1 // SPDX-License-Identifier: GPL-2.0
3 * Block multiqueue core code
5 * Copyright (C) 2013-2014 Jens Axboe
6 * Copyright (C) 2013-2014 Christoph Hellwig
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/backing-dev.h>
11 #include <linux/bio.h>
12 #include <linux/blkdev.h>
13 #include <linux/blk-integrity.h>
14 #include <linux/kmemleak.h>
16 #include <linux/init.h>
17 #include <linux/slab.h>
18 #include <linux/workqueue.h>
19 #include <linux/smp.h>
20 #include <linux/interrupt.h>
21 #include <linux/llist.h>
22 #include <linux/cpu.h>
23 #include <linux/cache.h>
24 #include <linux/sched/sysctl.h>
25 #include <linux/sched/topology.h>
26 #include <linux/sched/signal.h>
27 #include <linux/delay.h>
28 #include <linux/crash_dump.h>
29 #include <linux/prefetch.h>
30 #include <linux/blk-crypto.h>
31 #include <linux/part_stat.h>
33 #include <trace/events/block.h>
35 #include <linux/blk-mq.h>
36 #include <linux/t10-pi.h>
39 #include "blk-mq-debugfs.h"
40 #include "blk-mq-tag.h"
43 #include "blk-mq-sched.h"
44 #include "blk-rq-qos.h"
45 #include "blk-ioprio.h"
47 static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
49 static void blk_mq_poll_stats_start(struct request_queue *q);
50 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
52 static int blk_mq_poll_stats_bkt(const struct request *rq)
54 int ddir, sectors, bucket;
56 ddir = rq_data_dir(rq);
57 sectors = blk_rq_stats_sectors(rq);
59 bucket = ddir + 2 * ilog2(sectors);
63 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
64 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
69 #define BLK_QC_T_SHIFT 16
70 #define BLK_QC_T_INTERNAL (1U << 31)
72 static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q,
75 return xa_load(&q->hctx_table,
76 (qc & ~BLK_QC_T_INTERNAL) >> BLK_QC_T_SHIFT);
79 static inline struct request *blk_qc_to_rq(struct blk_mq_hw_ctx *hctx,
82 unsigned int tag = qc & ((1U << BLK_QC_T_SHIFT) - 1);
84 if (qc & BLK_QC_T_INTERNAL)
85 return blk_mq_tag_to_rq(hctx->sched_tags, tag);
86 return blk_mq_tag_to_rq(hctx->tags, tag);
89 static inline blk_qc_t blk_rq_to_qc(struct request *rq)
91 return (rq->mq_hctx->queue_num << BLK_QC_T_SHIFT) |
93 rq->tag : (rq->internal_tag | BLK_QC_T_INTERNAL));
97 * Check if any of the ctx, dispatch list or elevator
98 * have pending work in this hardware queue.
100 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
102 return !list_empty_careful(&hctx->dispatch) ||
103 sbitmap_any_bit_set(&hctx->ctx_map) ||
104 blk_mq_sched_has_work(hctx);
108 * Mark this ctx as having pending work in this hardware queue
110 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
111 struct blk_mq_ctx *ctx)
113 const int bit = ctx->index_hw[hctx->type];
115 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
116 sbitmap_set_bit(&hctx->ctx_map, bit);
119 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
120 struct blk_mq_ctx *ctx)
122 const int bit = ctx->index_hw[hctx->type];
124 sbitmap_clear_bit(&hctx->ctx_map, bit);
128 struct block_device *part;
129 unsigned int inflight[2];
132 static bool blk_mq_check_inflight(struct request *rq, void *priv)
134 struct mq_inflight *mi = priv;
136 if (rq->part && blk_do_io_stat(rq) &&
137 (!mi->part->bd_partno || rq->part == mi->part) &&
138 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
139 mi->inflight[rq_data_dir(rq)]++;
144 unsigned int blk_mq_in_flight(struct request_queue *q,
145 struct block_device *part)
147 struct mq_inflight mi = { .part = part };
149 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
151 return mi.inflight[0] + mi.inflight[1];
154 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
155 unsigned int inflight[2])
157 struct mq_inflight mi = { .part = part };
159 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
160 inflight[0] = mi.inflight[0];
161 inflight[1] = mi.inflight[1];
164 void blk_freeze_queue_start(struct request_queue *q)
166 mutex_lock(&q->mq_freeze_lock);
167 if (++q->mq_freeze_depth == 1) {
168 percpu_ref_kill(&q->q_usage_counter);
169 mutex_unlock(&q->mq_freeze_lock);
171 blk_mq_run_hw_queues(q, false);
173 mutex_unlock(&q->mq_freeze_lock);
176 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
178 void blk_mq_freeze_queue_wait(struct request_queue *q)
180 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
182 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
184 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
185 unsigned long timeout)
187 return wait_event_timeout(q->mq_freeze_wq,
188 percpu_ref_is_zero(&q->q_usage_counter),
191 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
194 * Guarantee no request is in use, so we can change any data structure of
195 * the queue afterward.
197 void blk_freeze_queue(struct request_queue *q)
200 * In the !blk_mq case we are only calling this to kill the
201 * q_usage_counter, otherwise this increases the freeze depth
202 * and waits for it to return to zero. For this reason there is
203 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
204 * exported to drivers as the only user for unfreeze is blk_mq.
206 blk_freeze_queue_start(q);
207 blk_mq_freeze_queue_wait(q);
210 void blk_mq_freeze_queue(struct request_queue *q)
213 * ...just an alias to keep freeze and unfreeze actions balanced
214 * in the blk_mq_* namespace
218 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
220 void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
222 mutex_lock(&q->mq_freeze_lock);
224 q->q_usage_counter.data->force_atomic = true;
225 q->mq_freeze_depth--;
226 WARN_ON_ONCE(q->mq_freeze_depth < 0);
227 if (!q->mq_freeze_depth) {
228 percpu_ref_resurrect(&q->q_usage_counter);
229 wake_up_all(&q->mq_freeze_wq);
231 mutex_unlock(&q->mq_freeze_lock);
234 void blk_mq_unfreeze_queue(struct request_queue *q)
236 __blk_mq_unfreeze_queue(q, false);
238 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
241 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
242 * mpt3sas driver such that this function can be removed.
244 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
248 spin_lock_irqsave(&q->queue_lock, flags);
249 if (!q->quiesce_depth++)
250 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
251 spin_unlock_irqrestore(&q->queue_lock, flags);
253 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
256 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
259 * Note: it is driver's responsibility for making sure that quiesce has
262 void blk_mq_wait_quiesce_done(struct request_queue *q)
264 if (blk_queue_has_srcu(q))
265 synchronize_srcu(q->srcu);
269 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
272 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
275 * Note: this function does not prevent that the struct request end_io()
276 * callback function is invoked. Once this function is returned, we make
277 * sure no dispatch can happen until the queue is unquiesced via
278 * blk_mq_unquiesce_queue().
280 void blk_mq_quiesce_queue(struct request_queue *q)
282 blk_mq_quiesce_queue_nowait(q);
283 blk_mq_wait_quiesce_done(q);
285 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
288 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
291 * This function recovers queue into the state before quiescing
292 * which is done by blk_mq_quiesce_queue.
294 void blk_mq_unquiesce_queue(struct request_queue *q)
297 bool run_queue = false;
299 spin_lock_irqsave(&q->queue_lock, flags);
300 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
302 } else if (!--q->quiesce_depth) {
303 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
306 spin_unlock_irqrestore(&q->queue_lock, flags);
308 /* dispatch requests which are inserted during quiescing */
310 blk_mq_run_hw_queues(q, true);
312 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
314 void blk_mq_wake_waiters(struct request_queue *q)
316 struct blk_mq_hw_ctx *hctx;
319 queue_for_each_hw_ctx(q, hctx, i)
320 if (blk_mq_hw_queue_mapped(hctx))
321 blk_mq_tag_wakeup_all(hctx->tags, true);
324 void blk_rq_init(struct request_queue *q, struct request *rq)
326 memset(rq, 0, sizeof(*rq));
328 INIT_LIST_HEAD(&rq->queuelist);
330 rq->__sector = (sector_t) -1;
331 INIT_HLIST_NODE(&rq->hash);
332 RB_CLEAR_NODE(&rq->rb_node);
333 rq->tag = BLK_MQ_NO_TAG;
334 rq->internal_tag = BLK_MQ_NO_TAG;
335 rq->start_time_ns = ktime_get_ns();
337 blk_crypto_rq_set_defaults(rq);
339 EXPORT_SYMBOL(blk_rq_init);
341 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
342 struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns)
344 struct blk_mq_ctx *ctx = data->ctx;
345 struct blk_mq_hw_ctx *hctx = data->hctx;
346 struct request_queue *q = data->q;
347 struct request *rq = tags->static_rqs[tag];
352 rq->cmd_flags = data->cmd_flags;
354 if (data->flags & BLK_MQ_REQ_PM)
355 data->rq_flags |= RQF_PM;
356 if (blk_queue_io_stat(q))
357 data->rq_flags |= RQF_IO_STAT;
358 rq->rq_flags = data->rq_flags;
360 if (!(data->rq_flags & RQF_ELV)) {
362 rq->internal_tag = BLK_MQ_NO_TAG;
364 rq->tag = BLK_MQ_NO_TAG;
365 rq->internal_tag = tag;
369 if (blk_mq_need_time_stamp(rq))
370 rq->start_time_ns = ktime_get_ns();
372 rq->start_time_ns = 0;
374 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
375 rq->alloc_time_ns = alloc_time_ns;
377 rq->io_start_time_ns = 0;
378 rq->stats_sectors = 0;
379 rq->nr_phys_segments = 0;
380 #if defined(CONFIG_BLK_DEV_INTEGRITY)
381 rq->nr_integrity_segments = 0;
384 rq->end_io_data = NULL;
386 blk_crypto_rq_set_defaults(rq);
387 INIT_LIST_HEAD(&rq->queuelist);
388 /* tag was already set */
389 WRITE_ONCE(rq->deadline, 0);
392 if (rq->rq_flags & RQF_ELV) {
393 struct elevator_queue *e = data->q->elevator;
395 INIT_HLIST_NODE(&rq->hash);
396 RB_CLEAR_NODE(&rq->rb_node);
398 if (!op_is_flush(data->cmd_flags) &&
399 e->type->ops.prepare_request) {
400 e->type->ops.prepare_request(rq);
401 rq->rq_flags |= RQF_ELVPRIV;
408 static inline struct request *
409 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data,
412 unsigned int tag, tag_offset;
413 struct blk_mq_tags *tags;
415 unsigned long tag_mask;
418 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
419 if (unlikely(!tag_mask))
422 tags = blk_mq_tags_from_data(data);
423 for (i = 0; tag_mask; i++) {
424 if (!(tag_mask & (1UL << i)))
426 tag = tag_offset + i;
427 prefetch(tags->static_rqs[tag]);
428 tag_mask &= ~(1UL << i);
429 rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns);
430 rq_list_add(data->cached_rq, rq);
433 /* caller already holds a reference, add for remainder */
434 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
437 return rq_list_pop(data->cached_rq);
440 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
442 struct request_queue *q = data->q;
443 u64 alloc_time_ns = 0;
447 /* alloc_time includes depth and tag waits */
448 if (blk_queue_rq_alloc_time(q))
449 alloc_time_ns = ktime_get_ns();
451 if (data->cmd_flags & REQ_NOWAIT)
452 data->flags |= BLK_MQ_REQ_NOWAIT;
455 struct elevator_queue *e = q->elevator;
457 data->rq_flags |= RQF_ELV;
460 * Flush/passthrough requests are special and go directly to the
461 * dispatch list. Don't include reserved tags in the
462 * limiting, as it isn't useful.
464 if (!op_is_flush(data->cmd_flags) &&
465 !blk_op_is_passthrough(data->cmd_flags) &&
466 e->type->ops.limit_depth &&
467 !(data->flags & BLK_MQ_REQ_RESERVED))
468 e->type->ops.limit_depth(data->cmd_flags, data);
472 data->ctx = blk_mq_get_ctx(q);
473 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
474 if (!(data->rq_flags & RQF_ELV))
475 blk_mq_tag_busy(data->hctx);
477 if (data->flags & BLK_MQ_REQ_RESERVED)
478 data->rq_flags |= RQF_RESV;
481 * Try batched alloc if we want more than 1 tag.
483 if (data->nr_tags > 1) {
484 rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns);
491 * Waiting allocations only fail because of an inactive hctx. In that
492 * case just retry the hctx assignment and tag allocation as CPU hotplug
493 * should have migrated us to an online CPU by now.
495 tag = blk_mq_get_tag(data);
496 if (tag == BLK_MQ_NO_TAG) {
497 if (data->flags & BLK_MQ_REQ_NOWAIT)
500 * Give up the CPU and sleep for a random short time to
501 * ensure that thread using a realtime scheduling class
502 * are migrated off the CPU, and thus off the hctx that
509 return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag,
513 static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
514 struct blk_plug *plug,
516 blk_mq_req_flags_t flags)
518 struct blk_mq_alloc_data data = {
522 .nr_tags = plug->nr_ios,
523 .cached_rq = &plug->cached_rq,
527 if (blk_queue_enter(q, flags))
532 rq = __blk_mq_alloc_requests(&data);
538 static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
540 blk_mq_req_flags_t flags)
542 struct blk_plug *plug = current->plug;
547 if (rq_list_empty(plug->cached_rq)) {
548 if (plug->nr_ios == 1)
550 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
555 rq = rq_list_peek(&plug->cached_rq);
556 if (!rq || rq->q != q)
559 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
561 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
564 plug->cached_rq = rq_list_next(rq);
567 INIT_LIST_HEAD(&rq->queuelist);
571 struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
572 blk_mq_req_flags_t flags)
576 rq = blk_mq_alloc_cached_request(q, opf, flags);
578 struct blk_mq_alloc_data data = {
586 ret = blk_queue_enter(q, flags);
590 rq = __blk_mq_alloc_requests(&data);
595 rq->__sector = (sector_t) -1;
596 rq->bio = rq->biotail = NULL;
600 return ERR_PTR(-EWOULDBLOCK);
602 EXPORT_SYMBOL(blk_mq_alloc_request);
604 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
605 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
607 struct blk_mq_alloc_data data = {
613 u64 alloc_time_ns = 0;
618 /* alloc_time includes depth and tag waits */
619 if (blk_queue_rq_alloc_time(q))
620 alloc_time_ns = ktime_get_ns();
623 * If the tag allocator sleeps we could get an allocation for a
624 * different hardware context. No need to complicate the low level
625 * allocator for this for the rare use case of a command tied to
628 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
629 return ERR_PTR(-EINVAL);
631 if (hctx_idx >= q->nr_hw_queues)
632 return ERR_PTR(-EIO);
634 ret = blk_queue_enter(q, flags);
639 * Check if the hardware context is actually mapped to anything.
640 * If not tell the caller that it should skip this queue.
643 data.hctx = xa_load(&q->hctx_table, hctx_idx);
644 if (!blk_mq_hw_queue_mapped(data.hctx))
646 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
647 if (cpu >= nr_cpu_ids)
649 data.ctx = __blk_mq_get_ctx(q, cpu);
652 blk_mq_tag_busy(data.hctx);
654 data.rq_flags |= RQF_ELV;
656 if (flags & BLK_MQ_REQ_RESERVED)
657 data.rq_flags |= RQF_RESV;
660 tag = blk_mq_get_tag(&data);
661 if (tag == BLK_MQ_NO_TAG)
663 return blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag,
670 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
672 static void __blk_mq_free_request(struct request *rq)
674 struct request_queue *q = rq->q;
675 struct blk_mq_ctx *ctx = rq->mq_ctx;
676 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
677 const int sched_tag = rq->internal_tag;
679 blk_crypto_free_request(rq);
680 blk_pm_mark_last_busy(rq);
682 if (rq->tag != BLK_MQ_NO_TAG)
683 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
684 if (sched_tag != BLK_MQ_NO_TAG)
685 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
686 blk_mq_sched_restart(hctx);
690 void blk_mq_free_request(struct request *rq)
692 struct request_queue *q = rq->q;
693 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
695 if ((rq->rq_flags & RQF_ELVPRIV) &&
696 q->elevator->type->ops.finish_request)
697 q->elevator->type->ops.finish_request(rq);
699 if (rq->rq_flags & RQF_MQ_INFLIGHT)
700 __blk_mq_dec_active_requests(hctx);
702 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
703 laptop_io_completion(q->disk->bdi);
707 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
708 if (req_ref_put_and_test(rq))
709 __blk_mq_free_request(rq);
711 EXPORT_SYMBOL_GPL(blk_mq_free_request);
713 void blk_mq_free_plug_rqs(struct blk_plug *plug)
717 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
718 blk_mq_free_request(rq);
721 void blk_dump_rq_flags(struct request *rq, char *msg)
723 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
724 rq->q->disk ? rq->q->disk->disk_name : "?",
725 (__force unsigned long long) rq->cmd_flags);
727 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
728 (unsigned long long)blk_rq_pos(rq),
729 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
730 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
731 rq->bio, rq->biotail, blk_rq_bytes(rq));
733 EXPORT_SYMBOL(blk_dump_rq_flags);
735 static void req_bio_endio(struct request *rq, struct bio *bio,
736 unsigned int nbytes, blk_status_t error)
738 if (unlikely(error)) {
739 bio->bi_status = error;
740 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
742 * Partial zone append completions cannot be supported as the
743 * BIO fragments may end up not being written sequentially.
745 if (bio->bi_iter.bi_size != nbytes)
746 bio->bi_status = BLK_STS_IOERR;
748 bio->bi_iter.bi_sector = rq->__sector;
751 bio_advance(bio, nbytes);
753 if (unlikely(rq->rq_flags & RQF_QUIET))
754 bio_set_flag(bio, BIO_QUIET);
755 /* don't actually finish bio if it's part of flush sequence */
756 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
760 static void blk_account_io_completion(struct request *req, unsigned int bytes)
762 if (req->part && blk_do_io_stat(req)) {
763 const int sgrp = op_stat_group(req_op(req));
766 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
771 static void blk_print_req_error(struct request *req, blk_status_t status)
773 printk_ratelimited(KERN_ERR
774 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
775 "phys_seg %u prio class %u\n",
776 blk_status_to_str(status),
777 req->q->disk ? req->q->disk->disk_name : "?",
778 blk_rq_pos(req), (__force u32)req_op(req),
779 blk_op_str(req_op(req)),
780 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
781 req->nr_phys_segments,
782 IOPRIO_PRIO_CLASS(req->ioprio));
786 * Fully end IO on a request. Does not support partial completions, or
789 static void blk_complete_request(struct request *req)
791 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
792 int total_bytes = blk_rq_bytes(req);
793 struct bio *bio = req->bio;
795 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
800 #ifdef CONFIG_BLK_DEV_INTEGRITY
801 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
802 req->q->integrity.profile->complete_fn(req, total_bytes);
805 blk_account_io_completion(req, total_bytes);
808 struct bio *next = bio->bi_next;
810 /* Completion has already been traced */
811 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
813 if (req_op(req) == REQ_OP_ZONE_APPEND)
814 bio->bi_iter.bi_sector = req->__sector;
822 * Reset counters so that the request stacking driver
823 * can find how many bytes remain in the request
831 * blk_update_request - Complete multiple bytes without completing the request
832 * @req: the request being processed
833 * @error: block status code
834 * @nr_bytes: number of bytes to complete for @req
837 * Ends I/O on a number of bytes attached to @req, but doesn't complete
838 * the request structure even if @req doesn't have leftover.
839 * If @req has leftover, sets it up for the next range of segments.
841 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
842 * %false return from this function.
845 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
846 * except in the consistency check at the end of this function.
849 * %false - this request doesn't have any more data
850 * %true - this request has more data
852 bool blk_update_request(struct request *req, blk_status_t error,
853 unsigned int nr_bytes)
857 trace_block_rq_complete(req, error, nr_bytes);
862 #ifdef CONFIG_BLK_DEV_INTEGRITY
863 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
865 req->q->integrity.profile->complete_fn(req, nr_bytes);
868 if (unlikely(error && !blk_rq_is_passthrough(req) &&
869 !(req->rq_flags & RQF_QUIET)) &&
870 !test_bit(GD_DEAD, &req->q->disk->state)) {
871 blk_print_req_error(req, error);
872 trace_block_rq_error(req, error, nr_bytes);
875 blk_account_io_completion(req, nr_bytes);
879 struct bio *bio = req->bio;
880 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
882 if (bio_bytes == bio->bi_iter.bi_size)
883 req->bio = bio->bi_next;
885 /* Completion has already been traced */
886 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
887 req_bio_endio(req, bio, bio_bytes, error);
889 total_bytes += bio_bytes;
890 nr_bytes -= bio_bytes;
901 * Reset counters so that the request stacking driver
902 * can find how many bytes remain in the request
909 req->__data_len -= total_bytes;
911 /* update sector only for requests with clear definition of sector */
912 if (!blk_rq_is_passthrough(req))
913 req->__sector += total_bytes >> 9;
915 /* mixed attributes always follow the first bio */
916 if (req->rq_flags & RQF_MIXED_MERGE) {
917 req->cmd_flags &= ~REQ_FAILFAST_MASK;
918 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
921 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
923 * If total number of sectors is less than the first segment
924 * size, something has gone terribly wrong.
926 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
927 blk_dump_rq_flags(req, "request botched");
928 req->__data_len = blk_rq_cur_bytes(req);
931 /* recalculate the number of segments */
932 req->nr_phys_segments = blk_recalc_rq_segments(req);
937 EXPORT_SYMBOL_GPL(blk_update_request);
939 static void __blk_account_io_done(struct request *req, u64 now)
941 const int sgrp = op_stat_group(req_op(req));
944 update_io_ticks(req->part, jiffies, true);
945 part_stat_inc(req->part, ios[sgrp]);
946 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
950 static inline void blk_account_io_done(struct request *req, u64 now)
953 * Account IO completion. flush_rq isn't accounted as a
954 * normal IO on queueing nor completion. Accounting the
955 * containing request is enough.
957 if (blk_do_io_stat(req) && req->part &&
958 !(req->rq_flags & RQF_FLUSH_SEQ))
959 __blk_account_io_done(req, now);
962 static void __blk_account_io_start(struct request *rq)
965 * All non-passthrough requests are created from a bio with one
966 * exception: when a flush command that is part of a flush sequence
967 * generated by the state machine in blk-flush.c is cloned onto the
968 * lower device by dm-multipath we can get here without a bio.
971 rq->part = rq->bio->bi_bdev;
973 rq->part = rq->q->disk->part0;
976 update_io_ticks(rq->part, jiffies, false);
980 static inline void blk_account_io_start(struct request *req)
982 if (blk_do_io_stat(req))
983 __blk_account_io_start(req);
986 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
988 if (rq->rq_flags & RQF_STATS) {
989 blk_mq_poll_stats_start(rq->q);
990 blk_stat_add(rq, now);
993 blk_mq_sched_completed_request(rq, now);
994 blk_account_io_done(rq, now);
997 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
999 if (blk_mq_need_time_stamp(rq))
1000 __blk_mq_end_request_acct(rq, ktime_get_ns());
1003 rq_qos_done(rq->q, rq);
1004 rq->end_io(rq, error);
1006 blk_mq_free_request(rq);
1009 EXPORT_SYMBOL(__blk_mq_end_request);
1011 void blk_mq_end_request(struct request *rq, blk_status_t error)
1013 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1015 __blk_mq_end_request(rq, error);
1017 EXPORT_SYMBOL(blk_mq_end_request);
1019 #define TAG_COMP_BATCH 32
1021 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1022 int *tag_array, int nr_tags)
1024 struct request_queue *q = hctx->queue;
1027 * All requests should have been marked as RQF_MQ_INFLIGHT, so
1028 * update hctx->nr_active in batch
1030 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1031 __blk_mq_sub_active_requests(hctx, nr_tags);
1033 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1034 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1037 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1039 int tags[TAG_COMP_BATCH], nr_tags = 0;
1040 struct blk_mq_hw_ctx *cur_hctx = NULL;
1045 now = ktime_get_ns();
1047 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1049 prefetch(rq->rq_next);
1051 blk_complete_request(rq);
1053 __blk_mq_end_request_acct(rq, now);
1055 rq_qos_done(rq->q, rq);
1057 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1058 if (!req_ref_put_and_test(rq))
1061 blk_crypto_free_request(rq);
1062 blk_pm_mark_last_busy(rq);
1064 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1066 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1068 cur_hctx = rq->mq_hctx;
1070 tags[nr_tags++] = rq->tag;
1074 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1076 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1078 static void blk_complete_reqs(struct llist_head *list)
1080 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1081 struct request *rq, *next;
1083 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1084 rq->q->mq_ops->complete(rq);
1087 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1089 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1092 static int blk_softirq_cpu_dead(unsigned int cpu)
1094 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1098 static void __blk_mq_complete_request_remote(void *data)
1100 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1103 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1105 int cpu = raw_smp_processor_id();
1107 if (!IS_ENABLED(CONFIG_SMP) ||
1108 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1111 * With force threaded interrupts enabled, raising softirq from an SMP
1112 * function call will always result in waking the ksoftirqd thread.
1113 * This is probably worse than completing the request on a different
1116 if (force_irqthreads())
1119 /* same CPU or cache domain? Complete locally */
1120 if (cpu == rq->mq_ctx->cpu ||
1121 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1122 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1125 /* don't try to IPI to an offline CPU */
1126 return cpu_online(rq->mq_ctx->cpu);
1129 static void blk_mq_complete_send_ipi(struct request *rq)
1131 struct llist_head *list;
1134 cpu = rq->mq_ctx->cpu;
1135 list = &per_cpu(blk_cpu_done, cpu);
1136 if (llist_add(&rq->ipi_list, list)) {
1137 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1138 smp_call_function_single_async(cpu, &rq->csd);
1142 static void blk_mq_raise_softirq(struct request *rq)
1144 struct llist_head *list;
1147 list = this_cpu_ptr(&blk_cpu_done);
1148 if (llist_add(&rq->ipi_list, list))
1149 raise_softirq(BLOCK_SOFTIRQ);
1153 bool blk_mq_complete_request_remote(struct request *rq)
1155 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1158 * For request which hctx has only one ctx mapping,
1159 * or a polled request, always complete locally,
1160 * it's pointless to redirect the completion.
1162 if (rq->mq_hctx->nr_ctx == 1 ||
1163 rq->cmd_flags & REQ_POLLED)
1166 if (blk_mq_complete_need_ipi(rq)) {
1167 blk_mq_complete_send_ipi(rq);
1171 if (rq->q->nr_hw_queues == 1) {
1172 blk_mq_raise_softirq(rq);
1177 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1180 * blk_mq_complete_request - end I/O on a request
1181 * @rq: the request being processed
1184 * Complete a request by scheduling the ->complete_rq operation.
1186 void blk_mq_complete_request(struct request *rq)
1188 if (!blk_mq_complete_request_remote(rq))
1189 rq->q->mq_ops->complete(rq);
1191 EXPORT_SYMBOL(blk_mq_complete_request);
1194 * blk_mq_start_request - Start processing a request
1195 * @rq: Pointer to request to be started
1197 * Function used by device drivers to notify the block layer that a request
1198 * is going to be processed now, so blk layer can do proper initializations
1199 * such as starting the timeout timer.
1201 void blk_mq_start_request(struct request *rq)
1203 struct request_queue *q = rq->q;
1205 trace_block_rq_issue(rq);
1207 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1208 rq->io_start_time_ns = ktime_get_ns();
1209 rq->stats_sectors = blk_rq_sectors(rq);
1210 rq->rq_flags |= RQF_STATS;
1211 rq_qos_issue(q, rq);
1214 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1217 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1219 #ifdef CONFIG_BLK_DEV_INTEGRITY
1220 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1221 q->integrity.profile->prepare_fn(rq);
1223 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1224 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1226 EXPORT_SYMBOL(blk_mq_start_request);
1229 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1230 * queues. This is important for md arrays to benefit from merging
1233 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1235 if (plug->multiple_queues)
1236 return BLK_MAX_REQUEST_COUNT * 2;
1237 return BLK_MAX_REQUEST_COUNT;
1240 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1242 struct request *last = rq_list_peek(&plug->mq_list);
1244 if (!plug->rq_count) {
1245 trace_block_plug(rq->q);
1246 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1247 (!blk_queue_nomerges(rq->q) &&
1248 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1249 blk_mq_flush_plug_list(plug, false);
1250 trace_block_plug(rq->q);
1253 if (!plug->multiple_queues && last && last->q != rq->q)
1254 plug->multiple_queues = true;
1255 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1256 plug->has_elevator = true;
1258 rq_list_add(&plug->mq_list, rq);
1263 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1264 * @rq: request to insert
1265 * @at_head: insert request at head or tail of queue
1268 * Insert a fully prepared request at the back of the I/O scheduler queue
1269 * for execution. Don't wait for completion.
1272 * This function will invoke @done directly if the queue is dead.
1274 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1276 WARN_ON(irqs_disabled());
1277 WARN_ON(!blk_rq_is_passthrough(rq));
1279 blk_account_io_start(rq);
1282 * As plugging can be enabled for passthrough requests on a zoned
1283 * device, directly accessing the plug instead of using blk_mq_plug()
1284 * should not have any consequences.
1287 blk_add_rq_to_plug(current->plug, rq);
1289 blk_mq_sched_insert_request(rq, at_head, true, false);
1291 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1293 struct blk_rq_wait {
1294 struct completion done;
1298 static void blk_end_sync_rq(struct request *rq, blk_status_t ret)
1300 struct blk_rq_wait *wait = rq->end_io_data;
1303 complete(&wait->done);
1306 bool blk_rq_is_poll(struct request *rq)
1310 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1312 if (WARN_ON_ONCE(!rq->bio))
1316 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1318 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1321 bio_poll(rq->bio, NULL, 0);
1323 } while (!completion_done(wait));
1327 * blk_execute_rq - insert a request into queue for execution
1328 * @rq: request to insert
1329 * @at_head: insert request at head or tail of queue
1332 * Insert a fully prepared request at the back of the I/O scheduler queue
1333 * for execution and wait for completion.
1334 * Return: The blk_status_t result provided to blk_mq_end_request().
1336 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1338 struct blk_rq_wait wait = {
1339 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1342 WARN_ON(irqs_disabled());
1343 WARN_ON(!blk_rq_is_passthrough(rq));
1345 rq->end_io_data = &wait;
1346 rq->end_io = blk_end_sync_rq;
1348 blk_account_io_start(rq);
1349 blk_mq_sched_insert_request(rq, at_head, true, false);
1351 if (blk_rq_is_poll(rq)) {
1352 blk_rq_poll_completion(rq, &wait.done);
1355 * Prevent hang_check timer from firing at us during very long
1358 unsigned long hang_check = sysctl_hung_task_timeout_secs;
1361 while (!wait_for_completion_io_timeout(&wait.done,
1362 hang_check * (HZ/2)))
1365 wait_for_completion_io(&wait.done);
1370 EXPORT_SYMBOL(blk_execute_rq);
1372 static void __blk_mq_requeue_request(struct request *rq)
1374 struct request_queue *q = rq->q;
1376 blk_mq_put_driver_tag(rq);
1378 trace_block_rq_requeue(rq);
1379 rq_qos_requeue(q, rq);
1381 if (blk_mq_request_started(rq)) {
1382 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1383 rq->rq_flags &= ~RQF_TIMED_OUT;
1387 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1389 __blk_mq_requeue_request(rq);
1391 /* this request will be re-inserted to io scheduler queue */
1392 blk_mq_sched_requeue_request(rq);
1394 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1396 EXPORT_SYMBOL(blk_mq_requeue_request);
1398 static void blk_mq_requeue_work(struct work_struct *work)
1400 struct request_queue *q =
1401 container_of(work, struct request_queue, requeue_work.work);
1403 struct request *rq, *next;
1405 spin_lock_irq(&q->requeue_lock);
1406 list_splice_init(&q->requeue_list, &rq_list);
1407 spin_unlock_irq(&q->requeue_lock);
1409 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1410 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1413 rq->rq_flags &= ~RQF_SOFTBARRIER;
1414 list_del_init(&rq->queuelist);
1416 * If RQF_DONTPREP, rq has contained some driver specific
1417 * data, so insert it to hctx dispatch list to avoid any
1420 if (rq->rq_flags & RQF_DONTPREP)
1421 blk_mq_request_bypass_insert(rq, false, false);
1423 blk_mq_sched_insert_request(rq, true, false, false);
1426 while (!list_empty(&rq_list)) {
1427 rq = list_entry(rq_list.next, struct request, queuelist);
1428 list_del_init(&rq->queuelist);
1429 blk_mq_sched_insert_request(rq, false, false, false);
1432 blk_mq_run_hw_queues(q, false);
1435 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1436 bool kick_requeue_list)
1438 struct request_queue *q = rq->q;
1439 unsigned long flags;
1442 * We abuse this flag that is otherwise used by the I/O scheduler to
1443 * request head insertion from the workqueue.
1445 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1447 spin_lock_irqsave(&q->requeue_lock, flags);
1449 rq->rq_flags |= RQF_SOFTBARRIER;
1450 list_add(&rq->queuelist, &q->requeue_list);
1452 list_add_tail(&rq->queuelist, &q->requeue_list);
1454 spin_unlock_irqrestore(&q->requeue_lock, flags);
1456 if (kick_requeue_list)
1457 blk_mq_kick_requeue_list(q);
1460 void blk_mq_kick_requeue_list(struct request_queue *q)
1462 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1464 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1466 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1467 unsigned long msecs)
1469 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1470 msecs_to_jiffies(msecs));
1472 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1474 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1477 * If we find a request that isn't idle we know the queue is busy
1478 * as it's checked in the iter.
1479 * Return false to stop the iteration.
1481 if (blk_mq_request_started(rq)) {
1491 bool blk_mq_queue_inflight(struct request_queue *q)
1495 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1498 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1500 static void blk_mq_rq_timed_out(struct request *req)
1502 req->rq_flags |= RQF_TIMED_OUT;
1503 if (req->q->mq_ops->timeout) {
1504 enum blk_eh_timer_return ret;
1506 ret = req->q->mq_ops->timeout(req);
1507 if (ret == BLK_EH_DONE)
1509 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1515 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
1517 unsigned long deadline;
1519 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1521 if (rq->rq_flags & RQF_TIMED_OUT)
1524 deadline = READ_ONCE(rq->deadline);
1525 if (time_after_eq(jiffies, deadline))
1530 else if (time_after(*next, deadline))
1535 void blk_mq_put_rq_ref(struct request *rq)
1537 if (is_flush_rq(rq))
1539 else if (req_ref_put_and_test(rq))
1540 __blk_mq_free_request(rq);
1543 static bool blk_mq_check_expired(struct request *rq, void *priv)
1545 unsigned long *next = priv;
1548 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1549 * be reallocated underneath the timeout handler's processing, then
1550 * the expire check is reliable. If the request is not expired, then
1551 * it was completed and reallocated as a new request after returning
1552 * from blk_mq_check_expired().
1554 if (blk_mq_req_expired(rq, next))
1555 blk_mq_rq_timed_out(rq);
1559 static void blk_mq_timeout_work(struct work_struct *work)
1561 struct request_queue *q =
1562 container_of(work, struct request_queue, timeout_work);
1563 unsigned long next = 0;
1564 struct blk_mq_hw_ctx *hctx;
1567 /* A deadlock might occur if a request is stuck requiring a
1568 * timeout at the same time a queue freeze is waiting
1569 * completion, since the timeout code would not be able to
1570 * acquire the queue reference here.
1572 * That's why we don't use blk_queue_enter here; instead, we use
1573 * percpu_ref_tryget directly, because we need to be able to
1574 * obtain a reference even in the short window between the queue
1575 * starting to freeze, by dropping the first reference in
1576 * blk_freeze_queue_start, and the moment the last request is
1577 * consumed, marked by the instant q_usage_counter reaches
1580 if (!percpu_ref_tryget(&q->q_usage_counter))
1583 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
1586 mod_timer(&q->timeout, next);
1589 * Request timeouts are handled as a forward rolling timer. If
1590 * we end up here it means that no requests are pending and
1591 * also that no request has been pending for a while. Mark
1592 * each hctx as idle.
1594 queue_for_each_hw_ctx(q, hctx, i) {
1595 /* the hctx may be unmapped, so check it here */
1596 if (blk_mq_hw_queue_mapped(hctx))
1597 blk_mq_tag_idle(hctx);
1603 struct flush_busy_ctx_data {
1604 struct blk_mq_hw_ctx *hctx;
1605 struct list_head *list;
1608 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1610 struct flush_busy_ctx_data *flush_data = data;
1611 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1612 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1613 enum hctx_type type = hctx->type;
1615 spin_lock(&ctx->lock);
1616 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1617 sbitmap_clear_bit(sb, bitnr);
1618 spin_unlock(&ctx->lock);
1623 * Process software queues that have been marked busy, splicing them
1624 * to the for-dispatch
1626 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1628 struct flush_busy_ctx_data data = {
1633 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1635 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1637 struct dispatch_rq_data {
1638 struct blk_mq_hw_ctx *hctx;
1642 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1645 struct dispatch_rq_data *dispatch_data = data;
1646 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1647 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1648 enum hctx_type type = hctx->type;
1650 spin_lock(&ctx->lock);
1651 if (!list_empty(&ctx->rq_lists[type])) {
1652 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1653 list_del_init(&dispatch_data->rq->queuelist);
1654 if (list_empty(&ctx->rq_lists[type]))
1655 sbitmap_clear_bit(sb, bitnr);
1657 spin_unlock(&ctx->lock);
1659 return !dispatch_data->rq;
1662 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1663 struct blk_mq_ctx *start)
1665 unsigned off = start ? start->index_hw[hctx->type] : 0;
1666 struct dispatch_rq_data data = {
1671 __sbitmap_for_each_set(&hctx->ctx_map, off,
1672 dispatch_rq_from_ctx, &data);
1677 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1679 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1680 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1683 blk_mq_tag_busy(rq->mq_hctx);
1685 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1686 bt = &rq->mq_hctx->tags->breserved_tags;
1689 if (!hctx_may_queue(rq->mq_hctx, bt))
1693 tag = __sbitmap_queue_get(bt);
1694 if (tag == BLK_MQ_NO_TAG)
1697 rq->tag = tag + tag_offset;
1701 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1703 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1706 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1707 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1708 rq->rq_flags |= RQF_MQ_INFLIGHT;
1709 __blk_mq_inc_active_requests(hctx);
1711 hctx->tags->rqs[rq->tag] = rq;
1715 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1716 int flags, void *key)
1718 struct blk_mq_hw_ctx *hctx;
1720 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1722 spin_lock(&hctx->dispatch_wait_lock);
1723 if (!list_empty(&wait->entry)) {
1724 struct sbitmap_queue *sbq;
1726 list_del_init(&wait->entry);
1727 sbq = &hctx->tags->bitmap_tags;
1728 atomic_dec(&sbq->ws_active);
1730 spin_unlock(&hctx->dispatch_wait_lock);
1732 blk_mq_run_hw_queue(hctx, true);
1737 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1738 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1739 * restart. For both cases, take care to check the condition again after
1740 * marking us as waiting.
1742 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1745 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1746 struct wait_queue_head *wq;
1747 wait_queue_entry_t *wait;
1750 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1751 blk_mq_sched_mark_restart_hctx(hctx);
1754 * It's possible that a tag was freed in the window between the
1755 * allocation failure and adding the hardware queue to the wait
1758 * Don't clear RESTART here, someone else could have set it.
1759 * At most this will cost an extra queue run.
1761 return blk_mq_get_driver_tag(rq);
1764 wait = &hctx->dispatch_wait;
1765 if (!list_empty_careful(&wait->entry))
1768 wq = &bt_wait_ptr(sbq, hctx)->wait;
1770 spin_lock_irq(&wq->lock);
1771 spin_lock(&hctx->dispatch_wait_lock);
1772 if (!list_empty(&wait->entry)) {
1773 spin_unlock(&hctx->dispatch_wait_lock);
1774 spin_unlock_irq(&wq->lock);
1778 atomic_inc(&sbq->ws_active);
1779 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1780 __add_wait_queue(wq, wait);
1783 * It's possible that a tag was freed in the window between the
1784 * allocation failure and adding the hardware queue to the wait
1787 ret = blk_mq_get_driver_tag(rq);
1789 spin_unlock(&hctx->dispatch_wait_lock);
1790 spin_unlock_irq(&wq->lock);
1795 * We got a tag, remove ourselves from the wait queue to ensure
1796 * someone else gets the wakeup.
1798 list_del_init(&wait->entry);
1799 atomic_dec(&sbq->ws_active);
1800 spin_unlock(&hctx->dispatch_wait_lock);
1801 spin_unlock_irq(&wq->lock);
1806 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1807 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1809 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1810 * - EWMA is one simple way to compute running average value
1811 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1812 * - take 4 as factor for avoiding to get too small(0) result, and this
1813 * factor doesn't matter because EWMA decreases exponentially
1815 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1819 ewma = hctx->dispatch_busy;
1824 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1826 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1827 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1829 hctx->dispatch_busy = ewma;
1832 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1834 static void blk_mq_handle_dev_resource(struct request *rq,
1835 struct list_head *list)
1837 struct request *next =
1838 list_first_entry_or_null(list, struct request, queuelist);
1841 * If an I/O scheduler has been configured and we got a driver tag for
1842 * the next request already, free it.
1845 blk_mq_put_driver_tag(next);
1847 list_add(&rq->queuelist, list);
1848 __blk_mq_requeue_request(rq);
1851 static void blk_mq_handle_zone_resource(struct request *rq,
1852 struct list_head *zone_list)
1855 * If we end up here it is because we cannot dispatch a request to a
1856 * specific zone due to LLD level zone-write locking or other zone
1857 * related resource not being available. In this case, set the request
1858 * aside in zone_list for retrying it later.
1860 list_add(&rq->queuelist, zone_list);
1861 __blk_mq_requeue_request(rq);
1864 enum prep_dispatch {
1866 PREP_DISPATCH_NO_TAG,
1867 PREP_DISPATCH_NO_BUDGET,
1870 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1873 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1874 int budget_token = -1;
1877 budget_token = blk_mq_get_dispatch_budget(rq->q);
1878 if (budget_token < 0) {
1879 blk_mq_put_driver_tag(rq);
1880 return PREP_DISPATCH_NO_BUDGET;
1882 blk_mq_set_rq_budget_token(rq, budget_token);
1885 if (!blk_mq_get_driver_tag(rq)) {
1887 * The initial allocation attempt failed, so we need to
1888 * rerun the hardware queue when a tag is freed. The
1889 * waitqueue takes care of that. If the queue is run
1890 * before we add this entry back on the dispatch list,
1891 * we'll re-run it below.
1893 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1895 * All budgets not got from this function will be put
1896 * together during handling partial dispatch
1899 blk_mq_put_dispatch_budget(rq->q, budget_token);
1900 return PREP_DISPATCH_NO_TAG;
1904 return PREP_DISPATCH_OK;
1907 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1908 static void blk_mq_release_budgets(struct request_queue *q,
1909 struct list_head *list)
1913 list_for_each_entry(rq, list, queuelist) {
1914 int budget_token = blk_mq_get_rq_budget_token(rq);
1916 if (budget_token >= 0)
1917 blk_mq_put_dispatch_budget(q, budget_token);
1922 * Returns true if we did some work AND can potentially do more.
1924 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1925 unsigned int nr_budgets)
1927 enum prep_dispatch prep;
1928 struct request_queue *q = hctx->queue;
1929 struct request *rq, *nxt;
1931 blk_status_t ret = BLK_STS_OK;
1932 LIST_HEAD(zone_list);
1933 bool needs_resource = false;
1935 if (list_empty(list))
1939 * Now process all the entries, sending them to the driver.
1941 errors = queued = 0;
1943 struct blk_mq_queue_data bd;
1945 rq = list_first_entry(list, struct request, queuelist);
1947 WARN_ON_ONCE(hctx != rq->mq_hctx);
1948 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1949 if (prep != PREP_DISPATCH_OK)
1952 list_del_init(&rq->queuelist);
1957 * Flag last if we have no more requests, or if we have more
1958 * but can't assign a driver tag to it.
1960 if (list_empty(list))
1963 nxt = list_first_entry(list, struct request, queuelist);
1964 bd.last = !blk_mq_get_driver_tag(nxt);
1968 * once the request is queued to lld, no need to cover the
1973 ret = q->mq_ops->queue_rq(hctx, &bd);
1978 case BLK_STS_RESOURCE:
1979 needs_resource = true;
1981 case BLK_STS_DEV_RESOURCE:
1982 blk_mq_handle_dev_resource(rq, list);
1984 case BLK_STS_ZONE_RESOURCE:
1986 * Move the request to zone_list and keep going through
1987 * the dispatch list to find more requests the drive can
1990 blk_mq_handle_zone_resource(rq, &zone_list);
1991 needs_resource = true;
1995 blk_mq_end_request(rq, ret);
1997 } while (!list_empty(list));
1999 if (!list_empty(&zone_list))
2000 list_splice_tail_init(&zone_list, list);
2002 /* If we didn't flush the entire list, we could have told the driver
2003 * there was more coming, but that turned out to be a lie.
2005 if ((!list_empty(list) || errors || needs_resource ||
2006 ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued)
2007 q->mq_ops->commit_rqs(hctx);
2009 * Any items that need requeuing? Stuff them into hctx->dispatch,
2010 * that is where we will continue on next queue run.
2012 if (!list_empty(list)) {
2014 /* For non-shared tags, the RESTART check will suffice */
2015 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2016 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
2019 blk_mq_release_budgets(q, list);
2021 spin_lock(&hctx->lock);
2022 list_splice_tail_init(list, &hctx->dispatch);
2023 spin_unlock(&hctx->lock);
2026 * Order adding requests to hctx->dispatch and checking
2027 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2028 * in blk_mq_sched_restart(). Avoid restart code path to
2029 * miss the new added requests to hctx->dispatch, meantime
2030 * SCHED_RESTART is observed here.
2035 * If SCHED_RESTART was set by the caller of this function and
2036 * it is no longer set that means that it was cleared by another
2037 * thread and hence that a queue rerun is needed.
2039 * If 'no_tag' is set, that means that we failed getting
2040 * a driver tag with an I/O scheduler attached. If our dispatch
2041 * waitqueue is no longer active, ensure that we run the queue
2042 * AFTER adding our entries back to the list.
2044 * If no I/O scheduler has been configured it is possible that
2045 * the hardware queue got stopped and restarted before requests
2046 * were pushed back onto the dispatch list. Rerun the queue to
2047 * avoid starvation. Notes:
2048 * - blk_mq_run_hw_queue() checks whether or not a queue has
2049 * been stopped before rerunning a queue.
2050 * - Some but not all block drivers stop a queue before
2051 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2054 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2055 * bit is set, run queue after a delay to avoid IO stalls
2056 * that could otherwise occur if the queue is idle. We'll do
2057 * similar if we couldn't get budget or couldn't lock a zone
2058 * and SCHED_RESTART is set.
2060 needs_restart = blk_mq_sched_needs_restart(hctx);
2061 if (prep == PREP_DISPATCH_NO_BUDGET)
2062 needs_resource = true;
2063 if (!needs_restart ||
2064 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2065 blk_mq_run_hw_queue(hctx, true);
2066 else if (needs_resource)
2067 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2069 blk_mq_update_dispatch_busy(hctx, true);
2072 blk_mq_update_dispatch_busy(hctx, false);
2074 return (queued + errors) != 0;
2078 * __blk_mq_run_hw_queue - Run a hardware queue.
2079 * @hctx: Pointer to the hardware queue to run.
2081 * Send pending requests to the hardware.
2083 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
2086 * We can't run the queue inline with ints disabled. Ensure that
2087 * we catch bad users of this early.
2089 WARN_ON_ONCE(in_interrupt());
2091 blk_mq_run_dispatch_ops(hctx->queue,
2092 blk_mq_sched_dispatch_requests(hctx));
2095 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2097 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2099 if (cpu >= nr_cpu_ids)
2100 cpu = cpumask_first(hctx->cpumask);
2105 * It'd be great if the workqueue API had a way to pass
2106 * in a mask and had some smarts for more clever placement.
2107 * For now we just round-robin here, switching for every
2108 * BLK_MQ_CPU_WORK_BATCH queued items.
2110 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2113 int next_cpu = hctx->next_cpu;
2115 if (hctx->queue->nr_hw_queues == 1)
2116 return WORK_CPU_UNBOUND;
2118 if (--hctx->next_cpu_batch <= 0) {
2120 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2122 if (next_cpu >= nr_cpu_ids)
2123 next_cpu = blk_mq_first_mapped_cpu(hctx);
2124 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2128 * Do unbound schedule if we can't find a online CPU for this hctx,
2129 * and it should only happen in the path of handling CPU DEAD.
2131 if (!cpu_online(next_cpu)) {
2138 * Make sure to re-select CPU next time once after CPUs
2139 * in hctx->cpumask become online again.
2141 hctx->next_cpu = next_cpu;
2142 hctx->next_cpu_batch = 1;
2143 return WORK_CPU_UNBOUND;
2146 hctx->next_cpu = next_cpu;
2151 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2152 * @hctx: Pointer to the hardware queue to run.
2153 * @async: If we want to run the queue asynchronously.
2154 * @msecs: Milliseconds of delay to wait before running the queue.
2156 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2157 * with a delay of @msecs.
2159 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2160 unsigned long msecs)
2162 if (unlikely(blk_mq_hctx_stopped(hctx)))
2165 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2166 if (cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2167 __blk_mq_run_hw_queue(hctx);
2172 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2173 msecs_to_jiffies(msecs));
2177 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2178 * @hctx: Pointer to the hardware queue to run.
2179 * @msecs: Milliseconds of delay to wait before running the queue.
2181 * Run a hardware queue asynchronously with a delay of @msecs.
2183 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2185 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2187 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2190 * blk_mq_run_hw_queue - Start to run a hardware queue.
2191 * @hctx: Pointer to the hardware queue to run.
2192 * @async: If we want to run the queue asynchronously.
2194 * Check if the request queue is not in a quiesced state and if there are
2195 * pending requests to be sent. If this is true, run the queue to send requests
2198 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2203 * When queue is quiesced, we may be switching io scheduler, or
2204 * updating nr_hw_queues, or other things, and we can't run queue
2205 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2207 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2210 __blk_mq_run_dispatch_ops(hctx->queue, false,
2211 need_run = !blk_queue_quiesced(hctx->queue) &&
2212 blk_mq_hctx_has_pending(hctx));
2215 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2217 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2220 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2223 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2225 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2227 * If the IO scheduler does not respect hardware queues when
2228 * dispatching, we just don't bother with multiple HW queues and
2229 * dispatch from hctx for the current CPU since running multiple queues
2230 * just causes lock contention inside the scheduler and pointless cache
2233 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2235 if (!blk_mq_hctx_stopped(hctx))
2241 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2242 * @q: Pointer to the request queue to run.
2243 * @async: If we want to run the queue asynchronously.
2245 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2247 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2251 if (blk_queue_sq_sched(q))
2252 sq_hctx = blk_mq_get_sq_hctx(q);
2253 queue_for_each_hw_ctx(q, hctx, i) {
2254 if (blk_mq_hctx_stopped(hctx))
2257 * Dispatch from this hctx either if there's no hctx preferred
2258 * by IO scheduler or if it has requests that bypass the
2261 if (!sq_hctx || sq_hctx == hctx ||
2262 !list_empty_careful(&hctx->dispatch))
2263 blk_mq_run_hw_queue(hctx, async);
2266 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2269 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2270 * @q: Pointer to the request queue to run.
2271 * @msecs: Milliseconds of delay to wait before running the queues.
2273 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2275 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2279 if (blk_queue_sq_sched(q))
2280 sq_hctx = blk_mq_get_sq_hctx(q);
2281 queue_for_each_hw_ctx(q, hctx, i) {
2282 if (blk_mq_hctx_stopped(hctx))
2285 * If there is already a run_work pending, leave the
2286 * pending delay untouched. Otherwise, a hctx can stall
2287 * if another hctx is re-delaying the other's work
2288 * before the work executes.
2290 if (delayed_work_pending(&hctx->run_work))
2293 * Dispatch from this hctx either if there's no hctx preferred
2294 * by IO scheduler or if it has requests that bypass the
2297 if (!sq_hctx || sq_hctx == hctx ||
2298 !list_empty_careful(&hctx->dispatch))
2299 blk_mq_delay_run_hw_queue(hctx, msecs);
2302 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2305 * This function is often used for pausing .queue_rq() by driver when
2306 * there isn't enough resource or some conditions aren't satisfied, and
2307 * BLK_STS_RESOURCE is usually returned.
2309 * We do not guarantee that dispatch can be drained or blocked
2310 * after blk_mq_stop_hw_queue() returns. Please use
2311 * blk_mq_quiesce_queue() for that requirement.
2313 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2315 cancel_delayed_work(&hctx->run_work);
2317 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2319 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2322 * This function is often used for pausing .queue_rq() by driver when
2323 * there isn't enough resource or some conditions aren't satisfied, and
2324 * BLK_STS_RESOURCE is usually returned.
2326 * We do not guarantee that dispatch can be drained or blocked
2327 * after blk_mq_stop_hw_queues() returns. Please use
2328 * blk_mq_quiesce_queue() for that requirement.
2330 void blk_mq_stop_hw_queues(struct request_queue *q)
2332 struct blk_mq_hw_ctx *hctx;
2335 queue_for_each_hw_ctx(q, hctx, i)
2336 blk_mq_stop_hw_queue(hctx);
2338 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2340 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2342 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2344 blk_mq_run_hw_queue(hctx, false);
2346 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2348 void blk_mq_start_hw_queues(struct request_queue *q)
2350 struct blk_mq_hw_ctx *hctx;
2353 queue_for_each_hw_ctx(q, hctx, i)
2354 blk_mq_start_hw_queue(hctx);
2356 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2358 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2360 if (!blk_mq_hctx_stopped(hctx))
2363 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2364 blk_mq_run_hw_queue(hctx, async);
2366 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2368 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2370 struct blk_mq_hw_ctx *hctx;
2373 queue_for_each_hw_ctx(q, hctx, i)
2374 blk_mq_start_stopped_hw_queue(hctx, async);
2376 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2378 static void blk_mq_run_work_fn(struct work_struct *work)
2380 struct blk_mq_hw_ctx *hctx;
2382 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2385 * If we are stopped, don't run the queue.
2387 if (blk_mq_hctx_stopped(hctx))
2390 __blk_mq_run_hw_queue(hctx);
2393 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2397 struct blk_mq_ctx *ctx = rq->mq_ctx;
2398 enum hctx_type type = hctx->type;
2400 lockdep_assert_held(&ctx->lock);
2402 trace_block_rq_insert(rq);
2405 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2407 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2410 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2413 struct blk_mq_ctx *ctx = rq->mq_ctx;
2415 lockdep_assert_held(&ctx->lock);
2417 __blk_mq_insert_req_list(hctx, rq, at_head);
2418 blk_mq_hctx_mark_pending(hctx, ctx);
2422 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2423 * @rq: Pointer to request to be inserted.
2424 * @at_head: true if the request should be inserted at the head of the list.
2425 * @run_queue: If we should run the hardware queue after inserting the request.
2427 * Should only be used carefully, when the caller knows we want to
2428 * bypass a potential IO scheduler on the target device.
2430 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2433 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2435 spin_lock(&hctx->lock);
2437 list_add(&rq->queuelist, &hctx->dispatch);
2439 list_add_tail(&rq->queuelist, &hctx->dispatch);
2440 spin_unlock(&hctx->lock);
2443 blk_mq_run_hw_queue(hctx, false);
2446 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2447 struct list_head *list)
2451 enum hctx_type type = hctx->type;
2454 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2457 list_for_each_entry(rq, list, queuelist) {
2458 BUG_ON(rq->mq_ctx != ctx);
2459 trace_block_rq_insert(rq);
2462 spin_lock(&ctx->lock);
2463 list_splice_tail_init(list, &ctx->rq_lists[type]);
2464 blk_mq_hctx_mark_pending(hctx, ctx);
2465 spin_unlock(&ctx->lock);
2468 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2471 if (hctx->queue->mq_ops->commit_rqs) {
2472 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2473 hctx->queue->mq_ops->commit_rqs(hctx);
2478 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2479 unsigned int nr_segs)
2483 if (bio->bi_opf & REQ_RAHEAD)
2484 rq->cmd_flags |= REQ_FAILFAST_MASK;
2486 rq->__sector = bio->bi_iter.bi_sector;
2487 blk_rq_bio_prep(rq, bio, nr_segs);
2489 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2490 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2493 blk_account_io_start(rq);
2496 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2497 struct request *rq, bool last)
2499 struct request_queue *q = rq->q;
2500 struct blk_mq_queue_data bd = {
2507 * For OK queue, we are done. For error, caller may kill it.
2508 * Any other error (busy), just add it to our list as we
2509 * previously would have done.
2511 ret = q->mq_ops->queue_rq(hctx, &bd);
2514 blk_mq_update_dispatch_busy(hctx, false);
2516 case BLK_STS_RESOURCE:
2517 case BLK_STS_DEV_RESOURCE:
2518 blk_mq_update_dispatch_busy(hctx, true);
2519 __blk_mq_requeue_request(rq);
2522 blk_mq_update_dispatch_busy(hctx, false);
2529 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2531 bool bypass_insert, bool last)
2533 struct request_queue *q = rq->q;
2534 bool run_queue = true;
2538 * RCU or SRCU read lock is needed before checking quiesced flag.
2540 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2541 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2542 * and avoid driver to try to dispatch again.
2544 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2546 bypass_insert = false;
2550 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2553 budget_token = blk_mq_get_dispatch_budget(q);
2554 if (budget_token < 0)
2557 blk_mq_set_rq_budget_token(rq, budget_token);
2559 if (!blk_mq_get_driver_tag(rq)) {
2560 blk_mq_put_dispatch_budget(q, budget_token);
2564 return __blk_mq_issue_directly(hctx, rq, last);
2567 return BLK_STS_RESOURCE;
2569 blk_mq_sched_insert_request(rq, false, run_queue, false);
2575 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2576 * @hctx: Pointer of the associated hardware queue.
2577 * @rq: Pointer to request to be sent.
2579 * If the device has enough resources to accept a new request now, send the
2580 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2581 * we can try send it another time in the future. Requests inserted at this
2582 * queue have higher priority.
2584 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2588 __blk_mq_try_issue_directly(hctx, rq, false, true);
2590 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2591 blk_mq_request_bypass_insert(rq, false, true);
2592 else if (ret != BLK_STS_OK)
2593 blk_mq_end_request(rq, ret);
2596 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2598 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2601 static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2603 struct blk_mq_hw_ctx *hctx = NULL;
2608 while ((rq = rq_list_pop(&plug->mq_list))) {
2609 bool last = rq_list_empty(plug->mq_list);
2612 if (hctx != rq->mq_hctx) {
2614 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2618 ret = blk_mq_request_issue_directly(rq, last);
2623 case BLK_STS_RESOURCE:
2624 case BLK_STS_DEV_RESOURCE:
2625 blk_mq_request_bypass_insert(rq, false, true);
2626 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2629 blk_mq_end_request(rq, ret);
2636 * If we didn't flush the entire list, we could have told the driver
2637 * there was more coming, but that turned out to be a lie.
2640 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2643 static void __blk_mq_flush_plug_list(struct request_queue *q,
2644 struct blk_plug *plug)
2646 if (blk_queue_quiesced(q))
2648 q->mq_ops->queue_rqs(&plug->mq_list);
2651 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2653 struct blk_mq_hw_ctx *this_hctx = NULL;
2654 struct blk_mq_ctx *this_ctx = NULL;
2655 struct request *requeue_list = NULL;
2656 unsigned int depth = 0;
2660 struct request *rq = rq_list_pop(&plug->mq_list);
2663 this_hctx = rq->mq_hctx;
2664 this_ctx = rq->mq_ctx;
2665 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2666 rq_list_add(&requeue_list, rq);
2669 list_add_tail(&rq->queuelist, &list);
2671 } while (!rq_list_empty(plug->mq_list));
2673 plug->mq_list = requeue_list;
2674 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2675 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched);
2678 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2682 if (rq_list_empty(plug->mq_list))
2686 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2687 struct request_queue *q;
2689 rq = rq_list_peek(&plug->mq_list);
2693 * Peek first request and see if we have a ->queue_rqs() hook.
2694 * If we do, we can dispatch the whole plug list in one go. We
2695 * already know at this point that all requests belong to the
2696 * same queue, caller must ensure that's the case.
2698 * Since we pass off the full list to the driver at this point,
2699 * we do not increment the active request count for the queue.
2700 * Bypass shared tags for now because of that.
2702 if (q->mq_ops->queue_rqs &&
2703 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2704 blk_mq_run_dispatch_ops(q,
2705 __blk_mq_flush_plug_list(q, plug));
2706 if (rq_list_empty(plug->mq_list))
2710 blk_mq_run_dispatch_ops(q,
2711 blk_mq_plug_issue_direct(plug, false));
2712 if (rq_list_empty(plug->mq_list))
2717 blk_mq_dispatch_plug_list(plug, from_schedule);
2718 } while (!rq_list_empty(plug->mq_list));
2721 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2722 struct list_head *list)
2727 while (!list_empty(list)) {
2729 struct request *rq = list_first_entry(list, struct request,
2732 list_del_init(&rq->queuelist);
2733 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2734 if (ret != BLK_STS_OK) {
2736 if (ret == BLK_STS_RESOURCE ||
2737 ret == BLK_STS_DEV_RESOURCE) {
2738 blk_mq_request_bypass_insert(rq, false,
2742 blk_mq_end_request(rq, ret);
2748 * If we didn't flush the entire list, we could have told
2749 * the driver there was more coming, but that turned out to
2752 if ((!list_empty(list) || errors) &&
2753 hctx->queue->mq_ops->commit_rqs && queued)
2754 hctx->queue->mq_ops->commit_rqs(hctx);
2757 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2758 struct bio *bio, unsigned int nr_segs)
2760 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2761 if (blk_attempt_plug_merge(q, bio, nr_segs))
2763 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2769 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2770 struct blk_plug *plug,
2774 struct blk_mq_alloc_data data = {
2777 .cmd_flags = bio->bi_opf,
2781 if (unlikely(bio_queue_enter(bio)))
2784 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2787 rq_qos_throttle(q, bio);
2790 data.nr_tags = plug->nr_ios;
2792 data.cached_rq = &plug->cached_rq;
2795 rq = __blk_mq_alloc_requests(&data);
2798 rq_qos_cleanup(q, bio);
2799 if (bio->bi_opf & REQ_NOWAIT)
2800 bio_wouldblock_error(bio);
2806 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2807 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2813 rq = rq_list_peek(&plug->cached_rq);
2814 if (!rq || rq->q != q)
2817 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2822 if (blk_mq_get_hctx_type((*bio)->bi_opf) != rq->mq_hctx->type)
2824 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2828 * If any qos ->throttle() end up blocking, we will have flushed the
2829 * plug and hence killed the cached_rq list as well. Pop this entry
2830 * before we throttle.
2832 plug->cached_rq = rq_list_next(rq);
2833 rq_qos_throttle(q, *bio);
2835 rq->cmd_flags = (*bio)->bi_opf;
2836 INIT_LIST_HEAD(&rq->queuelist);
2840 static void bio_set_ioprio(struct bio *bio)
2842 /* Nobody set ioprio so far? Initialize it based on task's nice value */
2843 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2844 bio->bi_ioprio = get_current_ioprio();
2845 blkcg_set_ioprio(bio);
2849 * blk_mq_submit_bio - Create and send a request to block device.
2850 * @bio: Bio pointer.
2852 * Builds up a request structure from @q and @bio and send to the device. The
2853 * request may not be queued directly to hardware if:
2854 * * This request can be merged with another one
2855 * * We want to place request at plug queue for possible future merging
2856 * * There is an IO scheduler active at this queue
2858 * It will not queue the request if there is an error with the bio, or at the
2861 void blk_mq_submit_bio(struct bio *bio)
2863 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2864 struct blk_plug *plug = blk_mq_plug(bio);
2865 const int is_sync = op_is_sync(bio->bi_opf);
2867 unsigned int nr_segs = 1;
2870 bio = blk_queue_bounce(bio, q);
2871 if (bio_may_exceed_limits(bio, &q->limits))
2872 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2874 if (!bio_integrity_prep(bio))
2877 bio_set_ioprio(bio);
2879 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2883 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2888 trace_block_getrq(bio);
2890 rq_qos_track(q, rq, bio);
2892 blk_mq_bio_to_request(rq, bio, nr_segs);
2894 ret = blk_crypto_init_request(rq);
2895 if (ret != BLK_STS_OK) {
2896 bio->bi_status = ret;
2898 blk_mq_free_request(rq);
2902 if (op_is_flush(bio->bi_opf)) {
2903 blk_insert_flush(rq);
2908 blk_add_rq_to_plug(plug, rq);
2909 else if ((rq->rq_flags & RQF_ELV) ||
2910 (rq->mq_hctx->dispatch_busy &&
2911 (q->nr_hw_queues == 1 || !is_sync)))
2912 blk_mq_sched_insert_request(rq, false, true, true);
2914 blk_mq_run_dispatch_ops(rq->q,
2915 blk_mq_try_issue_directly(rq->mq_hctx, rq));
2918 #ifdef CONFIG_BLK_MQ_STACKING
2920 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
2921 * @rq: the request being queued
2923 blk_status_t blk_insert_cloned_request(struct request *rq)
2925 struct request_queue *q = rq->q;
2926 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
2929 if (blk_rq_sectors(rq) > max_sectors) {
2931 * SCSI device does not have a good way to return if
2932 * Write Same/Zero is actually supported. If a device rejects
2933 * a non-read/write command (discard, write same,etc.) the
2934 * low-level device driver will set the relevant queue limit to
2935 * 0 to prevent blk-lib from issuing more of the offending
2936 * operations. Commands queued prior to the queue limit being
2937 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
2938 * errors being propagated to upper layers.
2940 if (max_sectors == 0)
2941 return BLK_STS_NOTSUPP;
2943 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
2944 __func__, blk_rq_sectors(rq), max_sectors);
2945 return BLK_STS_IOERR;
2949 * The queue settings related to segment counting may differ from the
2952 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
2953 if (rq->nr_phys_segments > queue_max_segments(q)) {
2954 printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n",
2955 __func__, rq->nr_phys_segments, queue_max_segments(q));
2956 return BLK_STS_IOERR;
2959 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
2960 return BLK_STS_IOERR;
2962 if (blk_crypto_insert_cloned_request(rq))
2963 return BLK_STS_IOERR;
2965 blk_account_io_start(rq);
2968 * Since we have a scheduler attached on the top device,
2969 * bypass a potential scheduler on the bottom device for
2972 blk_mq_run_dispatch_ops(q,
2973 ret = blk_mq_request_issue_directly(rq, true));
2975 blk_account_io_done(rq, ktime_get_ns());
2978 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
2981 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
2982 * @rq: the clone request to be cleaned up
2985 * Free all bios in @rq for a cloned request.
2987 void blk_rq_unprep_clone(struct request *rq)
2991 while ((bio = rq->bio) != NULL) {
2992 rq->bio = bio->bi_next;
2997 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3000 * blk_rq_prep_clone - Helper function to setup clone request
3001 * @rq: the request to be setup
3002 * @rq_src: original request to be cloned
3003 * @bs: bio_set that bios for clone are allocated from
3004 * @gfp_mask: memory allocation mask for bio
3005 * @bio_ctr: setup function to be called for each clone bio.
3006 * Returns %0 for success, non %0 for failure.
3007 * @data: private data to be passed to @bio_ctr
3010 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3011 * Also, pages which the original bios are pointing to are not copied
3012 * and the cloned bios just point same pages.
3013 * So cloned bios must be completed before original bios, which means
3014 * the caller must complete @rq before @rq_src.
3016 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3017 struct bio_set *bs, gfp_t gfp_mask,
3018 int (*bio_ctr)(struct bio *, struct bio *, void *),
3021 struct bio *bio, *bio_src;
3026 __rq_for_each_bio(bio_src, rq_src) {
3027 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3032 if (bio_ctr && bio_ctr(bio, bio_src, data))
3036 rq->biotail->bi_next = bio;
3039 rq->bio = rq->biotail = bio;
3044 /* Copy attributes of the original request to the clone request. */
3045 rq->__sector = blk_rq_pos(rq_src);
3046 rq->__data_len = blk_rq_bytes(rq_src);
3047 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3048 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3049 rq->special_vec = rq_src->special_vec;
3051 rq->nr_phys_segments = rq_src->nr_phys_segments;
3052 rq->ioprio = rq_src->ioprio;
3054 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3062 blk_rq_unprep_clone(rq);
3066 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3067 #endif /* CONFIG_BLK_MQ_STACKING */
3070 * Steal bios from a request and add them to a bio list.
3071 * The request must not have been partially completed before.
3073 void blk_steal_bios(struct bio_list *list, struct request *rq)
3077 list->tail->bi_next = rq->bio;
3079 list->head = rq->bio;
3080 list->tail = rq->biotail;
3088 EXPORT_SYMBOL_GPL(blk_steal_bios);
3090 static size_t order_to_size(unsigned int order)
3092 return (size_t)PAGE_SIZE << order;
3095 /* called before freeing request pool in @tags */
3096 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3097 struct blk_mq_tags *tags)
3100 unsigned long flags;
3102 /* There is no need to clear a driver tags own mapping */
3103 if (drv_tags == tags)
3106 list_for_each_entry(page, &tags->page_list, lru) {
3107 unsigned long start = (unsigned long)page_address(page);
3108 unsigned long end = start + order_to_size(page->private);
3111 for (i = 0; i < drv_tags->nr_tags; i++) {
3112 struct request *rq = drv_tags->rqs[i];
3113 unsigned long rq_addr = (unsigned long)rq;
3115 if (rq_addr >= start && rq_addr < end) {
3116 WARN_ON_ONCE(req_ref_read(rq) != 0);
3117 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3123 * Wait until all pending iteration is done.
3125 * Request reference is cleared and it is guaranteed to be observed
3126 * after the ->lock is released.
3128 spin_lock_irqsave(&drv_tags->lock, flags);
3129 spin_unlock_irqrestore(&drv_tags->lock, flags);
3132 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3133 unsigned int hctx_idx)
3135 struct blk_mq_tags *drv_tags;
3138 if (list_empty(&tags->page_list))
3141 if (blk_mq_is_shared_tags(set->flags))
3142 drv_tags = set->shared_tags;
3144 drv_tags = set->tags[hctx_idx];
3146 if (tags->static_rqs && set->ops->exit_request) {
3149 for (i = 0; i < tags->nr_tags; i++) {
3150 struct request *rq = tags->static_rqs[i];
3154 set->ops->exit_request(set, rq, hctx_idx);
3155 tags->static_rqs[i] = NULL;
3159 blk_mq_clear_rq_mapping(drv_tags, tags);
3161 while (!list_empty(&tags->page_list)) {
3162 page = list_first_entry(&tags->page_list, struct page, lru);
3163 list_del_init(&page->lru);
3165 * Remove kmemleak object previously allocated in
3166 * blk_mq_alloc_rqs().
3168 kmemleak_free(page_address(page));
3169 __free_pages(page, page->private);
3173 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3177 kfree(tags->static_rqs);
3178 tags->static_rqs = NULL;
3180 blk_mq_free_tags(tags);
3183 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3184 unsigned int hctx_idx)
3188 for (i = 0; i < set->nr_maps; i++) {
3189 unsigned int start = set->map[i].queue_offset;
3190 unsigned int end = start + set->map[i].nr_queues;
3192 if (hctx_idx >= start && hctx_idx < end)
3196 if (i >= set->nr_maps)
3197 i = HCTX_TYPE_DEFAULT;
3202 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3203 unsigned int hctx_idx)
3205 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3207 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3210 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3211 unsigned int hctx_idx,
3212 unsigned int nr_tags,
3213 unsigned int reserved_tags)
3215 int node = blk_mq_get_hctx_node(set, hctx_idx);
3216 struct blk_mq_tags *tags;
3218 if (node == NUMA_NO_NODE)
3219 node = set->numa_node;
3221 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3222 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3226 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3227 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3230 blk_mq_free_tags(tags);
3234 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3235 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3237 if (!tags->static_rqs) {
3239 blk_mq_free_tags(tags);
3246 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3247 unsigned int hctx_idx, int node)
3251 if (set->ops->init_request) {
3252 ret = set->ops->init_request(set, rq, hctx_idx, node);
3257 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3261 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3262 struct blk_mq_tags *tags,
3263 unsigned int hctx_idx, unsigned int depth)
3265 unsigned int i, j, entries_per_page, max_order = 4;
3266 int node = blk_mq_get_hctx_node(set, hctx_idx);
3267 size_t rq_size, left;
3269 if (node == NUMA_NO_NODE)
3270 node = set->numa_node;
3272 INIT_LIST_HEAD(&tags->page_list);
3275 * rq_size is the size of the request plus driver payload, rounded
3276 * to the cacheline size
3278 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3280 left = rq_size * depth;
3282 for (i = 0; i < depth; ) {
3283 int this_order = max_order;
3288 while (this_order && left < order_to_size(this_order - 1))
3292 page = alloc_pages_node(node,
3293 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3299 if (order_to_size(this_order) < rq_size)
3306 page->private = this_order;
3307 list_add_tail(&page->lru, &tags->page_list);
3309 p = page_address(page);
3311 * Allow kmemleak to scan these pages as they contain pointers
3312 * to additional allocations like via ops->init_request().
3314 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3315 entries_per_page = order_to_size(this_order) / rq_size;
3316 to_do = min(entries_per_page, depth - i);
3317 left -= to_do * rq_size;
3318 for (j = 0; j < to_do; j++) {
3319 struct request *rq = p;
3321 tags->static_rqs[i] = rq;
3322 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3323 tags->static_rqs[i] = NULL;
3334 blk_mq_free_rqs(set, tags, hctx_idx);
3338 struct rq_iter_data {
3339 struct blk_mq_hw_ctx *hctx;
3343 static bool blk_mq_has_request(struct request *rq, void *data)
3345 struct rq_iter_data *iter_data = data;
3347 if (rq->mq_hctx != iter_data->hctx)
3349 iter_data->has_rq = true;
3353 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3355 struct blk_mq_tags *tags = hctx->sched_tags ?
3356 hctx->sched_tags : hctx->tags;
3357 struct rq_iter_data data = {
3361 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3365 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3366 struct blk_mq_hw_ctx *hctx)
3368 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3370 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3375 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3377 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3378 struct blk_mq_hw_ctx, cpuhp_online);
3380 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3381 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3385 * Prevent new request from being allocated on the current hctx.
3387 * The smp_mb__after_atomic() Pairs with the implied barrier in
3388 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3389 * seen once we return from the tag allocator.
3391 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3392 smp_mb__after_atomic();
3395 * Try to grab a reference to the queue and wait for any outstanding
3396 * requests. If we could not grab a reference the queue has been
3397 * frozen and there are no requests.
3399 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3400 while (blk_mq_hctx_has_requests(hctx))
3402 percpu_ref_put(&hctx->queue->q_usage_counter);
3408 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3410 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3411 struct blk_mq_hw_ctx, cpuhp_online);
3413 if (cpumask_test_cpu(cpu, hctx->cpumask))
3414 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3419 * 'cpu' is going away. splice any existing rq_list entries from this
3420 * software queue to the hw queue dispatch list, and ensure that it
3423 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3425 struct blk_mq_hw_ctx *hctx;
3426 struct blk_mq_ctx *ctx;
3428 enum hctx_type type;
3430 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3431 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3434 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3437 spin_lock(&ctx->lock);
3438 if (!list_empty(&ctx->rq_lists[type])) {
3439 list_splice_init(&ctx->rq_lists[type], &tmp);
3440 blk_mq_hctx_clear_pending(hctx, ctx);
3442 spin_unlock(&ctx->lock);
3444 if (list_empty(&tmp))
3447 spin_lock(&hctx->lock);
3448 list_splice_tail_init(&tmp, &hctx->dispatch);
3449 spin_unlock(&hctx->lock);
3451 blk_mq_run_hw_queue(hctx, true);
3455 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3457 if (!(hctx->flags & BLK_MQ_F_STACKING))
3458 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3459 &hctx->cpuhp_online);
3460 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3465 * Before freeing hw queue, clearing the flush request reference in
3466 * tags->rqs[] for avoiding potential UAF.
3468 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3469 unsigned int queue_depth, struct request *flush_rq)
3472 unsigned long flags;
3474 /* The hw queue may not be mapped yet */
3478 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3480 for (i = 0; i < queue_depth; i++)
3481 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3484 * Wait until all pending iteration is done.
3486 * Request reference is cleared and it is guaranteed to be observed
3487 * after the ->lock is released.
3489 spin_lock_irqsave(&tags->lock, flags);
3490 spin_unlock_irqrestore(&tags->lock, flags);
3493 /* hctx->ctxs will be freed in queue's release handler */
3494 static void blk_mq_exit_hctx(struct request_queue *q,
3495 struct blk_mq_tag_set *set,
3496 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3498 struct request *flush_rq = hctx->fq->flush_rq;
3500 if (blk_mq_hw_queue_mapped(hctx))
3501 blk_mq_tag_idle(hctx);
3503 if (blk_queue_init_done(q))
3504 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3505 set->queue_depth, flush_rq);
3506 if (set->ops->exit_request)
3507 set->ops->exit_request(set, flush_rq, hctx_idx);
3509 if (set->ops->exit_hctx)
3510 set->ops->exit_hctx(hctx, hctx_idx);
3512 blk_mq_remove_cpuhp(hctx);
3514 xa_erase(&q->hctx_table, hctx_idx);
3516 spin_lock(&q->unused_hctx_lock);
3517 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3518 spin_unlock(&q->unused_hctx_lock);
3521 static void blk_mq_exit_hw_queues(struct request_queue *q,
3522 struct blk_mq_tag_set *set, int nr_queue)
3524 struct blk_mq_hw_ctx *hctx;
3527 queue_for_each_hw_ctx(q, hctx, i) {
3530 blk_mq_exit_hctx(q, set, hctx, i);
3534 static int blk_mq_init_hctx(struct request_queue *q,
3535 struct blk_mq_tag_set *set,
3536 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3538 hctx->queue_num = hctx_idx;
3540 if (!(hctx->flags & BLK_MQ_F_STACKING))
3541 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3542 &hctx->cpuhp_online);
3543 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3545 hctx->tags = set->tags[hctx_idx];
3547 if (set->ops->init_hctx &&
3548 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3549 goto unregister_cpu_notifier;
3551 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3555 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3561 if (set->ops->exit_request)
3562 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3564 if (set->ops->exit_hctx)
3565 set->ops->exit_hctx(hctx, hctx_idx);
3566 unregister_cpu_notifier:
3567 blk_mq_remove_cpuhp(hctx);
3571 static struct blk_mq_hw_ctx *
3572 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3575 struct blk_mq_hw_ctx *hctx;
3576 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3578 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3580 goto fail_alloc_hctx;
3582 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3585 atomic_set(&hctx->nr_active, 0);
3586 if (node == NUMA_NO_NODE)
3587 node = set->numa_node;
3588 hctx->numa_node = node;
3590 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3591 spin_lock_init(&hctx->lock);
3592 INIT_LIST_HEAD(&hctx->dispatch);
3594 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3596 INIT_LIST_HEAD(&hctx->hctx_list);
3599 * Allocate space for all possible cpus to avoid allocation at
3602 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3607 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3608 gfp, node, false, false))
3612 spin_lock_init(&hctx->dispatch_wait_lock);
3613 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3614 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3616 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3620 blk_mq_hctx_kobj_init(hctx);
3625 sbitmap_free(&hctx->ctx_map);
3629 free_cpumask_var(hctx->cpumask);
3636 static void blk_mq_init_cpu_queues(struct request_queue *q,
3637 unsigned int nr_hw_queues)
3639 struct blk_mq_tag_set *set = q->tag_set;
3642 for_each_possible_cpu(i) {
3643 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3644 struct blk_mq_hw_ctx *hctx;
3648 spin_lock_init(&__ctx->lock);
3649 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3650 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3655 * Set local node, IFF we have more than one hw queue. If
3656 * not, we remain on the home node of the device
3658 for (j = 0; j < set->nr_maps; j++) {
3659 hctx = blk_mq_map_queue_type(q, j, i);
3660 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3661 hctx->numa_node = cpu_to_node(i);
3666 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3667 unsigned int hctx_idx,
3670 struct blk_mq_tags *tags;
3673 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3677 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3679 blk_mq_free_rq_map(tags);
3686 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3689 if (blk_mq_is_shared_tags(set->flags)) {
3690 set->tags[hctx_idx] = set->shared_tags;
3695 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3698 return set->tags[hctx_idx];
3701 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3702 struct blk_mq_tags *tags,
3703 unsigned int hctx_idx)
3706 blk_mq_free_rqs(set, tags, hctx_idx);
3707 blk_mq_free_rq_map(tags);
3711 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3712 unsigned int hctx_idx)
3714 if (!blk_mq_is_shared_tags(set->flags))
3715 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3717 set->tags[hctx_idx] = NULL;
3720 static void blk_mq_map_swqueue(struct request_queue *q)
3722 unsigned int j, hctx_idx;
3724 struct blk_mq_hw_ctx *hctx;
3725 struct blk_mq_ctx *ctx;
3726 struct blk_mq_tag_set *set = q->tag_set;
3728 queue_for_each_hw_ctx(q, hctx, i) {
3729 cpumask_clear(hctx->cpumask);
3731 hctx->dispatch_from = NULL;
3735 * Map software to hardware queues.
3737 * If the cpu isn't present, the cpu is mapped to first hctx.
3739 for_each_possible_cpu(i) {
3741 ctx = per_cpu_ptr(q->queue_ctx, i);
3742 for (j = 0; j < set->nr_maps; j++) {
3743 if (!set->map[j].nr_queues) {
3744 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3745 HCTX_TYPE_DEFAULT, i);
3748 hctx_idx = set->map[j].mq_map[i];
3749 /* unmapped hw queue can be remapped after CPU topo changed */
3750 if (!set->tags[hctx_idx] &&
3751 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3753 * If tags initialization fail for some hctx,
3754 * that hctx won't be brought online. In this
3755 * case, remap the current ctx to hctx[0] which
3756 * is guaranteed to always have tags allocated
3758 set->map[j].mq_map[i] = 0;
3761 hctx = blk_mq_map_queue_type(q, j, i);
3762 ctx->hctxs[j] = hctx;
3764 * If the CPU is already set in the mask, then we've
3765 * mapped this one already. This can happen if
3766 * devices share queues across queue maps.
3768 if (cpumask_test_cpu(i, hctx->cpumask))
3771 cpumask_set_cpu(i, hctx->cpumask);
3773 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3774 hctx->ctxs[hctx->nr_ctx++] = ctx;
3777 * If the nr_ctx type overflows, we have exceeded the
3778 * amount of sw queues we can support.
3780 BUG_ON(!hctx->nr_ctx);
3783 for (; j < HCTX_MAX_TYPES; j++)
3784 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3785 HCTX_TYPE_DEFAULT, i);
3788 queue_for_each_hw_ctx(q, hctx, i) {
3790 * If no software queues are mapped to this hardware queue,
3791 * disable it and free the request entries.
3793 if (!hctx->nr_ctx) {
3794 /* Never unmap queue 0. We need it as a
3795 * fallback in case of a new remap fails
3799 __blk_mq_free_map_and_rqs(set, i);
3805 hctx->tags = set->tags[i];
3806 WARN_ON(!hctx->tags);
3809 * Set the map size to the number of mapped software queues.
3810 * This is more accurate and more efficient than looping
3811 * over all possibly mapped software queues.
3813 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3816 * Initialize batch roundrobin counts
3818 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3819 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3824 * Caller needs to ensure that we're either frozen/quiesced, or that
3825 * the queue isn't live yet.
3827 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3829 struct blk_mq_hw_ctx *hctx;
3832 queue_for_each_hw_ctx(q, hctx, i) {
3834 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3836 blk_mq_tag_idle(hctx);
3837 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3842 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3845 struct request_queue *q;
3847 lockdep_assert_held(&set->tag_list_lock);
3849 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3850 blk_mq_freeze_queue(q);
3851 queue_set_hctx_shared(q, shared);
3852 blk_mq_unfreeze_queue(q);
3856 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3858 struct blk_mq_tag_set *set = q->tag_set;
3860 mutex_lock(&set->tag_list_lock);
3861 list_del(&q->tag_set_list);
3862 if (list_is_singular(&set->tag_list)) {
3863 /* just transitioned to unshared */
3864 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3865 /* update existing queue */
3866 blk_mq_update_tag_set_shared(set, false);
3868 mutex_unlock(&set->tag_list_lock);
3869 INIT_LIST_HEAD(&q->tag_set_list);
3872 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3873 struct request_queue *q)
3875 mutex_lock(&set->tag_list_lock);
3878 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3880 if (!list_empty(&set->tag_list) &&
3881 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3882 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3883 /* update existing queue */
3884 blk_mq_update_tag_set_shared(set, true);
3886 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3887 queue_set_hctx_shared(q, true);
3888 list_add_tail(&q->tag_set_list, &set->tag_list);
3890 mutex_unlock(&set->tag_list_lock);
3893 /* All allocations will be freed in release handler of q->mq_kobj */
3894 static int blk_mq_alloc_ctxs(struct request_queue *q)
3896 struct blk_mq_ctxs *ctxs;
3899 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3903 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3904 if (!ctxs->queue_ctx)
3907 for_each_possible_cpu(cpu) {
3908 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3912 q->mq_kobj = &ctxs->kobj;
3913 q->queue_ctx = ctxs->queue_ctx;
3922 * It is the actual release handler for mq, but we do it from
3923 * request queue's release handler for avoiding use-after-free
3924 * and headache because q->mq_kobj shouldn't have been introduced,
3925 * but we can't group ctx/kctx kobj without it.
3927 void blk_mq_release(struct request_queue *q)
3929 struct blk_mq_hw_ctx *hctx, *next;
3932 queue_for_each_hw_ctx(q, hctx, i)
3933 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3935 /* all hctx are in .unused_hctx_list now */
3936 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3937 list_del_init(&hctx->hctx_list);
3938 kobject_put(&hctx->kobj);
3941 xa_destroy(&q->hctx_table);
3944 * release .mq_kobj and sw queue's kobject now because
3945 * both share lifetime with request queue.
3947 blk_mq_sysfs_deinit(q);
3950 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3953 struct request_queue *q;
3956 q = blk_alloc_queue(set->numa_node, set->flags & BLK_MQ_F_BLOCKING);
3958 return ERR_PTR(-ENOMEM);
3959 q->queuedata = queuedata;
3960 ret = blk_mq_init_allocated_queue(set, q);
3963 return ERR_PTR(ret);
3968 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3970 return blk_mq_init_queue_data(set, NULL);
3972 EXPORT_SYMBOL(blk_mq_init_queue);
3975 * blk_mq_destroy_queue - shutdown a request queue
3976 * @q: request queue to shutdown
3978 * This shuts down a request queue allocated by blk_mq_init_queue() and drops
3979 * the initial reference. All future requests will failed with -ENODEV.
3981 * Context: can sleep
3983 void blk_mq_destroy_queue(struct request_queue *q)
3985 WARN_ON_ONCE(!queue_is_mq(q));
3986 WARN_ON_ONCE(blk_queue_registered(q));
3990 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
3991 blk_queue_start_drain(q);
3992 blk_freeze_queue(q);
3995 blk_mq_cancel_work_sync(q);
3996 blk_mq_exit_queue(q);
3998 /* @q is and will stay empty, shutdown and put */
4001 EXPORT_SYMBOL(blk_mq_destroy_queue);
4003 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4004 struct lock_class_key *lkclass)
4006 struct request_queue *q;
4007 struct gendisk *disk;
4009 q = blk_mq_init_queue_data(set, queuedata);
4013 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4015 blk_mq_destroy_queue(q);
4016 return ERR_PTR(-ENOMEM);
4018 set_bit(GD_OWNS_QUEUE, &disk->state);
4021 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4023 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4024 struct lock_class_key *lkclass)
4026 if (!blk_get_queue(q))
4028 return __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4030 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4032 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4033 struct blk_mq_tag_set *set, struct request_queue *q,
4034 int hctx_idx, int node)
4036 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4038 /* reuse dead hctx first */
4039 spin_lock(&q->unused_hctx_lock);
4040 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4041 if (tmp->numa_node == node) {
4047 list_del_init(&hctx->hctx_list);
4048 spin_unlock(&q->unused_hctx_lock);
4051 hctx = blk_mq_alloc_hctx(q, set, node);
4055 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4061 kobject_put(&hctx->kobj);
4066 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4067 struct request_queue *q)
4069 struct blk_mq_hw_ctx *hctx;
4072 /* protect against switching io scheduler */
4073 mutex_lock(&q->sysfs_lock);
4074 for (i = 0; i < set->nr_hw_queues; i++) {
4076 int node = blk_mq_get_hctx_node(set, i);
4077 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4080 old_node = old_hctx->numa_node;
4081 blk_mq_exit_hctx(q, set, old_hctx, i);
4084 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4087 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4089 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4090 WARN_ON_ONCE(!hctx);
4094 * Increasing nr_hw_queues fails. Free the newly allocated
4095 * hctxs and keep the previous q->nr_hw_queues.
4097 if (i != set->nr_hw_queues) {
4098 j = q->nr_hw_queues;
4101 q->nr_hw_queues = set->nr_hw_queues;
4104 xa_for_each_start(&q->hctx_table, j, hctx, j)
4105 blk_mq_exit_hctx(q, set, hctx, j);
4106 mutex_unlock(&q->sysfs_lock);
4109 static void blk_mq_update_poll_flag(struct request_queue *q)
4111 struct blk_mq_tag_set *set = q->tag_set;
4113 if (set->nr_maps > HCTX_TYPE_POLL &&
4114 set->map[HCTX_TYPE_POLL].nr_queues)
4115 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4117 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4120 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4121 struct request_queue *q)
4123 WARN_ON_ONCE(blk_queue_has_srcu(q) !=
4124 !!(set->flags & BLK_MQ_F_BLOCKING));
4126 /* mark the queue as mq asap */
4127 q->mq_ops = set->ops;
4129 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
4130 blk_mq_poll_stats_bkt,
4131 BLK_MQ_POLL_STATS_BKTS, q);
4135 if (blk_mq_alloc_ctxs(q))
4138 /* init q->mq_kobj and sw queues' kobjects */
4139 blk_mq_sysfs_init(q);
4141 INIT_LIST_HEAD(&q->unused_hctx_list);
4142 spin_lock_init(&q->unused_hctx_lock);
4144 xa_init(&q->hctx_table);
4146 blk_mq_realloc_hw_ctxs(set, q);
4147 if (!q->nr_hw_queues)
4150 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4151 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4155 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4156 blk_mq_update_poll_flag(q);
4158 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4159 INIT_LIST_HEAD(&q->requeue_list);
4160 spin_lock_init(&q->requeue_lock);
4162 q->nr_requests = set->queue_depth;
4165 * Default to classic polling
4167 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
4169 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4170 blk_mq_add_queue_tag_set(set, q);
4171 blk_mq_map_swqueue(q);
4175 xa_destroy(&q->hctx_table);
4176 q->nr_hw_queues = 0;
4177 blk_mq_sysfs_deinit(q);
4179 blk_stat_free_callback(q->poll_cb);
4185 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4187 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4188 void blk_mq_exit_queue(struct request_queue *q)
4190 struct blk_mq_tag_set *set = q->tag_set;
4192 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4193 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4194 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4195 blk_mq_del_queue_tag_set(q);
4198 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4202 if (blk_mq_is_shared_tags(set->flags)) {
4203 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4206 if (!set->shared_tags)
4210 for (i = 0; i < set->nr_hw_queues; i++) {
4211 if (!__blk_mq_alloc_map_and_rqs(set, i))
4220 __blk_mq_free_map_and_rqs(set, i);
4222 if (blk_mq_is_shared_tags(set->flags)) {
4223 blk_mq_free_map_and_rqs(set, set->shared_tags,
4224 BLK_MQ_NO_HCTX_IDX);
4231 * Allocate the request maps associated with this tag_set. Note that this
4232 * may reduce the depth asked for, if memory is tight. set->queue_depth
4233 * will be updated to reflect the allocated depth.
4235 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4240 depth = set->queue_depth;
4242 err = __blk_mq_alloc_rq_maps(set);
4246 set->queue_depth >>= 1;
4247 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4251 } while (set->queue_depth);
4253 if (!set->queue_depth || err) {
4254 pr_err("blk-mq: failed to allocate request map\n");
4258 if (depth != set->queue_depth)
4259 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4260 depth, set->queue_depth);
4265 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4268 * blk_mq_map_queues() and multiple .map_queues() implementations
4269 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4270 * number of hardware queues.
4272 if (set->nr_maps == 1)
4273 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4275 if (set->ops->map_queues && !is_kdump_kernel()) {
4279 * transport .map_queues is usually done in the following
4282 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4283 * mask = get_cpu_mask(queue)
4284 * for_each_cpu(cpu, mask)
4285 * set->map[x].mq_map[cpu] = queue;
4288 * When we need to remap, the table has to be cleared for
4289 * killing stale mapping since one CPU may not be mapped
4292 for (i = 0; i < set->nr_maps; i++)
4293 blk_mq_clear_mq_map(&set->map[i]);
4295 set->ops->map_queues(set);
4297 BUG_ON(set->nr_maps > 1);
4298 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4302 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4303 int cur_nr_hw_queues, int new_nr_hw_queues)
4305 struct blk_mq_tags **new_tags;
4307 if (cur_nr_hw_queues >= new_nr_hw_queues)
4310 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4311 GFP_KERNEL, set->numa_node);
4316 memcpy(new_tags, set->tags, cur_nr_hw_queues *
4317 sizeof(*set->tags));
4319 set->tags = new_tags;
4320 set->nr_hw_queues = new_nr_hw_queues;
4325 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
4326 int new_nr_hw_queues)
4328 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
4332 * Alloc a tag set to be associated with one or more request queues.
4333 * May fail with EINVAL for various error conditions. May adjust the
4334 * requested depth down, if it's too large. In that case, the set
4335 * value will be stored in set->queue_depth.
4337 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4341 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4343 if (!set->nr_hw_queues)
4345 if (!set->queue_depth)
4347 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4350 if (!set->ops->queue_rq)
4353 if (!set->ops->get_budget ^ !set->ops->put_budget)
4356 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4357 pr_info("blk-mq: reduced tag depth to %u\n",
4359 set->queue_depth = BLK_MQ_MAX_DEPTH;
4364 else if (set->nr_maps > HCTX_MAX_TYPES)
4368 * If a crashdump is active, then we are potentially in a very
4369 * memory constrained environment. Limit us to 1 queue and
4370 * 64 tags to prevent using too much memory.
4372 if (is_kdump_kernel()) {
4373 set->nr_hw_queues = 1;
4375 set->queue_depth = min(64U, set->queue_depth);
4378 * There is no use for more h/w queues than cpus if we just have
4381 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4382 set->nr_hw_queues = nr_cpu_ids;
4384 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
4388 for (i = 0; i < set->nr_maps; i++) {
4389 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4390 sizeof(set->map[i].mq_map[0]),
4391 GFP_KERNEL, set->numa_node);
4392 if (!set->map[i].mq_map)
4393 goto out_free_mq_map;
4394 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4397 blk_mq_update_queue_map(set);
4399 ret = blk_mq_alloc_set_map_and_rqs(set);
4401 goto out_free_mq_map;
4403 mutex_init(&set->tag_list_lock);
4404 INIT_LIST_HEAD(&set->tag_list);
4409 for (i = 0; i < set->nr_maps; i++) {
4410 kfree(set->map[i].mq_map);
4411 set->map[i].mq_map = NULL;
4417 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4419 /* allocate and initialize a tagset for a simple single-queue device */
4420 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4421 const struct blk_mq_ops *ops, unsigned int queue_depth,
4422 unsigned int set_flags)
4424 memset(set, 0, sizeof(*set));
4426 set->nr_hw_queues = 1;
4428 set->queue_depth = queue_depth;
4429 set->numa_node = NUMA_NO_NODE;
4430 set->flags = set_flags;
4431 return blk_mq_alloc_tag_set(set);
4433 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4435 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4439 for (i = 0; i < set->nr_hw_queues; i++)
4440 __blk_mq_free_map_and_rqs(set, i);
4442 if (blk_mq_is_shared_tags(set->flags)) {
4443 blk_mq_free_map_and_rqs(set, set->shared_tags,
4444 BLK_MQ_NO_HCTX_IDX);
4447 for (j = 0; j < set->nr_maps; j++) {
4448 kfree(set->map[j].mq_map);
4449 set->map[j].mq_map = NULL;
4455 EXPORT_SYMBOL(blk_mq_free_tag_set);
4457 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4459 struct blk_mq_tag_set *set = q->tag_set;
4460 struct blk_mq_hw_ctx *hctx;
4467 if (q->nr_requests == nr)
4470 blk_mq_freeze_queue(q);
4471 blk_mq_quiesce_queue(q);
4474 queue_for_each_hw_ctx(q, hctx, i) {
4478 * If we're using an MQ scheduler, just update the scheduler
4479 * queue depth. This is similar to what the old code would do.
4481 if (hctx->sched_tags) {
4482 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4485 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4490 if (q->elevator && q->elevator->type->ops.depth_updated)
4491 q->elevator->type->ops.depth_updated(hctx);
4494 q->nr_requests = nr;
4495 if (blk_mq_is_shared_tags(set->flags)) {
4497 blk_mq_tag_update_sched_shared_tags(q);
4499 blk_mq_tag_resize_shared_tags(set, nr);
4503 blk_mq_unquiesce_queue(q);
4504 blk_mq_unfreeze_queue(q);
4510 * request_queue and elevator_type pair.
4511 * It is just used by __blk_mq_update_nr_hw_queues to cache
4512 * the elevator_type associated with a request_queue.
4514 struct blk_mq_qe_pair {
4515 struct list_head node;
4516 struct request_queue *q;
4517 struct elevator_type *type;
4521 * Cache the elevator_type in qe pair list and switch the
4522 * io scheduler to 'none'
4524 static bool blk_mq_elv_switch_none(struct list_head *head,
4525 struct request_queue *q)
4527 struct blk_mq_qe_pair *qe;
4532 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4536 /* q->elevator needs protection from ->sysfs_lock */
4537 mutex_lock(&q->sysfs_lock);
4539 INIT_LIST_HEAD(&qe->node);
4541 qe->type = q->elevator->type;
4542 list_add(&qe->node, head);
4545 * After elevator_switch, the previous elevator_queue will be
4546 * released by elevator_release. The reference of the io scheduler
4547 * module get by elevator_get will also be put. So we need to get
4548 * a reference of the io scheduler module here to prevent it to be
4551 __module_get(qe->type->elevator_owner);
4552 elevator_switch(q, NULL);
4553 mutex_unlock(&q->sysfs_lock);
4558 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4559 struct request_queue *q)
4561 struct blk_mq_qe_pair *qe;
4563 list_for_each_entry(qe, head, node)
4570 static void blk_mq_elv_switch_back(struct list_head *head,
4571 struct request_queue *q)
4573 struct blk_mq_qe_pair *qe;
4574 struct elevator_type *t;
4576 qe = blk_lookup_qe_pair(head, q);
4580 list_del(&qe->node);
4583 mutex_lock(&q->sysfs_lock);
4584 elevator_switch(q, t);
4585 mutex_unlock(&q->sysfs_lock);
4588 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4591 struct request_queue *q;
4593 int prev_nr_hw_queues;
4595 lockdep_assert_held(&set->tag_list_lock);
4597 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4598 nr_hw_queues = nr_cpu_ids;
4599 if (nr_hw_queues < 1)
4601 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4604 list_for_each_entry(q, &set->tag_list, tag_set_list)
4605 blk_mq_freeze_queue(q);
4607 * Switch IO scheduler to 'none', cleaning up the data associated
4608 * with the previous scheduler. We will switch back once we are done
4609 * updating the new sw to hw queue mappings.
4611 list_for_each_entry(q, &set->tag_list, tag_set_list)
4612 if (!blk_mq_elv_switch_none(&head, q))
4615 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4616 blk_mq_debugfs_unregister_hctxs(q);
4617 blk_mq_sysfs_unregister_hctxs(q);
4620 prev_nr_hw_queues = set->nr_hw_queues;
4621 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
4625 set->nr_hw_queues = nr_hw_queues;
4627 blk_mq_update_queue_map(set);
4628 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4629 blk_mq_realloc_hw_ctxs(set, q);
4630 blk_mq_update_poll_flag(q);
4631 if (q->nr_hw_queues != set->nr_hw_queues) {
4632 int i = prev_nr_hw_queues;
4634 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4635 nr_hw_queues, prev_nr_hw_queues);
4636 for (; i < set->nr_hw_queues; i++)
4637 __blk_mq_free_map_and_rqs(set, i);
4639 set->nr_hw_queues = prev_nr_hw_queues;
4640 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4643 blk_mq_map_swqueue(q);
4647 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4648 blk_mq_sysfs_register_hctxs(q);
4649 blk_mq_debugfs_register_hctxs(q);
4653 list_for_each_entry(q, &set->tag_list, tag_set_list)
4654 blk_mq_elv_switch_back(&head, q);
4656 list_for_each_entry(q, &set->tag_list, tag_set_list)
4657 blk_mq_unfreeze_queue(q);
4660 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4662 mutex_lock(&set->tag_list_lock);
4663 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4664 mutex_unlock(&set->tag_list_lock);
4666 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4668 /* Enable polling stats and return whether they were already enabled. */
4669 static bool blk_poll_stats_enable(struct request_queue *q)
4674 return blk_stats_alloc_enable(q);
4677 static void blk_mq_poll_stats_start(struct request_queue *q)
4680 * We don't arm the callback if polling stats are not enabled or the
4681 * callback is already active.
4683 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
4686 blk_stat_activate_msecs(q->poll_cb, 100);
4689 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4691 struct request_queue *q = cb->data;
4694 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4695 if (cb->stat[bucket].nr_samples)
4696 q->poll_stat[bucket] = cb->stat[bucket];
4700 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4703 unsigned long ret = 0;
4707 * If stats collection isn't on, don't sleep but turn it on for
4710 if (!blk_poll_stats_enable(q))
4714 * As an optimistic guess, use half of the mean service time
4715 * for this type of request. We can (and should) make this smarter.
4716 * For instance, if the completion latencies are tight, we can
4717 * get closer than just half the mean. This is especially
4718 * important on devices where the completion latencies are longer
4719 * than ~10 usec. We do use the stats for the relevant IO size
4720 * if available which does lead to better estimates.
4722 bucket = blk_mq_poll_stats_bkt(rq);
4726 if (q->poll_stat[bucket].nr_samples)
4727 ret = (q->poll_stat[bucket].mean + 1) / 2;
4732 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4734 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4735 struct request *rq = blk_qc_to_rq(hctx, qc);
4736 struct hrtimer_sleeper hs;
4737 enum hrtimer_mode mode;
4742 * If a request has completed on queue that uses an I/O scheduler, we
4743 * won't get back a request from blk_qc_to_rq.
4745 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4749 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4751 * 0: use half of prev avg
4752 * >0: use this specific value
4754 if (q->poll_nsec > 0)
4755 nsecs = q->poll_nsec;
4757 nsecs = blk_mq_poll_nsecs(q, rq);
4762 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4765 * This will be replaced with the stats tracking code, using
4766 * 'avg_completion_time / 2' as the pre-sleep target.
4770 mode = HRTIMER_MODE_REL;
4771 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4772 hrtimer_set_expires(&hs.timer, kt);
4775 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4777 set_current_state(TASK_UNINTERRUPTIBLE);
4778 hrtimer_sleeper_start_expires(&hs, mode);
4781 hrtimer_cancel(&hs.timer);
4782 mode = HRTIMER_MODE_ABS;
4783 } while (hs.task && !signal_pending(current));
4785 __set_current_state(TASK_RUNNING);
4786 destroy_hrtimer_on_stack(&hs.timer);
4789 * If we sleep, have the caller restart the poll loop to reset the
4790 * state. Like for the other success return cases, the caller is
4791 * responsible for checking if the IO completed. If the IO isn't
4792 * complete, we'll get called again and will go straight to the busy
4798 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4799 struct io_comp_batch *iob, unsigned int flags)
4801 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4802 long state = get_current_state();
4806 ret = q->mq_ops->poll(hctx, iob);
4808 __set_current_state(TASK_RUNNING);
4812 if (signal_pending_state(state, current))
4813 __set_current_state(TASK_RUNNING);
4814 if (task_is_running(current))
4817 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4820 } while (!need_resched());
4822 __set_current_state(TASK_RUNNING);
4826 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4829 if (!(flags & BLK_POLL_NOSLEEP) &&
4830 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4831 if (blk_mq_poll_hybrid(q, cookie))
4834 return blk_mq_poll_classic(q, cookie, iob, flags);
4837 unsigned int blk_mq_rq_cpu(struct request *rq)
4839 return rq->mq_ctx->cpu;
4841 EXPORT_SYMBOL(blk_mq_rq_cpu);
4843 void blk_mq_cancel_work_sync(struct request_queue *q)
4845 if (queue_is_mq(q)) {
4846 struct blk_mq_hw_ctx *hctx;
4849 cancel_delayed_work_sync(&q->requeue_work);
4851 queue_for_each_hw_ctx(q, hctx, i)
4852 cancel_delayed_work_sync(&hctx->run_work);
4856 static int __init blk_mq_init(void)
4860 for_each_possible_cpu(i)
4861 init_llist_head(&per_cpu(blk_cpu_done, i));
4862 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4864 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4865 "block/softirq:dead", NULL,
4866 blk_softirq_cpu_dead);
4867 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4868 blk_mq_hctx_notify_dead);
4869 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4870 blk_mq_hctx_notify_online,
4871 blk_mq_hctx_notify_offline);
4874 subsys_initcall(blk_mq_init);