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 struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
514 blk_mq_req_flags_t flags)
516 struct blk_mq_alloc_data data = {
525 ret = blk_queue_enter(q, flags);
529 rq = __blk_mq_alloc_requests(&data);
533 rq->__sector = (sector_t) -1;
534 rq->bio = rq->biotail = NULL;
538 return ERR_PTR(-EWOULDBLOCK);
540 EXPORT_SYMBOL(blk_mq_alloc_request);
542 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
543 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
545 struct blk_mq_alloc_data data = {
551 u64 alloc_time_ns = 0;
556 /* alloc_time includes depth and tag waits */
557 if (blk_queue_rq_alloc_time(q))
558 alloc_time_ns = ktime_get_ns();
561 * If the tag allocator sleeps we could get an allocation for a
562 * different hardware context. No need to complicate the low level
563 * allocator for this for the rare use case of a command tied to
566 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
567 return ERR_PTR(-EINVAL);
569 if (hctx_idx >= q->nr_hw_queues)
570 return ERR_PTR(-EIO);
572 ret = blk_queue_enter(q, flags);
577 * Check if the hardware context is actually mapped to anything.
578 * If not tell the caller that it should skip this queue.
581 data.hctx = xa_load(&q->hctx_table, hctx_idx);
582 if (!blk_mq_hw_queue_mapped(data.hctx))
584 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
585 if (cpu >= nr_cpu_ids)
587 data.ctx = __blk_mq_get_ctx(q, cpu);
590 blk_mq_tag_busy(data.hctx);
592 data.rq_flags |= RQF_ELV;
594 if (flags & BLK_MQ_REQ_RESERVED)
595 data.rq_flags |= RQF_RESV;
598 tag = blk_mq_get_tag(&data);
599 if (tag == BLK_MQ_NO_TAG)
601 return blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag,
608 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
610 static void __blk_mq_free_request(struct request *rq)
612 struct request_queue *q = rq->q;
613 struct blk_mq_ctx *ctx = rq->mq_ctx;
614 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
615 const int sched_tag = rq->internal_tag;
617 blk_crypto_free_request(rq);
618 blk_pm_mark_last_busy(rq);
620 if (rq->tag != BLK_MQ_NO_TAG)
621 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
622 if (sched_tag != BLK_MQ_NO_TAG)
623 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
624 blk_mq_sched_restart(hctx);
628 void blk_mq_free_request(struct request *rq)
630 struct request_queue *q = rq->q;
631 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
633 if ((rq->rq_flags & RQF_ELVPRIV) &&
634 q->elevator->type->ops.finish_request)
635 q->elevator->type->ops.finish_request(rq);
637 if (rq->rq_flags & RQF_MQ_INFLIGHT)
638 __blk_mq_dec_active_requests(hctx);
640 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
641 laptop_io_completion(q->disk->bdi);
645 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
646 if (req_ref_put_and_test(rq))
647 __blk_mq_free_request(rq);
649 EXPORT_SYMBOL_GPL(blk_mq_free_request);
651 void blk_mq_free_plug_rqs(struct blk_plug *plug)
655 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
656 blk_mq_free_request(rq);
659 void blk_dump_rq_flags(struct request *rq, char *msg)
661 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
662 rq->q->disk ? rq->q->disk->disk_name : "?",
663 (__force unsigned long long) rq->cmd_flags);
665 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
666 (unsigned long long)blk_rq_pos(rq),
667 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
668 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
669 rq->bio, rq->biotail, blk_rq_bytes(rq));
671 EXPORT_SYMBOL(blk_dump_rq_flags);
673 static void req_bio_endio(struct request *rq, struct bio *bio,
674 unsigned int nbytes, blk_status_t error)
676 if (unlikely(error)) {
677 bio->bi_status = error;
678 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
680 * Partial zone append completions cannot be supported as the
681 * BIO fragments may end up not being written sequentially.
683 if (bio->bi_iter.bi_size != nbytes)
684 bio->bi_status = BLK_STS_IOERR;
686 bio->bi_iter.bi_sector = rq->__sector;
689 bio_advance(bio, nbytes);
691 if (unlikely(rq->rq_flags & RQF_QUIET))
692 bio_set_flag(bio, BIO_QUIET);
693 /* don't actually finish bio if it's part of flush sequence */
694 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
698 static void blk_account_io_completion(struct request *req, unsigned int bytes)
700 if (req->part && blk_do_io_stat(req)) {
701 const int sgrp = op_stat_group(req_op(req));
704 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
709 static void blk_print_req_error(struct request *req, blk_status_t status)
711 printk_ratelimited(KERN_ERR
712 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
713 "phys_seg %u prio class %u\n",
714 blk_status_to_str(status),
715 req->q->disk ? req->q->disk->disk_name : "?",
716 blk_rq_pos(req), (__force u32)req_op(req),
717 blk_op_str(req_op(req)),
718 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
719 req->nr_phys_segments,
720 IOPRIO_PRIO_CLASS(req->ioprio));
724 * Fully end IO on a request. Does not support partial completions, or
727 static void blk_complete_request(struct request *req)
729 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
730 int total_bytes = blk_rq_bytes(req);
731 struct bio *bio = req->bio;
733 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
738 #ifdef CONFIG_BLK_DEV_INTEGRITY
739 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
740 req->q->integrity.profile->complete_fn(req, total_bytes);
743 blk_account_io_completion(req, total_bytes);
746 struct bio *next = bio->bi_next;
748 /* Completion has already been traced */
749 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
751 if (req_op(req) == REQ_OP_ZONE_APPEND)
752 bio->bi_iter.bi_sector = req->__sector;
760 * Reset counters so that the request stacking driver
761 * can find how many bytes remain in the request
769 * blk_update_request - Complete multiple bytes without completing the request
770 * @req: the request being processed
771 * @error: block status code
772 * @nr_bytes: number of bytes to complete for @req
775 * Ends I/O on a number of bytes attached to @req, but doesn't complete
776 * the request structure even if @req doesn't have leftover.
777 * If @req has leftover, sets it up for the next range of segments.
779 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
780 * %false return from this function.
783 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
784 * except in the consistency check at the end of this function.
787 * %false - this request doesn't have any more data
788 * %true - this request has more data
790 bool blk_update_request(struct request *req, blk_status_t error,
791 unsigned int nr_bytes)
795 trace_block_rq_complete(req, error, nr_bytes);
800 #ifdef CONFIG_BLK_DEV_INTEGRITY
801 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
803 req->q->integrity.profile->complete_fn(req, nr_bytes);
806 if (unlikely(error && !blk_rq_is_passthrough(req) &&
807 !(req->rq_flags & RQF_QUIET)) &&
808 !test_bit(GD_DEAD, &req->q->disk->state)) {
809 blk_print_req_error(req, error);
810 trace_block_rq_error(req, error, nr_bytes);
813 blk_account_io_completion(req, nr_bytes);
817 struct bio *bio = req->bio;
818 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
820 if (bio_bytes == bio->bi_iter.bi_size)
821 req->bio = bio->bi_next;
823 /* Completion has already been traced */
824 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
825 req_bio_endio(req, bio, bio_bytes, error);
827 total_bytes += bio_bytes;
828 nr_bytes -= bio_bytes;
839 * Reset counters so that the request stacking driver
840 * can find how many bytes remain in the request
847 req->__data_len -= total_bytes;
849 /* update sector only for requests with clear definition of sector */
850 if (!blk_rq_is_passthrough(req))
851 req->__sector += total_bytes >> 9;
853 /* mixed attributes always follow the first bio */
854 if (req->rq_flags & RQF_MIXED_MERGE) {
855 req->cmd_flags &= ~REQ_FAILFAST_MASK;
856 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
859 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
861 * If total number of sectors is less than the first segment
862 * size, something has gone terribly wrong.
864 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
865 blk_dump_rq_flags(req, "request botched");
866 req->__data_len = blk_rq_cur_bytes(req);
869 /* recalculate the number of segments */
870 req->nr_phys_segments = blk_recalc_rq_segments(req);
875 EXPORT_SYMBOL_GPL(blk_update_request);
877 static void __blk_account_io_done(struct request *req, u64 now)
879 const int sgrp = op_stat_group(req_op(req));
882 update_io_ticks(req->part, jiffies, true);
883 part_stat_inc(req->part, ios[sgrp]);
884 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
888 static inline void blk_account_io_done(struct request *req, u64 now)
891 * Account IO completion. flush_rq isn't accounted as a
892 * normal IO on queueing nor completion. Accounting the
893 * containing request is enough.
895 if (blk_do_io_stat(req) && req->part &&
896 !(req->rq_flags & RQF_FLUSH_SEQ))
897 __blk_account_io_done(req, now);
900 static void __blk_account_io_start(struct request *rq)
903 * All non-passthrough requests are created from a bio with one
904 * exception: when a flush command that is part of a flush sequence
905 * generated by the state machine in blk-flush.c is cloned onto the
906 * lower device by dm-multipath we can get here without a bio.
909 rq->part = rq->bio->bi_bdev;
911 rq->part = rq->q->disk->part0;
914 update_io_ticks(rq->part, jiffies, false);
918 static inline void blk_account_io_start(struct request *req)
920 if (blk_do_io_stat(req))
921 __blk_account_io_start(req);
924 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
926 if (rq->rq_flags & RQF_STATS) {
927 blk_mq_poll_stats_start(rq->q);
928 blk_stat_add(rq, now);
931 blk_mq_sched_completed_request(rq, now);
932 blk_account_io_done(rq, now);
935 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
937 if (blk_mq_need_time_stamp(rq))
938 __blk_mq_end_request_acct(rq, ktime_get_ns());
941 rq_qos_done(rq->q, rq);
942 rq->end_io(rq, error);
944 blk_mq_free_request(rq);
947 EXPORT_SYMBOL(__blk_mq_end_request);
949 void blk_mq_end_request(struct request *rq, blk_status_t error)
951 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
953 __blk_mq_end_request(rq, error);
955 EXPORT_SYMBOL(blk_mq_end_request);
957 #define TAG_COMP_BATCH 32
959 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
960 int *tag_array, int nr_tags)
962 struct request_queue *q = hctx->queue;
965 * All requests should have been marked as RQF_MQ_INFLIGHT, so
966 * update hctx->nr_active in batch
968 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
969 __blk_mq_sub_active_requests(hctx, nr_tags);
971 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
972 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
975 void blk_mq_end_request_batch(struct io_comp_batch *iob)
977 int tags[TAG_COMP_BATCH], nr_tags = 0;
978 struct blk_mq_hw_ctx *cur_hctx = NULL;
983 now = ktime_get_ns();
985 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
987 prefetch(rq->rq_next);
989 blk_complete_request(rq);
991 __blk_mq_end_request_acct(rq, now);
993 rq_qos_done(rq->q, rq);
995 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
996 if (!req_ref_put_and_test(rq))
999 blk_crypto_free_request(rq);
1000 blk_pm_mark_last_busy(rq);
1002 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1004 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1006 cur_hctx = rq->mq_hctx;
1008 tags[nr_tags++] = rq->tag;
1012 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1014 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1016 static void blk_complete_reqs(struct llist_head *list)
1018 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1019 struct request *rq, *next;
1021 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1022 rq->q->mq_ops->complete(rq);
1025 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1027 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1030 static int blk_softirq_cpu_dead(unsigned int cpu)
1032 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1036 static void __blk_mq_complete_request_remote(void *data)
1038 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1041 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1043 int cpu = raw_smp_processor_id();
1045 if (!IS_ENABLED(CONFIG_SMP) ||
1046 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1049 * With force threaded interrupts enabled, raising softirq from an SMP
1050 * function call will always result in waking the ksoftirqd thread.
1051 * This is probably worse than completing the request on a different
1054 if (force_irqthreads())
1057 /* same CPU or cache domain? Complete locally */
1058 if (cpu == rq->mq_ctx->cpu ||
1059 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1060 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1063 /* don't try to IPI to an offline CPU */
1064 return cpu_online(rq->mq_ctx->cpu);
1067 static void blk_mq_complete_send_ipi(struct request *rq)
1069 struct llist_head *list;
1072 cpu = rq->mq_ctx->cpu;
1073 list = &per_cpu(blk_cpu_done, cpu);
1074 if (llist_add(&rq->ipi_list, list)) {
1075 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1076 smp_call_function_single_async(cpu, &rq->csd);
1080 static void blk_mq_raise_softirq(struct request *rq)
1082 struct llist_head *list;
1085 list = this_cpu_ptr(&blk_cpu_done);
1086 if (llist_add(&rq->ipi_list, list))
1087 raise_softirq(BLOCK_SOFTIRQ);
1091 bool blk_mq_complete_request_remote(struct request *rq)
1093 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1096 * For request which hctx has only one ctx mapping,
1097 * or a polled request, always complete locally,
1098 * it's pointless to redirect the completion.
1100 if (rq->mq_hctx->nr_ctx == 1 ||
1101 rq->cmd_flags & REQ_POLLED)
1104 if (blk_mq_complete_need_ipi(rq)) {
1105 blk_mq_complete_send_ipi(rq);
1109 if (rq->q->nr_hw_queues == 1) {
1110 blk_mq_raise_softirq(rq);
1115 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1118 * blk_mq_complete_request - end I/O on a request
1119 * @rq: the request being processed
1122 * Complete a request by scheduling the ->complete_rq operation.
1124 void blk_mq_complete_request(struct request *rq)
1126 if (!blk_mq_complete_request_remote(rq))
1127 rq->q->mq_ops->complete(rq);
1129 EXPORT_SYMBOL(blk_mq_complete_request);
1132 * blk_mq_start_request - Start processing a request
1133 * @rq: Pointer to request to be started
1135 * Function used by device drivers to notify the block layer that a request
1136 * is going to be processed now, so blk layer can do proper initializations
1137 * such as starting the timeout timer.
1139 void blk_mq_start_request(struct request *rq)
1141 struct request_queue *q = rq->q;
1143 trace_block_rq_issue(rq);
1145 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1146 rq->io_start_time_ns = ktime_get_ns();
1147 rq->stats_sectors = blk_rq_sectors(rq);
1148 rq->rq_flags |= RQF_STATS;
1149 rq_qos_issue(q, rq);
1152 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1155 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1157 #ifdef CONFIG_BLK_DEV_INTEGRITY
1158 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1159 q->integrity.profile->prepare_fn(rq);
1161 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1162 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1164 EXPORT_SYMBOL(blk_mq_start_request);
1167 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1168 * queues. This is important for md arrays to benefit from merging
1171 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1173 if (plug->multiple_queues)
1174 return BLK_MAX_REQUEST_COUNT * 2;
1175 return BLK_MAX_REQUEST_COUNT;
1178 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1180 struct request *last = rq_list_peek(&plug->mq_list);
1182 if (!plug->rq_count) {
1183 trace_block_plug(rq->q);
1184 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1185 (!blk_queue_nomerges(rq->q) &&
1186 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1187 blk_mq_flush_plug_list(plug, false);
1188 trace_block_plug(rq->q);
1191 if (!plug->multiple_queues && last && last->q != rq->q)
1192 plug->multiple_queues = true;
1193 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1194 plug->has_elevator = true;
1196 rq_list_add(&plug->mq_list, rq);
1201 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1202 * @rq: request to insert
1203 * @at_head: insert request at head or tail of queue
1206 * Insert a fully prepared request at the back of the I/O scheduler queue
1207 * for execution. Don't wait for completion.
1210 * This function will invoke @done directly if the queue is dead.
1212 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1214 WARN_ON(irqs_disabled());
1215 WARN_ON(!blk_rq_is_passthrough(rq));
1217 blk_account_io_start(rq);
1220 * As plugging can be enabled for passthrough requests on a zoned
1221 * device, directly accessing the plug instead of using blk_mq_plug()
1222 * should not have any consequences.
1225 blk_add_rq_to_plug(current->plug, rq);
1227 blk_mq_sched_insert_request(rq, at_head, true, false);
1229 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1231 struct blk_rq_wait {
1232 struct completion done;
1236 static void blk_end_sync_rq(struct request *rq, blk_status_t ret)
1238 struct blk_rq_wait *wait = rq->end_io_data;
1241 complete(&wait->done);
1244 bool blk_rq_is_poll(struct request *rq)
1248 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1250 if (WARN_ON_ONCE(!rq->bio))
1254 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1256 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1259 bio_poll(rq->bio, NULL, 0);
1261 } while (!completion_done(wait));
1265 * blk_execute_rq - insert a request into queue for execution
1266 * @rq: request to insert
1267 * @at_head: insert request at head or tail of queue
1270 * Insert a fully prepared request at the back of the I/O scheduler queue
1271 * for execution and wait for completion.
1272 * Return: The blk_status_t result provided to blk_mq_end_request().
1274 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1276 struct blk_rq_wait wait = {
1277 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1280 WARN_ON(irqs_disabled());
1281 WARN_ON(!blk_rq_is_passthrough(rq));
1283 rq->end_io_data = &wait;
1284 rq->end_io = blk_end_sync_rq;
1286 blk_account_io_start(rq);
1287 blk_mq_sched_insert_request(rq, at_head, true, false);
1289 if (blk_rq_is_poll(rq)) {
1290 blk_rq_poll_completion(rq, &wait.done);
1293 * Prevent hang_check timer from firing at us during very long
1296 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1299 while (!wait_for_completion_io_timeout(&wait.done,
1300 hang_check * (HZ/2)))
1303 wait_for_completion_io(&wait.done);
1308 EXPORT_SYMBOL(blk_execute_rq);
1310 static void __blk_mq_requeue_request(struct request *rq)
1312 struct request_queue *q = rq->q;
1314 blk_mq_put_driver_tag(rq);
1316 trace_block_rq_requeue(rq);
1317 rq_qos_requeue(q, rq);
1319 if (blk_mq_request_started(rq)) {
1320 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1321 rq->rq_flags &= ~RQF_TIMED_OUT;
1325 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1327 __blk_mq_requeue_request(rq);
1329 /* this request will be re-inserted to io scheduler queue */
1330 blk_mq_sched_requeue_request(rq);
1332 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1334 EXPORT_SYMBOL(blk_mq_requeue_request);
1336 static void blk_mq_requeue_work(struct work_struct *work)
1338 struct request_queue *q =
1339 container_of(work, struct request_queue, requeue_work.work);
1341 struct request *rq, *next;
1343 spin_lock_irq(&q->requeue_lock);
1344 list_splice_init(&q->requeue_list, &rq_list);
1345 spin_unlock_irq(&q->requeue_lock);
1347 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1348 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1351 rq->rq_flags &= ~RQF_SOFTBARRIER;
1352 list_del_init(&rq->queuelist);
1354 * If RQF_DONTPREP, rq has contained some driver specific
1355 * data, so insert it to hctx dispatch list to avoid any
1358 if (rq->rq_flags & RQF_DONTPREP)
1359 blk_mq_request_bypass_insert(rq, false, false);
1361 blk_mq_sched_insert_request(rq, true, false, false);
1364 while (!list_empty(&rq_list)) {
1365 rq = list_entry(rq_list.next, struct request, queuelist);
1366 list_del_init(&rq->queuelist);
1367 blk_mq_sched_insert_request(rq, false, false, false);
1370 blk_mq_run_hw_queues(q, false);
1373 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1374 bool kick_requeue_list)
1376 struct request_queue *q = rq->q;
1377 unsigned long flags;
1380 * We abuse this flag that is otherwise used by the I/O scheduler to
1381 * request head insertion from the workqueue.
1383 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1385 spin_lock_irqsave(&q->requeue_lock, flags);
1387 rq->rq_flags |= RQF_SOFTBARRIER;
1388 list_add(&rq->queuelist, &q->requeue_list);
1390 list_add_tail(&rq->queuelist, &q->requeue_list);
1392 spin_unlock_irqrestore(&q->requeue_lock, flags);
1394 if (kick_requeue_list)
1395 blk_mq_kick_requeue_list(q);
1398 void blk_mq_kick_requeue_list(struct request_queue *q)
1400 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1402 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1404 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1405 unsigned long msecs)
1407 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1408 msecs_to_jiffies(msecs));
1410 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1412 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1415 * If we find a request that isn't idle we know the queue is busy
1416 * as it's checked in the iter.
1417 * Return false to stop the iteration.
1419 if (blk_mq_request_started(rq)) {
1429 bool blk_mq_queue_inflight(struct request_queue *q)
1433 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1436 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1438 static void blk_mq_rq_timed_out(struct request *req)
1440 req->rq_flags |= RQF_TIMED_OUT;
1441 if (req->q->mq_ops->timeout) {
1442 enum blk_eh_timer_return ret;
1444 ret = req->q->mq_ops->timeout(req);
1445 if (ret == BLK_EH_DONE)
1447 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1453 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
1455 unsigned long deadline;
1457 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1459 if (rq->rq_flags & RQF_TIMED_OUT)
1462 deadline = READ_ONCE(rq->deadline);
1463 if (time_after_eq(jiffies, deadline))
1468 else if (time_after(*next, deadline))
1473 void blk_mq_put_rq_ref(struct request *rq)
1475 if (is_flush_rq(rq))
1477 else if (req_ref_put_and_test(rq))
1478 __blk_mq_free_request(rq);
1481 static bool blk_mq_check_expired(struct request *rq, void *priv)
1483 unsigned long *next = priv;
1486 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1487 * be reallocated underneath the timeout handler's processing, then
1488 * the expire check is reliable. If the request is not expired, then
1489 * it was completed and reallocated as a new request after returning
1490 * from blk_mq_check_expired().
1492 if (blk_mq_req_expired(rq, next))
1493 blk_mq_rq_timed_out(rq);
1497 static void blk_mq_timeout_work(struct work_struct *work)
1499 struct request_queue *q =
1500 container_of(work, struct request_queue, timeout_work);
1501 unsigned long next = 0;
1502 struct blk_mq_hw_ctx *hctx;
1505 /* A deadlock might occur if a request is stuck requiring a
1506 * timeout at the same time a queue freeze is waiting
1507 * completion, since the timeout code would not be able to
1508 * acquire the queue reference here.
1510 * That's why we don't use blk_queue_enter here; instead, we use
1511 * percpu_ref_tryget directly, because we need to be able to
1512 * obtain a reference even in the short window between the queue
1513 * starting to freeze, by dropping the first reference in
1514 * blk_freeze_queue_start, and the moment the last request is
1515 * consumed, marked by the instant q_usage_counter reaches
1518 if (!percpu_ref_tryget(&q->q_usage_counter))
1521 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
1524 mod_timer(&q->timeout, next);
1527 * Request timeouts are handled as a forward rolling timer. If
1528 * we end up here it means that no requests are pending and
1529 * also that no request has been pending for a while. Mark
1530 * each hctx as idle.
1532 queue_for_each_hw_ctx(q, hctx, i) {
1533 /* the hctx may be unmapped, so check it here */
1534 if (blk_mq_hw_queue_mapped(hctx))
1535 blk_mq_tag_idle(hctx);
1541 struct flush_busy_ctx_data {
1542 struct blk_mq_hw_ctx *hctx;
1543 struct list_head *list;
1546 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1548 struct flush_busy_ctx_data *flush_data = data;
1549 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1550 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1551 enum hctx_type type = hctx->type;
1553 spin_lock(&ctx->lock);
1554 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1555 sbitmap_clear_bit(sb, bitnr);
1556 spin_unlock(&ctx->lock);
1561 * Process software queues that have been marked busy, splicing them
1562 * to the for-dispatch
1564 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1566 struct flush_busy_ctx_data data = {
1571 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1573 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1575 struct dispatch_rq_data {
1576 struct blk_mq_hw_ctx *hctx;
1580 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1583 struct dispatch_rq_data *dispatch_data = data;
1584 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1585 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1586 enum hctx_type type = hctx->type;
1588 spin_lock(&ctx->lock);
1589 if (!list_empty(&ctx->rq_lists[type])) {
1590 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1591 list_del_init(&dispatch_data->rq->queuelist);
1592 if (list_empty(&ctx->rq_lists[type]))
1593 sbitmap_clear_bit(sb, bitnr);
1595 spin_unlock(&ctx->lock);
1597 return !dispatch_data->rq;
1600 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1601 struct blk_mq_ctx *start)
1603 unsigned off = start ? start->index_hw[hctx->type] : 0;
1604 struct dispatch_rq_data data = {
1609 __sbitmap_for_each_set(&hctx->ctx_map, off,
1610 dispatch_rq_from_ctx, &data);
1615 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1617 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1618 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1621 blk_mq_tag_busy(rq->mq_hctx);
1623 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1624 bt = &rq->mq_hctx->tags->breserved_tags;
1627 if (!hctx_may_queue(rq->mq_hctx, bt))
1631 tag = __sbitmap_queue_get(bt);
1632 if (tag == BLK_MQ_NO_TAG)
1635 rq->tag = tag + tag_offset;
1639 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1641 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1644 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1645 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1646 rq->rq_flags |= RQF_MQ_INFLIGHT;
1647 __blk_mq_inc_active_requests(hctx);
1649 hctx->tags->rqs[rq->tag] = rq;
1653 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1654 int flags, void *key)
1656 struct blk_mq_hw_ctx *hctx;
1658 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1660 spin_lock(&hctx->dispatch_wait_lock);
1661 if (!list_empty(&wait->entry)) {
1662 struct sbitmap_queue *sbq;
1664 list_del_init(&wait->entry);
1665 sbq = &hctx->tags->bitmap_tags;
1666 atomic_dec(&sbq->ws_active);
1668 spin_unlock(&hctx->dispatch_wait_lock);
1670 blk_mq_run_hw_queue(hctx, true);
1675 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1676 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1677 * restart. For both cases, take care to check the condition again after
1678 * marking us as waiting.
1680 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1683 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1684 struct wait_queue_head *wq;
1685 wait_queue_entry_t *wait;
1688 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1689 blk_mq_sched_mark_restart_hctx(hctx);
1692 * It's possible that a tag was freed in the window between the
1693 * allocation failure and adding the hardware queue to the wait
1696 * Don't clear RESTART here, someone else could have set it.
1697 * At most this will cost an extra queue run.
1699 return blk_mq_get_driver_tag(rq);
1702 wait = &hctx->dispatch_wait;
1703 if (!list_empty_careful(&wait->entry))
1706 wq = &bt_wait_ptr(sbq, hctx)->wait;
1708 spin_lock_irq(&wq->lock);
1709 spin_lock(&hctx->dispatch_wait_lock);
1710 if (!list_empty(&wait->entry)) {
1711 spin_unlock(&hctx->dispatch_wait_lock);
1712 spin_unlock_irq(&wq->lock);
1716 atomic_inc(&sbq->ws_active);
1717 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1718 __add_wait_queue(wq, wait);
1721 * It's possible that a tag was freed in the window between the
1722 * allocation failure and adding the hardware queue to the wait
1725 ret = blk_mq_get_driver_tag(rq);
1727 spin_unlock(&hctx->dispatch_wait_lock);
1728 spin_unlock_irq(&wq->lock);
1733 * We got a tag, remove ourselves from the wait queue to ensure
1734 * someone else gets the wakeup.
1736 list_del_init(&wait->entry);
1737 atomic_dec(&sbq->ws_active);
1738 spin_unlock(&hctx->dispatch_wait_lock);
1739 spin_unlock_irq(&wq->lock);
1744 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1745 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1747 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1748 * - EWMA is one simple way to compute running average value
1749 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1750 * - take 4 as factor for avoiding to get too small(0) result, and this
1751 * factor doesn't matter because EWMA decreases exponentially
1753 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1757 ewma = hctx->dispatch_busy;
1762 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1764 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1765 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1767 hctx->dispatch_busy = ewma;
1770 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1772 static void blk_mq_handle_dev_resource(struct request *rq,
1773 struct list_head *list)
1775 struct request *next =
1776 list_first_entry_or_null(list, struct request, queuelist);
1779 * If an I/O scheduler has been configured and we got a driver tag for
1780 * the next request already, free it.
1783 blk_mq_put_driver_tag(next);
1785 list_add(&rq->queuelist, list);
1786 __blk_mq_requeue_request(rq);
1789 static void blk_mq_handle_zone_resource(struct request *rq,
1790 struct list_head *zone_list)
1793 * If we end up here it is because we cannot dispatch a request to a
1794 * specific zone due to LLD level zone-write locking or other zone
1795 * related resource not being available. In this case, set the request
1796 * aside in zone_list for retrying it later.
1798 list_add(&rq->queuelist, zone_list);
1799 __blk_mq_requeue_request(rq);
1802 enum prep_dispatch {
1804 PREP_DISPATCH_NO_TAG,
1805 PREP_DISPATCH_NO_BUDGET,
1808 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1811 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1812 int budget_token = -1;
1815 budget_token = blk_mq_get_dispatch_budget(rq->q);
1816 if (budget_token < 0) {
1817 blk_mq_put_driver_tag(rq);
1818 return PREP_DISPATCH_NO_BUDGET;
1820 blk_mq_set_rq_budget_token(rq, budget_token);
1823 if (!blk_mq_get_driver_tag(rq)) {
1825 * The initial allocation attempt failed, so we need to
1826 * rerun the hardware queue when a tag is freed. The
1827 * waitqueue takes care of that. If the queue is run
1828 * before we add this entry back on the dispatch list,
1829 * we'll re-run it below.
1831 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1833 * All budgets not got from this function will be put
1834 * together during handling partial dispatch
1837 blk_mq_put_dispatch_budget(rq->q, budget_token);
1838 return PREP_DISPATCH_NO_TAG;
1842 return PREP_DISPATCH_OK;
1845 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1846 static void blk_mq_release_budgets(struct request_queue *q,
1847 struct list_head *list)
1851 list_for_each_entry(rq, list, queuelist) {
1852 int budget_token = blk_mq_get_rq_budget_token(rq);
1854 if (budget_token >= 0)
1855 blk_mq_put_dispatch_budget(q, budget_token);
1860 * Returns true if we did some work AND can potentially do more.
1862 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1863 unsigned int nr_budgets)
1865 enum prep_dispatch prep;
1866 struct request_queue *q = hctx->queue;
1867 struct request *rq, *nxt;
1869 blk_status_t ret = BLK_STS_OK;
1870 LIST_HEAD(zone_list);
1871 bool needs_resource = false;
1873 if (list_empty(list))
1877 * Now process all the entries, sending them to the driver.
1879 errors = queued = 0;
1881 struct blk_mq_queue_data bd;
1883 rq = list_first_entry(list, struct request, queuelist);
1885 WARN_ON_ONCE(hctx != rq->mq_hctx);
1886 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1887 if (prep != PREP_DISPATCH_OK)
1890 list_del_init(&rq->queuelist);
1895 * Flag last if we have no more requests, or if we have more
1896 * but can't assign a driver tag to it.
1898 if (list_empty(list))
1901 nxt = list_first_entry(list, struct request, queuelist);
1902 bd.last = !blk_mq_get_driver_tag(nxt);
1906 * once the request is queued to lld, no need to cover the
1911 ret = q->mq_ops->queue_rq(hctx, &bd);
1916 case BLK_STS_RESOURCE:
1917 needs_resource = true;
1919 case BLK_STS_DEV_RESOURCE:
1920 blk_mq_handle_dev_resource(rq, list);
1922 case BLK_STS_ZONE_RESOURCE:
1924 * Move the request to zone_list and keep going through
1925 * the dispatch list to find more requests the drive can
1928 blk_mq_handle_zone_resource(rq, &zone_list);
1929 needs_resource = true;
1933 blk_mq_end_request(rq, ret);
1935 } while (!list_empty(list));
1937 if (!list_empty(&zone_list))
1938 list_splice_tail_init(&zone_list, list);
1940 /* If we didn't flush the entire list, we could have told the driver
1941 * there was more coming, but that turned out to be a lie.
1943 if ((!list_empty(list) || errors || needs_resource ||
1944 ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued)
1945 q->mq_ops->commit_rqs(hctx);
1947 * Any items that need requeuing? Stuff them into hctx->dispatch,
1948 * that is where we will continue on next queue run.
1950 if (!list_empty(list)) {
1952 /* For non-shared tags, the RESTART check will suffice */
1953 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1954 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1957 blk_mq_release_budgets(q, list);
1959 spin_lock(&hctx->lock);
1960 list_splice_tail_init(list, &hctx->dispatch);
1961 spin_unlock(&hctx->lock);
1964 * Order adding requests to hctx->dispatch and checking
1965 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1966 * in blk_mq_sched_restart(). Avoid restart code path to
1967 * miss the new added requests to hctx->dispatch, meantime
1968 * SCHED_RESTART is observed here.
1973 * If SCHED_RESTART was set by the caller of this function and
1974 * it is no longer set that means that it was cleared by another
1975 * thread and hence that a queue rerun is needed.
1977 * If 'no_tag' is set, that means that we failed getting
1978 * a driver tag with an I/O scheduler attached. If our dispatch
1979 * waitqueue is no longer active, ensure that we run the queue
1980 * AFTER adding our entries back to the list.
1982 * If no I/O scheduler has been configured it is possible that
1983 * the hardware queue got stopped and restarted before requests
1984 * were pushed back onto the dispatch list. Rerun the queue to
1985 * avoid starvation. Notes:
1986 * - blk_mq_run_hw_queue() checks whether or not a queue has
1987 * been stopped before rerunning a queue.
1988 * - Some but not all block drivers stop a queue before
1989 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1992 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1993 * bit is set, run queue after a delay to avoid IO stalls
1994 * that could otherwise occur if the queue is idle. We'll do
1995 * similar if we couldn't get budget or couldn't lock a zone
1996 * and SCHED_RESTART is set.
1998 needs_restart = blk_mq_sched_needs_restart(hctx);
1999 if (prep == PREP_DISPATCH_NO_BUDGET)
2000 needs_resource = true;
2001 if (!needs_restart ||
2002 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2003 blk_mq_run_hw_queue(hctx, true);
2004 else if (needs_resource)
2005 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2007 blk_mq_update_dispatch_busy(hctx, true);
2010 blk_mq_update_dispatch_busy(hctx, false);
2012 return (queued + errors) != 0;
2016 * __blk_mq_run_hw_queue - Run a hardware queue.
2017 * @hctx: Pointer to the hardware queue to run.
2019 * Send pending requests to the hardware.
2021 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
2024 * We can't run the queue inline with ints disabled. Ensure that
2025 * we catch bad users of this early.
2027 WARN_ON_ONCE(in_interrupt());
2029 blk_mq_run_dispatch_ops(hctx->queue,
2030 blk_mq_sched_dispatch_requests(hctx));
2033 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2035 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2037 if (cpu >= nr_cpu_ids)
2038 cpu = cpumask_first(hctx->cpumask);
2043 * It'd be great if the workqueue API had a way to pass
2044 * in a mask and had some smarts for more clever placement.
2045 * For now we just round-robin here, switching for every
2046 * BLK_MQ_CPU_WORK_BATCH queued items.
2048 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2051 int next_cpu = hctx->next_cpu;
2053 if (hctx->queue->nr_hw_queues == 1)
2054 return WORK_CPU_UNBOUND;
2056 if (--hctx->next_cpu_batch <= 0) {
2058 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2060 if (next_cpu >= nr_cpu_ids)
2061 next_cpu = blk_mq_first_mapped_cpu(hctx);
2062 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2066 * Do unbound schedule if we can't find a online CPU for this hctx,
2067 * and it should only happen in the path of handling CPU DEAD.
2069 if (!cpu_online(next_cpu)) {
2076 * Make sure to re-select CPU next time once after CPUs
2077 * in hctx->cpumask become online again.
2079 hctx->next_cpu = next_cpu;
2080 hctx->next_cpu_batch = 1;
2081 return WORK_CPU_UNBOUND;
2084 hctx->next_cpu = next_cpu;
2089 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2090 * @hctx: Pointer to the hardware queue to run.
2091 * @async: If we want to run the queue asynchronously.
2092 * @msecs: Milliseconds of delay to wait before running the queue.
2094 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2095 * with a delay of @msecs.
2097 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2098 unsigned long msecs)
2100 if (unlikely(blk_mq_hctx_stopped(hctx)))
2103 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2104 if (cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2105 __blk_mq_run_hw_queue(hctx);
2110 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2111 msecs_to_jiffies(msecs));
2115 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2116 * @hctx: Pointer to the hardware queue to run.
2117 * @msecs: Milliseconds of delay to wait before running the queue.
2119 * Run a hardware queue asynchronously with a delay of @msecs.
2121 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2123 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2125 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2128 * blk_mq_run_hw_queue - Start to run a hardware queue.
2129 * @hctx: Pointer to the hardware queue to run.
2130 * @async: If we want to run the queue asynchronously.
2132 * Check if the request queue is not in a quiesced state and if there are
2133 * pending requests to be sent. If this is true, run the queue to send requests
2136 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2141 * When queue is quiesced, we may be switching io scheduler, or
2142 * updating nr_hw_queues, or other things, and we can't run queue
2143 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2145 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2148 __blk_mq_run_dispatch_ops(hctx->queue, false,
2149 need_run = !blk_queue_quiesced(hctx->queue) &&
2150 blk_mq_hctx_has_pending(hctx));
2153 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2155 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2158 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2161 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2163 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2165 * If the IO scheduler does not respect hardware queues when
2166 * dispatching, we just don't bother with multiple HW queues and
2167 * dispatch from hctx for the current CPU since running multiple queues
2168 * just causes lock contention inside the scheduler and pointless cache
2171 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2173 if (!blk_mq_hctx_stopped(hctx))
2179 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2180 * @q: Pointer to the request queue to run.
2181 * @async: If we want to run the queue asynchronously.
2183 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2185 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2189 if (blk_queue_sq_sched(q))
2190 sq_hctx = blk_mq_get_sq_hctx(q);
2191 queue_for_each_hw_ctx(q, hctx, i) {
2192 if (blk_mq_hctx_stopped(hctx))
2195 * Dispatch from this hctx either if there's no hctx preferred
2196 * by IO scheduler or if it has requests that bypass the
2199 if (!sq_hctx || sq_hctx == hctx ||
2200 !list_empty_careful(&hctx->dispatch))
2201 blk_mq_run_hw_queue(hctx, async);
2204 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2207 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2208 * @q: Pointer to the request queue to run.
2209 * @msecs: Milliseconds of delay to wait before running the queues.
2211 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2213 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2217 if (blk_queue_sq_sched(q))
2218 sq_hctx = blk_mq_get_sq_hctx(q);
2219 queue_for_each_hw_ctx(q, hctx, i) {
2220 if (blk_mq_hctx_stopped(hctx))
2223 * If there is already a run_work pending, leave the
2224 * pending delay untouched. Otherwise, a hctx can stall
2225 * if another hctx is re-delaying the other's work
2226 * before the work executes.
2228 if (delayed_work_pending(&hctx->run_work))
2231 * Dispatch from this hctx either if there's no hctx preferred
2232 * by IO scheduler or if it has requests that bypass the
2235 if (!sq_hctx || sq_hctx == hctx ||
2236 !list_empty_careful(&hctx->dispatch))
2237 blk_mq_delay_run_hw_queue(hctx, msecs);
2240 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2243 * This function is often used for pausing .queue_rq() by driver when
2244 * there isn't enough resource or some conditions aren't satisfied, and
2245 * BLK_STS_RESOURCE is usually returned.
2247 * We do not guarantee that dispatch can be drained or blocked
2248 * after blk_mq_stop_hw_queue() returns. Please use
2249 * blk_mq_quiesce_queue() for that requirement.
2251 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2253 cancel_delayed_work(&hctx->run_work);
2255 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2257 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2260 * This function is often used for pausing .queue_rq() by driver when
2261 * there isn't enough resource or some conditions aren't satisfied, and
2262 * BLK_STS_RESOURCE is usually returned.
2264 * We do not guarantee that dispatch can be drained or blocked
2265 * after blk_mq_stop_hw_queues() returns. Please use
2266 * blk_mq_quiesce_queue() for that requirement.
2268 void blk_mq_stop_hw_queues(struct request_queue *q)
2270 struct blk_mq_hw_ctx *hctx;
2273 queue_for_each_hw_ctx(q, hctx, i)
2274 blk_mq_stop_hw_queue(hctx);
2276 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2278 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2280 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2282 blk_mq_run_hw_queue(hctx, false);
2284 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2286 void blk_mq_start_hw_queues(struct request_queue *q)
2288 struct blk_mq_hw_ctx *hctx;
2291 queue_for_each_hw_ctx(q, hctx, i)
2292 blk_mq_start_hw_queue(hctx);
2294 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2296 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2298 if (!blk_mq_hctx_stopped(hctx))
2301 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2302 blk_mq_run_hw_queue(hctx, async);
2304 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2306 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2308 struct blk_mq_hw_ctx *hctx;
2311 queue_for_each_hw_ctx(q, hctx, i)
2312 blk_mq_start_stopped_hw_queue(hctx, async);
2314 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2316 static void blk_mq_run_work_fn(struct work_struct *work)
2318 struct blk_mq_hw_ctx *hctx;
2320 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2323 * If we are stopped, don't run the queue.
2325 if (blk_mq_hctx_stopped(hctx))
2328 __blk_mq_run_hw_queue(hctx);
2331 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2335 struct blk_mq_ctx *ctx = rq->mq_ctx;
2336 enum hctx_type type = hctx->type;
2338 lockdep_assert_held(&ctx->lock);
2340 trace_block_rq_insert(rq);
2343 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2345 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2348 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2351 struct blk_mq_ctx *ctx = rq->mq_ctx;
2353 lockdep_assert_held(&ctx->lock);
2355 __blk_mq_insert_req_list(hctx, rq, at_head);
2356 blk_mq_hctx_mark_pending(hctx, ctx);
2360 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2361 * @rq: Pointer to request to be inserted.
2362 * @at_head: true if the request should be inserted at the head of the list.
2363 * @run_queue: If we should run the hardware queue after inserting the request.
2365 * Should only be used carefully, when the caller knows we want to
2366 * bypass a potential IO scheduler on the target device.
2368 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2371 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2373 spin_lock(&hctx->lock);
2375 list_add(&rq->queuelist, &hctx->dispatch);
2377 list_add_tail(&rq->queuelist, &hctx->dispatch);
2378 spin_unlock(&hctx->lock);
2381 blk_mq_run_hw_queue(hctx, false);
2384 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2385 struct list_head *list)
2389 enum hctx_type type = hctx->type;
2392 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2395 list_for_each_entry(rq, list, queuelist) {
2396 BUG_ON(rq->mq_ctx != ctx);
2397 trace_block_rq_insert(rq);
2400 spin_lock(&ctx->lock);
2401 list_splice_tail_init(list, &ctx->rq_lists[type]);
2402 blk_mq_hctx_mark_pending(hctx, ctx);
2403 spin_unlock(&ctx->lock);
2406 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2409 if (hctx->queue->mq_ops->commit_rqs) {
2410 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2411 hctx->queue->mq_ops->commit_rqs(hctx);
2416 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2417 unsigned int nr_segs)
2421 if (bio->bi_opf & REQ_RAHEAD)
2422 rq->cmd_flags |= REQ_FAILFAST_MASK;
2424 rq->__sector = bio->bi_iter.bi_sector;
2425 blk_rq_bio_prep(rq, bio, nr_segs);
2427 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2428 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2431 blk_account_io_start(rq);
2434 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2435 struct request *rq, bool last)
2437 struct request_queue *q = rq->q;
2438 struct blk_mq_queue_data bd = {
2445 * For OK queue, we are done. For error, caller may kill it.
2446 * Any other error (busy), just add it to our list as we
2447 * previously would have done.
2449 ret = q->mq_ops->queue_rq(hctx, &bd);
2452 blk_mq_update_dispatch_busy(hctx, false);
2454 case BLK_STS_RESOURCE:
2455 case BLK_STS_DEV_RESOURCE:
2456 blk_mq_update_dispatch_busy(hctx, true);
2457 __blk_mq_requeue_request(rq);
2460 blk_mq_update_dispatch_busy(hctx, false);
2467 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2469 bool bypass_insert, bool last)
2471 struct request_queue *q = rq->q;
2472 bool run_queue = true;
2476 * RCU or SRCU read lock is needed before checking quiesced flag.
2478 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2479 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2480 * and avoid driver to try to dispatch again.
2482 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2484 bypass_insert = false;
2488 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2491 budget_token = blk_mq_get_dispatch_budget(q);
2492 if (budget_token < 0)
2495 blk_mq_set_rq_budget_token(rq, budget_token);
2497 if (!blk_mq_get_driver_tag(rq)) {
2498 blk_mq_put_dispatch_budget(q, budget_token);
2502 return __blk_mq_issue_directly(hctx, rq, last);
2505 return BLK_STS_RESOURCE;
2507 blk_mq_sched_insert_request(rq, false, run_queue, false);
2513 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2514 * @hctx: Pointer of the associated hardware queue.
2515 * @rq: Pointer to request to be sent.
2517 * If the device has enough resources to accept a new request now, send the
2518 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2519 * we can try send it another time in the future. Requests inserted at this
2520 * queue have higher priority.
2522 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2526 __blk_mq_try_issue_directly(hctx, rq, false, true);
2528 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2529 blk_mq_request_bypass_insert(rq, false, true);
2530 else if (ret != BLK_STS_OK)
2531 blk_mq_end_request(rq, ret);
2534 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2536 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2539 static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2541 struct blk_mq_hw_ctx *hctx = NULL;
2546 while ((rq = rq_list_pop(&plug->mq_list))) {
2547 bool last = rq_list_empty(plug->mq_list);
2550 if (hctx != rq->mq_hctx) {
2552 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2556 ret = blk_mq_request_issue_directly(rq, last);
2561 case BLK_STS_RESOURCE:
2562 case BLK_STS_DEV_RESOURCE:
2563 blk_mq_request_bypass_insert(rq, false, true);
2564 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2567 blk_mq_end_request(rq, ret);
2574 * If we didn't flush the entire list, we could have told the driver
2575 * there was more coming, but that turned out to be a lie.
2578 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2581 static void __blk_mq_flush_plug_list(struct request_queue *q,
2582 struct blk_plug *plug)
2584 if (blk_queue_quiesced(q))
2586 q->mq_ops->queue_rqs(&plug->mq_list);
2589 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2591 struct blk_mq_hw_ctx *this_hctx = NULL;
2592 struct blk_mq_ctx *this_ctx = NULL;
2593 struct request *requeue_list = NULL;
2594 unsigned int depth = 0;
2598 struct request *rq = rq_list_pop(&plug->mq_list);
2601 this_hctx = rq->mq_hctx;
2602 this_ctx = rq->mq_ctx;
2603 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2604 rq_list_add(&requeue_list, rq);
2607 list_add_tail(&rq->queuelist, &list);
2609 } while (!rq_list_empty(plug->mq_list));
2611 plug->mq_list = requeue_list;
2612 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2613 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched);
2616 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2620 if (rq_list_empty(plug->mq_list))
2624 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2625 struct request_queue *q;
2627 rq = rq_list_peek(&plug->mq_list);
2631 * Peek first request and see if we have a ->queue_rqs() hook.
2632 * If we do, we can dispatch the whole plug list in one go. We
2633 * already know at this point that all requests belong to the
2634 * same queue, caller must ensure that's the case.
2636 * Since we pass off the full list to the driver at this point,
2637 * we do not increment the active request count for the queue.
2638 * Bypass shared tags for now because of that.
2640 if (q->mq_ops->queue_rqs &&
2641 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2642 blk_mq_run_dispatch_ops(q,
2643 __blk_mq_flush_plug_list(q, plug));
2644 if (rq_list_empty(plug->mq_list))
2648 blk_mq_run_dispatch_ops(q,
2649 blk_mq_plug_issue_direct(plug, false));
2650 if (rq_list_empty(plug->mq_list))
2655 blk_mq_dispatch_plug_list(plug, from_schedule);
2656 } while (!rq_list_empty(plug->mq_list));
2659 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2660 struct list_head *list)
2665 while (!list_empty(list)) {
2667 struct request *rq = list_first_entry(list, struct request,
2670 list_del_init(&rq->queuelist);
2671 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2672 if (ret != BLK_STS_OK) {
2674 if (ret == BLK_STS_RESOURCE ||
2675 ret == BLK_STS_DEV_RESOURCE) {
2676 blk_mq_request_bypass_insert(rq, false,
2680 blk_mq_end_request(rq, ret);
2686 * If we didn't flush the entire list, we could have told
2687 * the driver there was more coming, but that turned out to
2690 if ((!list_empty(list) || errors) &&
2691 hctx->queue->mq_ops->commit_rqs && queued)
2692 hctx->queue->mq_ops->commit_rqs(hctx);
2695 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2696 struct bio *bio, unsigned int nr_segs)
2698 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2699 if (blk_attempt_plug_merge(q, bio, nr_segs))
2701 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2707 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2708 struct blk_plug *plug,
2712 struct blk_mq_alloc_data data = {
2715 .cmd_flags = bio->bi_opf,
2719 if (unlikely(bio_queue_enter(bio)))
2722 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2725 rq_qos_throttle(q, bio);
2728 data.nr_tags = plug->nr_ios;
2730 data.cached_rq = &plug->cached_rq;
2733 rq = __blk_mq_alloc_requests(&data);
2736 rq_qos_cleanup(q, bio);
2737 if (bio->bi_opf & REQ_NOWAIT)
2738 bio_wouldblock_error(bio);
2744 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2745 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2751 rq = rq_list_peek(&plug->cached_rq);
2752 if (!rq || rq->q != q)
2755 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2760 if (blk_mq_get_hctx_type((*bio)->bi_opf) != rq->mq_hctx->type)
2762 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2766 * If any qos ->throttle() end up blocking, we will have flushed the
2767 * plug and hence killed the cached_rq list as well. Pop this entry
2768 * before we throttle.
2770 plug->cached_rq = rq_list_next(rq);
2771 rq_qos_throttle(q, *bio);
2773 rq->cmd_flags = (*bio)->bi_opf;
2774 INIT_LIST_HEAD(&rq->queuelist);
2778 static void bio_set_ioprio(struct bio *bio)
2780 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2781 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2782 bio->bi_ioprio = get_current_ioprio();
2783 blkcg_set_ioprio(bio);
2787 * blk_mq_submit_bio - Create and send a request to block device.
2788 * @bio: Bio pointer.
2790 * Builds up a request structure from @q and @bio and send to the device. The
2791 * request may not be queued directly to hardware if:
2792 * * This request can be merged with another one
2793 * * We want to place request at plug queue for possible future merging
2794 * * There is an IO scheduler active at this queue
2796 * It will not queue the request if there is an error with the bio, or at the
2799 void blk_mq_submit_bio(struct bio *bio)
2801 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2802 struct blk_plug *plug = blk_mq_plug(bio);
2803 const int is_sync = op_is_sync(bio->bi_opf);
2805 unsigned int nr_segs = 1;
2808 bio = blk_queue_bounce(bio, q);
2809 if (bio_may_exceed_limits(bio, &q->limits))
2810 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2812 if (!bio_integrity_prep(bio))
2815 bio_set_ioprio(bio);
2817 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2821 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2826 trace_block_getrq(bio);
2828 rq_qos_track(q, rq, bio);
2830 blk_mq_bio_to_request(rq, bio, nr_segs);
2832 ret = blk_crypto_init_request(rq);
2833 if (ret != BLK_STS_OK) {
2834 bio->bi_status = ret;
2836 blk_mq_free_request(rq);
2840 if (op_is_flush(bio->bi_opf)) {
2841 blk_insert_flush(rq);
2846 blk_add_rq_to_plug(plug, rq);
2847 else if ((rq->rq_flags & RQF_ELV) ||
2848 (rq->mq_hctx->dispatch_busy &&
2849 (q->nr_hw_queues == 1 || !is_sync)))
2850 blk_mq_sched_insert_request(rq, false, true, true);
2852 blk_mq_run_dispatch_ops(rq->q,
2853 blk_mq_try_issue_directly(rq->mq_hctx, rq));
2856 #ifdef CONFIG_BLK_MQ_STACKING
2858 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
2859 * @rq: the request being queued
2861 blk_status_t blk_insert_cloned_request(struct request *rq)
2863 struct request_queue *q = rq->q;
2864 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
2867 if (blk_rq_sectors(rq) > max_sectors) {
2869 * SCSI device does not have a good way to return if
2870 * Write Same/Zero is actually supported. If a device rejects
2871 * a non-read/write command (discard, write same,etc.) the
2872 * low-level device driver will set the relevant queue limit to
2873 * 0 to prevent blk-lib from issuing more of the offending
2874 * operations. Commands queued prior to the queue limit being
2875 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
2876 * errors being propagated to upper layers.
2878 if (max_sectors == 0)
2879 return BLK_STS_NOTSUPP;
2881 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
2882 __func__, blk_rq_sectors(rq), max_sectors);
2883 return BLK_STS_IOERR;
2887 * The queue settings related to segment counting may differ from the
2890 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
2891 if (rq->nr_phys_segments > queue_max_segments(q)) {
2892 printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n",
2893 __func__, rq->nr_phys_segments, queue_max_segments(q));
2894 return BLK_STS_IOERR;
2897 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
2898 return BLK_STS_IOERR;
2900 if (blk_crypto_insert_cloned_request(rq))
2901 return BLK_STS_IOERR;
2903 blk_account_io_start(rq);
2906 * Since we have a scheduler attached on the top device,
2907 * bypass a potential scheduler on the bottom device for
2910 blk_mq_run_dispatch_ops(q,
2911 ret = blk_mq_request_issue_directly(rq, true));
2913 blk_account_io_done(rq, ktime_get_ns());
2916 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
2919 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
2920 * @rq: the clone request to be cleaned up
2923 * Free all bios in @rq for a cloned request.
2925 void blk_rq_unprep_clone(struct request *rq)
2929 while ((bio = rq->bio) != NULL) {
2930 rq->bio = bio->bi_next;
2935 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
2938 * blk_rq_prep_clone - Helper function to setup clone request
2939 * @rq: the request to be setup
2940 * @rq_src: original request to be cloned
2941 * @bs: bio_set that bios for clone are allocated from
2942 * @gfp_mask: memory allocation mask for bio
2943 * @bio_ctr: setup function to be called for each clone bio.
2944 * Returns %0 for success, non %0 for failure.
2945 * @data: private data to be passed to @bio_ctr
2948 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
2949 * Also, pages which the original bios are pointing to are not copied
2950 * and the cloned bios just point same pages.
2951 * So cloned bios must be completed before original bios, which means
2952 * the caller must complete @rq before @rq_src.
2954 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
2955 struct bio_set *bs, gfp_t gfp_mask,
2956 int (*bio_ctr)(struct bio *, struct bio *, void *),
2959 struct bio *bio, *bio_src;
2964 __rq_for_each_bio(bio_src, rq_src) {
2965 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
2970 if (bio_ctr && bio_ctr(bio, bio_src, data))
2974 rq->biotail->bi_next = bio;
2977 rq->bio = rq->biotail = bio;
2982 /* Copy attributes of the original request to the clone request. */
2983 rq->__sector = blk_rq_pos(rq_src);
2984 rq->__data_len = blk_rq_bytes(rq_src);
2985 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
2986 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
2987 rq->special_vec = rq_src->special_vec;
2989 rq->nr_phys_segments = rq_src->nr_phys_segments;
2990 rq->ioprio = rq_src->ioprio;
2992 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3000 blk_rq_unprep_clone(rq);
3004 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3005 #endif /* CONFIG_BLK_MQ_STACKING */
3008 * Steal bios from a request and add them to a bio list.
3009 * The request must not have been partially completed before.
3011 void blk_steal_bios(struct bio_list *list, struct request *rq)
3015 list->tail->bi_next = rq->bio;
3017 list->head = rq->bio;
3018 list->tail = rq->biotail;
3026 EXPORT_SYMBOL_GPL(blk_steal_bios);
3028 static size_t order_to_size(unsigned int order)
3030 return (size_t)PAGE_SIZE << order;
3033 /* called before freeing request pool in @tags */
3034 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3035 struct blk_mq_tags *tags)
3038 unsigned long flags;
3040 /* There is no need to clear a driver tags own mapping */
3041 if (drv_tags == tags)
3044 list_for_each_entry(page, &tags->page_list, lru) {
3045 unsigned long start = (unsigned long)page_address(page);
3046 unsigned long end = start + order_to_size(page->private);
3049 for (i = 0; i < drv_tags->nr_tags; i++) {
3050 struct request *rq = drv_tags->rqs[i];
3051 unsigned long rq_addr = (unsigned long)rq;
3053 if (rq_addr >= start && rq_addr < end) {
3054 WARN_ON_ONCE(req_ref_read(rq) != 0);
3055 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3061 * Wait until all pending iteration is done.
3063 * Request reference is cleared and it is guaranteed to be observed
3064 * after the ->lock is released.
3066 spin_lock_irqsave(&drv_tags->lock, flags);
3067 spin_unlock_irqrestore(&drv_tags->lock, flags);
3070 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3071 unsigned int hctx_idx)
3073 struct blk_mq_tags *drv_tags;
3076 if (list_empty(&tags->page_list))
3079 if (blk_mq_is_shared_tags(set->flags))
3080 drv_tags = set->shared_tags;
3082 drv_tags = set->tags[hctx_idx];
3084 if (tags->static_rqs && set->ops->exit_request) {
3087 for (i = 0; i < tags->nr_tags; i++) {
3088 struct request *rq = tags->static_rqs[i];
3092 set->ops->exit_request(set, rq, hctx_idx);
3093 tags->static_rqs[i] = NULL;
3097 blk_mq_clear_rq_mapping(drv_tags, tags);
3099 while (!list_empty(&tags->page_list)) {
3100 page = list_first_entry(&tags->page_list, struct page, lru);
3101 list_del_init(&page->lru);
3103 * Remove kmemleak object previously allocated in
3104 * blk_mq_alloc_rqs().
3106 kmemleak_free(page_address(page));
3107 __free_pages(page, page->private);
3111 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3115 kfree(tags->static_rqs);
3116 tags->static_rqs = NULL;
3118 blk_mq_free_tags(tags);
3121 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3122 unsigned int hctx_idx)
3126 for (i = 0; i < set->nr_maps; i++) {
3127 unsigned int start = set->map[i].queue_offset;
3128 unsigned int end = start + set->map[i].nr_queues;
3130 if (hctx_idx >= start && hctx_idx < end)
3134 if (i >= set->nr_maps)
3135 i = HCTX_TYPE_DEFAULT;
3140 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3141 unsigned int hctx_idx)
3143 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3145 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3148 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3149 unsigned int hctx_idx,
3150 unsigned int nr_tags,
3151 unsigned int reserved_tags)
3153 int node = blk_mq_get_hctx_node(set, hctx_idx);
3154 struct blk_mq_tags *tags;
3156 if (node == NUMA_NO_NODE)
3157 node = set->numa_node;
3159 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3160 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3164 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3165 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3168 blk_mq_free_tags(tags);
3172 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3173 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3175 if (!tags->static_rqs) {
3177 blk_mq_free_tags(tags);
3184 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3185 unsigned int hctx_idx, int node)
3189 if (set->ops->init_request) {
3190 ret = set->ops->init_request(set, rq, hctx_idx, node);
3195 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3199 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3200 struct blk_mq_tags *tags,
3201 unsigned int hctx_idx, unsigned int depth)
3203 unsigned int i, j, entries_per_page, max_order = 4;
3204 int node = blk_mq_get_hctx_node(set, hctx_idx);
3205 size_t rq_size, left;
3207 if (node == NUMA_NO_NODE)
3208 node = set->numa_node;
3210 INIT_LIST_HEAD(&tags->page_list);
3213 * rq_size is the size of the request plus driver payload, rounded
3214 * to the cacheline size
3216 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3218 left = rq_size * depth;
3220 for (i = 0; i < depth; ) {
3221 int this_order = max_order;
3226 while (this_order && left < order_to_size(this_order - 1))
3230 page = alloc_pages_node(node,
3231 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3237 if (order_to_size(this_order) < rq_size)
3244 page->private = this_order;
3245 list_add_tail(&page->lru, &tags->page_list);
3247 p = page_address(page);
3249 * Allow kmemleak to scan these pages as they contain pointers
3250 * to additional allocations like via ops->init_request().
3252 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3253 entries_per_page = order_to_size(this_order) / rq_size;
3254 to_do = min(entries_per_page, depth - i);
3255 left -= to_do * rq_size;
3256 for (j = 0; j < to_do; j++) {
3257 struct request *rq = p;
3259 tags->static_rqs[i] = rq;
3260 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3261 tags->static_rqs[i] = NULL;
3272 blk_mq_free_rqs(set, tags, hctx_idx);
3276 struct rq_iter_data {
3277 struct blk_mq_hw_ctx *hctx;
3281 static bool blk_mq_has_request(struct request *rq, void *data)
3283 struct rq_iter_data *iter_data = data;
3285 if (rq->mq_hctx != iter_data->hctx)
3287 iter_data->has_rq = true;
3291 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3293 struct blk_mq_tags *tags = hctx->sched_tags ?
3294 hctx->sched_tags : hctx->tags;
3295 struct rq_iter_data data = {
3299 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3303 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3304 struct blk_mq_hw_ctx *hctx)
3306 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3308 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3313 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3315 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3316 struct blk_mq_hw_ctx, cpuhp_online);
3318 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3319 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3323 * Prevent new request from being allocated on the current hctx.
3325 * The smp_mb__after_atomic() Pairs with the implied barrier in
3326 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3327 * seen once we return from the tag allocator.
3329 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3330 smp_mb__after_atomic();
3333 * Try to grab a reference to the queue and wait for any outstanding
3334 * requests. If we could not grab a reference the queue has been
3335 * frozen and there are no requests.
3337 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3338 while (blk_mq_hctx_has_requests(hctx))
3340 percpu_ref_put(&hctx->queue->q_usage_counter);
3346 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3348 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3349 struct blk_mq_hw_ctx, cpuhp_online);
3351 if (cpumask_test_cpu(cpu, hctx->cpumask))
3352 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3357 * 'cpu' is going away. splice any existing rq_list entries from this
3358 * software queue to the hw queue dispatch list, and ensure that it
3361 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3363 struct blk_mq_hw_ctx *hctx;
3364 struct blk_mq_ctx *ctx;
3366 enum hctx_type type;
3368 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3369 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3372 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3375 spin_lock(&ctx->lock);
3376 if (!list_empty(&ctx->rq_lists[type])) {
3377 list_splice_init(&ctx->rq_lists[type], &tmp);
3378 blk_mq_hctx_clear_pending(hctx, ctx);
3380 spin_unlock(&ctx->lock);
3382 if (list_empty(&tmp))
3385 spin_lock(&hctx->lock);
3386 list_splice_tail_init(&tmp, &hctx->dispatch);
3387 spin_unlock(&hctx->lock);
3389 blk_mq_run_hw_queue(hctx, true);
3393 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3395 if (!(hctx->flags & BLK_MQ_F_STACKING))
3396 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3397 &hctx->cpuhp_online);
3398 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3403 * Before freeing hw queue, clearing the flush request reference in
3404 * tags->rqs[] for avoiding potential UAF.
3406 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3407 unsigned int queue_depth, struct request *flush_rq)
3410 unsigned long flags;
3412 /* The hw queue may not be mapped yet */
3416 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3418 for (i = 0; i < queue_depth; i++)
3419 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3422 * Wait until all pending iteration is done.
3424 * Request reference is cleared and it is guaranteed to be observed
3425 * after the ->lock is released.
3427 spin_lock_irqsave(&tags->lock, flags);
3428 spin_unlock_irqrestore(&tags->lock, flags);
3431 /* hctx->ctxs will be freed in queue's release handler */
3432 static void blk_mq_exit_hctx(struct request_queue *q,
3433 struct blk_mq_tag_set *set,
3434 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3436 struct request *flush_rq = hctx->fq->flush_rq;
3438 if (blk_mq_hw_queue_mapped(hctx))
3439 blk_mq_tag_idle(hctx);
3441 if (blk_queue_init_done(q))
3442 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3443 set->queue_depth, flush_rq);
3444 if (set->ops->exit_request)
3445 set->ops->exit_request(set, flush_rq, hctx_idx);
3447 if (set->ops->exit_hctx)
3448 set->ops->exit_hctx(hctx, hctx_idx);
3450 blk_mq_remove_cpuhp(hctx);
3452 xa_erase(&q->hctx_table, hctx_idx);
3454 spin_lock(&q->unused_hctx_lock);
3455 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3456 spin_unlock(&q->unused_hctx_lock);
3459 static void blk_mq_exit_hw_queues(struct request_queue *q,
3460 struct blk_mq_tag_set *set, int nr_queue)
3462 struct blk_mq_hw_ctx *hctx;
3465 queue_for_each_hw_ctx(q, hctx, i) {
3468 blk_mq_exit_hctx(q, set, hctx, i);
3472 static int blk_mq_init_hctx(struct request_queue *q,
3473 struct blk_mq_tag_set *set,
3474 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3476 hctx->queue_num = hctx_idx;
3478 if (!(hctx->flags & BLK_MQ_F_STACKING))
3479 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3480 &hctx->cpuhp_online);
3481 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3483 hctx->tags = set->tags[hctx_idx];
3485 if (set->ops->init_hctx &&
3486 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3487 goto unregister_cpu_notifier;
3489 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3493 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3499 if (set->ops->exit_request)
3500 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3502 if (set->ops->exit_hctx)
3503 set->ops->exit_hctx(hctx, hctx_idx);
3504 unregister_cpu_notifier:
3505 blk_mq_remove_cpuhp(hctx);
3509 static struct blk_mq_hw_ctx *
3510 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3513 struct blk_mq_hw_ctx *hctx;
3514 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3516 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3518 goto fail_alloc_hctx;
3520 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3523 atomic_set(&hctx->nr_active, 0);
3524 if (node == NUMA_NO_NODE)
3525 node = set->numa_node;
3526 hctx->numa_node = node;
3528 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3529 spin_lock_init(&hctx->lock);
3530 INIT_LIST_HEAD(&hctx->dispatch);
3532 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3534 INIT_LIST_HEAD(&hctx->hctx_list);
3537 * Allocate space for all possible cpus to avoid allocation at
3540 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3545 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3546 gfp, node, false, false))
3550 spin_lock_init(&hctx->dispatch_wait_lock);
3551 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3552 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3554 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3558 blk_mq_hctx_kobj_init(hctx);
3563 sbitmap_free(&hctx->ctx_map);
3567 free_cpumask_var(hctx->cpumask);
3574 static void blk_mq_init_cpu_queues(struct request_queue *q,
3575 unsigned int nr_hw_queues)
3577 struct blk_mq_tag_set *set = q->tag_set;
3580 for_each_possible_cpu(i) {
3581 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3582 struct blk_mq_hw_ctx *hctx;
3586 spin_lock_init(&__ctx->lock);
3587 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3588 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3593 * Set local node, IFF we have more than one hw queue. If
3594 * not, we remain on the home node of the device
3596 for (j = 0; j < set->nr_maps; j++) {
3597 hctx = blk_mq_map_queue_type(q, j, i);
3598 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3599 hctx->numa_node = cpu_to_node(i);
3604 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3605 unsigned int hctx_idx,
3608 struct blk_mq_tags *tags;
3611 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3615 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3617 blk_mq_free_rq_map(tags);
3624 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3627 if (blk_mq_is_shared_tags(set->flags)) {
3628 set->tags[hctx_idx] = set->shared_tags;
3633 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3636 return set->tags[hctx_idx];
3639 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3640 struct blk_mq_tags *tags,
3641 unsigned int hctx_idx)
3644 blk_mq_free_rqs(set, tags, hctx_idx);
3645 blk_mq_free_rq_map(tags);
3649 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3650 unsigned int hctx_idx)
3652 if (!blk_mq_is_shared_tags(set->flags))
3653 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3655 set->tags[hctx_idx] = NULL;
3658 static void blk_mq_map_swqueue(struct request_queue *q)
3660 unsigned int j, hctx_idx;
3662 struct blk_mq_hw_ctx *hctx;
3663 struct blk_mq_ctx *ctx;
3664 struct blk_mq_tag_set *set = q->tag_set;
3666 queue_for_each_hw_ctx(q, hctx, i) {
3667 cpumask_clear(hctx->cpumask);
3669 hctx->dispatch_from = NULL;
3673 * Map software to hardware queues.
3675 * If the cpu isn't present, the cpu is mapped to first hctx.
3677 for_each_possible_cpu(i) {
3679 ctx = per_cpu_ptr(q->queue_ctx, i);
3680 for (j = 0; j < set->nr_maps; j++) {
3681 if (!set->map[j].nr_queues) {
3682 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3683 HCTX_TYPE_DEFAULT, i);
3686 hctx_idx = set->map[j].mq_map[i];
3687 /* unmapped hw queue can be remapped after CPU topo changed */
3688 if (!set->tags[hctx_idx] &&
3689 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3691 * If tags initialization fail for some hctx,
3692 * that hctx won't be brought online. In this
3693 * case, remap the current ctx to hctx[0] which
3694 * is guaranteed to always have tags allocated
3696 set->map[j].mq_map[i] = 0;
3699 hctx = blk_mq_map_queue_type(q, j, i);
3700 ctx->hctxs[j] = hctx;
3702 * If the CPU is already set in the mask, then we've
3703 * mapped this one already. This can happen if
3704 * devices share queues across queue maps.
3706 if (cpumask_test_cpu(i, hctx->cpumask))
3709 cpumask_set_cpu(i, hctx->cpumask);
3711 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3712 hctx->ctxs[hctx->nr_ctx++] = ctx;
3715 * If the nr_ctx type overflows, we have exceeded the
3716 * amount of sw queues we can support.
3718 BUG_ON(!hctx->nr_ctx);
3721 for (; j < HCTX_MAX_TYPES; j++)
3722 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3723 HCTX_TYPE_DEFAULT, i);
3726 queue_for_each_hw_ctx(q, hctx, i) {
3728 * If no software queues are mapped to this hardware queue,
3729 * disable it and free the request entries.
3731 if (!hctx->nr_ctx) {
3732 /* Never unmap queue 0. We need it as a
3733 * fallback in case of a new remap fails
3737 __blk_mq_free_map_and_rqs(set, i);
3743 hctx->tags = set->tags[i];
3744 WARN_ON(!hctx->tags);
3747 * Set the map size to the number of mapped software queues.
3748 * This is more accurate and more efficient than looping
3749 * over all possibly mapped software queues.
3751 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3754 * Initialize batch roundrobin counts
3756 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3757 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3762 * Caller needs to ensure that we're either frozen/quiesced, or that
3763 * the queue isn't live yet.
3765 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3767 struct blk_mq_hw_ctx *hctx;
3770 queue_for_each_hw_ctx(q, hctx, i) {
3772 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3774 blk_mq_tag_idle(hctx);
3775 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3780 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3783 struct request_queue *q;
3785 lockdep_assert_held(&set->tag_list_lock);
3787 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3788 blk_mq_freeze_queue(q);
3789 queue_set_hctx_shared(q, shared);
3790 blk_mq_unfreeze_queue(q);
3794 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3796 struct blk_mq_tag_set *set = q->tag_set;
3798 mutex_lock(&set->tag_list_lock);
3799 list_del(&q->tag_set_list);
3800 if (list_is_singular(&set->tag_list)) {
3801 /* just transitioned to unshared */
3802 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3803 /* update existing queue */
3804 blk_mq_update_tag_set_shared(set, false);
3806 mutex_unlock(&set->tag_list_lock);
3807 INIT_LIST_HEAD(&q->tag_set_list);
3810 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3811 struct request_queue *q)
3813 mutex_lock(&set->tag_list_lock);
3816 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3818 if (!list_empty(&set->tag_list) &&
3819 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3820 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3821 /* update existing queue */
3822 blk_mq_update_tag_set_shared(set, true);
3824 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3825 queue_set_hctx_shared(q, true);
3826 list_add_tail(&q->tag_set_list, &set->tag_list);
3828 mutex_unlock(&set->tag_list_lock);
3831 /* All allocations will be freed in release handler of q->mq_kobj */
3832 static int blk_mq_alloc_ctxs(struct request_queue *q)
3834 struct blk_mq_ctxs *ctxs;
3837 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3841 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3842 if (!ctxs->queue_ctx)
3845 for_each_possible_cpu(cpu) {
3846 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3850 q->mq_kobj = &ctxs->kobj;
3851 q->queue_ctx = ctxs->queue_ctx;
3860 * It is the actual release handler for mq, but we do it from
3861 * request queue's release handler for avoiding use-after-free
3862 * and headache because q->mq_kobj shouldn't have been introduced,
3863 * but we can't group ctx/kctx kobj without it.
3865 void blk_mq_release(struct request_queue *q)
3867 struct blk_mq_hw_ctx *hctx, *next;
3870 queue_for_each_hw_ctx(q, hctx, i)
3871 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3873 /* all hctx are in .unused_hctx_list now */
3874 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3875 list_del_init(&hctx->hctx_list);
3876 kobject_put(&hctx->kobj);
3879 xa_destroy(&q->hctx_table);
3882 * release .mq_kobj and sw queue's kobject now because
3883 * both share lifetime with request queue.
3885 blk_mq_sysfs_deinit(q);
3888 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3891 struct request_queue *q;
3894 q = blk_alloc_queue(set->numa_node, set->flags & BLK_MQ_F_BLOCKING);
3896 return ERR_PTR(-ENOMEM);
3897 q->queuedata = queuedata;
3898 ret = blk_mq_init_allocated_queue(set, q);
3901 return ERR_PTR(ret);
3906 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3908 return blk_mq_init_queue_data(set, NULL);
3910 EXPORT_SYMBOL(blk_mq_init_queue);
3913 * blk_mq_destroy_queue - shutdown a request queue
3914 * @q: request queue to shutdown
3916 * This shuts down a request queue allocated by blk_mq_init_queue() and drops
3917 * the initial reference. All future requests will failed with -ENODEV.
3919 * Context: can sleep
3921 void blk_mq_destroy_queue(struct request_queue *q)
3923 WARN_ON_ONCE(!queue_is_mq(q));
3924 WARN_ON_ONCE(blk_queue_registered(q));
3928 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
3929 blk_queue_start_drain(q);
3930 blk_freeze_queue(q);
3933 blk_mq_cancel_work_sync(q);
3934 blk_mq_exit_queue(q);
3936 /* @q is and will stay empty, shutdown and put */
3939 EXPORT_SYMBOL(blk_mq_destroy_queue);
3941 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
3942 struct lock_class_key *lkclass)
3944 struct request_queue *q;
3945 struct gendisk *disk;
3947 q = blk_mq_init_queue_data(set, queuedata);
3951 disk = __alloc_disk_node(q, set->numa_node, lkclass);
3953 blk_mq_destroy_queue(q);
3954 return ERR_PTR(-ENOMEM);
3956 set_bit(GD_OWNS_QUEUE, &disk->state);
3959 EXPORT_SYMBOL(__blk_mq_alloc_disk);
3961 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
3962 struct lock_class_key *lkclass)
3964 if (!blk_get_queue(q))
3966 return __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
3968 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
3970 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3971 struct blk_mq_tag_set *set, struct request_queue *q,
3972 int hctx_idx, int node)
3974 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3976 /* reuse dead hctx first */
3977 spin_lock(&q->unused_hctx_lock);
3978 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3979 if (tmp->numa_node == node) {
3985 list_del_init(&hctx->hctx_list);
3986 spin_unlock(&q->unused_hctx_lock);
3989 hctx = blk_mq_alloc_hctx(q, set, node);
3993 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3999 kobject_put(&hctx->kobj);
4004 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4005 struct request_queue *q)
4007 struct blk_mq_hw_ctx *hctx;
4010 /* protect against switching io scheduler */
4011 mutex_lock(&q->sysfs_lock);
4012 for (i = 0; i < set->nr_hw_queues; i++) {
4014 int node = blk_mq_get_hctx_node(set, i);
4015 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4018 old_node = old_hctx->numa_node;
4019 blk_mq_exit_hctx(q, set, old_hctx, i);
4022 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4025 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4027 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4028 WARN_ON_ONCE(!hctx);
4032 * Increasing nr_hw_queues fails. Free the newly allocated
4033 * hctxs and keep the previous q->nr_hw_queues.
4035 if (i != set->nr_hw_queues) {
4036 j = q->nr_hw_queues;
4039 q->nr_hw_queues = set->nr_hw_queues;
4042 xa_for_each_start(&q->hctx_table, j, hctx, j)
4043 blk_mq_exit_hctx(q, set, hctx, j);
4044 mutex_unlock(&q->sysfs_lock);
4047 static void blk_mq_update_poll_flag(struct request_queue *q)
4049 struct blk_mq_tag_set *set = q->tag_set;
4051 if (set->nr_maps > HCTX_TYPE_POLL &&
4052 set->map[HCTX_TYPE_POLL].nr_queues)
4053 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4055 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4058 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4059 struct request_queue *q)
4061 WARN_ON_ONCE(blk_queue_has_srcu(q) !=
4062 !!(set->flags & BLK_MQ_F_BLOCKING));
4064 /* mark the queue as mq asap */
4065 q->mq_ops = set->ops;
4067 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
4068 blk_mq_poll_stats_bkt,
4069 BLK_MQ_POLL_STATS_BKTS, q);
4073 if (blk_mq_alloc_ctxs(q))
4076 /* init q->mq_kobj and sw queues' kobjects */
4077 blk_mq_sysfs_init(q);
4079 INIT_LIST_HEAD(&q->unused_hctx_list);
4080 spin_lock_init(&q->unused_hctx_lock);
4082 xa_init(&q->hctx_table);
4084 blk_mq_realloc_hw_ctxs(set, q);
4085 if (!q->nr_hw_queues)
4088 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4089 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4093 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4094 blk_mq_update_poll_flag(q);
4096 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4097 INIT_LIST_HEAD(&q->requeue_list);
4098 spin_lock_init(&q->requeue_lock);
4100 q->nr_requests = set->queue_depth;
4103 * Default to classic polling
4105 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
4107 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4108 blk_mq_add_queue_tag_set(set, q);
4109 blk_mq_map_swqueue(q);
4113 xa_destroy(&q->hctx_table);
4114 q->nr_hw_queues = 0;
4115 blk_mq_sysfs_deinit(q);
4117 blk_stat_free_callback(q->poll_cb);
4123 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4125 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4126 void blk_mq_exit_queue(struct request_queue *q)
4128 struct blk_mq_tag_set *set = q->tag_set;
4130 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4131 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4132 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4133 blk_mq_del_queue_tag_set(q);
4136 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4140 if (blk_mq_is_shared_tags(set->flags)) {
4141 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4144 if (!set->shared_tags)
4148 for (i = 0; i < set->nr_hw_queues; i++) {
4149 if (!__blk_mq_alloc_map_and_rqs(set, i))
4158 __blk_mq_free_map_and_rqs(set, i);
4160 if (blk_mq_is_shared_tags(set->flags)) {
4161 blk_mq_free_map_and_rqs(set, set->shared_tags,
4162 BLK_MQ_NO_HCTX_IDX);
4169 * Allocate the request maps associated with this tag_set. Note that this
4170 * may reduce the depth asked for, if memory is tight. set->queue_depth
4171 * will be updated to reflect the allocated depth.
4173 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4178 depth = set->queue_depth;
4180 err = __blk_mq_alloc_rq_maps(set);
4184 set->queue_depth >>= 1;
4185 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4189 } while (set->queue_depth);
4191 if (!set->queue_depth || err) {
4192 pr_err("blk-mq: failed to allocate request map\n");
4196 if (depth != set->queue_depth)
4197 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4198 depth, set->queue_depth);
4203 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4206 * blk_mq_map_queues() and multiple .map_queues() implementations
4207 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4208 * number of hardware queues.
4210 if (set->nr_maps == 1)
4211 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4213 if (set->ops->map_queues && !is_kdump_kernel()) {
4217 * transport .map_queues is usually done in the following
4220 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4221 * mask = get_cpu_mask(queue)
4222 * for_each_cpu(cpu, mask)
4223 * set->map[x].mq_map[cpu] = queue;
4226 * When we need to remap, the table has to be cleared for
4227 * killing stale mapping since one CPU may not be mapped
4230 for (i = 0; i < set->nr_maps; i++)
4231 blk_mq_clear_mq_map(&set->map[i]);
4233 set->ops->map_queues(set);
4235 BUG_ON(set->nr_maps > 1);
4236 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4240 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4241 int cur_nr_hw_queues, int new_nr_hw_queues)
4243 struct blk_mq_tags **new_tags;
4245 if (cur_nr_hw_queues >= new_nr_hw_queues)
4248 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4249 GFP_KERNEL, set->numa_node);
4254 memcpy(new_tags, set->tags, cur_nr_hw_queues *
4255 sizeof(*set->tags));
4257 set->tags = new_tags;
4258 set->nr_hw_queues = new_nr_hw_queues;
4263 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
4264 int new_nr_hw_queues)
4266 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
4270 * Alloc a tag set to be associated with one or more request queues.
4271 * May fail with EINVAL for various error conditions. May adjust the
4272 * requested depth down, if it's too large. In that case, the set
4273 * value will be stored in set->queue_depth.
4275 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4279 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4281 if (!set->nr_hw_queues)
4283 if (!set->queue_depth)
4285 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4288 if (!set->ops->queue_rq)
4291 if (!set->ops->get_budget ^ !set->ops->put_budget)
4294 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4295 pr_info("blk-mq: reduced tag depth to %u\n",
4297 set->queue_depth = BLK_MQ_MAX_DEPTH;
4302 else if (set->nr_maps > HCTX_MAX_TYPES)
4306 * If a crashdump is active, then we are potentially in a very
4307 * memory constrained environment. Limit us to 1 queue and
4308 * 64 tags to prevent using too much memory.
4310 if (is_kdump_kernel()) {
4311 set->nr_hw_queues = 1;
4313 set->queue_depth = min(64U, set->queue_depth);
4316 * There is no use for more h/w queues than cpus if we just have
4319 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4320 set->nr_hw_queues = nr_cpu_ids;
4322 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
4326 for (i = 0; i < set->nr_maps; i++) {
4327 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4328 sizeof(set->map[i].mq_map[0]),
4329 GFP_KERNEL, set->numa_node);
4330 if (!set->map[i].mq_map)
4331 goto out_free_mq_map;
4332 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4335 blk_mq_update_queue_map(set);
4337 ret = blk_mq_alloc_set_map_and_rqs(set);
4339 goto out_free_mq_map;
4341 mutex_init(&set->tag_list_lock);
4342 INIT_LIST_HEAD(&set->tag_list);
4347 for (i = 0; i < set->nr_maps; i++) {
4348 kfree(set->map[i].mq_map);
4349 set->map[i].mq_map = NULL;
4355 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4357 /* allocate and initialize a tagset for a simple single-queue device */
4358 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4359 const struct blk_mq_ops *ops, unsigned int queue_depth,
4360 unsigned int set_flags)
4362 memset(set, 0, sizeof(*set));
4364 set->nr_hw_queues = 1;
4366 set->queue_depth = queue_depth;
4367 set->numa_node = NUMA_NO_NODE;
4368 set->flags = set_flags;
4369 return blk_mq_alloc_tag_set(set);
4371 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4373 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4377 for (i = 0; i < set->nr_hw_queues; i++)
4378 __blk_mq_free_map_and_rqs(set, i);
4380 if (blk_mq_is_shared_tags(set->flags)) {
4381 blk_mq_free_map_and_rqs(set, set->shared_tags,
4382 BLK_MQ_NO_HCTX_IDX);
4385 for (j = 0; j < set->nr_maps; j++) {
4386 kfree(set->map[j].mq_map);
4387 set->map[j].mq_map = NULL;
4393 EXPORT_SYMBOL(blk_mq_free_tag_set);
4395 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4397 struct blk_mq_tag_set *set = q->tag_set;
4398 struct blk_mq_hw_ctx *hctx;
4405 if (q->nr_requests == nr)
4408 blk_mq_freeze_queue(q);
4409 blk_mq_quiesce_queue(q);
4412 queue_for_each_hw_ctx(q, hctx, i) {
4416 * If we're using an MQ scheduler, just update the scheduler
4417 * queue depth. This is similar to what the old code would do.
4419 if (hctx->sched_tags) {
4420 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4423 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4428 if (q->elevator && q->elevator->type->ops.depth_updated)
4429 q->elevator->type->ops.depth_updated(hctx);
4432 q->nr_requests = nr;
4433 if (blk_mq_is_shared_tags(set->flags)) {
4435 blk_mq_tag_update_sched_shared_tags(q);
4437 blk_mq_tag_resize_shared_tags(set, nr);
4441 blk_mq_unquiesce_queue(q);
4442 blk_mq_unfreeze_queue(q);
4448 * request_queue and elevator_type pair.
4449 * It is just used by __blk_mq_update_nr_hw_queues to cache
4450 * the elevator_type associated with a request_queue.
4452 struct blk_mq_qe_pair {
4453 struct list_head node;
4454 struct request_queue *q;
4455 struct elevator_type *type;
4459 * Cache the elevator_type in qe pair list and switch the
4460 * io scheduler to 'none'
4462 static bool blk_mq_elv_switch_none(struct list_head *head,
4463 struct request_queue *q)
4465 struct blk_mq_qe_pair *qe;
4470 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4474 /* q->elevator needs protection from ->sysfs_lock */
4475 mutex_lock(&q->sysfs_lock);
4477 INIT_LIST_HEAD(&qe->node);
4479 qe->type = q->elevator->type;
4480 list_add(&qe->node, head);
4483 * After elevator_switch, the previous elevator_queue will be
4484 * released by elevator_release. The reference of the io scheduler
4485 * module get by elevator_get will also be put. So we need to get
4486 * a reference of the io scheduler module here to prevent it to be
4489 __module_get(qe->type->elevator_owner);
4490 elevator_switch(q, NULL);
4491 mutex_unlock(&q->sysfs_lock);
4496 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4497 struct request_queue *q)
4499 struct blk_mq_qe_pair *qe;
4501 list_for_each_entry(qe, head, node)
4508 static void blk_mq_elv_switch_back(struct list_head *head,
4509 struct request_queue *q)
4511 struct blk_mq_qe_pair *qe;
4512 struct elevator_type *t;
4514 qe = blk_lookup_qe_pair(head, q);
4518 list_del(&qe->node);
4521 mutex_lock(&q->sysfs_lock);
4522 elevator_switch(q, t);
4523 mutex_unlock(&q->sysfs_lock);
4526 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4529 struct request_queue *q;
4531 int prev_nr_hw_queues;
4533 lockdep_assert_held(&set->tag_list_lock);
4535 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4536 nr_hw_queues = nr_cpu_ids;
4537 if (nr_hw_queues < 1)
4539 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4542 list_for_each_entry(q, &set->tag_list, tag_set_list)
4543 blk_mq_freeze_queue(q);
4545 * Switch IO scheduler to 'none', cleaning up the data associated
4546 * with the previous scheduler. We will switch back once we are done
4547 * updating the new sw to hw queue mappings.
4549 list_for_each_entry(q, &set->tag_list, tag_set_list)
4550 if (!blk_mq_elv_switch_none(&head, q))
4553 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4554 blk_mq_debugfs_unregister_hctxs(q);
4555 blk_mq_sysfs_unregister_hctxs(q);
4558 prev_nr_hw_queues = set->nr_hw_queues;
4559 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
4563 set->nr_hw_queues = nr_hw_queues;
4565 blk_mq_update_queue_map(set);
4566 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4567 blk_mq_realloc_hw_ctxs(set, q);
4568 blk_mq_update_poll_flag(q);
4569 if (q->nr_hw_queues != set->nr_hw_queues) {
4570 int i = prev_nr_hw_queues;
4572 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4573 nr_hw_queues, prev_nr_hw_queues);
4574 for (; i < set->nr_hw_queues; i++)
4575 __blk_mq_free_map_and_rqs(set, i);
4577 set->nr_hw_queues = prev_nr_hw_queues;
4578 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4581 blk_mq_map_swqueue(q);
4585 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4586 blk_mq_sysfs_register_hctxs(q);
4587 blk_mq_debugfs_register_hctxs(q);
4591 list_for_each_entry(q, &set->tag_list, tag_set_list)
4592 blk_mq_elv_switch_back(&head, q);
4594 list_for_each_entry(q, &set->tag_list, tag_set_list)
4595 blk_mq_unfreeze_queue(q);
4598 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4600 mutex_lock(&set->tag_list_lock);
4601 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4602 mutex_unlock(&set->tag_list_lock);
4604 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4606 /* Enable polling stats and return whether they were already enabled. */
4607 static bool blk_poll_stats_enable(struct request_queue *q)
4612 return blk_stats_alloc_enable(q);
4615 static void blk_mq_poll_stats_start(struct request_queue *q)
4618 * We don't arm the callback if polling stats are not enabled or the
4619 * callback is already active.
4621 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
4624 blk_stat_activate_msecs(q->poll_cb, 100);
4627 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4629 struct request_queue *q = cb->data;
4632 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4633 if (cb->stat[bucket].nr_samples)
4634 q->poll_stat[bucket] = cb->stat[bucket];
4638 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4641 unsigned long ret = 0;
4645 * If stats collection isn't on, don't sleep but turn it on for
4648 if (!blk_poll_stats_enable(q))
4652 * As an optimistic guess, use half of the mean service time
4653 * for this type of request. We can (and should) make this smarter.
4654 * For instance, if the completion latencies are tight, we can
4655 * get closer than just half the mean. This is especially
4656 * important on devices where the completion latencies are longer
4657 * than ~10 usec. We do use the stats for the relevant IO size
4658 * if available which does lead to better estimates.
4660 bucket = blk_mq_poll_stats_bkt(rq);
4664 if (q->poll_stat[bucket].nr_samples)
4665 ret = (q->poll_stat[bucket].mean + 1) / 2;
4670 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4672 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4673 struct request *rq = blk_qc_to_rq(hctx, qc);
4674 struct hrtimer_sleeper hs;
4675 enum hrtimer_mode mode;
4680 * If a request has completed on queue that uses an I/O scheduler, we
4681 * won't get back a request from blk_qc_to_rq.
4683 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4687 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4689 * 0: use half of prev avg
4690 * >0: use this specific value
4692 if (q->poll_nsec > 0)
4693 nsecs = q->poll_nsec;
4695 nsecs = blk_mq_poll_nsecs(q, rq);
4700 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4703 * This will be replaced with the stats tracking code, using
4704 * 'avg_completion_time / 2' as the pre-sleep target.
4708 mode = HRTIMER_MODE_REL;
4709 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4710 hrtimer_set_expires(&hs.timer, kt);
4713 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4715 set_current_state(TASK_UNINTERRUPTIBLE);
4716 hrtimer_sleeper_start_expires(&hs, mode);
4719 hrtimer_cancel(&hs.timer);
4720 mode = HRTIMER_MODE_ABS;
4721 } while (hs.task && !signal_pending(current));
4723 __set_current_state(TASK_RUNNING);
4724 destroy_hrtimer_on_stack(&hs.timer);
4727 * If we sleep, have the caller restart the poll loop to reset the
4728 * state. Like for the other success return cases, the caller is
4729 * responsible for checking if the IO completed. If the IO isn't
4730 * complete, we'll get called again and will go straight to the busy
4736 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4737 struct io_comp_batch *iob, unsigned int flags)
4739 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4740 long state = get_current_state();
4744 ret = q->mq_ops->poll(hctx, iob);
4746 __set_current_state(TASK_RUNNING);
4750 if (signal_pending_state(state, current))
4751 __set_current_state(TASK_RUNNING);
4752 if (task_is_running(current))
4755 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4758 } while (!need_resched());
4760 __set_current_state(TASK_RUNNING);
4764 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4767 if (!(flags & BLK_POLL_NOSLEEP) &&
4768 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4769 if (blk_mq_poll_hybrid(q, cookie))
4772 return blk_mq_poll_classic(q, cookie, iob, flags);
4775 unsigned int blk_mq_rq_cpu(struct request *rq)
4777 return rq->mq_ctx->cpu;
4779 EXPORT_SYMBOL(blk_mq_rq_cpu);
4781 void blk_mq_cancel_work_sync(struct request_queue *q)
4783 if (queue_is_mq(q)) {
4784 struct blk_mq_hw_ctx *hctx;
4787 cancel_delayed_work_sync(&q->requeue_work);
4789 queue_for_each_hw_ctx(q, hctx, i)
4790 cancel_delayed_work_sync(&hctx->run_work);
4794 static int __init blk_mq_init(void)
4798 for_each_possible_cpu(i)
4799 init_llist_head(&per_cpu(blk_cpu_done, i));
4800 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4802 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4803 "block/softirq:dead", NULL,
4804 blk_softirq_cpu_dead);
4805 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4806 blk_mq_hctx_notify_dead);
4807 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4808 blk_mq_hctx_notify_online,
4809 blk_mq_hctx_notify_offline);
4812 subsys_initcall(blk_mq_init);