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
831 * blk_update_request - Complete multiple bytes without completing the request
832 * @req: the request being processed
833 * @error: block status code
834 * @nr_bytes: number of bytes to complete for @req
837 * Ends I/O on a number of bytes attached to @req, but doesn't complete
838 * the request structure even if @req doesn't have leftover.
839 * If @req has leftover, sets it up for the next range of segments.
841 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
842 * %false return from this function.
845 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
846 * except in the consistency check at the end of this function.
849 * %false - this request doesn't have any more data
850 * %true - this request has more data
852 bool blk_update_request(struct request *req, blk_status_t error,
853 unsigned int nr_bytes)
857 trace_block_rq_complete(req, error, nr_bytes);
862 #ifdef CONFIG_BLK_DEV_INTEGRITY
863 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
865 req->q->integrity.profile->complete_fn(req, nr_bytes);
868 if (unlikely(error && !blk_rq_is_passthrough(req) &&
869 !(req->rq_flags & RQF_QUIET)) &&
870 !test_bit(GD_DEAD, &req->q->disk->state)) {
871 blk_print_req_error(req, error);
872 trace_block_rq_error(req, error, nr_bytes);
875 blk_account_io_completion(req, nr_bytes);
879 struct bio *bio = req->bio;
880 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
882 if (bio_bytes == bio->bi_iter.bi_size)
883 req->bio = bio->bi_next;
885 /* Completion has already been traced */
886 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
887 req_bio_endio(req, bio, bio_bytes, error);
889 total_bytes += bio_bytes;
890 nr_bytes -= bio_bytes;
901 * Reset counters so that the request stacking driver
902 * can find how many bytes remain in the request
909 req->__data_len -= total_bytes;
911 /* update sector only for requests with clear definition of sector */
912 if (!blk_rq_is_passthrough(req))
913 req->__sector += total_bytes >> 9;
915 /* mixed attributes always follow the first bio */
916 if (req->rq_flags & RQF_MIXED_MERGE) {
917 req->cmd_flags &= ~REQ_FAILFAST_MASK;
918 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
921 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
923 * If total number of sectors is less than the first segment
924 * size, something has gone terribly wrong.
926 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
927 blk_dump_rq_flags(req, "request botched");
928 req->__data_len = blk_rq_cur_bytes(req);
931 /* recalculate the number of segments */
932 req->nr_phys_segments = blk_recalc_rq_segments(req);
937 EXPORT_SYMBOL_GPL(blk_update_request);
939 static void __blk_account_io_done(struct request *req, u64 now)
941 const int sgrp = op_stat_group(req_op(req));
944 update_io_ticks(req->part, jiffies, true);
945 part_stat_inc(req->part, ios[sgrp]);
946 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
950 static inline void blk_account_io_done(struct request *req, u64 now)
953 * Account IO completion. flush_rq isn't accounted as a
954 * normal IO on queueing nor completion. Accounting the
955 * containing request is enough.
957 if (blk_do_io_stat(req) && req->part &&
958 !(req->rq_flags & RQF_FLUSH_SEQ))
959 __blk_account_io_done(req, now);
962 static void __blk_account_io_start(struct request *rq)
965 * All non-passthrough requests are created from a bio with one
966 * exception: when a flush command that is part of a flush sequence
967 * generated by the state machine in blk-flush.c is cloned onto the
968 * lower device by dm-multipath we can get here without a bio.
971 rq->part = rq->bio->bi_bdev;
973 rq->part = rq->q->disk->part0;
976 update_io_ticks(rq->part, jiffies, false);
980 static inline void blk_account_io_start(struct request *req)
982 if (blk_do_io_stat(req))
983 __blk_account_io_start(req);
986 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
988 if (rq->rq_flags & RQF_STATS) {
989 blk_mq_poll_stats_start(rq->q);
990 blk_stat_add(rq, now);
993 blk_mq_sched_completed_request(rq, now);
994 blk_account_io_done(rq, now);
997 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
999 if (blk_mq_need_time_stamp(rq))
1000 __blk_mq_end_request_acct(rq, ktime_get_ns());
1003 rq_qos_done(rq->q, rq);
1004 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1005 blk_mq_free_request(rq);
1007 blk_mq_free_request(rq);
1010 EXPORT_SYMBOL(__blk_mq_end_request);
1012 void blk_mq_end_request(struct request *rq, blk_status_t error)
1014 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1016 __blk_mq_end_request(rq, error);
1018 EXPORT_SYMBOL(blk_mq_end_request);
1020 #define TAG_COMP_BATCH 32
1022 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1023 int *tag_array, int nr_tags)
1025 struct request_queue *q = hctx->queue;
1028 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1029 * update hctx->nr_active in batch
1031 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1032 __blk_mq_sub_active_requests(hctx, nr_tags);
1034 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1035 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1038 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1040 int tags[TAG_COMP_BATCH], nr_tags = 0;
1041 struct blk_mq_hw_ctx *cur_hctx = NULL;
1046 now = ktime_get_ns();
1048 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1050 prefetch(rq->rq_next);
1052 blk_complete_request(rq);
1054 __blk_mq_end_request_acct(rq, now);
1056 rq_qos_done(rq->q, rq);
1058 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1059 if (!req_ref_put_and_test(rq))
1062 blk_crypto_free_request(rq);
1063 blk_pm_mark_last_busy(rq);
1065 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1067 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1069 cur_hctx = rq->mq_hctx;
1071 tags[nr_tags++] = rq->tag;
1075 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1077 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1079 static void blk_complete_reqs(struct llist_head *list)
1081 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1082 struct request *rq, *next;
1084 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1085 rq->q->mq_ops->complete(rq);
1088 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1090 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1093 static int blk_softirq_cpu_dead(unsigned int cpu)
1095 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1099 static void __blk_mq_complete_request_remote(void *data)
1101 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1104 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1106 int cpu = raw_smp_processor_id();
1108 if (!IS_ENABLED(CONFIG_SMP) ||
1109 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1112 * With force threaded interrupts enabled, raising softirq from an SMP
1113 * function call will always result in waking the ksoftirqd thread.
1114 * This is probably worse than completing the request on a different
1117 if (force_irqthreads())
1120 /* same CPU or cache domain? Complete locally */
1121 if (cpu == rq->mq_ctx->cpu ||
1122 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1123 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1126 /* don't try to IPI to an offline CPU */
1127 return cpu_online(rq->mq_ctx->cpu);
1130 static void blk_mq_complete_send_ipi(struct request *rq)
1132 struct llist_head *list;
1135 cpu = rq->mq_ctx->cpu;
1136 list = &per_cpu(blk_cpu_done, cpu);
1137 if (llist_add(&rq->ipi_list, list)) {
1138 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1139 smp_call_function_single_async(cpu, &rq->csd);
1143 static void blk_mq_raise_softirq(struct request *rq)
1145 struct llist_head *list;
1148 list = this_cpu_ptr(&blk_cpu_done);
1149 if (llist_add(&rq->ipi_list, list))
1150 raise_softirq(BLOCK_SOFTIRQ);
1154 bool blk_mq_complete_request_remote(struct request *rq)
1156 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1159 * For request which hctx has only one ctx mapping,
1160 * or a polled request, always complete locally,
1161 * it's pointless to redirect the completion.
1163 if (rq->mq_hctx->nr_ctx == 1 ||
1164 rq->cmd_flags & REQ_POLLED)
1167 if (blk_mq_complete_need_ipi(rq)) {
1168 blk_mq_complete_send_ipi(rq);
1172 if (rq->q->nr_hw_queues == 1) {
1173 blk_mq_raise_softirq(rq);
1178 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1181 * blk_mq_complete_request - end I/O on a request
1182 * @rq: the request being processed
1185 * Complete a request by scheduling the ->complete_rq operation.
1187 void blk_mq_complete_request(struct request *rq)
1189 if (!blk_mq_complete_request_remote(rq))
1190 rq->q->mq_ops->complete(rq);
1192 EXPORT_SYMBOL(blk_mq_complete_request);
1195 * blk_mq_start_request - Start processing a request
1196 * @rq: Pointer to request to be started
1198 * Function used by device drivers to notify the block layer that a request
1199 * is going to be processed now, so blk layer can do proper initializations
1200 * such as starting the timeout timer.
1202 void blk_mq_start_request(struct request *rq)
1204 struct request_queue *q = rq->q;
1206 trace_block_rq_issue(rq);
1208 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1209 rq->io_start_time_ns = ktime_get_ns();
1210 rq->stats_sectors = blk_rq_sectors(rq);
1211 rq->rq_flags |= RQF_STATS;
1212 rq_qos_issue(q, rq);
1215 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1218 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1220 #ifdef CONFIG_BLK_DEV_INTEGRITY
1221 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1222 q->integrity.profile->prepare_fn(rq);
1224 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1225 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1227 EXPORT_SYMBOL(blk_mq_start_request);
1230 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1231 * queues. This is important for md arrays to benefit from merging
1234 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1236 if (plug->multiple_queues)
1237 return BLK_MAX_REQUEST_COUNT * 2;
1238 return BLK_MAX_REQUEST_COUNT;
1241 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1243 struct request *last = rq_list_peek(&plug->mq_list);
1245 if (!plug->rq_count) {
1246 trace_block_plug(rq->q);
1247 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1248 (!blk_queue_nomerges(rq->q) &&
1249 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1250 blk_mq_flush_plug_list(plug, false);
1251 trace_block_plug(rq->q);
1254 if (!plug->multiple_queues && last && last->q != rq->q)
1255 plug->multiple_queues = true;
1256 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1257 plug->has_elevator = true;
1259 rq_list_add(&plug->mq_list, rq);
1264 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1265 * @rq: request to insert
1266 * @at_head: insert request at head or tail of queue
1269 * Insert a fully prepared request at the back of the I/O scheduler queue
1270 * for execution. Don't wait for completion.
1273 * This function will invoke @done directly if the queue is dead.
1275 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1277 WARN_ON(irqs_disabled());
1278 WARN_ON(!blk_rq_is_passthrough(rq));
1280 blk_account_io_start(rq);
1283 * As plugging can be enabled for passthrough requests on a zoned
1284 * device, directly accessing the plug instead of using blk_mq_plug()
1285 * should not have any consequences.
1288 blk_add_rq_to_plug(current->plug, rq);
1290 blk_mq_sched_insert_request(rq, at_head, true, false);
1292 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1294 struct blk_rq_wait {
1295 struct completion done;
1299 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1301 struct blk_rq_wait *wait = rq->end_io_data;
1304 complete(&wait->done);
1305 return RQ_END_IO_NONE;
1308 bool blk_rq_is_poll(struct request *rq)
1312 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1314 if (WARN_ON_ONCE(!rq->bio))
1318 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1320 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1323 bio_poll(rq->bio, NULL, 0);
1325 } while (!completion_done(wait));
1329 * blk_execute_rq - insert a request into queue for execution
1330 * @rq: request to insert
1331 * @at_head: insert request at head or tail of queue
1334 * Insert a fully prepared request at the back of the I/O scheduler queue
1335 * for execution and wait for completion.
1336 * Return: The blk_status_t result provided to blk_mq_end_request().
1338 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1340 struct blk_rq_wait wait = {
1341 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1344 WARN_ON(irqs_disabled());
1345 WARN_ON(!blk_rq_is_passthrough(rq));
1347 rq->end_io_data = &wait;
1348 rq->end_io = blk_end_sync_rq;
1350 blk_account_io_start(rq);
1351 blk_mq_sched_insert_request(rq, at_head, true, false);
1353 if (blk_rq_is_poll(rq)) {
1354 blk_rq_poll_completion(rq, &wait.done);
1357 * Prevent hang_check timer from firing at us during very long
1360 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1363 while (!wait_for_completion_io_timeout(&wait.done,
1364 hang_check * (HZ/2)))
1367 wait_for_completion_io(&wait.done);
1372 EXPORT_SYMBOL(blk_execute_rq);
1374 static void __blk_mq_requeue_request(struct request *rq)
1376 struct request_queue *q = rq->q;
1378 blk_mq_put_driver_tag(rq);
1380 trace_block_rq_requeue(rq);
1381 rq_qos_requeue(q, rq);
1383 if (blk_mq_request_started(rq)) {
1384 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1385 rq->rq_flags &= ~RQF_TIMED_OUT;
1389 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1391 __blk_mq_requeue_request(rq);
1393 /* this request will be re-inserted to io scheduler queue */
1394 blk_mq_sched_requeue_request(rq);
1396 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1398 EXPORT_SYMBOL(blk_mq_requeue_request);
1400 static void blk_mq_requeue_work(struct work_struct *work)
1402 struct request_queue *q =
1403 container_of(work, struct request_queue, requeue_work.work);
1405 struct request *rq, *next;
1407 spin_lock_irq(&q->requeue_lock);
1408 list_splice_init(&q->requeue_list, &rq_list);
1409 spin_unlock_irq(&q->requeue_lock);
1411 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1412 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1415 rq->rq_flags &= ~RQF_SOFTBARRIER;
1416 list_del_init(&rq->queuelist);
1418 * If RQF_DONTPREP, rq has contained some driver specific
1419 * data, so insert it to hctx dispatch list to avoid any
1422 if (rq->rq_flags & RQF_DONTPREP)
1423 blk_mq_request_bypass_insert(rq, false, false);
1425 blk_mq_sched_insert_request(rq, true, false, false);
1428 while (!list_empty(&rq_list)) {
1429 rq = list_entry(rq_list.next, struct request, queuelist);
1430 list_del_init(&rq->queuelist);
1431 blk_mq_sched_insert_request(rq, false, false, false);
1434 blk_mq_run_hw_queues(q, false);
1437 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1438 bool kick_requeue_list)
1440 struct request_queue *q = rq->q;
1441 unsigned long flags;
1444 * We abuse this flag that is otherwise used by the I/O scheduler to
1445 * request head insertion from the workqueue.
1447 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1449 spin_lock_irqsave(&q->requeue_lock, flags);
1451 rq->rq_flags |= RQF_SOFTBARRIER;
1452 list_add(&rq->queuelist, &q->requeue_list);
1454 list_add_tail(&rq->queuelist, &q->requeue_list);
1456 spin_unlock_irqrestore(&q->requeue_lock, flags);
1458 if (kick_requeue_list)
1459 blk_mq_kick_requeue_list(q);
1462 void blk_mq_kick_requeue_list(struct request_queue *q)
1464 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1466 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1468 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1469 unsigned long msecs)
1471 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1472 msecs_to_jiffies(msecs));
1474 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1476 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1479 * If we find a request that isn't idle we know the queue is busy
1480 * as it's checked in the iter.
1481 * Return false to stop the iteration.
1483 if (blk_mq_request_started(rq)) {
1493 bool blk_mq_queue_inflight(struct request_queue *q)
1497 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1500 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1502 static void blk_mq_rq_timed_out(struct request *req)
1504 req->rq_flags |= RQF_TIMED_OUT;
1505 if (req->q->mq_ops->timeout) {
1506 enum blk_eh_timer_return ret;
1508 ret = req->q->mq_ops->timeout(req);
1509 if (ret == BLK_EH_DONE)
1511 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1517 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
1519 unsigned long deadline;
1521 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1523 if (rq->rq_flags & RQF_TIMED_OUT)
1526 deadline = READ_ONCE(rq->deadline);
1527 if (time_after_eq(jiffies, deadline))
1532 else if (time_after(*next, deadline))
1537 void blk_mq_put_rq_ref(struct request *rq)
1539 if (is_flush_rq(rq)) {
1540 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1541 blk_mq_free_request(rq);
1542 } else if (req_ref_put_and_test(rq)) {
1543 __blk_mq_free_request(rq);
1547 static bool blk_mq_check_expired(struct request *rq, void *priv)
1549 unsigned long *next = priv;
1552 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1553 * be reallocated underneath the timeout handler's processing, then
1554 * the expire check is reliable. If the request is not expired, then
1555 * it was completed and reallocated as a new request after returning
1556 * from blk_mq_check_expired().
1558 if (blk_mq_req_expired(rq, next))
1559 blk_mq_rq_timed_out(rq);
1563 static void blk_mq_timeout_work(struct work_struct *work)
1565 struct request_queue *q =
1566 container_of(work, struct request_queue, timeout_work);
1567 unsigned long next = 0;
1568 struct blk_mq_hw_ctx *hctx;
1571 /* A deadlock might occur if a request is stuck requiring a
1572 * timeout at the same time a queue freeze is waiting
1573 * completion, since the timeout code would not be able to
1574 * acquire the queue reference here.
1576 * That's why we don't use blk_queue_enter here; instead, we use
1577 * percpu_ref_tryget directly, because we need to be able to
1578 * obtain a reference even in the short window between the queue
1579 * starting to freeze, by dropping the first reference in
1580 * blk_freeze_queue_start, and the moment the last request is
1581 * consumed, marked by the instant q_usage_counter reaches
1584 if (!percpu_ref_tryget(&q->q_usage_counter))
1587 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
1590 mod_timer(&q->timeout, next);
1593 * Request timeouts are handled as a forward rolling timer. If
1594 * we end up here it means that no requests are pending and
1595 * also that no request has been pending for a while. Mark
1596 * each hctx as idle.
1598 queue_for_each_hw_ctx(q, hctx, i) {
1599 /* the hctx may be unmapped, so check it here */
1600 if (blk_mq_hw_queue_mapped(hctx))
1601 blk_mq_tag_idle(hctx);
1607 struct flush_busy_ctx_data {
1608 struct blk_mq_hw_ctx *hctx;
1609 struct list_head *list;
1612 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1614 struct flush_busy_ctx_data *flush_data = data;
1615 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1616 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1617 enum hctx_type type = hctx->type;
1619 spin_lock(&ctx->lock);
1620 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1621 sbitmap_clear_bit(sb, bitnr);
1622 spin_unlock(&ctx->lock);
1627 * Process software queues that have been marked busy, splicing them
1628 * to the for-dispatch
1630 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1632 struct flush_busy_ctx_data data = {
1637 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1639 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1641 struct dispatch_rq_data {
1642 struct blk_mq_hw_ctx *hctx;
1646 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1649 struct dispatch_rq_data *dispatch_data = data;
1650 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1651 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1652 enum hctx_type type = hctx->type;
1654 spin_lock(&ctx->lock);
1655 if (!list_empty(&ctx->rq_lists[type])) {
1656 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1657 list_del_init(&dispatch_data->rq->queuelist);
1658 if (list_empty(&ctx->rq_lists[type]))
1659 sbitmap_clear_bit(sb, bitnr);
1661 spin_unlock(&ctx->lock);
1663 return !dispatch_data->rq;
1666 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1667 struct blk_mq_ctx *start)
1669 unsigned off = start ? start->index_hw[hctx->type] : 0;
1670 struct dispatch_rq_data data = {
1675 __sbitmap_for_each_set(&hctx->ctx_map, off,
1676 dispatch_rq_from_ctx, &data);
1681 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1683 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1684 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1687 blk_mq_tag_busy(rq->mq_hctx);
1689 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1690 bt = &rq->mq_hctx->tags->breserved_tags;
1693 if (!hctx_may_queue(rq->mq_hctx, bt))
1697 tag = __sbitmap_queue_get(bt);
1698 if (tag == BLK_MQ_NO_TAG)
1701 rq->tag = tag + tag_offset;
1705 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1707 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1710 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1711 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1712 rq->rq_flags |= RQF_MQ_INFLIGHT;
1713 __blk_mq_inc_active_requests(hctx);
1715 hctx->tags->rqs[rq->tag] = rq;
1719 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1720 int flags, void *key)
1722 struct blk_mq_hw_ctx *hctx;
1724 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1726 spin_lock(&hctx->dispatch_wait_lock);
1727 if (!list_empty(&wait->entry)) {
1728 struct sbitmap_queue *sbq;
1730 list_del_init(&wait->entry);
1731 sbq = &hctx->tags->bitmap_tags;
1732 atomic_dec(&sbq->ws_active);
1734 spin_unlock(&hctx->dispatch_wait_lock);
1736 blk_mq_run_hw_queue(hctx, true);
1741 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1742 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1743 * restart. For both cases, take care to check the condition again after
1744 * marking us as waiting.
1746 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1749 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1750 struct wait_queue_head *wq;
1751 wait_queue_entry_t *wait;
1754 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1755 blk_mq_sched_mark_restart_hctx(hctx);
1758 * It's possible that a tag was freed in the window between the
1759 * allocation failure and adding the hardware queue to the wait
1762 * Don't clear RESTART here, someone else could have set it.
1763 * At most this will cost an extra queue run.
1765 return blk_mq_get_driver_tag(rq);
1768 wait = &hctx->dispatch_wait;
1769 if (!list_empty_careful(&wait->entry))
1772 wq = &bt_wait_ptr(sbq, hctx)->wait;
1774 spin_lock_irq(&wq->lock);
1775 spin_lock(&hctx->dispatch_wait_lock);
1776 if (!list_empty(&wait->entry)) {
1777 spin_unlock(&hctx->dispatch_wait_lock);
1778 spin_unlock_irq(&wq->lock);
1782 atomic_inc(&sbq->ws_active);
1783 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1784 __add_wait_queue(wq, wait);
1787 * It's possible that a tag was freed in the window between the
1788 * allocation failure and adding the hardware queue to the wait
1791 ret = blk_mq_get_driver_tag(rq);
1793 spin_unlock(&hctx->dispatch_wait_lock);
1794 spin_unlock_irq(&wq->lock);
1799 * We got a tag, remove ourselves from the wait queue to ensure
1800 * someone else gets the wakeup.
1802 list_del_init(&wait->entry);
1803 atomic_dec(&sbq->ws_active);
1804 spin_unlock(&hctx->dispatch_wait_lock);
1805 spin_unlock_irq(&wq->lock);
1810 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1811 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1813 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1814 * - EWMA is one simple way to compute running average value
1815 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1816 * - take 4 as factor for avoiding to get too small(0) result, and this
1817 * factor doesn't matter because EWMA decreases exponentially
1819 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1823 ewma = hctx->dispatch_busy;
1828 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1830 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1831 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1833 hctx->dispatch_busy = ewma;
1836 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1838 static void blk_mq_handle_dev_resource(struct request *rq,
1839 struct list_head *list)
1841 struct request *next =
1842 list_first_entry_or_null(list, struct request, queuelist);
1845 * If an I/O scheduler has been configured and we got a driver tag for
1846 * the next request already, free it.
1849 blk_mq_put_driver_tag(next);
1851 list_add(&rq->queuelist, list);
1852 __blk_mq_requeue_request(rq);
1855 static void blk_mq_handle_zone_resource(struct request *rq,
1856 struct list_head *zone_list)
1859 * If we end up here it is because we cannot dispatch a request to a
1860 * specific zone due to LLD level zone-write locking or other zone
1861 * related resource not being available. In this case, set the request
1862 * aside in zone_list for retrying it later.
1864 list_add(&rq->queuelist, zone_list);
1865 __blk_mq_requeue_request(rq);
1868 enum prep_dispatch {
1870 PREP_DISPATCH_NO_TAG,
1871 PREP_DISPATCH_NO_BUDGET,
1874 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1877 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1878 int budget_token = -1;
1881 budget_token = blk_mq_get_dispatch_budget(rq->q);
1882 if (budget_token < 0) {
1883 blk_mq_put_driver_tag(rq);
1884 return PREP_DISPATCH_NO_BUDGET;
1886 blk_mq_set_rq_budget_token(rq, budget_token);
1889 if (!blk_mq_get_driver_tag(rq)) {
1891 * The initial allocation attempt failed, so we need to
1892 * rerun the hardware queue when a tag is freed. The
1893 * waitqueue takes care of that. If the queue is run
1894 * before we add this entry back on the dispatch list,
1895 * we'll re-run it below.
1897 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1899 * All budgets not got from this function will be put
1900 * together during handling partial dispatch
1903 blk_mq_put_dispatch_budget(rq->q, budget_token);
1904 return PREP_DISPATCH_NO_TAG;
1908 return PREP_DISPATCH_OK;
1911 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1912 static void blk_mq_release_budgets(struct request_queue *q,
1913 struct list_head *list)
1917 list_for_each_entry(rq, list, queuelist) {
1918 int budget_token = blk_mq_get_rq_budget_token(rq);
1920 if (budget_token >= 0)
1921 blk_mq_put_dispatch_budget(q, budget_token);
1926 * Returns true if we did some work AND can potentially do more.
1928 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1929 unsigned int nr_budgets)
1931 enum prep_dispatch prep;
1932 struct request_queue *q = hctx->queue;
1933 struct request *rq, *nxt;
1935 blk_status_t ret = BLK_STS_OK;
1936 LIST_HEAD(zone_list);
1937 bool needs_resource = false;
1939 if (list_empty(list))
1943 * Now process all the entries, sending them to the driver.
1945 errors = queued = 0;
1947 struct blk_mq_queue_data bd;
1949 rq = list_first_entry(list, struct request, queuelist);
1951 WARN_ON_ONCE(hctx != rq->mq_hctx);
1952 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1953 if (prep != PREP_DISPATCH_OK)
1956 list_del_init(&rq->queuelist);
1961 * Flag last if we have no more requests, or if we have more
1962 * but can't assign a driver tag to it.
1964 if (list_empty(list))
1967 nxt = list_first_entry(list, struct request, queuelist);
1968 bd.last = !blk_mq_get_driver_tag(nxt);
1972 * once the request is queued to lld, no need to cover the
1977 ret = q->mq_ops->queue_rq(hctx, &bd);
1982 case BLK_STS_RESOURCE:
1983 needs_resource = true;
1985 case BLK_STS_DEV_RESOURCE:
1986 blk_mq_handle_dev_resource(rq, list);
1988 case BLK_STS_ZONE_RESOURCE:
1990 * Move the request to zone_list and keep going through
1991 * the dispatch list to find more requests the drive can
1994 blk_mq_handle_zone_resource(rq, &zone_list);
1995 needs_resource = true;
1999 blk_mq_end_request(rq, ret);
2001 } while (!list_empty(list));
2003 if (!list_empty(&zone_list))
2004 list_splice_tail_init(&zone_list, list);
2006 /* If we didn't flush the entire list, we could have told the driver
2007 * there was more coming, but that turned out to be a lie.
2009 if ((!list_empty(list) || errors || needs_resource ||
2010 ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued)
2011 q->mq_ops->commit_rqs(hctx);
2013 * Any items that need requeuing? Stuff them into hctx->dispatch,
2014 * that is where we will continue on next queue run.
2016 if (!list_empty(list)) {
2018 /* For non-shared tags, the RESTART check will suffice */
2019 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2020 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
2023 blk_mq_release_budgets(q, list);
2025 spin_lock(&hctx->lock);
2026 list_splice_tail_init(list, &hctx->dispatch);
2027 spin_unlock(&hctx->lock);
2030 * Order adding requests to hctx->dispatch and checking
2031 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2032 * in blk_mq_sched_restart(). Avoid restart code path to
2033 * miss the new added requests to hctx->dispatch, meantime
2034 * SCHED_RESTART is observed here.
2039 * If SCHED_RESTART was set by the caller of this function and
2040 * it is no longer set that means that it was cleared by another
2041 * thread and hence that a queue rerun is needed.
2043 * If 'no_tag' is set, that means that we failed getting
2044 * a driver tag with an I/O scheduler attached. If our dispatch
2045 * waitqueue is no longer active, ensure that we run the queue
2046 * AFTER adding our entries back to the list.
2048 * If no I/O scheduler has been configured it is possible that
2049 * the hardware queue got stopped and restarted before requests
2050 * were pushed back onto the dispatch list. Rerun the queue to
2051 * avoid starvation. Notes:
2052 * - blk_mq_run_hw_queue() checks whether or not a queue has
2053 * been stopped before rerunning a queue.
2054 * - Some but not all block drivers stop a queue before
2055 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2058 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2059 * bit is set, run queue after a delay to avoid IO stalls
2060 * that could otherwise occur if the queue is idle. We'll do
2061 * similar if we couldn't get budget or couldn't lock a zone
2062 * and SCHED_RESTART is set.
2064 needs_restart = blk_mq_sched_needs_restart(hctx);
2065 if (prep == PREP_DISPATCH_NO_BUDGET)
2066 needs_resource = true;
2067 if (!needs_restart ||
2068 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2069 blk_mq_run_hw_queue(hctx, true);
2070 else if (needs_resource)
2071 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2073 blk_mq_update_dispatch_busy(hctx, true);
2076 blk_mq_update_dispatch_busy(hctx, false);
2078 return (queued + errors) != 0;
2082 * __blk_mq_run_hw_queue - Run a hardware queue.
2083 * @hctx: Pointer to the hardware queue to run.
2085 * Send pending requests to the hardware.
2087 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
2090 * We can't run the queue inline with ints disabled. Ensure that
2091 * we catch bad users of this early.
2093 WARN_ON_ONCE(in_interrupt());
2095 blk_mq_run_dispatch_ops(hctx->queue,
2096 blk_mq_sched_dispatch_requests(hctx));
2099 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2101 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2103 if (cpu >= nr_cpu_ids)
2104 cpu = cpumask_first(hctx->cpumask);
2109 * It'd be great if the workqueue API had a way to pass
2110 * in a mask and had some smarts for more clever placement.
2111 * For now we just round-robin here, switching for every
2112 * BLK_MQ_CPU_WORK_BATCH queued items.
2114 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2117 int next_cpu = hctx->next_cpu;
2119 if (hctx->queue->nr_hw_queues == 1)
2120 return WORK_CPU_UNBOUND;
2122 if (--hctx->next_cpu_batch <= 0) {
2124 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2126 if (next_cpu >= nr_cpu_ids)
2127 next_cpu = blk_mq_first_mapped_cpu(hctx);
2128 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2132 * Do unbound schedule if we can't find a online CPU for this hctx,
2133 * and it should only happen in the path of handling CPU DEAD.
2135 if (!cpu_online(next_cpu)) {
2142 * Make sure to re-select CPU next time once after CPUs
2143 * in hctx->cpumask become online again.
2145 hctx->next_cpu = next_cpu;
2146 hctx->next_cpu_batch = 1;
2147 return WORK_CPU_UNBOUND;
2150 hctx->next_cpu = next_cpu;
2155 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2156 * @hctx: Pointer to the hardware queue to run.
2157 * @async: If we want to run the queue asynchronously.
2158 * @msecs: Milliseconds of delay to wait before running the queue.
2160 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2161 * with a delay of @msecs.
2163 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2164 unsigned long msecs)
2166 if (unlikely(blk_mq_hctx_stopped(hctx)))
2169 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2170 if (cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2171 __blk_mq_run_hw_queue(hctx);
2176 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2177 msecs_to_jiffies(msecs));
2181 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2182 * @hctx: Pointer to the hardware queue to run.
2183 * @msecs: Milliseconds of delay to wait before running the queue.
2185 * Run a hardware queue asynchronously with a delay of @msecs.
2187 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2189 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2191 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2194 * blk_mq_run_hw_queue - Start to run a hardware queue.
2195 * @hctx: Pointer to the hardware queue to run.
2196 * @async: If we want to run the queue asynchronously.
2198 * Check if the request queue is not in a quiesced state and if there are
2199 * pending requests to be sent. If this is true, run the queue to send requests
2202 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2207 * When queue is quiesced, we may be switching io scheduler, or
2208 * updating nr_hw_queues, or other things, and we can't run queue
2209 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2211 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2214 __blk_mq_run_dispatch_ops(hctx->queue, false,
2215 need_run = !blk_queue_quiesced(hctx->queue) &&
2216 blk_mq_hctx_has_pending(hctx));
2219 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2221 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2224 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2227 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2229 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2231 * If the IO scheduler does not respect hardware queues when
2232 * dispatching, we just don't bother with multiple HW queues and
2233 * dispatch from hctx for the current CPU since running multiple queues
2234 * just causes lock contention inside the scheduler and pointless cache
2237 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2239 if (!blk_mq_hctx_stopped(hctx))
2245 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2246 * @q: Pointer to the request queue to run.
2247 * @async: If we want to run the queue asynchronously.
2249 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2251 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2255 if (blk_queue_sq_sched(q))
2256 sq_hctx = blk_mq_get_sq_hctx(q);
2257 queue_for_each_hw_ctx(q, hctx, i) {
2258 if (blk_mq_hctx_stopped(hctx))
2261 * Dispatch from this hctx either if there's no hctx preferred
2262 * by IO scheduler or if it has requests that bypass the
2265 if (!sq_hctx || sq_hctx == hctx ||
2266 !list_empty_careful(&hctx->dispatch))
2267 blk_mq_run_hw_queue(hctx, async);
2270 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2273 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2274 * @q: Pointer to the request queue to run.
2275 * @msecs: Milliseconds of delay to wait before running the queues.
2277 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2279 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2283 if (blk_queue_sq_sched(q))
2284 sq_hctx = blk_mq_get_sq_hctx(q);
2285 queue_for_each_hw_ctx(q, hctx, i) {
2286 if (blk_mq_hctx_stopped(hctx))
2289 * If there is already a run_work pending, leave the
2290 * pending delay untouched. Otherwise, a hctx can stall
2291 * if another hctx is re-delaying the other's work
2292 * before the work executes.
2294 if (delayed_work_pending(&hctx->run_work))
2297 * Dispatch from this hctx either if there's no hctx preferred
2298 * by IO scheduler or if it has requests that bypass the
2301 if (!sq_hctx || sq_hctx == hctx ||
2302 !list_empty_careful(&hctx->dispatch))
2303 blk_mq_delay_run_hw_queue(hctx, msecs);
2306 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2309 * This function is often used for pausing .queue_rq() by driver when
2310 * there isn't enough resource or some conditions aren't satisfied, and
2311 * BLK_STS_RESOURCE is usually returned.
2313 * We do not guarantee that dispatch can be drained or blocked
2314 * after blk_mq_stop_hw_queue() returns. Please use
2315 * blk_mq_quiesce_queue() for that requirement.
2317 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2319 cancel_delayed_work(&hctx->run_work);
2321 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2323 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2326 * This function is often used for pausing .queue_rq() by driver when
2327 * there isn't enough resource or some conditions aren't satisfied, and
2328 * BLK_STS_RESOURCE is usually returned.
2330 * We do not guarantee that dispatch can be drained or blocked
2331 * after blk_mq_stop_hw_queues() returns. Please use
2332 * blk_mq_quiesce_queue() for that requirement.
2334 void blk_mq_stop_hw_queues(struct request_queue *q)
2336 struct blk_mq_hw_ctx *hctx;
2339 queue_for_each_hw_ctx(q, hctx, i)
2340 blk_mq_stop_hw_queue(hctx);
2342 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2344 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2346 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2348 blk_mq_run_hw_queue(hctx, false);
2350 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2352 void blk_mq_start_hw_queues(struct request_queue *q)
2354 struct blk_mq_hw_ctx *hctx;
2357 queue_for_each_hw_ctx(q, hctx, i)
2358 blk_mq_start_hw_queue(hctx);
2360 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2362 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2364 if (!blk_mq_hctx_stopped(hctx))
2367 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2368 blk_mq_run_hw_queue(hctx, async);
2370 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2372 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2374 struct blk_mq_hw_ctx *hctx;
2377 queue_for_each_hw_ctx(q, hctx, i)
2378 blk_mq_start_stopped_hw_queue(hctx, async);
2380 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2382 static void blk_mq_run_work_fn(struct work_struct *work)
2384 struct blk_mq_hw_ctx *hctx;
2386 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2389 * If we are stopped, don't run the queue.
2391 if (blk_mq_hctx_stopped(hctx))
2394 __blk_mq_run_hw_queue(hctx);
2397 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2401 struct blk_mq_ctx *ctx = rq->mq_ctx;
2402 enum hctx_type type = hctx->type;
2404 lockdep_assert_held(&ctx->lock);
2406 trace_block_rq_insert(rq);
2409 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2411 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2414 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2417 struct blk_mq_ctx *ctx = rq->mq_ctx;
2419 lockdep_assert_held(&ctx->lock);
2421 __blk_mq_insert_req_list(hctx, rq, at_head);
2422 blk_mq_hctx_mark_pending(hctx, ctx);
2426 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2427 * @rq: Pointer to request to be inserted.
2428 * @at_head: true if the request should be inserted at the head of the list.
2429 * @run_queue: If we should run the hardware queue after inserting the request.
2431 * Should only be used carefully, when the caller knows we want to
2432 * bypass a potential IO scheduler on the target device.
2434 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2437 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2439 spin_lock(&hctx->lock);
2441 list_add(&rq->queuelist, &hctx->dispatch);
2443 list_add_tail(&rq->queuelist, &hctx->dispatch);
2444 spin_unlock(&hctx->lock);
2447 blk_mq_run_hw_queue(hctx, false);
2450 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2451 struct list_head *list)
2455 enum hctx_type type = hctx->type;
2458 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2461 list_for_each_entry(rq, list, queuelist) {
2462 BUG_ON(rq->mq_ctx != ctx);
2463 trace_block_rq_insert(rq);
2466 spin_lock(&ctx->lock);
2467 list_splice_tail_init(list, &ctx->rq_lists[type]);
2468 blk_mq_hctx_mark_pending(hctx, ctx);
2469 spin_unlock(&ctx->lock);
2472 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2475 if (hctx->queue->mq_ops->commit_rqs) {
2476 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2477 hctx->queue->mq_ops->commit_rqs(hctx);
2482 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2483 unsigned int nr_segs)
2487 if (bio->bi_opf & REQ_RAHEAD)
2488 rq->cmd_flags |= REQ_FAILFAST_MASK;
2490 rq->__sector = bio->bi_iter.bi_sector;
2491 blk_rq_bio_prep(rq, bio, nr_segs);
2493 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2494 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2497 blk_account_io_start(rq);
2500 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2501 struct request *rq, bool last)
2503 struct request_queue *q = rq->q;
2504 struct blk_mq_queue_data bd = {
2511 * For OK queue, we are done. For error, caller may kill it.
2512 * Any other error (busy), just add it to our list as we
2513 * previously would have done.
2515 ret = q->mq_ops->queue_rq(hctx, &bd);
2518 blk_mq_update_dispatch_busy(hctx, false);
2520 case BLK_STS_RESOURCE:
2521 case BLK_STS_DEV_RESOURCE:
2522 blk_mq_update_dispatch_busy(hctx, true);
2523 __blk_mq_requeue_request(rq);
2526 blk_mq_update_dispatch_busy(hctx, false);
2533 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2535 bool bypass_insert, bool last)
2537 struct request_queue *q = rq->q;
2538 bool run_queue = true;
2542 * RCU or SRCU read lock is needed before checking quiesced flag.
2544 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2545 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2546 * and avoid driver to try to dispatch again.
2548 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2550 bypass_insert = false;
2554 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2557 budget_token = blk_mq_get_dispatch_budget(q);
2558 if (budget_token < 0)
2561 blk_mq_set_rq_budget_token(rq, budget_token);
2563 if (!blk_mq_get_driver_tag(rq)) {
2564 blk_mq_put_dispatch_budget(q, budget_token);
2568 return __blk_mq_issue_directly(hctx, rq, last);
2571 return BLK_STS_RESOURCE;
2573 blk_mq_sched_insert_request(rq, false, run_queue, false);
2579 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2580 * @hctx: Pointer of the associated hardware queue.
2581 * @rq: Pointer to request to be sent.
2583 * If the device has enough resources to accept a new request now, send the
2584 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2585 * we can try send it another time in the future. Requests inserted at this
2586 * queue have higher priority.
2588 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2592 __blk_mq_try_issue_directly(hctx, rq, false, true);
2594 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2595 blk_mq_request_bypass_insert(rq, false, true);
2596 else if (ret != BLK_STS_OK)
2597 blk_mq_end_request(rq, ret);
2600 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2602 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2605 static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2607 struct blk_mq_hw_ctx *hctx = NULL;
2612 while ((rq = rq_list_pop(&plug->mq_list))) {
2613 bool last = rq_list_empty(plug->mq_list);
2616 if (hctx != rq->mq_hctx) {
2618 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2622 ret = blk_mq_request_issue_directly(rq, last);
2627 case BLK_STS_RESOURCE:
2628 case BLK_STS_DEV_RESOURCE:
2629 blk_mq_request_bypass_insert(rq, false, true);
2630 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2633 blk_mq_end_request(rq, ret);
2640 * If we didn't flush the entire list, we could have told the driver
2641 * there was more coming, but that turned out to be a lie.
2644 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2647 static void __blk_mq_flush_plug_list(struct request_queue *q,
2648 struct blk_plug *plug)
2650 if (blk_queue_quiesced(q))
2652 q->mq_ops->queue_rqs(&plug->mq_list);
2655 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2657 struct blk_mq_hw_ctx *this_hctx = NULL;
2658 struct blk_mq_ctx *this_ctx = NULL;
2659 struct request *requeue_list = NULL;
2660 unsigned int depth = 0;
2664 struct request *rq = rq_list_pop(&plug->mq_list);
2667 this_hctx = rq->mq_hctx;
2668 this_ctx = rq->mq_ctx;
2669 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2670 rq_list_add(&requeue_list, rq);
2673 list_add_tail(&rq->queuelist, &list);
2675 } while (!rq_list_empty(plug->mq_list));
2677 plug->mq_list = requeue_list;
2678 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2679 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched);
2682 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2686 if (rq_list_empty(plug->mq_list))
2690 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2691 struct request_queue *q;
2693 rq = rq_list_peek(&plug->mq_list);
2697 * Peek first request and see if we have a ->queue_rqs() hook.
2698 * If we do, we can dispatch the whole plug list in one go. We
2699 * already know at this point that all requests belong to the
2700 * same queue, caller must ensure that's the case.
2702 * Since we pass off the full list to the driver at this point,
2703 * we do not increment the active request count for the queue.
2704 * Bypass shared tags for now because of that.
2706 if (q->mq_ops->queue_rqs &&
2707 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2708 blk_mq_run_dispatch_ops(q,
2709 __blk_mq_flush_plug_list(q, plug));
2710 if (rq_list_empty(plug->mq_list))
2714 blk_mq_run_dispatch_ops(q,
2715 blk_mq_plug_issue_direct(plug, false));
2716 if (rq_list_empty(plug->mq_list))
2721 blk_mq_dispatch_plug_list(plug, from_schedule);
2722 } while (!rq_list_empty(plug->mq_list));
2725 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2726 struct list_head *list)
2731 while (!list_empty(list)) {
2733 struct request *rq = list_first_entry(list, struct request,
2736 list_del_init(&rq->queuelist);
2737 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2738 if (ret != BLK_STS_OK) {
2740 if (ret == BLK_STS_RESOURCE ||
2741 ret == BLK_STS_DEV_RESOURCE) {
2742 blk_mq_request_bypass_insert(rq, false,
2746 blk_mq_end_request(rq, ret);
2752 * If we didn't flush the entire list, we could have told
2753 * the driver there was more coming, but that turned out to
2756 if ((!list_empty(list) || errors) &&
2757 hctx->queue->mq_ops->commit_rqs && queued)
2758 hctx->queue->mq_ops->commit_rqs(hctx);
2761 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2762 struct bio *bio, unsigned int nr_segs)
2764 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2765 if (blk_attempt_plug_merge(q, bio, nr_segs))
2767 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2773 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2774 struct blk_plug *plug,
2778 struct blk_mq_alloc_data data = {
2781 .cmd_flags = bio->bi_opf,
2785 if (unlikely(bio_queue_enter(bio)))
2788 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2791 rq_qos_throttle(q, bio);
2794 data.nr_tags = plug->nr_ios;
2796 data.cached_rq = &plug->cached_rq;
2799 rq = __blk_mq_alloc_requests(&data);
2802 rq_qos_cleanup(q, bio);
2803 if (bio->bi_opf & REQ_NOWAIT)
2804 bio_wouldblock_error(bio);
2810 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2811 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2817 rq = rq_list_peek(&plug->cached_rq);
2818 if (!rq || rq->q != q)
2821 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2826 if (blk_mq_get_hctx_type((*bio)->bi_opf) != rq->mq_hctx->type)
2828 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2832 * If any qos ->throttle() end up blocking, we will have flushed the
2833 * plug and hence killed the cached_rq list as well. Pop this entry
2834 * before we throttle.
2836 plug->cached_rq = rq_list_next(rq);
2837 rq_qos_throttle(q, *bio);
2839 rq->cmd_flags = (*bio)->bi_opf;
2840 INIT_LIST_HEAD(&rq->queuelist);
2844 static void bio_set_ioprio(struct bio *bio)
2846 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2847 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2848 bio->bi_ioprio = get_current_ioprio();
2849 blkcg_set_ioprio(bio);
2853 * blk_mq_submit_bio - Create and send a request to block device.
2854 * @bio: Bio pointer.
2856 * Builds up a request structure from @q and @bio and send to the device. The
2857 * request may not be queued directly to hardware if:
2858 * * This request can be merged with another one
2859 * * We want to place request at plug queue for possible future merging
2860 * * There is an IO scheduler active at this queue
2862 * It will not queue the request if there is an error with the bio, or at the
2865 void blk_mq_submit_bio(struct bio *bio)
2867 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2868 struct blk_plug *plug = blk_mq_plug(bio);
2869 const int is_sync = op_is_sync(bio->bi_opf);
2871 unsigned int nr_segs = 1;
2874 bio = blk_queue_bounce(bio, q);
2875 if (bio_may_exceed_limits(bio, &q->limits))
2876 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2878 if (!bio_integrity_prep(bio))
2881 bio_set_ioprio(bio);
2883 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2887 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2892 trace_block_getrq(bio);
2894 rq_qos_track(q, rq, bio);
2896 blk_mq_bio_to_request(rq, bio, nr_segs);
2898 ret = blk_crypto_init_request(rq);
2899 if (ret != BLK_STS_OK) {
2900 bio->bi_status = ret;
2902 blk_mq_free_request(rq);
2906 if (op_is_flush(bio->bi_opf)) {
2907 blk_insert_flush(rq);
2912 blk_add_rq_to_plug(plug, rq);
2913 else if ((rq->rq_flags & RQF_ELV) ||
2914 (rq->mq_hctx->dispatch_busy &&
2915 (q->nr_hw_queues == 1 || !is_sync)))
2916 blk_mq_sched_insert_request(rq, false, true, true);
2918 blk_mq_run_dispatch_ops(rq->q,
2919 blk_mq_try_issue_directly(rq->mq_hctx, rq));
2922 #ifdef CONFIG_BLK_MQ_STACKING
2924 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
2925 * @rq: the request being queued
2927 blk_status_t blk_insert_cloned_request(struct request *rq)
2929 struct request_queue *q = rq->q;
2930 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
2933 if (blk_rq_sectors(rq) > max_sectors) {
2935 * SCSI device does not have a good way to return if
2936 * Write Same/Zero is actually supported. If a device rejects
2937 * a non-read/write command (discard, write same,etc.) the
2938 * low-level device driver will set the relevant queue limit to
2939 * 0 to prevent blk-lib from issuing more of the offending
2940 * operations. Commands queued prior to the queue limit being
2941 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
2942 * errors being propagated to upper layers.
2944 if (max_sectors == 0)
2945 return BLK_STS_NOTSUPP;
2947 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
2948 __func__, blk_rq_sectors(rq), max_sectors);
2949 return BLK_STS_IOERR;
2953 * The queue settings related to segment counting may differ from the
2956 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
2957 if (rq->nr_phys_segments > queue_max_segments(q)) {
2958 printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n",
2959 __func__, rq->nr_phys_segments, queue_max_segments(q));
2960 return BLK_STS_IOERR;
2963 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
2964 return BLK_STS_IOERR;
2966 if (blk_crypto_insert_cloned_request(rq))
2967 return BLK_STS_IOERR;
2969 blk_account_io_start(rq);
2972 * Since we have a scheduler attached on the top device,
2973 * bypass a potential scheduler on the bottom device for
2976 blk_mq_run_dispatch_ops(q,
2977 ret = blk_mq_request_issue_directly(rq, true));
2979 blk_account_io_done(rq, ktime_get_ns());
2982 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
2985 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
2986 * @rq: the clone request to be cleaned up
2989 * Free all bios in @rq for a cloned request.
2991 void blk_rq_unprep_clone(struct request *rq)
2995 while ((bio = rq->bio) != NULL) {
2996 rq->bio = bio->bi_next;
3001 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3004 * blk_rq_prep_clone - Helper function to setup clone request
3005 * @rq: the request to be setup
3006 * @rq_src: original request to be cloned
3007 * @bs: bio_set that bios for clone are allocated from
3008 * @gfp_mask: memory allocation mask for bio
3009 * @bio_ctr: setup function to be called for each clone bio.
3010 * Returns %0 for success, non %0 for failure.
3011 * @data: private data to be passed to @bio_ctr
3014 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3015 * Also, pages which the original bios are pointing to are not copied
3016 * and the cloned bios just point same pages.
3017 * So cloned bios must be completed before original bios, which means
3018 * the caller must complete @rq before @rq_src.
3020 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3021 struct bio_set *bs, gfp_t gfp_mask,
3022 int (*bio_ctr)(struct bio *, struct bio *, void *),
3025 struct bio *bio, *bio_src;
3030 __rq_for_each_bio(bio_src, rq_src) {
3031 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3036 if (bio_ctr && bio_ctr(bio, bio_src, data))
3040 rq->biotail->bi_next = bio;
3043 rq->bio = rq->biotail = bio;
3048 /* Copy attributes of the original request to the clone request. */
3049 rq->__sector = blk_rq_pos(rq_src);
3050 rq->__data_len = blk_rq_bytes(rq_src);
3051 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3052 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3053 rq->special_vec = rq_src->special_vec;
3055 rq->nr_phys_segments = rq_src->nr_phys_segments;
3056 rq->ioprio = rq_src->ioprio;
3058 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3066 blk_rq_unprep_clone(rq);
3070 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3071 #endif /* CONFIG_BLK_MQ_STACKING */
3074 * Steal bios from a request and add them to a bio list.
3075 * The request must not have been partially completed before.
3077 void blk_steal_bios(struct bio_list *list, struct request *rq)
3081 list->tail->bi_next = rq->bio;
3083 list->head = rq->bio;
3084 list->tail = rq->biotail;
3092 EXPORT_SYMBOL_GPL(blk_steal_bios);
3094 static size_t order_to_size(unsigned int order)
3096 return (size_t)PAGE_SIZE << order;
3099 /* called before freeing request pool in @tags */
3100 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3101 struct blk_mq_tags *tags)
3104 unsigned long flags;
3106 /* There is no need to clear a driver tags own mapping */
3107 if (drv_tags == tags)
3110 list_for_each_entry(page, &tags->page_list, lru) {
3111 unsigned long start = (unsigned long)page_address(page);
3112 unsigned long end = start + order_to_size(page->private);
3115 for (i = 0; i < drv_tags->nr_tags; i++) {
3116 struct request *rq = drv_tags->rqs[i];
3117 unsigned long rq_addr = (unsigned long)rq;
3119 if (rq_addr >= start && rq_addr < end) {
3120 WARN_ON_ONCE(req_ref_read(rq) != 0);
3121 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3127 * Wait until all pending iteration is done.
3129 * Request reference is cleared and it is guaranteed to be observed
3130 * after the ->lock is released.
3132 spin_lock_irqsave(&drv_tags->lock, flags);
3133 spin_unlock_irqrestore(&drv_tags->lock, flags);
3136 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3137 unsigned int hctx_idx)
3139 struct blk_mq_tags *drv_tags;
3142 if (list_empty(&tags->page_list))
3145 if (blk_mq_is_shared_tags(set->flags))
3146 drv_tags = set->shared_tags;
3148 drv_tags = set->tags[hctx_idx];
3150 if (tags->static_rqs && set->ops->exit_request) {
3153 for (i = 0; i < tags->nr_tags; i++) {
3154 struct request *rq = tags->static_rqs[i];
3158 set->ops->exit_request(set, rq, hctx_idx);
3159 tags->static_rqs[i] = NULL;
3163 blk_mq_clear_rq_mapping(drv_tags, tags);
3165 while (!list_empty(&tags->page_list)) {
3166 page = list_first_entry(&tags->page_list, struct page, lru);
3167 list_del_init(&page->lru);
3169 * Remove kmemleak object previously allocated in
3170 * blk_mq_alloc_rqs().
3172 kmemleak_free(page_address(page));
3173 __free_pages(page, page->private);
3177 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3181 kfree(tags->static_rqs);
3182 tags->static_rqs = NULL;
3184 blk_mq_free_tags(tags);
3187 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3188 unsigned int hctx_idx)
3192 for (i = 0; i < set->nr_maps; i++) {
3193 unsigned int start = set->map[i].queue_offset;
3194 unsigned int end = start + set->map[i].nr_queues;
3196 if (hctx_idx >= start && hctx_idx < end)
3200 if (i >= set->nr_maps)
3201 i = HCTX_TYPE_DEFAULT;
3206 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3207 unsigned int hctx_idx)
3209 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3211 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3214 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3215 unsigned int hctx_idx,
3216 unsigned int nr_tags,
3217 unsigned int reserved_tags)
3219 int node = blk_mq_get_hctx_node(set, hctx_idx);
3220 struct blk_mq_tags *tags;
3222 if (node == NUMA_NO_NODE)
3223 node = set->numa_node;
3225 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3226 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3230 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3231 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3234 blk_mq_free_tags(tags);
3238 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3239 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3241 if (!tags->static_rqs) {
3243 blk_mq_free_tags(tags);
3250 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3251 unsigned int hctx_idx, int node)
3255 if (set->ops->init_request) {
3256 ret = set->ops->init_request(set, rq, hctx_idx, node);
3261 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3265 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3266 struct blk_mq_tags *tags,
3267 unsigned int hctx_idx, unsigned int depth)
3269 unsigned int i, j, entries_per_page, max_order = 4;
3270 int node = blk_mq_get_hctx_node(set, hctx_idx);
3271 size_t rq_size, left;
3273 if (node == NUMA_NO_NODE)
3274 node = set->numa_node;
3276 INIT_LIST_HEAD(&tags->page_list);
3279 * rq_size is the size of the request plus driver payload, rounded
3280 * to the cacheline size
3282 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3284 left = rq_size * depth;
3286 for (i = 0; i < depth; ) {
3287 int this_order = max_order;
3292 while (this_order && left < order_to_size(this_order - 1))
3296 page = alloc_pages_node(node,
3297 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3303 if (order_to_size(this_order) < rq_size)
3310 page->private = this_order;
3311 list_add_tail(&page->lru, &tags->page_list);
3313 p = page_address(page);
3315 * Allow kmemleak to scan these pages as they contain pointers
3316 * to additional allocations like via ops->init_request().
3318 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3319 entries_per_page = order_to_size(this_order) / rq_size;
3320 to_do = min(entries_per_page, depth - i);
3321 left -= to_do * rq_size;
3322 for (j = 0; j < to_do; j++) {
3323 struct request *rq = p;
3325 tags->static_rqs[i] = rq;
3326 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3327 tags->static_rqs[i] = NULL;
3338 blk_mq_free_rqs(set, tags, hctx_idx);
3342 struct rq_iter_data {
3343 struct blk_mq_hw_ctx *hctx;
3347 static bool blk_mq_has_request(struct request *rq, void *data)
3349 struct rq_iter_data *iter_data = data;
3351 if (rq->mq_hctx != iter_data->hctx)
3353 iter_data->has_rq = true;
3357 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3359 struct blk_mq_tags *tags = hctx->sched_tags ?
3360 hctx->sched_tags : hctx->tags;
3361 struct rq_iter_data data = {
3365 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3369 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3370 struct blk_mq_hw_ctx *hctx)
3372 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3374 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3379 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3381 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3382 struct blk_mq_hw_ctx, cpuhp_online);
3384 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3385 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3389 * Prevent new request from being allocated on the current hctx.
3391 * The smp_mb__after_atomic() Pairs with the implied barrier in
3392 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3393 * seen once we return from the tag allocator.
3395 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3396 smp_mb__after_atomic();
3399 * Try to grab a reference to the queue and wait for any outstanding
3400 * requests. If we could not grab a reference the queue has been
3401 * frozen and there are no requests.
3403 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3404 while (blk_mq_hctx_has_requests(hctx))
3406 percpu_ref_put(&hctx->queue->q_usage_counter);
3412 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3414 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3415 struct blk_mq_hw_ctx, cpuhp_online);
3417 if (cpumask_test_cpu(cpu, hctx->cpumask))
3418 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3423 * 'cpu' is going away. splice any existing rq_list entries from this
3424 * software queue to the hw queue dispatch list, and ensure that it
3427 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3429 struct blk_mq_hw_ctx *hctx;
3430 struct blk_mq_ctx *ctx;
3432 enum hctx_type type;
3434 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3435 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3438 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3441 spin_lock(&ctx->lock);
3442 if (!list_empty(&ctx->rq_lists[type])) {
3443 list_splice_init(&ctx->rq_lists[type], &tmp);
3444 blk_mq_hctx_clear_pending(hctx, ctx);
3446 spin_unlock(&ctx->lock);
3448 if (list_empty(&tmp))
3451 spin_lock(&hctx->lock);
3452 list_splice_tail_init(&tmp, &hctx->dispatch);
3453 spin_unlock(&hctx->lock);
3455 blk_mq_run_hw_queue(hctx, true);
3459 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3461 if (!(hctx->flags & BLK_MQ_F_STACKING))
3462 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3463 &hctx->cpuhp_online);
3464 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3469 * Before freeing hw queue, clearing the flush request reference in
3470 * tags->rqs[] for avoiding potential UAF.
3472 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3473 unsigned int queue_depth, struct request *flush_rq)
3476 unsigned long flags;
3478 /* The hw queue may not be mapped yet */
3482 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3484 for (i = 0; i < queue_depth; i++)
3485 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3488 * Wait until all pending iteration is done.
3490 * Request reference is cleared and it is guaranteed to be observed
3491 * after the ->lock is released.
3493 spin_lock_irqsave(&tags->lock, flags);
3494 spin_unlock_irqrestore(&tags->lock, flags);
3497 /* hctx->ctxs will be freed in queue's release handler */
3498 static void blk_mq_exit_hctx(struct request_queue *q,
3499 struct blk_mq_tag_set *set,
3500 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3502 struct request *flush_rq = hctx->fq->flush_rq;
3504 if (blk_mq_hw_queue_mapped(hctx))
3505 blk_mq_tag_idle(hctx);
3507 if (blk_queue_init_done(q))
3508 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3509 set->queue_depth, flush_rq);
3510 if (set->ops->exit_request)
3511 set->ops->exit_request(set, flush_rq, hctx_idx);
3513 if (set->ops->exit_hctx)
3514 set->ops->exit_hctx(hctx, hctx_idx);
3516 blk_mq_remove_cpuhp(hctx);
3518 xa_erase(&q->hctx_table, hctx_idx);
3520 spin_lock(&q->unused_hctx_lock);
3521 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3522 spin_unlock(&q->unused_hctx_lock);
3525 static void blk_mq_exit_hw_queues(struct request_queue *q,
3526 struct blk_mq_tag_set *set, int nr_queue)
3528 struct blk_mq_hw_ctx *hctx;
3531 queue_for_each_hw_ctx(q, hctx, i) {
3534 blk_mq_exit_hctx(q, set, hctx, i);
3538 static int blk_mq_init_hctx(struct request_queue *q,
3539 struct blk_mq_tag_set *set,
3540 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3542 hctx->queue_num = hctx_idx;
3544 if (!(hctx->flags & BLK_MQ_F_STACKING))
3545 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3546 &hctx->cpuhp_online);
3547 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3549 hctx->tags = set->tags[hctx_idx];
3551 if (set->ops->init_hctx &&
3552 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3553 goto unregister_cpu_notifier;
3555 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3559 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3565 if (set->ops->exit_request)
3566 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3568 if (set->ops->exit_hctx)
3569 set->ops->exit_hctx(hctx, hctx_idx);
3570 unregister_cpu_notifier:
3571 blk_mq_remove_cpuhp(hctx);
3575 static struct blk_mq_hw_ctx *
3576 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3579 struct blk_mq_hw_ctx *hctx;
3580 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3582 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3584 goto fail_alloc_hctx;
3586 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3589 atomic_set(&hctx->nr_active, 0);
3590 if (node == NUMA_NO_NODE)
3591 node = set->numa_node;
3592 hctx->numa_node = node;
3594 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3595 spin_lock_init(&hctx->lock);
3596 INIT_LIST_HEAD(&hctx->dispatch);
3598 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3600 INIT_LIST_HEAD(&hctx->hctx_list);
3603 * Allocate space for all possible cpus to avoid allocation at
3606 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3611 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3612 gfp, node, false, false))
3616 spin_lock_init(&hctx->dispatch_wait_lock);
3617 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3618 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3620 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3624 blk_mq_hctx_kobj_init(hctx);
3629 sbitmap_free(&hctx->ctx_map);
3633 free_cpumask_var(hctx->cpumask);
3640 static void blk_mq_init_cpu_queues(struct request_queue *q,
3641 unsigned int nr_hw_queues)
3643 struct blk_mq_tag_set *set = q->tag_set;
3646 for_each_possible_cpu(i) {
3647 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3648 struct blk_mq_hw_ctx *hctx;
3652 spin_lock_init(&__ctx->lock);
3653 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3654 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3659 * Set local node, IFF we have more than one hw queue. If
3660 * not, we remain on the home node of the device
3662 for (j = 0; j < set->nr_maps; j++) {
3663 hctx = blk_mq_map_queue_type(q, j, i);
3664 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3665 hctx->numa_node = cpu_to_node(i);
3670 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3671 unsigned int hctx_idx,
3674 struct blk_mq_tags *tags;
3677 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3681 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3683 blk_mq_free_rq_map(tags);
3690 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3693 if (blk_mq_is_shared_tags(set->flags)) {
3694 set->tags[hctx_idx] = set->shared_tags;
3699 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3702 return set->tags[hctx_idx];
3705 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3706 struct blk_mq_tags *tags,
3707 unsigned int hctx_idx)
3710 blk_mq_free_rqs(set, tags, hctx_idx);
3711 blk_mq_free_rq_map(tags);
3715 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3716 unsigned int hctx_idx)
3718 if (!blk_mq_is_shared_tags(set->flags))
3719 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3721 set->tags[hctx_idx] = NULL;
3724 static void blk_mq_map_swqueue(struct request_queue *q)
3726 unsigned int j, hctx_idx;
3728 struct blk_mq_hw_ctx *hctx;
3729 struct blk_mq_ctx *ctx;
3730 struct blk_mq_tag_set *set = q->tag_set;
3732 queue_for_each_hw_ctx(q, hctx, i) {
3733 cpumask_clear(hctx->cpumask);
3735 hctx->dispatch_from = NULL;
3739 * Map software to hardware queues.
3741 * If the cpu isn't present, the cpu is mapped to first hctx.
3743 for_each_possible_cpu(i) {
3745 ctx = per_cpu_ptr(q->queue_ctx, i);
3746 for (j = 0; j < set->nr_maps; j++) {
3747 if (!set->map[j].nr_queues) {
3748 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3749 HCTX_TYPE_DEFAULT, i);
3752 hctx_idx = set->map[j].mq_map[i];
3753 /* unmapped hw queue can be remapped after CPU topo changed */
3754 if (!set->tags[hctx_idx] &&
3755 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3757 * If tags initialization fail for some hctx,
3758 * that hctx won't be brought online. In this
3759 * case, remap the current ctx to hctx[0] which
3760 * is guaranteed to always have tags allocated
3762 set->map[j].mq_map[i] = 0;
3765 hctx = blk_mq_map_queue_type(q, j, i);
3766 ctx->hctxs[j] = hctx;
3768 * If the CPU is already set in the mask, then we've
3769 * mapped this one already. This can happen if
3770 * devices share queues across queue maps.
3772 if (cpumask_test_cpu(i, hctx->cpumask))
3775 cpumask_set_cpu(i, hctx->cpumask);
3777 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3778 hctx->ctxs[hctx->nr_ctx++] = ctx;
3781 * If the nr_ctx type overflows, we have exceeded the
3782 * amount of sw queues we can support.
3784 BUG_ON(!hctx->nr_ctx);
3787 for (; j < HCTX_MAX_TYPES; j++)
3788 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3789 HCTX_TYPE_DEFAULT, i);
3792 queue_for_each_hw_ctx(q, hctx, i) {
3794 * If no software queues are mapped to this hardware queue,
3795 * disable it and free the request entries.
3797 if (!hctx->nr_ctx) {
3798 /* Never unmap queue 0. We need it as a
3799 * fallback in case of a new remap fails
3803 __blk_mq_free_map_and_rqs(set, i);
3809 hctx->tags = set->tags[i];
3810 WARN_ON(!hctx->tags);
3813 * Set the map size to the number of mapped software queues.
3814 * This is more accurate and more efficient than looping
3815 * over all possibly mapped software queues.
3817 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3820 * Initialize batch roundrobin counts
3822 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3823 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3828 * Caller needs to ensure that we're either frozen/quiesced, or that
3829 * the queue isn't live yet.
3831 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3833 struct blk_mq_hw_ctx *hctx;
3836 queue_for_each_hw_ctx(q, hctx, i) {
3838 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3840 blk_mq_tag_idle(hctx);
3841 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3846 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3849 struct request_queue *q;
3851 lockdep_assert_held(&set->tag_list_lock);
3853 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3854 blk_mq_freeze_queue(q);
3855 queue_set_hctx_shared(q, shared);
3856 blk_mq_unfreeze_queue(q);
3860 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3862 struct blk_mq_tag_set *set = q->tag_set;
3864 mutex_lock(&set->tag_list_lock);
3865 list_del(&q->tag_set_list);
3866 if (list_is_singular(&set->tag_list)) {
3867 /* just transitioned to unshared */
3868 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3869 /* update existing queue */
3870 blk_mq_update_tag_set_shared(set, false);
3872 mutex_unlock(&set->tag_list_lock);
3873 INIT_LIST_HEAD(&q->tag_set_list);
3876 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3877 struct request_queue *q)
3879 mutex_lock(&set->tag_list_lock);
3882 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3884 if (!list_empty(&set->tag_list) &&
3885 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3886 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3887 /* update existing queue */
3888 blk_mq_update_tag_set_shared(set, true);
3890 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3891 queue_set_hctx_shared(q, true);
3892 list_add_tail(&q->tag_set_list, &set->tag_list);
3894 mutex_unlock(&set->tag_list_lock);
3897 /* All allocations will be freed in release handler of q->mq_kobj */
3898 static int blk_mq_alloc_ctxs(struct request_queue *q)
3900 struct blk_mq_ctxs *ctxs;
3903 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3907 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3908 if (!ctxs->queue_ctx)
3911 for_each_possible_cpu(cpu) {
3912 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3916 q->mq_kobj = &ctxs->kobj;
3917 q->queue_ctx = ctxs->queue_ctx;
3926 * It is the actual release handler for mq, but we do it from
3927 * request queue's release handler for avoiding use-after-free
3928 * and headache because q->mq_kobj shouldn't have been introduced,
3929 * but we can't group ctx/kctx kobj without it.
3931 void blk_mq_release(struct request_queue *q)
3933 struct blk_mq_hw_ctx *hctx, *next;
3936 queue_for_each_hw_ctx(q, hctx, i)
3937 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3939 /* all hctx are in .unused_hctx_list now */
3940 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3941 list_del_init(&hctx->hctx_list);
3942 kobject_put(&hctx->kobj);
3945 xa_destroy(&q->hctx_table);
3948 * release .mq_kobj and sw queue's kobject now because
3949 * both share lifetime with request queue.
3951 blk_mq_sysfs_deinit(q);
3954 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3957 struct request_queue *q;
3960 q = blk_alloc_queue(set->numa_node, set->flags & BLK_MQ_F_BLOCKING);
3962 return ERR_PTR(-ENOMEM);
3963 q->queuedata = queuedata;
3964 ret = blk_mq_init_allocated_queue(set, q);
3967 return ERR_PTR(ret);
3972 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3974 return blk_mq_init_queue_data(set, NULL);
3976 EXPORT_SYMBOL(blk_mq_init_queue);
3979 * blk_mq_destroy_queue - shutdown a request queue
3980 * @q: request queue to shutdown
3982 * This shuts down a request queue allocated by blk_mq_init_queue() and drops
3983 * the initial reference. All future requests will failed with -ENODEV.
3985 * Context: can sleep
3987 void blk_mq_destroy_queue(struct request_queue *q)
3989 WARN_ON_ONCE(!queue_is_mq(q));
3990 WARN_ON_ONCE(blk_queue_registered(q));
3994 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
3995 blk_queue_start_drain(q);
3996 blk_freeze_queue(q);
3999 blk_mq_cancel_work_sync(q);
4000 blk_mq_exit_queue(q);
4002 /* @q is and will stay empty, shutdown and put */
4005 EXPORT_SYMBOL(blk_mq_destroy_queue);
4007 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4008 struct lock_class_key *lkclass)
4010 struct request_queue *q;
4011 struct gendisk *disk;
4013 q = blk_mq_init_queue_data(set, queuedata);
4017 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4019 blk_mq_destroy_queue(q);
4020 return ERR_PTR(-ENOMEM);
4022 set_bit(GD_OWNS_QUEUE, &disk->state);
4025 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4027 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4028 struct lock_class_key *lkclass)
4030 if (!blk_get_queue(q))
4032 return __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4034 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4036 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4037 struct blk_mq_tag_set *set, struct request_queue *q,
4038 int hctx_idx, int node)
4040 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4042 /* reuse dead hctx first */
4043 spin_lock(&q->unused_hctx_lock);
4044 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4045 if (tmp->numa_node == node) {
4051 list_del_init(&hctx->hctx_list);
4052 spin_unlock(&q->unused_hctx_lock);
4055 hctx = blk_mq_alloc_hctx(q, set, node);
4059 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4065 kobject_put(&hctx->kobj);
4070 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4071 struct request_queue *q)
4073 struct blk_mq_hw_ctx *hctx;
4076 /* protect against switching io scheduler */
4077 mutex_lock(&q->sysfs_lock);
4078 for (i = 0; i < set->nr_hw_queues; i++) {
4080 int node = blk_mq_get_hctx_node(set, i);
4081 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4084 old_node = old_hctx->numa_node;
4085 blk_mq_exit_hctx(q, set, old_hctx, i);
4088 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4091 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4093 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4094 WARN_ON_ONCE(!hctx);
4098 * Increasing nr_hw_queues fails. Free the newly allocated
4099 * hctxs and keep the previous q->nr_hw_queues.
4101 if (i != set->nr_hw_queues) {
4102 j = q->nr_hw_queues;
4105 q->nr_hw_queues = set->nr_hw_queues;
4108 xa_for_each_start(&q->hctx_table, j, hctx, j)
4109 blk_mq_exit_hctx(q, set, hctx, j);
4110 mutex_unlock(&q->sysfs_lock);
4113 static void blk_mq_update_poll_flag(struct request_queue *q)
4115 struct blk_mq_tag_set *set = q->tag_set;
4117 if (set->nr_maps > HCTX_TYPE_POLL &&
4118 set->map[HCTX_TYPE_POLL].nr_queues)
4119 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4121 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4124 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4125 struct request_queue *q)
4127 WARN_ON_ONCE(blk_queue_has_srcu(q) !=
4128 !!(set->flags & BLK_MQ_F_BLOCKING));
4130 /* mark the queue as mq asap */
4131 q->mq_ops = set->ops;
4133 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
4134 blk_mq_poll_stats_bkt,
4135 BLK_MQ_POLL_STATS_BKTS, q);
4139 if (blk_mq_alloc_ctxs(q))
4142 /* init q->mq_kobj and sw queues' kobjects */
4143 blk_mq_sysfs_init(q);
4145 INIT_LIST_HEAD(&q->unused_hctx_list);
4146 spin_lock_init(&q->unused_hctx_lock);
4148 xa_init(&q->hctx_table);
4150 blk_mq_realloc_hw_ctxs(set, q);
4151 if (!q->nr_hw_queues)
4154 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4155 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4159 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4160 blk_mq_update_poll_flag(q);
4162 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4163 INIT_LIST_HEAD(&q->requeue_list);
4164 spin_lock_init(&q->requeue_lock);
4166 q->nr_requests = set->queue_depth;
4169 * Default to classic polling
4171 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
4173 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4174 blk_mq_add_queue_tag_set(set, q);
4175 blk_mq_map_swqueue(q);
4179 xa_destroy(&q->hctx_table);
4180 q->nr_hw_queues = 0;
4181 blk_mq_sysfs_deinit(q);
4183 blk_stat_free_callback(q->poll_cb);
4189 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4191 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4192 void blk_mq_exit_queue(struct request_queue *q)
4194 struct blk_mq_tag_set *set = q->tag_set;
4196 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4197 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4198 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4199 blk_mq_del_queue_tag_set(q);
4202 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4206 if (blk_mq_is_shared_tags(set->flags)) {
4207 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4210 if (!set->shared_tags)
4214 for (i = 0; i < set->nr_hw_queues; i++) {
4215 if (!__blk_mq_alloc_map_and_rqs(set, i))
4224 __blk_mq_free_map_and_rqs(set, i);
4226 if (blk_mq_is_shared_tags(set->flags)) {
4227 blk_mq_free_map_and_rqs(set, set->shared_tags,
4228 BLK_MQ_NO_HCTX_IDX);
4235 * Allocate the request maps associated with this tag_set. Note that this
4236 * may reduce the depth asked for, if memory is tight. set->queue_depth
4237 * will be updated to reflect the allocated depth.
4239 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4244 depth = set->queue_depth;
4246 err = __blk_mq_alloc_rq_maps(set);
4250 set->queue_depth >>= 1;
4251 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4255 } while (set->queue_depth);
4257 if (!set->queue_depth || err) {
4258 pr_err("blk-mq: failed to allocate request map\n");
4262 if (depth != set->queue_depth)
4263 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4264 depth, set->queue_depth);
4269 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4272 * blk_mq_map_queues() and multiple .map_queues() implementations
4273 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4274 * number of hardware queues.
4276 if (set->nr_maps == 1)
4277 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4279 if (set->ops->map_queues && !is_kdump_kernel()) {
4283 * transport .map_queues is usually done in the following
4286 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4287 * mask = get_cpu_mask(queue)
4288 * for_each_cpu(cpu, mask)
4289 * set->map[x].mq_map[cpu] = queue;
4292 * When we need to remap, the table has to be cleared for
4293 * killing stale mapping since one CPU may not be mapped
4296 for (i = 0; i < set->nr_maps; i++)
4297 blk_mq_clear_mq_map(&set->map[i]);
4299 set->ops->map_queues(set);
4301 BUG_ON(set->nr_maps > 1);
4302 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4306 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4307 int cur_nr_hw_queues, int new_nr_hw_queues)
4309 struct blk_mq_tags **new_tags;
4311 if (cur_nr_hw_queues >= new_nr_hw_queues)
4314 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4315 GFP_KERNEL, set->numa_node);
4320 memcpy(new_tags, set->tags, cur_nr_hw_queues *
4321 sizeof(*set->tags));
4323 set->tags = new_tags;
4324 set->nr_hw_queues = new_nr_hw_queues;
4329 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
4330 int new_nr_hw_queues)
4332 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
4336 * Alloc a tag set to be associated with one or more request queues.
4337 * May fail with EINVAL for various error conditions. May adjust the
4338 * requested depth down, if it's too large. In that case, the set
4339 * value will be stored in set->queue_depth.
4341 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4345 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4347 if (!set->nr_hw_queues)
4349 if (!set->queue_depth)
4351 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4354 if (!set->ops->queue_rq)
4357 if (!set->ops->get_budget ^ !set->ops->put_budget)
4360 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4361 pr_info("blk-mq: reduced tag depth to %u\n",
4363 set->queue_depth = BLK_MQ_MAX_DEPTH;
4368 else if (set->nr_maps > HCTX_MAX_TYPES)
4372 * If a crashdump is active, then we are potentially in a very
4373 * memory constrained environment. Limit us to 1 queue and
4374 * 64 tags to prevent using too much memory.
4376 if (is_kdump_kernel()) {
4377 set->nr_hw_queues = 1;
4379 set->queue_depth = min(64U, set->queue_depth);
4382 * There is no use for more h/w queues than cpus if we just have
4385 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4386 set->nr_hw_queues = nr_cpu_ids;
4388 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
4392 for (i = 0; i < set->nr_maps; i++) {
4393 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4394 sizeof(set->map[i].mq_map[0]),
4395 GFP_KERNEL, set->numa_node);
4396 if (!set->map[i].mq_map)
4397 goto out_free_mq_map;
4398 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4401 blk_mq_update_queue_map(set);
4403 ret = blk_mq_alloc_set_map_and_rqs(set);
4405 goto out_free_mq_map;
4407 mutex_init(&set->tag_list_lock);
4408 INIT_LIST_HEAD(&set->tag_list);
4413 for (i = 0; i < set->nr_maps; i++) {
4414 kfree(set->map[i].mq_map);
4415 set->map[i].mq_map = NULL;
4421 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4423 /* allocate and initialize a tagset for a simple single-queue device */
4424 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4425 const struct blk_mq_ops *ops, unsigned int queue_depth,
4426 unsigned int set_flags)
4428 memset(set, 0, sizeof(*set));
4430 set->nr_hw_queues = 1;
4432 set->queue_depth = queue_depth;
4433 set->numa_node = NUMA_NO_NODE;
4434 set->flags = set_flags;
4435 return blk_mq_alloc_tag_set(set);
4437 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4439 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4443 for (i = 0; i < set->nr_hw_queues; i++)
4444 __blk_mq_free_map_and_rqs(set, i);
4446 if (blk_mq_is_shared_tags(set->flags)) {
4447 blk_mq_free_map_and_rqs(set, set->shared_tags,
4448 BLK_MQ_NO_HCTX_IDX);
4451 for (j = 0; j < set->nr_maps; j++) {
4452 kfree(set->map[j].mq_map);
4453 set->map[j].mq_map = NULL;
4459 EXPORT_SYMBOL(blk_mq_free_tag_set);
4461 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4463 struct blk_mq_tag_set *set = q->tag_set;
4464 struct blk_mq_hw_ctx *hctx;
4471 if (q->nr_requests == nr)
4474 blk_mq_freeze_queue(q);
4475 blk_mq_quiesce_queue(q);
4478 queue_for_each_hw_ctx(q, hctx, i) {
4482 * If we're using an MQ scheduler, just update the scheduler
4483 * queue depth. This is similar to what the old code would do.
4485 if (hctx->sched_tags) {
4486 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4489 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4494 if (q->elevator && q->elevator->type->ops.depth_updated)
4495 q->elevator->type->ops.depth_updated(hctx);
4498 q->nr_requests = nr;
4499 if (blk_mq_is_shared_tags(set->flags)) {
4501 blk_mq_tag_update_sched_shared_tags(q);
4503 blk_mq_tag_resize_shared_tags(set, nr);
4507 blk_mq_unquiesce_queue(q);
4508 blk_mq_unfreeze_queue(q);
4514 * request_queue and elevator_type pair.
4515 * It is just used by __blk_mq_update_nr_hw_queues to cache
4516 * the elevator_type associated with a request_queue.
4518 struct blk_mq_qe_pair {
4519 struct list_head node;
4520 struct request_queue *q;
4521 struct elevator_type *type;
4525 * Cache the elevator_type in qe pair list and switch the
4526 * io scheduler to 'none'
4528 static bool blk_mq_elv_switch_none(struct list_head *head,
4529 struct request_queue *q)
4531 struct blk_mq_qe_pair *qe;
4536 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4540 /* q->elevator needs protection from ->sysfs_lock */
4541 mutex_lock(&q->sysfs_lock);
4543 INIT_LIST_HEAD(&qe->node);
4545 qe->type = q->elevator->type;
4546 list_add(&qe->node, head);
4549 * After elevator_switch, the previous elevator_queue will be
4550 * released by elevator_release. The reference of the io scheduler
4551 * module get by elevator_get will also be put. So we need to get
4552 * a reference of the io scheduler module here to prevent it to be
4555 __module_get(qe->type->elevator_owner);
4556 elevator_switch(q, NULL);
4557 mutex_unlock(&q->sysfs_lock);
4562 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4563 struct request_queue *q)
4565 struct blk_mq_qe_pair *qe;
4567 list_for_each_entry(qe, head, node)
4574 static void blk_mq_elv_switch_back(struct list_head *head,
4575 struct request_queue *q)
4577 struct blk_mq_qe_pair *qe;
4578 struct elevator_type *t;
4580 qe = blk_lookup_qe_pair(head, q);
4584 list_del(&qe->node);
4587 mutex_lock(&q->sysfs_lock);
4588 elevator_switch(q, t);
4589 mutex_unlock(&q->sysfs_lock);
4592 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4595 struct request_queue *q;
4597 int prev_nr_hw_queues;
4599 lockdep_assert_held(&set->tag_list_lock);
4601 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4602 nr_hw_queues = nr_cpu_ids;
4603 if (nr_hw_queues < 1)
4605 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4608 list_for_each_entry(q, &set->tag_list, tag_set_list)
4609 blk_mq_freeze_queue(q);
4611 * Switch IO scheduler to 'none', cleaning up the data associated
4612 * with the previous scheduler. We will switch back once we are done
4613 * updating the new sw to hw queue mappings.
4615 list_for_each_entry(q, &set->tag_list, tag_set_list)
4616 if (!blk_mq_elv_switch_none(&head, q))
4619 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4620 blk_mq_debugfs_unregister_hctxs(q);
4621 blk_mq_sysfs_unregister_hctxs(q);
4624 prev_nr_hw_queues = set->nr_hw_queues;
4625 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
4629 set->nr_hw_queues = nr_hw_queues;
4631 blk_mq_update_queue_map(set);
4632 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4633 blk_mq_realloc_hw_ctxs(set, q);
4634 blk_mq_update_poll_flag(q);
4635 if (q->nr_hw_queues != set->nr_hw_queues) {
4636 int i = prev_nr_hw_queues;
4638 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4639 nr_hw_queues, prev_nr_hw_queues);
4640 for (; i < set->nr_hw_queues; i++)
4641 __blk_mq_free_map_and_rqs(set, i);
4643 set->nr_hw_queues = prev_nr_hw_queues;
4644 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4647 blk_mq_map_swqueue(q);
4651 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4652 blk_mq_sysfs_register_hctxs(q);
4653 blk_mq_debugfs_register_hctxs(q);
4657 list_for_each_entry(q, &set->tag_list, tag_set_list)
4658 blk_mq_elv_switch_back(&head, q);
4660 list_for_each_entry(q, &set->tag_list, tag_set_list)
4661 blk_mq_unfreeze_queue(q);
4664 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4666 mutex_lock(&set->tag_list_lock);
4667 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4668 mutex_unlock(&set->tag_list_lock);
4670 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4672 /* Enable polling stats and return whether they were already enabled. */
4673 static bool blk_poll_stats_enable(struct request_queue *q)
4678 return blk_stats_alloc_enable(q);
4681 static void blk_mq_poll_stats_start(struct request_queue *q)
4684 * We don't arm the callback if polling stats are not enabled or the
4685 * callback is already active.
4687 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
4690 blk_stat_activate_msecs(q->poll_cb, 100);
4693 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4695 struct request_queue *q = cb->data;
4698 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4699 if (cb->stat[bucket].nr_samples)
4700 q->poll_stat[bucket] = cb->stat[bucket];
4704 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4707 unsigned long ret = 0;
4711 * If stats collection isn't on, don't sleep but turn it on for
4714 if (!blk_poll_stats_enable(q))
4718 * As an optimistic guess, use half of the mean service time
4719 * for this type of request. We can (and should) make this smarter.
4720 * For instance, if the completion latencies are tight, we can
4721 * get closer than just half the mean. This is especially
4722 * important on devices where the completion latencies are longer
4723 * than ~10 usec. We do use the stats for the relevant IO size
4724 * if available which does lead to better estimates.
4726 bucket = blk_mq_poll_stats_bkt(rq);
4730 if (q->poll_stat[bucket].nr_samples)
4731 ret = (q->poll_stat[bucket].mean + 1) / 2;
4736 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4738 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4739 struct request *rq = blk_qc_to_rq(hctx, qc);
4740 struct hrtimer_sleeper hs;
4741 enum hrtimer_mode mode;
4746 * If a request has completed on queue that uses an I/O scheduler, we
4747 * won't get back a request from blk_qc_to_rq.
4749 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4753 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4755 * 0: use half of prev avg
4756 * >0: use this specific value
4758 if (q->poll_nsec > 0)
4759 nsecs = q->poll_nsec;
4761 nsecs = blk_mq_poll_nsecs(q, rq);
4766 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4769 * This will be replaced with the stats tracking code, using
4770 * 'avg_completion_time / 2' as the pre-sleep target.
4774 mode = HRTIMER_MODE_REL;
4775 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4776 hrtimer_set_expires(&hs.timer, kt);
4779 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4781 set_current_state(TASK_UNINTERRUPTIBLE);
4782 hrtimer_sleeper_start_expires(&hs, mode);
4785 hrtimer_cancel(&hs.timer);
4786 mode = HRTIMER_MODE_ABS;
4787 } while (hs.task && !signal_pending(current));
4789 __set_current_state(TASK_RUNNING);
4790 destroy_hrtimer_on_stack(&hs.timer);
4793 * If we sleep, have the caller restart the poll loop to reset the
4794 * state. Like for the other success return cases, the caller is
4795 * responsible for checking if the IO completed. If the IO isn't
4796 * complete, we'll get called again and will go straight to the busy
4802 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4803 struct io_comp_batch *iob, unsigned int flags)
4805 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4806 long state = get_current_state();
4810 ret = q->mq_ops->poll(hctx, iob);
4812 __set_current_state(TASK_RUNNING);
4816 if (signal_pending_state(state, current))
4817 __set_current_state(TASK_RUNNING);
4818 if (task_is_running(current))
4821 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4824 } while (!need_resched());
4826 __set_current_state(TASK_RUNNING);
4830 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4833 if (!(flags & BLK_POLL_NOSLEEP) &&
4834 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4835 if (blk_mq_poll_hybrid(q, cookie))
4838 return blk_mq_poll_classic(q, cookie, iob, flags);
4841 unsigned int blk_mq_rq_cpu(struct request *rq)
4843 return rq->mq_ctx->cpu;
4845 EXPORT_SYMBOL(blk_mq_rq_cpu);
4847 void blk_mq_cancel_work_sync(struct request_queue *q)
4849 if (queue_is_mq(q)) {
4850 struct blk_mq_hw_ctx *hctx;
4853 cancel_delayed_work_sync(&q->requeue_work);
4855 queue_for_each_hw_ctx(q, hctx, i)
4856 cancel_delayed_work_sync(&hctx->run_work);
4860 static int __init blk_mq_init(void)
4864 for_each_possible_cpu(i)
4865 init_llist_head(&per_cpu(blk_cpu_done, i));
4866 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4868 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4869 "block/softirq:dead", NULL,
4870 blk_softirq_cpu_dead);
4871 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4872 blk_mq_hctx_notify_dead);
4873 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4874 blk_mq_hctx_notify_online,
4875 blk_mq_hctx_notify_offline);
4878 subsys_initcall(blk_mq_init);