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"
46 static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
48 static void blk_mq_poll_stats_start(struct request_queue *q);
49 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
51 static int blk_mq_poll_stats_bkt(const struct request *rq)
53 int ddir, sectors, bucket;
55 ddir = rq_data_dir(rq);
56 sectors = blk_rq_stats_sectors(rq);
58 bucket = ddir + 2 * ilog2(sectors);
62 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
63 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
68 #define BLK_QC_T_SHIFT 16
69 #define BLK_QC_T_INTERNAL (1U << 31)
71 static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q,
74 return xa_load(&q->hctx_table,
75 (qc & ~BLK_QC_T_INTERNAL) >> BLK_QC_T_SHIFT);
78 static inline struct request *blk_qc_to_rq(struct blk_mq_hw_ctx *hctx,
81 unsigned int tag = qc & ((1U << BLK_QC_T_SHIFT) - 1);
83 if (qc & BLK_QC_T_INTERNAL)
84 return blk_mq_tag_to_rq(hctx->sched_tags, tag);
85 return blk_mq_tag_to_rq(hctx->tags, tag);
88 static inline blk_qc_t blk_rq_to_qc(struct request *rq)
90 return (rq->mq_hctx->queue_num << BLK_QC_T_SHIFT) |
92 rq->tag : (rq->internal_tag | BLK_QC_T_INTERNAL));
96 * Check if any of the ctx, dispatch list or elevator
97 * have pending work in this hardware queue.
99 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
101 return !list_empty_careful(&hctx->dispatch) ||
102 sbitmap_any_bit_set(&hctx->ctx_map) ||
103 blk_mq_sched_has_work(hctx);
107 * Mark this ctx as having pending work in this hardware queue
109 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
110 struct blk_mq_ctx *ctx)
112 const int bit = ctx->index_hw[hctx->type];
114 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
115 sbitmap_set_bit(&hctx->ctx_map, bit);
118 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
119 struct blk_mq_ctx *ctx)
121 const int bit = ctx->index_hw[hctx->type];
123 sbitmap_clear_bit(&hctx->ctx_map, bit);
127 struct block_device *part;
128 unsigned int inflight[2];
131 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);
478 * Try batched alloc if we want more than 1 tag.
480 if (data->nr_tags > 1) {
481 rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns);
488 * Waiting allocations only fail because of an inactive hctx. In that
489 * case just retry the hctx assignment and tag allocation as CPU hotplug
490 * should have migrated us to an online CPU by now.
492 tag = blk_mq_get_tag(data);
493 if (tag == BLK_MQ_NO_TAG) {
494 if (data->flags & BLK_MQ_REQ_NOWAIT)
497 * Give up the CPU and sleep for a random short time to
498 * ensure that thread using a realtime scheduling class
499 * are migrated off the CPU, and thus off the hctx that
506 return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag,
510 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
511 blk_mq_req_flags_t flags)
513 struct blk_mq_alloc_data data = {
522 ret = blk_queue_enter(q, flags);
526 rq = __blk_mq_alloc_requests(&data);
530 rq->__sector = (sector_t) -1;
531 rq->bio = rq->biotail = NULL;
535 return ERR_PTR(-EWOULDBLOCK);
537 EXPORT_SYMBOL(blk_mq_alloc_request);
539 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
540 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
542 struct blk_mq_alloc_data data = {
548 u64 alloc_time_ns = 0;
553 /* alloc_time includes depth and tag waits */
554 if (blk_queue_rq_alloc_time(q))
555 alloc_time_ns = ktime_get_ns();
558 * If the tag allocator sleeps we could get an allocation for a
559 * different hardware context. No need to complicate the low level
560 * allocator for this for the rare use case of a command tied to
563 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
564 return ERR_PTR(-EINVAL);
566 if (hctx_idx >= q->nr_hw_queues)
567 return ERR_PTR(-EIO);
569 ret = blk_queue_enter(q, flags);
574 * Check if the hardware context is actually mapped to anything.
575 * If not tell the caller that it should skip this queue.
578 data.hctx = xa_load(&q->hctx_table, hctx_idx);
579 if (!blk_mq_hw_queue_mapped(data.hctx))
581 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
582 if (cpu >= nr_cpu_ids)
584 data.ctx = __blk_mq_get_ctx(q, cpu);
587 blk_mq_tag_busy(data.hctx);
589 data.rq_flags |= RQF_ELV;
592 tag = blk_mq_get_tag(&data);
593 if (tag == BLK_MQ_NO_TAG)
595 return blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag,
602 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
604 static void __blk_mq_free_request(struct request *rq)
606 struct request_queue *q = rq->q;
607 struct blk_mq_ctx *ctx = rq->mq_ctx;
608 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
609 const int sched_tag = rq->internal_tag;
611 blk_crypto_free_request(rq);
612 blk_pm_mark_last_busy(rq);
614 if (rq->tag != BLK_MQ_NO_TAG)
615 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
616 if (sched_tag != BLK_MQ_NO_TAG)
617 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
618 blk_mq_sched_restart(hctx);
622 void blk_mq_free_request(struct request *rq)
624 struct request_queue *q = rq->q;
625 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
627 if ((rq->rq_flags & RQF_ELVPRIV) &&
628 q->elevator->type->ops.finish_request)
629 q->elevator->type->ops.finish_request(rq);
631 if (rq->rq_flags & RQF_MQ_INFLIGHT)
632 __blk_mq_dec_active_requests(hctx);
634 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
635 laptop_io_completion(q->disk->bdi);
639 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
640 if (req_ref_put_and_test(rq))
641 __blk_mq_free_request(rq);
643 EXPORT_SYMBOL_GPL(blk_mq_free_request);
645 void blk_mq_free_plug_rqs(struct blk_plug *plug)
649 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
650 blk_mq_free_request(rq);
653 void blk_dump_rq_flags(struct request *rq, char *msg)
655 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
656 rq->q->disk ? rq->q->disk->disk_name : "?",
657 (unsigned long long) rq->cmd_flags);
659 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
660 (unsigned long long)blk_rq_pos(rq),
661 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
662 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
663 rq->bio, rq->biotail, blk_rq_bytes(rq));
665 EXPORT_SYMBOL(blk_dump_rq_flags);
667 static void req_bio_endio(struct request *rq, struct bio *bio,
668 unsigned int nbytes, blk_status_t error)
670 if (unlikely(error)) {
671 bio->bi_status = error;
672 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
674 * Partial zone append completions cannot be supported as the
675 * BIO fragments may end up not being written sequentially.
677 if (bio->bi_iter.bi_size != nbytes)
678 bio->bi_status = BLK_STS_IOERR;
680 bio->bi_iter.bi_sector = rq->__sector;
683 bio_advance(bio, nbytes);
685 if (unlikely(rq->rq_flags & RQF_QUIET))
686 bio_set_flag(bio, BIO_QUIET);
687 /* don't actually finish bio if it's part of flush sequence */
688 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
692 static void blk_account_io_completion(struct request *req, unsigned int bytes)
694 if (req->part && blk_do_io_stat(req)) {
695 const int sgrp = op_stat_group(req_op(req));
698 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
703 static void blk_print_req_error(struct request *req, blk_status_t status)
705 printk_ratelimited(KERN_ERR
706 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
707 "phys_seg %u prio class %u\n",
708 blk_status_to_str(status),
709 req->q->disk ? req->q->disk->disk_name : "?",
710 blk_rq_pos(req), req_op(req), blk_op_str(req_op(req)),
711 req->cmd_flags & ~REQ_OP_MASK,
712 req->nr_phys_segments,
713 IOPRIO_PRIO_CLASS(req->ioprio));
717 * Fully end IO on a request. Does not support partial completions, or
720 static void blk_complete_request(struct request *req)
722 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
723 int total_bytes = blk_rq_bytes(req);
724 struct bio *bio = req->bio;
726 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
731 #ifdef CONFIG_BLK_DEV_INTEGRITY
732 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
733 req->q->integrity.profile->complete_fn(req, total_bytes);
736 blk_account_io_completion(req, total_bytes);
739 struct bio *next = bio->bi_next;
741 /* Completion has already been traced */
742 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
744 if (req_op(req) == REQ_OP_ZONE_APPEND)
745 bio->bi_iter.bi_sector = req->__sector;
753 * Reset counters so that the request stacking driver
754 * can find how many bytes remain in the request
762 * blk_update_request - Complete multiple bytes without completing the request
763 * @req: the request being processed
764 * @error: block status code
765 * @nr_bytes: number of bytes to complete for @req
768 * Ends I/O on a number of bytes attached to @req, but doesn't complete
769 * the request structure even if @req doesn't have leftover.
770 * If @req has leftover, sets it up for the next range of segments.
772 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
773 * %false return from this function.
776 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
777 * except in the consistency check at the end of this function.
780 * %false - this request doesn't have any more data
781 * %true - this request has more data
783 bool blk_update_request(struct request *req, blk_status_t error,
784 unsigned int nr_bytes)
788 trace_block_rq_complete(req, error, nr_bytes);
793 #ifdef CONFIG_BLK_DEV_INTEGRITY
794 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
796 req->q->integrity.profile->complete_fn(req, nr_bytes);
799 if (unlikely(error && !blk_rq_is_passthrough(req) &&
800 !(req->rq_flags & RQF_QUIET)) &&
801 !test_bit(GD_DEAD, &req->q->disk->state)) {
802 blk_print_req_error(req, error);
803 trace_block_rq_error(req, error, nr_bytes);
806 blk_account_io_completion(req, nr_bytes);
810 struct bio *bio = req->bio;
811 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
813 if (bio_bytes == bio->bi_iter.bi_size)
814 req->bio = bio->bi_next;
816 /* Completion has already been traced */
817 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
818 req_bio_endio(req, bio, bio_bytes, error);
820 total_bytes += bio_bytes;
821 nr_bytes -= bio_bytes;
832 * Reset counters so that the request stacking driver
833 * can find how many bytes remain in the request
840 req->__data_len -= total_bytes;
842 /* update sector only for requests with clear definition of sector */
843 if (!blk_rq_is_passthrough(req))
844 req->__sector += total_bytes >> 9;
846 /* mixed attributes always follow the first bio */
847 if (req->rq_flags & RQF_MIXED_MERGE) {
848 req->cmd_flags &= ~REQ_FAILFAST_MASK;
849 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
852 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
854 * If total number of sectors is less than the first segment
855 * size, something has gone terribly wrong.
857 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
858 blk_dump_rq_flags(req, "request botched");
859 req->__data_len = blk_rq_cur_bytes(req);
862 /* recalculate the number of segments */
863 req->nr_phys_segments = blk_recalc_rq_segments(req);
868 EXPORT_SYMBOL_GPL(blk_update_request);
870 static void __blk_account_io_done(struct request *req, u64 now)
872 const int sgrp = op_stat_group(req_op(req));
875 update_io_ticks(req->part, jiffies, true);
876 part_stat_inc(req->part, ios[sgrp]);
877 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
881 static inline void blk_account_io_done(struct request *req, u64 now)
884 * Account IO completion. flush_rq isn't accounted as a
885 * normal IO on queueing nor completion. Accounting the
886 * containing request is enough.
888 if (blk_do_io_stat(req) && req->part &&
889 !(req->rq_flags & RQF_FLUSH_SEQ))
890 __blk_account_io_done(req, now);
893 static void __blk_account_io_start(struct request *rq)
896 * All non-passthrough requests are created from a bio with one
897 * exception: when a flush command that is part of a flush sequence
898 * generated by the state machine in blk-flush.c is cloned onto the
899 * lower device by dm-multipath we can get here without a bio.
902 rq->part = rq->bio->bi_bdev;
904 rq->part = rq->q->disk->part0;
907 update_io_ticks(rq->part, jiffies, false);
911 static inline void blk_account_io_start(struct request *req)
913 if (blk_do_io_stat(req))
914 __blk_account_io_start(req);
917 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
919 if (rq->rq_flags & RQF_STATS) {
920 blk_mq_poll_stats_start(rq->q);
921 blk_stat_add(rq, now);
924 blk_mq_sched_completed_request(rq, now);
925 blk_account_io_done(rq, now);
928 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
930 if (blk_mq_need_time_stamp(rq))
931 __blk_mq_end_request_acct(rq, ktime_get_ns());
934 rq_qos_done(rq->q, rq);
935 rq->end_io(rq, error);
937 blk_mq_free_request(rq);
940 EXPORT_SYMBOL(__blk_mq_end_request);
942 void blk_mq_end_request(struct request *rq, blk_status_t error)
944 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
946 __blk_mq_end_request(rq, error);
948 EXPORT_SYMBOL(blk_mq_end_request);
950 #define TAG_COMP_BATCH 32
952 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
953 int *tag_array, int nr_tags)
955 struct request_queue *q = hctx->queue;
958 * All requests should have been marked as RQF_MQ_INFLIGHT, so
959 * update hctx->nr_active in batch
961 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
962 __blk_mq_sub_active_requests(hctx, nr_tags);
964 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
965 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
968 void blk_mq_end_request_batch(struct io_comp_batch *iob)
970 int tags[TAG_COMP_BATCH], nr_tags = 0;
971 struct blk_mq_hw_ctx *cur_hctx = NULL;
976 now = ktime_get_ns();
978 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
980 prefetch(rq->rq_next);
982 blk_complete_request(rq);
984 __blk_mq_end_request_acct(rq, now);
986 rq_qos_done(rq->q, rq);
988 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
989 if (!req_ref_put_and_test(rq))
992 blk_crypto_free_request(rq);
993 blk_pm_mark_last_busy(rq);
995 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
997 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
999 cur_hctx = rq->mq_hctx;
1001 tags[nr_tags++] = rq->tag;
1005 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1007 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1009 static void blk_complete_reqs(struct llist_head *list)
1011 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1012 struct request *rq, *next;
1014 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1015 rq->q->mq_ops->complete(rq);
1018 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1020 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1023 static int blk_softirq_cpu_dead(unsigned int cpu)
1025 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1029 static void __blk_mq_complete_request_remote(void *data)
1031 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1034 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1036 int cpu = raw_smp_processor_id();
1038 if (!IS_ENABLED(CONFIG_SMP) ||
1039 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1042 * With force threaded interrupts enabled, raising softirq from an SMP
1043 * function call will always result in waking the ksoftirqd thread.
1044 * This is probably worse than completing the request on a different
1047 if (force_irqthreads())
1050 /* same CPU or cache domain? Complete locally */
1051 if (cpu == rq->mq_ctx->cpu ||
1052 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1053 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1056 /* don't try to IPI to an offline CPU */
1057 return cpu_online(rq->mq_ctx->cpu);
1060 static void blk_mq_complete_send_ipi(struct request *rq)
1062 struct llist_head *list;
1065 cpu = rq->mq_ctx->cpu;
1066 list = &per_cpu(blk_cpu_done, cpu);
1067 if (llist_add(&rq->ipi_list, list)) {
1068 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1069 smp_call_function_single_async(cpu, &rq->csd);
1073 static void blk_mq_raise_softirq(struct request *rq)
1075 struct llist_head *list;
1078 list = this_cpu_ptr(&blk_cpu_done);
1079 if (llist_add(&rq->ipi_list, list))
1080 raise_softirq(BLOCK_SOFTIRQ);
1084 bool blk_mq_complete_request_remote(struct request *rq)
1086 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1089 * For a polled request, always complete locally, it's pointless
1090 * to redirect the completion.
1092 if (rq->cmd_flags & REQ_POLLED)
1095 if (blk_mq_complete_need_ipi(rq)) {
1096 blk_mq_complete_send_ipi(rq);
1100 if (rq->q->nr_hw_queues == 1) {
1101 blk_mq_raise_softirq(rq);
1106 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1109 * blk_mq_complete_request - end I/O on a request
1110 * @rq: the request being processed
1113 * Complete a request by scheduling the ->complete_rq operation.
1115 void blk_mq_complete_request(struct request *rq)
1117 if (!blk_mq_complete_request_remote(rq))
1118 rq->q->mq_ops->complete(rq);
1120 EXPORT_SYMBOL(blk_mq_complete_request);
1123 * blk_mq_start_request - Start processing a request
1124 * @rq: Pointer to request to be started
1126 * Function used by device drivers to notify the block layer that a request
1127 * is going to be processed now, so blk layer can do proper initializations
1128 * such as starting the timeout timer.
1130 void blk_mq_start_request(struct request *rq)
1132 struct request_queue *q = rq->q;
1134 trace_block_rq_issue(rq);
1136 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1137 rq->io_start_time_ns = ktime_get_ns();
1138 rq->stats_sectors = blk_rq_sectors(rq);
1139 rq->rq_flags |= RQF_STATS;
1140 rq_qos_issue(q, rq);
1143 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1146 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1148 #ifdef CONFIG_BLK_DEV_INTEGRITY
1149 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1150 q->integrity.profile->prepare_fn(rq);
1152 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1153 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1155 EXPORT_SYMBOL(blk_mq_start_request);
1158 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1159 * queues. This is important for md arrays to benefit from merging
1162 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1164 if (plug->multiple_queues)
1165 return BLK_MAX_REQUEST_COUNT * 2;
1166 return BLK_MAX_REQUEST_COUNT;
1169 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1171 struct request *last = rq_list_peek(&plug->mq_list);
1173 if (!plug->rq_count) {
1174 trace_block_plug(rq->q);
1175 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1176 (!blk_queue_nomerges(rq->q) &&
1177 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1178 blk_mq_flush_plug_list(plug, false);
1179 trace_block_plug(rq->q);
1182 if (!plug->multiple_queues && last && last->q != rq->q)
1183 plug->multiple_queues = true;
1184 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1185 plug->has_elevator = true;
1187 rq_list_add(&plug->mq_list, rq);
1192 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1193 * @rq: request to insert
1194 * @at_head: insert request at head or tail of queue
1197 * Insert a fully prepared request at the back of the I/O scheduler queue
1198 * for execution. Don't wait for completion.
1201 * This function will invoke @done directly if the queue is dead.
1203 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1205 WARN_ON(irqs_disabled());
1206 WARN_ON(!blk_rq_is_passthrough(rq));
1208 blk_account_io_start(rq);
1210 blk_add_rq_to_plug(current->plug, rq);
1212 blk_mq_sched_insert_request(rq, at_head, true, false);
1214 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1216 struct blk_rq_wait {
1217 struct completion done;
1221 static void blk_end_sync_rq(struct request *rq, blk_status_t ret)
1223 struct blk_rq_wait *wait = rq->end_io_data;
1226 complete(&wait->done);
1229 static bool blk_rq_is_poll(struct request *rq)
1233 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1235 if (WARN_ON_ONCE(!rq->bio))
1240 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1243 bio_poll(rq->bio, NULL, 0);
1245 } while (!completion_done(wait));
1249 * blk_execute_rq - insert a request into queue for execution
1250 * @rq: request to insert
1251 * @at_head: insert request at head or tail of queue
1254 * Insert a fully prepared request at the back of the I/O scheduler queue
1255 * for execution and wait for completion.
1256 * Return: The blk_status_t result provided to blk_mq_end_request().
1258 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1260 struct blk_rq_wait wait = {
1261 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1264 WARN_ON(irqs_disabled());
1265 WARN_ON(!blk_rq_is_passthrough(rq));
1267 rq->end_io_data = &wait;
1268 rq->end_io = blk_end_sync_rq;
1270 blk_account_io_start(rq);
1271 blk_mq_sched_insert_request(rq, at_head, true, false);
1273 if (blk_rq_is_poll(rq)) {
1274 blk_rq_poll_completion(rq, &wait.done);
1277 * Prevent hang_check timer from firing at us during very long
1280 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1283 while (!wait_for_completion_io_timeout(&wait.done,
1284 hang_check * (HZ/2)))
1287 wait_for_completion_io(&wait.done);
1292 EXPORT_SYMBOL(blk_execute_rq);
1294 static void __blk_mq_requeue_request(struct request *rq)
1296 struct request_queue *q = rq->q;
1298 blk_mq_put_driver_tag(rq);
1300 trace_block_rq_requeue(rq);
1301 rq_qos_requeue(q, rq);
1303 if (blk_mq_request_started(rq)) {
1304 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1305 rq->rq_flags &= ~RQF_TIMED_OUT;
1309 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1311 __blk_mq_requeue_request(rq);
1313 /* this request will be re-inserted to io scheduler queue */
1314 blk_mq_sched_requeue_request(rq);
1316 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1318 EXPORT_SYMBOL(blk_mq_requeue_request);
1320 static void blk_mq_requeue_work(struct work_struct *work)
1322 struct request_queue *q =
1323 container_of(work, struct request_queue, requeue_work.work);
1325 struct request *rq, *next;
1327 spin_lock_irq(&q->requeue_lock);
1328 list_splice_init(&q->requeue_list, &rq_list);
1329 spin_unlock_irq(&q->requeue_lock);
1331 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1332 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1335 rq->rq_flags &= ~RQF_SOFTBARRIER;
1336 list_del_init(&rq->queuelist);
1338 * If RQF_DONTPREP, rq has contained some driver specific
1339 * data, so insert it to hctx dispatch list to avoid any
1342 if (rq->rq_flags & RQF_DONTPREP)
1343 blk_mq_request_bypass_insert(rq, false, false);
1345 blk_mq_sched_insert_request(rq, true, false, false);
1348 while (!list_empty(&rq_list)) {
1349 rq = list_entry(rq_list.next, struct request, queuelist);
1350 list_del_init(&rq->queuelist);
1351 blk_mq_sched_insert_request(rq, false, false, false);
1354 blk_mq_run_hw_queues(q, false);
1357 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1358 bool kick_requeue_list)
1360 struct request_queue *q = rq->q;
1361 unsigned long flags;
1364 * We abuse this flag that is otherwise used by the I/O scheduler to
1365 * request head insertion from the workqueue.
1367 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1369 spin_lock_irqsave(&q->requeue_lock, flags);
1371 rq->rq_flags |= RQF_SOFTBARRIER;
1372 list_add(&rq->queuelist, &q->requeue_list);
1374 list_add_tail(&rq->queuelist, &q->requeue_list);
1376 spin_unlock_irqrestore(&q->requeue_lock, flags);
1378 if (kick_requeue_list)
1379 blk_mq_kick_requeue_list(q);
1382 void blk_mq_kick_requeue_list(struct request_queue *q)
1384 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1386 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1388 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1389 unsigned long msecs)
1391 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1392 msecs_to_jiffies(msecs));
1394 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1396 static bool blk_mq_rq_inflight(struct request *rq, void *priv,
1400 * If we find a request that isn't idle we know the queue is busy
1401 * as it's checked in the iter.
1402 * Return false to stop the iteration.
1404 if (blk_mq_request_started(rq)) {
1414 bool blk_mq_queue_inflight(struct request_queue *q)
1418 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1421 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1423 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
1425 req->rq_flags |= RQF_TIMED_OUT;
1426 if (req->q->mq_ops->timeout) {
1427 enum blk_eh_timer_return ret;
1429 ret = req->q->mq_ops->timeout(req, reserved);
1430 if (ret == BLK_EH_DONE)
1432 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1438 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
1440 unsigned long deadline;
1442 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1444 if (rq->rq_flags & RQF_TIMED_OUT)
1447 deadline = READ_ONCE(rq->deadline);
1448 if (time_after_eq(jiffies, deadline))
1453 else if (time_after(*next, deadline))
1458 void blk_mq_put_rq_ref(struct request *rq)
1460 if (is_flush_rq(rq))
1462 else if (req_ref_put_and_test(rq))
1463 __blk_mq_free_request(rq);
1466 static bool blk_mq_check_expired(struct request *rq, void *priv, bool reserved)
1468 unsigned long *next = priv;
1471 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1472 * be reallocated underneath the timeout handler's processing, then
1473 * the expire check is reliable. If the request is not expired, then
1474 * it was completed and reallocated as a new request after returning
1475 * from blk_mq_check_expired().
1477 if (blk_mq_req_expired(rq, next))
1478 blk_mq_rq_timed_out(rq, reserved);
1482 static void blk_mq_timeout_work(struct work_struct *work)
1484 struct request_queue *q =
1485 container_of(work, struct request_queue, timeout_work);
1486 unsigned long next = 0;
1487 struct blk_mq_hw_ctx *hctx;
1490 /* A deadlock might occur if a request is stuck requiring a
1491 * timeout at the same time a queue freeze is waiting
1492 * completion, since the timeout code would not be able to
1493 * acquire the queue reference here.
1495 * That's why we don't use blk_queue_enter here; instead, we use
1496 * percpu_ref_tryget directly, because we need to be able to
1497 * obtain a reference even in the short window between the queue
1498 * starting to freeze, by dropping the first reference in
1499 * blk_freeze_queue_start, and the moment the last request is
1500 * consumed, marked by the instant q_usage_counter reaches
1503 if (!percpu_ref_tryget(&q->q_usage_counter))
1506 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
1509 mod_timer(&q->timeout, next);
1512 * Request timeouts are handled as a forward rolling timer. If
1513 * we end up here it means that no requests are pending and
1514 * also that no request has been pending for a while. Mark
1515 * each hctx as idle.
1517 queue_for_each_hw_ctx(q, hctx, i) {
1518 /* the hctx may be unmapped, so check it here */
1519 if (blk_mq_hw_queue_mapped(hctx))
1520 blk_mq_tag_idle(hctx);
1526 struct flush_busy_ctx_data {
1527 struct blk_mq_hw_ctx *hctx;
1528 struct list_head *list;
1531 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1533 struct flush_busy_ctx_data *flush_data = data;
1534 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1535 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1536 enum hctx_type type = hctx->type;
1538 spin_lock(&ctx->lock);
1539 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1540 sbitmap_clear_bit(sb, bitnr);
1541 spin_unlock(&ctx->lock);
1546 * Process software queues that have been marked busy, splicing them
1547 * to the for-dispatch
1549 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1551 struct flush_busy_ctx_data data = {
1556 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1558 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1560 struct dispatch_rq_data {
1561 struct blk_mq_hw_ctx *hctx;
1565 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1568 struct dispatch_rq_data *dispatch_data = data;
1569 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1570 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1571 enum hctx_type type = hctx->type;
1573 spin_lock(&ctx->lock);
1574 if (!list_empty(&ctx->rq_lists[type])) {
1575 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1576 list_del_init(&dispatch_data->rq->queuelist);
1577 if (list_empty(&ctx->rq_lists[type]))
1578 sbitmap_clear_bit(sb, bitnr);
1580 spin_unlock(&ctx->lock);
1582 return !dispatch_data->rq;
1585 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1586 struct blk_mq_ctx *start)
1588 unsigned off = start ? start->index_hw[hctx->type] : 0;
1589 struct dispatch_rq_data data = {
1594 __sbitmap_for_each_set(&hctx->ctx_map, off,
1595 dispatch_rq_from_ctx, &data);
1600 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1602 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1603 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1606 blk_mq_tag_busy(rq->mq_hctx);
1608 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1609 bt = &rq->mq_hctx->tags->breserved_tags;
1612 if (!hctx_may_queue(rq->mq_hctx, bt))
1616 tag = __sbitmap_queue_get(bt);
1617 if (tag == BLK_MQ_NO_TAG)
1620 rq->tag = tag + tag_offset;
1624 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1626 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1629 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1630 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1631 rq->rq_flags |= RQF_MQ_INFLIGHT;
1632 __blk_mq_inc_active_requests(hctx);
1634 hctx->tags->rqs[rq->tag] = rq;
1638 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1639 int flags, void *key)
1641 struct blk_mq_hw_ctx *hctx;
1643 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1645 spin_lock(&hctx->dispatch_wait_lock);
1646 if (!list_empty(&wait->entry)) {
1647 struct sbitmap_queue *sbq;
1649 list_del_init(&wait->entry);
1650 sbq = &hctx->tags->bitmap_tags;
1651 atomic_dec(&sbq->ws_active);
1653 spin_unlock(&hctx->dispatch_wait_lock);
1655 blk_mq_run_hw_queue(hctx, true);
1660 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1661 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1662 * restart. For both cases, take care to check the condition again after
1663 * marking us as waiting.
1665 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1668 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1669 struct wait_queue_head *wq;
1670 wait_queue_entry_t *wait;
1673 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1674 blk_mq_sched_mark_restart_hctx(hctx);
1677 * It's possible that a tag was freed in the window between the
1678 * allocation failure and adding the hardware queue to the wait
1681 * Don't clear RESTART here, someone else could have set it.
1682 * At most this will cost an extra queue run.
1684 return blk_mq_get_driver_tag(rq);
1687 wait = &hctx->dispatch_wait;
1688 if (!list_empty_careful(&wait->entry))
1691 wq = &bt_wait_ptr(sbq, hctx)->wait;
1693 spin_lock_irq(&wq->lock);
1694 spin_lock(&hctx->dispatch_wait_lock);
1695 if (!list_empty(&wait->entry)) {
1696 spin_unlock(&hctx->dispatch_wait_lock);
1697 spin_unlock_irq(&wq->lock);
1701 atomic_inc(&sbq->ws_active);
1702 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1703 __add_wait_queue(wq, wait);
1706 * It's possible that a tag was freed in the window between the
1707 * allocation failure and adding the hardware queue to the wait
1710 ret = blk_mq_get_driver_tag(rq);
1712 spin_unlock(&hctx->dispatch_wait_lock);
1713 spin_unlock_irq(&wq->lock);
1718 * We got a tag, remove ourselves from the wait queue to ensure
1719 * someone else gets the wakeup.
1721 list_del_init(&wait->entry);
1722 atomic_dec(&sbq->ws_active);
1723 spin_unlock(&hctx->dispatch_wait_lock);
1724 spin_unlock_irq(&wq->lock);
1729 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1730 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1732 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1733 * - EWMA is one simple way to compute running average value
1734 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1735 * - take 4 as factor for avoiding to get too small(0) result, and this
1736 * factor doesn't matter because EWMA decreases exponentially
1738 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1742 ewma = hctx->dispatch_busy;
1747 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1749 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1750 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1752 hctx->dispatch_busy = ewma;
1755 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1757 static void blk_mq_handle_dev_resource(struct request *rq,
1758 struct list_head *list)
1760 struct request *next =
1761 list_first_entry_or_null(list, struct request, queuelist);
1764 * If an I/O scheduler has been configured and we got a driver tag for
1765 * the next request already, free it.
1768 blk_mq_put_driver_tag(next);
1770 list_add(&rq->queuelist, list);
1771 __blk_mq_requeue_request(rq);
1774 static void blk_mq_handle_zone_resource(struct request *rq,
1775 struct list_head *zone_list)
1778 * If we end up here it is because we cannot dispatch a request to a
1779 * specific zone due to LLD level zone-write locking or other zone
1780 * related resource not being available. In this case, set the request
1781 * aside in zone_list for retrying it later.
1783 list_add(&rq->queuelist, zone_list);
1784 __blk_mq_requeue_request(rq);
1787 enum prep_dispatch {
1789 PREP_DISPATCH_NO_TAG,
1790 PREP_DISPATCH_NO_BUDGET,
1793 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1796 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1797 int budget_token = -1;
1800 budget_token = blk_mq_get_dispatch_budget(rq->q);
1801 if (budget_token < 0) {
1802 blk_mq_put_driver_tag(rq);
1803 return PREP_DISPATCH_NO_BUDGET;
1805 blk_mq_set_rq_budget_token(rq, budget_token);
1808 if (!blk_mq_get_driver_tag(rq)) {
1810 * The initial allocation attempt failed, so we need to
1811 * rerun the hardware queue when a tag is freed. The
1812 * waitqueue takes care of that. If the queue is run
1813 * before we add this entry back on the dispatch list,
1814 * we'll re-run it below.
1816 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1818 * All budgets not got from this function will be put
1819 * together during handling partial dispatch
1822 blk_mq_put_dispatch_budget(rq->q, budget_token);
1823 return PREP_DISPATCH_NO_TAG;
1827 return PREP_DISPATCH_OK;
1830 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1831 static void blk_mq_release_budgets(struct request_queue *q,
1832 struct list_head *list)
1836 list_for_each_entry(rq, list, queuelist) {
1837 int budget_token = blk_mq_get_rq_budget_token(rq);
1839 if (budget_token >= 0)
1840 blk_mq_put_dispatch_budget(q, budget_token);
1845 * Returns true if we did some work AND can potentially do more.
1847 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1848 unsigned int nr_budgets)
1850 enum prep_dispatch prep;
1851 struct request_queue *q = hctx->queue;
1852 struct request *rq, *nxt;
1854 blk_status_t ret = BLK_STS_OK;
1855 LIST_HEAD(zone_list);
1856 bool needs_resource = false;
1858 if (list_empty(list))
1862 * Now process all the entries, sending them to the driver.
1864 errors = queued = 0;
1866 struct blk_mq_queue_data bd;
1868 rq = list_first_entry(list, struct request, queuelist);
1870 WARN_ON_ONCE(hctx != rq->mq_hctx);
1871 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1872 if (prep != PREP_DISPATCH_OK)
1875 list_del_init(&rq->queuelist);
1880 * Flag last if we have no more requests, or if we have more
1881 * but can't assign a driver tag to it.
1883 if (list_empty(list))
1886 nxt = list_first_entry(list, struct request, queuelist);
1887 bd.last = !blk_mq_get_driver_tag(nxt);
1891 * once the request is queued to lld, no need to cover the
1896 ret = q->mq_ops->queue_rq(hctx, &bd);
1901 case BLK_STS_RESOURCE:
1902 needs_resource = true;
1904 case BLK_STS_DEV_RESOURCE:
1905 blk_mq_handle_dev_resource(rq, list);
1907 case BLK_STS_ZONE_RESOURCE:
1909 * Move the request to zone_list and keep going through
1910 * the dispatch list to find more requests the drive can
1913 blk_mq_handle_zone_resource(rq, &zone_list);
1914 needs_resource = true;
1918 blk_mq_end_request(rq, ret);
1920 } while (!list_empty(list));
1922 if (!list_empty(&zone_list))
1923 list_splice_tail_init(&zone_list, list);
1925 /* If we didn't flush the entire list, we could have told the driver
1926 * there was more coming, but that turned out to be a lie.
1928 if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued)
1929 q->mq_ops->commit_rqs(hctx);
1931 * Any items that need requeuing? Stuff them into hctx->dispatch,
1932 * that is where we will continue on next queue run.
1934 if (!list_empty(list)) {
1936 /* For non-shared tags, the RESTART check will suffice */
1937 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1938 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1941 blk_mq_release_budgets(q, list);
1943 spin_lock(&hctx->lock);
1944 list_splice_tail_init(list, &hctx->dispatch);
1945 spin_unlock(&hctx->lock);
1948 * Order adding requests to hctx->dispatch and checking
1949 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1950 * in blk_mq_sched_restart(). Avoid restart code path to
1951 * miss the new added requests to hctx->dispatch, meantime
1952 * SCHED_RESTART is observed here.
1957 * If SCHED_RESTART was set by the caller of this function and
1958 * it is no longer set that means that it was cleared by another
1959 * thread and hence that a queue rerun is needed.
1961 * If 'no_tag' is set, that means that we failed getting
1962 * a driver tag with an I/O scheduler attached. If our dispatch
1963 * waitqueue is no longer active, ensure that we run the queue
1964 * AFTER adding our entries back to the list.
1966 * If no I/O scheduler has been configured it is possible that
1967 * the hardware queue got stopped and restarted before requests
1968 * were pushed back onto the dispatch list. Rerun the queue to
1969 * avoid starvation. Notes:
1970 * - blk_mq_run_hw_queue() checks whether or not a queue has
1971 * been stopped before rerunning a queue.
1972 * - Some but not all block drivers stop a queue before
1973 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1976 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1977 * bit is set, run queue after a delay to avoid IO stalls
1978 * that could otherwise occur if the queue is idle. We'll do
1979 * similar if we couldn't get budget or couldn't lock a zone
1980 * and SCHED_RESTART is set.
1982 needs_restart = blk_mq_sched_needs_restart(hctx);
1983 if (prep == PREP_DISPATCH_NO_BUDGET)
1984 needs_resource = true;
1985 if (!needs_restart ||
1986 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1987 blk_mq_run_hw_queue(hctx, true);
1988 else if (needs_restart && needs_resource)
1989 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1991 blk_mq_update_dispatch_busy(hctx, true);
1994 blk_mq_update_dispatch_busy(hctx, false);
1996 return (queued + errors) != 0;
2000 * __blk_mq_run_hw_queue - Run a hardware queue.
2001 * @hctx: Pointer to the hardware queue to run.
2003 * Send pending requests to the hardware.
2005 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
2008 * We can't run the queue inline with ints disabled. Ensure that
2009 * we catch bad users of this early.
2011 WARN_ON_ONCE(in_interrupt());
2013 blk_mq_run_dispatch_ops(hctx->queue,
2014 blk_mq_sched_dispatch_requests(hctx));
2017 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2019 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2021 if (cpu >= nr_cpu_ids)
2022 cpu = cpumask_first(hctx->cpumask);
2027 * It'd be great if the workqueue API had a way to pass
2028 * in a mask and had some smarts for more clever placement.
2029 * For now we just round-robin here, switching for every
2030 * BLK_MQ_CPU_WORK_BATCH queued items.
2032 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2035 int next_cpu = hctx->next_cpu;
2037 if (hctx->queue->nr_hw_queues == 1)
2038 return WORK_CPU_UNBOUND;
2040 if (--hctx->next_cpu_batch <= 0) {
2042 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2044 if (next_cpu >= nr_cpu_ids)
2045 next_cpu = blk_mq_first_mapped_cpu(hctx);
2046 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2050 * Do unbound schedule if we can't find a online CPU for this hctx,
2051 * and it should only happen in the path of handling CPU DEAD.
2053 if (!cpu_online(next_cpu)) {
2060 * Make sure to re-select CPU next time once after CPUs
2061 * in hctx->cpumask become online again.
2063 hctx->next_cpu = next_cpu;
2064 hctx->next_cpu_batch = 1;
2065 return WORK_CPU_UNBOUND;
2068 hctx->next_cpu = next_cpu;
2073 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2074 * @hctx: Pointer to the hardware queue to run.
2075 * @async: If we want to run the queue asynchronously.
2076 * @msecs: Milliseconds of delay to wait before running the queue.
2078 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2079 * with a delay of @msecs.
2081 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2082 unsigned long msecs)
2084 if (unlikely(blk_mq_hctx_stopped(hctx)))
2087 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2088 int cpu = get_cpu();
2089 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
2090 __blk_mq_run_hw_queue(hctx);
2098 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2099 msecs_to_jiffies(msecs));
2103 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2104 * @hctx: Pointer to the hardware queue to run.
2105 * @msecs: Milliseconds of delay to wait before running the queue.
2107 * Run a hardware queue asynchronously with a delay of @msecs.
2109 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2111 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2113 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2116 * blk_mq_run_hw_queue - Start to run a hardware queue.
2117 * @hctx: Pointer to the hardware queue to run.
2118 * @async: If we want to run the queue asynchronously.
2120 * Check if the request queue is not in a quiesced state and if there are
2121 * pending requests to be sent. If this is true, run the queue to send requests
2124 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2129 * When queue is quiesced, we may be switching io scheduler, or
2130 * updating nr_hw_queues, or other things, and we can't run queue
2131 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2133 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2136 __blk_mq_run_dispatch_ops(hctx->queue, false,
2137 need_run = !blk_queue_quiesced(hctx->queue) &&
2138 blk_mq_hctx_has_pending(hctx));
2141 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2143 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2146 * Is the request queue handled by an IO scheduler that does not respect
2147 * hardware queues when dispatching?
2149 static bool blk_mq_has_sqsched(struct request_queue *q)
2151 struct elevator_queue *e = q->elevator;
2153 if (e && e->type->ops.dispatch_request &&
2154 !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE))
2160 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2163 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2165 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2167 * If the IO scheduler does not respect hardware queues when
2168 * dispatching, we just don't bother with multiple HW queues and
2169 * dispatch from hctx for the current CPU since running multiple queues
2170 * just causes lock contention inside the scheduler and pointless cache
2173 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, 0, ctx);
2175 if (!blk_mq_hctx_stopped(hctx))
2181 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2182 * @q: Pointer to the request queue to run.
2183 * @async: If we want to run the queue asynchronously.
2185 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2187 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2191 if (blk_mq_has_sqsched(q))
2192 sq_hctx = blk_mq_get_sq_hctx(q);
2193 queue_for_each_hw_ctx(q, hctx, i) {
2194 if (blk_mq_hctx_stopped(hctx))
2197 * Dispatch from this hctx either if there's no hctx preferred
2198 * by IO scheduler or if it has requests that bypass the
2201 if (!sq_hctx || sq_hctx == hctx ||
2202 !list_empty_careful(&hctx->dispatch))
2203 blk_mq_run_hw_queue(hctx, async);
2206 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2209 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2210 * @q: Pointer to the request queue to run.
2211 * @msecs: Milliseconds of delay to wait before running the queues.
2213 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2215 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2219 if (blk_mq_has_sqsched(q))
2220 sq_hctx = blk_mq_get_sq_hctx(q);
2221 queue_for_each_hw_ctx(q, hctx, i) {
2222 if (blk_mq_hctx_stopped(hctx))
2225 * If there is already a run_work pending, leave the
2226 * pending delay untouched. Otherwise, a hctx can stall
2227 * if another hctx is re-delaying the other's work
2228 * before the work executes.
2230 if (delayed_work_pending(&hctx->run_work))
2233 * Dispatch from this hctx either if there's no hctx preferred
2234 * by IO scheduler or if it has requests that bypass the
2237 if (!sq_hctx || sq_hctx == hctx ||
2238 !list_empty_careful(&hctx->dispatch))
2239 blk_mq_delay_run_hw_queue(hctx, msecs);
2242 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2245 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
2246 * @q: request queue.
2248 * The caller is responsible for serializing this function against
2249 * blk_mq_{start,stop}_hw_queue().
2251 bool blk_mq_queue_stopped(struct request_queue *q)
2253 struct blk_mq_hw_ctx *hctx;
2256 queue_for_each_hw_ctx(q, hctx, i)
2257 if (blk_mq_hctx_stopped(hctx))
2262 EXPORT_SYMBOL(blk_mq_queue_stopped);
2265 * This function is often used for pausing .queue_rq() by driver when
2266 * there isn't enough resource or some conditions aren't satisfied, and
2267 * BLK_STS_RESOURCE is usually returned.
2269 * We do not guarantee that dispatch can be drained or blocked
2270 * after blk_mq_stop_hw_queue() returns. Please use
2271 * blk_mq_quiesce_queue() for that requirement.
2273 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2275 cancel_delayed_work(&hctx->run_work);
2277 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2279 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2282 * This function is often used for pausing .queue_rq() by driver when
2283 * there isn't enough resource or some conditions aren't satisfied, and
2284 * BLK_STS_RESOURCE is usually returned.
2286 * We do not guarantee that dispatch can be drained or blocked
2287 * after blk_mq_stop_hw_queues() returns. Please use
2288 * blk_mq_quiesce_queue() for that requirement.
2290 void blk_mq_stop_hw_queues(struct request_queue *q)
2292 struct blk_mq_hw_ctx *hctx;
2295 queue_for_each_hw_ctx(q, hctx, i)
2296 blk_mq_stop_hw_queue(hctx);
2298 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2300 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2302 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2304 blk_mq_run_hw_queue(hctx, false);
2306 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2308 void blk_mq_start_hw_queues(struct request_queue *q)
2310 struct blk_mq_hw_ctx *hctx;
2313 queue_for_each_hw_ctx(q, hctx, i)
2314 blk_mq_start_hw_queue(hctx);
2316 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2318 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2320 if (!blk_mq_hctx_stopped(hctx))
2323 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2324 blk_mq_run_hw_queue(hctx, async);
2326 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2328 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2330 struct blk_mq_hw_ctx *hctx;
2333 queue_for_each_hw_ctx(q, hctx, i)
2334 blk_mq_start_stopped_hw_queue(hctx, async);
2336 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2338 static void blk_mq_run_work_fn(struct work_struct *work)
2340 struct blk_mq_hw_ctx *hctx;
2342 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2345 * If we are stopped, don't run the queue.
2347 if (blk_mq_hctx_stopped(hctx))
2350 __blk_mq_run_hw_queue(hctx);
2353 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2357 struct blk_mq_ctx *ctx = rq->mq_ctx;
2358 enum hctx_type type = hctx->type;
2360 lockdep_assert_held(&ctx->lock);
2362 trace_block_rq_insert(rq);
2365 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2367 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2370 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2373 struct blk_mq_ctx *ctx = rq->mq_ctx;
2375 lockdep_assert_held(&ctx->lock);
2377 __blk_mq_insert_req_list(hctx, rq, at_head);
2378 blk_mq_hctx_mark_pending(hctx, ctx);
2382 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2383 * @rq: Pointer to request to be inserted.
2384 * @at_head: true if the request should be inserted at the head of the list.
2385 * @run_queue: If we should run the hardware queue after inserting the request.
2387 * Should only be used carefully, when the caller knows we want to
2388 * bypass a potential IO scheduler on the target device.
2390 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2393 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2395 spin_lock(&hctx->lock);
2397 list_add(&rq->queuelist, &hctx->dispatch);
2399 list_add_tail(&rq->queuelist, &hctx->dispatch);
2400 spin_unlock(&hctx->lock);
2403 blk_mq_run_hw_queue(hctx, false);
2406 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2407 struct list_head *list)
2411 enum hctx_type type = hctx->type;
2414 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2417 list_for_each_entry(rq, list, queuelist) {
2418 BUG_ON(rq->mq_ctx != ctx);
2419 trace_block_rq_insert(rq);
2422 spin_lock(&ctx->lock);
2423 list_splice_tail_init(list, &ctx->rq_lists[type]);
2424 blk_mq_hctx_mark_pending(hctx, ctx);
2425 spin_unlock(&ctx->lock);
2428 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2431 if (hctx->queue->mq_ops->commit_rqs) {
2432 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2433 hctx->queue->mq_ops->commit_rqs(hctx);
2438 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2439 unsigned int nr_segs)
2443 if (bio->bi_opf & REQ_RAHEAD)
2444 rq->cmd_flags |= REQ_FAILFAST_MASK;
2446 rq->__sector = bio->bi_iter.bi_sector;
2447 blk_rq_bio_prep(rq, bio, nr_segs);
2449 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2450 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2453 blk_account_io_start(rq);
2456 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2457 struct request *rq, bool last)
2459 struct request_queue *q = rq->q;
2460 struct blk_mq_queue_data bd = {
2467 * For OK queue, we are done. For error, caller may kill it.
2468 * Any other error (busy), just add it to our list as we
2469 * previously would have done.
2471 ret = q->mq_ops->queue_rq(hctx, &bd);
2474 blk_mq_update_dispatch_busy(hctx, false);
2476 case BLK_STS_RESOURCE:
2477 case BLK_STS_DEV_RESOURCE:
2478 blk_mq_update_dispatch_busy(hctx, true);
2479 __blk_mq_requeue_request(rq);
2482 blk_mq_update_dispatch_busy(hctx, false);
2489 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2491 bool bypass_insert, bool last)
2493 struct request_queue *q = rq->q;
2494 bool run_queue = true;
2498 * RCU or SRCU read lock is needed before checking quiesced flag.
2500 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2501 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2502 * and avoid driver to try to dispatch again.
2504 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2506 bypass_insert = false;
2510 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2513 budget_token = blk_mq_get_dispatch_budget(q);
2514 if (budget_token < 0)
2517 blk_mq_set_rq_budget_token(rq, budget_token);
2519 if (!blk_mq_get_driver_tag(rq)) {
2520 blk_mq_put_dispatch_budget(q, budget_token);
2524 return __blk_mq_issue_directly(hctx, rq, last);
2527 return BLK_STS_RESOURCE;
2529 blk_mq_sched_insert_request(rq, false, run_queue, false);
2535 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2536 * @hctx: Pointer of the associated hardware queue.
2537 * @rq: Pointer to request to be sent.
2539 * If the device has enough resources to accept a new request now, send the
2540 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2541 * we can try send it another time in the future. Requests inserted at this
2542 * queue have higher priority.
2544 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2548 __blk_mq_try_issue_directly(hctx, rq, false, true);
2550 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2551 blk_mq_request_bypass_insert(rq, false, true);
2552 else if (ret != BLK_STS_OK)
2553 blk_mq_end_request(rq, ret);
2556 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2558 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2561 static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2563 struct blk_mq_hw_ctx *hctx = NULL;
2568 while ((rq = rq_list_pop(&plug->mq_list))) {
2569 bool last = rq_list_empty(plug->mq_list);
2572 if (hctx != rq->mq_hctx) {
2574 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2578 ret = blk_mq_request_issue_directly(rq, last);
2583 case BLK_STS_RESOURCE:
2584 case BLK_STS_DEV_RESOURCE:
2585 blk_mq_request_bypass_insert(rq, false, last);
2586 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2589 blk_mq_end_request(rq, ret);
2596 * If we didn't flush the entire list, we could have told the driver
2597 * there was more coming, but that turned out to be a lie.
2600 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2603 static void __blk_mq_flush_plug_list(struct request_queue *q,
2604 struct blk_plug *plug)
2606 if (blk_queue_quiesced(q))
2608 q->mq_ops->queue_rqs(&plug->mq_list);
2611 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2613 struct blk_mq_hw_ctx *this_hctx = NULL;
2614 struct blk_mq_ctx *this_ctx = NULL;
2615 struct request *requeue_list = NULL;
2616 unsigned int depth = 0;
2620 struct request *rq = rq_list_pop(&plug->mq_list);
2623 this_hctx = rq->mq_hctx;
2624 this_ctx = rq->mq_ctx;
2625 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2626 rq_list_add(&requeue_list, rq);
2629 list_add_tail(&rq->queuelist, &list);
2631 } while (!rq_list_empty(plug->mq_list));
2633 plug->mq_list = requeue_list;
2634 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2635 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched);
2638 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2642 if (rq_list_empty(plug->mq_list))
2646 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2647 struct request_queue *q;
2649 rq = rq_list_peek(&plug->mq_list);
2653 * Peek first request and see if we have a ->queue_rqs() hook.
2654 * If we do, we can dispatch the whole plug list in one go. We
2655 * already know at this point that all requests belong to the
2656 * same queue, caller must ensure that's the case.
2658 * Since we pass off the full list to the driver at this point,
2659 * we do not increment the active request count for the queue.
2660 * Bypass shared tags for now because of that.
2662 if (q->mq_ops->queue_rqs &&
2663 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2664 blk_mq_run_dispatch_ops(q,
2665 __blk_mq_flush_plug_list(q, plug));
2666 if (rq_list_empty(plug->mq_list))
2670 blk_mq_run_dispatch_ops(q,
2671 blk_mq_plug_issue_direct(plug, false));
2672 if (rq_list_empty(plug->mq_list))
2677 blk_mq_dispatch_plug_list(plug, from_schedule);
2678 } while (!rq_list_empty(plug->mq_list));
2681 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2682 struct list_head *list)
2687 while (!list_empty(list)) {
2689 struct request *rq = list_first_entry(list, struct request,
2692 list_del_init(&rq->queuelist);
2693 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2694 if (ret != BLK_STS_OK) {
2695 if (ret == BLK_STS_RESOURCE ||
2696 ret == BLK_STS_DEV_RESOURCE) {
2697 blk_mq_request_bypass_insert(rq, false,
2701 blk_mq_end_request(rq, ret);
2708 * If we didn't flush the entire list, we could have told
2709 * the driver there was more coming, but that turned out to
2712 if ((!list_empty(list) || errors) &&
2713 hctx->queue->mq_ops->commit_rqs && queued)
2714 hctx->queue->mq_ops->commit_rqs(hctx);
2717 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2718 struct bio *bio, unsigned int nr_segs)
2720 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2721 if (blk_attempt_plug_merge(q, bio, nr_segs))
2723 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2729 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2730 struct blk_plug *plug,
2734 struct blk_mq_alloc_data data = {
2737 .cmd_flags = bio->bi_opf,
2741 if (unlikely(bio_queue_enter(bio)))
2744 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2747 rq_qos_throttle(q, bio);
2750 data.nr_tags = plug->nr_ios;
2752 data.cached_rq = &plug->cached_rq;
2755 rq = __blk_mq_alloc_requests(&data);
2758 rq_qos_cleanup(q, bio);
2759 if (bio->bi_opf & REQ_NOWAIT)
2760 bio_wouldblock_error(bio);
2766 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2767 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2773 rq = rq_list_peek(&plug->cached_rq);
2774 if (!rq || rq->q != q)
2777 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2782 rq_qos_throttle(q, *bio);
2784 if (blk_mq_get_hctx_type((*bio)->bi_opf) != rq->mq_hctx->type)
2786 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2789 rq->cmd_flags = (*bio)->bi_opf;
2790 plug->cached_rq = rq_list_next(rq);
2791 INIT_LIST_HEAD(&rq->queuelist);
2796 * blk_mq_submit_bio - Create and send a request to block device.
2797 * @bio: Bio pointer.
2799 * Builds up a request structure from @q and @bio and send to the device. The
2800 * request may not be queued directly to hardware if:
2801 * * This request can be merged with another one
2802 * * We want to place request at plug queue for possible future merging
2803 * * There is an IO scheduler active at this queue
2805 * It will not queue the request if there is an error with the bio, or at the
2808 void blk_mq_submit_bio(struct bio *bio)
2810 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2811 struct blk_plug *plug = blk_mq_plug(q, bio);
2812 const int is_sync = op_is_sync(bio->bi_opf);
2814 unsigned int nr_segs = 1;
2817 blk_queue_bounce(q, &bio);
2818 if (blk_may_split(q, bio))
2819 __blk_queue_split(q, &bio, &nr_segs);
2821 if (!bio_integrity_prep(bio))
2824 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2828 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2833 trace_block_getrq(bio);
2835 rq_qos_track(q, rq, bio);
2837 blk_mq_bio_to_request(rq, bio, nr_segs);
2839 ret = blk_crypto_init_request(rq);
2840 if (ret != BLK_STS_OK) {
2841 bio->bi_status = ret;
2843 blk_mq_free_request(rq);
2847 if (op_is_flush(bio->bi_opf)) {
2848 blk_insert_flush(rq);
2853 blk_add_rq_to_plug(plug, rq);
2854 else if ((rq->rq_flags & RQF_ELV) ||
2855 (rq->mq_hctx->dispatch_busy &&
2856 (q->nr_hw_queues == 1 || !is_sync)))
2857 blk_mq_sched_insert_request(rq, false, true, true);
2859 blk_mq_run_dispatch_ops(rq->q,
2860 blk_mq_try_issue_directly(rq->mq_hctx, rq));
2863 #ifdef CONFIG_BLK_MQ_STACKING
2865 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
2866 * @rq: the request being queued
2868 blk_status_t blk_insert_cloned_request(struct request *rq)
2870 struct request_queue *q = rq->q;
2871 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
2874 if (blk_rq_sectors(rq) > max_sectors) {
2876 * SCSI device does not have a good way to return if
2877 * Write Same/Zero is actually supported. If a device rejects
2878 * a non-read/write command (discard, write same,etc.) the
2879 * low-level device driver will set the relevant queue limit to
2880 * 0 to prevent blk-lib from issuing more of the offending
2881 * operations. Commands queued prior to the queue limit being
2882 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
2883 * errors being propagated to upper layers.
2885 if (max_sectors == 0)
2886 return BLK_STS_NOTSUPP;
2888 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
2889 __func__, blk_rq_sectors(rq), max_sectors);
2890 return BLK_STS_IOERR;
2894 * The queue settings related to segment counting may differ from the
2897 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
2898 if (rq->nr_phys_segments > queue_max_segments(q)) {
2899 printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n",
2900 __func__, rq->nr_phys_segments, queue_max_segments(q));
2901 return BLK_STS_IOERR;
2904 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
2905 return BLK_STS_IOERR;
2907 if (blk_crypto_insert_cloned_request(rq))
2908 return BLK_STS_IOERR;
2910 blk_account_io_start(rq);
2913 * Since we have a scheduler attached on the top device,
2914 * bypass a potential scheduler on the bottom device for
2917 blk_mq_run_dispatch_ops(q,
2918 ret = blk_mq_request_issue_directly(rq, true));
2920 blk_account_io_done(rq, ktime_get_ns());
2923 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
2926 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
2927 * @rq: the clone request to be cleaned up
2930 * Free all bios in @rq for a cloned request.
2932 void blk_rq_unprep_clone(struct request *rq)
2936 while ((bio = rq->bio) != NULL) {
2937 rq->bio = bio->bi_next;
2942 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
2945 * blk_rq_prep_clone - Helper function to setup clone request
2946 * @rq: the request to be setup
2947 * @rq_src: original request to be cloned
2948 * @bs: bio_set that bios for clone are allocated from
2949 * @gfp_mask: memory allocation mask for bio
2950 * @bio_ctr: setup function to be called for each clone bio.
2951 * Returns %0 for success, non %0 for failure.
2952 * @data: private data to be passed to @bio_ctr
2955 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
2956 * Also, pages which the original bios are pointing to are not copied
2957 * and the cloned bios just point same pages.
2958 * So cloned bios must be completed before original bios, which means
2959 * the caller must complete @rq before @rq_src.
2961 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
2962 struct bio_set *bs, gfp_t gfp_mask,
2963 int (*bio_ctr)(struct bio *, struct bio *, void *),
2966 struct bio *bio, *bio_src;
2971 __rq_for_each_bio(bio_src, rq_src) {
2972 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
2977 if (bio_ctr && bio_ctr(bio, bio_src, data))
2981 rq->biotail->bi_next = bio;
2984 rq->bio = rq->biotail = bio;
2989 /* Copy attributes of the original request to the clone request. */
2990 rq->__sector = blk_rq_pos(rq_src);
2991 rq->__data_len = blk_rq_bytes(rq_src);
2992 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
2993 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
2994 rq->special_vec = rq_src->special_vec;
2996 rq->nr_phys_segments = rq_src->nr_phys_segments;
2997 rq->ioprio = rq_src->ioprio;
2999 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3007 blk_rq_unprep_clone(rq);
3011 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3012 #endif /* CONFIG_BLK_MQ_STACKING */
3015 * Steal bios from a request and add them to a bio list.
3016 * The request must not have been partially completed before.
3018 void blk_steal_bios(struct bio_list *list, struct request *rq)
3022 list->tail->bi_next = rq->bio;
3024 list->head = rq->bio;
3025 list->tail = rq->biotail;
3033 EXPORT_SYMBOL_GPL(blk_steal_bios);
3035 static size_t order_to_size(unsigned int order)
3037 return (size_t)PAGE_SIZE << order;
3040 /* called before freeing request pool in @tags */
3041 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3042 struct blk_mq_tags *tags)
3045 unsigned long flags;
3047 /* There is no need to clear a driver tags own mapping */
3048 if (drv_tags == tags)
3051 list_for_each_entry(page, &tags->page_list, lru) {
3052 unsigned long start = (unsigned long)page_address(page);
3053 unsigned long end = start + order_to_size(page->private);
3056 for (i = 0; i < drv_tags->nr_tags; i++) {
3057 struct request *rq = drv_tags->rqs[i];
3058 unsigned long rq_addr = (unsigned long)rq;
3060 if (rq_addr >= start && rq_addr < end) {
3061 WARN_ON_ONCE(req_ref_read(rq) != 0);
3062 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3068 * Wait until all pending iteration is done.
3070 * Request reference is cleared and it is guaranteed to be observed
3071 * after the ->lock is released.
3073 spin_lock_irqsave(&drv_tags->lock, flags);
3074 spin_unlock_irqrestore(&drv_tags->lock, flags);
3077 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3078 unsigned int hctx_idx)
3080 struct blk_mq_tags *drv_tags;
3083 if (list_empty(&tags->page_list))
3086 if (blk_mq_is_shared_tags(set->flags))
3087 drv_tags = set->shared_tags;
3089 drv_tags = set->tags[hctx_idx];
3091 if (tags->static_rqs && set->ops->exit_request) {
3094 for (i = 0; i < tags->nr_tags; i++) {
3095 struct request *rq = tags->static_rqs[i];
3099 set->ops->exit_request(set, rq, hctx_idx);
3100 tags->static_rqs[i] = NULL;
3104 blk_mq_clear_rq_mapping(drv_tags, tags);
3106 while (!list_empty(&tags->page_list)) {
3107 page = list_first_entry(&tags->page_list, struct page, lru);
3108 list_del_init(&page->lru);
3110 * Remove kmemleak object previously allocated in
3111 * blk_mq_alloc_rqs().
3113 kmemleak_free(page_address(page));
3114 __free_pages(page, page->private);
3118 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3122 kfree(tags->static_rqs);
3123 tags->static_rqs = NULL;
3125 blk_mq_free_tags(tags);
3128 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3129 unsigned int hctx_idx)
3133 for (i = 0; i < set->nr_maps; i++) {
3134 unsigned int start = set->map[i].queue_offset;
3135 unsigned int end = start + set->map[i].nr_queues;
3137 if (hctx_idx >= start && hctx_idx < end)
3141 if (i >= set->nr_maps)
3142 i = HCTX_TYPE_DEFAULT;
3147 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3148 unsigned int hctx_idx)
3150 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3152 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3155 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3156 unsigned int hctx_idx,
3157 unsigned int nr_tags,
3158 unsigned int reserved_tags)
3160 int node = blk_mq_get_hctx_node(set, hctx_idx);
3161 struct blk_mq_tags *tags;
3163 if (node == NUMA_NO_NODE)
3164 node = set->numa_node;
3166 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3167 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3171 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3172 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3175 blk_mq_free_tags(tags);
3179 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3180 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3182 if (!tags->static_rqs) {
3184 blk_mq_free_tags(tags);
3191 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3192 unsigned int hctx_idx, int node)
3196 if (set->ops->init_request) {
3197 ret = set->ops->init_request(set, rq, hctx_idx, node);
3202 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3206 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3207 struct blk_mq_tags *tags,
3208 unsigned int hctx_idx, unsigned int depth)
3210 unsigned int i, j, entries_per_page, max_order = 4;
3211 int node = blk_mq_get_hctx_node(set, hctx_idx);
3212 size_t rq_size, left;
3214 if (node == NUMA_NO_NODE)
3215 node = set->numa_node;
3217 INIT_LIST_HEAD(&tags->page_list);
3220 * rq_size is the size of the request plus driver payload, rounded
3221 * to the cacheline size
3223 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3225 left = rq_size * depth;
3227 for (i = 0; i < depth; ) {
3228 int this_order = max_order;
3233 while (this_order && left < order_to_size(this_order - 1))
3237 page = alloc_pages_node(node,
3238 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3244 if (order_to_size(this_order) < rq_size)
3251 page->private = this_order;
3252 list_add_tail(&page->lru, &tags->page_list);
3254 p = page_address(page);
3256 * Allow kmemleak to scan these pages as they contain pointers
3257 * to additional allocations like via ops->init_request().
3259 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3260 entries_per_page = order_to_size(this_order) / rq_size;
3261 to_do = min(entries_per_page, depth - i);
3262 left -= to_do * rq_size;
3263 for (j = 0; j < to_do; j++) {
3264 struct request *rq = p;
3266 tags->static_rqs[i] = rq;
3267 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3268 tags->static_rqs[i] = NULL;
3279 blk_mq_free_rqs(set, tags, hctx_idx);
3283 struct rq_iter_data {
3284 struct blk_mq_hw_ctx *hctx;
3288 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
3290 struct rq_iter_data *iter_data = data;
3292 if (rq->mq_hctx != iter_data->hctx)
3294 iter_data->has_rq = true;
3298 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3300 struct blk_mq_tags *tags = hctx->sched_tags ?
3301 hctx->sched_tags : hctx->tags;
3302 struct rq_iter_data data = {
3306 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3310 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3311 struct blk_mq_hw_ctx *hctx)
3313 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3315 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3320 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3322 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3323 struct blk_mq_hw_ctx, cpuhp_online);
3325 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3326 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3330 * Prevent new request from being allocated on the current hctx.
3332 * The smp_mb__after_atomic() Pairs with the implied barrier in
3333 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3334 * seen once we return from the tag allocator.
3336 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3337 smp_mb__after_atomic();
3340 * Try to grab a reference to the queue and wait for any outstanding
3341 * requests. If we could not grab a reference the queue has been
3342 * frozen and there are no requests.
3344 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3345 while (blk_mq_hctx_has_requests(hctx))
3347 percpu_ref_put(&hctx->queue->q_usage_counter);
3353 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3355 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3356 struct blk_mq_hw_ctx, cpuhp_online);
3358 if (cpumask_test_cpu(cpu, hctx->cpumask))
3359 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3364 * 'cpu' is going away. splice any existing rq_list entries from this
3365 * software queue to the hw queue dispatch list, and ensure that it
3368 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3370 struct blk_mq_hw_ctx *hctx;
3371 struct blk_mq_ctx *ctx;
3373 enum hctx_type type;
3375 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3376 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3379 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3382 spin_lock(&ctx->lock);
3383 if (!list_empty(&ctx->rq_lists[type])) {
3384 list_splice_init(&ctx->rq_lists[type], &tmp);
3385 blk_mq_hctx_clear_pending(hctx, ctx);
3387 spin_unlock(&ctx->lock);
3389 if (list_empty(&tmp))
3392 spin_lock(&hctx->lock);
3393 list_splice_tail_init(&tmp, &hctx->dispatch);
3394 spin_unlock(&hctx->lock);
3396 blk_mq_run_hw_queue(hctx, true);
3400 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3402 if (!(hctx->flags & BLK_MQ_F_STACKING))
3403 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3404 &hctx->cpuhp_online);
3405 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3410 * Before freeing hw queue, clearing the flush request reference in
3411 * tags->rqs[] for avoiding potential UAF.
3413 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3414 unsigned int queue_depth, struct request *flush_rq)
3417 unsigned long flags;
3419 /* The hw queue may not be mapped yet */
3423 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3425 for (i = 0; i < queue_depth; i++)
3426 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3429 * Wait until all pending iteration is done.
3431 * Request reference is cleared and it is guaranteed to be observed
3432 * after the ->lock is released.
3434 spin_lock_irqsave(&tags->lock, flags);
3435 spin_unlock_irqrestore(&tags->lock, flags);
3438 /* hctx->ctxs will be freed in queue's release handler */
3439 static void blk_mq_exit_hctx(struct request_queue *q,
3440 struct blk_mq_tag_set *set,
3441 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3443 struct request *flush_rq = hctx->fq->flush_rq;
3445 if (blk_mq_hw_queue_mapped(hctx))
3446 blk_mq_tag_idle(hctx);
3448 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3449 set->queue_depth, flush_rq);
3450 if (set->ops->exit_request)
3451 set->ops->exit_request(set, flush_rq, hctx_idx);
3453 if (set->ops->exit_hctx)
3454 set->ops->exit_hctx(hctx, hctx_idx);
3456 blk_mq_remove_cpuhp(hctx);
3458 xa_erase(&q->hctx_table, hctx_idx);
3460 spin_lock(&q->unused_hctx_lock);
3461 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3462 spin_unlock(&q->unused_hctx_lock);
3465 static void blk_mq_exit_hw_queues(struct request_queue *q,
3466 struct blk_mq_tag_set *set, int nr_queue)
3468 struct blk_mq_hw_ctx *hctx;
3471 queue_for_each_hw_ctx(q, hctx, i) {
3474 blk_mq_exit_hctx(q, set, hctx, i);
3478 static int blk_mq_init_hctx(struct request_queue *q,
3479 struct blk_mq_tag_set *set,
3480 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3482 hctx->queue_num = hctx_idx;
3484 if (!(hctx->flags & BLK_MQ_F_STACKING))
3485 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3486 &hctx->cpuhp_online);
3487 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3489 hctx->tags = set->tags[hctx_idx];
3491 if (set->ops->init_hctx &&
3492 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3493 goto unregister_cpu_notifier;
3495 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3499 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3505 if (set->ops->exit_request)
3506 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3508 if (set->ops->exit_hctx)
3509 set->ops->exit_hctx(hctx, hctx_idx);
3510 unregister_cpu_notifier:
3511 blk_mq_remove_cpuhp(hctx);
3515 static struct blk_mq_hw_ctx *
3516 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3519 struct blk_mq_hw_ctx *hctx;
3520 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3522 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3524 goto fail_alloc_hctx;
3526 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3529 atomic_set(&hctx->nr_active, 0);
3530 if (node == NUMA_NO_NODE)
3531 node = set->numa_node;
3532 hctx->numa_node = node;
3534 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3535 spin_lock_init(&hctx->lock);
3536 INIT_LIST_HEAD(&hctx->dispatch);
3538 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3540 INIT_LIST_HEAD(&hctx->hctx_list);
3543 * Allocate space for all possible cpus to avoid allocation at
3546 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3551 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3552 gfp, node, false, false))
3556 spin_lock_init(&hctx->dispatch_wait_lock);
3557 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3558 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3560 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3564 blk_mq_hctx_kobj_init(hctx);
3569 sbitmap_free(&hctx->ctx_map);
3573 free_cpumask_var(hctx->cpumask);
3580 static void blk_mq_init_cpu_queues(struct request_queue *q,
3581 unsigned int nr_hw_queues)
3583 struct blk_mq_tag_set *set = q->tag_set;
3586 for_each_possible_cpu(i) {
3587 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3588 struct blk_mq_hw_ctx *hctx;
3592 spin_lock_init(&__ctx->lock);
3593 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3594 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3599 * Set local node, IFF we have more than one hw queue. If
3600 * not, we remain on the home node of the device
3602 for (j = 0; j < set->nr_maps; j++) {
3603 hctx = blk_mq_map_queue_type(q, j, i);
3604 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3605 hctx->numa_node = cpu_to_node(i);
3610 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3611 unsigned int hctx_idx,
3614 struct blk_mq_tags *tags;
3617 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3621 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3623 blk_mq_free_rq_map(tags);
3630 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3633 if (blk_mq_is_shared_tags(set->flags)) {
3634 set->tags[hctx_idx] = set->shared_tags;
3639 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3642 return set->tags[hctx_idx];
3645 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3646 struct blk_mq_tags *tags,
3647 unsigned int hctx_idx)
3650 blk_mq_free_rqs(set, tags, hctx_idx);
3651 blk_mq_free_rq_map(tags);
3655 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3656 unsigned int hctx_idx)
3658 if (!blk_mq_is_shared_tags(set->flags))
3659 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3661 set->tags[hctx_idx] = NULL;
3664 static void blk_mq_map_swqueue(struct request_queue *q)
3666 unsigned int j, hctx_idx;
3668 struct blk_mq_hw_ctx *hctx;
3669 struct blk_mq_ctx *ctx;
3670 struct blk_mq_tag_set *set = q->tag_set;
3672 queue_for_each_hw_ctx(q, hctx, i) {
3673 cpumask_clear(hctx->cpumask);
3675 hctx->dispatch_from = NULL;
3679 * Map software to hardware queues.
3681 * If the cpu isn't present, the cpu is mapped to first hctx.
3683 for_each_possible_cpu(i) {
3685 ctx = per_cpu_ptr(q->queue_ctx, i);
3686 for (j = 0; j < set->nr_maps; j++) {
3687 if (!set->map[j].nr_queues) {
3688 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3689 HCTX_TYPE_DEFAULT, i);
3692 hctx_idx = set->map[j].mq_map[i];
3693 /* unmapped hw queue can be remapped after CPU topo changed */
3694 if (!set->tags[hctx_idx] &&
3695 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3697 * If tags initialization fail for some hctx,
3698 * that hctx won't be brought online. In this
3699 * case, remap the current ctx to hctx[0] which
3700 * is guaranteed to always have tags allocated
3702 set->map[j].mq_map[i] = 0;
3705 hctx = blk_mq_map_queue_type(q, j, i);
3706 ctx->hctxs[j] = hctx;
3708 * If the CPU is already set in the mask, then we've
3709 * mapped this one already. This can happen if
3710 * devices share queues across queue maps.
3712 if (cpumask_test_cpu(i, hctx->cpumask))
3715 cpumask_set_cpu(i, hctx->cpumask);
3717 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3718 hctx->ctxs[hctx->nr_ctx++] = ctx;
3721 * If the nr_ctx type overflows, we have exceeded the
3722 * amount of sw queues we can support.
3724 BUG_ON(!hctx->nr_ctx);
3727 for (; j < HCTX_MAX_TYPES; j++)
3728 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3729 HCTX_TYPE_DEFAULT, i);
3732 queue_for_each_hw_ctx(q, hctx, i) {
3734 * If no software queues are mapped to this hardware queue,
3735 * disable it and free the request entries.
3737 if (!hctx->nr_ctx) {
3738 /* Never unmap queue 0. We need it as a
3739 * fallback in case of a new remap fails
3743 __blk_mq_free_map_and_rqs(set, i);
3749 hctx->tags = set->tags[i];
3750 WARN_ON(!hctx->tags);
3753 * Set the map size to the number of mapped software queues.
3754 * This is more accurate and more efficient than looping
3755 * over all possibly mapped software queues.
3757 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3760 * Initialize batch roundrobin counts
3762 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3763 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3768 * Caller needs to ensure that we're either frozen/quiesced, or that
3769 * the queue isn't live yet.
3771 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3773 struct blk_mq_hw_ctx *hctx;
3776 queue_for_each_hw_ctx(q, hctx, i) {
3778 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3780 blk_mq_tag_idle(hctx);
3781 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3786 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3789 struct request_queue *q;
3791 lockdep_assert_held(&set->tag_list_lock);
3793 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3794 blk_mq_freeze_queue(q);
3795 queue_set_hctx_shared(q, shared);
3796 blk_mq_unfreeze_queue(q);
3800 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3802 struct blk_mq_tag_set *set = q->tag_set;
3804 mutex_lock(&set->tag_list_lock);
3805 list_del(&q->tag_set_list);
3806 if (list_is_singular(&set->tag_list)) {
3807 /* just transitioned to unshared */
3808 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3809 /* update existing queue */
3810 blk_mq_update_tag_set_shared(set, false);
3812 mutex_unlock(&set->tag_list_lock);
3813 INIT_LIST_HEAD(&q->tag_set_list);
3816 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3817 struct request_queue *q)
3819 mutex_lock(&set->tag_list_lock);
3822 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3824 if (!list_empty(&set->tag_list) &&
3825 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3826 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3827 /* update existing queue */
3828 blk_mq_update_tag_set_shared(set, true);
3830 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3831 queue_set_hctx_shared(q, true);
3832 list_add_tail(&q->tag_set_list, &set->tag_list);
3834 mutex_unlock(&set->tag_list_lock);
3837 /* All allocations will be freed in release handler of q->mq_kobj */
3838 static int blk_mq_alloc_ctxs(struct request_queue *q)
3840 struct blk_mq_ctxs *ctxs;
3843 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3847 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3848 if (!ctxs->queue_ctx)
3851 for_each_possible_cpu(cpu) {
3852 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3856 q->mq_kobj = &ctxs->kobj;
3857 q->queue_ctx = ctxs->queue_ctx;
3866 * It is the actual release handler for mq, but we do it from
3867 * request queue's release handler for avoiding use-after-free
3868 * and headache because q->mq_kobj shouldn't have been introduced,
3869 * but we can't group ctx/kctx kobj without it.
3871 void blk_mq_release(struct request_queue *q)
3873 struct blk_mq_hw_ctx *hctx, *next;
3876 queue_for_each_hw_ctx(q, hctx, i)
3877 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3879 /* all hctx are in .unused_hctx_list now */
3880 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3881 list_del_init(&hctx->hctx_list);
3882 kobject_put(&hctx->kobj);
3885 xa_destroy(&q->hctx_table);
3888 * release .mq_kobj and sw queue's kobject now because
3889 * both share lifetime with request queue.
3891 blk_mq_sysfs_deinit(q);
3894 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3897 struct request_queue *q;
3900 q = blk_alloc_queue(set->numa_node, set->flags & BLK_MQ_F_BLOCKING);
3902 return ERR_PTR(-ENOMEM);
3903 q->queuedata = queuedata;
3904 ret = blk_mq_init_allocated_queue(set, q);
3906 blk_cleanup_queue(q);
3907 return ERR_PTR(ret);
3912 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3914 return blk_mq_init_queue_data(set, NULL);
3916 EXPORT_SYMBOL(blk_mq_init_queue);
3918 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
3919 struct lock_class_key *lkclass)
3921 struct request_queue *q;
3922 struct gendisk *disk;
3924 q = blk_mq_init_queue_data(set, queuedata);
3928 disk = __alloc_disk_node(q, set->numa_node, lkclass);
3930 blk_cleanup_queue(q);
3931 return ERR_PTR(-ENOMEM);
3935 EXPORT_SYMBOL(__blk_mq_alloc_disk);
3937 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3938 struct blk_mq_tag_set *set, struct request_queue *q,
3939 int hctx_idx, int node)
3941 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3943 /* reuse dead hctx first */
3944 spin_lock(&q->unused_hctx_lock);
3945 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3946 if (tmp->numa_node == node) {
3952 list_del_init(&hctx->hctx_list);
3953 spin_unlock(&q->unused_hctx_lock);
3956 hctx = blk_mq_alloc_hctx(q, set, node);
3960 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3966 kobject_put(&hctx->kobj);
3971 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3972 struct request_queue *q)
3974 struct blk_mq_hw_ctx *hctx;
3977 /* protect against switching io scheduler */
3978 mutex_lock(&q->sysfs_lock);
3979 for (i = 0; i < set->nr_hw_queues; i++) {
3981 int node = blk_mq_get_hctx_node(set, i);
3982 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
3985 old_node = old_hctx->numa_node;
3986 blk_mq_exit_hctx(q, set, old_hctx, i);
3989 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
3992 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
3994 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
3995 WARN_ON_ONCE(!hctx);
3999 * Increasing nr_hw_queues fails. Free the newly allocated
4000 * hctxs and keep the previous q->nr_hw_queues.
4002 if (i != set->nr_hw_queues) {
4003 j = q->nr_hw_queues;
4006 q->nr_hw_queues = set->nr_hw_queues;
4009 xa_for_each_start(&q->hctx_table, j, hctx, j)
4010 blk_mq_exit_hctx(q, set, hctx, j);
4011 mutex_unlock(&q->sysfs_lock);
4014 static void blk_mq_update_poll_flag(struct request_queue *q)
4016 struct blk_mq_tag_set *set = q->tag_set;
4018 if (set->nr_maps > HCTX_TYPE_POLL &&
4019 set->map[HCTX_TYPE_POLL].nr_queues)
4020 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4022 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4025 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4026 struct request_queue *q)
4028 WARN_ON_ONCE(blk_queue_has_srcu(q) !=
4029 !!(set->flags & BLK_MQ_F_BLOCKING));
4031 /* mark the queue as mq asap */
4032 q->mq_ops = set->ops;
4034 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
4035 blk_mq_poll_stats_bkt,
4036 BLK_MQ_POLL_STATS_BKTS, q);
4040 if (blk_mq_alloc_ctxs(q))
4043 /* init q->mq_kobj and sw queues' kobjects */
4044 blk_mq_sysfs_init(q);
4046 INIT_LIST_HEAD(&q->unused_hctx_list);
4047 spin_lock_init(&q->unused_hctx_lock);
4049 xa_init(&q->hctx_table);
4051 blk_mq_realloc_hw_ctxs(set, q);
4052 if (!q->nr_hw_queues)
4055 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4056 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4060 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4061 blk_mq_update_poll_flag(q);
4063 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4064 INIT_LIST_HEAD(&q->requeue_list);
4065 spin_lock_init(&q->requeue_lock);
4067 q->nr_requests = set->queue_depth;
4070 * Default to classic polling
4072 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
4074 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4075 blk_mq_add_queue_tag_set(set, q);
4076 blk_mq_map_swqueue(q);
4080 xa_destroy(&q->hctx_table);
4081 q->nr_hw_queues = 0;
4082 blk_mq_sysfs_deinit(q);
4084 blk_stat_free_callback(q->poll_cb);
4090 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4092 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4093 void blk_mq_exit_queue(struct request_queue *q)
4095 struct blk_mq_tag_set *set = q->tag_set;
4097 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4098 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4099 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4100 blk_mq_del_queue_tag_set(q);
4103 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4107 if (blk_mq_is_shared_tags(set->flags)) {
4108 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4111 if (!set->shared_tags)
4115 for (i = 0; i < set->nr_hw_queues; i++) {
4116 if (!__blk_mq_alloc_map_and_rqs(set, i))
4125 __blk_mq_free_map_and_rqs(set, i);
4127 if (blk_mq_is_shared_tags(set->flags)) {
4128 blk_mq_free_map_and_rqs(set, set->shared_tags,
4129 BLK_MQ_NO_HCTX_IDX);
4136 * Allocate the request maps associated with this tag_set. Note that this
4137 * may reduce the depth asked for, if memory is tight. set->queue_depth
4138 * will be updated to reflect the allocated depth.
4140 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4145 depth = set->queue_depth;
4147 err = __blk_mq_alloc_rq_maps(set);
4151 set->queue_depth >>= 1;
4152 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4156 } while (set->queue_depth);
4158 if (!set->queue_depth || err) {
4159 pr_err("blk-mq: failed to allocate request map\n");
4163 if (depth != set->queue_depth)
4164 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4165 depth, set->queue_depth);
4170 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4173 * blk_mq_map_queues() and multiple .map_queues() implementations
4174 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4175 * number of hardware queues.
4177 if (set->nr_maps == 1)
4178 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4180 if (set->ops->map_queues && !is_kdump_kernel()) {
4184 * transport .map_queues is usually done in the following
4187 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4188 * mask = get_cpu_mask(queue)
4189 * for_each_cpu(cpu, mask)
4190 * set->map[x].mq_map[cpu] = queue;
4193 * When we need to remap, the table has to be cleared for
4194 * killing stale mapping since one CPU may not be mapped
4197 for (i = 0; i < set->nr_maps; i++)
4198 blk_mq_clear_mq_map(&set->map[i]);
4200 return set->ops->map_queues(set);
4202 BUG_ON(set->nr_maps > 1);
4203 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4207 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4208 int cur_nr_hw_queues, int new_nr_hw_queues)
4210 struct blk_mq_tags **new_tags;
4212 if (cur_nr_hw_queues >= new_nr_hw_queues)
4215 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4216 GFP_KERNEL, set->numa_node);
4221 memcpy(new_tags, set->tags, cur_nr_hw_queues *
4222 sizeof(*set->tags));
4224 set->tags = new_tags;
4225 set->nr_hw_queues = new_nr_hw_queues;
4230 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
4231 int new_nr_hw_queues)
4233 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
4237 * Alloc a tag set to be associated with one or more request queues.
4238 * May fail with EINVAL for various error conditions. May adjust the
4239 * requested depth down, if it's too large. In that case, the set
4240 * value will be stored in set->queue_depth.
4242 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4246 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4248 if (!set->nr_hw_queues)
4250 if (!set->queue_depth)
4252 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4255 if (!set->ops->queue_rq)
4258 if (!set->ops->get_budget ^ !set->ops->put_budget)
4261 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4262 pr_info("blk-mq: reduced tag depth to %u\n",
4264 set->queue_depth = BLK_MQ_MAX_DEPTH;
4269 else if (set->nr_maps > HCTX_MAX_TYPES)
4273 * If a crashdump is active, then we are potentially in a very
4274 * memory constrained environment. Limit us to 1 queue and
4275 * 64 tags to prevent using too much memory.
4277 if (is_kdump_kernel()) {
4278 set->nr_hw_queues = 1;
4280 set->queue_depth = min(64U, set->queue_depth);
4283 * There is no use for more h/w queues than cpus if we just have
4286 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4287 set->nr_hw_queues = nr_cpu_ids;
4289 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
4293 for (i = 0; i < set->nr_maps; i++) {
4294 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4295 sizeof(set->map[i].mq_map[0]),
4296 GFP_KERNEL, set->numa_node);
4297 if (!set->map[i].mq_map)
4298 goto out_free_mq_map;
4299 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4302 ret = blk_mq_update_queue_map(set);
4304 goto out_free_mq_map;
4306 ret = blk_mq_alloc_set_map_and_rqs(set);
4308 goto out_free_mq_map;
4310 mutex_init(&set->tag_list_lock);
4311 INIT_LIST_HEAD(&set->tag_list);
4316 for (i = 0; i < set->nr_maps; i++) {
4317 kfree(set->map[i].mq_map);
4318 set->map[i].mq_map = NULL;
4324 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4326 /* allocate and initialize a tagset for a simple single-queue device */
4327 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4328 const struct blk_mq_ops *ops, unsigned int queue_depth,
4329 unsigned int set_flags)
4331 memset(set, 0, sizeof(*set));
4333 set->nr_hw_queues = 1;
4335 set->queue_depth = queue_depth;
4336 set->numa_node = NUMA_NO_NODE;
4337 set->flags = set_flags;
4338 return blk_mq_alloc_tag_set(set);
4340 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4342 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4346 for (i = 0; i < set->nr_hw_queues; i++)
4347 __blk_mq_free_map_and_rqs(set, i);
4349 if (blk_mq_is_shared_tags(set->flags)) {
4350 blk_mq_free_map_and_rqs(set, set->shared_tags,
4351 BLK_MQ_NO_HCTX_IDX);
4354 for (j = 0; j < set->nr_maps; j++) {
4355 kfree(set->map[j].mq_map);
4356 set->map[j].mq_map = NULL;
4362 EXPORT_SYMBOL(blk_mq_free_tag_set);
4364 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4366 struct blk_mq_tag_set *set = q->tag_set;
4367 struct blk_mq_hw_ctx *hctx;
4374 if (q->nr_requests == nr)
4377 blk_mq_freeze_queue(q);
4378 blk_mq_quiesce_queue(q);
4381 queue_for_each_hw_ctx(q, hctx, i) {
4385 * If we're using an MQ scheduler, just update the scheduler
4386 * queue depth. This is similar to what the old code would do.
4388 if (hctx->sched_tags) {
4389 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4392 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4397 if (q->elevator && q->elevator->type->ops.depth_updated)
4398 q->elevator->type->ops.depth_updated(hctx);
4401 q->nr_requests = nr;
4402 if (blk_mq_is_shared_tags(set->flags)) {
4404 blk_mq_tag_update_sched_shared_tags(q);
4406 blk_mq_tag_resize_shared_tags(set, nr);
4410 blk_mq_unquiesce_queue(q);
4411 blk_mq_unfreeze_queue(q);
4417 * request_queue and elevator_type pair.
4418 * It is just used by __blk_mq_update_nr_hw_queues to cache
4419 * the elevator_type associated with a request_queue.
4421 struct blk_mq_qe_pair {
4422 struct list_head node;
4423 struct request_queue *q;
4424 struct elevator_type *type;
4428 * Cache the elevator_type in qe pair list and switch the
4429 * io scheduler to 'none'
4431 static bool blk_mq_elv_switch_none(struct list_head *head,
4432 struct request_queue *q)
4434 struct blk_mq_qe_pair *qe;
4439 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4443 INIT_LIST_HEAD(&qe->node);
4445 qe->type = q->elevator->type;
4446 list_add(&qe->node, head);
4448 mutex_lock(&q->sysfs_lock);
4450 * After elevator_switch_mq, the previous elevator_queue will be
4451 * released by elevator_release. The reference of the io scheduler
4452 * module get by elevator_get will also be put. So we need to get
4453 * a reference of the io scheduler module here to prevent it to be
4456 __module_get(qe->type->elevator_owner);
4457 elevator_switch_mq(q, NULL);
4458 mutex_unlock(&q->sysfs_lock);
4463 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4464 struct request_queue *q)
4466 struct blk_mq_qe_pair *qe;
4468 list_for_each_entry(qe, head, node)
4475 static void blk_mq_elv_switch_back(struct list_head *head,
4476 struct request_queue *q)
4478 struct blk_mq_qe_pair *qe;
4479 struct elevator_type *t;
4481 qe = blk_lookup_qe_pair(head, q);
4485 list_del(&qe->node);
4488 mutex_lock(&q->sysfs_lock);
4489 elevator_switch_mq(q, t);
4490 mutex_unlock(&q->sysfs_lock);
4493 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4496 struct request_queue *q;
4498 int prev_nr_hw_queues;
4500 lockdep_assert_held(&set->tag_list_lock);
4502 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4503 nr_hw_queues = nr_cpu_ids;
4504 if (nr_hw_queues < 1)
4506 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4509 list_for_each_entry(q, &set->tag_list, tag_set_list)
4510 blk_mq_freeze_queue(q);
4512 * Switch IO scheduler to 'none', cleaning up the data associated
4513 * with the previous scheduler. We will switch back once we are done
4514 * updating the new sw to hw queue mappings.
4516 list_for_each_entry(q, &set->tag_list, tag_set_list)
4517 if (!blk_mq_elv_switch_none(&head, q))
4520 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4521 blk_mq_debugfs_unregister_hctxs(q);
4522 blk_mq_sysfs_unregister(q);
4525 prev_nr_hw_queues = set->nr_hw_queues;
4526 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
4530 set->nr_hw_queues = nr_hw_queues;
4532 blk_mq_update_queue_map(set);
4533 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4534 blk_mq_realloc_hw_ctxs(set, q);
4535 blk_mq_update_poll_flag(q);
4536 if (q->nr_hw_queues != set->nr_hw_queues) {
4537 int i = prev_nr_hw_queues;
4539 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4540 nr_hw_queues, prev_nr_hw_queues);
4541 for (; i < set->nr_hw_queues; i++)
4542 __blk_mq_free_map_and_rqs(set, i);
4544 set->nr_hw_queues = prev_nr_hw_queues;
4545 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4548 blk_mq_map_swqueue(q);
4552 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4553 blk_mq_sysfs_register(q);
4554 blk_mq_debugfs_register_hctxs(q);
4558 list_for_each_entry(q, &set->tag_list, tag_set_list)
4559 blk_mq_elv_switch_back(&head, q);
4561 list_for_each_entry(q, &set->tag_list, tag_set_list)
4562 blk_mq_unfreeze_queue(q);
4565 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4567 mutex_lock(&set->tag_list_lock);
4568 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4569 mutex_unlock(&set->tag_list_lock);
4571 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4573 /* Enable polling stats and return whether they were already enabled. */
4574 static bool blk_poll_stats_enable(struct request_queue *q)
4579 return blk_stats_alloc_enable(q);
4582 static void blk_mq_poll_stats_start(struct request_queue *q)
4585 * We don't arm the callback if polling stats are not enabled or the
4586 * callback is already active.
4588 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
4591 blk_stat_activate_msecs(q->poll_cb, 100);
4594 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4596 struct request_queue *q = cb->data;
4599 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4600 if (cb->stat[bucket].nr_samples)
4601 q->poll_stat[bucket] = cb->stat[bucket];
4605 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4608 unsigned long ret = 0;
4612 * If stats collection isn't on, don't sleep but turn it on for
4615 if (!blk_poll_stats_enable(q))
4619 * As an optimistic guess, use half of the mean service time
4620 * for this type of request. We can (and should) make this smarter.
4621 * For instance, if the completion latencies are tight, we can
4622 * get closer than just half the mean. This is especially
4623 * important on devices where the completion latencies are longer
4624 * than ~10 usec. We do use the stats for the relevant IO size
4625 * if available which does lead to better estimates.
4627 bucket = blk_mq_poll_stats_bkt(rq);
4631 if (q->poll_stat[bucket].nr_samples)
4632 ret = (q->poll_stat[bucket].mean + 1) / 2;
4637 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4639 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4640 struct request *rq = blk_qc_to_rq(hctx, qc);
4641 struct hrtimer_sleeper hs;
4642 enum hrtimer_mode mode;
4647 * If a request has completed on queue that uses an I/O scheduler, we
4648 * won't get back a request from blk_qc_to_rq.
4650 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4654 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4656 * 0: use half of prev avg
4657 * >0: use this specific value
4659 if (q->poll_nsec > 0)
4660 nsecs = q->poll_nsec;
4662 nsecs = blk_mq_poll_nsecs(q, rq);
4667 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4670 * This will be replaced with the stats tracking code, using
4671 * 'avg_completion_time / 2' as the pre-sleep target.
4675 mode = HRTIMER_MODE_REL;
4676 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4677 hrtimer_set_expires(&hs.timer, kt);
4680 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4682 set_current_state(TASK_UNINTERRUPTIBLE);
4683 hrtimer_sleeper_start_expires(&hs, mode);
4686 hrtimer_cancel(&hs.timer);
4687 mode = HRTIMER_MODE_ABS;
4688 } while (hs.task && !signal_pending(current));
4690 __set_current_state(TASK_RUNNING);
4691 destroy_hrtimer_on_stack(&hs.timer);
4694 * If we sleep, have the caller restart the poll loop to reset the
4695 * state. Like for the other success return cases, the caller is
4696 * responsible for checking if the IO completed. If the IO isn't
4697 * complete, we'll get called again and will go straight to the busy
4703 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4704 struct io_comp_batch *iob, unsigned int flags)
4706 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4707 long state = get_current_state();
4711 ret = q->mq_ops->poll(hctx, iob);
4713 __set_current_state(TASK_RUNNING);
4717 if (signal_pending_state(state, current))
4718 __set_current_state(TASK_RUNNING);
4719 if (task_is_running(current))
4722 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4725 } while (!need_resched());
4727 __set_current_state(TASK_RUNNING);
4731 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4734 if (!(flags & BLK_POLL_NOSLEEP) &&
4735 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4736 if (blk_mq_poll_hybrid(q, cookie))
4739 return blk_mq_poll_classic(q, cookie, iob, flags);
4742 unsigned int blk_mq_rq_cpu(struct request *rq)
4744 return rq->mq_ctx->cpu;
4746 EXPORT_SYMBOL(blk_mq_rq_cpu);
4748 void blk_mq_cancel_work_sync(struct request_queue *q)
4750 if (queue_is_mq(q)) {
4751 struct blk_mq_hw_ctx *hctx;
4754 cancel_delayed_work_sync(&q->requeue_work);
4756 queue_for_each_hw_ctx(q, hctx, i)
4757 cancel_delayed_work_sync(&hctx->run_work);
4761 static int __init blk_mq_init(void)
4765 for_each_possible_cpu(i)
4766 init_llist_head(&per_cpu(blk_cpu_done, i));
4767 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4769 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4770 "block/softirq:dead", NULL,
4771 blk_softirq_cpu_dead);
4772 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4773 blk_mq_hctx_notify_dead);
4774 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4775 blk_mq_hctx_notify_online,
4776 blk_mq_hctx_notify_offline);
4779 subsys_initcall(blk_mq_init);