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 a polled request, always complete locally, it's pointless
1097 * to redirect the completion.
1099 if (rq->cmd_flags & REQ_POLLED)
1102 if (blk_mq_complete_need_ipi(rq)) {
1103 blk_mq_complete_send_ipi(rq);
1107 if (rq->q->nr_hw_queues == 1) {
1108 blk_mq_raise_softirq(rq);
1113 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1116 * blk_mq_complete_request - end I/O on a request
1117 * @rq: the request being processed
1120 * Complete a request by scheduling the ->complete_rq operation.
1122 void blk_mq_complete_request(struct request *rq)
1124 if (!blk_mq_complete_request_remote(rq))
1125 rq->q->mq_ops->complete(rq);
1127 EXPORT_SYMBOL(blk_mq_complete_request);
1130 * blk_mq_start_request - Start processing a request
1131 * @rq: Pointer to request to be started
1133 * Function used by device drivers to notify the block layer that a request
1134 * is going to be processed now, so blk layer can do proper initializations
1135 * such as starting the timeout timer.
1137 void blk_mq_start_request(struct request *rq)
1139 struct request_queue *q = rq->q;
1141 trace_block_rq_issue(rq);
1143 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1144 rq->io_start_time_ns = ktime_get_ns();
1145 rq->stats_sectors = blk_rq_sectors(rq);
1146 rq->rq_flags |= RQF_STATS;
1147 rq_qos_issue(q, rq);
1150 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1153 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1155 #ifdef CONFIG_BLK_DEV_INTEGRITY
1156 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1157 q->integrity.profile->prepare_fn(rq);
1159 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1160 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1162 EXPORT_SYMBOL(blk_mq_start_request);
1165 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1166 * queues. This is important for md arrays to benefit from merging
1169 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1171 if (plug->multiple_queues)
1172 return BLK_MAX_REQUEST_COUNT * 2;
1173 return BLK_MAX_REQUEST_COUNT;
1176 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1178 struct request *last = rq_list_peek(&plug->mq_list);
1180 if (!plug->rq_count) {
1181 trace_block_plug(rq->q);
1182 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1183 (!blk_queue_nomerges(rq->q) &&
1184 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1185 blk_mq_flush_plug_list(plug, false);
1187 trace_block_plug(rq->q);
1190 if (!plug->multiple_queues && last && last->q != rq->q)
1191 plug->multiple_queues = true;
1192 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1193 plug->has_elevator = true;
1195 rq_list_add(&plug->mq_list, rq);
1200 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1201 * @rq: request to insert
1202 * @at_head: insert request at head or tail of queue
1205 * Insert a fully prepared request at the back of the I/O scheduler queue
1206 * for execution. Don't wait for completion.
1209 * This function will invoke @done directly if the queue is dead.
1211 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1213 WARN_ON(irqs_disabled());
1214 WARN_ON(!blk_rq_is_passthrough(rq));
1216 blk_account_io_start(rq);
1218 blk_add_rq_to_plug(current->plug, rq);
1220 blk_mq_sched_insert_request(rq, at_head, true, false);
1222 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1224 struct blk_rq_wait {
1225 struct completion done;
1229 static void blk_end_sync_rq(struct request *rq, blk_status_t ret)
1231 struct blk_rq_wait *wait = rq->end_io_data;
1234 complete(&wait->done);
1237 static bool blk_rq_is_poll(struct request *rq)
1241 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1243 if (WARN_ON_ONCE(!rq->bio))
1248 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1251 bio_poll(rq->bio, NULL, 0);
1253 } while (!completion_done(wait));
1257 * blk_execute_rq - insert a request into queue for execution
1258 * @rq: request to insert
1259 * @at_head: insert request at head or tail of queue
1262 * Insert a fully prepared request at the back of the I/O scheduler queue
1263 * for execution and wait for completion.
1264 * Return: The blk_status_t result provided to blk_mq_end_request().
1266 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1268 struct blk_rq_wait wait = {
1269 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1272 WARN_ON(irqs_disabled());
1273 WARN_ON(!blk_rq_is_passthrough(rq));
1275 rq->end_io_data = &wait;
1276 rq->end_io = blk_end_sync_rq;
1278 blk_account_io_start(rq);
1279 blk_mq_sched_insert_request(rq, at_head, true, false);
1281 if (blk_rq_is_poll(rq)) {
1282 blk_rq_poll_completion(rq, &wait.done);
1285 * Prevent hang_check timer from firing at us during very long
1288 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1291 while (!wait_for_completion_io_timeout(&wait.done,
1292 hang_check * (HZ/2)))
1295 wait_for_completion_io(&wait.done);
1300 EXPORT_SYMBOL(blk_execute_rq);
1302 static void __blk_mq_requeue_request(struct request *rq)
1304 struct request_queue *q = rq->q;
1306 blk_mq_put_driver_tag(rq);
1308 trace_block_rq_requeue(rq);
1309 rq_qos_requeue(q, rq);
1311 if (blk_mq_request_started(rq)) {
1312 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1313 rq->rq_flags &= ~RQF_TIMED_OUT;
1317 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1319 __blk_mq_requeue_request(rq);
1321 /* this request will be re-inserted to io scheduler queue */
1322 blk_mq_sched_requeue_request(rq);
1324 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1326 EXPORT_SYMBOL(blk_mq_requeue_request);
1328 static void blk_mq_requeue_work(struct work_struct *work)
1330 struct request_queue *q =
1331 container_of(work, struct request_queue, requeue_work.work);
1333 struct request *rq, *next;
1335 spin_lock_irq(&q->requeue_lock);
1336 list_splice_init(&q->requeue_list, &rq_list);
1337 spin_unlock_irq(&q->requeue_lock);
1339 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1340 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1343 rq->rq_flags &= ~RQF_SOFTBARRIER;
1344 list_del_init(&rq->queuelist);
1346 * If RQF_DONTPREP, rq has contained some driver specific
1347 * data, so insert it to hctx dispatch list to avoid any
1350 if (rq->rq_flags & RQF_DONTPREP)
1351 blk_mq_request_bypass_insert(rq, false, false);
1353 blk_mq_sched_insert_request(rq, true, false, false);
1356 while (!list_empty(&rq_list)) {
1357 rq = list_entry(rq_list.next, struct request, queuelist);
1358 list_del_init(&rq->queuelist);
1359 blk_mq_sched_insert_request(rq, false, false, false);
1362 blk_mq_run_hw_queues(q, false);
1365 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1366 bool kick_requeue_list)
1368 struct request_queue *q = rq->q;
1369 unsigned long flags;
1372 * We abuse this flag that is otherwise used by the I/O scheduler to
1373 * request head insertion from the workqueue.
1375 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1377 spin_lock_irqsave(&q->requeue_lock, flags);
1379 rq->rq_flags |= RQF_SOFTBARRIER;
1380 list_add(&rq->queuelist, &q->requeue_list);
1382 list_add_tail(&rq->queuelist, &q->requeue_list);
1384 spin_unlock_irqrestore(&q->requeue_lock, flags);
1386 if (kick_requeue_list)
1387 blk_mq_kick_requeue_list(q);
1390 void blk_mq_kick_requeue_list(struct request_queue *q)
1392 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1394 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1396 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1397 unsigned long msecs)
1399 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1400 msecs_to_jiffies(msecs));
1402 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1404 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1407 * If we find a request that isn't idle we know the queue is busy
1408 * as it's checked in the iter.
1409 * Return false to stop the iteration.
1411 if (blk_mq_request_started(rq)) {
1421 bool blk_mq_queue_inflight(struct request_queue *q)
1425 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1428 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1430 static void blk_mq_rq_timed_out(struct request *req)
1432 req->rq_flags |= RQF_TIMED_OUT;
1433 if (req->q->mq_ops->timeout) {
1434 enum blk_eh_timer_return ret;
1436 ret = req->q->mq_ops->timeout(req);
1437 if (ret == BLK_EH_DONE)
1439 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1445 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
1447 unsigned long deadline;
1449 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1451 if (rq->rq_flags & RQF_TIMED_OUT)
1454 deadline = READ_ONCE(rq->deadline);
1455 if (time_after_eq(jiffies, deadline))
1460 else if (time_after(*next, deadline))
1465 void blk_mq_put_rq_ref(struct request *rq)
1467 if (is_flush_rq(rq))
1469 else if (req_ref_put_and_test(rq))
1470 __blk_mq_free_request(rq);
1473 static bool blk_mq_check_expired(struct request *rq, void *priv)
1475 unsigned long *next = priv;
1478 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1479 * be reallocated underneath the timeout handler's processing, then
1480 * the expire check is reliable. If the request is not expired, then
1481 * it was completed and reallocated as a new request after returning
1482 * from blk_mq_check_expired().
1484 if (blk_mq_req_expired(rq, next))
1485 blk_mq_rq_timed_out(rq);
1489 static void blk_mq_timeout_work(struct work_struct *work)
1491 struct request_queue *q =
1492 container_of(work, struct request_queue, timeout_work);
1493 unsigned long next = 0;
1494 struct blk_mq_hw_ctx *hctx;
1497 /* A deadlock might occur if a request is stuck requiring a
1498 * timeout at the same time a queue freeze is waiting
1499 * completion, since the timeout code would not be able to
1500 * acquire the queue reference here.
1502 * That's why we don't use blk_queue_enter here; instead, we use
1503 * percpu_ref_tryget directly, because we need to be able to
1504 * obtain a reference even in the short window between the queue
1505 * starting to freeze, by dropping the first reference in
1506 * blk_freeze_queue_start, and the moment the last request is
1507 * consumed, marked by the instant q_usage_counter reaches
1510 if (!percpu_ref_tryget(&q->q_usage_counter))
1513 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
1516 mod_timer(&q->timeout, next);
1519 * Request timeouts are handled as a forward rolling timer. If
1520 * we end up here it means that no requests are pending and
1521 * also that no request has been pending for a while. Mark
1522 * each hctx as idle.
1524 queue_for_each_hw_ctx(q, hctx, i) {
1525 /* the hctx may be unmapped, so check it here */
1526 if (blk_mq_hw_queue_mapped(hctx))
1527 blk_mq_tag_idle(hctx);
1533 struct flush_busy_ctx_data {
1534 struct blk_mq_hw_ctx *hctx;
1535 struct list_head *list;
1538 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1540 struct flush_busy_ctx_data *flush_data = data;
1541 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1542 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1543 enum hctx_type type = hctx->type;
1545 spin_lock(&ctx->lock);
1546 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1547 sbitmap_clear_bit(sb, bitnr);
1548 spin_unlock(&ctx->lock);
1553 * Process software queues that have been marked busy, splicing them
1554 * to the for-dispatch
1556 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1558 struct flush_busy_ctx_data data = {
1563 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1565 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1567 struct dispatch_rq_data {
1568 struct blk_mq_hw_ctx *hctx;
1572 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1575 struct dispatch_rq_data *dispatch_data = data;
1576 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1577 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1578 enum hctx_type type = hctx->type;
1580 spin_lock(&ctx->lock);
1581 if (!list_empty(&ctx->rq_lists[type])) {
1582 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1583 list_del_init(&dispatch_data->rq->queuelist);
1584 if (list_empty(&ctx->rq_lists[type]))
1585 sbitmap_clear_bit(sb, bitnr);
1587 spin_unlock(&ctx->lock);
1589 return !dispatch_data->rq;
1592 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1593 struct blk_mq_ctx *start)
1595 unsigned off = start ? start->index_hw[hctx->type] : 0;
1596 struct dispatch_rq_data data = {
1601 __sbitmap_for_each_set(&hctx->ctx_map, off,
1602 dispatch_rq_from_ctx, &data);
1607 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1609 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1610 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1613 blk_mq_tag_busy(rq->mq_hctx);
1615 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1616 bt = &rq->mq_hctx->tags->breserved_tags;
1619 if (!hctx_may_queue(rq->mq_hctx, bt))
1623 tag = __sbitmap_queue_get(bt);
1624 if (tag == BLK_MQ_NO_TAG)
1627 rq->tag = tag + tag_offset;
1631 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1633 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1636 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1637 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1638 rq->rq_flags |= RQF_MQ_INFLIGHT;
1639 __blk_mq_inc_active_requests(hctx);
1641 hctx->tags->rqs[rq->tag] = rq;
1645 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1646 int flags, void *key)
1648 struct blk_mq_hw_ctx *hctx;
1650 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1652 spin_lock(&hctx->dispatch_wait_lock);
1653 if (!list_empty(&wait->entry)) {
1654 struct sbitmap_queue *sbq;
1656 list_del_init(&wait->entry);
1657 sbq = &hctx->tags->bitmap_tags;
1658 atomic_dec(&sbq->ws_active);
1660 spin_unlock(&hctx->dispatch_wait_lock);
1662 blk_mq_run_hw_queue(hctx, true);
1667 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1668 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1669 * restart. For both cases, take care to check the condition again after
1670 * marking us as waiting.
1672 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1675 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1676 struct wait_queue_head *wq;
1677 wait_queue_entry_t *wait;
1680 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1681 blk_mq_sched_mark_restart_hctx(hctx);
1684 * It's possible that a tag was freed in the window between the
1685 * allocation failure and adding the hardware queue to the wait
1688 * Don't clear RESTART here, someone else could have set it.
1689 * At most this will cost an extra queue run.
1691 return blk_mq_get_driver_tag(rq);
1694 wait = &hctx->dispatch_wait;
1695 if (!list_empty_careful(&wait->entry))
1698 wq = &bt_wait_ptr(sbq, hctx)->wait;
1700 spin_lock_irq(&wq->lock);
1701 spin_lock(&hctx->dispatch_wait_lock);
1702 if (!list_empty(&wait->entry)) {
1703 spin_unlock(&hctx->dispatch_wait_lock);
1704 spin_unlock_irq(&wq->lock);
1708 atomic_inc(&sbq->ws_active);
1709 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1710 __add_wait_queue(wq, wait);
1713 * It's possible that a tag was freed in the window between the
1714 * allocation failure and adding the hardware queue to the wait
1717 ret = blk_mq_get_driver_tag(rq);
1719 spin_unlock(&hctx->dispatch_wait_lock);
1720 spin_unlock_irq(&wq->lock);
1725 * We got a tag, remove ourselves from the wait queue to ensure
1726 * someone else gets the wakeup.
1728 list_del_init(&wait->entry);
1729 atomic_dec(&sbq->ws_active);
1730 spin_unlock(&hctx->dispatch_wait_lock);
1731 spin_unlock_irq(&wq->lock);
1736 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1737 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1739 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1740 * - EWMA is one simple way to compute running average value
1741 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1742 * - take 4 as factor for avoiding to get too small(0) result, and this
1743 * factor doesn't matter because EWMA decreases exponentially
1745 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1749 ewma = hctx->dispatch_busy;
1754 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1756 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1757 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1759 hctx->dispatch_busy = ewma;
1762 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1764 static void blk_mq_handle_dev_resource(struct request *rq,
1765 struct list_head *list)
1767 struct request *next =
1768 list_first_entry_or_null(list, struct request, queuelist);
1771 * If an I/O scheduler has been configured and we got a driver tag for
1772 * the next request already, free it.
1775 blk_mq_put_driver_tag(next);
1777 list_add(&rq->queuelist, list);
1778 __blk_mq_requeue_request(rq);
1781 static void blk_mq_handle_zone_resource(struct request *rq,
1782 struct list_head *zone_list)
1785 * If we end up here it is because we cannot dispatch a request to a
1786 * specific zone due to LLD level zone-write locking or other zone
1787 * related resource not being available. In this case, set the request
1788 * aside in zone_list for retrying it later.
1790 list_add(&rq->queuelist, zone_list);
1791 __blk_mq_requeue_request(rq);
1794 enum prep_dispatch {
1796 PREP_DISPATCH_NO_TAG,
1797 PREP_DISPATCH_NO_BUDGET,
1800 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1803 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1804 int budget_token = -1;
1807 budget_token = blk_mq_get_dispatch_budget(rq->q);
1808 if (budget_token < 0) {
1809 blk_mq_put_driver_tag(rq);
1810 return PREP_DISPATCH_NO_BUDGET;
1812 blk_mq_set_rq_budget_token(rq, budget_token);
1815 if (!blk_mq_get_driver_tag(rq)) {
1817 * The initial allocation attempt failed, so we need to
1818 * rerun the hardware queue when a tag is freed. The
1819 * waitqueue takes care of that. If the queue is run
1820 * before we add this entry back on the dispatch list,
1821 * we'll re-run it below.
1823 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1825 * All budgets not got from this function will be put
1826 * together during handling partial dispatch
1829 blk_mq_put_dispatch_budget(rq->q, budget_token);
1830 return PREP_DISPATCH_NO_TAG;
1834 return PREP_DISPATCH_OK;
1837 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1838 static void blk_mq_release_budgets(struct request_queue *q,
1839 struct list_head *list)
1843 list_for_each_entry(rq, list, queuelist) {
1844 int budget_token = blk_mq_get_rq_budget_token(rq);
1846 if (budget_token >= 0)
1847 blk_mq_put_dispatch_budget(q, budget_token);
1852 * Returns true if we did some work AND can potentially do more.
1854 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1855 unsigned int nr_budgets)
1857 enum prep_dispatch prep;
1858 struct request_queue *q = hctx->queue;
1859 struct request *rq, *nxt;
1861 blk_status_t ret = BLK_STS_OK;
1862 LIST_HEAD(zone_list);
1863 bool needs_resource = false;
1865 if (list_empty(list))
1869 * Now process all the entries, sending them to the driver.
1871 errors = queued = 0;
1873 struct blk_mq_queue_data bd;
1875 rq = list_first_entry(list, struct request, queuelist);
1877 WARN_ON_ONCE(hctx != rq->mq_hctx);
1878 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1879 if (prep != PREP_DISPATCH_OK)
1882 list_del_init(&rq->queuelist);
1887 * Flag last if we have no more requests, or if we have more
1888 * but can't assign a driver tag to it.
1890 if (list_empty(list))
1893 nxt = list_first_entry(list, struct request, queuelist);
1894 bd.last = !blk_mq_get_driver_tag(nxt);
1898 * once the request is queued to lld, no need to cover the
1903 ret = q->mq_ops->queue_rq(hctx, &bd);
1908 case BLK_STS_RESOURCE:
1909 needs_resource = true;
1911 case BLK_STS_DEV_RESOURCE:
1912 blk_mq_handle_dev_resource(rq, list);
1914 case BLK_STS_ZONE_RESOURCE:
1916 * Move the request to zone_list and keep going through
1917 * the dispatch list to find more requests the drive can
1920 blk_mq_handle_zone_resource(rq, &zone_list);
1921 needs_resource = true;
1925 blk_mq_end_request(rq, ret);
1927 } while (!list_empty(list));
1929 if (!list_empty(&zone_list))
1930 list_splice_tail_init(&zone_list, list);
1932 /* If we didn't flush the entire list, we could have told the driver
1933 * there was more coming, but that turned out to be a lie.
1935 if ((!list_empty(list) || errors || needs_resource ||
1936 ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued)
1937 q->mq_ops->commit_rqs(hctx);
1939 * Any items that need requeuing? Stuff them into hctx->dispatch,
1940 * that is where we will continue on next queue run.
1942 if (!list_empty(list)) {
1944 /* For non-shared tags, the RESTART check will suffice */
1945 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1946 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1949 blk_mq_release_budgets(q, list);
1951 spin_lock(&hctx->lock);
1952 list_splice_tail_init(list, &hctx->dispatch);
1953 spin_unlock(&hctx->lock);
1956 * Order adding requests to hctx->dispatch and checking
1957 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1958 * in blk_mq_sched_restart(). Avoid restart code path to
1959 * miss the new added requests to hctx->dispatch, meantime
1960 * SCHED_RESTART is observed here.
1965 * If SCHED_RESTART was set by the caller of this function and
1966 * it is no longer set that means that it was cleared by another
1967 * thread and hence that a queue rerun is needed.
1969 * If 'no_tag' is set, that means that we failed getting
1970 * a driver tag with an I/O scheduler attached. If our dispatch
1971 * waitqueue is no longer active, ensure that we run the queue
1972 * AFTER adding our entries back to the list.
1974 * If no I/O scheduler has been configured it is possible that
1975 * the hardware queue got stopped and restarted before requests
1976 * were pushed back onto the dispatch list. Rerun the queue to
1977 * avoid starvation. Notes:
1978 * - blk_mq_run_hw_queue() checks whether or not a queue has
1979 * been stopped before rerunning a queue.
1980 * - Some but not all block drivers stop a queue before
1981 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1984 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1985 * bit is set, run queue after a delay to avoid IO stalls
1986 * that could otherwise occur if the queue is idle. We'll do
1987 * similar if we couldn't get budget or couldn't lock a zone
1988 * and SCHED_RESTART is set.
1990 needs_restart = blk_mq_sched_needs_restart(hctx);
1991 if (prep == PREP_DISPATCH_NO_BUDGET)
1992 needs_resource = true;
1993 if (!needs_restart ||
1994 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1995 blk_mq_run_hw_queue(hctx, true);
1996 else if (needs_restart && needs_resource)
1997 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1999 blk_mq_update_dispatch_busy(hctx, true);
2002 blk_mq_update_dispatch_busy(hctx, false);
2004 return (queued + errors) != 0;
2008 * __blk_mq_run_hw_queue - Run a hardware queue.
2009 * @hctx: Pointer to the hardware queue to run.
2011 * Send pending requests to the hardware.
2013 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
2016 * We can't run the queue inline with ints disabled. Ensure that
2017 * we catch bad users of this early.
2019 WARN_ON_ONCE(in_interrupt());
2021 blk_mq_run_dispatch_ops(hctx->queue,
2022 blk_mq_sched_dispatch_requests(hctx));
2025 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2027 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2029 if (cpu >= nr_cpu_ids)
2030 cpu = cpumask_first(hctx->cpumask);
2035 * It'd be great if the workqueue API had a way to pass
2036 * in a mask and had some smarts for more clever placement.
2037 * For now we just round-robin here, switching for every
2038 * BLK_MQ_CPU_WORK_BATCH queued items.
2040 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2043 int next_cpu = hctx->next_cpu;
2045 if (hctx->queue->nr_hw_queues == 1)
2046 return WORK_CPU_UNBOUND;
2048 if (--hctx->next_cpu_batch <= 0) {
2050 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2052 if (next_cpu >= nr_cpu_ids)
2053 next_cpu = blk_mq_first_mapped_cpu(hctx);
2054 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2058 * Do unbound schedule if we can't find a online CPU for this hctx,
2059 * and it should only happen in the path of handling CPU DEAD.
2061 if (!cpu_online(next_cpu)) {
2068 * Make sure to re-select CPU next time once after CPUs
2069 * in hctx->cpumask become online again.
2071 hctx->next_cpu = next_cpu;
2072 hctx->next_cpu_batch = 1;
2073 return WORK_CPU_UNBOUND;
2076 hctx->next_cpu = next_cpu;
2081 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2082 * @hctx: Pointer to the hardware queue to run.
2083 * @async: If we want to run the queue asynchronously.
2084 * @msecs: Milliseconds of delay to wait before running the queue.
2086 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2087 * with a delay of @msecs.
2089 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2090 unsigned long msecs)
2092 if (unlikely(blk_mq_hctx_stopped(hctx)))
2095 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2096 if (cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2097 __blk_mq_run_hw_queue(hctx);
2102 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2103 msecs_to_jiffies(msecs));
2107 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2108 * @hctx: Pointer to the hardware queue to run.
2109 * @msecs: Milliseconds of delay to wait before running the queue.
2111 * Run a hardware queue asynchronously with a delay of @msecs.
2113 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2115 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2117 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2120 * blk_mq_run_hw_queue - Start to run a hardware queue.
2121 * @hctx: Pointer to the hardware queue to run.
2122 * @async: If we want to run the queue asynchronously.
2124 * Check if the request queue is not in a quiesced state and if there are
2125 * pending requests to be sent. If this is true, run the queue to send requests
2128 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2133 * When queue is quiesced, we may be switching io scheduler, or
2134 * updating nr_hw_queues, or other things, and we can't run queue
2135 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2137 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2140 __blk_mq_run_dispatch_ops(hctx->queue, false,
2141 need_run = !blk_queue_quiesced(hctx->queue) &&
2142 blk_mq_hctx_has_pending(hctx));
2145 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2147 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2150 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2153 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2155 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2157 * If the IO scheduler does not respect hardware queues when
2158 * dispatching, we just don't bother with multiple HW queues and
2159 * dispatch from hctx for the current CPU since running multiple queues
2160 * just causes lock contention inside the scheduler and pointless cache
2163 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2165 if (!blk_mq_hctx_stopped(hctx))
2171 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2172 * @q: Pointer to the request queue to run.
2173 * @async: If we want to run the queue asynchronously.
2175 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2177 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2181 if (blk_queue_sq_sched(q))
2182 sq_hctx = blk_mq_get_sq_hctx(q);
2183 queue_for_each_hw_ctx(q, hctx, i) {
2184 if (blk_mq_hctx_stopped(hctx))
2187 * Dispatch from this hctx either if there's no hctx preferred
2188 * by IO scheduler or if it has requests that bypass the
2191 if (!sq_hctx || sq_hctx == hctx ||
2192 !list_empty_careful(&hctx->dispatch))
2193 blk_mq_run_hw_queue(hctx, async);
2196 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2199 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2200 * @q: Pointer to the request queue to run.
2201 * @msecs: Milliseconds of delay to wait before running the queues.
2203 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2205 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2209 if (blk_queue_sq_sched(q))
2210 sq_hctx = blk_mq_get_sq_hctx(q);
2211 queue_for_each_hw_ctx(q, hctx, i) {
2212 if (blk_mq_hctx_stopped(hctx))
2215 * If there is already a run_work pending, leave the
2216 * pending delay untouched. Otherwise, a hctx can stall
2217 * if another hctx is re-delaying the other's work
2218 * before the work executes.
2220 if (delayed_work_pending(&hctx->run_work))
2223 * Dispatch from this hctx either if there's no hctx preferred
2224 * by IO scheduler or if it has requests that bypass the
2227 if (!sq_hctx || sq_hctx == hctx ||
2228 !list_empty_careful(&hctx->dispatch))
2229 blk_mq_delay_run_hw_queue(hctx, msecs);
2232 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2235 * This function is often used for pausing .queue_rq() by driver when
2236 * there isn't enough resource or some conditions aren't satisfied, and
2237 * BLK_STS_RESOURCE is usually returned.
2239 * We do not guarantee that dispatch can be drained or blocked
2240 * after blk_mq_stop_hw_queue() returns. Please use
2241 * blk_mq_quiesce_queue() for that requirement.
2243 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2245 cancel_delayed_work(&hctx->run_work);
2247 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2249 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2252 * This function is often used for pausing .queue_rq() by driver when
2253 * there isn't enough resource or some conditions aren't satisfied, and
2254 * BLK_STS_RESOURCE is usually returned.
2256 * We do not guarantee that dispatch can be drained or blocked
2257 * after blk_mq_stop_hw_queues() returns. Please use
2258 * blk_mq_quiesce_queue() for that requirement.
2260 void blk_mq_stop_hw_queues(struct request_queue *q)
2262 struct blk_mq_hw_ctx *hctx;
2265 queue_for_each_hw_ctx(q, hctx, i)
2266 blk_mq_stop_hw_queue(hctx);
2268 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2270 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2272 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2274 blk_mq_run_hw_queue(hctx, false);
2276 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2278 void blk_mq_start_hw_queues(struct request_queue *q)
2280 struct blk_mq_hw_ctx *hctx;
2283 queue_for_each_hw_ctx(q, hctx, i)
2284 blk_mq_start_hw_queue(hctx);
2286 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2288 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2290 if (!blk_mq_hctx_stopped(hctx))
2293 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2294 blk_mq_run_hw_queue(hctx, async);
2296 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2298 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2300 struct blk_mq_hw_ctx *hctx;
2303 queue_for_each_hw_ctx(q, hctx, i)
2304 blk_mq_start_stopped_hw_queue(hctx, async);
2306 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2308 static void blk_mq_run_work_fn(struct work_struct *work)
2310 struct blk_mq_hw_ctx *hctx;
2312 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2315 * If we are stopped, don't run the queue.
2317 if (blk_mq_hctx_stopped(hctx))
2320 __blk_mq_run_hw_queue(hctx);
2323 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2327 struct blk_mq_ctx *ctx = rq->mq_ctx;
2328 enum hctx_type type = hctx->type;
2330 lockdep_assert_held(&ctx->lock);
2332 trace_block_rq_insert(rq);
2335 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2337 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2340 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2343 struct blk_mq_ctx *ctx = rq->mq_ctx;
2345 lockdep_assert_held(&ctx->lock);
2347 __blk_mq_insert_req_list(hctx, rq, at_head);
2348 blk_mq_hctx_mark_pending(hctx, ctx);
2352 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2353 * @rq: Pointer to request to be inserted.
2354 * @at_head: true if the request should be inserted at the head of the list.
2355 * @run_queue: If we should run the hardware queue after inserting the request.
2357 * Should only be used carefully, when the caller knows we want to
2358 * bypass a potential IO scheduler on the target device.
2360 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2363 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2365 spin_lock(&hctx->lock);
2367 list_add(&rq->queuelist, &hctx->dispatch);
2369 list_add_tail(&rq->queuelist, &hctx->dispatch);
2370 spin_unlock(&hctx->lock);
2373 blk_mq_run_hw_queue(hctx, false);
2376 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2377 struct list_head *list)
2381 enum hctx_type type = hctx->type;
2384 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2387 list_for_each_entry(rq, list, queuelist) {
2388 BUG_ON(rq->mq_ctx != ctx);
2389 trace_block_rq_insert(rq);
2392 spin_lock(&ctx->lock);
2393 list_splice_tail_init(list, &ctx->rq_lists[type]);
2394 blk_mq_hctx_mark_pending(hctx, ctx);
2395 spin_unlock(&ctx->lock);
2398 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2401 if (hctx->queue->mq_ops->commit_rqs) {
2402 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2403 hctx->queue->mq_ops->commit_rqs(hctx);
2408 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2409 unsigned int nr_segs)
2413 if (bio->bi_opf & REQ_RAHEAD)
2414 rq->cmd_flags |= REQ_FAILFAST_MASK;
2416 rq->__sector = bio->bi_iter.bi_sector;
2417 blk_rq_bio_prep(rq, bio, nr_segs);
2419 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2420 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2423 blk_account_io_start(rq);
2426 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2427 struct request *rq, bool last)
2429 struct request_queue *q = rq->q;
2430 struct blk_mq_queue_data bd = {
2437 * For OK queue, we are done. For error, caller may kill it.
2438 * Any other error (busy), just add it to our list as we
2439 * previously would have done.
2441 ret = q->mq_ops->queue_rq(hctx, &bd);
2444 blk_mq_update_dispatch_busy(hctx, false);
2446 case BLK_STS_RESOURCE:
2447 case BLK_STS_DEV_RESOURCE:
2448 blk_mq_update_dispatch_busy(hctx, true);
2449 __blk_mq_requeue_request(rq);
2452 blk_mq_update_dispatch_busy(hctx, false);
2459 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2461 bool bypass_insert, bool last)
2463 struct request_queue *q = rq->q;
2464 bool run_queue = true;
2468 * RCU or SRCU read lock is needed before checking quiesced flag.
2470 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2471 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2472 * and avoid driver to try to dispatch again.
2474 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2476 bypass_insert = false;
2480 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2483 budget_token = blk_mq_get_dispatch_budget(q);
2484 if (budget_token < 0)
2487 blk_mq_set_rq_budget_token(rq, budget_token);
2489 if (!blk_mq_get_driver_tag(rq)) {
2490 blk_mq_put_dispatch_budget(q, budget_token);
2494 return __blk_mq_issue_directly(hctx, rq, last);
2497 return BLK_STS_RESOURCE;
2499 blk_mq_sched_insert_request(rq, false, run_queue, false);
2505 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2506 * @hctx: Pointer of the associated hardware queue.
2507 * @rq: Pointer to request to be sent.
2509 * If the device has enough resources to accept a new request now, send the
2510 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2511 * we can try send it another time in the future. Requests inserted at this
2512 * queue have higher priority.
2514 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2518 __blk_mq_try_issue_directly(hctx, rq, false, true);
2520 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2521 blk_mq_request_bypass_insert(rq, false, true);
2522 else if (ret != BLK_STS_OK)
2523 blk_mq_end_request(rq, ret);
2526 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2528 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2531 static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2533 struct blk_mq_hw_ctx *hctx = NULL;
2538 while ((rq = rq_list_pop(&plug->mq_list))) {
2539 bool last = rq_list_empty(plug->mq_list);
2542 if (hctx != rq->mq_hctx) {
2544 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2548 ret = blk_mq_request_issue_directly(rq, last);
2553 case BLK_STS_RESOURCE:
2554 case BLK_STS_DEV_RESOURCE:
2555 blk_mq_request_bypass_insert(rq, false, true);
2556 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2559 blk_mq_end_request(rq, ret);
2566 * If we didn't flush the entire list, we could have told the driver
2567 * there was more coming, but that turned out to be a lie.
2570 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2573 static void __blk_mq_flush_plug_list(struct request_queue *q,
2574 struct blk_plug *plug)
2576 if (blk_queue_quiesced(q))
2578 q->mq_ops->queue_rqs(&plug->mq_list);
2581 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2583 struct blk_mq_hw_ctx *this_hctx = NULL;
2584 struct blk_mq_ctx *this_ctx = NULL;
2585 struct request *requeue_list = NULL;
2586 unsigned int depth = 0;
2590 struct request *rq = rq_list_pop(&plug->mq_list);
2593 this_hctx = rq->mq_hctx;
2594 this_ctx = rq->mq_ctx;
2595 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2596 rq_list_add(&requeue_list, rq);
2599 list_add_tail(&rq->queuelist, &list);
2601 } while (!rq_list_empty(plug->mq_list));
2603 plug->mq_list = requeue_list;
2604 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2605 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched);
2608 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2612 if (rq_list_empty(plug->mq_list))
2616 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2617 struct request_queue *q;
2619 rq = rq_list_peek(&plug->mq_list);
2623 * Peek first request and see if we have a ->queue_rqs() hook.
2624 * If we do, we can dispatch the whole plug list in one go. We
2625 * already know at this point that all requests belong to the
2626 * same queue, caller must ensure that's the case.
2628 * Since we pass off the full list to the driver at this point,
2629 * we do not increment the active request count for the queue.
2630 * Bypass shared tags for now because of that.
2632 if (q->mq_ops->queue_rqs &&
2633 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2634 blk_mq_run_dispatch_ops(q,
2635 __blk_mq_flush_plug_list(q, plug));
2636 if (rq_list_empty(plug->mq_list))
2640 blk_mq_run_dispatch_ops(q,
2641 blk_mq_plug_issue_direct(plug, false));
2642 if (rq_list_empty(plug->mq_list))
2647 blk_mq_dispatch_plug_list(plug, from_schedule);
2648 } while (!rq_list_empty(plug->mq_list));
2651 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2652 struct list_head *list)
2657 while (!list_empty(list)) {
2659 struct request *rq = list_first_entry(list, struct request,
2662 list_del_init(&rq->queuelist);
2663 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2664 if (ret != BLK_STS_OK) {
2666 if (ret == BLK_STS_RESOURCE ||
2667 ret == BLK_STS_DEV_RESOURCE) {
2668 blk_mq_request_bypass_insert(rq, false,
2672 blk_mq_end_request(rq, ret);
2678 * If we didn't flush the entire list, we could have told
2679 * the driver there was more coming, but that turned out to
2682 if ((!list_empty(list) || errors) &&
2683 hctx->queue->mq_ops->commit_rqs && queued)
2684 hctx->queue->mq_ops->commit_rqs(hctx);
2687 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2688 struct bio *bio, unsigned int nr_segs)
2690 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2691 if (blk_attempt_plug_merge(q, bio, nr_segs))
2693 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2699 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2700 struct blk_plug *plug,
2704 struct blk_mq_alloc_data data = {
2707 .cmd_flags = bio->bi_opf,
2711 if (unlikely(bio_queue_enter(bio)))
2714 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2717 rq_qos_throttle(q, bio);
2720 data.nr_tags = plug->nr_ios;
2722 data.cached_rq = &plug->cached_rq;
2725 rq = __blk_mq_alloc_requests(&data);
2728 rq_qos_cleanup(q, bio);
2729 if (bio->bi_opf & REQ_NOWAIT)
2730 bio_wouldblock_error(bio);
2736 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2737 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2743 rq = rq_list_peek(&plug->cached_rq);
2744 if (!rq || rq->q != q)
2747 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2752 if (blk_mq_get_hctx_type((*bio)->bi_opf) != rq->mq_hctx->type)
2754 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2758 * If any qos ->throttle() end up blocking, we will have flushed the
2759 * plug and hence killed the cached_rq list as well. Pop this entry
2760 * before we throttle.
2762 plug->cached_rq = rq_list_next(rq);
2763 rq_qos_throttle(q, *bio);
2765 rq->cmd_flags = (*bio)->bi_opf;
2766 INIT_LIST_HEAD(&rq->queuelist);
2770 static void bio_set_ioprio(struct bio *bio)
2772 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2773 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2774 bio->bi_ioprio = get_current_ioprio();
2775 blkcg_set_ioprio(bio);
2779 * blk_mq_submit_bio - Create and send a request to block device.
2780 * @bio: Bio pointer.
2782 * Builds up a request structure from @q and @bio and send to the device. The
2783 * request may not be queued directly to hardware if:
2784 * * This request can be merged with another one
2785 * * We want to place request at plug queue for possible future merging
2786 * * There is an IO scheduler active at this queue
2788 * It will not queue the request if there is an error with the bio, or at the
2791 void blk_mq_submit_bio(struct bio *bio)
2793 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2794 struct blk_plug *plug = blk_mq_plug(bio);
2795 const int is_sync = op_is_sync(bio->bi_opf);
2797 unsigned int nr_segs = 1;
2800 bio = blk_queue_bounce(bio, q);
2801 if (bio_may_exceed_limits(bio, &q->limits))
2802 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2804 if (!bio_integrity_prep(bio))
2807 bio_set_ioprio(bio);
2809 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2813 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2818 trace_block_getrq(bio);
2820 rq_qos_track(q, rq, bio);
2822 blk_mq_bio_to_request(rq, bio, nr_segs);
2824 ret = blk_crypto_init_request(rq);
2825 if (ret != BLK_STS_OK) {
2826 bio->bi_status = ret;
2828 blk_mq_free_request(rq);
2832 if (op_is_flush(bio->bi_opf)) {
2833 blk_insert_flush(rq);
2838 blk_add_rq_to_plug(plug, rq);
2839 else if ((rq->rq_flags & RQF_ELV) ||
2840 (rq->mq_hctx->dispatch_busy &&
2841 (q->nr_hw_queues == 1 || !is_sync)))
2842 blk_mq_sched_insert_request(rq, false, true, true);
2844 blk_mq_run_dispatch_ops(rq->q,
2845 blk_mq_try_issue_directly(rq->mq_hctx, rq));
2848 #ifdef CONFIG_BLK_MQ_STACKING
2850 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
2851 * @rq: the request being queued
2853 blk_status_t blk_insert_cloned_request(struct request *rq)
2855 struct request_queue *q = rq->q;
2856 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
2859 if (blk_rq_sectors(rq) > max_sectors) {
2861 * SCSI device does not have a good way to return if
2862 * Write Same/Zero is actually supported. If a device rejects
2863 * a non-read/write command (discard, write same,etc.) the
2864 * low-level device driver will set the relevant queue limit to
2865 * 0 to prevent blk-lib from issuing more of the offending
2866 * operations. Commands queued prior to the queue limit being
2867 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
2868 * errors being propagated to upper layers.
2870 if (max_sectors == 0)
2871 return BLK_STS_NOTSUPP;
2873 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
2874 __func__, blk_rq_sectors(rq), max_sectors);
2875 return BLK_STS_IOERR;
2879 * The queue settings related to segment counting may differ from the
2882 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
2883 if (rq->nr_phys_segments > queue_max_segments(q)) {
2884 printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n",
2885 __func__, rq->nr_phys_segments, queue_max_segments(q));
2886 return BLK_STS_IOERR;
2889 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
2890 return BLK_STS_IOERR;
2892 if (blk_crypto_insert_cloned_request(rq))
2893 return BLK_STS_IOERR;
2895 blk_account_io_start(rq);
2898 * Since we have a scheduler attached on the top device,
2899 * bypass a potential scheduler on the bottom device for
2902 blk_mq_run_dispatch_ops(q,
2903 ret = blk_mq_request_issue_directly(rq, true));
2905 blk_account_io_done(rq, ktime_get_ns());
2908 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
2911 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
2912 * @rq: the clone request to be cleaned up
2915 * Free all bios in @rq for a cloned request.
2917 void blk_rq_unprep_clone(struct request *rq)
2921 while ((bio = rq->bio) != NULL) {
2922 rq->bio = bio->bi_next;
2927 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
2930 * blk_rq_prep_clone - Helper function to setup clone request
2931 * @rq: the request to be setup
2932 * @rq_src: original request to be cloned
2933 * @bs: bio_set that bios for clone are allocated from
2934 * @gfp_mask: memory allocation mask for bio
2935 * @bio_ctr: setup function to be called for each clone bio.
2936 * Returns %0 for success, non %0 for failure.
2937 * @data: private data to be passed to @bio_ctr
2940 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
2941 * Also, pages which the original bios are pointing to are not copied
2942 * and the cloned bios just point same pages.
2943 * So cloned bios must be completed before original bios, which means
2944 * the caller must complete @rq before @rq_src.
2946 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
2947 struct bio_set *bs, gfp_t gfp_mask,
2948 int (*bio_ctr)(struct bio *, struct bio *, void *),
2951 struct bio *bio, *bio_src;
2956 __rq_for_each_bio(bio_src, rq_src) {
2957 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
2962 if (bio_ctr && bio_ctr(bio, bio_src, data))
2966 rq->biotail->bi_next = bio;
2969 rq->bio = rq->biotail = bio;
2974 /* Copy attributes of the original request to the clone request. */
2975 rq->__sector = blk_rq_pos(rq_src);
2976 rq->__data_len = blk_rq_bytes(rq_src);
2977 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
2978 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
2979 rq->special_vec = rq_src->special_vec;
2981 rq->nr_phys_segments = rq_src->nr_phys_segments;
2982 rq->ioprio = rq_src->ioprio;
2984 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
2992 blk_rq_unprep_clone(rq);
2996 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
2997 #endif /* CONFIG_BLK_MQ_STACKING */
3000 * Steal bios from a request and add them to a bio list.
3001 * The request must not have been partially completed before.
3003 void blk_steal_bios(struct bio_list *list, struct request *rq)
3007 list->tail->bi_next = rq->bio;
3009 list->head = rq->bio;
3010 list->tail = rq->biotail;
3018 EXPORT_SYMBOL_GPL(blk_steal_bios);
3020 static size_t order_to_size(unsigned int order)
3022 return (size_t)PAGE_SIZE << order;
3025 /* called before freeing request pool in @tags */
3026 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3027 struct blk_mq_tags *tags)
3030 unsigned long flags;
3033 * There is no need to clear mapping if driver tags is not initialized
3034 * or the mapping belongs to the driver tags.
3036 if (!drv_tags || drv_tags == tags)
3039 list_for_each_entry(page, &tags->page_list, lru) {
3040 unsigned long start = (unsigned long)page_address(page);
3041 unsigned long end = start + order_to_size(page->private);
3044 for (i = 0; i < drv_tags->nr_tags; i++) {
3045 struct request *rq = drv_tags->rqs[i];
3046 unsigned long rq_addr = (unsigned long)rq;
3048 if (rq_addr >= start && rq_addr < end) {
3049 WARN_ON_ONCE(req_ref_read(rq) != 0);
3050 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3056 * Wait until all pending iteration is done.
3058 * Request reference is cleared and it is guaranteed to be observed
3059 * after the ->lock is released.
3061 spin_lock_irqsave(&drv_tags->lock, flags);
3062 spin_unlock_irqrestore(&drv_tags->lock, flags);
3065 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3066 unsigned int hctx_idx)
3068 struct blk_mq_tags *drv_tags;
3071 if (list_empty(&tags->page_list))
3074 if (blk_mq_is_shared_tags(set->flags))
3075 drv_tags = set->shared_tags;
3077 drv_tags = set->tags[hctx_idx];
3079 if (tags->static_rqs && set->ops->exit_request) {
3082 for (i = 0; i < tags->nr_tags; i++) {
3083 struct request *rq = tags->static_rqs[i];
3087 set->ops->exit_request(set, rq, hctx_idx);
3088 tags->static_rqs[i] = NULL;
3092 blk_mq_clear_rq_mapping(drv_tags, tags);
3094 while (!list_empty(&tags->page_list)) {
3095 page = list_first_entry(&tags->page_list, struct page, lru);
3096 list_del_init(&page->lru);
3098 * Remove kmemleak object previously allocated in
3099 * blk_mq_alloc_rqs().
3101 kmemleak_free(page_address(page));
3102 __free_pages(page, page->private);
3106 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3110 kfree(tags->static_rqs);
3111 tags->static_rqs = NULL;
3113 blk_mq_free_tags(tags);
3116 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3117 unsigned int hctx_idx)
3121 for (i = 0; i < set->nr_maps; i++) {
3122 unsigned int start = set->map[i].queue_offset;
3123 unsigned int end = start + set->map[i].nr_queues;
3125 if (hctx_idx >= start && hctx_idx < end)
3129 if (i >= set->nr_maps)
3130 i = HCTX_TYPE_DEFAULT;
3135 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3136 unsigned int hctx_idx)
3138 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3140 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3143 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3144 unsigned int hctx_idx,
3145 unsigned int nr_tags,
3146 unsigned int reserved_tags)
3148 int node = blk_mq_get_hctx_node(set, hctx_idx);
3149 struct blk_mq_tags *tags;
3151 if (node == NUMA_NO_NODE)
3152 node = set->numa_node;
3154 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3155 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3159 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3160 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3163 blk_mq_free_tags(tags);
3167 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3168 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3170 if (!tags->static_rqs) {
3172 blk_mq_free_tags(tags);
3179 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3180 unsigned int hctx_idx, int node)
3184 if (set->ops->init_request) {
3185 ret = set->ops->init_request(set, rq, hctx_idx, node);
3190 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3194 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3195 struct blk_mq_tags *tags,
3196 unsigned int hctx_idx, unsigned int depth)
3198 unsigned int i, j, entries_per_page, max_order = 4;
3199 int node = blk_mq_get_hctx_node(set, hctx_idx);
3200 size_t rq_size, left;
3202 if (node == NUMA_NO_NODE)
3203 node = set->numa_node;
3205 INIT_LIST_HEAD(&tags->page_list);
3208 * rq_size is the size of the request plus driver payload, rounded
3209 * to the cacheline size
3211 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3213 left = rq_size * depth;
3215 for (i = 0; i < depth; ) {
3216 int this_order = max_order;
3221 while (this_order && left < order_to_size(this_order - 1))
3225 page = alloc_pages_node(node,
3226 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3232 if (order_to_size(this_order) < rq_size)
3239 page->private = this_order;
3240 list_add_tail(&page->lru, &tags->page_list);
3242 p = page_address(page);
3244 * Allow kmemleak to scan these pages as they contain pointers
3245 * to additional allocations like via ops->init_request().
3247 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3248 entries_per_page = order_to_size(this_order) / rq_size;
3249 to_do = min(entries_per_page, depth - i);
3250 left -= to_do * rq_size;
3251 for (j = 0; j < to_do; j++) {
3252 struct request *rq = p;
3254 tags->static_rqs[i] = rq;
3255 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3256 tags->static_rqs[i] = NULL;
3267 blk_mq_free_rqs(set, tags, hctx_idx);
3271 struct rq_iter_data {
3272 struct blk_mq_hw_ctx *hctx;
3276 static bool blk_mq_has_request(struct request *rq, void *data)
3278 struct rq_iter_data *iter_data = data;
3280 if (rq->mq_hctx != iter_data->hctx)
3282 iter_data->has_rq = true;
3286 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3288 struct blk_mq_tags *tags = hctx->sched_tags ?
3289 hctx->sched_tags : hctx->tags;
3290 struct rq_iter_data data = {
3294 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3298 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3299 struct blk_mq_hw_ctx *hctx)
3301 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3303 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3308 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3310 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3311 struct blk_mq_hw_ctx, cpuhp_online);
3313 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3314 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3318 * Prevent new request from being allocated on the current hctx.
3320 * The smp_mb__after_atomic() Pairs with the implied barrier in
3321 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3322 * seen once we return from the tag allocator.
3324 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3325 smp_mb__after_atomic();
3328 * Try to grab a reference to the queue and wait for any outstanding
3329 * requests. If we could not grab a reference the queue has been
3330 * frozen and there are no requests.
3332 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3333 while (blk_mq_hctx_has_requests(hctx))
3335 percpu_ref_put(&hctx->queue->q_usage_counter);
3341 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3343 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3344 struct blk_mq_hw_ctx, cpuhp_online);
3346 if (cpumask_test_cpu(cpu, hctx->cpumask))
3347 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3352 * 'cpu' is going away. splice any existing rq_list entries from this
3353 * software queue to the hw queue dispatch list, and ensure that it
3356 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3358 struct blk_mq_hw_ctx *hctx;
3359 struct blk_mq_ctx *ctx;
3361 enum hctx_type type;
3363 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3364 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3367 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3370 spin_lock(&ctx->lock);
3371 if (!list_empty(&ctx->rq_lists[type])) {
3372 list_splice_init(&ctx->rq_lists[type], &tmp);
3373 blk_mq_hctx_clear_pending(hctx, ctx);
3375 spin_unlock(&ctx->lock);
3377 if (list_empty(&tmp))
3380 spin_lock(&hctx->lock);
3381 list_splice_tail_init(&tmp, &hctx->dispatch);
3382 spin_unlock(&hctx->lock);
3384 blk_mq_run_hw_queue(hctx, true);
3388 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3390 if (!(hctx->flags & BLK_MQ_F_STACKING))
3391 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3392 &hctx->cpuhp_online);
3393 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3398 * Before freeing hw queue, clearing the flush request reference in
3399 * tags->rqs[] for avoiding potential UAF.
3401 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3402 unsigned int queue_depth, struct request *flush_rq)
3405 unsigned long flags;
3407 /* The hw queue may not be mapped yet */
3411 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3413 for (i = 0; i < queue_depth; i++)
3414 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3417 * Wait until all pending iteration is done.
3419 * Request reference is cleared and it is guaranteed to be observed
3420 * after the ->lock is released.
3422 spin_lock_irqsave(&tags->lock, flags);
3423 spin_unlock_irqrestore(&tags->lock, flags);
3426 /* hctx->ctxs will be freed in queue's release handler */
3427 static void blk_mq_exit_hctx(struct request_queue *q,
3428 struct blk_mq_tag_set *set,
3429 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3431 struct request *flush_rq = hctx->fq->flush_rq;
3433 if (blk_mq_hw_queue_mapped(hctx))
3434 blk_mq_tag_idle(hctx);
3436 if (blk_queue_init_done(q))
3437 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3438 set->queue_depth, flush_rq);
3439 if (set->ops->exit_request)
3440 set->ops->exit_request(set, flush_rq, hctx_idx);
3442 if (set->ops->exit_hctx)
3443 set->ops->exit_hctx(hctx, hctx_idx);
3445 blk_mq_remove_cpuhp(hctx);
3447 xa_erase(&q->hctx_table, hctx_idx);
3449 spin_lock(&q->unused_hctx_lock);
3450 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3451 spin_unlock(&q->unused_hctx_lock);
3454 static void blk_mq_exit_hw_queues(struct request_queue *q,
3455 struct blk_mq_tag_set *set, int nr_queue)
3457 struct blk_mq_hw_ctx *hctx;
3460 queue_for_each_hw_ctx(q, hctx, i) {
3463 blk_mq_exit_hctx(q, set, hctx, i);
3467 static int blk_mq_init_hctx(struct request_queue *q,
3468 struct blk_mq_tag_set *set,
3469 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3471 hctx->queue_num = hctx_idx;
3473 if (!(hctx->flags & BLK_MQ_F_STACKING))
3474 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3475 &hctx->cpuhp_online);
3476 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3478 hctx->tags = set->tags[hctx_idx];
3480 if (set->ops->init_hctx &&
3481 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3482 goto unregister_cpu_notifier;
3484 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3488 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3494 if (set->ops->exit_request)
3495 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3497 if (set->ops->exit_hctx)
3498 set->ops->exit_hctx(hctx, hctx_idx);
3499 unregister_cpu_notifier:
3500 blk_mq_remove_cpuhp(hctx);
3504 static struct blk_mq_hw_ctx *
3505 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3508 struct blk_mq_hw_ctx *hctx;
3509 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3511 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3513 goto fail_alloc_hctx;
3515 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3518 atomic_set(&hctx->nr_active, 0);
3519 if (node == NUMA_NO_NODE)
3520 node = set->numa_node;
3521 hctx->numa_node = node;
3523 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3524 spin_lock_init(&hctx->lock);
3525 INIT_LIST_HEAD(&hctx->dispatch);
3527 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3529 INIT_LIST_HEAD(&hctx->hctx_list);
3532 * Allocate space for all possible cpus to avoid allocation at
3535 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3540 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3541 gfp, node, false, false))
3545 spin_lock_init(&hctx->dispatch_wait_lock);
3546 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3547 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3549 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3553 blk_mq_hctx_kobj_init(hctx);
3558 sbitmap_free(&hctx->ctx_map);
3562 free_cpumask_var(hctx->cpumask);
3569 static void blk_mq_init_cpu_queues(struct request_queue *q,
3570 unsigned int nr_hw_queues)
3572 struct blk_mq_tag_set *set = q->tag_set;
3575 for_each_possible_cpu(i) {
3576 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3577 struct blk_mq_hw_ctx *hctx;
3581 spin_lock_init(&__ctx->lock);
3582 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3583 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3588 * Set local node, IFF we have more than one hw queue. If
3589 * not, we remain on the home node of the device
3591 for (j = 0; j < set->nr_maps; j++) {
3592 hctx = blk_mq_map_queue_type(q, j, i);
3593 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3594 hctx->numa_node = cpu_to_node(i);
3599 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3600 unsigned int hctx_idx,
3603 struct blk_mq_tags *tags;
3606 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3610 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3612 blk_mq_free_rq_map(tags);
3619 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3622 if (blk_mq_is_shared_tags(set->flags)) {
3623 set->tags[hctx_idx] = set->shared_tags;
3628 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3631 return set->tags[hctx_idx];
3634 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3635 struct blk_mq_tags *tags,
3636 unsigned int hctx_idx)
3639 blk_mq_free_rqs(set, tags, hctx_idx);
3640 blk_mq_free_rq_map(tags);
3644 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3645 unsigned int hctx_idx)
3647 if (!blk_mq_is_shared_tags(set->flags))
3648 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3650 set->tags[hctx_idx] = NULL;
3653 static void blk_mq_map_swqueue(struct request_queue *q)
3655 unsigned int j, hctx_idx;
3657 struct blk_mq_hw_ctx *hctx;
3658 struct blk_mq_ctx *ctx;
3659 struct blk_mq_tag_set *set = q->tag_set;
3661 queue_for_each_hw_ctx(q, hctx, i) {
3662 cpumask_clear(hctx->cpumask);
3664 hctx->dispatch_from = NULL;
3668 * Map software to hardware queues.
3670 * If the cpu isn't present, the cpu is mapped to first hctx.
3672 for_each_possible_cpu(i) {
3674 ctx = per_cpu_ptr(q->queue_ctx, i);
3675 for (j = 0; j < set->nr_maps; j++) {
3676 if (!set->map[j].nr_queues) {
3677 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3678 HCTX_TYPE_DEFAULT, i);
3681 hctx_idx = set->map[j].mq_map[i];
3682 /* unmapped hw queue can be remapped after CPU topo changed */
3683 if (!set->tags[hctx_idx] &&
3684 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3686 * If tags initialization fail for some hctx,
3687 * that hctx won't be brought online. In this
3688 * case, remap the current ctx to hctx[0] which
3689 * is guaranteed to always have tags allocated
3691 set->map[j].mq_map[i] = 0;
3694 hctx = blk_mq_map_queue_type(q, j, i);
3695 ctx->hctxs[j] = hctx;
3697 * If the CPU is already set in the mask, then we've
3698 * mapped this one already. This can happen if
3699 * devices share queues across queue maps.
3701 if (cpumask_test_cpu(i, hctx->cpumask))
3704 cpumask_set_cpu(i, hctx->cpumask);
3706 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3707 hctx->ctxs[hctx->nr_ctx++] = ctx;
3710 * If the nr_ctx type overflows, we have exceeded the
3711 * amount of sw queues we can support.
3713 BUG_ON(!hctx->nr_ctx);
3716 for (; j < HCTX_MAX_TYPES; j++)
3717 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3718 HCTX_TYPE_DEFAULT, i);
3721 queue_for_each_hw_ctx(q, hctx, i) {
3723 * If no software queues are mapped to this hardware queue,
3724 * disable it and free the request entries.
3726 if (!hctx->nr_ctx) {
3727 /* Never unmap queue 0. We need it as a
3728 * fallback in case of a new remap fails
3732 __blk_mq_free_map_and_rqs(set, i);
3738 hctx->tags = set->tags[i];
3739 WARN_ON(!hctx->tags);
3742 * Set the map size to the number of mapped software queues.
3743 * This is more accurate and more efficient than looping
3744 * over all possibly mapped software queues.
3746 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3749 * Initialize batch roundrobin counts
3751 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3752 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3757 * Caller needs to ensure that we're either frozen/quiesced, or that
3758 * the queue isn't live yet.
3760 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3762 struct blk_mq_hw_ctx *hctx;
3765 queue_for_each_hw_ctx(q, hctx, i) {
3767 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3769 blk_mq_tag_idle(hctx);
3770 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3775 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3778 struct request_queue *q;
3780 lockdep_assert_held(&set->tag_list_lock);
3782 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3783 blk_mq_freeze_queue(q);
3784 queue_set_hctx_shared(q, shared);
3785 blk_mq_unfreeze_queue(q);
3789 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3791 struct blk_mq_tag_set *set = q->tag_set;
3793 mutex_lock(&set->tag_list_lock);
3794 list_del(&q->tag_set_list);
3795 if (list_is_singular(&set->tag_list)) {
3796 /* just transitioned to unshared */
3797 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3798 /* update existing queue */
3799 blk_mq_update_tag_set_shared(set, false);
3801 mutex_unlock(&set->tag_list_lock);
3802 INIT_LIST_HEAD(&q->tag_set_list);
3805 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3806 struct request_queue *q)
3808 mutex_lock(&set->tag_list_lock);
3811 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3813 if (!list_empty(&set->tag_list) &&
3814 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3815 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3816 /* update existing queue */
3817 blk_mq_update_tag_set_shared(set, true);
3819 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3820 queue_set_hctx_shared(q, true);
3821 list_add_tail(&q->tag_set_list, &set->tag_list);
3823 mutex_unlock(&set->tag_list_lock);
3826 /* All allocations will be freed in release handler of q->mq_kobj */
3827 static int blk_mq_alloc_ctxs(struct request_queue *q)
3829 struct blk_mq_ctxs *ctxs;
3832 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3836 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3837 if (!ctxs->queue_ctx)
3840 for_each_possible_cpu(cpu) {
3841 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3845 q->mq_kobj = &ctxs->kobj;
3846 q->queue_ctx = ctxs->queue_ctx;
3855 * It is the actual release handler for mq, but we do it from
3856 * request queue's release handler for avoiding use-after-free
3857 * and headache because q->mq_kobj shouldn't have been introduced,
3858 * but we can't group ctx/kctx kobj without it.
3860 void blk_mq_release(struct request_queue *q)
3862 struct blk_mq_hw_ctx *hctx, *next;
3865 queue_for_each_hw_ctx(q, hctx, i)
3866 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3868 /* all hctx are in .unused_hctx_list now */
3869 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3870 list_del_init(&hctx->hctx_list);
3871 kobject_put(&hctx->kobj);
3874 xa_destroy(&q->hctx_table);
3877 * release .mq_kobj and sw queue's kobject now because
3878 * both share lifetime with request queue.
3880 blk_mq_sysfs_deinit(q);
3883 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3886 struct request_queue *q;
3889 q = blk_alloc_queue(set->numa_node, set->flags & BLK_MQ_F_BLOCKING);
3891 return ERR_PTR(-ENOMEM);
3892 q->queuedata = queuedata;
3893 ret = blk_mq_init_allocated_queue(set, q);
3896 return ERR_PTR(ret);
3901 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3903 return blk_mq_init_queue_data(set, NULL);
3905 EXPORT_SYMBOL(blk_mq_init_queue);
3908 * blk_mq_destroy_queue - shutdown a request queue
3909 * @q: request queue to shutdown
3911 * This shuts down a request queue allocated by blk_mq_init_queue() and drops
3912 * the initial reference. All future requests will failed with -ENODEV.
3914 * Context: can sleep
3916 void blk_mq_destroy_queue(struct request_queue *q)
3918 WARN_ON_ONCE(!queue_is_mq(q));
3919 WARN_ON_ONCE(blk_queue_registered(q));
3923 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
3924 blk_queue_start_drain(q);
3925 blk_freeze_queue(q);
3928 blk_mq_cancel_work_sync(q);
3929 blk_mq_exit_queue(q);
3931 /* @q is and will stay empty, shutdown and put */
3934 EXPORT_SYMBOL(blk_mq_destroy_queue);
3936 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
3937 struct lock_class_key *lkclass)
3939 struct request_queue *q;
3940 struct gendisk *disk;
3942 q = blk_mq_init_queue_data(set, queuedata);
3946 disk = __alloc_disk_node(q, set->numa_node, lkclass);
3948 blk_mq_destroy_queue(q);
3949 return ERR_PTR(-ENOMEM);
3951 set_bit(GD_OWNS_QUEUE, &disk->state);
3954 EXPORT_SYMBOL(__blk_mq_alloc_disk);
3956 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
3957 struct lock_class_key *lkclass)
3959 struct gendisk *disk;
3961 if (!blk_get_queue(q))
3963 disk = __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);
4115 blk_stat_free_callback(q->poll_cb);
4121 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4123 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4124 void blk_mq_exit_queue(struct request_queue *q)
4126 struct blk_mq_tag_set *set = q->tag_set;
4128 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4129 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4130 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4131 blk_mq_del_queue_tag_set(q);
4134 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4138 if (blk_mq_is_shared_tags(set->flags)) {
4139 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4142 if (!set->shared_tags)
4146 for (i = 0; i < set->nr_hw_queues; i++) {
4147 if (!__blk_mq_alloc_map_and_rqs(set, i))
4156 __blk_mq_free_map_and_rqs(set, i);
4158 if (blk_mq_is_shared_tags(set->flags)) {
4159 blk_mq_free_map_and_rqs(set, set->shared_tags,
4160 BLK_MQ_NO_HCTX_IDX);
4167 * Allocate the request maps associated with this tag_set. Note that this
4168 * may reduce the depth asked for, if memory is tight. set->queue_depth
4169 * will be updated to reflect the allocated depth.
4171 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4176 depth = set->queue_depth;
4178 err = __blk_mq_alloc_rq_maps(set);
4182 set->queue_depth >>= 1;
4183 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4187 } while (set->queue_depth);
4189 if (!set->queue_depth || err) {
4190 pr_err("blk-mq: failed to allocate request map\n");
4194 if (depth != set->queue_depth)
4195 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4196 depth, set->queue_depth);
4201 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4204 * blk_mq_map_queues() and multiple .map_queues() implementations
4205 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4206 * number of hardware queues.
4208 if (set->nr_maps == 1)
4209 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4211 if (set->ops->map_queues && !is_kdump_kernel()) {
4215 * transport .map_queues is usually done in the following
4218 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4219 * mask = get_cpu_mask(queue)
4220 * for_each_cpu(cpu, mask)
4221 * set->map[x].mq_map[cpu] = queue;
4224 * When we need to remap, the table has to be cleared for
4225 * killing stale mapping since one CPU may not be mapped
4228 for (i = 0; i < set->nr_maps; i++)
4229 blk_mq_clear_mq_map(&set->map[i]);
4231 return set->ops->map_queues(set);
4233 BUG_ON(set->nr_maps > 1);
4234 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4238 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4239 int cur_nr_hw_queues, int new_nr_hw_queues)
4241 struct blk_mq_tags **new_tags;
4243 if (cur_nr_hw_queues >= new_nr_hw_queues)
4246 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4247 GFP_KERNEL, set->numa_node);
4252 memcpy(new_tags, set->tags, cur_nr_hw_queues *
4253 sizeof(*set->tags));
4255 set->tags = new_tags;
4256 set->nr_hw_queues = new_nr_hw_queues;
4261 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
4262 int new_nr_hw_queues)
4264 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
4268 * Alloc a tag set to be associated with one or more request queues.
4269 * May fail with EINVAL for various error conditions. May adjust the
4270 * requested depth down, if it's too large. In that case, the set
4271 * value will be stored in set->queue_depth.
4273 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4277 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4279 if (!set->nr_hw_queues)
4281 if (!set->queue_depth)
4283 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4286 if (!set->ops->queue_rq)
4289 if (!set->ops->get_budget ^ !set->ops->put_budget)
4292 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4293 pr_info("blk-mq: reduced tag depth to %u\n",
4295 set->queue_depth = BLK_MQ_MAX_DEPTH;
4300 else if (set->nr_maps > HCTX_MAX_TYPES)
4304 * If a crashdump is active, then we are potentially in a very
4305 * memory constrained environment. Limit us to 1 queue and
4306 * 64 tags to prevent using too much memory.
4308 if (is_kdump_kernel()) {
4309 set->nr_hw_queues = 1;
4311 set->queue_depth = min(64U, set->queue_depth);
4314 * There is no use for more h/w queues than cpus if we just have
4317 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4318 set->nr_hw_queues = nr_cpu_ids;
4320 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
4324 for (i = 0; i < set->nr_maps; i++) {
4325 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4326 sizeof(set->map[i].mq_map[0]),
4327 GFP_KERNEL, set->numa_node);
4328 if (!set->map[i].mq_map)
4329 goto out_free_mq_map;
4330 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4333 ret = blk_mq_update_queue_map(set);
4335 goto out_free_mq_map;
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);