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 !test_bit(GD_DEAD, &req->q->disk->state)) {
799 blk_print_req_error(req, error);
800 trace_block_rq_error(req, error, nr_bytes);
803 blk_account_io_completion(req, nr_bytes);
807 struct bio *bio = req->bio;
808 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
810 if (bio_bytes == bio->bi_iter.bi_size)
811 req->bio = bio->bi_next;
813 /* Completion has already been traced */
814 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
815 req_bio_endio(req, bio, bio_bytes, error);
817 total_bytes += bio_bytes;
818 nr_bytes -= bio_bytes;
829 * Reset counters so that the request stacking driver
830 * can find how many bytes remain in the request
837 req->__data_len -= total_bytes;
839 /* update sector only for requests with clear definition of sector */
840 if (!blk_rq_is_passthrough(req))
841 req->__sector += total_bytes >> 9;
843 /* mixed attributes always follow the first bio */
844 if (req->rq_flags & RQF_MIXED_MERGE) {
845 req->cmd_flags &= ~REQ_FAILFAST_MASK;
846 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
849 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
851 * If total number of sectors is less than the first segment
852 * size, something has gone terribly wrong.
854 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
855 blk_dump_rq_flags(req, "request botched");
856 req->__data_len = blk_rq_cur_bytes(req);
859 /* recalculate the number of segments */
860 req->nr_phys_segments = blk_recalc_rq_segments(req);
865 EXPORT_SYMBOL_GPL(blk_update_request);
867 static void __blk_account_io_done(struct request *req, u64 now)
869 const int sgrp = op_stat_group(req_op(req));
872 update_io_ticks(req->part, jiffies, true);
873 part_stat_inc(req->part, ios[sgrp]);
874 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
878 static inline void blk_account_io_done(struct request *req, u64 now)
881 * Account IO completion. flush_rq isn't accounted as a
882 * normal IO on queueing nor completion. Accounting the
883 * containing request is enough.
885 if (blk_do_io_stat(req) && req->part &&
886 !(req->rq_flags & RQF_FLUSH_SEQ))
887 __blk_account_io_done(req, now);
890 static void __blk_account_io_start(struct request *rq)
893 * All non-passthrough requests are created from a bio with one
894 * exception: when a flush command that is part of a flush sequence
895 * generated by the state machine in blk-flush.c is cloned onto the
896 * lower device by dm-multipath we can get here without a bio.
899 rq->part = rq->bio->bi_bdev;
901 rq->part = rq->q->disk->part0;
904 update_io_ticks(rq->part, jiffies, false);
908 static inline void blk_account_io_start(struct request *req)
910 if (blk_do_io_stat(req))
911 __blk_account_io_start(req);
914 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
916 if (rq->rq_flags & RQF_STATS) {
917 blk_mq_poll_stats_start(rq->q);
918 blk_stat_add(rq, now);
921 blk_mq_sched_completed_request(rq, now);
922 blk_account_io_done(rq, now);
925 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
927 if (blk_mq_need_time_stamp(rq))
928 __blk_mq_end_request_acct(rq, ktime_get_ns());
931 rq_qos_done(rq->q, rq);
932 rq->end_io(rq, error);
934 blk_mq_free_request(rq);
937 EXPORT_SYMBOL(__blk_mq_end_request);
939 void blk_mq_end_request(struct request *rq, blk_status_t error)
941 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
943 __blk_mq_end_request(rq, error);
945 EXPORT_SYMBOL(blk_mq_end_request);
947 #define TAG_COMP_BATCH 32
949 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
950 int *tag_array, int nr_tags)
952 struct request_queue *q = hctx->queue;
955 * All requests should have been marked as RQF_MQ_INFLIGHT, so
956 * update hctx->nr_active in batch
958 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
959 __blk_mq_sub_active_requests(hctx, nr_tags);
961 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
962 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
965 void blk_mq_end_request_batch(struct io_comp_batch *iob)
967 int tags[TAG_COMP_BATCH], nr_tags = 0;
968 struct blk_mq_hw_ctx *cur_hctx = NULL;
973 now = ktime_get_ns();
975 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
977 prefetch(rq->rq_next);
979 blk_complete_request(rq);
981 __blk_mq_end_request_acct(rq, now);
983 rq_qos_done(rq->q, rq);
985 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
986 if (!req_ref_put_and_test(rq))
989 blk_crypto_free_request(rq);
990 blk_pm_mark_last_busy(rq);
992 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
994 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
996 cur_hctx = rq->mq_hctx;
998 tags[nr_tags++] = rq->tag;
1002 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1004 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1006 static void blk_complete_reqs(struct llist_head *list)
1008 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1009 struct request *rq, *next;
1011 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1012 rq->q->mq_ops->complete(rq);
1015 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1017 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1020 static int blk_softirq_cpu_dead(unsigned int cpu)
1022 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1026 static void __blk_mq_complete_request_remote(void *data)
1028 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1031 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1033 int cpu = raw_smp_processor_id();
1035 if (!IS_ENABLED(CONFIG_SMP) ||
1036 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1039 * With force threaded interrupts enabled, raising softirq from an SMP
1040 * function call will always result in waking the ksoftirqd thread.
1041 * This is probably worse than completing the request on a different
1044 if (force_irqthreads())
1047 /* same CPU or cache domain? Complete locally */
1048 if (cpu == rq->mq_ctx->cpu ||
1049 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1050 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1053 /* don't try to IPI to an offline CPU */
1054 return cpu_online(rq->mq_ctx->cpu);
1057 static void blk_mq_complete_send_ipi(struct request *rq)
1059 struct llist_head *list;
1062 cpu = rq->mq_ctx->cpu;
1063 list = &per_cpu(blk_cpu_done, cpu);
1064 if (llist_add(&rq->ipi_list, list)) {
1065 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1066 smp_call_function_single_async(cpu, &rq->csd);
1070 static void blk_mq_raise_softirq(struct request *rq)
1072 struct llist_head *list;
1075 list = this_cpu_ptr(&blk_cpu_done);
1076 if (llist_add(&rq->ipi_list, list))
1077 raise_softirq(BLOCK_SOFTIRQ);
1081 bool blk_mq_complete_request_remote(struct request *rq)
1083 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1086 * For a polled request, always complete locally, it's pointless
1087 * to redirect the completion.
1089 if (rq->cmd_flags & REQ_POLLED)
1092 if (blk_mq_complete_need_ipi(rq)) {
1093 blk_mq_complete_send_ipi(rq);
1097 if (rq->q->nr_hw_queues == 1) {
1098 blk_mq_raise_softirq(rq);
1103 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1106 * blk_mq_complete_request - end I/O on a request
1107 * @rq: the request being processed
1110 * Complete a request by scheduling the ->complete_rq operation.
1112 void blk_mq_complete_request(struct request *rq)
1114 if (!blk_mq_complete_request_remote(rq))
1115 rq->q->mq_ops->complete(rq);
1117 EXPORT_SYMBOL(blk_mq_complete_request);
1120 * blk_mq_start_request - Start processing a request
1121 * @rq: Pointer to request to be started
1123 * Function used by device drivers to notify the block layer that a request
1124 * is going to be processed now, so blk layer can do proper initializations
1125 * such as starting the timeout timer.
1127 void blk_mq_start_request(struct request *rq)
1129 struct request_queue *q = rq->q;
1131 trace_block_rq_issue(rq);
1133 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1134 rq->io_start_time_ns = ktime_get_ns();
1135 rq->stats_sectors = blk_rq_sectors(rq);
1136 rq->rq_flags |= RQF_STATS;
1137 rq_qos_issue(q, rq);
1140 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1143 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1145 #ifdef CONFIG_BLK_DEV_INTEGRITY
1146 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1147 q->integrity.profile->prepare_fn(rq);
1149 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1150 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1152 EXPORT_SYMBOL(blk_mq_start_request);
1155 * blk_end_sync_rq - executes a completion event on a request
1156 * @rq: request to complete
1157 * @error: end I/O status of the request
1159 static void blk_end_sync_rq(struct request *rq, blk_status_t error)
1161 struct completion *waiting = rq->end_io_data;
1163 rq->end_io_data = (void *)(uintptr_t)error;
1166 * complete last, if this is a stack request the process (and thus
1167 * the rq pointer) could be invalid right after this complete()
1173 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1174 * queues. This is important for md arrays to benefit from merging
1177 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1179 if (plug->multiple_queues)
1180 return BLK_MAX_REQUEST_COUNT * 2;
1181 return BLK_MAX_REQUEST_COUNT;
1184 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1186 struct request *last = rq_list_peek(&plug->mq_list);
1188 if (!plug->rq_count) {
1189 trace_block_plug(rq->q);
1190 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1191 (!blk_queue_nomerges(rq->q) &&
1192 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1193 blk_mq_flush_plug_list(plug, false);
1194 trace_block_plug(rq->q);
1197 if (!plug->multiple_queues && last && last->q != rq->q)
1198 plug->multiple_queues = true;
1199 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
1200 plug->has_elevator = true;
1202 rq_list_add(&plug->mq_list, rq);
1206 static void __blk_execute_rq_nowait(struct request *rq, bool at_head,
1207 rq_end_io_fn *done, bool use_plug)
1209 WARN_ON(irqs_disabled());
1210 WARN_ON(!blk_rq_is_passthrough(rq));
1214 blk_account_io_start(rq);
1216 if (use_plug && current->plug) {
1217 blk_add_rq_to_plug(current->plug, rq);
1221 * don't check dying flag for MQ because the request won't
1222 * be reused after dying flag is set
1224 blk_mq_sched_insert_request(rq, at_head, true, false);
1229 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1230 * @rq: request to insert
1231 * @at_head: insert request at head or tail of queue
1232 * @done: I/O completion handler
1235 * Insert a fully prepared request at the back of the I/O scheduler queue
1236 * for execution. Don't wait for completion.
1239 * This function will invoke @done directly if the queue is dead.
1241 void blk_execute_rq_nowait(struct request *rq, bool at_head, rq_end_io_fn *done)
1243 __blk_execute_rq_nowait(rq, at_head, done, true);
1246 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1248 static bool blk_rq_is_poll(struct request *rq)
1252 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1254 if (WARN_ON_ONCE(!rq->bio))
1259 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1262 bio_poll(rq->bio, NULL, 0);
1264 } while (!completion_done(wait));
1268 * blk_execute_rq - insert a request into queue for execution
1269 * @rq: request to insert
1270 * @at_head: insert request at head or tail of queue
1273 * Insert a fully prepared request at the back of the I/O scheduler queue
1274 * for execution and wait for completion.
1275 * Return: The blk_status_t result provided to blk_mq_end_request().
1277 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1279 DECLARE_COMPLETION_ONSTACK(wait);
1280 unsigned long hang_check;
1283 * iopoll requires request to be submitted to driver, so can't
1286 rq->end_io_data = &wait;
1287 __blk_execute_rq_nowait(rq, at_head, blk_end_sync_rq,
1288 !blk_rq_is_poll(rq));
1290 /* Prevent hang_check timer from firing at us during very long I/O */
1291 hang_check = sysctl_hung_task_timeout_secs;
1293 if (blk_rq_is_poll(rq))
1294 blk_rq_poll_completion(rq, &wait);
1295 else if (hang_check)
1296 while (!wait_for_completion_io_timeout(&wait,
1297 hang_check * (HZ/2)))
1300 wait_for_completion_io(&wait);
1302 return (blk_status_t)(uintptr_t)rq->end_io_data;
1304 EXPORT_SYMBOL(blk_execute_rq);
1306 static void __blk_mq_requeue_request(struct request *rq)
1308 struct request_queue *q = rq->q;
1310 blk_mq_put_driver_tag(rq);
1312 trace_block_rq_requeue(rq);
1313 rq_qos_requeue(q, rq);
1315 if (blk_mq_request_started(rq)) {
1316 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1317 rq->rq_flags &= ~RQF_TIMED_OUT;
1321 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1323 __blk_mq_requeue_request(rq);
1325 /* this request will be re-inserted to io scheduler queue */
1326 blk_mq_sched_requeue_request(rq);
1328 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1330 EXPORT_SYMBOL(blk_mq_requeue_request);
1332 static void blk_mq_requeue_work(struct work_struct *work)
1334 struct request_queue *q =
1335 container_of(work, struct request_queue, requeue_work.work);
1337 struct request *rq, *next;
1339 spin_lock_irq(&q->requeue_lock);
1340 list_splice_init(&q->requeue_list, &rq_list);
1341 spin_unlock_irq(&q->requeue_lock);
1343 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1344 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1347 rq->rq_flags &= ~RQF_SOFTBARRIER;
1348 list_del_init(&rq->queuelist);
1350 * If RQF_DONTPREP, rq has contained some driver specific
1351 * data, so insert it to hctx dispatch list to avoid any
1354 if (rq->rq_flags & RQF_DONTPREP)
1355 blk_mq_request_bypass_insert(rq, false, false);
1357 blk_mq_sched_insert_request(rq, true, false, false);
1360 while (!list_empty(&rq_list)) {
1361 rq = list_entry(rq_list.next, struct request, queuelist);
1362 list_del_init(&rq->queuelist);
1363 blk_mq_sched_insert_request(rq, false, false, false);
1366 blk_mq_run_hw_queues(q, false);
1369 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1370 bool kick_requeue_list)
1372 struct request_queue *q = rq->q;
1373 unsigned long flags;
1376 * We abuse this flag that is otherwise used by the I/O scheduler to
1377 * request head insertion from the workqueue.
1379 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1381 spin_lock_irqsave(&q->requeue_lock, flags);
1383 rq->rq_flags |= RQF_SOFTBARRIER;
1384 list_add(&rq->queuelist, &q->requeue_list);
1386 list_add_tail(&rq->queuelist, &q->requeue_list);
1388 spin_unlock_irqrestore(&q->requeue_lock, flags);
1390 if (kick_requeue_list)
1391 blk_mq_kick_requeue_list(q);
1394 void blk_mq_kick_requeue_list(struct request_queue *q)
1396 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1398 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1400 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1401 unsigned long msecs)
1403 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1404 msecs_to_jiffies(msecs));
1406 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1408 static bool blk_mq_rq_inflight(struct request *rq, void *priv,
1412 * If we find a request that isn't idle we know the queue is busy
1413 * as it's checked in the iter.
1414 * Return false to stop the iteration.
1416 if (blk_mq_request_started(rq)) {
1426 bool blk_mq_queue_inflight(struct request_queue *q)
1430 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1433 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1435 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
1437 req->rq_flags |= RQF_TIMED_OUT;
1438 if (req->q->mq_ops->timeout) {
1439 enum blk_eh_timer_return ret;
1441 ret = req->q->mq_ops->timeout(req, reserved);
1442 if (ret == BLK_EH_DONE)
1444 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1450 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
1452 unsigned long deadline;
1454 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1456 if (rq->rq_flags & RQF_TIMED_OUT)
1459 deadline = READ_ONCE(rq->deadline);
1460 if (time_after_eq(jiffies, deadline))
1465 else if (time_after(*next, deadline))
1470 void blk_mq_put_rq_ref(struct request *rq)
1472 if (is_flush_rq(rq))
1474 else if (req_ref_put_and_test(rq))
1475 __blk_mq_free_request(rq);
1478 static bool blk_mq_check_expired(struct request *rq, void *priv, bool reserved)
1480 unsigned long *next = priv;
1483 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1484 * be reallocated underneath the timeout handler's processing, then
1485 * the expire check is reliable. If the request is not expired, then
1486 * it was completed and reallocated as a new request after returning
1487 * from blk_mq_check_expired().
1489 if (blk_mq_req_expired(rq, next))
1490 blk_mq_rq_timed_out(rq, reserved);
1494 static void blk_mq_timeout_work(struct work_struct *work)
1496 struct request_queue *q =
1497 container_of(work, struct request_queue, timeout_work);
1498 unsigned long next = 0;
1499 struct blk_mq_hw_ctx *hctx;
1502 /* A deadlock might occur if a request is stuck requiring a
1503 * timeout at the same time a queue freeze is waiting
1504 * completion, since the timeout code would not be able to
1505 * acquire the queue reference here.
1507 * That's why we don't use blk_queue_enter here; instead, we use
1508 * percpu_ref_tryget directly, because we need to be able to
1509 * obtain a reference even in the short window between the queue
1510 * starting to freeze, by dropping the first reference in
1511 * blk_freeze_queue_start, and the moment the last request is
1512 * consumed, marked by the instant q_usage_counter reaches
1515 if (!percpu_ref_tryget(&q->q_usage_counter))
1518 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
1521 mod_timer(&q->timeout, next);
1524 * Request timeouts are handled as a forward rolling timer. If
1525 * we end up here it means that no requests are pending and
1526 * also that no request has been pending for a while. Mark
1527 * each hctx as idle.
1529 queue_for_each_hw_ctx(q, hctx, i) {
1530 /* the hctx may be unmapped, so check it here */
1531 if (blk_mq_hw_queue_mapped(hctx))
1532 blk_mq_tag_idle(hctx);
1538 struct flush_busy_ctx_data {
1539 struct blk_mq_hw_ctx *hctx;
1540 struct list_head *list;
1543 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1545 struct flush_busy_ctx_data *flush_data = data;
1546 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1547 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1548 enum hctx_type type = hctx->type;
1550 spin_lock(&ctx->lock);
1551 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1552 sbitmap_clear_bit(sb, bitnr);
1553 spin_unlock(&ctx->lock);
1558 * Process software queues that have been marked busy, splicing them
1559 * to the for-dispatch
1561 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1563 struct flush_busy_ctx_data data = {
1568 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1570 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1572 struct dispatch_rq_data {
1573 struct blk_mq_hw_ctx *hctx;
1577 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1580 struct dispatch_rq_data *dispatch_data = data;
1581 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1582 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1583 enum hctx_type type = hctx->type;
1585 spin_lock(&ctx->lock);
1586 if (!list_empty(&ctx->rq_lists[type])) {
1587 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1588 list_del_init(&dispatch_data->rq->queuelist);
1589 if (list_empty(&ctx->rq_lists[type]))
1590 sbitmap_clear_bit(sb, bitnr);
1592 spin_unlock(&ctx->lock);
1594 return !dispatch_data->rq;
1597 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1598 struct blk_mq_ctx *start)
1600 unsigned off = start ? start->index_hw[hctx->type] : 0;
1601 struct dispatch_rq_data data = {
1606 __sbitmap_for_each_set(&hctx->ctx_map, off,
1607 dispatch_rq_from_ctx, &data);
1612 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1614 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1615 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1618 blk_mq_tag_busy(rq->mq_hctx);
1620 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1621 bt = &rq->mq_hctx->tags->breserved_tags;
1624 if (!hctx_may_queue(rq->mq_hctx, bt))
1628 tag = __sbitmap_queue_get(bt);
1629 if (tag == BLK_MQ_NO_TAG)
1632 rq->tag = tag + tag_offset;
1636 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1638 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1641 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1642 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1643 rq->rq_flags |= RQF_MQ_INFLIGHT;
1644 __blk_mq_inc_active_requests(hctx);
1646 hctx->tags->rqs[rq->tag] = rq;
1650 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1651 int flags, void *key)
1653 struct blk_mq_hw_ctx *hctx;
1655 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1657 spin_lock(&hctx->dispatch_wait_lock);
1658 if (!list_empty(&wait->entry)) {
1659 struct sbitmap_queue *sbq;
1661 list_del_init(&wait->entry);
1662 sbq = &hctx->tags->bitmap_tags;
1663 atomic_dec(&sbq->ws_active);
1665 spin_unlock(&hctx->dispatch_wait_lock);
1667 blk_mq_run_hw_queue(hctx, true);
1672 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1673 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1674 * restart. For both cases, take care to check the condition again after
1675 * marking us as waiting.
1677 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1680 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1681 struct wait_queue_head *wq;
1682 wait_queue_entry_t *wait;
1685 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1686 blk_mq_sched_mark_restart_hctx(hctx);
1689 * It's possible that a tag was freed in the window between the
1690 * allocation failure and adding the hardware queue to the wait
1693 * Don't clear RESTART here, someone else could have set it.
1694 * At most this will cost an extra queue run.
1696 return blk_mq_get_driver_tag(rq);
1699 wait = &hctx->dispatch_wait;
1700 if (!list_empty_careful(&wait->entry))
1703 wq = &bt_wait_ptr(sbq, hctx)->wait;
1705 spin_lock_irq(&wq->lock);
1706 spin_lock(&hctx->dispatch_wait_lock);
1707 if (!list_empty(&wait->entry)) {
1708 spin_unlock(&hctx->dispatch_wait_lock);
1709 spin_unlock_irq(&wq->lock);
1713 atomic_inc(&sbq->ws_active);
1714 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1715 __add_wait_queue(wq, wait);
1718 * It's possible that a tag was freed in the window between the
1719 * allocation failure and adding the hardware queue to the wait
1722 ret = blk_mq_get_driver_tag(rq);
1724 spin_unlock(&hctx->dispatch_wait_lock);
1725 spin_unlock_irq(&wq->lock);
1730 * We got a tag, remove ourselves from the wait queue to ensure
1731 * someone else gets the wakeup.
1733 list_del_init(&wait->entry);
1734 atomic_dec(&sbq->ws_active);
1735 spin_unlock(&hctx->dispatch_wait_lock);
1736 spin_unlock_irq(&wq->lock);
1741 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1742 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1744 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1745 * - EWMA is one simple way to compute running average value
1746 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1747 * - take 4 as factor for avoiding to get too small(0) result, and this
1748 * factor doesn't matter because EWMA decreases exponentially
1750 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1754 ewma = hctx->dispatch_busy;
1759 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1761 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1762 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1764 hctx->dispatch_busy = ewma;
1767 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1769 static void blk_mq_handle_dev_resource(struct request *rq,
1770 struct list_head *list)
1772 struct request *next =
1773 list_first_entry_or_null(list, struct request, queuelist);
1776 * If an I/O scheduler has been configured and we got a driver tag for
1777 * the next request already, free it.
1780 blk_mq_put_driver_tag(next);
1782 list_add(&rq->queuelist, list);
1783 __blk_mq_requeue_request(rq);
1786 static void blk_mq_handle_zone_resource(struct request *rq,
1787 struct list_head *zone_list)
1790 * If we end up here it is because we cannot dispatch a request to a
1791 * specific zone due to LLD level zone-write locking or other zone
1792 * related resource not being available. In this case, set the request
1793 * aside in zone_list for retrying it later.
1795 list_add(&rq->queuelist, zone_list);
1796 __blk_mq_requeue_request(rq);
1799 enum prep_dispatch {
1801 PREP_DISPATCH_NO_TAG,
1802 PREP_DISPATCH_NO_BUDGET,
1805 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1808 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1809 int budget_token = -1;
1812 budget_token = blk_mq_get_dispatch_budget(rq->q);
1813 if (budget_token < 0) {
1814 blk_mq_put_driver_tag(rq);
1815 return PREP_DISPATCH_NO_BUDGET;
1817 blk_mq_set_rq_budget_token(rq, budget_token);
1820 if (!blk_mq_get_driver_tag(rq)) {
1822 * The initial allocation attempt failed, so we need to
1823 * rerun the hardware queue when a tag is freed. The
1824 * waitqueue takes care of that. If the queue is run
1825 * before we add this entry back on the dispatch list,
1826 * we'll re-run it below.
1828 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1830 * All budgets not got from this function will be put
1831 * together during handling partial dispatch
1834 blk_mq_put_dispatch_budget(rq->q, budget_token);
1835 return PREP_DISPATCH_NO_TAG;
1839 return PREP_DISPATCH_OK;
1842 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1843 static void blk_mq_release_budgets(struct request_queue *q,
1844 struct list_head *list)
1848 list_for_each_entry(rq, list, queuelist) {
1849 int budget_token = blk_mq_get_rq_budget_token(rq);
1851 if (budget_token >= 0)
1852 blk_mq_put_dispatch_budget(q, budget_token);
1857 * Returns true if we did some work AND can potentially do more.
1859 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1860 unsigned int nr_budgets)
1862 enum prep_dispatch prep;
1863 struct request_queue *q = hctx->queue;
1864 struct request *rq, *nxt;
1866 blk_status_t ret = BLK_STS_OK;
1867 LIST_HEAD(zone_list);
1868 bool needs_resource = false;
1870 if (list_empty(list))
1874 * Now process all the entries, sending them to the driver.
1876 errors = queued = 0;
1878 struct blk_mq_queue_data bd;
1880 rq = list_first_entry(list, struct request, queuelist);
1882 WARN_ON_ONCE(hctx != rq->mq_hctx);
1883 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1884 if (prep != PREP_DISPATCH_OK)
1887 list_del_init(&rq->queuelist);
1892 * Flag last if we have no more requests, or if we have more
1893 * but can't assign a driver tag to it.
1895 if (list_empty(list))
1898 nxt = list_first_entry(list, struct request, queuelist);
1899 bd.last = !blk_mq_get_driver_tag(nxt);
1903 * once the request is queued to lld, no need to cover the
1908 ret = q->mq_ops->queue_rq(hctx, &bd);
1913 case BLK_STS_RESOURCE:
1914 needs_resource = true;
1916 case BLK_STS_DEV_RESOURCE:
1917 blk_mq_handle_dev_resource(rq, list);
1919 case BLK_STS_ZONE_RESOURCE:
1921 * Move the request to zone_list and keep going through
1922 * the dispatch list to find more requests the drive can
1925 blk_mq_handle_zone_resource(rq, &zone_list);
1926 needs_resource = true;
1930 blk_mq_end_request(rq, ret);
1932 } while (!list_empty(list));
1934 if (!list_empty(&zone_list))
1935 list_splice_tail_init(&zone_list, list);
1937 /* If we didn't flush the entire list, we could have told the driver
1938 * there was more coming, but that turned out to be a lie.
1940 if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued)
1941 q->mq_ops->commit_rqs(hctx);
1943 * Any items that need requeuing? Stuff them into hctx->dispatch,
1944 * that is where we will continue on next queue run.
1946 if (!list_empty(list)) {
1948 /* For non-shared tags, the RESTART check will suffice */
1949 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1950 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1953 blk_mq_release_budgets(q, list);
1955 spin_lock(&hctx->lock);
1956 list_splice_tail_init(list, &hctx->dispatch);
1957 spin_unlock(&hctx->lock);
1960 * Order adding requests to hctx->dispatch and checking
1961 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1962 * in blk_mq_sched_restart(). Avoid restart code path to
1963 * miss the new added requests to hctx->dispatch, meantime
1964 * SCHED_RESTART is observed here.
1969 * If SCHED_RESTART was set by the caller of this function and
1970 * it is no longer set that means that it was cleared by another
1971 * thread and hence that a queue rerun is needed.
1973 * If 'no_tag' is set, that means that we failed getting
1974 * a driver tag with an I/O scheduler attached. If our dispatch
1975 * waitqueue is no longer active, ensure that we run the queue
1976 * AFTER adding our entries back to the list.
1978 * If no I/O scheduler has been configured it is possible that
1979 * the hardware queue got stopped and restarted before requests
1980 * were pushed back onto the dispatch list. Rerun the queue to
1981 * avoid starvation. Notes:
1982 * - blk_mq_run_hw_queue() checks whether or not a queue has
1983 * been stopped before rerunning a queue.
1984 * - Some but not all block drivers stop a queue before
1985 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1988 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1989 * bit is set, run queue after a delay to avoid IO stalls
1990 * that could otherwise occur if the queue is idle. We'll do
1991 * similar if we couldn't get budget or couldn't lock a zone
1992 * and SCHED_RESTART is set.
1994 needs_restart = blk_mq_sched_needs_restart(hctx);
1995 if (prep == PREP_DISPATCH_NO_BUDGET)
1996 needs_resource = true;
1997 if (!needs_restart ||
1998 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1999 blk_mq_run_hw_queue(hctx, true);
2000 else if (needs_restart && needs_resource)
2001 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2003 blk_mq_update_dispatch_busy(hctx, true);
2006 blk_mq_update_dispatch_busy(hctx, false);
2008 return (queued + errors) != 0;
2012 * __blk_mq_run_hw_queue - Run a hardware queue.
2013 * @hctx: Pointer to the hardware queue to run.
2015 * Send pending requests to the hardware.
2017 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
2020 * We can't run the queue inline with ints disabled. Ensure that
2021 * we catch bad users of this early.
2023 WARN_ON_ONCE(in_interrupt());
2025 blk_mq_run_dispatch_ops(hctx->queue,
2026 blk_mq_sched_dispatch_requests(hctx));
2029 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2031 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2033 if (cpu >= nr_cpu_ids)
2034 cpu = cpumask_first(hctx->cpumask);
2039 * It'd be great if the workqueue API had a way to pass
2040 * in a mask and had some smarts for more clever placement.
2041 * For now we just round-robin here, switching for every
2042 * BLK_MQ_CPU_WORK_BATCH queued items.
2044 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2047 int next_cpu = hctx->next_cpu;
2049 if (hctx->queue->nr_hw_queues == 1)
2050 return WORK_CPU_UNBOUND;
2052 if (--hctx->next_cpu_batch <= 0) {
2054 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2056 if (next_cpu >= nr_cpu_ids)
2057 next_cpu = blk_mq_first_mapped_cpu(hctx);
2058 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2062 * Do unbound schedule if we can't find a online CPU for this hctx,
2063 * and it should only happen in the path of handling CPU DEAD.
2065 if (!cpu_online(next_cpu)) {
2072 * Make sure to re-select CPU next time once after CPUs
2073 * in hctx->cpumask become online again.
2075 hctx->next_cpu = next_cpu;
2076 hctx->next_cpu_batch = 1;
2077 return WORK_CPU_UNBOUND;
2080 hctx->next_cpu = next_cpu;
2085 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2086 * @hctx: Pointer to the hardware queue to run.
2087 * @async: If we want to run the queue asynchronously.
2088 * @msecs: Milliseconds of delay to wait before running the queue.
2090 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2091 * with a delay of @msecs.
2093 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2094 unsigned long msecs)
2096 if (unlikely(blk_mq_hctx_stopped(hctx)))
2099 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2100 int cpu = get_cpu();
2101 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
2102 __blk_mq_run_hw_queue(hctx);
2110 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2111 msecs_to_jiffies(msecs));
2115 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2116 * @hctx: Pointer to the hardware queue to run.
2117 * @msecs: Milliseconds of delay to wait before running the queue.
2119 * Run a hardware queue asynchronously with a delay of @msecs.
2121 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2123 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2125 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2128 * blk_mq_run_hw_queue - Start to run a hardware queue.
2129 * @hctx: Pointer to the hardware queue to run.
2130 * @async: If we want to run the queue asynchronously.
2132 * Check if the request queue is not in a quiesced state and if there are
2133 * pending requests to be sent. If this is true, run the queue to send requests
2136 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2141 * When queue is quiesced, we may be switching io scheduler, or
2142 * updating nr_hw_queues, or other things, and we can't run queue
2143 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2145 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2148 __blk_mq_run_dispatch_ops(hctx->queue, false,
2149 need_run = !blk_queue_quiesced(hctx->queue) &&
2150 blk_mq_hctx_has_pending(hctx));
2153 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2155 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2158 * Is the request queue handled by an IO scheduler that does not respect
2159 * hardware queues when dispatching?
2161 static bool blk_mq_has_sqsched(struct request_queue *q)
2163 struct elevator_queue *e = q->elevator;
2165 if (e && e->type->ops.dispatch_request &&
2166 !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE))
2172 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2175 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2177 struct blk_mq_hw_ctx *hctx;
2180 * If the IO scheduler does not respect hardware queues when
2181 * dispatching, we just don't bother with multiple HW queues and
2182 * dispatch from hctx for the current CPU since running multiple queues
2183 * just causes lock contention inside the scheduler and pointless cache
2186 hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT,
2187 raw_smp_processor_id());
2188 if (!blk_mq_hctx_stopped(hctx))
2194 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2195 * @q: Pointer to the request queue to run.
2196 * @async: If we want to run the queue asynchronously.
2198 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2200 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2204 if (blk_mq_has_sqsched(q))
2205 sq_hctx = blk_mq_get_sq_hctx(q);
2206 queue_for_each_hw_ctx(q, hctx, i) {
2207 if (blk_mq_hctx_stopped(hctx))
2210 * Dispatch from this hctx either if there's no hctx preferred
2211 * by IO scheduler or if it has requests that bypass the
2214 if (!sq_hctx || sq_hctx == hctx ||
2215 !list_empty_careful(&hctx->dispatch))
2216 blk_mq_run_hw_queue(hctx, async);
2219 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2222 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2223 * @q: Pointer to the request queue to run.
2224 * @msecs: Milliseconds of delay to wait before running the queues.
2226 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2228 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2232 if (blk_mq_has_sqsched(q))
2233 sq_hctx = blk_mq_get_sq_hctx(q);
2234 queue_for_each_hw_ctx(q, hctx, i) {
2235 if (blk_mq_hctx_stopped(hctx))
2238 * If there is already a run_work pending, leave the
2239 * pending delay untouched. Otherwise, a hctx can stall
2240 * if another hctx is re-delaying the other's work
2241 * before the work executes.
2243 if (delayed_work_pending(&hctx->run_work))
2246 * Dispatch from this hctx either if there's no hctx preferred
2247 * by IO scheduler or if it has requests that bypass the
2250 if (!sq_hctx || sq_hctx == hctx ||
2251 !list_empty_careful(&hctx->dispatch))
2252 blk_mq_delay_run_hw_queue(hctx, msecs);
2255 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2258 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
2259 * @q: request queue.
2261 * The caller is responsible for serializing this function against
2262 * blk_mq_{start,stop}_hw_queue().
2264 bool blk_mq_queue_stopped(struct request_queue *q)
2266 struct blk_mq_hw_ctx *hctx;
2269 queue_for_each_hw_ctx(q, hctx, i)
2270 if (blk_mq_hctx_stopped(hctx))
2275 EXPORT_SYMBOL(blk_mq_queue_stopped);
2278 * This function is often used for pausing .queue_rq() by driver when
2279 * there isn't enough resource or some conditions aren't satisfied, and
2280 * BLK_STS_RESOURCE is usually returned.
2282 * We do not guarantee that dispatch can be drained or blocked
2283 * after blk_mq_stop_hw_queue() returns. Please use
2284 * blk_mq_quiesce_queue() for that requirement.
2286 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2288 cancel_delayed_work(&hctx->run_work);
2290 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2292 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2295 * This function is often used for pausing .queue_rq() by driver when
2296 * there isn't enough resource or some conditions aren't satisfied, and
2297 * BLK_STS_RESOURCE is usually returned.
2299 * We do not guarantee that dispatch can be drained or blocked
2300 * after blk_mq_stop_hw_queues() returns. Please use
2301 * blk_mq_quiesce_queue() for that requirement.
2303 void blk_mq_stop_hw_queues(struct request_queue *q)
2305 struct blk_mq_hw_ctx *hctx;
2308 queue_for_each_hw_ctx(q, hctx, i)
2309 blk_mq_stop_hw_queue(hctx);
2311 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2313 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2315 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2317 blk_mq_run_hw_queue(hctx, false);
2319 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2321 void blk_mq_start_hw_queues(struct request_queue *q)
2323 struct blk_mq_hw_ctx *hctx;
2326 queue_for_each_hw_ctx(q, hctx, i)
2327 blk_mq_start_hw_queue(hctx);
2329 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2331 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2333 if (!blk_mq_hctx_stopped(hctx))
2336 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2337 blk_mq_run_hw_queue(hctx, async);
2339 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2341 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2343 struct blk_mq_hw_ctx *hctx;
2346 queue_for_each_hw_ctx(q, hctx, i)
2347 blk_mq_start_stopped_hw_queue(hctx, async);
2349 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2351 static void blk_mq_run_work_fn(struct work_struct *work)
2353 struct blk_mq_hw_ctx *hctx;
2355 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2358 * If we are stopped, don't run the queue.
2360 if (blk_mq_hctx_stopped(hctx))
2363 __blk_mq_run_hw_queue(hctx);
2366 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2370 struct blk_mq_ctx *ctx = rq->mq_ctx;
2371 enum hctx_type type = hctx->type;
2373 lockdep_assert_held(&ctx->lock);
2375 trace_block_rq_insert(rq);
2378 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2380 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2383 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2386 struct blk_mq_ctx *ctx = rq->mq_ctx;
2388 lockdep_assert_held(&ctx->lock);
2390 __blk_mq_insert_req_list(hctx, rq, at_head);
2391 blk_mq_hctx_mark_pending(hctx, ctx);
2395 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2396 * @rq: Pointer to request to be inserted.
2397 * @at_head: true if the request should be inserted at the head of the list.
2398 * @run_queue: If we should run the hardware queue after inserting the request.
2400 * Should only be used carefully, when the caller knows we want to
2401 * bypass a potential IO scheduler on the target device.
2403 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2406 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2408 spin_lock(&hctx->lock);
2410 list_add(&rq->queuelist, &hctx->dispatch);
2412 list_add_tail(&rq->queuelist, &hctx->dispatch);
2413 spin_unlock(&hctx->lock);
2416 blk_mq_run_hw_queue(hctx, false);
2419 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2420 struct list_head *list)
2424 enum hctx_type type = hctx->type;
2427 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2430 list_for_each_entry(rq, list, queuelist) {
2431 BUG_ON(rq->mq_ctx != ctx);
2432 trace_block_rq_insert(rq);
2435 spin_lock(&ctx->lock);
2436 list_splice_tail_init(list, &ctx->rq_lists[type]);
2437 blk_mq_hctx_mark_pending(hctx, ctx);
2438 spin_unlock(&ctx->lock);
2441 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2444 if (hctx->queue->mq_ops->commit_rqs) {
2445 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2446 hctx->queue->mq_ops->commit_rqs(hctx);
2451 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2452 unsigned int nr_segs)
2456 if (bio->bi_opf & REQ_RAHEAD)
2457 rq->cmd_flags |= REQ_FAILFAST_MASK;
2459 rq->__sector = bio->bi_iter.bi_sector;
2460 blk_rq_bio_prep(rq, bio, nr_segs);
2462 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2463 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2466 blk_account_io_start(rq);
2469 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2470 struct request *rq, bool last)
2472 struct request_queue *q = rq->q;
2473 struct blk_mq_queue_data bd = {
2480 * For OK queue, we are done. For error, caller may kill it.
2481 * Any other error (busy), just add it to our list as we
2482 * previously would have done.
2484 ret = q->mq_ops->queue_rq(hctx, &bd);
2487 blk_mq_update_dispatch_busy(hctx, false);
2489 case BLK_STS_RESOURCE:
2490 case BLK_STS_DEV_RESOURCE:
2491 blk_mq_update_dispatch_busy(hctx, true);
2492 __blk_mq_requeue_request(rq);
2495 blk_mq_update_dispatch_busy(hctx, false);
2502 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2504 bool bypass_insert, bool last)
2506 struct request_queue *q = rq->q;
2507 bool run_queue = true;
2511 * RCU or SRCU read lock is needed before checking quiesced flag.
2513 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2514 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2515 * and avoid driver to try to dispatch again.
2517 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2519 bypass_insert = false;
2523 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2526 budget_token = blk_mq_get_dispatch_budget(q);
2527 if (budget_token < 0)
2530 blk_mq_set_rq_budget_token(rq, budget_token);
2532 if (!blk_mq_get_driver_tag(rq)) {
2533 blk_mq_put_dispatch_budget(q, budget_token);
2537 return __blk_mq_issue_directly(hctx, rq, last);
2540 return BLK_STS_RESOURCE;
2542 blk_mq_sched_insert_request(rq, false, run_queue, false);
2548 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2549 * @hctx: Pointer of the associated hardware queue.
2550 * @rq: Pointer to request to be sent.
2552 * If the device has enough resources to accept a new request now, send the
2553 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2554 * we can try send it another time in the future. Requests inserted at this
2555 * queue have higher priority.
2557 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2561 __blk_mq_try_issue_directly(hctx, rq, false, true);
2563 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2564 blk_mq_request_bypass_insert(rq, false, true);
2565 else if (ret != BLK_STS_OK)
2566 blk_mq_end_request(rq, ret);
2569 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2571 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2574 static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2576 struct blk_mq_hw_ctx *hctx = NULL;
2581 while ((rq = rq_list_pop(&plug->mq_list))) {
2582 bool last = rq_list_empty(plug->mq_list);
2585 if (hctx != rq->mq_hctx) {
2587 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2591 ret = blk_mq_request_issue_directly(rq, last);
2596 case BLK_STS_RESOURCE:
2597 case BLK_STS_DEV_RESOURCE:
2598 blk_mq_request_bypass_insert(rq, false, last);
2599 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2602 blk_mq_end_request(rq, ret);
2609 * If we didn't flush the entire list, we could have told the driver
2610 * there was more coming, but that turned out to be a lie.
2613 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2616 static void __blk_mq_flush_plug_list(struct request_queue *q,
2617 struct blk_plug *plug)
2619 if (blk_queue_quiesced(q))
2621 q->mq_ops->queue_rqs(&plug->mq_list);
2624 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2626 struct blk_mq_hw_ctx *this_hctx = NULL;
2627 struct blk_mq_ctx *this_ctx = NULL;
2628 struct request *requeue_list = NULL;
2629 unsigned int depth = 0;
2633 struct request *rq = rq_list_pop(&plug->mq_list);
2636 this_hctx = rq->mq_hctx;
2637 this_ctx = rq->mq_ctx;
2638 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2639 rq_list_add(&requeue_list, rq);
2642 list_add_tail(&rq->queuelist, &list);
2644 } while (!rq_list_empty(plug->mq_list));
2646 plug->mq_list = requeue_list;
2647 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2648 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list, from_sched);
2651 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2655 if (rq_list_empty(plug->mq_list))
2659 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2660 struct request_queue *q;
2662 rq = rq_list_peek(&plug->mq_list);
2666 * Peek first request and see if we have a ->queue_rqs() hook.
2667 * If we do, we can dispatch the whole plug list in one go. We
2668 * already know at this point that all requests belong to the
2669 * same queue, caller must ensure that's the case.
2671 * Since we pass off the full list to the driver at this point,
2672 * we do not increment the active request count for the queue.
2673 * Bypass shared tags for now because of that.
2675 if (q->mq_ops->queue_rqs &&
2676 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2677 blk_mq_run_dispatch_ops(q,
2678 __blk_mq_flush_plug_list(q, plug));
2679 if (rq_list_empty(plug->mq_list))
2683 blk_mq_run_dispatch_ops(q,
2684 blk_mq_plug_issue_direct(plug, false));
2685 if (rq_list_empty(plug->mq_list))
2690 blk_mq_dispatch_plug_list(plug, from_schedule);
2691 } while (!rq_list_empty(plug->mq_list));
2694 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2695 struct list_head *list)
2700 while (!list_empty(list)) {
2702 struct request *rq = list_first_entry(list, struct request,
2705 list_del_init(&rq->queuelist);
2706 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2707 if (ret != BLK_STS_OK) {
2708 if (ret == BLK_STS_RESOURCE ||
2709 ret == BLK_STS_DEV_RESOURCE) {
2710 blk_mq_request_bypass_insert(rq, false,
2714 blk_mq_end_request(rq, ret);
2721 * If we didn't flush the entire list, we could have told
2722 * the driver there was more coming, but that turned out to
2725 if ((!list_empty(list) || errors) &&
2726 hctx->queue->mq_ops->commit_rqs && queued)
2727 hctx->queue->mq_ops->commit_rqs(hctx);
2730 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2731 struct bio *bio, unsigned int nr_segs)
2733 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2734 if (blk_attempt_plug_merge(q, bio, nr_segs))
2736 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2742 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2743 struct blk_plug *plug,
2747 struct blk_mq_alloc_data data = {
2750 .cmd_flags = bio->bi_opf,
2754 if (unlikely(bio_queue_enter(bio)))
2757 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2760 rq_qos_throttle(q, bio);
2763 data.nr_tags = plug->nr_ios;
2765 data.cached_rq = &plug->cached_rq;
2768 rq = __blk_mq_alloc_requests(&data);
2771 rq_qos_cleanup(q, bio);
2772 if (bio->bi_opf & REQ_NOWAIT)
2773 bio_wouldblock_error(bio);
2779 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2780 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2786 rq = rq_list_peek(&plug->cached_rq);
2787 if (!rq || rq->q != q)
2790 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2795 rq_qos_throttle(q, *bio);
2797 if (blk_mq_get_hctx_type((*bio)->bi_opf) != rq->mq_hctx->type)
2799 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2802 rq->cmd_flags = (*bio)->bi_opf;
2803 plug->cached_rq = rq_list_next(rq);
2804 INIT_LIST_HEAD(&rq->queuelist);
2809 * blk_mq_submit_bio - Create and send a request to block device.
2810 * @bio: Bio pointer.
2812 * Builds up a request structure from @q and @bio and send to the device. The
2813 * request may not be queued directly to hardware if:
2814 * * This request can be merged with another one
2815 * * We want to place request at plug queue for possible future merging
2816 * * There is an IO scheduler active at this queue
2818 * It will not queue the request if there is an error with the bio, or at the
2821 void blk_mq_submit_bio(struct bio *bio)
2823 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2824 struct blk_plug *plug = blk_mq_plug(q, bio);
2825 const int is_sync = op_is_sync(bio->bi_opf);
2827 unsigned int nr_segs = 1;
2830 blk_queue_bounce(q, &bio);
2831 if (blk_may_split(q, bio))
2832 __blk_queue_split(q, &bio, &nr_segs);
2834 if (!bio_integrity_prep(bio))
2837 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2841 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2846 trace_block_getrq(bio);
2848 rq_qos_track(q, rq, bio);
2850 blk_mq_bio_to_request(rq, bio, nr_segs);
2852 ret = blk_crypto_init_request(rq);
2853 if (ret != BLK_STS_OK) {
2854 bio->bi_status = ret;
2856 blk_mq_free_request(rq);
2860 if (op_is_flush(bio->bi_opf)) {
2861 blk_insert_flush(rq);
2866 blk_add_rq_to_plug(plug, rq);
2867 else if ((rq->rq_flags & RQF_ELV) ||
2868 (rq->mq_hctx->dispatch_busy &&
2869 (q->nr_hw_queues == 1 || !is_sync)))
2870 blk_mq_sched_insert_request(rq, false, true, true);
2872 blk_mq_run_dispatch_ops(rq->q,
2873 blk_mq_try_issue_directly(rq->mq_hctx, rq));
2876 #ifdef CONFIG_BLK_MQ_STACKING
2878 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
2879 * @rq: the request being queued
2881 blk_status_t blk_insert_cloned_request(struct request *rq)
2883 struct request_queue *q = rq->q;
2884 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
2887 if (blk_rq_sectors(rq) > max_sectors) {
2889 * SCSI device does not have a good way to return if
2890 * Write Same/Zero is actually supported. If a device rejects
2891 * a non-read/write command (discard, write same,etc.) the
2892 * low-level device driver will set the relevant queue limit to
2893 * 0 to prevent blk-lib from issuing more of the offending
2894 * operations. Commands queued prior to the queue limit being
2895 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
2896 * errors being propagated to upper layers.
2898 if (max_sectors == 0)
2899 return BLK_STS_NOTSUPP;
2901 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
2902 __func__, blk_rq_sectors(rq), max_sectors);
2903 return BLK_STS_IOERR;
2907 * The queue settings related to segment counting may differ from the
2910 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
2911 if (rq->nr_phys_segments > queue_max_segments(q)) {
2912 printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n",
2913 __func__, rq->nr_phys_segments, queue_max_segments(q));
2914 return BLK_STS_IOERR;
2917 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
2918 return BLK_STS_IOERR;
2920 if (blk_crypto_insert_cloned_request(rq))
2921 return BLK_STS_IOERR;
2923 blk_account_io_start(rq);
2926 * Since we have a scheduler attached on the top device,
2927 * bypass a potential scheduler on the bottom device for
2930 blk_mq_run_dispatch_ops(q,
2931 ret = blk_mq_request_issue_directly(rq, true));
2933 blk_account_io_done(rq, ktime_get_ns());
2936 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
2939 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
2940 * @rq: the clone request to be cleaned up
2943 * Free all bios in @rq for a cloned request.
2945 void blk_rq_unprep_clone(struct request *rq)
2949 while ((bio = rq->bio) != NULL) {
2950 rq->bio = bio->bi_next;
2955 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
2958 * blk_rq_prep_clone - Helper function to setup clone request
2959 * @rq: the request to be setup
2960 * @rq_src: original request to be cloned
2961 * @bs: bio_set that bios for clone are allocated from
2962 * @gfp_mask: memory allocation mask for bio
2963 * @bio_ctr: setup function to be called for each clone bio.
2964 * Returns %0 for success, non %0 for failure.
2965 * @data: private data to be passed to @bio_ctr
2968 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
2969 * Also, pages which the original bios are pointing to are not copied
2970 * and the cloned bios just point same pages.
2971 * So cloned bios must be completed before original bios, which means
2972 * the caller must complete @rq before @rq_src.
2974 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
2975 struct bio_set *bs, gfp_t gfp_mask,
2976 int (*bio_ctr)(struct bio *, struct bio *, void *),
2979 struct bio *bio, *bio_src;
2984 __rq_for_each_bio(bio_src, rq_src) {
2985 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
2990 if (bio_ctr && bio_ctr(bio, bio_src, data))
2994 rq->biotail->bi_next = bio;
2997 rq->bio = rq->biotail = bio;
3002 /* Copy attributes of the original request to the clone request. */
3003 rq->__sector = blk_rq_pos(rq_src);
3004 rq->__data_len = blk_rq_bytes(rq_src);
3005 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3006 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3007 rq->special_vec = rq_src->special_vec;
3009 rq->nr_phys_segments = rq_src->nr_phys_segments;
3010 rq->ioprio = rq_src->ioprio;
3012 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3020 blk_rq_unprep_clone(rq);
3024 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3025 #endif /* CONFIG_BLK_MQ_STACKING */
3028 * Steal bios from a request and add them to a bio list.
3029 * The request must not have been partially completed before.
3031 void blk_steal_bios(struct bio_list *list, struct request *rq)
3035 list->tail->bi_next = rq->bio;
3037 list->head = rq->bio;
3038 list->tail = rq->biotail;
3046 EXPORT_SYMBOL_GPL(blk_steal_bios);
3048 static size_t order_to_size(unsigned int order)
3050 return (size_t)PAGE_SIZE << order;
3053 /* called before freeing request pool in @tags */
3054 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3055 struct blk_mq_tags *tags)
3058 unsigned long flags;
3060 /* There is no need to clear a driver tags own mapping */
3061 if (drv_tags == tags)
3064 list_for_each_entry(page, &tags->page_list, lru) {
3065 unsigned long start = (unsigned long)page_address(page);
3066 unsigned long end = start + order_to_size(page->private);
3069 for (i = 0; i < drv_tags->nr_tags; i++) {
3070 struct request *rq = drv_tags->rqs[i];
3071 unsigned long rq_addr = (unsigned long)rq;
3073 if (rq_addr >= start && rq_addr < end) {
3074 WARN_ON_ONCE(req_ref_read(rq) != 0);
3075 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3081 * Wait until all pending iteration is done.
3083 * Request reference is cleared and it is guaranteed to be observed
3084 * after the ->lock is released.
3086 spin_lock_irqsave(&drv_tags->lock, flags);
3087 spin_unlock_irqrestore(&drv_tags->lock, flags);
3090 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3091 unsigned int hctx_idx)
3093 struct blk_mq_tags *drv_tags;
3096 if (list_empty(&tags->page_list))
3099 if (blk_mq_is_shared_tags(set->flags))
3100 drv_tags = set->shared_tags;
3102 drv_tags = set->tags[hctx_idx];
3104 if (tags->static_rqs && set->ops->exit_request) {
3107 for (i = 0; i < tags->nr_tags; i++) {
3108 struct request *rq = tags->static_rqs[i];
3112 set->ops->exit_request(set, rq, hctx_idx);
3113 tags->static_rqs[i] = NULL;
3117 blk_mq_clear_rq_mapping(drv_tags, tags);
3119 while (!list_empty(&tags->page_list)) {
3120 page = list_first_entry(&tags->page_list, struct page, lru);
3121 list_del_init(&page->lru);
3123 * Remove kmemleak object previously allocated in
3124 * blk_mq_alloc_rqs().
3126 kmemleak_free(page_address(page));
3127 __free_pages(page, page->private);
3131 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3135 kfree(tags->static_rqs);
3136 tags->static_rqs = NULL;
3138 blk_mq_free_tags(tags);
3141 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3142 unsigned int hctx_idx)
3146 for (i = 0; i < set->nr_maps; i++) {
3147 unsigned int start = set->map[i].queue_offset;
3148 unsigned int end = start + set->map[i].nr_queues;
3150 if (hctx_idx >= start && hctx_idx < end)
3154 if (i >= set->nr_maps)
3155 i = HCTX_TYPE_DEFAULT;
3160 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3161 unsigned int hctx_idx)
3163 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3165 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3168 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3169 unsigned int hctx_idx,
3170 unsigned int nr_tags,
3171 unsigned int reserved_tags)
3173 int node = blk_mq_get_hctx_node(set, hctx_idx);
3174 struct blk_mq_tags *tags;
3176 if (node == NUMA_NO_NODE)
3177 node = set->numa_node;
3179 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3180 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3184 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3185 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3188 blk_mq_free_tags(tags);
3192 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3193 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3195 if (!tags->static_rqs) {
3197 blk_mq_free_tags(tags);
3204 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3205 unsigned int hctx_idx, int node)
3209 if (set->ops->init_request) {
3210 ret = set->ops->init_request(set, rq, hctx_idx, node);
3215 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3219 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3220 struct blk_mq_tags *tags,
3221 unsigned int hctx_idx, unsigned int depth)
3223 unsigned int i, j, entries_per_page, max_order = 4;
3224 int node = blk_mq_get_hctx_node(set, hctx_idx);
3225 size_t rq_size, left;
3227 if (node == NUMA_NO_NODE)
3228 node = set->numa_node;
3230 INIT_LIST_HEAD(&tags->page_list);
3233 * rq_size is the size of the request plus driver payload, rounded
3234 * to the cacheline size
3236 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3238 left = rq_size * depth;
3240 for (i = 0; i < depth; ) {
3241 int this_order = max_order;
3246 while (this_order && left < order_to_size(this_order - 1))
3250 page = alloc_pages_node(node,
3251 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3257 if (order_to_size(this_order) < rq_size)
3264 page->private = this_order;
3265 list_add_tail(&page->lru, &tags->page_list);
3267 p = page_address(page);
3269 * Allow kmemleak to scan these pages as they contain pointers
3270 * to additional allocations like via ops->init_request().
3272 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3273 entries_per_page = order_to_size(this_order) / rq_size;
3274 to_do = min(entries_per_page, depth - i);
3275 left -= to_do * rq_size;
3276 for (j = 0; j < to_do; j++) {
3277 struct request *rq = p;
3279 tags->static_rqs[i] = rq;
3280 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3281 tags->static_rqs[i] = NULL;
3292 blk_mq_free_rqs(set, tags, hctx_idx);
3296 struct rq_iter_data {
3297 struct blk_mq_hw_ctx *hctx;
3301 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
3303 struct rq_iter_data *iter_data = data;
3305 if (rq->mq_hctx != iter_data->hctx)
3307 iter_data->has_rq = true;
3311 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3313 struct blk_mq_tags *tags = hctx->sched_tags ?
3314 hctx->sched_tags : hctx->tags;
3315 struct rq_iter_data data = {
3319 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3323 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3324 struct blk_mq_hw_ctx *hctx)
3326 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3328 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3333 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3335 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3336 struct blk_mq_hw_ctx, cpuhp_online);
3338 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3339 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3343 * Prevent new request from being allocated on the current hctx.
3345 * The smp_mb__after_atomic() Pairs with the implied barrier in
3346 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3347 * seen once we return from the tag allocator.
3349 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3350 smp_mb__after_atomic();
3353 * Try to grab a reference to the queue and wait for any outstanding
3354 * requests. If we could not grab a reference the queue has been
3355 * frozen and there are no requests.
3357 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3358 while (blk_mq_hctx_has_requests(hctx))
3360 percpu_ref_put(&hctx->queue->q_usage_counter);
3366 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3368 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3369 struct blk_mq_hw_ctx, cpuhp_online);
3371 if (cpumask_test_cpu(cpu, hctx->cpumask))
3372 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3377 * 'cpu' is going away. splice any existing rq_list entries from this
3378 * software queue to the hw queue dispatch list, and ensure that it
3381 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3383 struct blk_mq_hw_ctx *hctx;
3384 struct blk_mq_ctx *ctx;
3386 enum hctx_type type;
3388 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3389 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3392 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3395 spin_lock(&ctx->lock);
3396 if (!list_empty(&ctx->rq_lists[type])) {
3397 list_splice_init(&ctx->rq_lists[type], &tmp);
3398 blk_mq_hctx_clear_pending(hctx, ctx);
3400 spin_unlock(&ctx->lock);
3402 if (list_empty(&tmp))
3405 spin_lock(&hctx->lock);
3406 list_splice_tail_init(&tmp, &hctx->dispatch);
3407 spin_unlock(&hctx->lock);
3409 blk_mq_run_hw_queue(hctx, true);
3413 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3415 if (!(hctx->flags & BLK_MQ_F_STACKING))
3416 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3417 &hctx->cpuhp_online);
3418 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3423 * Before freeing hw queue, clearing the flush request reference in
3424 * tags->rqs[] for avoiding potential UAF.
3426 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3427 unsigned int queue_depth, struct request *flush_rq)
3430 unsigned long flags;
3432 /* The hw queue may not be mapped yet */
3436 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3438 for (i = 0; i < queue_depth; i++)
3439 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3442 * Wait until all pending iteration is done.
3444 * Request reference is cleared and it is guaranteed to be observed
3445 * after the ->lock is released.
3447 spin_lock_irqsave(&tags->lock, flags);
3448 spin_unlock_irqrestore(&tags->lock, flags);
3451 /* hctx->ctxs will be freed in queue's release handler */
3452 static void blk_mq_exit_hctx(struct request_queue *q,
3453 struct blk_mq_tag_set *set,
3454 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3456 struct request *flush_rq = hctx->fq->flush_rq;
3458 if (blk_mq_hw_queue_mapped(hctx))
3459 blk_mq_tag_idle(hctx);
3461 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3462 set->queue_depth, flush_rq);
3463 if (set->ops->exit_request)
3464 set->ops->exit_request(set, flush_rq, hctx_idx);
3466 if (set->ops->exit_hctx)
3467 set->ops->exit_hctx(hctx, hctx_idx);
3469 blk_mq_remove_cpuhp(hctx);
3471 xa_erase(&q->hctx_table, hctx_idx);
3473 spin_lock(&q->unused_hctx_lock);
3474 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3475 spin_unlock(&q->unused_hctx_lock);
3478 static void blk_mq_exit_hw_queues(struct request_queue *q,
3479 struct blk_mq_tag_set *set, int nr_queue)
3481 struct blk_mq_hw_ctx *hctx;
3484 queue_for_each_hw_ctx(q, hctx, i) {
3487 blk_mq_exit_hctx(q, set, hctx, i);
3491 static int blk_mq_init_hctx(struct request_queue *q,
3492 struct blk_mq_tag_set *set,
3493 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3495 hctx->queue_num = hctx_idx;
3497 if (!(hctx->flags & BLK_MQ_F_STACKING))
3498 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3499 &hctx->cpuhp_online);
3500 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3502 hctx->tags = set->tags[hctx_idx];
3504 if (set->ops->init_hctx &&
3505 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3506 goto unregister_cpu_notifier;
3508 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3512 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3518 if (set->ops->exit_request)
3519 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3521 if (set->ops->exit_hctx)
3522 set->ops->exit_hctx(hctx, hctx_idx);
3523 unregister_cpu_notifier:
3524 blk_mq_remove_cpuhp(hctx);
3528 static struct blk_mq_hw_ctx *
3529 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3532 struct blk_mq_hw_ctx *hctx;
3533 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3535 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3537 goto fail_alloc_hctx;
3539 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3542 atomic_set(&hctx->nr_active, 0);
3543 if (node == NUMA_NO_NODE)
3544 node = set->numa_node;
3545 hctx->numa_node = node;
3547 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3548 spin_lock_init(&hctx->lock);
3549 INIT_LIST_HEAD(&hctx->dispatch);
3551 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3553 INIT_LIST_HEAD(&hctx->hctx_list);
3556 * Allocate space for all possible cpus to avoid allocation at
3559 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3564 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3565 gfp, node, false, false))
3569 spin_lock_init(&hctx->dispatch_wait_lock);
3570 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3571 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3573 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3577 blk_mq_hctx_kobj_init(hctx);
3582 sbitmap_free(&hctx->ctx_map);
3586 free_cpumask_var(hctx->cpumask);
3593 static void blk_mq_init_cpu_queues(struct request_queue *q,
3594 unsigned int nr_hw_queues)
3596 struct blk_mq_tag_set *set = q->tag_set;
3599 for_each_possible_cpu(i) {
3600 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3601 struct blk_mq_hw_ctx *hctx;
3605 spin_lock_init(&__ctx->lock);
3606 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3607 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3612 * Set local node, IFF we have more than one hw queue. If
3613 * not, we remain on the home node of the device
3615 for (j = 0; j < set->nr_maps; j++) {
3616 hctx = blk_mq_map_queue_type(q, j, i);
3617 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3618 hctx->numa_node = cpu_to_node(i);
3623 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3624 unsigned int hctx_idx,
3627 struct blk_mq_tags *tags;
3630 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3634 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3636 blk_mq_free_rq_map(tags);
3643 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3646 if (blk_mq_is_shared_tags(set->flags)) {
3647 set->tags[hctx_idx] = set->shared_tags;
3652 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3655 return set->tags[hctx_idx];
3658 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3659 struct blk_mq_tags *tags,
3660 unsigned int hctx_idx)
3663 blk_mq_free_rqs(set, tags, hctx_idx);
3664 blk_mq_free_rq_map(tags);
3668 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3669 unsigned int hctx_idx)
3671 if (!blk_mq_is_shared_tags(set->flags))
3672 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3674 set->tags[hctx_idx] = NULL;
3677 static void blk_mq_map_swqueue(struct request_queue *q)
3679 unsigned int j, hctx_idx;
3681 struct blk_mq_hw_ctx *hctx;
3682 struct blk_mq_ctx *ctx;
3683 struct blk_mq_tag_set *set = q->tag_set;
3685 queue_for_each_hw_ctx(q, hctx, i) {
3686 cpumask_clear(hctx->cpumask);
3688 hctx->dispatch_from = NULL;
3692 * Map software to hardware queues.
3694 * If the cpu isn't present, the cpu is mapped to first hctx.
3696 for_each_possible_cpu(i) {
3698 ctx = per_cpu_ptr(q->queue_ctx, i);
3699 for (j = 0; j < set->nr_maps; j++) {
3700 if (!set->map[j].nr_queues) {
3701 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3702 HCTX_TYPE_DEFAULT, i);
3705 hctx_idx = set->map[j].mq_map[i];
3706 /* unmapped hw queue can be remapped after CPU topo changed */
3707 if (!set->tags[hctx_idx] &&
3708 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3710 * If tags initialization fail for some hctx,
3711 * that hctx won't be brought online. In this
3712 * case, remap the current ctx to hctx[0] which
3713 * is guaranteed to always have tags allocated
3715 set->map[j].mq_map[i] = 0;
3718 hctx = blk_mq_map_queue_type(q, j, i);
3719 ctx->hctxs[j] = hctx;
3721 * If the CPU is already set in the mask, then we've
3722 * mapped this one already. This can happen if
3723 * devices share queues across queue maps.
3725 if (cpumask_test_cpu(i, hctx->cpumask))
3728 cpumask_set_cpu(i, hctx->cpumask);
3730 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3731 hctx->ctxs[hctx->nr_ctx++] = ctx;
3734 * If the nr_ctx type overflows, we have exceeded the
3735 * amount of sw queues we can support.
3737 BUG_ON(!hctx->nr_ctx);
3740 for (; j < HCTX_MAX_TYPES; j++)
3741 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3742 HCTX_TYPE_DEFAULT, i);
3745 queue_for_each_hw_ctx(q, hctx, i) {
3747 * If no software queues are mapped to this hardware queue,
3748 * disable it and free the request entries.
3750 if (!hctx->nr_ctx) {
3751 /* Never unmap queue 0. We need it as a
3752 * fallback in case of a new remap fails
3756 __blk_mq_free_map_and_rqs(set, i);
3762 hctx->tags = set->tags[i];
3763 WARN_ON(!hctx->tags);
3766 * Set the map size to the number of mapped software queues.
3767 * This is more accurate and more efficient than looping
3768 * over all possibly mapped software queues.
3770 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3773 * Initialize batch roundrobin counts
3775 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3776 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3781 * Caller needs to ensure that we're either frozen/quiesced, or that
3782 * the queue isn't live yet.
3784 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3786 struct blk_mq_hw_ctx *hctx;
3789 queue_for_each_hw_ctx(q, hctx, i) {
3791 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3793 blk_mq_tag_idle(hctx);
3794 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3799 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3802 struct request_queue *q;
3804 lockdep_assert_held(&set->tag_list_lock);
3806 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3807 blk_mq_freeze_queue(q);
3808 queue_set_hctx_shared(q, shared);
3809 blk_mq_unfreeze_queue(q);
3813 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3815 struct blk_mq_tag_set *set = q->tag_set;
3817 mutex_lock(&set->tag_list_lock);
3818 list_del(&q->tag_set_list);
3819 if (list_is_singular(&set->tag_list)) {
3820 /* just transitioned to unshared */
3821 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3822 /* update existing queue */
3823 blk_mq_update_tag_set_shared(set, false);
3825 mutex_unlock(&set->tag_list_lock);
3826 INIT_LIST_HEAD(&q->tag_set_list);
3829 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3830 struct request_queue *q)
3832 mutex_lock(&set->tag_list_lock);
3835 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3837 if (!list_empty(&set->tag_list) &&
3838 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3839 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3840 /* update existing queue */
3841 blk_mq_update_tag_set_shared(set, true);
3843 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3844 queue_set_hctx_shared(q, true);
3845 list_add_tail(&q->tag_set_list, &set->tag_list);
3847 mutex_unlock(&set->tag_list_lock);
3850 /* All allocations will be freed in release handler of q->mq_kobj */
3851 static int blk_mq_alloc_ctxs(struct request_queue *q)
3853 struct blk_mq_ctxs *ctxs;
3856 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3860 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3861 if (!ctxs->queue_ctx)
3864 for_each_possible_cpu(cpu) {
3865 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3869 q->mq_kobj = &ctxs->kobj;
3870 q->queue_ctx = ctxs->queue_ctx;
3879 * It is the actual release handler for mq, but we do it from
3880 * request queue's release handler for avoiding use-after-free
3881 * and headache because q->mq_kobj shouldn't have been introduced,
3882 * but we can't group ctx/kctx kobj without it.
3884 void blk_mq_release(struct request_queue *q)
3886 struct blk_mq_hw_ctx *hctx, *next;
3889 queue_for_each_hw_ctx(q, hctx, i)
3890 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3892 /* all hctx are in .unused_hctx_list now */
3893 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3894 list_del_init(&hctx->hctx_list);
3895 kobject_put(&hctx->kobj);
3898 xa_destroy(&q->hctx_table);
3901 * release .mq_kobj and sw queue's kobject now because
3902 * both share lifetime with request queue.
3904 blk_mq_sysfs_deinit(q);
3907 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3910 struct request_queue *q;
3913 q = blk_alloc_queue(set->numa_node, set->flags & BLK_MQ_F_BLOCKING);
3915 return ERR_PTR(-ENOMEM);
3916 q->queuedata = queuedata;
3917 ret = blk_mq_init_allocated_queue(set, q);
3919 blk_cleanup_queue(q);
3920 return ERR_PTR(ret);
3925 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3927 return blk_mq_init_queue_data(set, NULL);
3929 EXPORT_SYMBOL(blk_mq_init_queue);
3931 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
3932 struct lock_class_key *lkclass)
3934 struct request_queue *q;
3935 struct gendisk *disk;
3937 q = blk_mq_init_queue_data(set, queuedata);
3941 disk = __alloc_disk_node(q, set->numa_node, lkclass);
3943 blk_cleanup_queue(q);
3944 return ERR_PTR(-ENOMEM);
3948 EXPORT_SYMBOL(__blk_mq_alloc_disk);
3950 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3951 struct blk_mq_tag_set *set, struct request_queue *q,
3952 int hctx_idx, int node)
3954 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3956 /* reuse dead hctx first */
3957 spin_lock(&q->unused_hctx_lock);
3958 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3959 if (tmp->numa_node == node) {
3965 list_del_init(&hctx->hctx_list);
3966 spin_unlock(&q->unused_hctx_lock);
3969 hctx = blk_mq_alloc_hctx(q, set, node);
3973 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3979 kobject_put(&hctx->kobj);
3984 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3985 struct request_queue *q)
3987 struct blk_mq_hw_ctx *hctx;
3990 /* protect against switching io scheduler */
3991 mutex_lock(&q->sysfs_lock);
3992 for (i = 0; i < set->nr_hw_queues; i++) {
3994 int node = blk_mq_get_hctx_node(set, i);
3995 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
3998 old_node = old_hctx->numa_node;
3999 blk_mq_exit_hctx(q, set, old_hctx, i);
4002 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4005 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4007 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4008 WARN_ON_ONCE(!hctx);
4012 * Increasing nr_hw_queues fails. Free the newly allocated
4013 * hctxs and keep the previous q->nr_hw_queues.
4015 if (i != set->nr_hw_queues) {
4016 j = q->nr_hw_queues;
4019 q->nr_hw_queues = set->nr_hw_queues;
4022 xa_for_each_start(&q->hctx_table, j, hctx, j)
4023 blk_mq_exit_hctx(q, set, hctx, j);
4024 mutex_unlock(&q->sysfs_lock);
4027 static void blk_mq_update_poll_flag(struct request_queue *q)
4029 struct blk_mq_tag_set *set = q->tag_set;
4031 if (set->nr_maps > HCTX_TYPE_POLL &&
4032 set->map[HCTX_TYPE_POLL].nr_queues)
4033 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4035 blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4038 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4039 struct request_queue *q)
4041 WARN_ON_ONCE(blk_queue_has_srcu(q) !=
4042 !!(set->flags & BLK_MQ_F_BLOCKING));
4044 /* mark the queue as mq asap */
4045 q->mq_ops = set->ops;
4047 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
4048 blk_mq_poll_stats_bkt,
4049 BLK_MQ_POLL_STATS_BKTS, q);
4053 if (blk_mq_alloc_ctxs(q))
4056 /* init q->mq_kobj and sw queues' kobjects */
4057 blk_mq_sysfs_init(q);
4059 INIT_LIST_HEAD(&q->unused_hctx_list);
4060 spin_lock_init(&q->unused_hctx_lock);
4062 xa_init(&q->hctx_table);
4064 blk_mq_realloc_hw_ctxs(set, q);
4065 if (!q->nr_hw_queues)
4068 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4069 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4073 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4074 blk_mq_update_poll_flag(q);
4076 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4077 INIT_LIST_HEAD(&q->requeue_list);
4078 spin_lock_init(&q->requeue_lock);
4080 q->nr_requests = set->queue_depth;
4083 * Default to classic polling
4085 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
4087 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4088 blk_mq_add_queue_tag_set(set, q);
4089 blk_mq_map_swqueue(q);
4093 xa_destroy(&q->hctx_table);
4094 q->nr_hw_queues = 0;
4095 blk_mq_sysfs_deinit(q);
4097 blk_stat_free_callback(q->poll_cb);
4103 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4105 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4106 void blk_mq_exit_queue(struct request_queue *q)
4108 struct blk_mq_tag_set *set = q->tag_set;
4110 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4111 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4112 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4113 blk_mq_del_queue_tag_set(q);
4116 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4120 if (blk_mq_is_shared_tags(set->flags)) {
4121 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4124 if (!set->shared_tags)
4128 for (i = 0; i < set->nr_hw_queues; i++) {
4129 if (!__blk_mq_alloc_map_and_rqs(set, i))
4138 __blk_mq_free_map_and_rqs(set, i);
4140 if (blk_mq_is_shared_tags(set->flags)) {
4141 blk_mq_free_map_and_rqs(set, set->shared_tags,
4142 BLK_MQ_NO_HCTX_IDX);
4149 * Allocate the request maps associated with this tag_set. Note that this
4150 * may reduce the depth asked for, if memory is tight. set->queue_depth
4151 * will be updated to reflect the allocated depth.
4153 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4158 depth = set->queue_depth;
4160 err = __blk_mq_alloc_rq_maps(set);
4164 set->queue_depth >>= 1;
4165 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4169 } while (set->queue_depth);
4171 if (!set->queue_depth || err) {
4172 pr_err("blk-mq: failed to allocate request map\n");
4176 if (depth != set->queue_depth)
4177 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4178 depth, set->queue_depth);
4183 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4186 * blk_mq_map_queues() and multiple .map_queues() implementations
4187 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4188 * number of hardware queues.
4190 if (set->nr_maps == 1)
4191 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4193 if (set->ops->map_queues && !is_kdump_kernel()) {
4197 * transport .map_queues is usually done in the following
4200 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4201 * mask = get_cpu_mask(queue)
4202 * for_each_cpu(cpu, mask)
4203 * set->map[x].mq_map[cpu] = queue;
4206 * When we need to remap, the table has to be cleared for
4207 * killing stale mapping since one CPU may not be mapped
4210 for (i = 0; i < set->nr_maps; i++)
4211 blk_mq_clear_mq_map(&set->map[i]);
4213 return set->ops->map_queues(set);
4215 BUG_ON(set->nr_maps > 1);
4216 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4220 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4221 int cur_nr_hw_queues, int new_nr_hw_queues)
4223 struct blk_mq_tags **new_tags;
4225 if (cur_nr_hw_queues >= new_nr_hw_queues)
4228 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4229 GFP_KERNEL, set->numa_node);
4234 memcpy(new_tags, set->tags, cur_nr_hw_queues *
4235 sizeof(*set->tags));
4237 set->tags = new_tags;
4238 set->nr_hw_queues = new_nr_hw_queues;
4243 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
4244 int new_nr_hw_queues)
4246 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
4250 * Alloc a tag set to be associated with one or more request queues.
4251 * May fail with EINVAL for various error conditions. May adjust the
4252 * requested depth down, if it's too large. In that case, the set
4253 * value will be stored in set->queue_depth.
4255 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4259 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4261 if (!set->nr_hw_queues)
4263 if (!set->queue_depth)
4265 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4268 if (!set->ops->queue_rq)
4271 if (!set->ops->get_budget ^ !set->ops->put_budget)
4274 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4275 pr_info("blk-mq: reduced tag depth to %u\n",
4277 set->queue_depth = BLK_MQ_MAX_DEPTH;
4282 else if (set->nr_maps > HCTX_MAX_TYPES)
4286 * If a crashdump is active, then we are potentially in a very
4287 * memory constrained environment. Limit us to 1 queue and
4288 * 64 tags to prevent using too much memory.
4290 if (is_kdump_kernel()) {
4291 set->nr_hw_queues = 1;
4293 set->queue_depth = min(64U, set->queue_depth);
4296 * There is no use for more h/w queues than cpus if we just have
4299 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4300 set->nr_hw_queues = nr_cpu_ids;
4302 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
4306 for (i = 0; i < set->nr_maps; i++) {
4307 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4308 sizeof(set->map[i].mq_map[0]),
4309 GFP_KERNEL, set->numa_node);
4310 if (!set->map[i].mq_map)
4311 goto out_free_mq_map;
4312 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4315 ret = blk_mq_update_queue_map(set);
4317 goto out_free_mq_map;
4319 ret = blk_mq_alloc_set_map_and_rqs(set);
4321 goto out_free_mq_map;
4323 mutex_init(&set->tag_list_lock);
4324 INIT_LIST_HEAD(&set->tag_list);
4329 for (i = 0; i < set->nr_maps; i++) {
4330 kfree(set->map[i].mq_map);
4331 set->map[i].mq_map = NULL;
4337 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4339 /* allocate and initialize a tagset for a simple single-queue device */
4340 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4341 const struct blk_mq_ops *ops, unsigned int queue_depth,
4342 unsigned int set_flags)
4344 memset(set, 0, sizeof(*set));
4346 set->nr_hw_queues = 1;
4348 set->queue_depth = queue_depth;
4349 set->numa_node = NUMA_NO_NODE;
4350 set->flags = set_flags;
4351 return blk_mq_alloc_tag_set(set);
4353 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4355 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4359 for (i = 0; i < set->nr_hw_queues; i++)
4360 __blk_mq_free_map_and_rqs(set, i);
4362 if (blk_mq_is_shared_tags(set->flags)) {
4363 blk_mq_free_map_and_rqs(set, set->shared_tags,
4364 BLK_MQ_NO_HCTX_IDX);
4367 for (j = 0; j < set->nr_maps; j++) {
4368 kfree(set->map[j].mq_map);
4369 set->map[j].mq_map = NULL;
4375 EXPORT_SYMBOL(blk_mq_free_tag_set);
4377 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4379 struct blk_mq_tag_set *set = q->tag_set;
4380 struct blk_mq_hw_ctx *hctx;
4387 if (q->nr_requests == nr)
4390 blk_mq_freeze_queue(q);
4391 blk_mq_quiesce_queue(q);
4394 queue_for_each_hw_ctx(q, hctx, i) {
4398 * If we're using an MQ scheduler, just update the scheduler
4399 * queue depth. This is similar to what the old code would do.
4401 if (hctx->sched_tags) {
4402 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4405 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4410 if (q->elevator && q->elevator->type->ops.depth_updated)
4411 q->elevator->type->ops.depth_updated(hctx);
4414 q->nr_requests = nr;
4415 if (blk_mq_is_shared_tags(set->flags)) {
4417 blk_mq_tag_update_sched_shared_tags(q);
4419 blk_mq_tag_resize_shared_tags(set, nr);
4423 blk_mq_unquiesce_queue(q);
4424 blk_mq_unfreeze_queue(q);
4430 * request_queue and elevator_type pair.
4431 * It is just used by __blk_mq_update_nr_hw_queues to cache
4432 * the elevator_type associated with a request_queue.
4434 struct blk_mq_qe_pair {
4435 struct list_head node;
4436 struct request_queue *q;
4437 struct elevator_type *type;
4441 * Cache the elevator_type in qe pair list and switch the
4442 * io scheduler to 'none'
4444 static bool blk_mq_elv_switch_none(struct list_head *head,
4445 struct request_queue *q)
4447 struct blk_mq_qe_pair *qe;
4452 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4456 INIT_LIST_HEAD(&qe->node);
4458 qe->type = q->elevator->type;
4459 list_add(&qe->node, head);
4461 mutex_lock(&q->sysfs_lock);
4463 * After elevator_switch_mq, the previous elevator_queue will be
4464 * released by elevator_release. The reference of the io scheduler
4465 * module get by elevator_get will also be put. So we need to get
4466 * a reference of the io scheduler module here to prevent it to be
4469 __module_get(qe->type->elevator_owner);
4470 elevator_switch_mq(q, NULL);
4471 mutex_unlock(&q->sysfs_lock);
4476 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4477 struct request_queue *q)
4479 struct blk_mq_qe_pair *qe;
4481 list_for_each_entry(qe, head, node)
4488 static void blk_mq_elv_switch_back(struct list_head *head,
4489 struct request_queue *q)
4491 struct blk_mq_qe_pair *qe;
4492 struct elevator_type *t;
4494 qe = blk_lookup_qe_pair(head, q);
4498 list_del(&qe->node);
4501 mutex_lock(&q->sysfs_lock);
4502 elevator_switch_mq(q, t);
4503 mutex_unlock(&q->sysfs_lock);
4506 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4509 struct request_queue *q;
4511 int prev_nr_hw_queues;
4513 lockdep_assert_held(&set->tag_list_lock);
4515 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4516 nr_hw_queues = nr_cpu_ids;
4517 if (nr_hw_queues < 1)
4519 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4522 list_for_each_entry(q, &set->tag_list, tag_set_list)
4523 blk_mq_freeze_queue(q);
4525 * Switch IO scheduler to 'none', cleaning up the data associated
4526 * with the previous scheduler. We will switch back once we are done
4527 * updating the new sw to hw queue mappings.
4529 list_for_each_entry(q, &set->tag_list, tag_set_list)
4530 if (!blk_mq_elv_switch_none(&head, q))
4533 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4534 blk_mq_debugfs_unregister_hctxs(q);
4535 blk_mq_sysfs_unregister(q);
4538 prev_nr_hw_queues = set->nr_hw_queues;
4539 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
4543 set->nr_hw_queues = nr_hw_queues;
4545 blk_mq_update_queue_map(set);
4546 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4547 blk_mq_realloc_hw_ctxs(set, q);
4548 blk_mq_update_poll_flag(q);
4549 if (q->nr_hw_queues != set->nr_hw_queues) {
4550 int i = prev_nr_hw_queues;
4552 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4553 nr_hw_queues, prev_nr_hw_queues);
4554 for (; i < set->nr_hw_queues; i++)
4555 __blk_mq_free_map_and_rqs(set, i);
4557 set->nr_hw_queues = prev_nr_hw_queues;
4558 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4561 blk_mq_map_swqueue(q);
4565 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4566 blk_mq_sysfs_register(q);
4567 blk_mq_debugfs_register_hctxs(q);
4571 list_for_each_entry(q, &set->tag_list, tag_set_list)
4572 blk_mq_elv_switch_back(&head, q);
4574 list_for_each_entry(q, &set->tag_list, tag_set_list)
4575 blk_mq_unfreeze_queue(q);
4578 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4580 mutex_lock(&set->tag_list_lock);
4581 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4582 mutex_unlock(&set->tag_list_lock);
4584 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4586 /* Enable polling stats and return whether they were already enabled. */
4587 static bool blk_poll_stats_enable(struct request_queue *q)
4592 return blk_stats_alloc_enable(q);
4595 static void blk_mq_poll_stats_start(struct request_queue *q)
4598 * We don't arm the callback if polling stats are not enabled or the
4599 * callback is already active.
4601 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
4604 blk_stat_activate_msecs(q->poll_cb, 100);
4607 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4609 struct request_queue *q = cb->data;
4612 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4613 if (cb->stat[bucket].nr_samples)
4614 q->poll_stat[bucket] = cb->stat[bucket];
4618 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4621 unsigned long ret = 0;
4625 * If stats collection isn't on, don't sleep but turn it on for
4628 if (!blk_poll_stats_enable(q))
4632 * As an optimistic guess, use half of the mean service time
4633 * for this type of request. We can (and should) make this smarter.
4634 * For instance, if the completion latencies are tight, we can
4635 * get closer than just half the mean. This is especially
4636 * important on devices where the completion latencies are longer
4637 * than ~10 usec. We do use the stats for the relevant IO size
4638 * if available which does lead to better estimates.
4640 bucket = blk_mq_poll_stats_bkt(rq);
4644 if (q->poll_stat[bucket].nr_samples)
4645 ret = (q->poll_stat[bucket].mean + 1) / 2;
4650 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4652 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4653 struct request *rq = blk_qc_to_rq(hctx, qc);
4654 struct hrtimer_sleeper hs;
4655 enum hrtimer_mode mode;
4660 * If a request has completed on queue that uses an I/O scheduler, we
4661 * won't get back a request from blk_qc_to_rq.
4663 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4667 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4669 * 0: use half of prev avg
4670 * >0: use this specific value
4672 if (q->poll_nsec > 0)
4673 nsecs = q->poll_nsec;
4675 nsecs = blk_mq_poll_nsecs(q, rq);
4680 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4683 * This will be replaced with the stats tracking code, using
4684 * 'avg_completion_time / 2' as the pre-sleep target.
4688 mode = HRTIMER_MODE_REL;
4689 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4690 hrtimer_set_expires(&hs.timer, kt);
4693 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4695 set_current_state(TASK_UNINTERRUPTIBLE);
4696 hrtimer_sleeper_start_expires(&hs, mode);
4699 hrtimer_cancel(&hs.timer);
4700 mode = HRTIMER_MODE_ABS;
4701 } while (hs.task && !signal_pending(current));
4703 __set_current_state(TASK_RUNNING);
4704 destroy_hrtimer_on_stack(&hs.timer);
4707 * If we sleep, have the caller restart the poll loop to reset the
4708 * state. Like for the other success return cases, the caller is
4709 * responsible for checking if the IO completed. If the IO isn't
4710 * complete, we'll get called again and will go straight to the busy
4716 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4717 struct io_comp_batch *iob, unsigned int flags)
4719 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4720 long state = get_current_state();
4724 ret = q->mq_ops->poll(hctx, iob);
4726 __set_current_state(TASK_RUNNING);
4730 if (signal_pending_state(state, current))
4731 __set_current_state(TASK_RUNNING);
4732 if (task_is_running(current))
4735 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4738 } while (!need_resched());
4740 __set_current_state(TASK_RUNNING);
4744 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4747 if (!(flags & BLK_POLL_NOSLEEP) &&
4748 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4749 if (blk_mq_poll_hybrid(q, cookie))
4752 return blk_mq_poll_classic(q, cookie, iob, flags);
4755 unsigned int blk_mq_rq_cpu(struct request *rq)
4757 return rq->mq_ctx->cpu;
4759 EXPORT_SYMBOL(blk_mq_rq_cpu);
4761 void blk_mq_cancel_work_sync(struct request_queue *q)
4763 if (queue_is_mq(q)) {
4764 struct blk_mq_hw_ctx *hctx;
4767 cancel_delayed_work_sync(&q->requeue_work);
4769 queue_for_each_hw_ctx(q, hctx, i)
4770 cancel_delayed_work_sync(&hctx->run_work);
4774 static int __init blk_mq_init(void)
4778 for_each_possible_cpu(i)
4779 init_llist_head(&per_cpu(blk_cpu_done, i));
4780 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4782 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4783 "block/softirq:dead", NULL,
4784 blk_softirq_cpu_dead);
4785 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4786 blk_mq_hctx_notify_dead);
4787 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4788 blk_mq_hctx_notify_online,
4789 blk_mq_hctx_notify_offline);
4792 subsys_initcall(blk_mq_init);