1 // SPDX-License-Identifier: GPL-2.0
3 * Block multiqueue core code
5 * Copyright (C) 2013-2014 Jens Axboe
6 * Copyright (C) 2013-2014 Christoph Hellwig
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/backing-dev.h>
11 #include <linux/bio.h>
12 #include <linux/blkdev.h>
13 #include <linux/blk-integrity.h>
14 #include <linux/kmemleak.h>
16 #include <linux/init.h>
17 #include <linux/slab.h>
18 #include <linux/workqueue.h>
19 #include <linux/smp.h>
20 #include <linux/interrupt.h>
21 #include <linux/llist.h>
22 #include <linux/cpu.h>
23 #include <linux/cache.h>
24 #include <linux/sched/sysctl.h>
25 #include <linux/sched/topology.h>
26 #include <linux/sched/signal.h>
27 #include <linux/delay.h>
28 #include <linux/crash_dump.h>
29 #include <linux/prefetch.h>
30 #include <linux/blk-crypto.h>
31 #include <linux/part_stat.h>
33 #include <trace/events/block.h>
35 #include <linux/blk-mq.h>
36 #include <linux/t10-pi.h>
39 #include "blk-mq-debugfs.h"
40 #include "blk-mq-tag.h"
43 #include "blk-mq-sched.h"
44 #include "blk-rq-qos.h"
46 static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
48 static void blk_mq_poll_stats_start(struct request_queue *q);
49 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
51 static int blk_mq_poll_stats_bkt(const struct request *rq)
53 int ddir, sectors, bucket;
55 ddir = rq_data_dir(rq);
56 sectors = blk_rq_stats_sectors(rq);
58 bucket = ddir + 2 * ilog2(sectors);
62 else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
63 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
68 #define BLK_QC_T_SHIFT 16
69 #define BLK_QC_T_INTERNAL (1U << 31)
71 static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q,
74 return xa_load(&q->hctx_table,
75 (qc & ~BLK_QC_T_INTERNAL) >> BLK_QC_T_SHIFT);
78 static inline struct request *blk_qc_to_rq(struct blk_mq_hw_ctx *hctx,
81 unsigned int tag = qc & ((1U << BLK_QC_T_SHIFT) - 1);
83 if (qc & BLK_QC_T_INTERNAL)
84 return blk_mq_tag_to_rq(hctx->sched_tags, tag);
85 return blk_mq_tag_to_rq(hctx->tags, tag);
88 static inline blk_qc_t blk_rq_to_qc(struct request *rq)
90 return (rq->mq_hctx->queue_num << BLK_QC_T_SHIFT) |
92 rq->tag : (rq->internal_tag | BLK_QC_T_INTERNAL));
96 * Check if any of the ctx, dispatch list or elevator
97 * have pending work in this hardware queue.
99 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
101 return !list_empty_careful(&hctx->dispatch) ||
102 sbitmap_any_bit_set(&hctx->ctx_map) ||
103 blk_mq_sched_has_work(hctx);
107 * Mark this ctx as having pending work in this hardware queue
109 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
110 struct blk_mq_ctx *ctx)
112 const int bit = ctx->index_hw[hctx->type];
114 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
115 sbitmap_set_bit(&hctx->ctx_map, bit);
118 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
119 struct blk_mq_ctx *ctx)
121 const int bit = ctx->index_hw[hctx->type];
123 sbitmap_clear_bit(&hctx->ctx_map, bit);
127 struct block_device *part;
128 unsigned int inflight[2];
131 static bool blk_mq_check_inflight(struct request *rq, void *priv,
134 struct mq_inflight *mi = priv;
136 if ((!mi->part->bd_partno || rq->part == mi->part) &&
137 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
138 mi->inflight[rq_data_dir(rq)]++;
143 unsigned int blk_mq_in_flight(struct request_queue *q,
144 struct block_device *part)
146 struct mq_inflight mi = { .part = part };
148 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
150 return mi.inflight[0] + mi.inflight[1];
153 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
154 unsigned int inflight[2])
156 struct mq_inflight mi = { .part = part };
158 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
159 inflight[0] = mi.inflight[0];
160 inflight[1] = mi.inflight[1];
163 void blk_freeze_queue_start(struct request_queue *q)
165 mutex_lock(&q->mq_freeze_lock);
166 if (++q->mq_freeze_depth == 1) {
167 percpu_ref_kill(&q->q_usage_counter);
168 mutex_unlock(&q->mq_freeze_lock);
170 blk_mq_run_hw_queues(q, false);
172 mutex_unlock(&q->mq_freeze_lock);
175 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
177 void blk_mq_freeze_queue_wait(struct request_queue *q)
179 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
181 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
183 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
184 unsigned long timeout)
186 return wait_event_timeout(q->mq_freeze_wq,
187 percpu_ref_is_zero(&q->q_usage_counter),
190 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
193 * Guarantee no request is in use, so we can change any data structure of
194 * the queue afterward.
196 void blk_freeze_queue(struct request_queue *q)
199 * In the !blk_mq case we are only calling this to kill the
200 * q_usage_counter, otherwise this increases the freeze depth
201 * and waits for it to return to zero. For this reason there is
202 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
203 * exported to drivers as the only user for unfreeze is blk_mq.
205 blk_freeze_queue_start(q);
206 blk_mq_freeze_queue_wait(q);
209 void blk_mq_freeze_queue(struct request_queue *q)
212 * ...just an alias to keep freeze and unfreeze actions balanced
213 * in the blk_mq_* namespace
217 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
219 void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
221 mutex_lock(&q->mq_freeze_lock);
223 q->q_usage_counter.data->force_atomic = true;
224 q->mq_freeze_depth--;
225 WARN_ON_ONCE(q->mq_freeze_depth < 0);
226 if (!q->mq_freeze_depth) {
227 percpu_ref_resurrect(&q->q_usage_counter);
228 wake_up_all(&q->mq_freeze_wq);
230 mutex_unlock(&q->mq_freeze_lock);
233 void blk_mq_unfreeze_queue(struct request_queue *q)
235 __blk_mq_unfreeze_queue(q, false);
237 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
240 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
241 * mpt3sas driver such that this function can be removed.
243 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
247 spin_lock_irqsave(&q->queue_lock, flags);
248 if (!q->quiesce_depth++)
249 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
250 spin_unlock_irqrestore(&q->queue_lock, flags);
252 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
255 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
258 * Note: it is driver's responsibility for making sure that quiesce has
261 void blk_mq_wait_quiesce_done(struct request_queue *q)
263 if (blk_queue_has_srcu(q))
264 synchronize_srcu(q->srcu);
268 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
271 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
274 * Note: this function does not prevent that the struct request end_io()
275 * callback function is invoked. Once this function is returned, we make
276 * sure no dispatch can happen until the queue is unquiesced via
277 * blk_mq_unquiesce_queue().
279 void blk_mq_quiesce_queue(struct request_queue *q)
281 blk_mq_quiesce_queue_nowait(q);
282 blk_mq_wait_quiesce_done(q);
284 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
287 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
290 * This function recovers queue into the state before quiescing
291 * which is done by blk_mq_quiesce_queue.
293 void blk_mq_unquiesce_queue(struct request_queue *q)
296 bool run_queue = false;
298 spin_lock_irqsave(&q->queue_lock, flags);
299 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
301 } else if (!--q->quiesce_depth) {
302 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
305 spin_unlock_irqrestore(&q->queue_lock, flags);
307 /* dispatch requests which are inserted during quiescing */
309 blk_mq_run_hw_queues(q, true);
311 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
313 void blk_mq_wake_waiters(struct request_queue *q)
315 struct blk_mq_hw_ctx *hctx;
318 queue_for_each_hw_ctx(q, hctx, i)
319 if (blk_mq_hw_queue_mapped(hctx))
320 blk_mq_tag_wakeup_all(hctx->tags, true);
323 void blk_rq_init(struct request_queue *q, struct request *rq)
325 memset(rq, 0, sizeof(*rq));
327 INIT_LIST_HEAD(&rq->queuelist);
329 rq->__sector = (sector_t) -1;
330 INIT_HLIST_NODE(&rq->hash);
331 RB_CLEAR_NODE(&rq->rb_node);
332 rq->tag = BLK_MQ_NO_TAG;
333 rq->internal_tag = BLK_MQ_NO_TAG;
334 rq->start_time_ns = ktime_get_ns();
336 blk_crypto_rq_set_defaults(rq);
338 EXPORT_SYMBOL(blk_rq_init);
340 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
341 struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns)
343 struct blk_mq_ctx *ctx = data->ctx;
344 struct blk_mq_hw_ctx *hctx = data->hctx;
345 struct request_queue *q = data->q;
346 struct request *rq = tags->static_rqs[tag];
351 rq->cmd_flags = data->cmd_flags;
353 if (data->flags & BLK_MQ_REQ_PM)
354 data->rq_flags |= RQF_PM;
355 if (blk_queue_io_stat(q))
356 data->rq_flags |= RQF_IO_STAT;
357 rq->rq_flags = data->rq_flags;
359 if (!(data->rq_flags & RQF_ELV)) {
361 rq->internal_tag = BLK_MQ_NO_TAG;
363 rq->tag = BLK_MQ_NO_TAG;
364 rq->internal_tag = tag;
368 if (blk_mq_need_time_stamp(rq))
369 rq->start_time_ns = ktime_get_ns();
371 rq->start_time_ns = 0;
373 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
374 rq->alloc_time_ns = alloc_time_ns;
376 rq->io_start_time_ns = 0;
377 rq->stats_sectors = 0;
378 rq->nr_phys_segments = 0;
379 #if defined(CONFIG_BLK_DEV_INTEGRITY)
380 rq->nr_integrity_segments = 0;
383 rq->end_io_data = NULL;
385 blk_crypto_rq_set_defaults(rq);
386 INIT_LIST_HEAD(&rq->queuelist);
387 /* tag was already set */
388 WRITE_ONCE(rq->deadline, 0);
391 if (rq->rq_flags & RQF_ELV) {
392 struct elevator_queue *e = data->q->elevator;
394 INIT_HLIST_NODE(&rq->hash);
395 RB_CLEAR_NODE(&rq->rb_node);
397 if (!op_is_flush(data->cmd_flags) &&
398 e->type->ops.prepare_request) {
399 e->type->ops.prepare_request(rq);
400 rq->rq_flags |= RQF_ELVPRIV;
407 static inline struct request *
408 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data,
411 unsigned int tag, tag_offset;
412 struct blk_mq_tags *tags;
414 unsigned long tag_mask;
417 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
418 if (unlikely(!tag_mask))
421 tags = blk_mq_tags_from_data(data);
422 for (i = 0; tag_mask; i++) {
423 if (!(tag_mask & (1UL << i)))
425 tag = tag_offset + i;
426 prefetch(tags->static_rqs[tag]);
427 tag_mask &= ~(1UL << i);
428 rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns);
429 rq_list_add(data->cached_rq, rq);
432 /* caller already holds a reference, add for remainder */
433 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
436 return rq_list_pop(data->cached_rq);
439 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
441 struct request_queue *q = data->q;
442 u64 alloc_time_ns = 0;
446 /* alloc_time includes depth and tag waits */
447 if (blk_queue_rq_alloc_time(q))
448 alloc_time_ns = ktime_get_ns();
450 if (data->cmd_flags & REQ_NOWAIT)
451 data->flags |= BLK_MQ_REQ_NOWAIT;
454 struct elevator_queue *e = q->elevator;
456 data->rq_flags |= RQF_ELV;
459 * Flush/passthrough requests are special and go directly to the
460 * dispatch list. Don't include reserved tags in the
461 * limiting, as it isn't useful.
463 if (!op_is_flush(data->cmd_flags) &&
464 !blk_op_is_passthrough(data->cmd_flags) &&
465 e->type->ops.limit_depth &&
466 !(data->flags & BLK_MQ_REQ_RESERVED))
467 e->type->ops.limit_depth(data->cmd_flags, data);
471 data->ctx = blk_mq_get_ctx(q);
472 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
473 if (!(data->rq_flags & RQF_ELV))
474 blk_mq_tag_busy(data->hctx);
477 * Try batched alloc if we want more than 1 tag.
479 if (data->nr_tags > 1) {
480 rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns);
487 * Waiting allocations only fail because of an inactive hctx. In that
488 * case just retry the hctx assignment and tag allocation as CPU hotplug
489 * should have migrated us to an online CPU by now.
491 tag = blk_mq_get_tag(data);
492 if (tag == BLK_MQ_NO_TAG) {
493 if (data->flags & BLK_MQ_REQ_NOWAIT)
496 * Give up the CPU and sleep for a random short time to
497 * ensure that thread using a realtime scheduling class
498 * are migrated off the CPU, and thus off the hctx that
505 return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag,
509 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
510 blk_mq_req_flags_t flags)
512 struct blk_mq_alloc_data data = {
521 ret = blk_queue_enter(q, flags);
525 rq = __blk_mq_alloc_requests(&data);
529 rq->__sector = (sector_t) -1;
530 rq->bio = rq->biotail = NULL;
534 return ERR_PTR(-EWOULDBLOCK);
536 EXPORT_SYMBOL(blk_mq_alloc_request);
538 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
539 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
541 struct blk_mq_alloc_data data = {
547 u64 alloc_time_ns = 0;
552 /* alloc_time includes depth and tag waits */
553 if (blk_queue_rq_alloc_time(q))
554 alloc_time_ns = ktime_get_ns();
557 * If the tag allocator sleeps we could get an allocation for a
558 * different hardware context. No need to complicate the low level
559 * allocator for this for the rare use case of a command tied to
562 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
563 return ERR_PTR(-EINVAL);
565 if (hctx_idx >= q->nr_hw_queues)
566 return ERR_PTR(-EIO);
568 ret = blk_queue_enter(q, flags);
573 * Check if the hardware context is actually mapped to anything.
574 * If not tell the caller that it should skip this queue.
577 data.hctx = xa_load(&q->hctx_table, hctx_idx);
578 if (!blk_mq_hw_queue_mapped(data.hctx))
580 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
581 data.ctx = __blk_mq_get_ctx(q, cpu);
584 blk_mq_tag_busy(data.hctx);
586 data.rq_flags |= RQF_ELV;
589 tag = blk_mq_get_tag(&data);
590 if (tag == BLK_MQ_NO_TAG)
592 return blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag,
599 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
601 static void __blk_mq_free_request(struct request *rq)
603 struct request_queue *q = rq->q;
604 struct blk_mq_ctx *ctx = rq->mq_ctx;
605 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
606 const int sched_tag = rq->internal_tag;
608 blk_crypto_free_request(rq);
609 blk_pm_mark_last_busy(rq);
611 if (rq->tag != BLK_MQ_NO_TAG)
612 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
613 if (sched_tag != BLK_MQ_NO_TAG)
614 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
615 blk_mq_sched_restart(hctx);
619 void blk_mq_free_request(struct request *rq)
621 struct request_queue *q = rq->q;
622 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
624 if ((rq->rq_flags & RQF_ELVPRIV) &&
625 q->elevator->type->ops.finish_request)
626 q->elevator->type->ops.finish_request(rq);
628 if (rq->rq_flags & RQF_MQ_INFLIGHT)
629 __blk_mq_dec_active_requests(hctx);
631 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
632 laptop_io_completion(q->disk->bdi);
636 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
637 if (req_ref_put_and_test(rq))
638 __blk_mq_free_request(rq);
640 EXPORT_SYMBOL_GPL(blk_mq_free_request);
642 void blk_mq_free_plug_rqs(struct blk_plug *plug)
646 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
647 blk_mq_free_request(rq);
650 void blk_dump_rq_flags(struct request *rq, char *msg)
652 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
653 rq->q->disk ? rq->q->disk->disk_name : "?",
654 (unsigned long long) rq->cmd_flags);
656 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
657 (unsigned long long)blk_rq_pos(rq),
658 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
659 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
660 rq->bio, rq->biotail, blk_rq_bytes(rq));
662 EXPORT_SYMBOL(blk_dump_rq_flags);
664 static void req_bio_endio(struct request *rq, struct bio *bio,
665 unsigned int nbytes, blk_status_t error)
667 if (unlikely(error)) {
668 bio->bi_status = error;
669 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
671 * Partial zone append completions cannot be supported as the
672 * BIO fragments may end up not being written sequentially.
674 if (bio->bi_iter.bi_size != nbytes)
675 bio->bi_status = BLK_STS_IOERR;
677 bio->bi_iter.bi_sector = rq->__sector;
680 bio_advance(bio, nbytes);
682 if (unlikely(rq->rq_flags & RQF_QUIET))
683 bio_set_flag(bio, BIO_QUIET);
684 /* don't actually finish bio if it's part of flush sequence */
685 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
689 static void blk_account_io_completion(struct request *req, unsigned int bytes)
691 if (req->part && blk_do_io_stat(req)) {
692 const int sgrp = op_stat_group(req_op(req));
695 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
700 static void blk_print_req_error(struct request *req, blk_status_t status)
702 printk_ratelimited(KERN_ERR
703 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
704 "phys_seg %u prio class %u\n",
705 blk_status_to_str(status),
706 req->q->disk ? req->q->disk->disk_name : "?",
707 blk_rq_pos(req), req_op(req), blk_op_str(req_op(req)),
708 req->cmd_flags & ~REQ_OP_MASK,
709 req->nr_phys_segments,
710 IOPRIO_PRIO_CLASS(req->ioprio));
714 * Fully end IO on a request. Does not support partial completions, or
717 static void blk_complete_request(struct request *req)
719 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
720 int total_bytes = blk_rq_bytes(req);
721 struct bio *bio = req->bio;
723 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
728 #ifdef CONFIG_BLK_DEV_INTEGRITY
729 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
730 req->q->integrity.profile->complete_fn(req, total_bytes);
733 blk_account_io_completion(req, total_bytes);
736 struct bio *next = bio->bi_next;
738 /* Completion has already been traced */
739 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
741 if (req_op(req) == REQ_OP_ZONE_APPEND)
742 bio->bi_iter.bi_sector = req->__sector;
750 * Reset counters so that the request stacking driver
751 * can find how many bytes remain in the request
759 * blk_update_request - Complete multiple bytes without completing the request
760 * @req: the request being processed
761 * @error: block status code
762 * @nr_bytes: number of bytes to complete for @req
765 * Ends I/O on a number of bytes attached to @req, but doesn't complete
766 * the request structure even if @req doesn't have leftover.
767 * If @req has leftover, sets it up for the next range of segments.
769 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
770 * %false return from this function.
773 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
774 * except in the consistency check at the end of this function.
777 * %false - this request doesn't have any more data
778 * %true - this request has more data
780 bool blk_update_request(struct request *req, blk_status_t error,
781 unsigned int nr_bytes)
785 trace_block_rq_complete(req, error, nr_bytes);
790 #ifdef CONFIG_BLK_DEV_INTEGRITY
791 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
793 req->q->integrity.profile->complete_fn(req, nr_bytes);
796 if (unlikely(error && !blk_rq_is_passthrough(req) &&
797 !(req->rq_flags & RQF_QUIET))) {
798 blk_print_req_error(req, error);
799 trace_block_rq_error(req, error, nr_bytes);
802 blk_account_io_completion(req, nr_bytes);
806 struct bio *bio = req->bio;
807 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
809 if (bio_bytes == bio->bi_iter.bi_size)
810 req->bio = bio->bi_next;
812 /* Completion has already been traced */
813 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
814 req_bio_endio(req, bio, bio_bytes, error);
816 total_bytes += bio_bytes;
817 nr_bytes -= bio_bytes;
828 * Reset counters so that the request stacking driver
829 * can find how many bytes remain in the request
836 req->__data_len -= total_bytes;
838 /* update sector only for requests with clear definition of sector */
839 if (!blk_rq_is_passthrough(req))
840 req->__sector += total_bytes >> 9;
842 /* mixed attributes always follow the first bio */
843 if (req->rq_flags & RQF_MIXED_MERGE) {
844 req->cmd_flags &= ~REQ_FAILFAST_MASK;
845 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
848 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
850 * If total number of sectors is less than the first segment
851 * size, something has gone terribly wrong.
853 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
854 blk_dump_rq_flags(req, "request botched");
855 req->__data_len = blk_rq_cur_bytes(req);
858 /* recalculate the number of segments */
859 req->nr_phys_segments = blk_recalc_rq_segments(req);
864 EXPORT_SYMBOL_GPL(blk_update_request);
866 static void __blk_account_io_done(struct request *req, u64 now)
868 const int sgrp = op_stat_group(req_op(req));
871 update_io_ticks(req->part, jiffies, true);
872 part_stat_inc(req->part, ios[sgrp]);
873 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
877 static inline void blk_account_io_done(struct request *req, u64 now)
880 * Account IO completion. flush_rq isn't accounted as a
881 * normal IO on queueing nor completion. Accounting the
882 * containing request is enough.
884 if (blk_do_io_stat(req) && req->part &&
885 !(req->rq_flags & RQF_FLUSH_SEQ))
886 __blk_account_io_done(req, now);
889 static void __blk_account_io_start(struct request *rq)
892 * All non-passthrough requests are created from a bio with one
893 * exception: when a flush command that is part of a flush sequence
894 * generated by the state machine in blk-flush.c is cloned onto the
895 * lower device by dm-multipath we can get here without a bio.
898 rq->part = rq->bio->bi_bdev;
900 rq->part = rq->q->disk->part0;
903 update_io_ticks(rq->part, jiffies, false);
907 static inline void blk_account_io_start(struct request *req)
909 if (blk_do_io_stat(req))
910 __blk_account_io_start(req);
913 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
915 if (rq->rq_flags & RQF_STATS) {
916 blk_mq_poll_stats_start(rq->q);
917 blk_stat_add(rq, now);
920 blk_mq_sched_completed_request(rq, now);
921 blk_account_io_done(rq, now);
924 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
926 if (blk_mq_need_time_stamp(rq))
927 __blk_mq_end_request_acct(rq, ktime_get_ns());
930 rq_qos_done(rq->q, rq);
931 rq->end_io(rq, error);
933 blk_mq_free_request(rq);
936 EXPORT_SYMBOL(__blk_mq_end_request);
938 void blk_mq_end_request(struct request *rq, blk_status_t error)
940 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
942 __blk_mq_end_request(rq, error);
944 EXPORT_SYMBOL(blk_mq_end_request);
946 #define TAG_COMP_BATCH 32
948 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
949 int *tag_array, int nr_tags)
951 struct request_queue *q = hctx->queue;
954 * All requests should have been marked as RQF_MQ_INFLIGHT, so
955 * update hctx->nr_active in batch
957 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
958 __blk_mq_sub_active_requests(hctx, nr_tags);
960 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
961 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
964 void blk_mq_end_request_batch(struct io_comp_batch *iob)
966 int tags[TAG_COMP_BATCH], nr_tags = 0;
967 struct blk_mq_hw_ctx *cur_hctx = NULL;
972 now = ktime_get_ns();
974 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
976 prefetch(rq->rq_next);
978 blk_complete_request(rq);
980 __blk_mq_end_request_acct(rq, now);
982 rq_qos_done(rq->q, rq);
984 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
985 if (!req_ref_put_and_test(rq))
988 blk_crypto_free_request(rq);
989 blk_pm_mark_last_busy(rq);
991 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
993 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
995 cur_hctx = rq->mq_hctx;
997 tags[nr_tags++] = rq->tag;
1001 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1003 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1005 static void blk_complete_reqs(struct llist_head *list)
1007 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1008 struct request *rq, *next;
1010 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1011 rq->q->mq_ops->complete(rq);
1014 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1016 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1019 static int blk_softirq_cpu_dead(unsigned int cpu)
1021 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1025 static void __blk_mq_complete_request_remote(void *data)
1027 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1030 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1032 int cpu = raw_smp_processor_id();
1034 if (!IS_ENABLED(CONFIG_SMP) ||
1035 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1038 * With force threaded interrupts enabled, raising softirq from an SMP
1039 * function call will always result in waking the ksoftirqd thread.
1040 * This is probably worse than completing the request on a different
1043 if (force_irqthreads())
1046 /* same CPU or cache domain? Complete locally */
1047 if (cpu == rq->mq_ctx->cpu ||
1048 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1049 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1052 /* don't try to IPI to an offline CPU */
1053 return cpu_online(rq->mq_ctx->cpu);
1056 static void blk_mq_complete_send_ipi(struct request *rq)
1058 struct llist_head *list;
1061 cpu = rq->mq_ctx->cpu;
1062 list = &per_cpu(blk_cpu_done, cpu);
1063 if (llist_add(&rq->ipi_list, list)) {
1064 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1065 smp_call_function_single_async(cpu, &rq->csd);
1069 static void blk_mq_raise_softirq(struct request *rq)
1071 struct llist_head *list;
1074 list = this_cpu_ptr(&blk_cpu_done);
1075 if (llist_add(&rq->ipi_list, list))
1076 raise_softirq(BLOCK_SOFTIRQ);
1080 bool blk_mq_complete_request_remote(struct request *rq)
1082 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1085 * For a polled request, always complete locallly, it's pointless
1086 * to redirect the completion.
1088 if (rq->cmd_flags & REQ_POLLED)
1091 if (blk_mq_complete_need_ipi(rq)) {
1092 blk_mq_complete_send_ipi(rq);
1096 if (rq->q->nr_hw_queues == 1) {
1097 blk_mq_raise_softirq(rq);
1102 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1105 * blk_mq_complete_request - end I/O on a request
1106 * @rq: the request being processed
1109 * Complete a request by scheduling the ->complete_rq operation.
1111 void blk_mq_complete_request(struct request *rq)
1113 if (!blk_mq_complete_request_remote(rq))
1114 rq->q->mq_ops->complete(rq);
1116 EXPORT_SYMBOL(blk_mq_complete_request);
1119 * blk_mq_start_request - Start processing a request
1120 * @rq: Pointer to request to be started
1122 * Function used by device drivers to notify the block layer that a request
1123 * is going to be processed now, so blk layer can do proper initializations
1124 * such as starting the timeout timer.
1126 void blk_mq_start_request(struct request *rq)
1128 struct request_queue *q = rq->q;
1130 trace_block_rq_issue(rq);
1132 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1134 #ifdef CONFIG_BLK_CGROUP
1136 start_time = bio_issue_time(&rq->bio->bi_issue);
1139 start_time = ktime_get_ns();
1140 rq->io_start_time_ns = start_time;
1141 rq->stats_sectors = blk_rq_sectors(rq);
1142 rq->rq_flags |= RQF_STATS;
1143 rq_qos_issue(q, rq);
1146 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1149 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1151 #ifdef CONFIG_BLK_DEV_INTEGRITY
1152 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1153 q->integrity.profile->prepare_fn(rq);
1155 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1156 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1158 EXPORT_SYMBOL(blk_mq_start_request);
1161 * blk_end_sync_rq - executes a completion event on a request
1162 * @rq: request to complete
1163 * @error: end I/O status of the request
1165 static void blk_end_sync_rq(struct request *rq, blk_status_t error)
1167 struct completion *waiting = rq->end_io_data;
1169 rq->end_io_data = (void *)(uintptr_t)error;
1172 * complete last, if this is a stack request the process (and thus
1173 * the rq pointer) could be invalid right after this complete()
1179 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1180 * @rq: request to insert
1181 * @at_head: insert request at head or tail of queue
1182 * @done: I/O completion handler
1185 * Insert a fully prepared request at the back of the I/O scheduler queue
1186 * for execution. Don't wait for completion.
1189 * This function will invoke @done directly if the queue is dead.
1191 void blk_execute_rq_nowait(struct request *rq, bool at_head, rq_end_io_fn *done)
1193 WARN_ON(irqs_disabled());
1194 WARN_ON(!blk_rq_is_passthrough(rq));
1198 blk_account_io_start(rq);
1201 * don't check dying flag for MQ because the request won't
1202 * be reused after dying flag is set
1204 blk_mq_sched_insert_request(rq, at_head, true, false);
1206 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1208 static bool blk_rq_is_poll(struct request *rq)
1212 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1214 if (WARN_ON_ONCE(!rq->bio))
1219 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1222 bio_poll(rq->bio, NULL, 0);
1224 } while (!completion_done(wait));
1228 * blk_execute_rq - insert a request into queue for execution
1229 * @rq: request to insert
1230 * @at_head: insert request at head or tail of queue
1233 * Insert a fully prepared request at the back of the I/O scheduler queue
1234 * for execution and wait for completion.
1235 * Return: The blk_status_t result provided to blk_mq_end_request().
1237 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1239 DECLARE_COMPLETION_ONSTACK(wait);
1240 unsigned long hang_check;
1242 rq->end_io_data = &wait;
1243 blk_execute_rq_nowait(rq, at_head, blk_end_sync_rq);
1245 /* Prevent hang_check timer from firing at us during very long I/O */
1246 hang_check = sysctl_hung_task_timeout_secs;
1248 if (blk_rq_is_poll(rq))
1249 blk_rq_poll_completion(rq, &wait);
1250 else if (hang_check)
1251 while (!wait_for_completion_io_timeout(&wait,
1252 hang_check * (HZ/2)))
1255 wait_for_completion_io(&wait);
1257 return (blk_status_t)(uintptr_t)rq->end_io_data;
1259 EXPORT_SYMBOL(blk_execute_rq);
1261 static void __blk_mq_requeue_request(struct request *rq)
1263 struct request_queue *q = rq->q;
1265 blk_mq_put_driver_tag(rq);
1267 trace_block_rq_requeue(rq);
1268 rq_qos_requeue(q, rq);
1270 if (blk_mq_request_started(rq)) {
1271 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1272 rq->rq_flags &= ~RQF_TIMED_OUT;
1276 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1278 __blk_mq_requeue_request(rq);
1280 /* this request will be re-inserted to io scheduler queue */
1281 blk_mq_sched_requeue_request(rq);
1283 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1285 EXPORT_SYMBOL(blk_mq_requeue_request);
1287 static void blk_mq_requeue_work(struct work_struct *work)
1289 struct request_queue *q =
1290 container_of(work, struct request_queue, requeue_work.work);
1292 struct request *rq, *next;
1294 spin_lock_irq(&q->requeue_lock);
1295 list_splice_init(&q->requeue_list, &rq_list);
1296 spin_unlock_irq(&q->requeue_lock);
1298 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1299 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1302 rq->rq_flags &= ~RQF_SOFTBARRIER;
1303 list_del_init(&rq->queuelist);
1305 * If RQF_DONTPREP, rq has contained some driver specific
1306 * data, so insert it to hctx dispatch list to avoid any
1309 if (rq->rq_flags & RQF_DONTPREP)
1310 blk_mq_request_bypass_insert(rq, false, false);
1312 blk_mq_sched_insert_request(rq, true, false, false);
1315 while (!list_empty(&rq_list)) {
1316 rq = list_entry(rq_list.next, struct request, queuelist);
1317 list_del_init(&rq->queuelist);
1318 blk_mq_sched_insert_request(rq, false, false, false);
1321 blk_mq_run_hw_queues(q, false);
1324 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1325 bool kick_requeue_list)
1327 struct request_queue *q = rq->q;
1328 unsigned long flags;
1331 * We abuse this flag that is otherwise used by the I/O scheduler to
1332 * request head insertion from the workqueue.
1334 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1336 spin_lock_irqsave(&q->requeue_lock, flags);
1338 rq->rq_flags |= RQF_SOFTBARRIER;
1339 list_add(&rq->queuelist, &q->requeue_list);
1341 list_add_tail(&rq->queuelist, &q->requeue_list);
1343 spin_unlock_irqrestore(&q->requeue_lock, flags);
1345 if (kick_requeue_list)
1346 blk_mq_kick_requeue_list(q);
1349 void blk_mq_kick_requeue_list(struct request_queue *q)
1351 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1353 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1355 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1356 unsigned long msecs)
1358 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1359 msecs_to_jiffies(msecs));
1361 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1363 static bool blk_mq_rq_inflight(struct request *rq, void *priv,
1367 * If we find a request that isn't idle we know the queue is busy
1368 * as it's checked in the iter.
1369 * Return false to stop the iteration.
1371 if (blk_mq_request_started(rq)) {
1381 bool blk_mq_queue_inflight(struct request_queue *q)
1385 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1388 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1390 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
1392 req->rq_flags |= RQF_TIMED_OUT;
1393 if (req->q->mq_ops->timeout) {
1394 enum blk_eh_timer_return ret;
1396 ret = req->q->mq_ops->timeout(req, reserved);
1397 if (ret == BLK_EH_DONE)
1399 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1405 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
1407 unsigned long deadline;
1409 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1411 if (rq->rq_flags & RQF_TIMED_OUT)
1414 deadline = READ_ONCE(rq->deadline);
1415 if (time_after_eq(jiffies, deadline))
1420 else if (time_after(*next, deadline))
1425 void blk_mq_put_rq_ref(struct request *rq)
1427 if (is_flush_rq(rq))
1429 else if (req_ref_put_and_test(rq))
1430 __blk_mq_free_request(rq);
1433 static bool blk_mq_check_expired(struct request *rq, void *priv, bool reserved)
1435 unsigned long *next = priv;
1438 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1439 * be reallocated underneath the timeout handler's processing, then
1440 * the expire check is reliable. If the request is not expired, then
1441 * it was completed and reallocated as a new request after returning
1442 * from blk_mq_check_expired().
1444 if (blk_mq_req_expired(rq, next))
1445 blk_mq_rq_timed_out(rq, reserved);
1449 static void blk_mq_timeout_work(struct work_struct *work)
1451 struct request_queue *q =
1452 container_of(work, struct request_queue, timeout_work);
1453 unsigned long next = 0;
1454 struct blk_mq_hw_ctx *hctx;
1457 /* A deadlock might occur if a request is stuck requiring a
1458 * timeout at the same time a queue freeze is waiting
1459 * completion, since the timeout code would not be able to
1460 * acquire the queue reference here.
1462 * That's why we don't use blk_queue_enter here; instead, we use
1463 * percpu_ref_tryget directly, because we need to be able to
1464 * obtain a reference even in the short window between the queue
1465 * starting to freeze, by dropping the first reference in
1466 * blk_freeze_queue_start, and the moment the last request is
1467 * consumed, marked by the instant q_usage_counter reaches
1470 if (!percpu_ref_tryget(&q->q_usage_counter))
1473 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
1476 mod_timer(&q->timeout, next);
1479 * Request timeouts are handled as a forward rolling timer. If
1480 * we end up here it means that no requests are pending and
1481 * also that no request has been pending for a while. Mark
1482 * each hctx as idle.
1484 queue_for_each_hw_ctx(q, hctx, i) {
1485 /* the hctx may be unmapped, so check it here */
1486 if (blk_mq_hw_queue_mapped(hctx))
1487 blk_mq_tag_idle(hctx);
1493 struct flush_busy_ctx_data {
1494 struct blk_mq_hw_ctx *hctx;
1495 struct list_head *list;
1498 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1500 struct flush_busy_ctx_data *flush_data = data;
1501 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1502 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1503 enum hctx_type type = hctx->type;
1505 spin_lock(&ctx->lock);
1506 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1507 sbitmap_clear_bit(sb, bitnr);
1508 spin_unlock(&ctx->lock);
1513 * Process software queues that have been marked busy, splicing them
1514 * to the for-dispatch
1516 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1518 struct flush_busy_ctx_data data = {
1523 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1525 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1527 struct dispatch_rq_data {
1528 struct blk_mq_hw_ctx *hctx;
1532 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1535 struct dispatch_rq_data *dispatch_data = data;
1536 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1537 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1538 enum hctx_type type = hctx->type;
1540 spin_lock(&ctx->lock);
1541 if (!list_empty(&ctx->rq_lists[type])) {
1542 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1543 list_del_init(&dispatch_data->rq->queuelist);
1544 if (list_empty(&ctx->rq_lists[type]))
1545 sbitmap_clear_bit(sb, bitnr);
1547 spin_unlock(&ctx->lock);
1549 return !dispatch_data->rq;
1552 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1553 struct blk_mq_ctx *start)
1555 unsigned off = start ? start->index_hw[hctx->type] : 0;
1556 struct dispatch_rq_data data = {
1561 __sbitmap_for_each_set(&hctx->ctx_map, off,
1562 dispatch_rq_from_ctx, &data);
1567 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1569 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1570 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1573 blk_mq_tag_busy(rq->mq_hctx);
1575 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1576 bt = &rq->mq_hctx->tags->breserved_tags;
1579 if (!hctx_may_queue(rq->mq_hctx, bt))
1583 tag = __sbitmap_queue_get(bt);
1584 if (tag == BLK_MQ_NO_TAG)
1587 rq->tag = tag + tag_offset;
1591 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1593 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1596 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1597 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1598 rq->rq_flags |= RQF_MQ_INFLIGHT;
1599 __blk_mq_inc_active_requests(hctx);
1601 hctx->tags->rqs[rq->tag] = rq;
1605 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1606 int flags, void *key)
1608 struct blk_mq_hw_ctx *hctx;
1610 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1612 spin_lock(&hctx->dispatch_wait_lock);
1613 if (!list_empty(&wait->entry)) {
1614 struct sbitmap_queue *sbq;
1616 list_del_init(&wait->entry);
1617 sbq = &hctx->tags->bitmap_tags;
1618 atomic_dec(&sbq->ws_active);
1620 spin_unlock(&hctx->dispatch_wait_lock);
1622 blk_mq_run_hw_queue(hctx, true);
1627 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1628 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1629 * restart. For both cases, take care to check the condition again after
1630 * marking us as waiting.
1632 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1635 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1636 struct wait_queue_head *wq;
1637 wait_queue_entry_t *wait;
1640 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1641 blk_mq_sched_mark_restart_hctx(hctx);
1644 * It's possible that a tag was freed in the window between the
1645 * allocation failure and adding the hardware queue to the wait
1648 * Don't clear RESTART here, someone else could have set it.
1649 * At most this will cost an extra queue run.
1651 return blk_mq_get_driver_tag(rq);
1654 wait = &hctx->dispatch_wait;
1655 if (!list_empty_careful(&wait->entry))
1658 wq = &bt_wait_ptr(sbq, hctx)->wait;
1660 spin_lock_irq(&wq->lock);
1661 spin_lock(&hctx->dispatch_wait_lock);
1662 if (!list_empty(&wait->entry)) {
1663 spin_unlock(&hctx->dispatch_wait_lock);
1664 spin_unlock_irq(&wq->lock);
1668 atomic_inc(&sbq->ws_active);
1669 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1670 __add_wait_queue(wq, wait);
1673 * It's possible that a tag was freed in the window between the
1674 * allocation failure and adding the hardware queue to the wait
1677 ret = blk_mq_get_driver_tag(rq);
1679 spin_unlock(&hctx->dispatch_wait_lock);
1680 spin_unlock_irq(&wq->lock);
1685 * We got a tag, remove ourselves from the wait queue to ensure
1686 * someone else gets the wakeup.
1688 list_del_init(&wait->entry);
1689 atomic_dec(&sbq->ws_active);
1690 spin_unlock(&hctx->dispatch_wait_lock);
1691 spin_unlock_irq(&wq->lock);
1696 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1697 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1699 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1700 * - EWMA is one simple way to compute running average value
1701 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1702 * - take 4 as factor for avoiding to get too small(0) result, and this
1703 * factor doesn't matter because EWMA decreases exponentially
1705 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1709 ewma = hctx->dispatch_busy;
1714 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1716 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1717 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1719 hctx->dispatch_busy = ewma;
1722 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1724 static void blk_mq_handle_dev_resource(struct request *rq,
1725 struct list_head *list)
1727 struct request *next =
1728 list_first_entry_or_null(list, struct request, queuelist);
1731 * If an I/O scheduler has been configured and we got a driver tag for
1732 * the next request already, free it.
1735 blk_mq_put_driver_tag(next);
1737 list_add(&rq->queuelist, list);
1738 __blk_mq_requeue_request(rq);
1741 static void blk_mq_handle_zone_resource(struct request *rq,
1742 struct list_head *zone_list)
1745 * If we end up here it is because we cannot dispatch a request to a
1746 * specific zone due to LLD level zone-write locking or other zone
1747 * related resource not being available. In this case, set the request
1748 * aside in zone_list for retrying it later.
1750 list_add(&rq->queuelist, zone_list);
1751 __blk_mq_requeue_request(rq);
1754 enum prep_dispatch {
1756 PREP_DISPATCH_NO_TAG,
1757 PREP_DISPATCH_NO_BUDGET,
1760 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1763 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1764 int budget_token = -1;
1767 budget_token = blk_mq_get_dispatch_budget(rq->q);
1768 if (budget_token < 0) {
1769 blk_mq_put_driver_tag(rq);
1770 return PREP_DISPATCH_NO_BUDGET;
1772 blk_mq_set_rq_budget_token(rq, budget_token);
1775 if (!blk_mq_get_driver_tag(rq)) {
1777 * The initial allocation attempt failed, so we need to
1778 * rerun the hardware queue when a tag is freed. The
1779 * waitqueue takes care of that. If the queue is run
1780 * before we add this entry back on the dispatch list,
1781 * we'll re-run it below.
1783 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1785 * All budgets not got from this function will be put
1786 * together during handling partial dispatch
1789 blk_mq_put_dispatch_budget(rq->q, budget_token);
1790 return PREP_DISPATCH_NO_TAG;
1794 return PREP_DISPATCH_OK;
1797 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1798 static void blk_mq_release_budgets(struct request_queue *q,
1799 struct list_head *list)
1803 list_for_each_entry(rq, list, queuelist) {
1804 int budget_token = blk_mq_get_rq_budget_token(rq);
1806 if (budget_token >= 0)
1807 blk_mq_put_dispatch_budget(q, budget_token);
1812 * Returns true if we did some work AND can potentially do more.
1814 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1815 unsigned int nr_budgets)
1817 enum prep_dispatch prep;
1818 struct request_queue *q = hctx->queue;
1819 struct request *rq, *nxt;
1821 blk_status_t ret = BLK_STS_OK;
1822 LIST_HEAD(zone_list);
1823 bool needs_resource = false;
1825 if (list_empty(list))
1829 * Now process all the entries, sending them to the driver.
1831 errors = queued = 0;
1833 struct blk_mq_queue_data bd;
1835 rq = list_first_entry(list, struct request, queuelist);
1837 WARN_ON_ONCE(hctx != rq->mq_hctx);
1838 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1839 if (prep != PREP_DISPATCH_OK)
1842 list_del_init(&rq->queuelist);
1847 * Flag last if we have no more requests, or if we have more
1848 * but can't assign a driver tag to it.
1850 if (list_empty(list))
1853 nxt = list_first_entry(list, struct request, queuelist);
1854 bd.last = !blk_mq_get_driver_tag(nxt);
1858 * once the request is queued to lld, no need to cover the
1863 ret = q->mq_ops->queue_rq(hctx, &bd);
1868 case BLK_STS_RESOURCE:
1869 needs_resource = true;
1871 case BLK_STS_DEV_RESOURCE:
1872 blk_mq_handle_dev_resource(rq, list);
1874 case BLK_STS_ZONE_RESOURCE:
1876 * Move the request to zone_list and keep going through
1877 * the dispatch list to find more requests the drive can
1880 blk_mq_handle_zone_resource(rq, &zone_list);
1881 needs_resource = true;
1885 blk_mq_end_request(rq, ret);
1887 } while (!list_empty(list));
1889 if (!list_empty(&zone_list))
1890 list_splice_tail_init(&zone_list, list);
1892 /* If we didn't flush the entire list, we could have told the driver
1893 * there was more coming, but that turned out to be a lie.
1895 if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued)
1896 q->mq_ops->commit_rqs(hctx);
1898 * Any items that need requeuing? Stuff them into hctx->dispatch,
1899 * that is where we will continue on next queue run.
1901 if (!list_empty(list)) {
1903 /* For non-shared tags, the RESTART check will suffice */
1904 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1905 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1908 blk_mq_release_budgets(q, list);
1910 spin_lock(&hctx->lock);
1911 list_splice_tail_init(list, &hctx->dispatch);
1912 spin_unlock(&hctx->lock);
1915 * Order adding requests to hctx->dispatch and checking
1916 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1917 * in blk_mq_sched_restart(). Avoid restart code path to
1918 * miss the new added requests to hctx->dispatch, meantime
1919 * SCHED_RESTART is observed here.
1924 * If SCHED_RESTART was set by the caller of this function and
1925 * it is no longer set that means that it was cleared by another
1926 * thread and hence that a queue rerun is needed.
1928 * If 'no_tag' is set, that means that we failed getting
1929 * a driver tag with an I/O scheduler attached. If our dispatch
1930 * waitqueue is no longer active, ensure that we run the queue
1931 * AFTER adding our entries back to the list.
1933 * If no I/O scheduler has been configured it is possible that
1934 * the hardware queue got stopped and restarted before requests
1935 * were pushed back onto the dispatch list. Rerun the queue to
1936 * avoid starvation. Notes:
1937 * - blk_mq_run_hw_queue() checks whether or not a queue has
1938 * been stopped before rerunning a queue.
1939 * - Some but not all block drivers stop a queue before
1940 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1943 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1944 * bit is set, run queue after a delay to avoid IO stalls
1945 * that could otherwise occur if the queue is idle. We'll do
1946 * similar if we couldn't get budget or couldn't lock a zone
1947 * and SCHED_RESTART is set.
1949 needs_restart = blk_mq_sched_needs_restart(hctx);
1950 if (prep == PREP_DISPATCH_NO_BUDGET)
1951 needs_resource = true;
1952 if (!needs_restart ||
1953 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1954 blk_mq_run_hw_queue(hctx, true);
1955 else if (needs_restart && needs_resource)
1956 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1958 blk_mq_update_dispatch_busy(hctx, true);
1961 blk_mq_update_dispatch_busy(hctx, false);
1963 return (queued + errors) != 0;
1967 * __blk_mq_run_hw_queue - Run a hardware queue.
1968 * @hctx: Pointer to the hardware queue to run.
1970 * Send pending requests to the hardware.
1972 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1975 * We can't run the queue inline with ints disabled. Ensure that
1976 * we catch bad users of this early.
1978 WARN_ON_ONCE(in_interrupt());
1980 blk_mq_run_dispatch_ops(hctx->queue,
1981 blk_mq_sched_dispatch_requests(hctx));
1984 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1986 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1988 if (cpu >= nr_cpu_ids)
1989 cpu = cpumask_first(hctx->cpumask);
1994 * It'd be great if the workqueue API had a way to pass
1995 * in a mask and had some smarts for more clever placement.
1996 * For now we just round-robin here, switching for every
1997 * BLK_MQ_CPU_WORK_BATCH queued items.
1999 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2002 int next_cpu = hctx->next_cpu;
2004 if (hctx->queue->nr_hw_queues == 1)
2005 return WORK_CPU_UNBOUND;
2007 if (--hctx->next_cpu_batch <= 0) {
2009 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2011 if (next_cpu >= nr_cpu_ids)
2012 next_cpu = blk_mq_first_mapped_cpu(hctx);
2013 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2017 * Do unbound schedule if we can't find a online CPU for this hctx,
2018 * and it should only happen in the path of handling CPU DEAD.
2020 if (!cpu_online(next_cpu)) {
2027 * Make sure to re-select CPU next time once after CPUs
2028 * in hctx->cpumask become online again.
2030 hctx->next_cpu = next_cpu;
2031 hctx->next_cpu_batch = 1;
2032 return WORK_CPU_UNBOUND;
2035 hctx->next_cpu = next_cpu;
2040 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2041 * @hctx: Pointer to the hardware queue to run.
2042 * @async: If we want to run the queue asynchronously.
2043 * @msecs: Milliseconds of delay to wait before running the queue.
2045 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2046 * with a delay of @msecs.
2048 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2049 unsigned long msecs)
2051 if (unlikely(blk_mq_hctx_stopped(hctx)))
2054 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2055 int cpu = get_cpu();
2056 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
2057 __blk_mq_run_hw_queue(hctx);
2065 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2066 msecs_to_jiffies(msecs));
2070 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2071 * @hctx: Pointer to the hardware queue to run.
2072 * @msecs: Milliseconds of delay to wait before running the queue.
2074 * Run a hardware queue asynchronously with a delay of @msecs.
2076 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2078 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2080 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2083 * blk_mq_run_hw_queue - Start to run a hardware queue.
2084 * @hctx: Pointer to the hardware queue to run.
2085 * @async: If we want to run the queue asynchronously.
2087 * Check if the request queue is not in a quiesced state and if there are
2088 * pending requests to be sent. If this is true, run the queue to send requests
2091 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2096 * When queue is quiesced, we may be switching io scheduler, or
2097 * updating nr_hw_queues, or other things, and we can't run queue
2098 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2100 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2103 __blk_mq_run_dispatch_ops(hctx->queue, false,
2104 need_run = !blk_queue_quiesced(hctx->queue) &&
2105 blk_mq_hctx_has_pending(hctx));
2108 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2110 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2113 * Is the request queue handled by an IO scheduler that does not respect
2114 * hardware queues when dispatching?
2116 static bool blk_mq_has_sqsched(struct request_queue *q)
2118 struct elevator_queue *e = q->elevator;
2120 if (e && e->type->ops.dispatch_request &&
2121 !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE))
2127 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2130 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2132 struct blk_mq_hw_ctx *hctx;
2135 * If the IO scheduler does not respect hardware queues when
2136 * dispatching, we just don't bother with multiple HW queues and
2137 * dispatch from hctx for the current CPU since running multiple queues
2138 * just causes lock contention inside the scheduler and pointless cache
2141 hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT,
2142 raw_smp_processor_id());
2143 if (!blk_mq_hctx_stopped(hctx))
2149 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2150 * @q: Pointer to the request queue to run.
2151 * @async: If we want to run the queue asynchronously.
2153 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2155 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2159 if (blk_mq_has_sqsched(q))
2160 sq_hctx = blk_mq_get_sq_hctx(q);
2161 queue_for_each_hw_ctx(q, hctx, i) {
2162 if (blk_mq_hctx_stopped(hctx))
2165 * Dispatch from this hctx either if there's no hctx preferred
2166 * by IO scheduler or if it has requests that bypass the
2169 if (!sq_hctx || sq_hctx == hctx ||
2170 !list_empty_careful(&hctx->dispatch))
2171 blk_mq_run_hw_queue(hctx, async);
2174 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2177 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2178 * @q: Pointer to the request queue to run.
2179 * @msecs: Milliseconds of delay to wait before running the queues.
2181 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2183 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2187 if (blk_mq_has_sqsched(q))
2188 sq_hctx = blk_mq_get_sq_hctx(q);
2189 queue_for_each_hw_ctx(q, hctx, i) {
2190 if (blk_mq_hctx_stopped(hctx))
2193 * If there is already a run_work pending, leave the
2194 * pending delay untouched. Otherwise, a hctx can stall
2195 * if another hctx is re-delaying the other's work
2196 * before the work executes.
2198 if (delayed_work_pending(&hctx->run_work))
2201 * Dispatch from this hctx either if there's no hctx preferred
2202 * by IO scheduler or if it has requests that bypass the
2205 if (!sq_hctx || sq_hctx == hctx ||
2206 !list_empty_careful(&hctx->dispatch))
2207 blk_mq_delay_run_hw_queue(hctx, msecs);
2210 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2213 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
2214 * @q: request queue.
2216 * The caller is responsible for serializing this function against
2217 * blk_mq_{start,stop}_hw_queue().
2219 bool blk_mq_queue_stopped(struct request_queue *q)
2221 struct blk_mq_hw_ctx *hctx;
2224 queue_for_each_hw_ctx(q, hctx, i)
2225 if (blk_mq_hctx_stopped(hctx))
2230 EXPORT_SYMBOL(blk_mq_queue_stopped);
2233 * This function is often used for pausing .queue_rq() by driver when
2234 * there isn't enough resource or some conditions aren't satisfied, and
2235 * BLK_STS_RESOURCE is usually returned.
2237 * We do not guarantee that dispatch can be drained or blocked
2238 * after blk_mq_stop_hw_queue() returns. Please use
2239 * blk_mq_quiesce_queue() for that requirement.
2241 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2243 cancel_delayed_work(&hctx->run_work);
2245 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2247 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2250 * This function is often used for pausing .queue_rq() by driver when
2251 * there isn't enough resource or some conditions aren't satisfied, and
2252 * BLK_STS_RESOURCE is usually returned.
2254 * We do not guarantee that dispatch can be drained or blocked
2255 * after blk_mq_stop_hw_queues() returns. Please use
2256 * blk_mq_quiesce_queue() for that requirement.
2258 void blk_mq_stop_hw_queues(struct request_queue *q)
2260 struct blk_mq_hw_ctx *hctx;
2263 queue_for_each_hw_ctx(q, hctx, i)
2264 blk_mq_stop_hw_queue(hctx);
2266 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2268 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2270 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2272 blk_mq_run_hw_queue(hctx, false);
2274 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2276 void blk_mq_start_hw_queues(struct request_queue *q)
2278 struct blk_mq_hw_ctx *hctx;
2281 queue_for_each_hw_ctx(q, hctx, i)
2282 blk_mq_start_hw_queue(hctx);
2284 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2286 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2288 if (!blk_mq_hctx_stopped(hctx))
2291 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2292 blk_mq_run_hw_queue(hctx, async);
2294 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2296 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2298 struct blk_mq_hw_ctx *hctx;
2301 queue_for_each_hw_ctx(q, hctx, i)
2302 blk_mq_start_stopped_hw_queue(hctx, async);
2304 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2306 static void blk_mq_run_work_fn(struct work_struct *work)
2308 struct blk_mq_hw_ctx *hctx;
2310 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2313 * If we are stopped, don't run the queue.
2315 if (blk_mq_hctx_stopped(hctx))
2318 __blk_mq_run_hw_queue(hctx);
2321 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2325 struct blk_mq_ctx *ctx = rq->mq_ctx;
2326 enum hctx_type type = hctx->type;
2328 lockdep_assert_held(&ctx->lock);
2330 trace_block_rq_insert(rq);
2333 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2335 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2338 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2341 struct blk_mq_ctx *ctx = rq->mq_ctx;
2343 lockdep_assert_held(&ctx->lock);
2345 __blk_mq_insert_req_list(hctx, rq, at_head);
2346 blk_mq_hctx_mark_pending(hctx, ctx);
2350 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2351 * @rq: Pointer to request to be inserted.
2352 * @at_head: true if the request should be inserted at the head of the list.
2353 * @run_queue: If we should run the hardware queue after inserting the request.
2355 * Should only be used carefully, when the caller knows we want to
2356 * bypass a potential IO scheduler on the target device.
2358 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2361 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2363 spin_lock(&hctx->lock);
2365 list_add(&rq->queuelist, &hctx->dispatch);
2367 list_add_tail(&rq->queuelist, &hctx->dispatch);
2368 spin_unlock(&hctx->lock);
2371 blk_mq_run_hw_queue(hctx, false);
2374 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2375 struct list_head *list)
2379 enum hctx_type type = hctx->type;
2382 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2385 list_for_each_entry(rq, list, queuelist) {
2386 BUG_ON(rq->mq_ctx != ctx);
2387 trace_block_rq_insert(rq);
2390 spin_lock(&ctx->lock);
2391 list_splice_tail_init(list, &ctx->rq_lists[type]);
2392 blk_mq_hctx_mark_pending(hctx, ctx);
2393 spin_unlock(&ctx->lock);
2396 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2399 if (hctx->queue->mq_ops->commit_rqs) {
2400 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2401 hctx->queue->mq_ops->commit_rqs(hctx);
2406 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2407 unsigned int nr_segs)
2411 if (bio->bi_opf & REQ_RAHEAD)
2412 rq->cmd_flags |= REQ_FAILFAST_MASK;
2414 rq->__sector = bio->bi_iter.bi_sector;
2415 blk_rq_bio_prep(rq, bio, nr_segs);
2417 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2418 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2421 blk_account_io_start(rq);
2424 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2425 struct request *rq, bool last)
2427 struct request_queue *q = rq->q;
2428 struct blk_mq_queue_data bd = {
2435 * For OK queue, we are done. For error, caller may kill it.
2436 * Any other error (busy), just add it to our list as we
2437 * previously would have done.
2439 ret = q->mq_ops->queue_rq(hctx, &bd);
2442 blk_mq_update_dispatch_busy(hctx, false);
2444 case BLK_STS_RESOURCE:
2445 case BLK_STS_DEV_RESOURCE:
2446 blk_mq_update_dispatch_busy(hctx, true);
2447 __blk_mq_requeue_request(rq);
2450 blk_mq_update_dispatch_busy(hctx, false);
2457 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2459 bool bypass_insert, bool last)
2461 struct request_queue *q = rq->q;
2462 bool run_queue = true;
2466 * RCU or SRCU read lock is needed before checking quiesced flag.
2468 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2469 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2470 * and avoid driver to try to dispatch again.
2472 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2474 bypass_insert = false;
2478 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2481 budget_token = blk_mq_get_dispatch_budget(q);
2482 if (budget_token < 0)
2485 blk_mq_set_rq_budget_token(rq, budget_token);
2487 if (!blk_mq_get_driver_tag(rq)) {
2488 blk_mq_put_dispatch_budget(q, budget_token);
2492 return __blk_mq_issue_directly(hctx, rq, last);
2495 return BLK_STS_RESOURCE;
2497 blk_mq_sched_insert_request(rq, false, run_queue, false);
2503 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2504 * @hctx: Pointer of the associated hardware queue.
2505 * @rq: Pointer to request to be sent.
2507 * If the device has enough resources to accept a new request now, send the
2508 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2509 * we can try send it another time in the future. Requests inserted at this
2510 * queue have higher priority.
2512 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2516 __blk_mq_try_issue_directly(hctx, rq, false, true);
2518 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2519 blk_mq_request_bypass_insert(rq, false, true);
2520 else if (ret != BLK_STS_OK)
2521 blk_mq_end_request(rq, ret);
2524 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2526 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2529 static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2531 struct blk_mq_hw_ctx *hctx = NULL;
2536 while ((rq = rq_list_pop(&plug->mq_list))) {
2537 bool last = rq_list_empty(plug->mq_list);
2540 if (hctx != rq->mq_hctx) {
2542 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2546 ret = blk_mq_request_issue_directly(rq, last);
2551 case BLK_STS_RESOURCE:
2552 case BLK_STS_DEV_RESOURCE:
2553 blk_mq_request_bypass_insert(rq, false, last);
2554 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2557 blk_mq_end_request(rq, ret);
2564 * If we didn't flush the entire list, we could have told the driver
2565 * there was more coming, but that turned out to be a lie.
2568 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2571 static void __blk_mq_flush_plug_list(struct request_queue *q,
2572 struct blk_plug *plug)
2574 if (blk_queue_quiesced(q))
2576 q->mq_ops->queue_rqs(&plug->mq_list);
2579 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2581 struct blk_mq_hw_ctx *this_hctx = NULL;
2582 struct blk_mq_ctx *this_ctx = NULL;
2583 struct request *requeue_list = NULL;
2584 unsigned int depth = 0;
2588 struct request *rq = rq_list_pop(&plug->mq_list);
2591 this_hctx = rq->mq_hctx;
2592 this_ctx = rq->mq_ctx;
2593 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2594 rq_list_add(&requeue_list, rq);
2597 list_add_tail(&rq->queuelist, &list);
2599 } while (!rq_list_empty(plug->mq_list));
2601 plug->mq_list = requeue_list;
2602 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2603 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched);
2606 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2610 if (rq_list_empty(plug->mq_list))
2614 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2615 struct request_queue *q;
2617 rq = rq_list_peek(&plug->mq_list);
2621 * Peek first request and see if we have a ->queue_rqs() hook.
2622 * If we do, we can dispatch the whole plug list in one go. We
2623 * already know at this point that all requests belong to the
2624 * same queue, caller must ensure that's the case.
2626 * Since we pass off the full list to the driver at this point,
2627 * we do not increment the active request count for the queue.
2628 * Bypass shared tags for now because of that.
2630 if (q->mq_ops->queue_rqs &&
2631 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2632 blk_mq_run_dispatch_ops(q,
2633 __blk_mq_flush_plug_list(q, plug));
2634 if (rq_list_empty(plug->mq_list))
2638 blk_mq_run_dispatch_ops(q,
2639 blk_mq_plug_issue_direct(plug, false));
2640 if (rq_list_empty(plug->mq_list))
2645 blk_mq_dispatch_plug_list(plug, from_schedule);
2646 } while (!rq_list_empty(plug->mq_list));
2649 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2650 struct list_head *list)
2655 while (!list_empty(list)) {
2657 struct request *rq = list_first_entry(list, struct request,
2660 list_del_init(&rq->queuelist);
2661 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2662 if (ret != BLK_STS_OK) {
2663 if (ret == BLK_STS_RESOURCE ||
2664 ret == BLK_STS_DEV_RESOURCE) {
2665 blk_mq_request_bypass_insert(rq, false,
2669 blk_mq_end_request(rq, ret);
2676 * If we didn't flush the entire list, we could have told
2677 * the driver there was more coming, but that turned out to
2680 if ((!list_empty(list) || errors) &&
2681 hctx->queue->mq_ops->commit_rqs && queued)
2682 hctx->queue->mq_ops->commit_rqs(hctx);
2686 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
2687 * queues. This is important for md arrays to benefit from merging
2690 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
2692 if (plug->multiple_queues)
2693 return BLK_MAX_REQUEST_COUNT * 2;
2694 return BLK_MAX_REQUEST_COUNT;
2697 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2699 struct request *last = rq_list_peek(&plug->mq_list);
2701 if (!plug->rq_count) {
2702 trace_block_plug(rq->q);
2703 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
2704 (!blk_queue_nomerges(rq->q) &&
2705 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2706 blk_mq_flush_plug_list(plug, false);
2707 trace_block_plug(rq->q);
2710 if (!plug->multiple_queues && last && last->q != rq->q)
2711 plug->multiple_queues = true;
2712 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
2713 plug->has_elevator = true;
2715 rq_list_add(&plug->mq_list, rq);
2719 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2720 struct bio *bio, unsigned int nr_segs)
2722 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2723 if (blk_attempt_plug_merge(q, bio, nr_segs))
2725 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2731 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2732 struct blk_plug *plug,
2736 struct blk_mq_alloc_data data = {
2739 .cmd_flags = bio->bi_opf,
2743 if (unlikely(bio_queue_enter(bio)))
2746 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2749 rq_qos_throttle(q, bio);
2752 data.nr_tags = plug->nr_ios;
2754 data.cached_rq = &plug->cached_rq;
2757 rq = __blk_mq_alloc_requests(&data);
2760 rq_qos_cleanup(q, bio);
2761 if (bio->bi_opf & REQ_NOWAIT)
2762 bio_wouldblock_error(bio);
2768 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2769 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2775 rq = rq_list_peek(&plug->cached_rq);
2776 if (!rq || rq->q != q)
2779 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2784 rq_qos_throttle(q, *bio);
2786 if (blk_mq_get_hctx_type((*bio)->bi_opf) != rq->mq_hctx->type)
2788 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2791 rq->cmd_flags = (*bio)->bi_opf;
2792 plug->cached_rq = rq_list_next(rq);
2793 INIT_LIST_HEAD(&rq->queuelist);
2798 * blk_mq_submit_bio - Create and send a request to block device.
2799 * @bio: Bio pointer.
2801 * Builds up a request structure from @q and @bio and send to the device. The
2802 * request may not be queued directly to hardware if:
2803 * * This request can be merged with another one
2804 * * We want to place request at plug queue for possible future merging
2805 * * There is an IO scheduler active at this queue
2807 * It will not queue the request if there is an error with the bio, or at the
2810 void blk_mq_submit_bio(struct bio *bio)
2812 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2813 struct blk_plug *plug = blk_mq_plug(q, bio);
2814 const int is_sync = op_is_sync(bio->bi_opf);
2816 unsigned int nr_segs = 1;
2819 blk_queue_bounce(q, &bio);
2820 if (blk_may_split(q, bio))
2821 __blk_queue_split(q, &bio, &nr_segs);
2823 if (!bio_integrity_prep(bio))
2826 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2830 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2835 trace_block_getrq(bio);
2837 rq_qos_track(q, rq, bio);
2839 blk_mq_bio_to_request(rq, bio, nr_segs);
2841 ret = blk_crypto_init_request(rq);
2842 if (ret != BLK_STS_OK) {
2843 bio->bi_status = ret;
2845 blk_mq_free_request(rq);
2849 if (op_is_flush(bio->bi_opf)) {
2850 blk_insert_flush(rq);
2855 blk_add_rq_to_plug(plug, rq);
2856 else if ((rq->rq_flags & RQF_ELV) ||
2857 (rq->mq_hctx->dispatch_busy &&
2858 (q->nr_hw_queues == 1 || !is_sync)))
2859 blk_mq_sched_insert_request(rq, false, true, true);
2861 blk_mq_run_dispatch_ops(rq->q,
2862 blk_mq_try_issue_directly(rq->mq_hctx, rq));
2865 #ifdef CONFIG_BLK_MQ_STACKING
2867 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
2868 * @rq: the request being queued
2870 blk_status_t blk_insert_cloned_request(struct request *rq)
2872 struct request_queue *q = rq->q;
2873 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
2876 if (blk_rq_sectors(rq) > max_sectors) {
2878 * SCSI device does not have a good way to return if
2879 * Write Same/Zero is actually supported. If a device rejects
2880 * a non-read/write command (discard, write same,etc.) the
2881 * low-level device driver will set the relevant queue limit to
2882 * 0 to prevent blk-lib from issuing more of the offending
2883 * operations. Commands queued prior to the queue limit being
2884 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
2885 * errors being propagated to upper layers.
2887 if (max_sectors == 0)
2888 return BLK_STS_NOTSUPP;
2890 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
2891 __func__, blk_rq_sectors(rq), max_sectors);
2892 return BLK_STS_IOERR;
2896 * The queue settings related to segment counting may differ from the
2899 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
2900 if (rq->nr_phys_segments > queue_max_segments(q)) {
2901 printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n",
2902 __func__, rq->nr_phys_segments, queue_max_segments(q));
2903 return BLK_STS_IOERR;
2906 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
2907 return BLK_STS_IOERR;
2909 if (blk_crypto_insert_cloned_request(rq))
2910 return BLK_STS_IOERR;
2912 blk_account_io_start(rq);
2915 * Since we have a scheduler attached on the top device,
2916 * bypass a potential scheduler on the bottom device for
2919 blk_mq_run_dispatch_ops(q,
2920 ret = blk_mq_request_issue_directly(rq, true));
2922 blk_account_io_done(rq, ktime_get_ns());
2925 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
2928 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
2929 * @rq: the clone request to be cleaned up
2932 * Free all bios in @rq for a cloned request.
2934 void blk_rq_unprep_clone(struct request *rq)
2938 while ((bio = rq->bio) != NULL) {
2939 rq->bio = bio->bi_next;
2944 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
2947 * blk_rq_prep_clone - Helper function to setup clone request
2948 * @rq: the request to be setup
2949 * @rq_src: original request to be cloned
2950 * @bs: bio_set that bios for clone are allocated from
2951 * @gfp_mask: memory allocation mask for bio
2952 * @bio_ctr: setup function to be called for each clone bio.
2953 * Returns %0 for success, non %0 for failure.
2954 * @data: private data to be passed to @bio_ctr
2957 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
2958 * Also, pages which the original bios are pointing to are not copied
2959 * and the cloned bios just point same pages.
2960 * So cloned bios must be completed before original bios, which means
2961 * the caller must complete @rq before @rq_src.
2963 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
2964 struct bio_set *bs, gfp_t gfp_mask,
2965 int (*bio_ctr)(struct bio *, struct bio *, void *),
2968 struct bio *bio, *bio_src;
2973 __rq_for_each_bio(bio_src, rq_src) {
2974 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
2979 if (bio_ctr && bio_ctr(bio, bio_src, data))
2983 rq->biotail->bi_next = bio;
2986 rq->bio = rq->biotail = bio;
2991 /* Copy attributes of the original request to the clone request. */
2992 rq->__sector = blk_rq_pos(rq_src);
2993 rq->__data_len = blk_rq_bytes(rq_src);
2994 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
2995 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
2996 rq->special_vec = rq_src->special_vec;
2998 rq->nr_phys_segments = rq_src->nr_phys_segments;
2999 rq->ioprio = rq_src->ioprio;
3001 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3009 blk_rq_unprep_clone(rq);
3013 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3014 #endif /* CONFIG_BLK_MQ_STACKING */
3017 * Steal bios from a request and add them to a bio list.
3018 * The request must not have been partially completed before.
3020 void blk_steal_bios(struct bio_list *list, struct request *rq)
3024 list->tail->bi_next = rq->bio;
3026 list->head = rq->bio;
3027 list->tail = rq->biotail;
3035 EXPORT_SYMBOL_GPL(blk_steal_bios);
3037 static size_t order_to_size(unsigned int order)
3039 return (size_t)PAGE_SIZE << order;
3042 /* called before freeing request pool in @tags */
3043 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3044 struct blk_mq_tags *tags)
3047 unsigned long flags;
3049 /* There is no need to clear a driver tags own mapping */
3050 if (drv_tags == tags)
3053 list_for_each_entry(page, &tags->page_list, lru) {
3054 unsigned long start = (unsigned long)page_address(page);
3055 unsigned long end = start + order_to_size(page->private);
3058 for (i = 0; i < drv_tags->nr_tags; i++) {
3059 struct request *rq = drv_tags->rqs[i];
3060 unsigned long rq_addr = (unsigned long)rq;
3062 if (rq_addr >= start && rq_addr < end) {
3063 WARN_ON_ONCE(req_ref_read(rq) != 0);
3064 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3070 * Wait until all pending iteration is done.
3072 * Request reference is cleared and it is guaranteed to be observed
3073 * after the ->lock is released.
3075 spin_lock_irqsave(&drv_tags->lock, flags);
3076 spin_unlock_irqrestore(&drv_tags->lock, flags);
3079 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3080 unsigned int hctx_idx)
3082 struct blk_mq_tags *drv_tags;
3085 if (list_empty(&tags->page_list))
3088 if (blk_mq_is_shared_tags(set->flags))
3089 drv_tags = set->shared_tags;
3091 drv_tags = set->tags[hctx_idx];
3093 if (tags->static_rqs && set->ops->exit_request) {
3096 for (i = 0; i < tags->nr_tags; i++) {
3097 struct request *rq = tags->static_rqs[i];
3101 set->ops->exit_request(set, rq, hctx_idx);
3102 tags->static_rqs[i] = NULL;
3106 blk_mq_clear_rq_mapping(drv_tags, tags);
3108 while (!list_empty(&tags->page_list)) {
3109 page = list_first_entry(&tags->page_list, struct page, lru);
3110 list_del_init(&page->lru);
3112 * Remove kmemleak object previously allocated in
3113 * blk_mq_alloc_rqs().
3115 kmemleak_free(page_address(page));
3116 __free_pages(page, page->private);
3120 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3124 kfree(tags->static_rqs);
3125 tags->static_rqs = NULL;
3127 blk_mq_free_tags(tags);
3130 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3131 unsigned int hctx_idx)
3135 for (i = 0; i < set->nr_maps; i++) {
3136 unsigned int start = set->map[i].queue_offset;
3137 unsigned int end = start + set->map[i].nr_queues;
3139 if (hctx_idx >= start && hctx_idx < end)
3143 if (i >= set->nr_maps)
3144 i = HCTX_TYPE_DEFAULT;
3149 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3150 unsigned int hctx_idx)
3152 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3154 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3157 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3158 unsigned int hctx_idx,
3159 unsigned int nr_tags,
3160 unsigned int reserved_tags)
3162 int node = blk_mq_get_hctx_node(set, hctx_idx);
3163 struct blk_mq_tags *tags;
3165 if (node == NUMA_NO_NODE)
3166 node = set->numa_node;
3168 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3169 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3173 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3174 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3177 blk_mq_free_tags(tags);
3181 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3182 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3184 if (!tags->static_rqs) {
3186 blk_mq_free_tags(tags);
3193 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3194 unsigned int hctx_idx, int node)
3198 if (set->ops->init_request) {
3199 ret = set->ops->init_request(set, rq, hctx_idx, node);
3204 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3208 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3209 struct blk_mq_tags *tags,
3210 unsigned int hctx_idx, unsigned int depth)
3212 unsigned int i, j, entries_per_page, max_order = 4;
3213 int node = blk_mq_get_hctx_node(set, hctx_idx);
3214 size_t rq_size, left;
3216 if (node == NUMA_NO_NODE)
3217 node = set->numa_node;
3219 INIT_LIST_HEAD(&tags->page_list);
3222 * rq_size is the size of the request plus driver payload, rounded
3223 * to the cacheline size
3225 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3227 left = rq_size * depth;
3229 for (i = 0; i < depth; ) {
3230 int this_order = max_order;
3235 while (this_order && left < order_to_size(this_order - 1))
3239 page = alloc_pages_node(node,
3240 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3246 if (order_to_size(this_order) < rq_size)
3253 page->private = this_order;
3254 list_add_tail(&page->lru, &tags->page_list);
3256 p = page_address(page);
3258 * Allow kmemleak to scan these pages as they contain pointers
3259 * to additional allocations like via ops->init_request().
3261 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3262 entries_per_page = order_to_size(this_order) / rq_size;
3263 to_do = min(entries_per_page, depth - i);
3264 left -= to_do * rq_size;
3265 for (j = 0; j < to_do; j++) {
3266 struct request *rq = p;
3268 tags->static_rqs[i] = rq;
3269 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3270 tags->static_rqs[i] = NULL;
3281 blk_mq_free_rqs(set, tags, hctx_idx);
3285 struct rq_iter_data {
3286 struct blk_mq_hw_ctx *hctx;
3290 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
3292 struct rq_iter_data *iter_data = data;
3294 if (rq->mq_hctx != iter_data->hctx)
3296 iter_data->has_rq = true;
3300 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3302 struct blk_mq_tags *tags = hctx->sched_tags ?
3303 hctx->sched_tags : hctx->tags;
3304 struct rq_iter_data data = {
3308 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3312 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3313 struct blk_mq_hw_ctx *hctx)
3315 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3317 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3322 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3324 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3325 struct blk_mq_hw_ctx, cpuhp_online);
3327 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3328 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3332 * Prevent new request from being allocated on the current hctx.
3334 * The smp_mb__after_atomic() Pairs with the implied barrier in
3335 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3336 * seen once we return from the tag allocator.
3338 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3339 smp_mb__after_atomic();
3342 * Try to grab a reference to the queue and wait for any outstanding
3343 * requests. If we could not grab a reference the queue has been
3344 * frozen and there are no requests.
3346 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3347 while (blk_mq_hctx_has_requests(hctx))
3349 percpu_ref_put(&hctx->queue->q_usage_counter);
3355 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3357 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3358 struct blk_mq_hw_ctx, cpuhp_online);
3360 if (cpumask_test_cpu(cpu, hctx->cpumask))
3361 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3366 * 'cpu' is going away. splice any existing rq_list entries from this
3367 * software queue to the hw queue dispatch list, and ensure that it
3370 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3372 struct blk_mq_hw_ctx *hctx;
3373 struct blk_mq_ctx *ctx;
3375 enum hctx_type type;
3377 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3378 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3381 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3384 spin_lock(&ctx->lock);
3385 if (!list_empty(&ctx->rq_lists[type])) {
3386 list_splice_init(&ctx->rq_lists[type], &tmp);
3387 blk_mq_hctx_clear_pending(hctx, ctx);
3389 spin_unlock(&ctx->lock);
3391 if (list_empty(&tmp))
3394 spin_lock(&hctx->lock);
3395 list_splice_tail_init(&tmp, &hctx->dispatch);
3396 spin_unlock(&hctx->lock);
3398 blk_mq_run_hw_queue(hctx, true);
3402 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3404 if (!(hctx->flags & BLK_MQ_F_STACKING))
3405 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3406 &hctx->cpuhp_online);
3407 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3412 * Before freeing hw queue, clearing the flush request reference in
3413 * tags->rqs[] for avoiding potential UAF.
3415 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3416 unsigned int queue_depth, struct request *flush_rq)
3419 unsigned long flags;
3421 /* The hw queue may not be mapped yet */
3425 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3427 for (i = 0; i < queue_depth; i++)
3428 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3431 * Wait until all pending iteration is done.
3433 * Request reference is cleared and it is guaranteed to be observed
3434 * after the ->lock is released.
3436 spin_lock_irqsave(&tags->lock, flags);
3437 spin_unlock_irqrestore(&tags->lock, flags);
3440 /* hctx->ctxs will be freed in queue's release handler */
3441 static void blk_mq_exit_hctx(struct request_queue *q,
3442 struct blk_mq_tag_set *set,
3443 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3445 struct request *flush_rq = hctx->fq->flush_rq;
3447 if (blk_mq_hw_queue_mapped(hctx))
3448 blk_mq_tag_idle(hctx);
3450 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3451 set->queue_depth, flush_rq);
3452 if (set->ops->exit_request)
3453 set->ops->exit_request(set, flush_rq, hctx_idx);
3455 if (set->ops->exit_hctx)
3456 set->ops->exit_hctx(hctx, hctx_idx);
3458 blk_mq_remove_cpuhp(hctx);
3460 xa_erase(&q->hctx_table, hctx_idx);
3462 spin_lock(&q->unused_hctx_lock);
3463 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3464 spin_unlock(&q->unused_hctx_lock);
3467 static void blk_mq_exit_hw_queues(struct request_queue *q,
3468 struct blk_mq_tag_set *set, int nr_queue)
3470 struct blk_mq_hw_ctx *hctx;
3473 queue_for_each_hw_ctx(q, hctx, i) {
3476 blk_mq_exit_hctx(q, set, hctx, i);
3480 static int blk_mq_init_hctx(struct request_queue *q,
3481 struct blk_mq_tag_set *set,
3482 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3484 hctx->queue_num = hctx_idx;
3486 if (!(hctx->flags & BLK_MQ_F_STACKING))
3487 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3488 &hctx->cpuhp_online);
3489 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3491 hctx->tags = set->tags[hctx_idx];
3493 if (set->ops->init_hctx &&
3494 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3495 goto unregister_cpu_notifier;
3497 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3501 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3507 if (set->ops->exit_request)
3508 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3510 if (set->ops->exit_hctx)
3511 set->ops->exit_hctx(hctx, hctx_idx);
3512 unregister_cpu_notifier:
3513 blk_mq_remove_cpuhp(hctx);
3517 static struct blk_mq_hw_ctx *
3518 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3521 struct blk_mq_hw_ctx *hctx;
3522 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3524 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3526 goto fail_alloc_hctx;
3528 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3531 atomic_set(&hctx->nr_active, 0);
3532 if (node == NUMA_NO_NODE)
3533 node = set->numa_node;
3534 hctx->numa_node = node;
3536 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3537 spin_lock_init(&hctx->lock);
3538 INIT_LIST_HEAD(&hctx->dispatch);
3540 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3542 INIT_LIST_HEAD(&hctx->hctx_list);
3545 * Allocate space for all possible cpus to avoid allocation at
3548 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3553 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3554 gfp, node, false, false))
3558 spin_lock_init(&hctx->dispatch_wait_lock);
3559 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3560 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3562 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3566 blk_mq_hctx_kobj_init(hctx);
3571 sbitmap_free(&hctx->ctx_map);
3575 free_cpumask_var(hctx->cpumask);
3582 static void blk_mq_init_cpu_queues(struct request_queue *q,
3583 unsigned int nr_hw_queues)
3585 struct blk_mq_tag_set *set = q->tag_set;
3588 for_each_possible_cpu(i) {
3589 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3590 struct blk_mq_hw_ctx *hctx;
3594 spin_lock_init(&__ctx->lock);
3595 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3596 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3601 * Set local node, IFF we have more than one hw queue. If
3602 * not, we remain on the home node of the device
3604 for (j = 0; j < set->nr_maps; j++) {
3605 hctx = blk_mq_map_queue_type(q, j, i);
3606 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3607 hctx->numa_node = cpu_to_node(i);
3612 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3613 unsigned int hctx_idx,
3616 struct blk_mq_tags *tags;
3619 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3623 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3625 blk_mq_free_rq_map(tags);
3632 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3635 if (blk_mq_is_shared_tags(set->flags)) {
3636 set->tags[hctx_idx] = set->shared_tags;
3641 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3644 return set->tags[hctx_idx];
3647 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3648 struct blk_mq_tags *tags,
3649 unsigned int hctx_idx)
3652 blk_mq_free_rqs(set, tags, hctx_idx);
3653 blk_mq_free_rq_map(tags);
3657 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3658 unsigned int hctx_idx)
3660 if (!blk_mq_is_shared_tags(set->flags))
3661 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3663 set->tags[hctx_idx] = NULL;
3666 static void blk_mq_map_swqueue(struct request_queue *q)
3668 unsigned int j, hctx_idx;
3670 struct blk_mq_hw_ctx *hctx;
3671 struct blk_mq_ctx *ctx;
3672 struct blk_mq_tag_set *set = q->tag_set;
3674 queue_for_each_hw_ctx(q, hctx, i) {
3675 cpumask_clear(hctx->cpumask);
3677 hctx->dispatch_from = NULL;
3681 * Map software to hardware queues.
3683 * If the cpu isn't present, the cpu is mapped to first hctx.
3685 for_each_possible_cpu(i) {
3687 ctx = per_cpu_ptr(q->queue_ctx, i);
3688 for (j = 0; j < set->nr_maps; j++) {
3689 if (!set->map[j].nr_queues) {
3690 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3691 HCTX_TYPE_DEFAULT, i);
3694 hctx_idx = set->map[j].mq_map[i];
3695 /* unmapped hw queue can be remapped after CPU topo changed */
3696 if (!set->tags[hctx_idx] &&
3697 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3699 * If tags initialization fail for some hctx,
3700 * that hctx won't be brought online. In this
3701 * case, remap the current ctx to hctx[0] which
3702 * is guaranteed to always have tags allocated
3704 set->map[j].mq_map[i] = 0;
3707 hctx = blk_mq_map_queue_type(q, j, i);
3708 ctx->hctxs[j] = hctx;
3710 * If the CPU is already set in the mask, then we've
3711 * mapped this one already. This can happen if
3712 * devices share queues across queue maps.
3714 if (cpumask_test_cpu(i, hctx->cpumask))
3717 cpumask_set_cpu(i, hctx->cpumask);
3719 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3720 hctx->ctxs[hctx->nr_ctx++] = ctx;
3723 * If the nr_ctx type overflows, we have exceeded the
3724 * amount of sw queues we can support.
3726 BUG_ON(!hctx->nr_ctx);
3729 for (; j < HCTX_MAX_TYPES; j++)
3730 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3731 HCTX_TYPE_DEFAULT, i);
3734 queue_for_each_hw_ctx(q, hctx, i) {
3736 * If no software queues are mapped to this hardware queue,
3737 * disable it and free the request entries.
3739 if (!hctx->nr_ctx) {
3740 /* Never unmap queue 0. We need it as a
3741 * fallback in case of a new remap fails
3745 __blk_mq_free_map_and_rqs(set, i);
3751 hctx->tags = set->tags[i];
3752 WARN_ON(!hctx->tags);
3755 * Set the map size to the number of mapped software queues.
3756 * This is more accurate and more efficient than looping
3757 * over all possibly mapped software queues.
3759 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3762 * Initialize batch roundrobin counts
3764 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3765 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3770 * Caller needs to ensure that we're either frozen/quiesced, or that
3771 * the queue isn't live yet.
3773 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3775 struct blk_mq_hw_ctx *hctx;
3778 queue_for_each_hw_ctx(q, hctx, i) {
3780 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3782 blk_mq_tag_idle(hctx);
3783 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3788 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3791 struct request_queue *q;
3793 lockdep_assert_held(&set->tag_list_lock);
3795 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3796 blk_mq_freeze_queue(q);
3797 queue_set_hctx_shared(q, shared);
3798 blk_mq_unfreeze_queue(q);
3802 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3804 struct blk_mq_tag_set *set = q->tag_set;
3806 mutex_lock(&set->tag_list_lock);
3807 list_del(&q->tag_set_list);
3808 if (list_is_singular(&set->tag_list)) {
3809 /* just transitioned to unshared */
3810 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3811 /* update existing queue */
3812 blk_mq_update_tag_set_shared(set, false);
3814 mutex_unlock(&set->tag_list_lock);
3815 INIT_LIST_HEAD(&q->tag_set_list);
3818 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3819 struct request_queue *q)
3821 mutex_lock(&set->tag_list_lock);
3824 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3826 if (!list_empty(&set->tag_list) &&
3827 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3828 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3829 /* update existing queue */
3830 blk_mq_update_tag_set_shared(set, true);
3832 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3833 queue_set_hctx_shared(q, true);
3834 list_add_tail(&q->tag_set_list, &set->tag_list);
3836 mutex_unlock(&set->tag_list_lock);
3839 /* All allocations will be freed in release handler of q->mq_kobj */
3840 static int blk_mq_alloc_ctxs(struct request_queue *q)
3842 struct blk_mq_ctxs *ctxs;
3845 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3849 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3850 if (!ctxs->queue_ctx)
3853 for_each_possible_cpu(cpu) {
3854 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3858 q->mq_kobj = &ctxs->kobj;
3859 q->queue_ctx = ctxs->queue_ctx;
3868 * It is the actual release handler for mq, but we do it from
3869 * request queue's release handler for avoiding use-after-free
3870 * and headache because q->mq_kobj shouldn't have been introduced,
3871 * but we can't group ctx/kctx kobj without it.
3873 void blk_mq_release(struct request_queue *q)
3875 struct blk_mq_hw_ctx *hctx, *next;
3878 queue_for_each_hw_ctx(q, hctx, i)
3879 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3881 /* all hctx are in .unused_hctx_list now */
3882 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3883 list_del_init(&hctx->hctx_list);
3884 kobject_put(&hctx->kobj);
3887 xa_destroy(&q->hctx_table);
3890 * release .mq_kobj and sw queue's kobject now because
3891 * both share lifetime with request queue.
3893 blk_mq_sysfs_deinit(q);
3896 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3899 struct request_queue *q;
3902 q = blk_alloc_queue(set->numa_node, set->flags & BLK_MQ_F_BLOCKING);
3904 return ERR_PTR(-ENOMEM);
3905 q->queuedata = queuedata;
3906 ret = blk_mq_init_allocated_queue(set, q);
3908 blk_cleanup_queue(q);
3909 return ERR_PTR(ret);
3914 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3916 return blk_mq_init_queue_data(set, NULL);
3918 EXPORT_SYMBOL(blk_mq_init_queue);
3920 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
3921 struct lock_class_key *lkclass)
3923 struct request_queue *q;
3924 struct gendisk *disk;
3926 q = blk_mq_init_queue_data(set, queuedata);
3930 disk = __alloc_disk_node(q, set->numa_node, lkclass);
3932 blk_cleanup_queue(q);
3933 return ERR_PTR(-ENOMEM);
3937 EXPORT_SYMBOL(__blk_mq_alloc_disk);
3939 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3940 struct blk_mq_tag_set *set, struct request_queue *q,
3941 int hctx_idx, int node)
3943 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3945 /* reuse dead hctx first */
3946 spin_lock(&q->unused_hctx_lock);
3947 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3948 if (tmp->numa_node == node) {
3954 list_del_init(&hctx->hctx_list);
3955 spin_unlock(&q->unused_hctx_lock);
3958 hctx = blk_mq_alloc_hctx(q, set, node);
3962 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3968 kobject_put(&hctx->kobj);
3973 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3974 struct request_queue *q)
3976 struct blk_mq_hw_ctx *hctx;
3979 /* protect against switching io scheduler */
3980 mutex_lock(&q->sysfs_lock);
3981 for (i = 0; i < set->nr_hw_queues; i++) {
3983 int node = blk_mq_get_hctx_node(set, i);
3984 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
3987 old_node = old_hctx->numa_node;
3988 blk_mq_exit_hctx(q, set, old_hctx, i);
3991 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
3994 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
3996 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
3997 WARN_ON_ONCE(!hctx);
4001 * Increasing nr_hw_queues fails. Free the newly allocated
4002 * hctxs and keep the previous q->nr_hw_queues.
4004 if (i != set->nr_hw_queues) {
4005 j = q->nr_hw_queues;
4008 q->nr_hw_queues = set->nr_hw_queues;
4011 xa_for_each_start(&q->hctx_table, j, hctx, j)
4012 blk_mq_exit_hctx(q, set, hctx, j);
4013 mutex_unlock(&q->sysfs_lock);
4016 static void blk_mq_update_poll_flag(struct request_queue *q)
4018 struct blk_mq_tag_set *set = q->tag_set;
4020 if (set->nr_maps > HCTX_TYPE_POLL &&
4021 set->map[HCTX_TYPE_POLL].nr_queues)
4022 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4024 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4027 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4028 struct request_queue *q)
4030 WARN_ON_ONCE(blk_queue_has_srcu(q) !=
4031 !!(set->flags & BLK_MQ_F_BLOCKING));
4033 /* mark the queue as mq asap */
4034 q->mq_ops = set->ops;
4036 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
4037 blk_mq_poll_stats_bkt,
4038 BLK_MQ_POLL_STATS_BKTS, q);
4042 if (blk_mq_alloc_ctxs(q))
4045 /* init q->mq_kobj and sw queues' kobjects */
4046 blk_mq_sysfs_init(q);
4048 INIT_LIST_HEAD(&q->unused_hctx_list);
4049 spin_lock_init(&q->unused_hctx_lock);
4051 xa_init(&q->hctx_table);
4053 blk_mq_realloc_hw_ctxs(set, q);
4054 if (!q->nr_hw_queues)
4057 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4058 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4062 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4063 blk_mq_update_poll_flag(q);
4065 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4066 INIT_LIST_HEAD(&q->requeue_list);
4067 spin_lock_init(&q->requeue_lock);
4069 q->nr_requests = set->queue_depth;
4072 * Default to classic polling
4074 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
4076 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4077 blk_mq_add_queue_tag_set(set, q);
4078 blk_mq_map_swqueue(q);
4082 xa_destroy(&q->hctx_table);
4083 q->nr_hw_queues = 0;
4084 blk_mq_sysfs_deinit(q);
4086 blk_stat_free_callback(q->poll_cb);
4092 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4094 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4095 void blk_mq_exit_queue(struct request_queue *q)
4097 struct blk_mq_tag_set *set = q->tag_set;
4099 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4100 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4101 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4102 blk_mq_del_queue_tag_set(q);
4105 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4109 if (blk_mq_is_shared_tags(set->flags)) {
4110 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4113 if (!set->shared_tags)
4117 for (i = 0; i < set->nr_hw_queues; i++) {
4118 if (!__blk_mq_alloc_map_and_rqs(set, i))
4127 __blk_mq_free_map_and_rqs(set, i);
4129 if (blk_mq_is_shared_tags(set->flags)) {
4130 blk_mq_free_map_and_rqs(set, set->shared_tags,
4131 BLK_MQ_NO_HCTX_IDX);
4138 * Allocate the request maps associated with this tag_set. Note that this
4139 * may reduce the depth asked for, if memory is tight. set->queue_depth
4140 * will be updated to reflect the allocated depth.
4142 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4147 depth = set->queue_depth;
4149 err = __blk_mq_alloc_rq_maps(set);
4153 set->queue_depth >>= 1;
4154 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4158 } while (set->queue_depth);
4160 if (!set->queue_depth || err) {
4161 pr_err("blk-mq: failed to allocate request map\n");
4165 if (depth != set->queue_depth)
4166 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4167 depth, set->queue_depth);
4172 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4175 * blk_mq_map_queues() and multiple .map_queues() implementations
4176 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4177 * number of hardware queues.
4179 if (set->nr_maps == 1)
4180 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4182 if (set->ops->map_queues && !is_kdump_kernel()) {
4186 * transport .map_queues is usually done in the following
4189 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4190 * mask = get_cpu_mask(queue)
4191 * for_each_cpu(cpu, mask)
4192 * set->map[x].mq_map[cpu] = queue;
4195 * When we need to remap, the table has to be cleared for
4196 * killing stale mapping since one CPU may not be mapped
4199 for (i = 0; i < set->nr_maps; i++)
4200 blk_mq_clear_mq_map(&set->map[i]);
4202 return set->ops->map_queues(set);
4204 BUG_ON(set->nr_maps > 1);
4205 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4209 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4210 int cur_nr_hw_queues, int new_nr_hw_queues)
4212 struct blk_mq_tags **new_tags;
4214 if (cur_nr_hw_queues >= new_nr_hw_queues)
4217 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4218 GFP_KERNEL, set->numa_node);
4223 memcpy(new_tags, set->tags, cur_nr_hw_queues *
4224 sizeof(*set->tags));
4226 set->tags = new_tags;
4227 set->nr_hw_queues = new_nr_hw_queues;
4232 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
4233 int new_nr_hw_queues)
4235 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
4239 * Alloc a tag set to be associated with one or more request queues.
4240 * May fail with EINVAL for various error conditions. May adjust the
4241 * requested depth down, if it's too large. In that case, the set
4242 * value will be stored in set->queue_depth.
4244 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4248 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4250 if (!set->nr_hw_queues)
4252 if (!set->queue_depth)
4254 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4257 if (!set->ops->queue_rq)
4260 if (!set->ops->get_budget ^ !set->ops->put_budget)
4263 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4264 pr_info("blk-mq: reduced tag depth to %u\n",
4266 set->queue_depth = BLK_MQ_MAX_DEPTH;
4271 else if (set->nr_maps > HCTX_MAX_TYPES)
4275 * If a crashdump is active, then we are potentially in a very
4276 * memory constrained environment. Limit us to 1 queue and
4277 * 64 tags to prevent using too much memory.
4279 if (is_kdump_kernel()) {
4280 set->nr_hw_queues = 1;
4282 set->queue_depth = min(64U, set->queue_depth);
4285 * There is no use for more h/w queues than cpus if we just have
4288 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4289 set->nr_hw_queues = nr_cpu_ids;
4291 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
4295 for (i = 0; i < set->nr_maps; i++) {
4296 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4297 sizeof(set->map[i].mq_map[0]),
4298 GFP_KERNEL, set->numa_node);
4299 if (!set->map[i].mq_map)
4300 goto out_free_mq_map;
4301 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4304 ret = blk_mq_update_queue_map(set);
4306 goto out_free_mq_map;
4308 ret = blk_mq_alloc_set_map_and_rqs(set);
4310 goto out_free_mq_map;
4312 mutex_init(&set->tag_list_lock);
4313 INIT_LIST_HEAD(&set->tag_list);
4318 for (i = 0; i < set->nr_maps; i++) {
4319 kfree(set->map[i].mq_map);
4320 set->map[i].mq_map = NULL;
4326 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4328 /* allocate and initialize a tagset for a simple single-queue device */
4329 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4330 const struct blk_mq_ops *ops, unsigned int queue_depth,
4331 unsigned int set_flags)
4333 memset(set, 0, sizeof(*set));
4335 set->nr_hw_queues = 1;
4337 set->queue_depth = queue_depth;
4338 set->numa_node = NUMA_NO_NODE;
4339 set->flags = set_flags;
4340 return blk_mq_alloc_tag_set(set);
4342 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4344 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4348 for (i = 0; i < set->nr_hw_queues; i++)
4349 __blk_mq_free_map_and_rqs(set, i);
4351 if (blk_mq_is_shared_tags(set->flags)) {
4352 blk_mq_free_map_and_rqs(set, set->shared_tags,
4353 BLK_MQ_NO_HCTX_IDX);
4356 for (j = 0; j < set->nr_maps; j++) {
4357 kfree(set->map[j].mq_map);
4358 set->map[j].mq_map = NULL;
4364 EXPORT_SYMBOL(blk_mq_free_tag_set);
4366 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4368 struct blk_mq_tag_set *set = q->tag_set;
4369 struct blk_mq_hw_ctx *hctx;
4376 if (q->nr_requests == nr)
4379 blk_mq_freeze_queue(q);
4380 blk_mq_quiesce_queue(q);
4383 queue_for_each_hw_ctx(q, hctx, i) {
4387 * If we're using an MQ scheduler, just update the scheduler
4388 * queue depth. This is similar to what the old code would do.
4390 if (hctx->sched_tags) {
4391 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4394 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4399 if (q->elevator && q->elevator->type->ops.depth_updated)
4400 q->elevator->type->ops.depth_updated(hctx);
4403 q->nr_requests = nr;
4404 if (blk_mq_is_shared_tags(set->flags)) {
4406 blk_mq_tag_update_sched_shared_tags(q);
4408 blk_mq_tag_resize_shared_tags(set, nr);
4412 blk_mq_unquiesce_queue(q);
4413 blk_mq_unfreeze_queue(q);
4419 * request_queue and elevator_type pair.
4420 * It is just used by __blk_mq_update_nr_hw_queues to cache
4421 * the elevator_type associated with a request_queue.
4423 struct blk_mq_qe_pair {
4424 struct list_head node;
4425 struct request_queue *q;
4426 struct elevator_type *type;
4430 * Cache the elevator_type in qe pair list and switch the
4431 * io scheduler to 'none'
4433 static bool blk_mq_elv_switch_none(struct list_head *head,
4434 struct request_queue *q)
4436 struct blk_mq_qe_pair *qe;
4441 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4445 INIT_LIST_HEAD(&qe->node);
4447 qe->type = q->elevator->type;
4448 list_add(&qe->node, head);
4450 mutex_lock(&q->sysfs_lock);
4452 * After elevator_switch_mq, the previous elevator_queue will be
4453 * released by elevator_release. The reference of the io scheduler
4454 * module get by elevator_get will also be put. So we need to get
4455 * a reference of the io scheduler module here to prevent it to be
4458 __module_get(qe->type->elevator_owner);
4459 elevator_switch_mq(q, NULL);
4460 mutex_unlock(&q->sysfs_lock);
4465 static void blk_mq_elv_switch_back(struct list_head *head,
4466 struct request_queue *q)
4468 struct blk_mq_qe_pair *qe;
4469 struct elevator_type *t = NULL;
4471 list_for_each_entry(qe, head, node)
4480 list_del(&qe->node);
4483 mutex_lock(&q->sysfs_lock);
4484 elevator_switch_mq(q, t);
4485 mutex_unlock(&q->sysfs_lock);
4488 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4491 struct request_queue *q;
4493 int prev_nr_hw_queues;
4495 lockdep_assert_held(&set->tag_list_lock);
4497 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4498 nr_hw_queues = nr_cpu_ids;
4499 if (nr_hw_queues < 1)
4501 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4504 list_for_each_entry(q, &set->tag_list, tag_set_list)
4505 blk_mq_freeze_queue(q);
4507 * Switch IO scheduler to 'none', cleaning up the data associated
4508 * with the previous scheduler. We will switch back once we are done
4509 * updating the new sw to hw queue mappings.
4511 list_for_each_entry(q, &set->tag_list, tag_set_list)
4512 if (!blk_mq_elv_switch_none(&head, q))
4515 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4516 blk_mq_debugfs_unregister_hctxs(q);
4517 blk_mq_sysfs_unregister(q);
4520 prev_nr_hw_queues = set->nr_hw_queues;
4521 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
4525 set->nr_hw_queues = nr_hw_queues;
4527 blk_mq_update_queue_map(set);
4528 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4529 blk_mq_realloc_hw_ctxs(set, q);
4530 blk_mq_update_poll_flag(q);
4531 if (q->nr_hw_queues != set->nr_hw_queues) {
4532 int i = prev_nr_hw_queues;
4534 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4535 nr_hw_queues, prev_nr_hw_queues);
4536 for (; i < set->nr_hw_queues; i++)
4537 __blk_mq_free_map_and_rqs(set, i);
4539 set->nr_hw_queues = prev_nr_hw_queues;
4540 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4543 blk_mq_map_swqueue(q);
4547 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4548 blk_mq_sysfs_register(q);
4549 blk_mq_debugfs_register_hctxs(q);
4553 list_for_each_entry(q, &set->tag_list, tag_set_list)
4554 blk_mq_elv_switch_back(&head, q);
4556 list_for_each_entry(q, &set->tag_list, tag_set_list)
4557 blk_mq_unfreeze_queue(q);
4560 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4562 mutex_lock(&set->tag_list_lock);
4563 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4564 mutex_unlock(&set->tag_list_lock);
4566 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4568 /* Enable polling stats and return whether they were already enabled. */
4569 static bool blk_poll_stats_enable(struct request_queue *q)
4574 return blk_stats_alloc_enable(q);
4577 static void blk_mq_poll_stats_start(struct request_queue *q)
4580 * We don't arm the callback if polling stats are not enabled or the
4581 * callback is already active.
4583 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
4586 blk_stat_activate_msecs(q->poll_cb, 100);
4589 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4591 struct request_queue *q = cb->data;
4594 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4595 if (cb->stat[bucket].nr_samples)
4596 q->poll_stat[bucket] = cb->stat[bucket];
4600 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4603 unsigned long ret = 0;
4607 * If stats collection isn't on, don't sleep but turn it on for
4610 if (!blk_poll_stats_enable(q))
4614 * As an optimistic guess, use half of the mean service time
4615 * for this type of request. We can (and should) make this smarter.
4616 * For instance, if the completion latencies are tight, we can
4617 * get closer than just half the mean. This is especially
4618 * important on devices where the completion latencies are longer
4619 * than ~10 usec. We do use the stats for the relevant IO size
4620 * if available which does lead to better estimates.
4622 bucket = blk_mq_poll_stats_bkt(rq);
4626 if (q->poll_stat[bucket].nr_samples)
4627 ret = (q->poll_stat[bucket].mean + 1) / 2;
4632 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4634 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4635 struct request *rq = blk_qc_to_rq(hctx, qc);
4636 struct hrtimer_sleeper hs;
4637 enum hrtimer_mode mode;
4642 * If a request has completed on queue that uses an I/O scheduler, we
4643 * won't get back a request from blk_qc_to_rq.
4645 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4649 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4651 * 0: use half of prev avg
4652 * >0: use this specific value
4654 if (q->poll_nsec > 0)
4655 nsecs = q->poll_nsec;
4657 nsecs = blk_mq_poll_nsecs(q, rq);
4662 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4665 * This will be replaced with the stats tracking code, using
4666 * 'avg_completion_time / 2' as the pre-sleep target.
4670 mode = HRTIMER_MODE_REL;
4671 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4672 hrtimer_set_expires(&hs.timer, kt);
4675 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4677 set_current_state(TASK_UNINTERRUPTIBLE);
4678 hrtimer_sleeper_start_expires(&hs, mode);
4681 hrtimer_cancel(&hs.timer);
4682 mode = HRTIMER_MODE_ABS;
4683 } while (hs.task && !signal_pending(current));
4685 __set_current_state(TASK_RUNNING);
4686 destroy_hrtimer_on_stack(&hs.timer);
4689 * If we sleep, have the caller restart the poll loop to reset the
4690 * state. Like for the other success return cases, the caller is
4691 * responsible for checking if the IO completed. If the IO isn't
4692 * complete, we'll get called again and will go straight to the busy
4698 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4699 struct io_comp_batch *iob, unsigned int flags)
4701 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4702 long state = get_current_state();
4706 ret = q->mq_ops->poll(hctx, iob);
4708 __set_current_state(TASK_RUNNING);
4712 if (signal_pending_state(state, current))
4713 __set_current_state(TASK_RUNNING);
4714 if (task_is_running(current))
4717 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4720 } while (!need_resched());
4722 __set_current_state(TASK_RUNNING);
4726 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4729 if (!(flags & BLK_POLL_NOSLEEP) &&
4730 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4731 if (blk_mq_poll_hybrid(q, cookie))
4734 return blk_mq_poll_classic(q, cookie, iob, flags);
4737 unsigned int blk_mq_rq_cpu(struct request *rq)
4739 return rq->mq_ctx->cpu;
4741 EXPORT_SYMBOL(blk_mq_rq_cpu);
4743 void blk_mq_cancel_work_sync(struct request_queue *q)
4745 if (queue_is_mq(q)) {
4746 struct blk_mq_hw_ctx *hctx;
4749 cancel_delayed_work_sync(&q->requeue_work);
4751 queue_for_each_hw_ctx(q, hctx, i)
4752 cancel_delayed_work_sync(&hctx->run_work);
4756 static int __init blk_mq_init(void)
4760 for_each_possible_cpu(i)
4761 init_llist_head(&per_cpu(blk_cpu_done, i));
4762 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4764 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4765 "block/softirq:dead", NULL,
4766 blk_softirq_cpu_dead);
4767 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4768 blk_mq_hctx_notify_dead);
4769 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4770 blk_mq_hctx_notify_online,
4771 blk_mq_hctx_notify_offline);
4774 subsys_initcall(blk_mq_init);