wait for completion.
1282 * This function will invoke @done directly if the queue is dead.
1284 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1286 WARN_ON(irqs_disabled());
1287 WARN_ON(!blk_rq_is_passthrough(rq));
1289 blk_account_io_start(rq);
1292 * As plugging can be enabled for passthrough requests on a zoned
1293 * device, directly accessing the plug instead of using blk_mq_plug()
1294 * should not have any consequences.
1297 blk_add_rq_to_plug(current->plug, rq);
1299 blk_mq_sched_insert_request(rq, at_head, true, false);
1301 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1303 struct blk_rq_wait {
1304 struct completion done;
1308 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1310 struct blk_rq_wait *wait = rq->end_io_data;
1313 complete(&wait->done);
1314 return RQ_END_IO_NONE;
1317 bool blk_rq_is_poll(struct request *rq)
1321 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1323 if (WARN_ON_ONCE(!rq->bio))
1327 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1329 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1332 bio_poll(rq->bio, NULL, 0);
1334 } while (!completion_done(wait));
1338 * blk_execute_rq - insert a request into queue for execution
1339 * @rq: request to insert
1340 * @at_head: insert request at head or tail of queue
1343 * Insert a fully prepared request at the back of the I/O scheduler queue
1344 * for execution and wait for completion.
1345 * Return: The blk_status_t result provided to blk_mq_end_request().
1347 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1349 struct blk_rq_wait wait = {
1350 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1353 WARN_ON(irqs_disabled());
1354 WARN_ON(!blk_rq_is_passthrough(rq));
1356 rq->end_io_data = &wait;
1357 rq->end_io = blk_end_sync_rq;
1359 blk_account_io_start(rq);
1360 blk_mq_sched_insert_request(rq, at_head, true, false);
1362 if (blk_rq_is_poll(rq)) {
1363 blk_rq_poll_completion(rq, &wait.done);
1366 * Prevent hang_check timer from firing at us during very long
1369 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1372 while (!wait_for_completion_io_timeout(&wait.done,
1373 hang_check * (HZ/2)))
1376 wait_for_completion_io(&wait.done);
1381 EXPORT_SYMBOL(blk_execute_rq);
1383 static void __blk_mq_requeue_request(struct request *rq)
1385 struct request_queue *q = rq->q;
1387 blk_mq_put_driver_tag(rq);
1389 trace_block_rq_requeue(rq);
1390 rq_qos_requeue(q, rq);
1392 if (blk_mq_request_started(rq)) {
1393 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1394 rq->rq_flags &= ~RQF_TIMED_OUT;
1398 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1400 __blk_mq_requeue_request(rq);
1402 /* this request will be re-inserted to io scheduler queue */
1403 blk_mq_sched_requeue_request(rq);
1405 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1407 EXPORT_SYMBOL(blk_mq_requeue_request);
1409 static void blk_mq_requeue_work(struct work_struct *work)
1411 struct request_queue *q =
1412 container_of(work, struct request_queue, requeue_work.work);
1414 struct request *rq, *next;
1416 spin_lock_irq(&q->requeue_lock);
1417 list_splice_init(&q->requeue_list, &rq_list);
1418 spin_unlock_irq(&q->requeue_lock);
1420 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1421 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1424 rq->rq_flags &= ~RQF_SOFTBARRIER;
1425 list_del_init(&rq->queuelist);
1427 * If RQF_DONTPREP, rq has contained some driver specific
1428 * data, so insert it to hctx dispatch list to avoid any
1431 if (rq->rq_flags & RQF_DONTPREP)
1432 blk_mq_request_bypass_insert(rq, false, false);
1434 blk_mq_sched_insert_request(rq, true, false, false);
1437 while (!list_empty(&rq_list)) {
1438 rq = list_entry(rq_list.next, struct request, queuelist);
1439 list_del_init(&rq->queuelist);
1440 blk_mq_sched_insert_request(rq, false, false, false);
1443 blk_mq_run_hw_queues(q, false);
1446 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1447 bool kick_requeue_list)
1449 struct request_queue *q = rq->q;
1450 unsigned long flags;
1453 * We abuse this flag that is otherwise used by the I/O scheduler to
1454 * request head insertion from the workqueue.
1456 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1458 spin_lock_irqsave(&q->requeue_lock, flags);
1460 rq->rq_flags |= RQF_SOFTBARRIER;
1461 list_add(&rq->queuelist, &q->requeue_list);
1463 list_add_tail(&rq->queuelist, &q->requeue_list);
1465 spin_unlock_irqrestore(&q->requeue_lock, flags);
1467 if (kick_requeue_list)
1468 blk_mq_kick_requeue_list(q);
1471 void blk_mq_kick_requeue_list(struct request_queue *q)
1473 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1475 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1477 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1478 unsigned long msecs)
1480 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1481 msecs_to_jiffies(msecs));
1483 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1485 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1488 * If we find a request that isn't idle we know the queue is busy
1489 * as it's checked in the iter.
1490 * Return false to stop the iteration.
1492 if (blk_mq_request_started(rq)) {
1502 bool blk_mq_queue_inflight(struct request_queue *q)
1506 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1509 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1511 static void blk_mq_rq_timed_out(struct request *req)
1513 req->rq_flags |= RQF_TIMED_OUT;
1514 if (req->q->mq_ops->timeout) {
1515 enum blk_eh_timer_return ret;
1517 ret = req->q->mq_ops->timeout(req);
1518 if (ret == BLK_EH_DONE)
1520 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1526 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
1528 unsigned long deadline;
1530 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1532 if (rq->rq_flags & RQF_TIMED_OUT)
1535 deadline = READ_ONCE(rq->deadline);
1536 if (time_after_eq(jiffies, deadline))
1541 else if (time_after(*next, deadline))
1546 void blk_mq_put_rq_ref(struct request *rq)
1548 if (is_flush_rq(rq)) {
1549 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1550 blk_mq_free_request(rq);
1551 } else if (req_ref_put_and_test(rq)) {
1552 __blk_mq_free_request(rq);
1556 static bool blk_mq_check_expired(struct request *rq, void *priv)
1558 unsigned long *next = priv;
1561 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1562 * be reallocated underneath the timeout handler's processing, then
1563 * the expire check is reliable. If the request is not expired, then
1564 * it was completed and reallocated as a new request after returning
1565 * from blk_mq_check_expired().
1567 if (blk_mq_req_expired(rq, next))
1568 blk_mq_rq_timed_out(rq);
1572 static void blk_mq_timeout_work(struct work_struct *work)
1574 struct request_queue *q =
1575 container_of(work, struct request_queue, timeout_work);
1576 unsigned long next = 0;
1577 struct blk_mq_hw_ctx *hctx;
1580 /* A deadlock might occur if a request is stuck requiring a
1581 * timeout at the same time a queue freeze is waiting
1582 * completion, since the timeout code would not be able to
1583 * acquire the queue reference here.
1585 * That's why we don't use blk_queue_enter here; instead, we use
1586 * percpu_ref_tryget directly, because we need to be able to
1587 * obtain a reference even in the short window between the queue
1588 * starting to freeze, by dropping the first reference in
1589 * blk_freeze_queue_start, and the moment the last request is
1590 * consumed, marked by the instant q_usage_counter reaches
1593 if (!percpu_ref_tryget(&q->q_usage_counter))
1596 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
1599 mod_timer(&q->timeout, next);
1602 * Request timeouts are handled as a forward rolling timer. If
1603 * we end up here it means that no requests are pending and
1604 * also that no request has been pending for a while. Mark
1605 * each hctx as idle.
1607 queue_for_each_hw_ctx(q, hctx, i) {
1608 /* the hctx may be unmapped, so check it here */
1609 if (blk_mq_hw_queue_mapped(hctx))
1610 blk_mq_tag_idle(hctx);
1616 struct flush_busy_ctx_data {
1617 struct blk_mq_hw_ctx *hctx;
1618 struct list_head *list;
1621 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1623 struct flush_busy_ctx_data *flush_data = data;
1624 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1625 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1626 enum hctx_type type = hctx->type;
1628 spin_lock(&ctx->lock);
1629 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1630 sbitmap_clear_bit(sb, bitnr);
1631 spin_unlock(&ctx->lock);
1636 * Process software queues that have been marked busy, splicing them
1637 * to the for-dispatch
1639 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1641 struct flush_busy_ctx_data data = {
1646 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1648 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1650 struct dispatch_rq_data {
1651 struct blk_mq_hw_ctx *hctx;
1655 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1658 struct dispatch_rq_data *dispatch_data = data;
1659 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1660 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1661 enum hctx_type type = hctx->type;
1663 spin_lock(&ctx->lock);
1664 if (!list_empty(&ctx->rq_lists[type])) {
1665 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1666 list_del_init(&dispatch_data->rq->queuelist);
1667 if (list_empty(&ctx->rq_lists[type]))
1668 sbitmap_clear_bit(sb, bitnr);
1670 spin_unlock(&ctx->lock);
1672 return !dispatch_data->rq;
1675 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1676 struct blk_mq_ctx *start)
1678 unsigned off = start ? start->index_hw[hctx->type] : 0;
1679 struct dispatch_rq_data data = {
1684 __sbitmap_for_each_set(&hctx->ctx_map, off,
1685 dispatch_rq_from_ctx, &data);
1690 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1692 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1693 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1696 blk_mq_tag_busy(rq->mq_hctx);
1698 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1699 bt = &rq->mq_hctx->tags->breserved_tags;
1702 if (!hctx_may_queue(rq->mq_hctx, bt))
1706 tag = __sbitmap_queue_get(bt);
1707 if (tag == BLK_MQ_NO_TAG)
1710 rq->tag = tag + tag_offset;
1714 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1716 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1719 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1720 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1721 rq->rq_flags |= RQF_MQ_INFLIGHT;
1722 __blk_mq_inc_active_requests(hctx);
1724 hctx->tags->rqs[rq->tag] = rq;
1728 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1729 int flags, void *key)
1731 struct blk_mq_hw_ctx *hctx;
1733 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1735 spin_lock(&hctx->dispatch_wait_lock);
1736 if (!list_empty(&wait->entry)) {
1737 struct sbitmap_queue *sbq;
1739 list_del_init(&wait->entry);
1740 sbq = &hctx->tags->bitmap_tags;
1741 atomic_dec(&sbq->ws_active);
1743 spin_unlock(&hctx->dispatch_wait_lock);
1745 blk_mq_run_hw_queue(hctx, true);
1750 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1751 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1752 * restart. For both cases, take care to check the condition again after
1753 * marking us as waiting.
1755 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1758 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1759 struct wait_queue_head *wq;
1760 wait_queue_entry_t *wait;
1763 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1764 blk_mq_sched_mark_restart_hctx(hctx);
1767 * It's possible that a tag was freed in the window between the
1768 * allocation failure and adding the hardware queue to the wait
1771 * Don't clear RESTART here, someone else could have set it.
1772 * At most this will cost an extra queue run.
1774 return blk_mq_get_driver_tag(rq);
1777 wait = &hctx->dispatch_wait;
1778 if (!list_empty_careful(&wait->entry))
1781 wq = &bt_wait_ptr(sbq, hctx)->wait;
1783 spin_lock_irq(&wq->lock);
1784 spin_lock(&hctx->dispatch_wait_lock);
1785 if (!list_empty(&wait->entry)) {
1786 spin_unlock(&hctx->dispatch_wait_lock);
1787 spin_unlock_irq(&wq->lock);
1791 atomic_inc(&sbq->ws_active);
1792 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1793 __add_wait_queue(wq, wait);
1796 * It's possible that a tag was freed in the window between the
1797 * allocation failure and adding the hardware queue to the wait
1800 ret = blk_mq_get_driver_tag(rq);
1802 spin_unlock(&hctx->dispatch_wait_lock);
1803 spin_unlock_irq(&wq->lock);
1808 * We got a tag, remove ourselves from the wait queue to ensure
1809 * someone else gets the wakeup.
1811 list_del_init(&wait->entry);
1812 atomic_dec(&sbq->ws_active);
1813 spin_unlock(&hctx->dispatch_wait_lock);
1814 spin_unlock_irq(&wq->lock);
1819 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1820 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1822 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1823 * - EWMA is one simple way to compute running average value
1824 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1825 * - take 4 as factor for avoiding to get too small(0) result, and this
1826 * factor doesn't matter because EWMA decreases exponentially
1828 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1832 ewma = hctx->dispatch_busy;
1837 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1839 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1840 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1842 hctx->dispatch_busy = ewma;
1845 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1847 static void blk_mq_handle_dev_resource(struct request *rq,
1848 struct list_head *list)
1850 struct request *next =
1851 list_first_entry_or_null(list, struct request, queuelist);
1854 * If an I/O scheduler has been configured and we got a driver tag for
1855 * the next request already, free it.
1858 blk_mq_put_driver_tag(next);
1860 list_add(&rq->queuelist, list);
1861 __blk_mq_requeue_request(rq);
1864 static void blk_mq_handle_zone_resource(struct request *rq,
1865 struct list_head *zone_list)
1868 * If we end up here it is because we cannot dispatch a request to a
1869 * specific zone due to LLD level zone-write locking or other zone
1870 * related resource not being available. In this case, set the request
1871 * aside in zone_list for retrying it later.
1873 list_add(&rq->queuelist, zone_list);
1874 __blk_mq_requeue_request(rq);
1877 enum prep_dispatch {
1879 PREP_DISPATCH_NO_TAG,
1880 PREP_DISPATCH_NO_BUDGET,
1883 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1886 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1887 int budget_token = -1;
1890 budget_token = blk_mq_get_dispatch_budget(rq->q);
1891 if (budget_token < 0) {
1892 blk_mq_put_driver_tag(rq);
1893 return PREP_DISPATCH_NO_BUDGET;
1895 blk_mq_set_rq_budget_token(rq, budget_token);
1898 if (!blk_mq_get_driver_tag(rq)) {
1900 * The initial allocation attempt failed, so we need to
1901 * rerun the hardware queue when a tag is freed. The
1902 * waitqueue takes care of that. If the queue is run
1903 * before we add this entry back on the dispatch list,
1904 * we'll re-run it below.
1906 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1908 * All budgets not got from this function will be put
1909 * together during handling partial dispatch
1912 blk_mq_put_dispatch_budget(rq->q, budget_token);
1913 return PREP_DISPATCH_NO_TAG;
1917 return PREP_DISPATCH_OK;
1920 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1921 static void blk_mq_release_budgets(struct request_queue *q,
1922 struct list_head *list)
1926 list_for_each_entry(rq, list, queuelist) {
1927 int budget_token = blk_mq_get_rq_budget_token(rq);
1929 if (budget_token >= 0)
1930 blk_mq_put_dispatch_budget(q, budget_token);
1935 * Returns true if we did some work AND can potentially do more.
1937 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1938 unsigned int nr_budgets)
1940 enum prep_dispatch prep;
1941 struct request_queue *q = hctx->queue;
1942 struct request *rq, *nxt;
1944 blk_status_t ret = BLK_STS_OK;
1945 LIST_HEAD(zone_list);
1946 bool needs_resource = false;
1948 if (list_empty(list))
1952 * Now process all the entries, sending them to the driver.
1954 errors = queued = 0;
1956 struct blk_mq_queue_data bd;
1958 rq = list_first_entry(list, struct request, queuelist);
1960 WARN_ON_ONCE(hctx != rq->mq_hctx);
1961 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1962 if (prep != PREP_DISPATCH_OK)
1965 list_del_init(&rq->queuelist);
1970 * Flag last if we have no more requests, or if we have more
1971 * but can't assign a driver tag to it.
1973 if (list_empty(list))
1976 nxt = list_first_entry(list, struct request, queuelist);
1977 bd.last = !blk_mq_get_driver_tag(nxt);
1981 * once the request is queued to lld, no need to cover the
1986 ret = q->mq_ops->queue_rq(hctx, &bd);
1991 case BLK_STS_RESOURCE:
1992 needs_resource = true;
1994 case BLK_STS_DEV_RESOURCE:
1995 blk_mq_handle_dev_resource(rq, list);
1997 case BLK_STS_ZONE_RESOURCE:
1999 * Move the request to zone_list and keep going through
2000 * the dispatch list to find more requests the drive can
2003 blk_mq_handle_zone_resource(rq, &zone_list);
2004 needs_resource = true;
2008 blk_mq_end_request(rq, ret);
2010 } while (!list_empty(list));
2012 if (!list_empty(&zone_list))
2013 list_splice_tail_init(&zone_list, list);
2015 /* If we didn't flush the entire list, we could have told the driver
2016 * there was more coming, but that turned out to be a lie.
2018 if ((!list_empty(list) || errors || needs_resource ||
2019 ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued)
2020 q->mq_ops->commit_rqs(hctx);
2022 * Any items that need requeuing? Stuff them into hctx->dispatch,
2023 * that is where we will continue on next queue run.
2025 if (!list_empty(list)) {
2027 /* For non-shared tags, the RESTART check will suffice */
2028 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2029 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
2032 blk_mq_release_budgets(q, list);
2034 spin_lock(&hctx->lock);
2035 list_splice_tail_init(list, &hctx->dispatch);
2036 spin_unlock(&hctx->lock);
2039 * Order adding requests to hctx->dispatch and checking
2040 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2041 * in blk_mq_sched_restart(). Avoid restart code path to
2042 * miss the new added requests to hctx->dispatch, meantime
2043 * SCHED_RESTART is observed here.
2048 * If SCHED_RESTART was set by the caller of this function and
2049 * it is no longer set that means that it was cleared by another
2050 * thread and hence that a queue rerun is needed.
2052 * If 'no_tag' is set, that means that we failed getting
2053 * a driver tag with an I/O scheduler attached. If our dispatch
2054 * waitqueue is no longer active, ensure that we run the queue
2055 * AFTER adding our entries back to the list.
2057 * If no I/O scheduler has been configured it is possible that
2058 * the hardware queue got stopped and restarted before requests
2059 * were pushed back onto the dispatch list. Rerun the queue to
2060 * avoid starvation. Notes:
2061 * - blk_mq_run_hw_queue() checks whether or not a queue has
2062 * been stopped before rerunning a queue.
2063 * - Some but not all block drivers stop a queue before
2064 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2067 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2068 * bit is set, run queue after a delay to avoid IO stalls
2069 * that could otherwise occur if the queue is idle. We'll do
2070 * similar if we couldn't get budget or couldn't lock a zone
2071 * and SCHED_RESTART is set.
2073 needs_restart = blk_mq_sched_needs_restart(hctx);
2074 if (prep == PREP_DISPATCH_NO_BUDGET)
2075 needs_resource = true;
2076 if (!needs_restart ||
2077 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2078 blk_mq_run_hw_queue(hctx, true);
2079 else if (needs_resource)
2080 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2082 blk_mq_update_dispatch_busy(hctx, true);
2085 blk_mq_update_dispatch_busy(hctx, false);
2087 return (queued + errors) != 0;
2091 * __blk_mq_run_hw_queue - Run a hardware queue.
2092 * @hctx: Pointer to the hardware queue to run.
2094 * Send pending requests to the hardware.
2096 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
2099 * We can't run the queue inline with ints disabled. Ensure that
2100 * we catch bad users of this early.
2102 WARN_ON_ONCE(in_interrupt());
2104 blk_mq_run_dispatch_ops(hctx->queue,
2105 blk_mq_sched_dispatch_requests(hctx));
2108 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2110 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2112 if (cpu >= nr_cpu_ids)
2113 cpu = cpumask_first(hctx->cpumask);
2118 * It'd be great if the workqueue API had a way to pass
2119 * in a mask and had some smarts for more clever placement.
2120 * For now we just round-robin here, switching for every
2121 * BLK_MQ_CPU_WORK_BATCH queued items.
2123 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2126 int next_cpu = hctx->next_cpu;
2128 if (hctx->queue->nr_hw_queues == 1)
2129 return WORK_CPU_UNBOUND;
2131 if (--hctx->next_cpu_batch <= 0) {
2133 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2135 if (next_cpu >= nr_cpu_ids)
2136 next_cpu = blk_mq_first_mapped_cpu(hctx);
2137 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2141 * Do unbound schedule if we can't find a online CPU for this hctx,
2142 * and it should only happen in the path of handling CPU DEAD.
2144 if (!cpu_online(next_cpu)) {
2151 * Make sure to re-select CPU next time once after CPUs
2152 * in hctx->cpumask become online again.
2154 hctx->next_cpu = next_cpu;
2155 hctx->next_cpu_batch = 1;
2156 return WORK_CPU_UNBOUND;
2159 hctx->next_cpu = next_cpu;
2164 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2165 * @hctx: Pointer to the hardware queue to run.
2166 * @async: If we want to run the queue asynchronously.
2167 * @msecs: Milliseconds of delay to wait before running the queue.
2169 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2170 * with a delay of @msecs.
2172 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2173 unsigned long msecs)
2175 if (unlikely(blk_mq_hctx_stopped(hctx)))
2178 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2179 if (cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2180 __blk_mq_run_hw_queue(hctx);
2185 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2186 msecs_to_jiffies(msecs));
2190 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2191 * @hctx: Pointer to the hardware queue to run.
2192 * @msecs: Milliseconds of delay to wait before running the queue.
2194 * Run a hardware queue asynchronously with a delay of @msecs.
2196 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2198 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2200 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2203 * blk_mq_run_hw_queue - Start to run a hardware queue.
2204 * @hctx: Pointer to the hardware queue to run.
2205 * @async: If we want to run the queue asynchronously.
2207 * Check if the request queue is not in a quiesced state and if there are
2208 * pending requests to be sent. If this is true, run the queue to send requests
2211 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2216 * When queue is quiesced, we may be switching io scheduler, or
2217 * updating nr_hw_queues, or other things, and we can't run queue
2218 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2220 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2223 __blk_mq_run_dispatch_ops(hctx->queue, false,
2224 need_run = !blk_queue_quiesced(hctx->queue) &&
2225 blk_mq_hctx_has_pending(hctx));
2228 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2230 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2233 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2236 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2238 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2240 * If the IO scheduler does not respect hardware queues when
2241 * dispatching, we just don't bother with multiple HW queues and
2242 * dispatch from hctx for the current CPU since running multiple queues
2243 * just causes lock contention inside the scheduler and pointless cache
2246 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2248 if (!blk_mq_hctx_stopped(hctx))
2254 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2255 * @q: Pointer to the request queue to run.
2256 * @async: If we want to run the queue asynchronously.
2258 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2260 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2264 if (blk_queue_sq_sched(q))
2265 sq_hctx = blk_mq_get_sq_hctx(q);
2266 queue_for_each_hw_ctx(q, hctx, i) {
2267 if (blk_mq_hctx_stopped(hctx))
2270 * Dispatch from this hctx either if there's no hctx preferred
2271 * by IO scheduler or if it has requests that bypass the
2274 if (!sq_hctx || sq_hctx == hctx ||
2275 !list_empty_careful(&hctx->dispatch))
2276 blk_mq_run_hw_queue(hctx, async);
2279 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2282 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2283 * @q: Pointer to the request queue to run.
2284 * @msecs: Milliseconds of delay to wait before running the queues.
2286 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2288 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2292 if (blk_queue_sq_sched(q))
2293 sq_hctx = blk_mq_get_sq_hctx(q);
2294 queue_for_each_hw_ctx(q, hctx, i) {
2295 if (blk_mq_hctx_stopped(hctx))
2298 * If there is already a run_work pending, leave the
2299 * pending delay untouched. Otherwise, a hctx can stall
2300 * if another hctx is re-delaying the other's work
2301 * before the work executes.
2303 if (delayed_work_pending(&hctx->run_work))
2306 * Dispatch from this hctx either if there's no hctx preferred
2307 * by IO scheduler or if it has requests that bypass the
2310 if (!sq_hctx || sq_hctx == hctx ||
2311 !list_empty_careful(&hctx->dispatch))
2312 blk_mq_delay_run_hw_queue(hctx, msecs);
2315 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2318 * This function is often used for pausing .queue_rq() by driver when
2319 * there isn't enough resource or some conditions aren't satisfied, and
2320 * BLK_STS_RESOURCE is usually returned.
2322 * We do not guarantee that dispatch can be drained or blocked
2323 * after blk_mq_stop_hw_queue() returns. Please use
2324 * blk_mq_quiesce_queue() for that requirement.
2326 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2328 cancel_delayed_work(&hctx->run_work);
2330 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2332 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2335 * This function is often used for pausing .queue_rq() by driver when
2336 * there isn't enough resource or some conditions aren't satisfied, and
2337 * BLK_STS_RESOURCE is usually returned.
2339 * We do not guarantee that dispatch can be drained or blocked
2340 * after blk_mq_stop_hw_queues() returns. Please use
2341 * blk_mq_quiesce_queue() for that requirement.
2343 void blk_mq_stop_hw_queues(struct request_queue *q)
2345 struct blk_mq_hw_ctx *hctx;
2348 queue_for_each_hw_ctx(q, hctx, i)
2349 blk_mq_stop_hw_queue(hctx);
2351 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2353 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2355 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2357 blk_mq_run_hw_queue(hctx, false);
2359 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2361 void blk_mq_start_hw_queues(struct request_queue *q)
2363 struct blk_mq_hw_ctx *hctx;
2366 queue_for_each_hw_ctx(q, hctx, i)
2367 blk_mq_start_hw_queue(hctx);
2369 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2371 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2373 if (!blk_mq_hctx_stopped(hctx))
2376 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2377 blk_mq_run_hw_queue(hctx, async);
2379 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2381 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2383 struct blk_mq_hw_ctx *hctx;
2386 queue_for_each_hw_ctx(q, hctx, i)
2387 blk_mq_start_stopped_hw_queue(hctx, async);
2389 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2391 static void blk_mq_run_work_fn(struct work_struct *work)
2393 struct blk_mq_hw_ctx *hctx;
2395 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2398 * If we are stopped, don't run the queue.
2400 if (blk_mq_hctx_stopped(hctx))
2403 __blk_mq_run_hw_queue(hctx);
2406 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2410 struct blk_mq_ctx *ctx = rq->mq_ctx;
2411 enum hctx_type type = hctx->type;
2413 lockdep_assert_held(&ctx->lock);
2415 trace_block_rq_insert(rq);
2418 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2420 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2423 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2426 struct blk_mq_ctx *ctx = rq->mq_ctx;
2428 lockdep_assert_held(&ctx->lock);
2430 __blk_mq_insert_req_list(hctx, rq, at_head);
2431 blk_mq_hctx_mark_pending(hctx, ctx);
2435 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2436 * @rq: Pointer to request to be inserted.
2437 * @at_head: true if the request should be inserted at the head of the list.
2438 * @run_queue: If we should run the hardware queue after inserting the request.
2440 * Should only be used carefully, when the caller knows we want to
2441 * bypass a potential IO scheduler on the target device.
2443 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2446 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2448 spin_lock(&hctx->lock);
2450 list_add(&rq->queuelist, &hctx->dispatch);
2452 list_add_tail(&rq->queuelist, &hctx->dispatch);
2453 spin_unlock(&hctx->lock);
2456 blk_mq_run_hw_queue(hctx, false);
2459 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2460 struct list_head *list)
2464 enum hctx_type type = hctx->type;
2467 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2470 list_for_each_entry(rq, list, queuelist) {
2471 BUG_ON(rq->mq_ctx != ctx);
2472 trace_block_rq_insert(rq);
2475 spin_lock(&ctx->lock);
2476 list_splice_tail_init(list, &ctx->rq_lists[type]);
2477 blk_mq_hctx_mark_pending(hctx, ctx);
2478 spin_unlock(&ctx->lock);
2481 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2484 if (hctx->queue->mq_ops->commit_rqs) {
2485 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2486 hctx->queue->mq_ops->commit_rqs(hctx);
2491 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2492 unsigned int nr_segs)
2496 if (bio->bi_opf & REQ_RAHEAD)
2497 rq->cmd_flags |= REQ_FAILFAST_MASK;
2499 rq->__sector = bio->bi_iter.bi_sector;
2500 blk_rq_bio_prep(rq, bio, nr_segs);
2502 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2503 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2506 blk_account_io_start(rq);
2509 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2510 struct request *rq, bool last)
2512 struct request_queue *q = rq->q;
2513 struct blk_mq_queue_data bd = {
2520 * For OK queue, we are done. For error, caller may kill it.
2521 * Any other error (busy), just add it to our list as we
2522 * previously would have done.
2524 ret = q->mq_ops->queue_rq(hctx, &bd);
2527 blk_mq_update_dispatch_busy(hctx, false);
2529 case BLK_STS_RESOURCE:
2530 case BLK_STS_DEV_RESOURCE:
2531 blk_mq_update_dispatch_busy(hctx, true);
2532 __blk_mq_requeue_request(rq);
2535 blk_mq_update_dispatch_busy(hctx, false);
2542 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2544 bool bypass_insert, bool last)
2546 struct request_queue *q = rq->q;
2547 bool run_queue = true;
2551 * RCU or SRCU read lock is needed before checking quiesced flag.
2553 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2554 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2555 * and avoid driver to try to dispatch again.
2557 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2559 bypass_insert = false;
2563 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2566 budget_token = blk_mq_get_dispatch_budget(q);
2567 if (budget_token < 0)
2570 blk_mq_set_rq_budget_token(rq, budget_token);
2572 if (!blk_mq_get_driver_tag(rq)) {
2573 blk_mq_put_dispatch_budget(q, budget_token);
2577 return __blk_mq_issue_directly(hctx, rq, last);
2580 return BLK_STS_RESOURCE;
2582 blk_mq_sched_insert_request(rq, false, run_queue, false);
2588 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2589 * @hctx: Pointer of the associated hardware queue.
2590 * @rq: Pointer to request to be sent.
2592 * If the device has enough resources to accept a new request now, send the
2593 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2594 * we can try send it another time in the future. Requests inserted at this
2595 * queue have higher priority.
2597 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2601 __blk_mq_try_issue_directly(hctx, rq, false, true);
2603 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2604 blk_mq_request_bypass_insert(rq, false, true);
2605 else if (ret != BLK_STS_OK)
2606 blk_mq_end_request(rq, ret);
2609 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2611 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2614 static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2616 struct blk_mq_hw_ctx *hctx = NULL;
2621 while ((rq = rq_list_pop(&plug->mq_list))) {
2622 bool last = rq_list_empty(plug->mq_list);
2625 if (hctx != rq->mq_hctx) {
2627 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2631 ret = blk_mq_request_issue_directly(rq, last);
2636 case BLK_STS_RESOURCE:
2637 case BLK_STS_DEV_RESOURCE:
2638 blk_mq_request_bypass_insert(rq, false, true);
2639 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2642 blk_mq_end_request(rq, ret);
2649 * If we didn't flush the entire list, we could have told the driver
2650 * there was more coming, but that turned out to be a lie.
2653 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2656 static void __blk_mq_flush_plug_list(struct request_queue *q,
2657 struct blk_plug *plug)
2659 if (blk_queue_quiesced(q))
2661 q->mq_ops->queue_rqs(&plug->mq_list);
2664 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2666 struct blk_mq_hw_ctx *this_hctx = NULL;
2667 struct blk_mq_ctx *this_ctx = NULL;
2668 struct request *requeue_list = NULL;
2669 unsigned int depth = 0;
2673 struct request *rq = rq_list_pop(&plug->mq_list);
2676 this_hctx = rq->mq_hctx;
2677 this_ctx = rq->mq_ctx;
2678 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2679 rq_list_add(&requeue_list, rq);
2682 list_add_tail(&rq->queuelist, &list);
2684 } while (!rq_list_empty(plug->mq_list));
2686 plug->mq_list = requeue_list;
2687 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2688 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched);
2691 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2695 if (rq_list_empty(plug->mq_list))
2699 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2700 struct request_queue *q;
2702 rq = rq_list_peek(&plug->mq_list);
2706 * Peek first request and see if we have a ->queue_rqs() hook.
2707 * If we do, we can dispatch the whole plug list in one go. We
2708 * already know at this point that all requests belong to the
2709 * same queue, caller must ensure that's the case.
2711 * Since we pass off the full list to the driver at this point,
2712 * we do not increment the active request count for the queue.
2713 * Bypass shared tags for now because of that.
2715 if (q->mq_ops->queue_rqs &&
2716 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2717 blk_mq_run_dispatch_ops(q,
2718 __blk_mq_flush_plug_list(q, plug));
2719 if (rq_list_empty(plug->mq_list))
2723 blk_mq_run_dispatch_ops(q,
2724 blk_mq_plug_issue_direct(plug, false));
2725 if (rq_list_empty(plug->mq_list))
2730 blk_mq_dispatch_plug_list(plug, from_schedule);
2731 } while (!rq_list_empty(plug->mq_list));
2734 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2735 struct list_head *list)
2740 while (!list_empty(list)) {
2742 struct request *rq = list_first_entry(list, struct request,
2745 list_del_init(&rq->queuelist);
2746 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2747 if (ret != BLK_STS_OK) {
2749 if (ret == BLK_STS_RESOURCE ||
2750 ret == BLK_STS_DEV_RESOURCE) {
2751 blk_mq_request_bypass_insert(rq, false,
2755 blk_mq_end_request(rq, ret);
2761 * If we didn't flush the entire list, we could have told
2762 * the driver there was more coming, but that turned out to
2765 if ((!list_empty(list) || errors) &&
2766 hctx->queue->mq_ops->commit_rqs && queued)
2767 hctx->queue->mq_ops->commit_rqs(hctx);
2770 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2771 struct bio *bio, unsigned int nr_segs)
2773 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2774 if (blk_attempt_plug_merge(q, bio, nr_segs))
2776 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2782 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2783 struct blk_plug *plug,
2787 struct blk_mq_alloc_data data = {
2790 .cmd_flags = bio->bi_opf,
2794 if (unlikely(bio_queue_enter(bio)))
2797 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2800 rq_qos_throttle(q, bio);
2803 data.nr_tags = plug->nr_ios;
2805 data.cached_rq = &plug->cached_rq;
2808 rq = __blk_mq_alloc_requests(&data);
2811 rq_qos_cleanup(q, bio);
2812 if (bio->bi_opf & REQ_NOWAIT)
2813 bio_wouldblock_error(bio);
2819 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2820 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2826 rq = rq_list_peek(&plug->cached_rq);
2827 if (!rq || rq->q != q)
2830 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2835 if (blk_mq_get_hctx_type((*bio)->bi_opf) != rq->mq_hctx->type)
2837 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2841 * If any qos ->throttle() end up blocking, we will have flushed the
2842 * plug and hence killed the cached_rq list as well. Pop this entry
2843 * before we throttle.
2845 plug->cached_rq = rq_list_next(rq);
2846 rq_qos_throttle(q, *bio);
2848 rq->cmd_flags = (*bio)->bi_opf;
2849 INIT_LIST_HEAD(&rq->queuelist);
2853 static void bio_set_ioprio(struct bio *bio)
2855 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2856 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2857 bio->bi_ioprio = get_current_ioprio();
2858 blkcg_set_ioprio(bio);
2862 * blk_mq_submit_bio - Create and send a request to block device.
2863 * @bio: Bio pointer.
2865 * Builds up a request structure from @q and @bio and send to the device. The
2866 * request may not be queued directly to hardware if:
2867 * * This request can be merged with another one
2868 * * We want to place request at plug queue for possible future merging
2869 * * There is an IO scheduler active at this queue
2871 * It will not queue the request if there is an error with the bio, or at the
2874 void blk_mq_submit_bio(struct bio *bio)
2876 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2877 struct blk_plug *plug = blk_mq_plug(bio);
2878 const int is_sync = op_is_sync(bio->bi_opf);
2880 unsigned int nr_segs = 1;
2883 bio = blk_queue_bounce(bio, q);
2884 if (bio_may_exceed_limits(bio, &q->limits))
2885 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2887 if (!bio_integrity_prep(bio))
2890 bio_set_ioprio(bio);
2892 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2896 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2901 trace_block_getrq(bio);
2903 rq_qos_track(q, rq, bio);
2905 blk_mq_bio_to_request(rq, bio, nr_segs);
2907 ret = blk_crypto_init_request(rq);
2908 if (ret != BLK_STS_OK) {
2909 bio->bi_status = ret;
2911 blk_mq_free_request(rq);
2915 if (op_is_flush(bio->bi_opf)) {
2916 blk_insert_flush(rq);
2921 blk_add_rq_to_plug(plug, rq);
2922 else if ((rq->rq_flags & RQF_ELV) ||
2923 (rq->mq_hctx->dispatch_busy &&
2924 (q->nr_hw_queues == 1 || !is_sync)))
2925 blk_mq_sched_insert_request(rq, false, true, true);
2927 blk_mq_run_dispatch_ops(rq->q,
2928 blk_mq_try_issue_directly(rq->mq_hctx, rq));
2931 #ifdef CONFIG_BLK_MQ_STACKING
2933 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
2934 * @rq: the request being queued
2936 blk_status_t blk_insert_cloned_request(struct request *rq)
2938 struct request_queue *q = rq->q;
2939 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
2942 if (blk_rq_sectors(rq) > max_sectors) {
2944 * SCSI device does not have a good way to return if
2945 * Write Same/Zero is actually supported. If a device rejects
2946 * a non-read/write command (discard, write same,etc.) the
2947 * low-level device driver will set the relevant queue limit to
2948 * 0 to prevent blk-lib from issuing more of the offending
2949 * operations. Commands queued prior to the queue limit being
2950 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
2951 * errors being propagated to upper layers.
2953 if (max_sectors == 0)
2954 return BLK_STS_NOTSUPP;
2956 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
2957 __func__, blk_rq_sectors(rq), max_sectors);
2958 return BLK_STS_IOERR;
2962 * The queue settings related to segment counting may differ from the
2965 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
2966 if (rq->nr_phys_segments > queue_max_segments(q)) {
2967 printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n",
2968 __func__, rq->nr_phys_segments, queue_max_segments(q));
2969 return BLK_STS_IOERR;
2972 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
2973 return BLK_STS_IOERR;
2975 if (blk_crypto_insert_cloned_request(rq))
2976 return BLK_STS_IOERR;
2978 blk_account_io_start(rq);
2981 * Since we have a scheduler attached on the top device,
2982 * bypass a potential scheduler on the bottom device for
2985 blk_mq_run_dispatch_ops(q,
2986 ret = blk_mq_request_issue_directly(rq, true));
2988 blk_account_io_done(rq, ktime_get_ns());
2991 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
2994 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
2995 * @rq: the clone request to be cleaned up
2998 * Free all bios in @rq for a cloned request.
3000 void blk_rq_unprep_clone(struct request *rq)
3004 while ((bio = rq->bio) != NULL) {
3005 rq->bio = bio->bi_next;
3010 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3013 * blk_rq_prep_clone - Helper function to setup clone request
3014 * @rq: the request to be setup
3015 * @rq_src: original request to be cloned
3016 * @bs: bio_set that bios for clone are allocated from
3017 * @gfp_mask: memory allocation mask for bio
3018 * @bio_ctr: setup function to be called for each clone bio.
3019 * Returns %0 for success, non %0 for failure.
3020 * @data: private data to be passed to @bio_ctr
3023 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3024 * Also, pages which the original bios are pointing to are not copied
3025 * and the cloned bios just point same pages.
3026 * So cloned bios must be completed before original bios, which means
3027 * the caller must complete @rq before @rq_src.
3029 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3030 struct bio_set *bs, gfp_t gfp_mask,
3031 int (*bio_ctr)(struct bio *, struct bio *, void *),
3034 struct bio *bio, *bio_src;
3039 __rq_for_each_bio(bio_src, rq_src) {
3040 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3045 if (bio_ctr && bio_ctr(bio, bio_src, data))
3049 rq->biotail->bi_next = bio;
3052 rq->bio = rq->biotail = bio;
3057 /* Copy attributes of the original request to the clone request. */
3058 rq->__sector = blk_rq_pos(rq_src);
3059 rq->__data_len = blk_rq_bytes(rq_src);
3060 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3061 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3062 rq->special_vec = rq_src->special_vec;
3064 rq->nr_phys_segments = rq_src->nr_phys_segments;
3065 rq->ioprio = rq_src->ioprio;
3067 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3075 blk_rq_unprep_clone(rq);
3079 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3080 #endif /* CONFIG_BLK_MQ_STACKING */
3083 * Steal bios from a request and add them to a bio list.
3084 * The request must not have been partially completed before.
3086 void blk_steal_bios(struct bio_list *list, struct request *rq)
3090 list->tail->bi_next = rq->bio;
3092 list->head = rq->bio;
3093 list->tail = rq->biotail;
3101 EXPORT_SYMBOL_GPL(blk_steal_bios);
3103 static size_t order_to_size(unsigned int order)
3105 return (size_t)PAGE_SIZE << order;
3108 /* called before freeing request pool in @tags */
3109 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3110 struct blk_mq_tags *tags)
3113 unsigned long flags;
3116 * There is no need to clear mapping if driver tags is not initialized
3117 * or the mapping belongs to the driver tags.
3119 if (!drv_tags || drv_tags == tags)
3122 list_for_each_entry(page, &tags->page_list, lru) {
3123 unsigned long start = (unsigned long)page_address(page);
3124 unsigned long end = start + order_to_size(page->private);
3127 for (i = 0; i < drv_tags->nr_tags; i++) {
3128 struct request *rq = drv_tags->rqs[i];
3129 unsigned long rq_addr = (unsigned long)rq;
3131 if (rq_addr >= start && rq_addr < end) {
3132 WARN_ON_ONCE(req_ref_read(rq) != 0);
3133 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3139 * Wait until all pending iteration is done.
3141 * Request reference is cleared and it is guaranteed to be observed
3142 * after the ->lock is released.
3144 spin_lock_irqsave(&drv_tags->lock, flags);
3145 spin_unlock_irqrestore(&drv_tags->lock, flags);
3148 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3149 unsigned int hctx_idx)
3151 struct blk_mq_tags *drv_tags;
3154 if (list_empty(&tags->page_list))
3157 if (blk_mq_is_shared_tags(set->flags))
3158 drv_tags = set->shared_tags;
3160 drv_tags = set->tags[hctx_idx];
3162 if (tags->static_rqs && set->ops->exit_request) {
3165 for (i = 0; i < tags->nr_tags; i++) {
3166 struct request *rq = tags->static_rqs[i];
3170 set->ops->exit_request(set, rq, hctx_idx);
3171 tags->static_rqs[i] = NULL;
3175 blk_mq_clear_rq_mapping(drv_tags, tags);
3177 while (!list_empty(&tags->page_list)) {
3178 page = list_first_entry(&tags->page_list, struct page, lru);
3179 list_del_init(&page->lru);
3181 * Remove kmemleak object previously allocated in
3182 * blk_mq_alloc_rqs().
3184 kmemleak_free(page_address(page));
3185 __free_pages(page, page->private);
3189 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3193 kfree(tags->static_rqs);
3194 tags->static_rqs = NULL;
3196 blk_mq_free_tags(tags);
3199 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3200 unsigned int hctx_idx)
3204 for (i = 0; i < set->nr_maps; i++) {
3205 unsigned int start = set->map[i].queue_offset;
3206 unsigned int end = start + set->map[i].nr_queues;
3208 if (hctx_idx >= start && hctx_idx < end)
3212 if (i >= set->nr_maps)
3213 i = HCTX_TYPE_DEFAULT;
3218 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3219 unsigned int hctx_idx)
3221 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3223 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3226 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3227 unsigned int hctx_idx,
3228 unsigned int nr_tags,
3229 unsigned int reserved_tags)
3231 int node = blk_mq_get_hctx_node(set, hctx_idx);
3232 struct blk_mq_tags *tags;
3234 if (node == NUMA_NO_NODE)
3235 node = set->numa_node;
3237 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3238 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3242 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3243 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3246 blk_mq_free_tags(tags);
3250 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3251 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3253 if (!tags->static_rqs) {
3255 blk_mq_free_tags(tags);
3262 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3263 unsigned int hctx_idx, int node)
3267 if (set->ops->init_request) {
3268 ret = set->ops->init_request(set, rq, hctx_idx, node);
3273 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3277 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3278 struct blk_mq_tags *tags,
3279 unsigned int hctx_idx, unsigned int depth)
3281 unsigned int i, j, entries_per_page, max_order = 4;
3282 int node = blk_mq_get_hctx_node(set, hctx_idx);
3283 size_t rq_size, left;
3285 if (node == NUMA_NO_NODE)
3286 node = set->numa_node;
3288 INIT_LIST_HEAD(&tags->page_list);
3291 * rq_size is the size of the request plus driver payload, rounded
3292 * to the cacheline size
3294 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3296 left = rq_size * depth;
3298 for (i = 0; i < depth; ) {
3299 int this_order = max_order;
3304 while (this_order && left < order_to_size(this_order - 1))
3308 page = alloc_pages_node(node,
3309 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3315 if (order_to_size(this_order) < rq_size)
3322 page->private = this_order;
3323 list_add_tail(&page->lru, &tags->page_list);
3325 p = page_address(page);
3327 * Allow kmemleak to scan these pages as they contain pointers
3328 * to additional allocations like via ops->init_request().
3330 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3331 entries_per_page = order_to_size(this_order) / rq_size;
3332 to_do = min(entries_per_page, depth - i);
3333 left -= to_do * rq_size;
3334 for (j = 0; j < to_do; j++) {
3335 struct request *rq = p;
3337 tags->static_rqs[i] = rq;
3338 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3339 tags->static_rqs[i] = NULL;
3350 blk_mq_free_rqs(set, tags, hctx_idx);
3354 struct rq_iter_data {
3355 struct blk_mq_hw_ctx *hctx;
3359 static bool blk_mq_has_request(struct request *rq, void *data)
3361 struct rq_iter_data *iter_data = data;
3363 if (rq->mq_hctx != iter_data->hctx)
3365 iter_data->has_rq = true;
3369 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3371 struct blk_mq_tags *tags = hctx->sched_tags ?
3372 hctx->sched_tags : hctx->tags;
3373 struct rq_iter_data data = {
3377 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3381 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3382 struct blk_mq_hw_ctx *hctx)
3384 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3386 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3391 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3393 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3394 struct blk_mq_hw_ctx, cpuhp_online);
3396 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3397 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3401 * Prevent new request from being allocated on the current hctx.
3403 * The smp_mb__after_atomic() Pairs with the implied barrier in
3404 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3405 * seen once we return from the tag allocator.
3407 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3408 smp_mb__after_atomic();
3411 * Try to grab a reference to the queue and wait for any outstanding
3412 * requests. If we could not grab a reference the queue has been
3413 * frozen and there are no requests.
3415 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3416 while (blk_mq_hctx_has_requests(hctx))
3418 percpu_ref_put(&hctx->queue->q_usage_counter);
3424 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3426 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3427 struct blk_mq_hw_ctx, cpuhp_online);
3429 if (cpumask_test_cpu(cpu, hctx->cpumask))
3430 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3435 * 'cpu' is going away. splice any existing rq_list entries from this
3436 * software queue to the hw queue dispatch list, and ensure that it
3439 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3441 struct blk_mq_hw_ctx *hctx;
3442 struct blk_mq_ctx *ctx;
3444 enum hctx_type type;
3446 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3447 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3450 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3453 spin_lock(&ctx->lock);
3454 if (!list_empty(&ctx->rq_lists[type])) {
3455 list_splice_init(&ctx->rq_lists[type], &tmp);
3456 blk_mq_hctx_clear_pending(hctx, ctx);
3458 spin_unlock(&ctx->lock);
3460 if (list_empty(&tmp))
3463 spin_lock(&hctx->lock);
3464 list_splice_tail_init(&tmp, &hctx->dispatch);
3465 spin_unlock(&hctx->lock);
3467 blk_mq_run_hw_queue(hctx, true);
3471 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3473 if (!(hctx->flags & BLK_MQ_F_STACKING))
3474 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3475 &hctx->cpuhp_online);
3476 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3481 * Before freeing hw queue, clearing the flush request reference in
3482 * tags->rqs[] for avoiding potential UAF.
3484 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3485 unsigned int queue_depth, struct request *flush_rq)
3488 unsigned long flags;
3490 /* The hw queue may not be mapped yet */
3494 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3496 for (i = 0; i < queue_depth; i++)
3497 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3500 * Wait until all pending iteration is done.
3502 * Request reference is cleared and it is guaranteed to be observed
3503 * after the ->lock is released.
3505 spin_lock_irqsave(&tags->lock, flags);
3506 spin_unlock_irqrestore(&tags->lock, flags);
3509 /* hctx->ctxs will be freed in queue's release handler */
3510 static void blk_mq_exit_hctx(struct request_queue *q,
3511 struct blk_mq_tag_set *set,
3512 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3514 struct request *flush_rq = hctx->fq->flush_rq;
3516 if (blk_mq_hw_queue_mapped(hctx))
3517 blk_mq_tag_idle(hctx);
3519 if (blk_queue_init_done(q))
3520 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3521 set->queue_depth, flush_rq);
3522 if (set->ops->exit_request)
3523 set->ops->exit_request(set, flush_rq, hctx_idx);
3525 if (set->ops->exit_hctx)
3526 set->ops->exit_hctx(hctx, hctx_idx);
3528 blk_mq_remove_cpuhp(hctx);
3530 xa_erase(&q->hctx_table, hctx_idx);
3532 spin_lock(&q->unused_hctx_lock);
3533 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3534 spin_unlock(&q->unused_hctx_lock);
3537 static void blk_mq_exit_hw_queues(struct request_queue *q,
3538 struct blk_mq_tag_set *set, int nr_queue)
3540 struct blk_mq_hw_ctx *hctx;
3543 queue_for_each_hw_ctx(q, hctx, i) {
3546 blk_mq_exit_hctx(q, set, hctx, i);
3550 static int blk_mq_init_hctx(struct request_queue *q,
3551 struct blk_mq_tag_set *set,
3552 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3554 hctx->queue_num = hctx_idx;
3556 if (!(hctx->flags & BLK_MQ_F_STACKING))
3557 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3558 &hctx->cpuhp_online);
3559 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3561 hctx->tags = set->tags[hctx_idx];
3563 if (set->ops->init_hctx &&
3564 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3565 goto unregister_cpu_notifier;
3567 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3571 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3577 if (set->ops->exit_request)
3578 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3580 if (set->ops->exit_hctx)
3581 set->ops->exit_hctx(hctx, hctx_idx);
3582 unregister_cpu_notifier:
3583 blk_mq_remove_cpuhp(hctx);
3587 static struct blk_mq_hw_ctx *
3588 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3591 struct blk_mq_hw_ctx *hctx;
3592 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3594 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3596 goto fail_alloc_hctx;
3598 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3601 atomic_set(&hctx->nr_active, 0);
3602 if (node == NUMA_NO_NODE)
3603 node = set->numa_node;
3604 hctx->numa_node = node;
3606 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3607 spin_lock_init(&hctx->lock);
3608 INIT_LIST_HEAD(&hctx->dispatch);
3610 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3612 INIT_LIST_HEAD(&hctx->hctx_list);
3615 * Allocate space for all possible cpus to avoid allocation at
3618 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3623 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3624 gfp, node, false, false))
3628 spin_lock_init(&hctx->dispatch_wait_lock);
3629 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3630 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3632 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3636 blk_mq_hctx_kobj_init(hctx);
3641 sbitmap_free(&hctx->ctx_map);
3645 free_cpumask_var(hctx->cpumask);
3652 static void blk_mq_init_cpu_queues(struct request_queue *q,
3653 unsigned int nr_hw_queues)
3655 struct blk_mq_tag_set *set = q->tag_set;
3658 for_each_possible_cpu(i) {
3659 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3660 struct blk_mq_hw_ctx *hctx;
3664 spin_lock_init(&__ctx->lock);
3665 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3666 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3671 * Set local node, IFF we have more than one hw queue. If
3672 * not, we remain on the home node of the device
3674 for (j = 0; j < set->nr_maps; j++) {
3675 hctx = blk_mq_map_queue_type(q, j, i);
3676 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3677 hctx->numa_node = cpu_to_node(i);
3682 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3683 unsigned int hctx_idx,
3686 struct blk_mq_tags *tags;
3689 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3693 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3695 blk_mq_free_rq_map(tags);
3702 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3705 if (blk_mq_is_shared_tags(set->flags)) {
3706 set->tags[hctx_idx] = set->shared_tags;
3711 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3714 return set->tags[hctx_idx];
3717 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3718 struct blk_mq_tags *tags,
3719 unsigned int hctx_idx)
3722 blk_mq_free_rqs(set, tags, hctx_idx);
3723 blk_mq_free_rq_map(tags);
3727 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3728 unsigned int hctx_idx)
3730 if (!blk_mq_is_shared_tags(set->flags))
3731 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3733 set->tags[hctx_idx] = NULL;
3736 static void blk_mq_map_swqueue(struct request_queue *q)
3738 unsigned int j, hctx_idx;
3740 struct blk_mq_hw_ctx *hctx;
3741 struct blk_mq_ctx *ctx;
3742 struct blk_mq_tag_set *set = q->tag_set;
3744 queue_for_each_hw_ctx(q, hctx, i) {
3745 cpumask_clear(hctx->cpumask);
3747 hctx->dispatch_from = NULL;
3751 * Map software to hardware queues.
3753 * If the cpu isn't present, the cpu is mapped to first hctx.
3755 for_each_possible_cpu(i) {
3757 ctx = per_cpu_ptr(q->queue_ctx, i);
3758 for (j = 0; j < set->nr_maps; j++) {
3759 if (!set->map[j].nr_queues) {
3760 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3761 HCTX_TYPE_DEFAULT, i);
3764 hctx_idx = set->map[j].mq_map[i];
3765 /* unmapped hw queue can be remapped after CPU topo changed */
3766 if (!set->tags[hctx_idx] &&
3767 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3769 * If tags initialization fail for some hctx,
3770 * that hctx won't be brought online. In this
3771 * case, remap the current ctx to hctx[0] which
3772 * is guaranteed to always have tags allocated
3774 set->map[j].mq_map[i] = 0;
3777 hctx = blk_mq_map_queue_type(q, j, i);
3778 ctx->hctxs[j] = hctx;
3780 * If the CPU is already set in the mask, then we've
3781 * mapped this one already. This can happen if
3782 * devices share queues across queue maps.
3784 if (cpumask_test_cpu(i, hctx->cpumask))
3787 cpumask_set_cpu(i, hctx->cpumask);
3789 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3790 hctx->ctxs[hctx->nr_ctx++] = ctx;
3793 * If the nr_ctx type overflows, we have exceeded the
3794 * amount of sw queues we can support.
3796 BUG_ON(!hctx->nr_ctx);
3799 for (; j < HCTX_MAX_TYPES; j++)
3800 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3801 HCTX_TYPE_DEFAULT, i);
3804 queue_for_each_hw_ctx(q, hctx, i) {
3806 * If no software queues are mapped to this hardware queue,
3807 * disable it and free the request entries.
3809 if (!hctx->nr_ctx) {
3810 /* Never unmap queue 0. We need it as a
3811 * fallback in case of a new remap fails
3815 __blk_mq_free_map_and_rqs(set, i);
3821 hctx->tags = set->tags[i];
3822 WARN_ON(!hctx->tags);
3825 * Set the map size to the number of mapped software queues.
3826 * This is more accurate and more efficient than looping
3827 * over all possibly mapped software queues.
3829 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3832 * Initialize batch roundrobin counts
3834 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3835 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3840 * Caller needs to ensure that we're either frozen/quiesced, or that
3841 * the queue isn't live yet.
3843 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3845 struct blk_mq_hw_ctx *hctx;
3848 queue_for_each_hw_ctx(q, hctx, i) {
3850 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3852 blk_mq_tag_idle(hctx);
3853 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3858 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3861 struct request_queue *q;
3863 lockdep_assert_held(&set->tag_list_lock);
3865 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3866 blk_mq_freeze_queue(q);
3867 queue_set_hctx_shared(q, shared);
3868 blk_mq_unfreeze_queue(q);
3872 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3874 struct blk_mq_tag_set *set = q->tag_set;
3876 mutex_lock(&set->tag_list_lock);
3877 list_del(&q->tag_set_list);
3878 if (list_is_singular(&set->tag_list)) {
3879 /* just transitioned to unshared */
3880 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3881 /* update existing queue */
3882 blk_mq_update_tag_set_shared(set, false);
3884 mutex_unlock(&set->tag_list_lock);
3885 INIT_LIST_HEAD(&q->tag_set_list);
3888 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3889 struct request_queue *q)
3891 mutex_lock(&set->tag_list_lock);
3894 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3896 if (!list_empty(&set->tag_list) &&
3897 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3898 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3899 /* update existing queue */
3900 blk_mq_update_tag_set_shared(set, true);
3902 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3903 queue_set_hctx_shared(q, true);
3904 list_add_tail(&q->tag_set_list, &set->tag_list);
3906 mutex_unlock(&set->tag_list_lock);
3909 /* All allocations will be freed in release handler of q->mq_kobj */
3910 static int blk_mq_alloc_ctxs(struct request_queue *q)
3912 struct blk_mq_ctxs *ctxs;
3915 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3919 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3920 if (!ctxs->queue_ctx)
3923 for_each_possible_cpu(cpu) {
3924 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3928 q->mq_kobj = &ctxs->kobj;
3929 q->queue_ctx = ctxs->queue_ctx;
3938 * It is the actual release handler for mq, but we do it from
3939 * request queue's release handler for avoiding use-after-free
3940 * and headache because q->mq_kobj shouldn't have been introduced,
3941 * but we can't group ctx/kctx kobj without it.
3943 void blk_mq_release(struct request_queue *q)
3945 struct blk_mq_hw_ctx *hctx, *next;
3948 queue_for_each_hw_ctx(q, hctx, i)
3949 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3951 /* all hctx are in .unused_hctx_list now */
3952 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3953 list_del_init(&hctx->hctx_list);
3954 kobject_put(&hctx->kobj);
3957 xa_destroy(&q->hctx_table);
3960 * release .mq_kobj and sw queue's kobject now because
3961 * both share lifetime with request queue.
3963 blk_mq_sysfs_deinit(q);
3966 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3969 struct request_queue *q;
3972 q = blk_alloc_queue(set->numa_node, set->flags & BLK_MQ_F_BLOCKING);
3974 return ERR_PTR(-ENOMEM);
3975 q->queuedata = queuedata;
3976 ret = blk_mq_init_allocated_queue(set, q);
3979 return ERR_PTR(ret);
3984 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3986 return blk_mq_init_queue_data(set, NULL);
3988 EXPORT_SYMBOL(blk_mq_init_queue);
3991 * blk_mq_destroy_queue - shutdown a request queue
3992 * @q: request queue to shutdown
3994 * This shuts down a request queue allocated by blk_mq_init_queue() and drops
3995 * the initial reference. All future requests will failed with -ENODEV.
3997 * Context: can sleep
3999 void blk_mq_destroy_queue(struct request_queue *q)
4001 WARN_ON_ONCE(!queue_is_mq(q));
4002 WARN_ON_ONCE(blk_queue_registered(q));
4006 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4007 blk_queue_start_drain(q);
4008 blk_freeze_queue(q);
4011 blk_mq_cancel_work_sync(q);
4012 blk_mq_exit_queue(q);
4014 /* @q is and will stay empty, shutdown and put */
4017 EXPORT_SYMBOL(blk_mq_destroy_queue);
4019 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4020 struct lock_class_key *lkclass)
4022 struct request_queue *q;
4023 struct gendisk *disk;
4025 q = blk_mq_init_queue_data(set, queuedata);
4029 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4031 blk_mq_destroy_queue(q);
4032 return ERR_PTR(-ENOMEM);
4034 set_bit(GD_OWNS_QUEUE, &disk->state);
4037 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4039 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4040 struct lock_class_key *lkclass)
4042 if (!blk_get_queue(q))
4044 return __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4046 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4048 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4049 struct blk_mq_tag_set *set, struct request_queue *q,
4050 int hctx_idx, int node)
4052 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4054 /* reuse dead hctx first */
4055 spin_lock(&q->unused_hctx_lock);
4056 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4057 if (tmp->numa_node == node) {
4063 list_del_init(&hctx->hctx_list);
4064 spin_unlock(&q->unused_hctx_lock);
4067 hctx = blk_mq_alloc_hctx(q, set, node);
4071 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4077 kobject_put(&hctx->kobj);
4082 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4083 struct request_queue *q)
4085 struct blk_mq_hw_ctx *hctx;
4088 /* protect against switching io scheduler */
4089 mutex_lock(&q->sysfs_lock);
4090 for (i = 0; i < set->nr_hw_queues; i++) {
4092 int node = blk_mq_get_hctx_node(set, i);
4093 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4096 old_node = old_hctx->numa_node;
4097 blk_mq_exit_hctx(q, set, old_hctx, i);
4100 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4103 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4105 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4106 WARN_ON_ONCE(!hctx);
4110 * Increasing nr_hw_queues fails. Free the newly allocated
4111 * hctxs and keep the previous q->nr_hw_queues.
4113 if (i != set->nr_hw_queues) {
4114 j = q->nr_hw_queues;
4117 q->nr_hw_queues = set->nr_hw_queues;
4120 xa_for_each_start(&q->hctx_table, j, hctx, j)
4121 blk_mq_exit_hctx(q, set, hctx, j);
4122 mutex_unlock(&q->sysfs_lock);
4125 static void blk_mq_update_poll_flag(struct request_queue *q)
4127 struct blk_mq_tag_set *set = q->tag_set;
4129 if (set->nr_maps > HCTX_TYPE_POLL &&
4130 set->map[HCTX_TYPE_POLL].nr_queues)
4131 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4133 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4136 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4137 struct request_queue *q)
4139 WARN_ON_ONCE(blk_queue_has_srcu(q) !=
4140 !!(set->flags & BLK_MQ_F_BLOCKING));
4142 /* mark the queue as mq asap */
4143 q->mq_ops = set->ops;
4145 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
4146 blk_mq_poll_stats_bkt,
4147 BLK_MQ_POLL_STATS_BKTS, q);
4151 if (blk_mq_alloc_ctxs(q))
4154 /* init q->mq_kobj and sw queues' kobjects */
4155 blk_mq_sysfs_init(q);
4157 INIT_LIST_HEAD(&q->unused_hctx_list);
4158 spin_lock_init(&q->unused_hctx_lock);
4160 xa_init(&q->hctx_table);
4162 blk_mq_realloc_hw_ctxs(set, q);
4163 if (!q->nr_hw_queues)
4166 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4167 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4171 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4172 blk_mq_update_poll_flag(q);
4174 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4175 INIT_LIST_HEAD(&q->requeue_list);
4176 spin_lock_init(&q->requeue_lock);
4178 q->nr_requests = set->queue_depth;
4181 * Default to classic polling
4183 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
4185 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4186 blk_mq_add_queue_tag_set(set, q);
4187 blk_mq_map_swqueue(q);
4191 xa_destroy(&q->hctx_table);
4192 q->nr_hw_queues = 0;
4193 blk_mq_sysfs_deinit(q);
4195 blk_stat_free_callback(q->poll_cb);
4201 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4203 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4204 void blk_mq_exit_queue(struct request_queue *q)
4206 struct blk_mq_tag_set *set = q->tag_set;
4208 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4209 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4210 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4211 blk_mq_del_queue_tag_set(q);
4214 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4218 if (blk_mq_is_shared_tags(set->flags)) {
4219 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4222 if (!set->shared_tags)
4226 for (i = 0; i < set->nr_hw_queues; i++) {
4227 if (!__blk_mq_alloc_map_and_rqs(set, i))
4236 __blk_mq_free_map_and_rqs(set, i);
4238 if (blk_mq_is_shared_tags(set->flags)) {
4239 blk_mq_free_map_and_rqs(set, set->shared_tags,
4240 BLK_MQ_NO_HCTX_IDX);
4247 * Allocate the request maps associated with this tag_set. Note that this
4248 * may reduce the depth asked for, if memory is tight. set->queue_depth
4249 * will be updated to reflect the allocated depth.
4251 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4256 depth = set->queue_depth;
4258 err = __blk_mq_alloc_rq_maps(set);
4262 set->queue_depth >>= 1;
4263 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4267 } while (set->queue_depth);
4269 if (!set->queue_depth || err) {
4270 pr_err("blk-mq: failed to allocate request map\n");
4274 if (depth != set->queue_depth)
4275 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4276 depth, set->queue_depth);
4281 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4284 * blk_mq_map_queues() and multiple .map_queues() implementations
4285 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4286 * number of hardware queues.
4288 if (set->nr_maps == 1)
4289 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4291 if (set->ops->map_queues && !is_kdump_kernel()) {
4295 * transport .map_queues is usually done in the following
4298 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4299 * mask = get_cpu_mask(queue)
4300 * for_each_cpu(cpu, mask)
4301 * set->map[x].mq_map[cpu] = queue;
4304 * When we need to remap, the table has to be cleared for
4305 * killing stale mapping since one CPU may not be mapped
4308 for (i = 0; i < set->nr_maps; i++)
4309 blk_mq_clear_mq_map(&set->map[i]);
4311 set->ops->map_queues(set);
4313 BUG_ON(set->nr_maps > 1);
4314 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4318 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4319 int cur_nr_hw_queues, int new_nr_hw_queues)
4321 struct blk_mq_tags **new_tags;
4323 if (cur_nr_hw_queues >= new_nr_hw_queues)
4326 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4327 GFP_KERNEL, set->numa_node);
4332 memcpy(new_tags, set->tags, cur_nr_hw_queues *
4333 sizeof(*set->tags));
4335 set->tags = new_tags;
4336 set->nr_hw_queues = new_nr_hw_queues;
4341 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
4342 int new_nr_hw_queues)
4344 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
4348 * Alloc a tag set to be associated with one or more request queues.
4349 * May fail with EINVAL for various error conditions. May adjust the
4350 * requested depth down, if it's too large. In that case, the set
4351 * value will be stored in set->queue_depth.
4353 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4357 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4359 if (!set->nr_hw_queues)
4361 if (!set->queue_depth)
4363 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4366 if (!set->ops->queue_rq)
4369 if (!set->ops->get_budget ^ !set->ops->put_budget)
4372 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4373 pr_info("blk-mq: reduced tag depth to %u\n",
4375 set->queue_depth = BLK_MQ_MAX_DEPTH;
4380 else if (set->nr_maps > HCTX_MAX_TYPES)
4384 * If a crashdump is active, then we are potentially in a very
4385 * memory constrained environment. Limit us to 1 queue and
4386 * 64 tags to prevent using too much memory.
4388 if (is_kdump_kernel()) {
4389 set->nr_hw_queues = 1;
4391 set->queue_depth = min(64U, set->queue_depth);
4394 * There is no use for more h/w queues than cpus if we just have
4397 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4398 set->nr_hw_queues = nr_cpu_ids;
4400 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
4404 for (i = 0; i < set->nr_maps; i++) {
4405 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4406 sizeof(set->map[i].mq_map[0]),
4407 GFP_KERNEL, set->numa_node);
4408 if (!set->map[i].mq_map)
4409 goto out_free_mq_map;
4410 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4413 blk_mq_update_queue_map(set);
4415 ret = blk_mq_alloc_set_map_and_rqs(set);
4417 goto out_free_mq_map;
4419 mutex_init(&set->tag_list_lock);
4420 INIT_LIST_HEAD(&set->tag_list);
4425 for (i = 0; i < set->nr_maps; i++) {
4426 kfree(set->map[i].mq_map);
4427 set->map[i].mq_map = NULL;
4433 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4435 /* allocate and initialize a tagset for a simple single-queue device */
4436 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4437 const struct blk_mq_ops *ops, unsigned int queue_depth,
4438 unsigned int set_flags)
4440 memset(set, 0, sizeof(*set));
4442 set->nr_hw_queues = 1;
4444 set->queue_depth = queue_depth;
4445 set->numa_node = NUMA_NO_NODE;
4446 set->flags = set_flags;
4447 return blk_mq_alloc_tag_set(set);
4449 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4451 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4455 for (i = 0; i < set->nr_hw_queues; i++)
4456 __blk_mq_free_map_and_rqs(set, i);
4458 if (blk_mq_is_shared_tags(set->flags)) {
4459 blk_mq_free_map_and_rqs(set, set->shared_tags,
4460 BLK_MQ_NO_HCTX_IDX);
4463 for (j = 0; j < set->nr_maps; j++) {
4464 kfree(set->map[j].mq_map);
4465 set->map[j].mq_map = NULL;
4471 EXPORT_SYMBOL(blk_mq_free_tag_set);
4473 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4475 struct blk_mq_tag_set *set = q->tag_set;
4476 struct blk_mq_hw_ctx *hctx;
4483 if (q->nr_requests == nr)
4486 blk_mq_freeze_queue(q);
4487 blk_mq_quiesce_queue(q);
4490 queue_for_each_hw_ctx(q, hctx, i) {
4494 * If we're using an MQ scheduler, just update the scheduler
4495 * queue depth. This is similar to what the old code would do.
4497 if (hctx->sched_tags) {
4498 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4501 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4506 if (q->elevator && q->elevator->type->ops.depth_updated)
4507 q->elevator->type->ops.depth_updated(hctx);
4510 q->nr_requests = nr;
4511 if (blk_mq_is_shared_tags(set->flags)) {
4513 blk_mq_tag_update_sched_shared_tags(q);
4515 blk_mq_tag_resize_shared_tags(set, nr);
4519 blk_mq_unquiesce_queue(q);
4520 blk_mq_unfreeze_queue(q);
4526 * request_queue and elevator_type pair.
4527 * It is just used by __blk_mq_update_nr_hw_queues to cache
4528 * the elevator_type associated with a request_queue.
4530 struct blk_mq_qe_pair {
4531 struct list_head node;
4532 struct request_queue *q;
4533 struct elevator_type *type;
4537 * Cache the elevator_type in qe pair list and switch the
4538 * io scheduler to 'none'
4540 static bool blk_mq_elv_switch_none(struct list_head *head,
4541 struct request_queue *q)
4543 struct blk_mq_qe_pair *qe;
4548 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4552 /* q->elevator needs protection from ->sysfs_lock */
4553 mutex_lock(&q->sysfs_lock);
4555 INIT_LIST_HEAD(&qe->node);
4557 qe->type = q->elevator->type;
4558 list_add(&qe->node, head);
4561 * After elevator_switch, the previous elevator_queue will be
4562 * released by elevator_release. The reference of the io scheduler
4563 * module get by elevator_get will also be put. So we need to get
4564 * a reference of the io scheduler module here to prevent it to be
4567 __module_get(qe->type->elevator_owner);
4568 elevator_switch(q, NULL);
4569 mutex_unlock(&q->sysfs_lock);
4574 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4575 struct request_queue *q)
4577 struct blk_mq_qe_pair *qe;
4579 list_for_each_entry(qe, head, node)
4586 static void blk_mq_elv_switch_back(struct list_head *head,
4587 struct request_queue *q)
4589 struct blk_mq_qe_pair *qe;
4590 struct elevator_type *t;
4592 qe = blk_lookup_qe_pair(head, q);
4596 list_del(&qe->node);
4599 mutex_lock(&q->sysfs_lock);
4600 elevator_switch(q, t);
4601 mutex_unlock(&q->sysfs_lock);
4604 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4607 struct request_queue *q;
4609 int prev_nr_hw_queues;
4611 lockdep_assert_held(&set->tag_list_lock);
4613 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4614 nr_hw_queues = nr_cpu_ids;
4615 if (nr_hw_queues < 1)
4617 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4620 list_for_each_entry(q, &set->tag_list, tag_set_list)
4621 blk_mq_freeze_queue(q);
4623 * Switch IO scheduler to 'none', cleaning up the data associated
4624 * with the previous scheduler. We will switch back once we are done
4625 * updating the new sw to hw queue mappings.
4627 list_for_each_entry(q, &set->tag_list, tag_set_list)
4628 if (!blk_mq_elv_switch_none(&head, q))
4631 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4632 blk_mq_debugfs_unregister_hctxs(q);
4633 blk_mq_sysfs_unregister_hctxs(q);
4636 prev_nr_hw_queues = set->nr_hw_queues;
4637 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
4641 set->nr_hw_queues = nr_hw_queues;
4643 blk_mq_update_queue_map(set);
4644 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4645 blk_mq_realloc_hw_ctxs(set, q);
4646 blk_mq_update_poll_flag(q);
4647 if (q->nr_hw_queues != set->nr_hw_queues) {
4648 int i = prev_nr_hw_queues;
4650 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4651 nr_hw_queues, prev_nr_hw_queues);
4652 for (; i < set->nr_hw_queues; i++)
4653 __blk_mq_free_map_and_rqs(set, i);
4655 set->nr_hw_queues = prev_nr_hw_queues;
4656 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4659 blk_mq_map_swqueue(q);
4663 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4664 blk_mq_sysfs_register_hctxs(q);
4665 blk_mq_debugfs_register_hctxs(q);
4669 list_for_each_entry(q, &set->tag_list, tag_set_list)
4670 blk_mq_elv_switch_back(&head, q);
4672 list_for_each_entry(q, &set->tag_list, tag_set_list)
4673 blk_mq_unfreeze_queue(q);
4676 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4678 mutex_lock(&set->tag_list_lock);
4679 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4680 mutex_unlock(&set->tag_list_lock);
4682 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4684 /* Enable polling stats and return whether they were already enabled. */
4685 static bool blk_poll_stats_enable(struct request_queue *q)
4690 return blk_stats_alloc_enable(q);
4693 static void blk_mq_poll_stats_start(struct request_queue *q)
4696 * We don't arm the callback if polling stats are not enabled or the
4697 * callback is already active.
4699 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
4702 blk_stat_activate_msecs(q->poll_cb, 100);
4705 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4707 struct request_queue *q = cb->data;
4710 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4711 if (cb->stat[bucket].nr_samples)
4712 q->poll_stat[bucket] = cb->stat[bucket];
4716 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4719 unsigned long ret = 0;
4723 * If stats collection isn't on, don't sleep but turn it on for
4726 if (!blk_poll_stats_enable(q))
4730 * As an optimistic guess, use half of the mean service time
4731 * for this type of request. We can (and should) make this smarter.
4732 * For instance, if the completion latencies are tight, we can
4733 * get closer than just half the mean. This is especially
4734 * important on devices where the completion latencies are longer
4735 * than ~10 usec. We do use the stats for the relevant IO size
4736 * if available which does lead to better estimates.
4738 bucket = blk_mq_poll_stats_bkt(rq);
4742 if (q->poll_stat[bucket].nr_samples)
4743 ret = (q->poll_stat[bucket].mean + 1) / 2;
4748 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4750 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4751 struct request *rq = blk_qc_to_rq(hctx, qc);
4752 struct hrtimer_sleeper hs;
4753 enum hrtimer_mode mode;
4758 * If a request has completed on queue that uses an I/O scheduler, we
4759 * won't get back a request from blk_qc_to_rq.
4761 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4765 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4767 * 0: use half of prev avg
4768 * >0: use this specific value
4770 if (q->poll_nsec > 0)
4771 nsecs = q->poll_nsec;
4773 nsecs = blk_mq_poll_nsecs(q, rq);
4778 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4781 * This will be replaced with the stats tracking code, using
4782 * 'avg_completion_time / 2' as the pre-sleep target.
4786 mode = HRTIMER_MODE_REL;
4787 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4788 hrtimer_set_expires(&hs.timer, kt);
4791 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4793 set_current_state(TASK_UNINTERRUPTIBLE);
4794 hrtimer_sleeper_start_expires(&hs, mode);
4797 hrtimer_cancel(&hs.timer);
4798 mode = HRTIMER_MODE_ABS;
4799 } while (hs.task && !signal_pending(current));
4801 __set_current_state(TASK_RUNNING);
4802 destroy_hrtimer_on_stack(&hs.timer);
4805 * If we sleep, have the caller restart the poll loop to reset the
4806 * state. Like for the other success return cases, the caller is
4807 * responsible for checking if the IO completed. If the IO isn't
4808 * complete, we'll get called again and will go straight to the busy
4814 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4815 struct io_comp_batch *iob, unsigned int flags)
4817 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4818 long state = get_current_state();
4822 ret = q->mq_ops->poll(hctx, iob);
4824 __set_current_state(TASK_RUNNING);
4828 if (signal_pending_state(state, current))
4829 __set_current_state(TASK_RUNNING);
4830 if (task_is_running(current))
4833 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4836 } while (!need_resched());
4838 __set_current_state(TASK_RUNNING);
4842 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4845 if (!(flags & BLK_POLL_NOSLEEP) &&
4846 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4847 if (blk_mq_poll_hybrid(q, cookie))
4850 return blk_mq_poll_classic(q, cookie, iob, flags);
4853 unsigned int blk_mq_rq_cpu(struct request *rq)
4855 return rq->mq_ctx->cpu;
4857 EXPORT_SYMBOL(blk_mq_rq_cpu);
4859 void blk_mq_cancel_work_sync(struct request_queue *q)
4861 if (queue_is_mq(q)) {
4862 struct blk_mq_hw_ctx *hctx;
4865 cancel_delayed_work_sync(&q->requeue_work);
4867 queue_for_each_hw_ctx(q, hctx, i)
4868 cancel_delayed_work_sync(&hctx->run_work);
4872 static int __init blk_mq_init(void)
4876 for_each_possible_cpu(i)
4877 init_llist_head(&per_cpu(blk_cpu_done, i));
4878 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4880 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4881 "block/softirq:dead", NULL,
4882 blk_softirq_cpu_dead);
4883 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4884 blk_mq_hctx_notify_dead);
4885 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4886 blk_mq_hctx_notify_online,
4887 blk_mq_hctx_notify_offline);
4890 subsys_initcall(blk_mq_init);