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 q->queue_hw_ctx[(qc & ~BLK_QC_T_INTERNAL) >> BLK_QC_T_SHIFT];
77 static inline struct request *blk_qc_to_rq(struct blk_mq_hw_ctx *hctx,
80 unsigned int tag = qc & ((1U << BLK_QC_T_SHIFT) - 1);
82 if (qc & BLK_QC_T_INTERNAL)
83 return blk_mq_tag_to_rq(hctx->sched_tags, tag);
84 return blk_mq_tag_to_rq(hctx->tags, tag);
87 static inline blk_qc_t blk_rq_to_qc(struct request *rq)
89 return (rq->mq_hctx->queue_num << BLK_QC_T_SHIFT) |
91 rq->tag : (rq->internal_tag | BLK_QC_T_INTERNAL));
95 * Check if any of the ctx, dispatch list or elevator
96 * have pending work in this hardware queue.
98 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
100 return !list_empty_careful(&hctx->dispatch) ||
101 sbitmap_any_bit_set(&hctx->ctx_map) ||
102 blk_mq_sched_has_work(hctx);
106 * Mark this ctx as having pending work in this hardware queue
108 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
109 struct blk_mq_ctx *ctx)
111 const int bit = ctx->index_hw[hctx->type];
113 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
114 sbitmap_set_bit(&hctx->ctx_map, bit);
117 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
118 struct blk_mq_ctx *ctx)
120 const int bit = ctx->index_hw[hctx->type];
122 sbitmap_clear_bit(&hctx->ctx_map, bit);
126 struct block_device *part;
127 unsigned int inflight[2];
130 static bool blk_mq_check_inflight(struct request *rq, void *priv,
133 struct mq_inflight *mi = priv;
135 if ((!mi->part->bd_partno || rq->part == mi->part) &&
136 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
137 mi->inflight[rq_data_dir(rq)]++;
142 unsigned int blk_mq_in_flight(struct request_queue *q,
143 struct block_device *part)
145 struct mq_inflight mi = { .part = part };
147 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
149 return mi.inflight[0] + mi.inflight[1];
152 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
153 unsigned int inflight[2])
155 struct mq_inflight mi = { .part = part };
157 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
158 inflight[0] = mi.inflight[0];
159 inflight[1] = mi.inflight[1];
162 void blk_freeze_queue_start(struct request_queue *q)
164 mutex_lock(&q->mq_freeze_lock);
165 if (++q->mq_freeze_depth == 1) {
166 percpu_ref_kill(&q->q_usage_counter);
167 mutex_unlock(&q->mq_freeze_lock);
169 blk_mq_run_hw_queues(q, false);
171 mutex_unlock(&q->mq_freeze_lock);
174 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
176 void blk_mq_freeze_queue_wait(struct request_queue *q)
178 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
180 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
182 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
183 unsigned long timeout)
185 return wait_event_timeout(q->mq_freeze_wq,
186 percpu_ref_is_zero(&q->q_usage_counter),
189 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
192 * Guarantee no request is in use, so we can change any data structure of
193 * the queue afterward.
195 void blk_freeze_queue(struct request_queue *q)
198 * In the !blk_mq case we are only calling this to kill the
199 * q_usage_counter, otherwise this increases the freeze depth
200 * and waits for it to return to zero. For this reason there is
201 * no blk_unfreeze_queue(), and blk_freeze_queue() is not
202 * exported to drivers as the only user for unfreeze is blk_mq.
204 blk_freeze_queue_start(q);
205 blk_mq_freeze_queue_wait(q);
208 void blk_mq_freeze_queue(struct request_queue *q)
211 * ...just an alias to keep freeze and unfreeze actions balanced
212 * in the blk_mq_* namespace
216 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
218 void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
220 mutex_lock(&q->mq_freeze_lock);
222 q->q_usage_counter.data->force_atomic = true;
223 q->mq_freeze_depth--;
224 WARN_ON_ONCE(q->mq_freeze_depth < 0);
225 if (!q->mq_freeze_depth) {
226 percpu_ref_resurrect(&q->q_usage_counter);
227 wake_up_all(&q->mq_freeze_wq);
229 mutex_unlock(&q->mq_freeze_lock);
232 void blk_mq_unfreeze_queue(struct request_queue *q)
234 __blk_mq_unfreeze_queue(q, false);
236 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
239 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
240 * mpt3sas driver such that this function can be removed.
242 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
246 spin_lock_irqsave(&q->queue_lock, flags);
247 if (!q->quiesce_depth++)
248 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
249 spin_unlock_irqrestore(&q->queue_lock, flags);
251 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
254 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
257 * Note: it is driver's responsibility for making sure that quiesce has
260 void blk_mq_wait_quiesce_done(struct request_queue *q)
262 if (blk_queue_has_srcu(q))
263 synchronize_srcu(q->srcu);
267 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
270 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
273 * Note: this function does not prevent that the struct request end_io()
274 * callback function is invoked. Once this function is returned, we make
275 * sure no dispatch can happen until the queue is unquiesced via
276 * blk_mq_unquiesce_queue().
278 void blk_mq_quiesce_queue(struct request_queue *q)
280 blk_mq_quiesce_queue_nowait(q);
281 blk_mq_wait_quiesce_done(q);
283 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
286 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
289 * This function recovers queue into the state before quiescing
290 * which is done by blk_mq_quiesce_queue.
292 void blk_mq_unquiesce_queue(struct request_queue *q)
295 bool run_queue = false;
297 spin_lock_irqsave(&q->queue_lock, flags);
298 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
300 } else if (!--q->quiesce_depth) {
301 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
304 spin_unlock_irqrestore(&q->queue_lock, flags);
306 /* dispatch requests which are inserted during quiescing */
308 blk_mq_run_hw_queues(q, true);
310 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
312 void blk_mq_wake_waiters(struct request_queue *q)
314 struct blk_mq_hw_ctx *hctx;
317 queue_for_each_hw_ctx(q, hctx, i)
318 if (blk_mq_hw_queue_mapped(hctx))
319 blk_mq_tag_wakeup_all(hctx->tags, true);
322 void blk_rq_init(struct request_queue *q, struct request *rq)
324 memset(rq, 0, sizeof(*rq));
326 INIT_LIST_HEAD(&rq->queuelist);
328 rq->__sector = (sector_t) -1;
329 INIT_HLIST_NODE(&rq->hash);
330 RB_CLEAR_NODE(&rq->rb_node);
331 rq->tag = BLK_MQ_NO_TAG;
332 rq->internal_tag = BLK_MQ_NO_TAG;
333 rq->start_time_ns = ktime_get_ns();
335 blk_crypto_rq_set_defaults(rq);
337 EXPORT_SYMBOL(blk_rq_init);
339 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
340 struct blk_mq_tags *tags, unsigned int tag, u64 alloc_time_ns)
342 struct blk_mq_ctx *ctx = data->ctx;
343 struct blk_mq_hw_ctx *hctx = data->hctx;
344 struct request_queue *q = data->q;
345 struct request *rq = tags->static_rqs[tag];
350 rq->cmd_flags = data->cmd_flags;
352 if (data->flags & BLK_MQ_REQ_PM)
353 data->rq_flags |= RQF_PM;
354 if (blk_queue_io_stat(q))
355 data->rq_flags |= RQF_IO_STAT;
356 rq->rq_flags = data->rq_flags;
358 if (!(data->rq_flags & RQF_ELV)) {
360 rq->internal_tag = BLK_MQ_NO_TAG;
362 rq->tag = BLK_MQ_NO_TAG;
363 rq->internal_tag = tag;
367 if (blk_mq_need_time_stamp(rq))
368 rq->start_time_ns = ktime_get_ns();
370 rq->start_time_ns = 0;
372 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
373 rq->alloc_time_ns = alloc_time_ns;
375 rq->io_start_time_ns = 0;
376 rq->stats_sectors = 0;
377 rq->nr_phys_segments = 0;
378 #if defined(CONFIG_BLK_DEV_INTEGRITY)
379 rq->nr_integrity_segments = 0;
382 rq->end_io_data = NULL;
384 blk_crypto_rq_set_defaults(rq);
385 INIT_LIST_HEAD(&rq->queuelist);
386 /* tag was already set */
387 WRITE_ONCE(rq->deadline, 0);
390 if (rq->rq_flags & RQF_ELV) {
391 struct elevator_queue *e = data->q->elevator;
393 INIT_HLIST_NODE(&rq->hash);
394 RB_CLEAR_NODE(&rq->rb_node);
396 if (!op_is_flush(data->cmd_flags) &&
397 e->type->ops.prepare_request) {
398 e->type->ops.prepare_request(rq);
399 rq->rq_flags |= RQF_ELVPRIV;
406 static inline struct request *
407 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data,
410 unsigned int tag, tag_offset;
411 struct blk_mq_tags *tags;
413 unsigned long tag_mask;
416 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
417 if (unlikely(!tag_mask))
420 tags = blk_mq_tags_from_data(data);
421 for (i = 0; tag_mask; i++) {
422 if (!(tag_mask & (1UL << i)))
424 tag = tag_offset + i;
425 prefetch(tags->static_rqs[tag]);
426 tag_mask &= ~(1UL << i);
427 rq = blk_mq_rq_ctx_init(data, tags, tag, alloc_time_ns);
428 rq_list_add(data->cached_rq, rq);
431 /* caller already holds a reference, add for remainder */
432 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
435 return rq_list_pop(data->cached_rq);
438 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
440 struct request_queue *q = data->q;
441 u64 alloc_time_ns = 0;
445 /* alloc_time includes depth and tag waits */
446 if (blk_queue_rq_alloc_time(q))
447 alloc_time_ns = ktime_get_ns();
449 if (data->cmd_flags & REQ_NOWAIT)
450 data->flags |= BLK_MQ_REQ_NOWAIT;
453 struct elevator_queue *e = q->elevator;
455 data->rq_flags |= RQF_ELV;
458 * Flush/passthrough requests are special and go directly to the
459 * dispatch list. Don't include reserved tags in the
460 * limiting, as it isn't useful.
462 if (!op_is_flush(data->cmd_flags) &&
463 !blk_op_is_passthrough(data->cmd_flags) &&
464 e->type->ops.limit_depth &&
465 !(data->flags & BLK_MQ_REQ_RESERVED))
466 e->type->ops.limit_depth(data->cmd_flags, data);
470 data->ctx = blk_mq_get_ctx(q);
471 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
472 if (!(data->rq_flags & RQF_ELV))
473 blk_mq_tag_busy(data->hctx);
476 * Try batched alloc if we want more than 1 tag.
478 if (data->nr_tags > 1) {
479 rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns);
486 * Waiting allocations only fail because of an inactive hctx. In that
487 * case just retry the hctx assignment and tag allocation as CPU hotplug
488 * should have migrated us to an online CPU by now.
490 tag = blk_mq_get_tag(data);
491 if (tag == BLK_MQ_NO_TAG) {
492 if (data->flags & BLK_MQ_REQ_NOWAIT)
495 * Give up the CPU and sleep for a random short time to
496 * ensure that thread using a realtime scheduling class
497 * are migrated off the CPU, and thus off the hctx that
504 return blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag,
508 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
509 blk_mq_req_flags_t flags)
511 struct blk_mq_alloc_data data = {
520 ret = blk_queue_enter(q, flags);
524 rq = __blk_mq_alloc_requests(&data);
528 rq->__sector = (sector_t) -1;
529 rq->bio = rq->biotail = NULL;
533 return ERR_PTR(-EWOULDBLOCK);
535 EXPORT_SYMBOL(blk_mq_alloc_request);
537 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
538 unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
540 struct blk_mq_alloc_data data = {
546 u64 alloc_time_ns = 0;
551 /* alloc_time includes depth and tag waits */
552 if (blk_queue_rq_alloc_time(q))
553 alloc_time_ns = ktime_get_ns();
556 * If the tag allocator sleeps we could get an allocation for a
557 * different hardware context. No need to complicate the low level
558 * allocator for this for the rare use case of a command tied to
561 if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED))))
562 return ERR_PTR(-EINVAL);
564 if (hctx_idx >= q->nr_hw_queues)
565 return ERR_PTR(-EIO);
567 ret = blk_queue_enter(q, flags);
572 * Check if the hardware context is actually mapped to anything.
573 * If not tell the caller that it should skip this queue.
576 data.hctx = q->queue_hw_ctx[hctx_idx];
577 if (!blk_mq_hw_queue_mapped(data.hctx))
579 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
580 data.ctx = __blk_mq_get_ctx(q, cpu);
583 blk_mq_tag_busy(data.hctx);
585 data.rq_flags |= RQF_ELV;
588 tag = blk_mq_get_tag(&data);
589 if (tag == BLK_MQ_NO_TAG)
591 return blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag,
598 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
600 static void __blk_mq_free_request(struct request *rq)
602 struct request_queue *q = rq->q;
603 struct blk_mq_ctx *ctx = rq->mq_ctx;
604 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
605 const int sched_tag = rq->internal_tag;
607 blk_crypto_free_request(rq);
608 blk_pm_mark_last_busy(rq);
610 if (rq->tag != BLK_MQ_NO_TAG)
611 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
612 if (sched_tag != BLK_MQ_NO_TAG)
613 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
614 blk_mq_sched_restart(hctx);
618 void blk_mq_free_request(struct request *rq)
620 struct request_queue *q = rq->q;
621 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
623 if ((rq->rq_flags & RQF_ELVPRIV) &&
624 q->elevator->type->ops.finish_request)
625 q->elevator->type->ops.finish_request(rq);
627 if (rq->rq_flags & RQF_MQ_INFLIGHT)
628 __blk_mq_dec_active_requests(hctx);
630 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
631 laptop_io_completion(q->disk->bdi);
635 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
636 if (req_ref_put_and_test(rq))
637 __blk_mq_free_request(rq);
639 EXPORT_SYMBOL_GPL(blk_mq_free_request);
641 void blk_mq_free_plug_rqs(struct blk_plug *plug)
645 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
646 blk_mq_free_request(rq);
649 void blk_dump_rq_flags(struct request *rq, char *msg)
651 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
652 rq->q->disk ? rq->q->disk->disk_name : "?",
653 (unsigned long long) rq->cmd_flags);
655 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
656 (unsigned long long)blk_rq_pos(rq),
657 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
658 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
659 rq->bio, rq->biotail, blk_rq_bytes(rq));
661 EXPORT_SYMBOL(blk_dump_rq_flags);
663 static void req_bio_endio(struct request *rq, struct bio *bio,
664 unsigned int nbytes, blk_status_t error)
666 if (unlikely(error)) {
667 bio->bi_status = error;
668 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
670 * Partial zone append completions cannot be supported as the
671 * BIO fragments may end up not being written sequentially.
673 if (bio->bi_iter.bi_size != nbytes)
674 bio->bi_status = BLK_STS_IOERR;
676 bio->bi_iter.bi_sector = rq->__sector;
679 bio_advance(bio, nbytes);
681 if (unlikely(rq->rq_flags & RQF_QUIET))
682 bio_set_flag(bio, BIO_QUIET);
683 /* don't actually finish bio if it's part of flush sequence */
684 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
688 static void blk_account_io_completion(struct request *req, unsigned int bytes)
690 if (req->part && blk_do_io_stat(req)) {
691 const int sgrp = op_stat_group(req_op(req));
694 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
699 static void blk_print_req_error(struct request *req, blk_status_t status)
701 printk_ratelimited(KERN_ERR
702 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
703 "phys_seg %u prio class %u\n",
704 blk_status_to_str(status),
705 req->q->disk ? req->q->disk->disk_name : "?",
706 blk_rq_pos(req), req_op(req), blk_op_str(req_op(req)),
707 req->cmd_flags & ~REQ_OP_MASK,
708 req->nr_phys_segments,
709 IOPRIO_PRIO_CLASS(req->ioprio));
713 * Fully end IO on a request. Does not support partial completions, or
716 static void blk_complete_request(struct request *req)
718 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
719 int total_bytes = blk_rq_bytes(req);
720 struct bio *bio = req->bio;
722 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
727 #ifdef CONFIG_BLK_DEV_INTEGRITY
728 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
729 req->q->integrity.profile->complete_fn(req, total_bytes);
732 blk_account_io_completion(req, total_bytes);
735 struct bio *next = bio->bi_next;
737 /* Completion has already been traced */
738 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
740 if (req_op(req) == REQ_OP_ZONE_APPEND)
741 bio->bi_iter.bi_sector = req->__sector;
749 * Reset counters so that the request stacking driver
750 * can find how many bytes remain in the request
758 * blk_update_request - Complete multiple bytes without completing the request
759 * @req: the request being processed
760 * @error: block status code
761 * @nr_bytes: number of bytes to complete for @req
764 * Ends I/O on a number of bytes attached to @req, but doesn't complete
765 * the request structure even if @req doesn't have leftover.
766 * If @req has leftover, sets it up for the next range of segments.
768 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
769 * %false return from this function.
772 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
773 * except in the consistency check at the end of this function.
776 * %false - this request doesn't have any more data
777 * %true - this request has more data
779 bool blk_update_request(struct request *req, blk_status_t error,
780 unsigned int nr_bytes)
784 trace_block_rq_complete(req, error, nr_bytes);
789 #ifdef CONFIG_BLK_DEV_INTEGRITY
790 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
792 req->q->integrity.profile->complete_fn(req, nr_bytes);
795 if (unlikely(error && !blk_rq_is_passthrough(req) &&
796 !(req->rq_flags & RQF_QUIET)))
797 blk_print_req_error(req, error);
799 blk_account_io_completion(req, nr_bytes);
803 struct bio *bio = req->bio;
804 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
806 if (bio_bytes == bio->bi_iter.bi_size)
807 req->bio = bio->bi_next;
809 /* Completion has already been traced */
810 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
811 req_bio_endio(req, bio, bio_bytes, error);
813 total_bytes += bio_bytes;
814 nr_bytes -= bio_bytes;
825 * Reset counters so that the request stacking driver
826 * can find how many bytes remain in the request
833 req->__data_len -= total_bytes;
835 /* update sector only for requests with clear definition of sector */
836 if (!blk_rq_is_passthrough(req))
837 req->__sector += total_bytes >> 9;
839 /* mixed attributes always follow the first bio */
840 if (req->rq_flags & RQF_MIXED_MERGE) {
841 req->cmd_flags &= ~REQ_FAILFAST_MASK;
842 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
845 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
847 * If total number of sectors is less than the first segment
848 * size, something has gone terribly wrong.
850 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
851 blk_dump_rq_flags(req, "request botched");
852 req->__data_len = blk_rq_cur_bytes(req);
855 /* recalculate the number of segments */
856 req->nr_phys_segments = blk_recalc_rq_segments(req);
861 EXPORT_SYMBOL_GPL(blk_update_request);
863 static void __blk_account_io_done(struct request *req, u64 now)
865 const int sgrp = op_stat_group(req_op(req));
868 update_io_ticks(req->part, jiffies, true);
869 part_stat_inc(req->part, ios[sgrp]);
870 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
874 static inline void blk_account_io_done(struct request *req, u64 now)
877 * Account IO completion. flush_rq isn't accounted as a
878 * normal IO on queueing nor completion. Accounting the
879 * containing request is enough.
881 if (blk_do_io_stat(req) && req->part &&
882 !(req->rq_flags & RQF_FLUSH_SEQ))
883 __blk_account_io_done(req, now);
886 static void __blk_account_io_start(struct request *rq)
888 /* passthrough requests can hold bios that do not have ->bi_bdev set */
889 if (rq->bio && rq->bio->bi_bdev)
890 rq->part = rq->bio->bi_bdev;
891 else if (rq->q->disk)
892 rq->part = rq->q->disk->part0;
895 update_io_ticks(rq->part, jiffies, false);
899 static inline void blk_account_io_start(struct request *req)
901 if (blk_do_io_stat(req))
902 __blk_account_io_start(req);
905 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
907 if (rq->rq_flags & RQF_STATS) {
908 blk_mq_poll_stats_start(rq->q);
909 blk_stat_add(rq, now);
912 blk_mq_sched_completed_request(rq, now);
913 blk_account_io_done(rq, now);
916 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
918 if (blk_mq_need_time_stamp(rq))
919 __blk_mq_end_request_acct(rq, ktime_get_ns());
922 rq_qos_done(rq->q, rq);
923 rq->end_io(rq, error);
925 blk_mq_free_request(rq);
928 EXPORT_SYMBOL(__blk_mq_end_request);
930 void blk_mq_end_request(struct request *rq, blk_status_t error)
932 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
934 __blk_mq_end_request(rq, error);
936 EXPORT_SYMBOL(blk_mq_end_request);
938 #define TAG_COMP_BATCH 32
940 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
941 int *tag_array, int nr_tags)
943 struct request_queue *q = hctx->queue;
946 * All requests should have been marked as RQF_MQ_INFLIGHT, so
947 * update hctx->nr_active in batch
949 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
950 __blk_mq_sub_active_requests(hctx, nr_tags);
952 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
953 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
956 void blk_mq_end_request_batch(struct io_comp_batch *iob)
958 int tags[TAG_COMP_BATCH], nr_tags = 0;
959 struct blk_mq_hw_ctx *cur_hctx = NULL;
964 now = ktime_get_ns();
966 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
968 prefetch(rq->rq_next);
970 blk_complete_request(rq);
972 __blk_mq_end_request_acct(rq, now);
974 rq_qos_done(rq->q, rq);
976 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
977 if (!req_ref_put_and_test(rq))
980 blk_crypto_free_request(rq);
981 blk_pm_mark_last_busy(rq);
983 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
985 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
987 cur_hctx = rq->mq_hctx;
989 tags[nr_tags++] = rq->tag;
993 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
995 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
997 static void blk_complete_reqs(struct llist_head *list)
999 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1000 struct request *rq, *next;
1002 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1003 rq->q->mq_ops->complete(rq);
1006 static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1008 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1011 static int blk_softirq_cpu_dead(unsigned int cpu)
1013 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1017 static void __blk_mq_complete_request_remote(void *data)
1019 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1022 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1024 int cpu = raw_smp_processor_id();
1026 if (!IS_ENABLED(CONFIG_SMP) ||
1027 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1030 * With force threaded interrupts enabled, raising softirq from an SMP
1031 * function call will always result in waking the ksoftirqd thread.
1032 * This is probably worse than completing the request on a different
1035 if (force_irqthreads())
1038 /* same CPU or cache domain? Complete locally */
1039 if (cpu == rq->mq_ctx->cpu ||
1040 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1041 cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1044 /* don't try to IPI to an offline CPU */
1045 return cpu_online(rq->mq_ctx->cpu);
1048 static void blk_mq_complete_send_ipi(struct request *rq)
1050 struct llist_head *list;
1053 cpu = rq->mq_ctx->cpu;
1054 list = &per_cpu(blk_cpu_done, cpu);
1055 if (llist_add(&rq->ipi_list, list)) {
1056 INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1057 smp_call_function_single_async(cpu, &rq->csd);
1061 static void blk_mq_raise_softirq(struct request *rq)
1063 struct llist_head *list;
1066 list = this_cpu_ptr(&blk_cpu_done);
1067 if (llist_add(&rq->ipi_list, list))
1068 raise_softirq(BLOCK_SOFTIRQ);
1072 bool blk_mq_complete_request_remote(struct request *rq)
1074 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1077 * For a polled request, always complete locallly, it's pointless
1078 * to redirect the completion.
1080 if (rq->cmd_flags & REQ_POLLED)
1083 if (blk_mq_complete_need_ipi(rq)) {
1084 blk_mq_complete_send_ipi(rq);
1088 if (rq->q->nr_hw_queues == 1) {
1089 blk_mq_raise_softirq(rq);
1094 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1097 * blk_mq_complete_request - end I/O on a request
1098 * @rq: the request being processed
1101 * Complete a request by scheduling the ->complete_rq operation.
1103 void blk_mq_complete_request(struct request *rq)
1105 if (!blk_mq_complete_request_remote(rq))
1106 rq->q->mq_ops->complete(rq);
1108 EXPORT_SYMBOL(blk_mq_complete_request);
1111 * blk_mq_start_request - Start processing a request
1112 * @rq: Pointer to request to be started
1114 * Function used by device drivers to notify the block layer that a request
1115 * is going to be processed now, so blk layer can do proper initializations
1116 * such as starting the timeout timer.
1118 void blk_mq_start_request(struct request *rq)
1120 struct request_queue *q = rq->q;
1122 trace_block_rq_issue(rq);
1124 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1126 #ifdef CONFIG_BLK_CGROUP
1128 start_time = bio_issue_time(&rq->bio->bi_issue);
1131 start_time = ktime_get_ns();
1132 rq->io_start_time_ns = start_time;
1133 rq->stats_sectors = blk_rq_sectors(rq);
1134 rq->rq_flags |= RQF_STATS;
1135 rq_qos_issue(q, rq);
1138 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1141 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1143 #ifdef CONFIG_BLK_DEV_INTEGRITY
1144 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1145 q->integrity.profile->prepare_fn(rq);
1147 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1148 WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq));
1150 EXPORT_SYMBOL(blk_mq_start_request);
1153 * blk_end_sync_rq - executes a completion event on a request
1154 * @rq: request to complete
1155 * @error: end I/O status of the request
1157 static void blk_end_sync_rq(struct request *rq, blk_status_t error)
1159 struct completion *waiting = rq->end_io_data;
1161 rq->end_io_data = (void *)(uintptr_t)error;
1164 * complete last, if this is a stack request the process (and thus
1165 * the rq pointer) could be invalid right after this complete()
1171 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1172 * @rq: request to insert
1173 * @at_head: insert request at head or tail of queue
1174 * @done: I/O completion handler
1177 * Insert a fully prepared request at the back of the I/O scheduler queue
1178 * for execution. Don't wait for completion.
1181 * This function will invoke @done directly if the queue is dead.
1183 void blk_execute_rq_nowait(struct request *rq, bool at_head, rq_end_io_fn *done)
1185 WARN_ON(irqs_disabled());
1186 WARN_ON(!blk_rq_is_passthrough(rq));
1190 blk_account_io_start(rq);
1193 * don't check dying flag for MQ because the request won't
1194 * be reused after dying flag is set
1196 blk_mq_sched_insert_request(rq, at_head, true, false);
1198 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1200 static bool blk_rq_is_poll(struct request *rq)
1204 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1206 if (WARN_ON_ONCE(!rq->bio))
1211 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1214 bio_poll(rq->bio, NULL, 0);
1216 } while (!completion_done(wait));
1220 * blk_execute_rq - insert a request into queue for execution
1221 * @rq: request to insert
1222 * @at_head: insert request at head or tail of queue
1225 * Insert a fully prepared request at the back of the I/O scheduler queue
1226 * for execution and wait for completion.
1227 * Return: The blk_status_t result provided to blk_mq_end_request().
1229 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1231 DECLARE_COMPLETION_ONSTACK(wait);
1232 unsigned long hang_check;
1234 rq->end_io_data = &wait;
1235 blk_execute_rq_nowait(rq, at_head, blk_end_sync_rq);
1237 /* Prevent hang_check timer from firing at us during very long I/O */
1238 hang_check = sysctl_hung_task_timeout_secs;
1240 if (blk_rq_is_poll(rq))
1241 blk_rq_poll_completion(rq, &wait);
1242 else if (hang_check)
1243 while (!wait_for_completion_io_timeout(&wait,
1244 hang_check * (HZ/2)))
1247 wait_for_completion_io(&wait);
1249 return (blk_status_t)(uintptr_t)rq->end_io_data;
1251 EXPORT_SYMBOL(blk_execute_rq);
1253 static void __blk_mq_requeue_request(struct request *rq)
1255 struct request_queue *q = rq->q;
1257 blk_mq_put_driver_tag(rq);
1259 trace_block_rq_requeue(rq);
1260 rq_qos_requeue(q, rq);
1262 if (blk_mq_request_started(rq)) {
1263 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1264 rq->rq_flags &= ~RQF_TIMED_OUT;
1268 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1270 __blk_mq_requeue_request(rq);
1272 /* this request will be re-inserted to io scheduler queue */
1273 blk_mq_sched_requeue_request(rq);
1275 blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
1277 EXPORT_SYMBOL(blk_mq_requeue_request);
1279 static void blk_mq_requeue_work(struct work_struct *work)
1281 struct request_queue *q =
1282 container_of(work, struct request_queue, requeue_work.work);
1284 struct request *rq, *next;
1286 spin_lock_irq(&q->requeue_lock);
1287 list_splice_init(&q->requeue_list, &rq_list);
1288 spin_unlock_irq(&q->requeue_lock);
1290 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
1291 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
1294 rq->rq_flags &= ~RQF_SOFTBARRIER;
1295 list_del_init(&rq->queuelist);
1297 * If RQF_DONTPREP, rq has contained some driver specific
1298 * data, so insert it to hctx dispatch list to avoid any
1301 if (rq->rq_flags & RQF_DONTPREP)
1302 blk_mq_request_bypass_insert(rq, false, false);
1304 blk_mq_sched_insert_request(rq, true, false, false);
1307 while (!list_empty(&rq_list)) {
1308 rq = list_entry(rq_list.next, struct request, queuelist);
1309 list_del_init(&rq->queuelist);
1310 blk_mq_sched_insert_request(rq, false, false, false);
1313 blk_mq_run_hw_queues(q, false);
1316 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
1317 bool kick_requeue_list)
1319 struct request_queue *q = rq->q;
1320 unsigned long flags;
1323 * We abuse this flag that is otherwise used by the I/O scheduler to
1324 * request head insertion from the workqueue.
1326 BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
1328 spin_lock_irqsave(&q->requeue_lock, flags);
1330 rq->rq_flags |= RQF_SOFTBARRIER;
1331 list_add(&rq->queuelist, &q->requeue_list);
1333 list_add_tail(&rq->queuelist, &q->requeue_list);
1335 spin_unlock_irqrestore(&q->requeue_lock, flags);
1337 if (kick_requeue_list)
1338 blk_mq_kick_requeue_list(q);
1341 void blk_mq_kick_requeue_list(struct request_queue *q)
1343 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1345 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1347 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1348 unsigned long msecs)
1350 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1351 msecs_to_jiffies(msecs));
1353 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1355 static bool blk_mq_rq_inflight(struct request *rq, void *priv,
1359 * If we find a request that isn't idle we know the queue is busy
1360 * as it's checked in the iter.
1361 * Return false to stop the iteration.
1363 if (blk_mq_request_started(rq)) {
1373 bool blk_mq_queue_inflight(struct request_queue *q)
1377 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1380 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1382 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
1384 req->rq_flags |= RQF_TIMED_OUT;
1385 if (req->q->mq_ops->timeout) {
1386 enum blk_eh_timer_return ret;
1388 ret = req->q->mq_ops->timeout(req, reserved);
1389 if (ret == BLK_EH_DONE)
1391 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1397 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
1399 unsigned long deadline;
1401 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1403 if (rq->rq_flags & RQF_TIMED_OUT)
1406 deadline = READ_ONCE(rq->deadline);
1407 if (time_after_eq(jiffies, deadline))
1412 else if (time_after(*next, deadline))
1417 void blk_mq_put_rq_ref(struct request *rq)
1419 if (is_flush_rq(rq))
1421 else if (req_ref_put_and_test(rq))
1422 __blk_mq_free_request(rq);
1425 static bool blk_mq_check_expired(struct request *rq, void *priv, bool reserved)
1427 unsigned long *next = priv;
1430 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1431 * be reallocated underneath the timeout handler's processing, then
1432 * the expire check is reliable. If the request is not expired, then
1433 * it was completed and reallocated as a new request after returning
1434 * from blk_mq_check_expired().
1436 if (blk_mq_req_expired(rq, next))
1437 blk_mq_rq_timed_out(rq, reserved);
1441 static void blk_mq_timeout_work(struct work_struct *work)
1443 struct request_queue *q =
1444 container_of(work, struct request_queue, timeout_work);
1445 unsigned long next = 0;
1446 struct blk_mq_hw_ctx *hctx;
1449 /* A deadlock might occur if a request is stuck requiring a
1450 * timeout at the same time a queue freeze is waiting
1451 * completion, since the timeout code would not be able to
1452 * acquire the queue reference here.
1454 * That's why we don't use blk_queue_enter here; instead, we use
1455 * percpu_ref_tryget directly, because we need to be able to
1456 * obtain a reference even in the short window between the queue
1457 * starting to freeze, by dropping the first reference in
1458 * blk_freeze_queue_start, and the moment the last request is
1459 * consumed, marked by the instant q_usage_counter reaches
1462 if (!percpu_ref_tryget(&q->q_usage_counter))
1465 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
1468 mod_timer(&q->timeout, next);
1471 * Request timeouts are handled as a forward rolling timer. If
1472 * we end up here it means that no requests are pending and
1473 * also that no request has been pending for a while. Mark
1474 * each hctx as idle.
1476 queue_for_each_hw_ctx(q, hctx, i) {
1477 /* the hctx may be unmapped, so check it here */
1478 if (blk_mq_hw_queue_mapped(hctx))
1479 blk_mq_tag_idle(hctx);
1485 struct flush_busy_ctx_data {
1486 struct blk_mq_hw_ctx *hctx;
1487 struct list_head *list;
1490 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1492 struct flush_busy_ctx_data *flush_data = data;
1493 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1494 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1495 enum hctx_type type = hctx->type;
1497 spin_lock(&ctx->lock);
1498 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1499 sbitmap_clear_bit(sb, bitnr);
1500 spin_unlock(&ctx->lock);
1505 * Process software queues that have been marked busy, splicing them
1506 * to the for-dispatch
1508 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1510 struct flush_busy_ctx_data data = {
1515 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1517 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1519 struct dispatch_rq_data {
1520 struct blk_mq_hw_ctx *hctx;
1524 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1527 struct dispatch_rq_data *dispatch_data = data;
1528 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1529 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1530 enum hctx_type type = hctx->type;
1532 spin_lock(&ctx->lock);
1533 if (!list_empty(&ctx->rq_lists[type])) {
1534 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1535 list_del_init(&dispatch_data->rq->queuelist);
1536 if (list_empty(&ctx->rq_lists[type]))
1537 sbitmap_clear_bit(sb, bitnr);
1539 spin_unlock(&ctx->lock);
1541 return !dispatch_data->rq;
1544 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1545 struct blk_mq_ctx *start)
1547 unsigned off = start ? start->index_hw[hctx->type] : 0;
1548 struct dispatch_rq_data data = {
1553 __sbitmap_for_each_set(&hctx->ctx_map, off,
1554 dispatch_rq_from_ctx, &data);
1559 static bool __blk_mq_alloc_driver_tag(struct request *rq)
1561 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1562 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1565 blk_mq_tag_busy(rq->mq_hctx);
1567 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1568 bt = &rq->mq_hctx->tags->breserved_tags;
1571 if (!hctx_may_queue(rq->mq_hctx, bt))
1575 tag = __sbitmap_queue_get(bt);
1576 if (tag == BLK_MQ_NO_TAG)
1579 rq->tag = tag + tag_offset;
1583 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1585 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1588 if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1589 !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1590 rq->rq_flags |= RQF_MQ_INFLIGHT;
1591 __blk_mq_inc_active_requests(hctx);
1593 hctx->tags->rqs[rq->tag] = rq;
1597 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1598 int flags, void *key)
1600 struct blk_mq_hw_ctx *hctx;
1602 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1604 spin_lock(&hctx->dispatch_wait_lock);
1605 if (!list_empty(&wait->entry)) {
1606 struct sbitmap_queue *sbq;
1608 list_del_init(&wait->entry);
1609 sbq = &hctx->tags->bitmap_tags;
1610 atomic_dec(&sbq->ws_active);
1612 spin_unlock(&hctx->dispatch_wait_lock);
1614 blk_mq_run_hw_queue(hctx, true);
1619 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1620 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1621 * restart. For both cases, take care to check the condition again after
1622 * marking us as waiting.
1624 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1627 struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1628 struct wait_queue_head *wq;
1629 wait_queue_entry_t *wait;
1632 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
1633 blk_mq_sched_mark_restart_hctx(hctx);
1636 * It's possible that a tag was freed in the window between the
1637 * allocation failure and adding the hardware queue to the wait
1640 * Don't clear RESTART here, someone else could have set it.
1641 * At most this will cost an extra queue run.
1643 return blk_mq_get_driver_tag(rq);
1646 wait = &hctx->dispatch_wait;
1647 if (!list_empty_careful(&wait->entry))
1650 wq = &bt_wait_ptr(sbq, hctx)->wait;
1652 spin_lock_irq(&wq->lock);
1653 spin_lock(&hctx->dispatch_wait_lock);
1654 if (!list_empty(&wait->entry)) {
1655 spin_unlock(&hctx->dispatch_wait_lock);
1656 spin_unlock_irq(&wq->lock);
1660 atomic_inc(&sbq->ws_active);
1661 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1662 __add_wait_queue(wq, wait);
1665 * It's possible that a tag was freed in the window between the
1666 * allocation failure and adding the hardware queue to the wait
1669 ret = blk_mq_get_driver_tag(rq);
1671 spin_unlock(&hctx->dispatch_wait_lock);
1672 spin_unlock_irq(&wq->lock);
1677 * We got a tag, remove ourselves from the wait queue to ensure
1678 * someone else gets the wakeup.
1680 list_del_init(&wait->entry);
1681 atomic_dec(&sbq->ws_active);
1682 spin_unlock(&hctx->dispatch_wait_lock);
1683 spin_unlock_irq(&wq->lock);
1688 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1689 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1691 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1692 * - EWMA is one simple way to compute running average value
1693 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1694 * - take 4 as factor for avoiding to get too small(0) result, and this
1695 * factor doesn't matter because EWMA decreases exponentially
1697 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1701 ewma = hctx->dispatch_busy;
1706 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1708 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1709 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1711 hctx->dispatch_busy = ewma;
1714 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1716 static void blk_mq_handle_dev_resource(struct request *rq,
1717 struct list_head *list)
1719 struct request *next =
1720 list_first_entry_or_null(list, struct request, queuelist);
1723 * If an I/O scheduler has been configured and we got a driver tag for
1724 * the next request already, free it.
1727 blk_mq_put_driver_tag(next);
1729 list_add(&rq->queuelist, list);
1730 __blk_mq_requeue_request(rq);
1733 static void blk_mq_handle_zone_resource(struct request *rq,
1734 struct list_head *zone_list)
1737 * If we end up here it is because we cannot dispatch a request to a
1738 * specific zone due to LLD level zone-write locking or other zone
1739 * related resource not being available. In this case, set the request
1740 * aside in zone_list for retrying it later.
1742 list_add(&rq->queuelist, zone_list);
1743 __blk_mq_requeue_request(rq);
1746 enum prep_dispatch {
1748 PREP_DISPATCH_NO_TAG,
1749 PREP_DISPATCH_NO_BUDGET,
1752 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1755 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1756 int budget_token = -1;
1759 budget_token = blk_mq_get_dispatch_budget(rq->q);
1760 if (budget_token < 0) {
1761 blk_mq_put_driver_tag(rq);
1762 return PREP_DISPATCH_NO_BUDGET;
1764 blk_mq_set_rq_budget_token(rq, budget_token);
1767 if (!blk_mq_get_driver_tag(rq)) {
1769 * The initial allocation attempt failed, so we need to
1770 * rerun the hardware queue when a tag is freed. The
1771 * waitqueue takes care of that. If the queue is run
1772 * before we add this entry back on the dispatch list,
1773 * we'll re-run it below.
1775 if (!blk_mq_mark_tag_wait(hctx, rq)) {
1777 * All budgets not got from this function will be put
1778 * together during handling partial dispatch
1781 blk_mq_put_dispatch_budget(rq->q, budget_token);
1782 return PREP_DISPATCH_NO_TAG;
1786 return PREP_DISPATCH_OK;
1789 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1790 static void blk_mq_release_budgets(struct request_queue *q,
1791 struct list_head *list)
1795 list_for_each_entry(rq, list, queuelist) {
1796 int budget_token = blk_mq_get_rq_budget_token(rq);
1798 if (budget_token >= 0)
1799 blk_mq_put_dispatch_budget(q, budget_token);
1804 * Returns true if we did some work AND can potentially do more.
1806 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1807 unsigned int nr_budgets)
1809 enum prep_dispatch prep;
1810 struct request_queue *q = hctx->queue;
1811 struct request *rq, *nxt;
1813 blk_status_t ret = BLK_STS_OK;
1814 LIST_HEAD(zone_list);
1815 bool needs_resource = false;
1817 if (list_empty(list))
1821 * Now process all the entries, sending them to the driver.
1823 errors = queued = 0;
1825 struct blk_mq_queue_data bd;
1827 rq = list_first_entry(list, struct request, queuelist);
1829 WARN_ON_ONCE(hctx != rq->mq_hctx);
1830 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
1831 if (prep != PREP_DISPATCH_OK)
1834 list_del_init(&rq->queuelist);
1839 * Flag last if we have no more requests, or if we have more
1840 * but can't assign a driver tag to it.
1842 if (list_empty(list))
1845 nxt = list_first_entry(list, struct request, queuelist);
1846 bd.last = !blk_mq_get_driver_tag(nxt);
1850 * once the request is queued to lld, no need to cover the
1855 ret = q->mq_ops->queue_rq(hctx, &bd);
1860 case BLK_STS_RESOURCE:
1861 needs_resource = true;
1863 case BLK_STS_DEV_RESOURCE:
1864 blk_mq_handle_dev_resource(rq, list);
1866 case BLK_STS_ZONE_RESOURCE:
1868 * Move the request to zone_list and keep going through
1869 * the dispatch list to find more requests the drive can
1872 blk_mq_handle_zone_resource(rq, &zone_list);
1873 needs_resource = true;
1877 blk_mq_end_request(rq, ret);
1879 } while (!list_empty(list));
1881 if (!list_empty(&zone_list))
1882 list_splice_tail_init(&zone_list, list);
1884 /* If we didn't flush the entire list, we could have told the driver
1885 * there was more coming, but that turned out to be a lie.
1887 if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued)
1888 q->mq_ops->commit_rqs(hctx);
1890 * Any items that need requeuing? Stuff them into hctx->dispatch,
1891 * that is where we will continue on next queue run.
1893 if (!list_empty(list)) {
1895 /* For non-shared tags, the RESTART check will suffice */
1896 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
1897 (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
1900 blk_mq_release_budgets(q, list);
1902 spin_lock(&hctx->lock);
1903 list_splice_tail_init(list, &hctx->dispatch);
1904 spin_unlock(&hctx->lock);
1907 * Order adding requests to hctx->dispatch and checking
1908 * SCHED_RESTART flag. The pair of this smp_mb() is the one
1909 * in blk_mq_sched_restart(). Avoid restart code path to
1910 * miss the new added requests to hctx->dispatch, meantime
1911 * SCHED_RESTART is observed here.
1916 * If SCHED_RESTART was set by the caller of this function and
1917 * it is no longer set that means that it was cleared by another
1918 * thread and hence that a queue rerun is needed.
1920 * If 'no_tag' is set, that means that we failed getting
1921 * a driver tag with an I/O scheduler attached. If our dispatch
1922 * waitqueue is no longer active, ensure that we run the queue
1923 * AFTER adding our entries back to the list.
1925 * If no I/O scheduler has been configured it is possible that
1926 * the hardware queue got stopped and restarted before requests
1927 * were pushed back onto the dispatch list. Rerun the queue to
1928 * avoid starvation. Notes:
1929 * - blk_mq_run_hw_queue() checks whether or not a queue has
1930 * been stopped before rerunning a queue.
1931 * - Some but not all block drivers stop a queue before
1932 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1935 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1936 * bit is set, run queue after a delay to avoid IO stalls
1937 * that could otherwise occur if the queue is idle. We'll do
1938 * similar if we couldn't get budget or couldn't lock a zone
1939 * and SCHED_RESTART is set.
1941 needs_restart = blk_mq_sched_needs_restart(hctx);
1942 if (prep == PREP_DISPATCH_NO_BUDGET)
1943 needs_resource = true;
1944 if (!needs_restart ||
1945 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1946 blk_mq_run_hw_queue(hctx, true);
1947 else if (needs_restart && needs_resource)
1948 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1950 blk_mq_update_dispatch_busy(hctx, true);
1953 blk_mq_update_dispatch_busy(hctx, false);
1955 return (queued + errors) != 0;
1959 * __blk_mq_run_hw_queue - Run a hardware queue.
1960 * @hctx: Pointer to the hardware queue to run.
1962 * Send pending requests to the hardware.
1964 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1967 * We can't run the queue inline with ints disabled. Ensure that
1968 * we catch bad users of this early.
1970 WARN_ON_ONCE(in_interrupt());
1972 blk_mq_run_dispatch_ops(hctx->queue,
1973 blk_mq_sched_dispatch_requests(hctx));
1976 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1978 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1980 if (cpu >= nr_cpu_ids)
1981 cpu = cpumask_first(hctx->cpumask);
1986 * It'd be great if the workqueue API had a way to pass
1987 * in a mask and had some smarts for more clever placement.
1988 * For now we just round-robin here, switching for every
1989 * BLK_MQ_CPU_WORK_BATCH queued items.
1991 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1994 int next_cpu = hctx->next_cpu;
1996 if (hctx->queue->nr_hw_queues == 1)
1997 return WORK_CPU_UNBOUND;
1999 if (--hctx->next_cpu_batch <= 0) {
2001 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2003 if (next_cpu >= nr_cpu_ids)
2004 next_cpu = blk_mq_first_mapped_cpu(hctx);
2005 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2009 * Do unbound schedule if we can't find a online CPU for this hctx,
2010 * and it should only happen in the path of handling CPU DEAD.
2012 if (!cpu_online(next_cpu)) {
2019 * Make sure to re-select CPU next time once after CPUs
2020 * in hctx->cpumask become online again.
2022 hctx->next_cpu = next_cpu;
2023 hctx->next_cpu_batch = 1;
2024 return WORK_CPU_UNBOUND;
2027 hctx->next_cpu = next_cpu;
2032 * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue.
2033 * @hctx: Pointer to the hardware queue to run.
2034 * @async: If we want to run the queue asynchronously.
2035 * @msecs: Milliseconds of delay to wait before running the queue.
2037 * If !@async, try to run the queue now. Else, run the queue asynchronously and
2038 * with a delay of @msecs.
2040 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
2041 unsigned long msecs)
2043 if (unlikely(blk_mq_hctx_stopped(hctx)))
2046 if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
2047 int cpu = get_cpu();
2048 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
2049 __blk_mq_run_hw_queue(hctx);
2057 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2058 msecs_to_jiffies(msecs));
2062 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2063 * @hctx: Pointer to the hardware queue to run.
2064 * @msecs: Milliseconds of delay to wait before running the queue.
2066 * Run a hardware queue asynchronously with a delay of @msecs.
2068 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2070 __blk_mq_delay_run_hw_queue(hctx, true, msecs);
2072 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2075 * blk_mq_run_hw_queue - Start to run a hardware queue.
2076 * @hctx: Pointer to the hardware queue to run.
2077 * @async: If we want to run the queue asynchronously.
2079 * Check if the request queue is not in a quiesced state and if there are
2080 * pending requests to be sent. If this is true, run the queue to send requests
2083 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2088 * When queue is quiesced, we may be switching io scheduler, or
2089 * updating nr_hw_queues, or other things, and we can't run queue
2090 * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2092 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2095 __blk_mq_run_dispatch_ops(hctx->queue, false,
2096 need_run = !blk_queue_quiesced(hctx->queue) &&
2097 blk_mq_hctx_has_pending(hctx));
2100 __blk_mq_delay_run_hw_queue(hctx, async, 0);
2102 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2105 * Is the request queue handled by an IO scheduler that does not respect
2106 * hardware queues when dispatching?
2108 static bool blk_mq_has_sqsched(struct request_queue *q)
2110 struct elevator_queue *e = q->elevator;
2112 if (e && e->type->ops.dispatch_request &&
2113 !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE))
2119 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2122 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2124 struct blk_mq_hw_ctx *hctx;
2127 * If the IO scheduler does not respect hardware queues when
2128 * dispatching, we just don't bother with multiple HW queues and
2129 * dispatch from hctx for the current CPU since running multiple queues
2130 * just causes lock contention inside the scheduler and pointless cache
2133 hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT,
2134 raw_smp_processor_id());
2135 if (!blk_mq_hctx_stopped(hctx))
2141 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2142 * @q: Pointer to the request queue to run.
2143 * @async: If we want to run the queue asynchronously.
2145 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2147 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2151 if (blk_mq_has_sqsched(q))
2152 sq_hctx = blk_mq_get_sq_hctx(q);
2153 queue_for_each_hw_ctx(q, hctx, i) {
2154 if (blk_mq_hctx_stopped(hctx))
2157 * Dispatch from this hctx either if there's no hctx preferred
2158 * by IO scheduler or if it has requests that bypass the
2161 if (!sq_hctx || sq_hctx == hctx ||
2162 !list_empty_careful(&hctx->dispatch))
2163 blk_mq_run_hw_queue(hctx, async);
2166 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2169 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2170 * @q: Pointer to the request queue to run.
2171 * @msecs: Milliseconds of delay to wait before running the queues.
2173 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2175 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2179 if (blk_mq_has_sqsched(q))
2180 sq_hctx = blk_mq_get_sq_hctx(q);
2181 queue_for_each_hw_ctx(q, hctx, i) {
2182 if (blk_mq_hctx_stopped(hctx))
2185 * Dispatch from this hctx either if there's no hctx preferred
2186 * by IO scheduler or if it has requests that bypass the
2189 if (!sq_hctx || sq_hctx == hctx ||
2190 !list_empty_careful(&hctx->dispatch))
2191 blk_mq_delay_run_hw_queue(hctx, msecs);
2194 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2197 * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
2198 * @q: request queue.
2200 * The caller is responsible for serializing this function against
2201 * blk_mq_{start,stop}_hw_queue().
2203 bool blk_mq_queue_stopped(struct request_queue *q)
2205 struct blk_mq_hw_ctx *hctx;
2208 queue_for_each_hw_ctx(q, hctx, i)
2209 if (blk_mq_hctx_stopped(hctx))
2214 EXPORT_SYMBOL(blk_mq_queue_stopped);
2217 * This function is often used for pausing .queue_rq() by driver when
2218 * there isn't enough resource or some conditions aren't satisfied, and
2219 * BLK_STS_RESOURCE is usually returned.
2221 * We do not guarantee that dispatch can be drained or blocked
2222 * after blk_mq_stop_hw_queue() returns. Please use
2223 * blk_mq_quiesce_queue() for that requirement.
2225 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2227 cancel_delayed_work(&hctx->run_work);
2229 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2231 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2234 * This function is often used for pausing .queue_rq() by driver when
2235 * there isn't enough resource or some conditions aren't satisfied, and
2236 * BLK_STS_RESOURCE is usually returned.
2238 * We do not guarantee that dispatch can be drained or blocked
2239 * after blk_mq_stop_hw_queues() returns. Please use
2240 * blk_mq_quiesce_queue() for that requirement.
2242 void blk_mq_stop_hw_queues(struct request_queue *q)
2244 struct blk_mq_hw_ctx *hctx;
2247 queue_for_each_hw_ctx(q, hctx, i)
2248 blk_mq_stop_hw_queue(hctx);
2250 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2252 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2254 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2256 blk_mq_run_hw_queue(hctx, false);
2258 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2260 void blk_mq_start_hw_queues(struct request_queue *q)
2262 struct blk_mq_hw_ctx *hctx;
2265 queue_for_each_hw_ctx(q, hctx, i)
2266 blk_mq_start_hw_queue(hctx);
2268 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2270 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2272 if (!blk_mq_hctx_stopped(hctx))
2275 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2276 blk_mq_run_hw_queue(hctx, async);
2278 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2280 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2282 struct blk_mq_hw_ctx *hctx;
2285 queue_for_each_hw_ctx(q, hctx, i)
2286 blk_mq_start_stopped_hw_queue(hctx, async);
2288 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2290 static void blk_mq_run_work_fn(struct work_struct *work)
2292 struct blk_mq_hw_ctx *hctx;
2294 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
2297 * If we are stopped, don't run the queue.
2299 if (blk_mq_hctx_stopped(hctx))
2302 __blk_mq_run_hw_queue(hctx);
2305 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
2309 struct blk_mq_ctx *ctx = rq->mq_ctx;
2310 enum hctx_type type = hctx->type;
2312 lockdep_assert_held(&ctx->lock);
2314 trace_block_rq_insert(rq);
2317 list_add(&rq->queuelist, &ctx->rq_lists[type]);
2319 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
2322 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
2325 struct blk_mq_ctx *ctx = rq->mq_ctx;
2327 lockdep_assert_held(&ctx->lock);
2329 __blk_mq_insert_req_list(hctx, rq, at_head);
2330 blk_mq_hctx_mark_pending(hctx, ctx);
2334 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2335 * @rq: Pointer to request to be inserted.
2336 * @at_head: true if the request should be inserted at the head of the list.
2337 * @run_queue: If we should run the hardware queue after inserting the request.
2339 * Should only be used carefully, when the caller knows we want to
2340 * bypass a potential IO scheduler on the target device.
2342 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
2345 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2347 spin_lock(&hctx->lock);
2349 list_add(&rq->queuelist, &hctx->dispatch);
2351 list_add_tail(&rq->queuelist, &hctx->dispatch);
2352 spin_unlock(&hctx->lock);
2355 blk_mq_run_hw_queue(hctx, false);
2358 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
2359 struct list_head *list)
2363 enum hctx_type type = hctx->type;
2366 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2369 list_for_each_entry(rq, list, queuelist) {
2370 BUG_ON(rq->mq_ctx != ctx);
2371 trace_block_rq_insert(rq);
2374 spin_lock(&ctx->lock);
2375 list_splice_tail_init(list, &ctx->rq_lists[type]);
2376 blk_mq_hctx_mark_pending(hctx, ctx);
2377 spin_unlock(&ctx->lock);
2380 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int *queued,
2383 if (hctx->queue->mq_ops->commit_rqs) {
2384 trace_block_unplug(hctx->queue, *queued, !from_schedule);
2385 hctx->queue->mq_ops->commit_rqs(hctx);
2390 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2391 unsigned int nr_segs)
2395 if (bio->bi_opf & REQ_RAHEAD)
2396 rq->cmd_flags |= REQ_FAILFAST_MASK;
2398 rq->__sector = bio->bi_iter.bi_sector;
2399 rq->write_hint = bio->bi_write_hint;
2400 blk_rq_bio_prep(rq, bio, nr_segs);
2402 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2403 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2406 blk_account_io_start(rq);
2409 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2410 struct request *rq, bool last)
2412 struct request_queue *q = rq->q;
2413 struct blk_mq_queue_data bd = {
2420 * For OK queue, we are done. For error, caller may kill it.
2421 * Any other error (busy), just add it to our list as we
2422 * previously would have done.
2424 ret = q->mq_ops->queue_rq(hctx, &bd);
2427 blk_mq_update_dispatch_busy(hctx, false);
2429 case BLK_STS_RESOURCE:
2430 case BLK_STS_DEV_RESOURCE:
2431 blk_mq_update_dispatch_busy(hctx, true);
2432 __blk_mq_requeue_request(rq);
2435 blk_mq_update_dispatch_busy(hctx, false);
2442 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2444 bool bypass_insert, bool last)
2446 struct request_queue *q = rq->q;
2447 bool run_queue = true;
2451 * RCU or SRCU read lock is needed before checking quiesced flag.
2453 * When queue is stopped or quiesced, ignore 'bypass_insert' from
2454 * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
2455 * and avoid driver to try to dispatch again.
2457 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
2459 bypass_insert = false;
2463 if ((rq->rq_flags & RQF_ELV) && !bypass_insert)
2466 budget_token = blk_mq_get_dispatch_budget(q);
2467 if (budget_token < 0)
2470 blk_mq_set_rq_budget_token(rq, budget_token);
2472 if (!blk_mq_get_driver_tag(rq)) {
2473 blk_mq_put_dispatch_budget(q, budget_token);
2477 return __blk_mq_issue_directly(hctx, rq, last);
2480 return BLK_STS_RESOURCE;
2482 blk_mq_sched_insert_request(rq, false, run_queue, false);
2488 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2489 * @hctx: Pointer of the associated hardware queue.
2490 * @rq: Pointer to request to be sent.
2492 * If the device has enough resources to accept a new request now, send the
2493 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2494 * we can try send it another time in the future. Requests inserted at this
2495 * queue have higher priority.
2497 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2501 __blk_mq_try_issue_directly(hctx, rq, false, true);
2503 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
2504 blk_mq_request_bypass_insert(rq, false, true);
2505 else if (ret != BLK_STS_OK)
2506 blk_mq_end_request(rq, ret);
2509 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2511 return __blk_mq_try_issue_directly(rq->mq_hctx, rq, true, last);
2514 static void blk_mq_plug_issue_direct(struct blk_plug *plug, bool from_schedule)
2516 struct blk_mq_hw_ctx *hctx = NULL;
2521 while ((rq = rq_list_pop(&plug->mq_list))) {
2522 bool last = rq_list_empty(plug->mq_list);
2525 if (hctx != rq->mq_hctx) {
2527 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2531 ret = blk_mq_request_issue_directly(rq, last);
2536 case BLK_STS_RESOURCE:
2537 case BLK_STS_DEV_RESOURCE:
2538 blk_mq_request_bypass_insert(rq, false, last);
2539 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2542 blk_mq_end_request(rq, ret);
2549 * If we didn't flush the entire list, we could have told the driver
2550 * there was more coming, but that turned out to be a lie.
2553 blk_mq_commit_rqs(hctx, &queued, from_schedule);
2556 static void __blk_mq_flush_plug_list(struct request_queue *q,
2557 struct blk_plug *plug)
2559 if (blk_queue_quiesced(q))
2561 q->mq_ops->queue_rqs(&plug->mq_list);
2564 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2566 struct blk_mq_hw_ctx *this_hctx;
2567 struct blk_mq_ctx *this_ctx;
2572 if (rq_list_empty(plug->mq_list))
2576 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2577 struct request_queue *q;
2579 rq = rq_list_peek(&plug->mq_list);
2583 * Peek first request and see if we have a ->queue_rqs() hook.
2584 * If we do, we can dispatch the whole plug list in one go. We
2585 * already know at this point that all requests belong to the
2586 * same queue, caller must ensure that's the case.
2588 * Since we pass off the full list to the driver at this point,
2589 * we do not increment the active request count for the queue.
2590 * Bypass shared tags for now because of that.
2592 if (q->mq_ops->queue_rqs &&
2593 !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2594 blk_mq_run_dispatch_ops(q,
2595 __blk_mq_flush_plug_list(q, plug));
2596 if (rq_list_empty(plug->mq_list))
2600 blk_mq_run_dispatch_ops(q,
2601 blk_mq_plug_issue_direct(plug, false));
2602 if (rq_list_empty(plug->mq_list))
2610 rq = rq_list_pop(&plug->mq_list);
2613 this_hctx = rq->mq_hctx;
2614 this_ctx = rq->mq_ctx;
2615 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx) {
2616 trace_block_unplug(this_hctx->queue, depth,
2618 blk_mq_sched_insert_requests(this_hctx, this_ctx,
2619 &list, from_schedule);
2621 this_hctx = rq->mq_hctx;
2622 this_ctx = rq->mq_ctx;
2626 list_add(&rq->queuelist, &list);
2628 } while (!rq_list_empty(plug->mq_list));
2630 if (!list_empty(&list)) {
2631 trace_block_unplug(this_hctx->queue, depth, !from_schedule);
2632 blk_mq_sched_insert_requests(this_hctx, this_ctx, &list,
2637 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2638 struct list_head *list)
2643 while (!list_empty(list)) {
2645 struct request *rq = list_first_entry(list, struct request,
2648 list_del_init(&rq->queuelist);
2649 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2650 if (ret != BLK_STS_OK) {
2651 if (ret == BLK_STS_RESOURCE ||
2652 ret == BLK_STS_DEV_RESOURCE) {
2653 blk_mq_request_bypass_insert(rq, false,
2657 blk_mq_end_request(rq, ret);
2664 * If we didn't flush the entire list, we could have told
2665 * the driver there was more coming, but that turned out to
2668 if ((!list_empty(list) || errors) &&
2669 hctx->queue->mq_ops->commit_rqs && queued)
2670 hctx->queue->mq_ops->commit_rqs(hctx);
2674 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
2675 * queues. This is important for md arrays to benefit from merging
2678 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
2680 if (plug->multiple_queues)
2681 return BLK_MAX_REQUEST_COUNT * 2;
2682 return BLK_MAX_REQUEST_COUNT;
2685 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
2687 struct request *last = rq_list_peek(&plug->mq_list);
2689 if (!plug->rq_count) {
2690 trace_block_plug(rq->q);
2691 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
2692 (!blk_queue_nomerges(rq->q) &&
2693 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2694 blk_mq_flush_plug_list(plug, false);
2695 trace_block_plug(rq->q);
2698 if (!plug->multiple_queues && last && last->q != rq->q)
2699 plug->multiple_queues = true;
2700 if (!plug->has_elevator && (rq->rq_flags & RQF_ELV))
2701 plug->has_elevator = true;
2703 rq_list_add(&plug->mq_list, rq);
2707 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2708 struct bio *bio, unsigned int nr_segs)
2710 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2711 if (blk_attempt_plug_merge(q, bio, nr_segs))
2713 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2719 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2720 struct blk_plug *plug,
2724 struct blk_mq_alloc_data data = {
2727 .cmd_flags = bio->bi_opf,
2731 if (unlikely(bio_queue_enter(bio)))
2734 if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2737 rq_qos_throttle(q, bio);
2740 data.nr_tags = plug->nr_ios;
2742 data.cached_rq = &plug->cached_rq;
2745 rq = __blk_mq_alloc_requests(&data);
2748 rq_qos_cleanup(q, bio);
2749 if (bio->bi_opf & REQ_NOWAIT)
2750 bio_wouldblock_error(bio);
2756 static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2757 struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2763 rq = rq_list_peek(&plug->cached_rq);
2764 if (!rq || rq->q != q)
2767 if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2772 rq_qos_throttle(q, *bio);
2774 if (blk_mq_get_hctx_type((*bio)->bi_opf) != rq->mq_hctx->type)
2776 if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2779 rq->cmd_flags = (*bio)->bi_opf;
2780 plug->cached_rq = rq_list_next(rq);
2781 INIT_LIST_HEAD(&rq->queuelist);
2786 * blk_mq_submit_bio - Create and send a request to block device.
2787 * @bio: Bio pointer.
2789 * Builds up a request structure from @q and @bio and send to the device. The
2790 * request may not be queued directly to hardware if:
2791 * * This request can be merged with another one
2792 * * We want to place request at plug queue for possible future merging
2793 * * There is an IO scheduler active at this queue
2795 * It will not queue the request if there is an error with the bio, or at the
2798 void blk_mq_submit_bio(struct bio *bio)
2800 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2801 struct blk_plug *plug = blk_mq_plug(q, bio);
2802 const int is_sync = op_is_sync(bio->bi_opf);
2804 unsigned int nr_segs = 1;
2807 if (unlikely(!blk_crypto_bio_prep(&bio)))
2810 blk_queue_bounce(q, &bio);
2811 if (blk_may_split(q, bio))
2812 __blk_queue_split(q, &bio, &nr_segs);
2814 if (!bio_integrity_prep(bio))
2817 rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2821 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2826 trace_block_getrq(bio);
2828 rq_qos_track(q, rq, bio);
2830 blk_mq_bio_to_request(rq, bio, nr_segs);
2832 ret = blk_crypto_init_request(rq);
2833 if (ret != BLK_STS_OK) {
2834 bio->bi_status = ret;
2836 blk_mq_free_request(rq);
2840 if (op_is_flush(bio->bi_opf)) {
2841 blk_insert_flush(rq);
2846 blk_add_rq_to_plug(plug, rq);
2847 else if ((rq->rq_flags & RQF_ELV) ||
2848 (rq->mq_hctx->dispatch_busy &&
2849 (q->nr_hw_queues == 1 || !is_sync)))
2850 blk_mq_sched_insert_request(rq, false, true, true);
2852 blk_mq_run_dispatch_ops(rq->q,
2853 blk_mq_try_issue_directly(rq->mq_hctx, rq));
2857 * blk_cloned_rq_check_limits - Helper function to check a cloned request
2858 * for the new queue limits
2860 * @rq: the request being checked
2863 * @rq may have been made based on weaker limitations of upper-level queues
2864 * in request stacking drivers, and it may violate the limitation of @q.
2865 * Since the block layer and the underlying device driver trust @rq
2866 * after it is inserted to @q, it should be checked against @q before
2867 * the insertion using this generic function.
2869 * Request stacking drivers like request-based dm may change the queue
2870 * limits when retrying requests on other queues. Those requests need
2871 * to be checked against the new queue limits again during dispatch.
2873 static blk_status_t blk_cloned_rq_check_limits(struct request_queue *q,
2876 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
2878 if (blk_rq_sectors(rq) > max_sectors) {
2880 * SCSI device does not have a good way to return if
2881 * Write Same/Zero is actually supported. If a device rejects
2882 * a non-read/write command (discard, write same,etc.) the
2883 * low-level device driver will set the relevant queue limit to
2884 * 0 to prevent blk-lib from issuing more of the offending
2885 * operations. Commands queued prior to the queue limit being
2886 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
2887 * errors being propagated to upper layers.
2889 if (max_sectors == 0)
2890 return BLK_STS_NOTSUPP;
2892 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
2893 __func__, blk_rq_sectors(rq), max_sectors);
2894 return BLK_STS_IOERR;
2898 * The queue settings related to segment counting may differ from the
2901 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
2902 if (rq->nr_phys_segments > queue_max_segments(q)) {
2903 printk(KERN_ERR "%s: over max segments limit. (%hu > %hu)\n",
2904 __func__, rq->nr_phys_segments, queue_max_segments(q));
2905 return BLK_STS_IOERR;
2912 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
2913 * @q: the queue to submit the request
2914 * @rq: the request being queued
2916 blk_status_t blk_insert_cloned_request(struct request_queue *q, struct request *rq)
2920 ret = blk_cloned_rq_check_limits(q, rq);
2921 if (ret != BLK_STS_OK)
2925 should_fail_request(rq->q->disk->part0, blk_rq_bytes(rq)))
2926 return BLK_STS_IOERR;
2928 if (blk_crypto_insert_cloned_request(rq))
2929 return BLK_STS_IOERR;
2931 blk_account_io_start(rq);
2934 * Since we have a scheduler attached on the top device,
2935 * bypass a potential scheduler on the bottom device for
2938 blk_mq_run_dispatch_ops(rq->q,
2939 ret = blk_mq_request_issue_directly(rq, true));
2941 blk_account_io_done(rq, ktime_get_ns());
2944 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
2947 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
2948 * @rq: the clone request to be cleaned up
2951 * Free all bios in @rq for a cloned request.
2953 void blk_rq_unprep_clone(struct request *rq)
2957 while ((bio = rq->bio) != NULL) {
2958 rq->bio = bio->bi_next;
2963 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
2966 * blk_rq_prep_clone - Helper function to setup clone request
2967 * @rq: the request to be setup
2968 * @rq_src: original request to be cloned
2969 * @bs: bio_set that bios for clone are allocated from
2970 * @gfp_mask: memory allocation mask for bio
2971 * @bio_ctr: setup function to be called for each clone bio.
2972 * Returns %0 for success, non %0 for failure.
2973 * @data: private data to be passed to @bio_ctr
2976 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
2977 * Also, pages which the original bios are pointing to are not copied
2978 * and the cloned bios just point same pages.
2979 * So cloned bios must be completed before original bios, which means
2980 * the caller must complete @rq before @rq_src.
2982 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
2983 struct bio_set *bs, gfp_t gfp_mask,
2984 int (*bio_ctr)(struct bio *, struct bio *, void *),
2987 struct bio *bio, *bio_src;
2992 __rq_for_each_bio(bio_src, rq_src) {
2993 bio = bio_clone_fast(bio_src, gfp_mask, bs);
2996 bio->bi_bdev = rq->q->disk->part0;
2998 if (bio_ctr && bio_ctr(bio, bio_src, data))
3002 rq->biotail->bi_next = bio;
3005 rq->bio = rq->biotail = bio;
3010 /* Copy attributes of the original request to the clone request. */
3011 rq->__sector = blk_rq_pos(rq_src);
3012 rq->__data_len = blk_rq_bytes(rq_src);
3013 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3014 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3015 rq->special_vec = rq_src->special_vec;
3017 rq->nr_phys_segments = rq_src->nr_phys_segments;
3018 rq->ioprio = rq_src->ioprio;
3020 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3028 blk_rq_unprep_clone(rq);
3032 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3035 * Steal bios from a request and add them to a bio list.
3036 * The request must not have been partially completed before.
3038 void blk_steal_bios(struct bio_list *list, struct request *rq)
3042 list->tail->bi_next = rq->bio;
3044 list->head = rq->bio;
3045 list->tail = rq->biotail;
3053 EXPORT_SYMBOL_GPL(blk_steal_bios);
3055 static size_t order_to_size(unsigned int order)
3057 return (size_t)PAGE_SIZE << order;
3060 /* called before freeing request pool in @tags */
3061 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3062 struct blk_mq_tags *tags)
3065 unsigned long flags;
3067 /* There is no need to clear a driver tags own mapping */
3068 if (drv_tags == tags)
3071 list_for_each_entry(page, &tags->page_list, lru) {
3072 unsigned long start = (unsigned long)page_address(page);
3073 unsigned long end = start + order_to_size(page->private);
3076 for (i = 0; i < drv_tags->nr_tags; i++) {
3077 struct request *rq = drv_tags->rqs[i];
3078 unsigned long rq_addr = (unsigned long)rq;
3080 if (rq_addr >= start && rq_addr < end) {
3081 WARN_ON_ONCE(req_ref_read(rq) != 0);
3082 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3088 * Wait until all pending iteration is done.
3090 * Request reference is cleared and it is guaranteed to be observed
3091 * after the ->lock is released.
3093 spin_lock_irqsave(&drv_tags->lock, flags);
3094 spin_unlock_irqrestore(&drv_tags->lock, flags);
3097 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3098 unsigned int hctx_idx)
3100 struct blk_mq_tags *drv_tags;
3103 if (blk_mq_is_shared_tags(set->flags))
3104 drv_tags = set->shared_tags;
3106 drv_tags = set->tags[hctx_idx];
3108 if (tags->static_rqs && set->ops->exit_request) {
3111 for (i = 0; i < tags->nr_tags; i++) {
3112 struct request *rq = tags->static_rqs[i];
3116 set->ops->exit_request(set, rq, hctx_idx);
3117 tags->static_rqs[i] = NULL;
3121 blk_mq_clear_rq_mapping(drv_tags, tags);
3123 while (!list_empty(&tags->page_list)) {
3124 page = list_first_entry(&tags->page_list, struct page, lru);
3125 list_del_init(&page->lru);
3127 * Remove kmemleak object previously allocated in
3128 * blk_mq_alloc_rqs().
3130 kmemleak_free(page_address(page));
3131 __free_pages(page, page->private);
3135 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3139 kfree(tags->static_rqs);
3140 tags->static_rqs = NULL;
3142 blk_mq_free_tags(tags);
3145 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3146 unsigned int hctx_idx,
3147 unsigned int nr_tags,
3148 unsigned int reserved_tags)
3150 struct blk_mq_tags *tags;
3153 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
3154 if (node == NUMA_NO_NODE)
3155 node = set->numa_node;
3157 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3158 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3162 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3163 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3166 blk_mq_free_tags(tags);
3170 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3171 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3173 if (!tags->static_rqs) {
3175 blk_mq_free_tags(tags);
3182 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3183 unsigned int hctx_idx, int node)
3187 if (set->ops->init_request) {
3188 ret = set->ops->init_request(set, rq, hctx_idx, node);
3193 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3197 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3198 struct blk_mq_tags *tags,
3199 unsigned int hctx_idx, unsigned int depth)
3201 unsigned int i, j, entries_per_page, max_order = 4;
3202 size_t rq_size, left;
3205 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
3206 if (node == NUMA_NO_NODE)
3207 node = set->numa_node;
3209 INIT_LIST_HEAD(&tags->page_list);
3212 * rq_size is the size of the request plus driver payload, rounded
3213 * to the cacheline size
3215 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3217 left = rq_size * depth;
3219 for (i = 0; i < depth; ) {
3220 int this_order = max_order;
3225 while (this_order && left < order_to_size(this_order - 1))
3229 page = alloc_pages_node(node,
3230 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3236 if (order_to_size(this_order) < rq_size)
3243 page->private = this_order;
3244 list_add_tail(&page->lru, &tags->page_list);
3246 p = page_address(page);
3248 * Allow kmemleak to scan these pages as they contain pointers
3249 * to additional allocations like via ops->init_request().
3251 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3252 entries_per_page = order_to_size(this_order) / rq_size;
3253 to_do = min(entries_per_page, depth - i);
3254 left -= to_do * rq_size;
3255 for (j = 0; j < to_do; j++) {
3256 struct request *rq = p;
3258 tags->static_rqs[i] = rq;
3259 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3260 tags->static_rqs[i] = NULL;
3271 blk_mq_free_rqs(set, tags, hctx_idx);
3275 struct rq_iter_data {
3276 struct blk_mq_hw_ctx *hctx;
3280 static bool blk_mq_has_request(struct request *rq, void *data, bool reserved)
3282 struct rq_iter_data *iter_data = data;
3284 if (rq->mq_hctx != iter_data->hctx)
3286 iter_data->has_rq = true;
3290 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3292 struct blk_mq_tags *tags = hctx->sched_tags ?
3293 hctx->sched_tags : hctx->tags;
3294 struct rq_iter_data data = {
3298 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3302 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3303 struct blk_mq_hw_ctx *hctx)
3305 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3307 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3312 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3314 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3315 struct blk_mq_hw_ctx, cpuhp_online);
3317 if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3318 !blk_mq_last_cpu_in_hctx(cpu, hctx))
3322 * Prevent new request from being allocated on the current hctx.
3324 * The smp_mb__after_atomic() Pairs with the implied barrier in
3325 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3326 * seen once we return from the tag allocator.
3328 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3329 smp_mb__after_atomic();
3332 * Try to grab a reference to the queue and wait for any outstanding
3333 * requests. If we could not grab a reference the queue has been
3334 * frozen and there are no requests.
3336 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3337 while (blk_mq_hctx_has_requests(hctx))
3339 percpu_ref_put(&hctx->queue->q_usage_counter);
3345 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3347 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3348 struct blk_mq_hw_ctx, cpuhp_online);
3350 if (cpumask_test_cpu(cpu, hctx->cpumask))
3351 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3356 * 'cpu' is going away. splice any existing rq_list entries from this
3357 * software queue to the hw queue dispatch list, and ensure that it
3360 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3362 struct blk_mq_hw_ctx *hctx;
3363 struct blk_mq_ctx *ctx;
3365 enum hctx_type type;
3367 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3368 if (!cpumask_test_cpu(cpu, hctx->cpumask))
3371 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3374 spin_lock(&ctx->lock);
3375 if (!list_empty(&ctx->rq_lists[type])) {
3376 list_splice_init(&ctx->rq_lists[type], &tmp);
3377 blk_mq_hctx_clear_pending(hctx, ctx);
3379 spin_unlock(&ctx->lock);
3381 if (list_empty(&tmp))
3384 spin_lock(&hctx->lock);
3385 list_splice_tail_init(&tmp, &hctx->dispatch);
3386 spin_unlock(&hctx->lock);
3388 blk_mq_run_hw_queue(hctx, true);
3392 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3394 if (!(hctx->flags & BLK_MQ_F_STACKING))
3395 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3396 &hctx->cpuhp_online);
3397 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3402 * Before freeing hw queue, clearing the flush request reference in
3403 * tags->rqs[] for avoiding potential UAF.
3405 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3406 unsigned int queue_depth, struct request *flush_rq)
3409 unsigned long flags;
3411 /* The hw queue may not be mapped yet */
3415 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3417 for (i = 0; i < queue_depth; i++)
3418 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3421 * Wait until all pending iteration is done.
3423 * Request reference is cleared and it is guaranteed to be observed
3424 * after the ->lock is released.
3426 spin_lock_irqsave(&tags->lock, flags);
3427 spin_unlock_irqrestore(&tags->lock, flags);
3430 /* hctx->ctxs will be freed in queue's release handler */
3431 static void blk_mq_exit_hctx(struct request_queue *q,
3432 struct blk_mq_tag_set *set,
3433 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3435 struct request *flush_rq = hctx->fq->flush_rq;
3437 if (blk_mq_hw_queue_mapped(hctx))
3438 blk_mq_tag_idle(hctx);
3440 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3441 set->queue_depth, flush_rq);
3442 if (set->ops->exit_request)
3443 set->ops->exit_request(set, flush_rq, hctx_idx);
3445 if (set->ops->exit_hctx)
3446 set->ops->exit_hctx(hctx, hctx_idx);
3448 blk_mq_remove_cpuhp(hctx);
3450 spin_lock(&q->unused_hctx_lock);
3451 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3452 spin_unlock(&q->unused_hctx_lock);
3455 static void blk_mq_exit_hw_queues(struct request_queue *q,
3456 struct blk_mq_tag_set *set, int nr_queue)
3458 struct blk_mq_hw_ctx *hctx;
3461 queue_for_each_hw_ctx(q, hctx, i) {
3464 blk_mq_debugfs_unregister_hctx(hctx);
3465 blk_mq_exit_hctx(q, set, hctx, i);
3469 static int blk_mq_init_hctx(struct request_queue *q,
3470 struct blk_mq_tag_set *set,
3471 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3473 hctx->queue_num = hctx_idx;
3475 if (!(hctx->flags & BLK_MQ_F_STACKING))
3476 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3477 &hctx->cpuhp_online);
3478 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3480 hctx->tags = set->tags[hctx_idx];
3482 if (set->ops->init_hctx &&
3483 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3484 goto unregister_cpu_notifier;
3486 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3492 if (set->ops->exit_hctx)
3493 set->ops->exit_hctx(hctx, hctx_idx);
3494 unregister_cpu_notifier:
3495 blk_mq_remove_cpuhp(hctx);
3499 static struct blk_mq_hw_ctx *
3500 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3503 struct blk_mq_hw_ctx *hctx;
3504 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3506 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3508 goto fail_alloc_hctx;
3510 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3513 atomic_set(&hctx->nr_active, 0);
3514 if (node == NUMA_NO_NODE)
3515 node = set->numa_node;
3516 hctx->numa_node = node;
3518 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3519 spin_lock_init(&hctx->lock);
3520 INIT_LIST_HEAD(&hctx->dispatch);
3522 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3524 INIT_LIST_HEAD(&hctx->hctx_list);
3527 * Allocate space for all possible cpus to avoid allocation at
3530 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3535 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3536 gfp, node, false, false))
3540 spin_lock_init(&hctx->dispatch_wait_lock);
3541 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3542 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3544 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3548 blk_mq_hctx_kobj_init(hctx);
3553 sbitmap_free(&hctx->ctx_map);
3557 free_cpumask_var(hctx->cpumask);
3564 static void blk_mq_init_cpu_queues(struct request_queue *q,
3565 unsigned int nr_hw_queues)
3567 struct blk_mq_tag_set *set = q->tag_set;
3570 for_each_possible_cpu(i) {
3571 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3572 struct blk_mq_hw_ctx *hctx;
3576 spin_lock_init(&__ctx->lock);
3577 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3578 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3583 * Set local node, IFF we have more than one hw queue. If
3584 * not, we remain on the home node of the device
3586 for (j = 0; j < set->nr_maps; j++) {
3587 hctx = blk_mq_map_queue_type(q, j, i);
3588 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3589 hctx->numa_node = cpu_to_node(i);
3594 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3595 unsigned int hctx_idx,
3598 struct blk_mq_tags *tags;
3601 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3605 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3607 blk_mq_free_rq_map(tags);
3614 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3617 if (blk_mq_is_shared_tags(set->flags)) {
3618 set->tags[hctx_idx] = set->shared_tags;
3623 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3626 return set->tags[hctx_idx];
3629 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3630 struct blk_mq_tags *tags,
3631 unsigned int hctx_idx)
3634 blk_mq_free_rqs(set, tags, hctx_idx);
3635 blk_mq_free_rq_map(tags);
3639 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3640 unsigned int hctx_idx)
3642 if (!blk_mq_is_shared_tags(set->flags))
3643 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3645 set->tags[hctx_idx] = NULL;
3648 static void blk_mq_map_swqueue(struct request_queue *q)
3650 unsigned int i, j, hctx_idx;
3651 struct blk_mq_hw_ctx *hctx;
3652 struct blk_mq_ctx *ctx;
3653 struct blk_mq_tag_set *set = q->tag_set;
3655 queue_for_each_hw_ctx(q, hctx, i) {
3656 cpumask_clear(hctx->cpumask);
3658 hctx->dispatch_from = NULL;
3662 * Map software to hardware queues.
3664 * If the cpu isn't present, the cpu is mapped to first hctx.
3666 for_each_possible_cpu(i) {
3668 ctx = per_cpu_ptr(q->queue_ctx, i);
3669 for (j = 0; j < set->nr_maps; j++) {
3670 if (!set->map[j].nr_queues) {
3671 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3672 HCTX_TYPE_DEFAULT, i);
3675 hctx_idx = set->map[j].mq_map[i];
3676 /* unmapped hw queue can be remapped after CPU topo changed */
3677 if (!set->tags[hctx_idx] &&
3678 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3680 * If tags initialization fail for some hctx,
3681 * that hctx won't be brought online. In this
3682 * case, remap the current ctx to hctx[0] which
3683 * is guaranteed to always have tags allocated
3685 set->map[j].mq_map[i] = 0;
3688 hctx = blk_mq_map_queue_type(q, j, i);
3689 ctx->hctxs[j] = hctx;
3691 * If the CPU is already set in the mask, then we've
3692 * mapped this one already. This can happen if
3693 * devices share queues across queue maps.
3695 if (cpumask_test_cpu(i, hctx->cpumask))
3698 cpumask_set_cpu(i, hctx->cpumask);
3700 ctx->index_hw[hctx->type] = hctx->nr_ctx;
3701 hctx->ctxs[hctx->nr_ctx++] = ctx;
3704 * If the nr_ctx type overflows, we have exceeded the
3705 * amount of sw queues we can support.
3707 BUG_ON(!hctx->nr_ctx);
3710 for (; j < HCTX_MAX_TYPES; j++)
3711 ctx->hctxs[j] = blk_mq_map_queue_type(q,
3712 HCTX_TYPE_DEFAULT, i);
3715 queue_for_each_hw_ctx(q, hctx, i) {
3717 * If no software queues are mapped to this hardware queue,
3718 * disable it and free the request entries.
3720 if (!hctx->nr_ctx) {
3721 /* Never unmap queue 0. We need it as a
3722 * fallback in case of a new remap fails
3726 __blk_mq_free_map_and_rqs(set, i);
3732 hctx->tags = set->tags[i];
3733 WARN_ON(!hctx->tags);
3736 * Set the map size to the number of mapped software queues.
3737 * This is more accurate and more efficient than looping
3738 * over all possibly mapped software queues.
3740 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3743 * Initialize batch roundrobin counts
3745 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3746 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3751 * Caller needs to ensure that we're either frozen/quiesced, or that
3752 * the queue isn't live yet.
3754 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3756 struct blk_mq_hw_ctx *hctx;
3759 queue_for_each_hw_ctx(q, hctx, i) {
3761 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3763 blk_mq_tag_idle(hctx);
3764 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3769 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3772 struct request_queue *q;
3774 lockdep_assert_held(&set->tag_list_lock);
3776 list_for_each_entry(q, &set->tag_list, tag_set_list) {
3777 blk_mq_freeze_queue(q);
3778 queue_set_hctx_shared(q, shared);
3779 blk_mq_unfreeze_queue(q);
3783 static void blk_mq_del_queue_tag_set(struct request_queue *q)
3785 struct blk_mq_tag_set *set = q->tag_set;
3787 mutex_lock(&set->tag_list_lock);
3788 list_del(&q->tag_set_list);
3789 if (list_is_singular(&set->tag_list)) {
3790 /* just transitioned to unshared */
3791 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3792 /* update existing queue */
3793 blk_mq_update_tag_set_shared(set, false);
3795 mutex_unlock(&set->tag_list_lock);
3796 INIT_LIST_HEAD(&q->tag_set_list);
3799 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3800 struct request_queue *q)
3802 mutex_lock(&set->tag_list_lock);
3805 * Check to see if we're transitioning to shared (from 1 to 2 queues).
3807 if (!list_empty(&set->tag_list) &&
3808 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3809 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3810 /* update existing queue */
3811 blk_mq_update_tag_set_shared(set, true);
3813 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3814 queue_set_hctx_shared(q, true);
3815 list_add_tail(&q->tag_set_list, &set->tag_list);
3817 mutex_unlock(&set->tag_list_lock);
3820 /* All allocations will be freed in release handler of q->mq_kobj */
3821 static int blk_mq_alloc_ctxs(struct request_queue *q)
3823 struct blk_mq_ctxs *ctxs;
3826 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3830 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3831 if (!ctxs->queue_ctx)
3834 for_each_possible_cpu(cpu) {
3835 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3839 q->mq_kobj = &ctxs->kobj;
3840 q->queue_ctx = ctxs->queue_ctx;
3849 * It is the actual release handler for mq, but we do it from
3850 * request queue's release handler for avoiding use-after-free
3851 * and headache because q->mq_kobj shouldn't have been introduced,
3852 * but we can't group ctx/kctx kobj without it.
3854 void blk_mq_release(struct request_queue *q)
3856 struct blk_mq_hw_ctx *hctx, *next;
3859 queue_for_each_hw_ctx(q, hctx, i)
3860 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
3862 /* all hctx are in .unused_hctx_list now */
3863 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
3864 list_del_init(&hctx->hctx_list);
3865 kobject_put(&hctx->kobj);
3868 kfree(q->queue_hw_ctx);
3871 * release .mq_kobj and sw queue's kobject now because
3872 * both share lifetime with request queue.
3874 blk_mq_sysfs_deinit(q);
3877 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
3880 struct request_queue *q;
3883 q = blk_alloc_queue(set->numa_node, set->flags & BLK_MQ_F_BLOCKING);
3885 return ERR_PTR(-ENOMEM);
3886 q->queuedata = queuedata;
3887 ret = blk_mq_init_allocated_queue(set, q);
3889 blk_cleanup_queue(q);
3890 return ERR_PTR(ret);
3895 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
3897 return blk_mq_init_queue_data(set, NULL);
3899 EXPORT_SYMBOL(blk_mq_init_queue);
3901 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
3902 struct lock_class_key *lkclass)
3904 struct request_queue *q;
3905 struct gendisk *disk;
3907 q = blk_mq_init_queue_data(set, queuedata);
3911 disk = __alloc_disk_node(q, set->numa_node, lkclass);
3913 blk_cleanup_queue(q);
3914 return ERR_PTR(-ENOMEM);
3918 EXPORT_SYMBOL(__blk_mq_alloc_disk);
3920 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
3921 struct blk_mq_tag_set *set, struct request_queue *q,
3922 int hctx_idx, int node)
3924 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
3926 /* reuse dead hctx first */
3927 spin_lock(&q->unused_hctx_lock);
3928 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
3929 if (tmp->numa_node == node) {
3935 list_del_init(&hctx->hctx_list);
3936 spin_unlock(&q->unused_hctx_lock);
3939 hctx = blk_mq_alloc_hctx(q, set, node);
3943 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
3949 kobject_put(&hctx->kobj);
3954 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
3955 struct request_queue *q)
3958 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
3960 if (q->nr_hw_queues < set->nr_hw_queues) {
3961 struct blk_mq_hw_ctx **new_hctxs;
3963 new_hctxs = kcalloc_node(set->nr_hw_queues,
3964 sizeof(*new_hctxs), GFP_KERNEL,
3969 memcpy(new_hctxs, hctxs, q->nr_hw_queues *
3971 q->queue_hw_ctx = new_hctxs;
3976 /* protect against switching io scheduler */
3977 mutex_lock(&q->sysfs_lock);
3978 for (i = 0; i < set->nr_hw_queues; i++) {
3980 struct blk_mq_hw_ctx *hctx;
3982 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
3984 * If the hw queue has been mapped to another numa node,
3985 * we need to realloc the hctx. If allocation fails, fallback
3986 * to use the previous one.
3988 if (hctxs[i] && (hctxs[i]->numa_node == node))
3991 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
3994 blk_mq_exit_hctx(q, set, hctxs[i], i);
3998 pr_warn("Allocate new hctx on node %d fails,\
3999 fallback to previous one on node %d\n",
4000 node, hctxs[i]->numa_node);
4006 * Increasing nr_hw_queues fails. Free the newly allocated
4007 * hctxs and keep the previous q->nr_hw_queues.
4009 if (i != set->nr_hw_queues) {
4010 j = q->nr_hw_queues;
4014 end = q->nr_hw_queues;
4015 q->nr_hw_queues = set->nr_hw_queues;
4018 for (; j < end; j++) {
4019 struct blk_mq_hw_ctx *hctx = hctxs[j];
4022 blk_mq_exit_hctx(q, set, hctx, j);
4026 mutex_unlock(&q->sysfs_lock);
4029 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4030 struct request_queue *q)
4032 WARN_ON_ONCE(blk_queue_has_srcu(q) !=
4033 !!(set->flags & BLK_MQ_F_BLOCKING));
4035 /* mark the queue as mq asap */
4036 q->mq_ops = set->ops;
4038 q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
4039 blk_mq_poll_stats_bkt,
4040 BLK_MQ_POLL_STATS_BKTS, q);
4044 if (blk_mq_alloc_ctxs(q))
4047 /* init q->mq_kobj and sw queues' kobjects */
4048 blk_mq_sysfs_init(q);
4050 INIT_LIST_HEAD(&q->unused_hctx_list);
4051 spin_lock_init(&q->unused_hctx_lock);
4053 blk_mq_realloc_hw_ctxs(set, q);
4054 if (!q->nr_hw_queues)
4057 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4058 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4062 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4063 if (set->nr_maps > HCTX_TYPE_POLL &&
4064 set->map[HCTX_TYPE_POLL].nr_queues)
4065 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4067 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4068 INIT_LIST_HEAD(&q->requeue_list);
4069 spin_lock_init(&q->requeue_lock);
4071 q->nr_requests = set->queue_depth;
4074 * Default to classic polling
4076 q->poll_nsec = BLK_MQ_POLL_CLASSIC;
4078 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4079 blk_mq_add_queue_tag_set(set, q);
4080 blk_mq_map_swqueue(q);
4084 kfree(q->queue_hw_ctx);
4085 q->nr_hw_queues = 0;
4086 blk_mq_sysfs_deinit(q);
4088 blk_stat_free_callback(q->poll_cb);
4094 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4096 /* tags can _not_ be used after returning from blk_mq_exit_queue */
4097 void blk_mq_exit_queue(struct request_queue *q)
4099 struct blk_mq_tag_set *set = q->tag_set;
4101 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4102 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4103 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4104 blk_mq_del_queue_tag_set(q);
4107 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4111 if (blk_mq_is_shared_tags(set->flags)) {
4112 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4115 if (!set->shared_tags)
4119 for (i = 0; i < set->nr_hw_queues; i++) {
4120 if (!__blk_mq_alloc_map_and_rqs(set, i))
4129 __blk_mq_free_map_and_rqs(set, i);
4131 if (blk_mq_is_shared_tags(set->flags)) {
4132 blk_mq_free_map_and_rqs(set, set->shared_tags,
4133 BLK_MQ_NO_HCTX_IDX);
4140 * Allocate the request maps associated with this tag_set. Note that this
4141 * may reduce the depth asked for, if memory is tight. set->queue_depth
4142 * will be updated to reflect the allocated depth.
4144 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4149 depth = set->queue_depth;
4151 err = __blk_mq_alloc_rq_maps(set);
4155 set->queue_depth >>= 1;
4156 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4160 } while (set->queue_depth);
4162 if (!set->queue_depth || err) {
4163 pr_err("blk-mq: failed to allocate request map\n");
4167 if (depth != set->queue_depth)
4168 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4169 depth, set->queue_depth);
4174 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4177 * blk_mq_map_queues() and multiple .map_queues() implementations
4178 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4179 * number of hardware queues.
4181 if (set->nr_maps == 1)
4182 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4184 if (set->ops->map_queues && !is_kdump_kernel()) {
4188 * transport .map_queues is usually done in the following
4191 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4192 * mask = get_cpu_mask(queue)
4193 * for_each_cpu(cpu, mask)
4194 * set->map[x].mq_map[cpu] = queue;
4197 * When we need to remap, the table has to be cleared for
4198 * killing stale mapping since one CPU may not be mapped
4201 for (i = 0; i < set->nr_maps; i++)
4202 blk_mq_clear_mq_map(&set->map[i]);
4204 return set->ops->map_queues(set);
4206 BUG_ON(set->nr_maps > 1);
4207 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4211 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4212 int cur_nr_hw_queues, int new_nr_hw_queues)
4214 struct blk_mq_tags **new_tags;
4216 if (cur_nr_hw_queues >= new_nr_hw_queues)
4219 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4220 GFP_KERNEL, set->numa_node);
4225 memcpy(new_tags, set->tags, cur_nr_hw_queues *
4226 sizeof(*set->tags));
4228 set->tags = new_tags;
4229 set->nr_hw_queues = new_nr_hw_queues;
4234 static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set,
4235 int new_nr_hw_queues)
4237 return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues);
4241 * Alloc a tag set to be associated with one or more request queues.
4242 * May fail with EINVAL for various error conditions. May adjust the
4243 * requested depth down, if it's too large. In that case, the set
4244 * value will be stored in set->queue_depth.
4246 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4250 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4252 if (!set->nr_hw_queues)
4254 if (!set->queue_depth)
4256 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4259 if (!set->ops->queue_rq)
4262 if (!set->ops->get_budget ^ !set->ops->put_budget)
4265 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4266 pr_info("blk-mq: reduced tag depth to %u\n",
4268 set->queue_depth = BLK_MQ_MAX_DEPTH;
4273 else if (set->nr_maps > HCTX_MAX_TYPES)
4277 * If a crashdump is active, then we are potentially in a very
4278 * memory constrained environment. Limit us to 1 queue and
4279 * 64 tags to prevent using too much memory.
4281 if (is_kdump_kernel()) {
4282 set->nr_hw_queues = 1;
4284 set->queue_depth = min(64U, set->queue_depth);
4287 * There is no use for more h/w queues than cpus if we just have
4290 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4291 set->nr_hw_queues = nr_cpu_ids;
4293 if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0)
4297 for (i = 0; i < set->nr_maps; i++) {
4298 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4299 sizeof(set->map[i].mq_map[0]),
4300 GFP_KERNEL, set->numa_node);
4301 if (!set->map[i].mq_map)
4302 goto out_free_mq_map;
4303 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4306 ret = blk_mq_update_queue_map(set);
4308 goto out_free_mq_map;
4310 ret = blk_mq_alloc_set_map_and_rqs(set);
4312 goto out_free_mq_map;
4314 mutex_init(&set->tag_list_lock);
4315 INIT_LIST_HEAD(&set->tag_list);
4320 for (i = 0; i < set->nr_maps; i++) {
4321 kfree(set->map[i].mq_map);
4322 set->map[i].mq_map = NULL;
4328 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4330 /* allocate and initialize a tagset for a simple single-queue device */
4331 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4332 const struct blk_mq_ops *ops, unsigned int queue_depth,
4333 unsigned int set_flags)
4335 memset(set, 0, sizeof(*set));
4337 set->nr_hw_queues = 1;
4339 set->queue_depth = queue_depth;
4340 set->numa_node = NUMA_NO_NODE;
4341 set->flags = set_flags;
4342 return blk_mq_alloc_tag_set(set);
4344 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4346 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4350 for (i = 0; i < set->nr_hw_queues; i++)
4351 __blk_mq_free_map_and_rqs(set, i);
4353 if (blk_mq_is_shared_tags(set->flags)) {
4354 blk_mq_free_map_and_rqs(set, set->shared_tags,
4355 BLK_MQ_NO_HCTX_IDX);
4358 for (j = 0; j < set->nr_maps; j++) {
4359 kfree(set->map[j].mq_map);
4360 set->map[j].mq_map = NULL;
4366 EXPORT_SYMBOL(blk_mq_free_tag_set);
4368 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4370 struct blk_mq_tag_set *set = q->tag_set;
4371 struct blk_mq_hw_ctx *hctx;
4377 if (q->nr_requests == nr)
4380 blk_mq_freeze_queue(q);
4381 blk_mq_quiesce_queue(q);
4384 queue_for_each_hw_ctx(q, hctx, i) {
4388 * If we're using an MQ scheduler, just update the scheduler
4389 * queue depth. This is similar to what the old code would do.
4391 if (hctx->sched_tags) {
4392 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4395 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4400 if (q->elevator && q->elevator->type->ops.depth_updated)
4401 q->elevator->type->ops.depth_updated(hctx);
4404 q->nr_requests = nr;
4405 if (blk_mq_is_shared_tags(set->flags)) {
4407 blk_mq_tag_update_sched_shared_tags(q);
4409 blk_mq_tag_resize_shared_tags(set, nr);
4413 blk_mq_unquiesce_queue(q);
4414 blk_mq_unfreeze_queue(q);
4420 * request_queue and elevator_type pair.
4421 * It is just used by __blk_mq_update_nr_hw_queues to cache
4422 * the elevator_type associated with a request_queue.
4424 struct blk_mq_qe_pair {
4425 struct list_head node;
4426 struct request_queue *q;
4427 struct elevator_type *type;
4431 * Cache the elevator_type in qe pair list and switch the
4432 * io scheduler to 'none'
4434 static bool blk_mq_elv_switch_none(struct list_head *head,
4435 struct request_queue *q)
4437 struct blk_mq_qe_pair *qe;
4442 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4446 INIT_LIST_HEAD(&qe->node);
4448 qe->type = q->elevator->type;
4449 list_add(&qe->node, head);
4451 mutex_lock(&q->sysfs_lock);
4453 * After elevator_switch_mq, the previous elevator_queue will be
4454 * released by elevator_release. The reference of the io scheduler
4455 * module get by elevator_get will also be put. So we need to get
4456 * a reference of the io scheduler module here to prevent it to be
4459 __module_get(qe->type->elevator_owner);
4460 elevator_switch_mq(q, NULL);
4461 mutex_unlock(&q->sysfs_lock);
4466 static void blk_mq_elv_switch_back(struct list_head *head,
4467 struct request_queue *q)
4469 struct blk_mq_qe_pair *qe;
4470 struct elevator_type *t = NULL;
4472 list_for_each_entry(qe, head, node)
4481 list_del(&qe->node);
4484 mutex_lock(&q->sysfs_lock);
4485 elevator_switch_mq(q, t);
4486 mutex_unlock(&q->sysfs_lock);
4489 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4492 struct request_queue *q;
4494 int prev_nr_hw_queues;
4496 lockdep_assert_held(&set->tag_list_lock);
4498 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4499 nr_hw_queues = nr_cpu_ids;
4500 if (nr_hw_queues < 1)
4502 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4505 list_for_each_entry(q, &set->tag_list, tag_set_list)
4506 blk_mq_freeze_queue(q);
4508 * Switch IO scheduler to 'none', cleaning up the data associated
4509 * with the previous scheduler. We will switch back once we are done
4510 * updating the new sw to hw queue mappings.
4512 list_for_each_entry(q, &set->tag_list, tag_set_list)
4513 if (!blk_mq_elv_switch_none(&head, q))
4516 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4517 blk_mq_debugfs_unregister_hctxs(q);
4518 blk_mq_sysfs_unregister(q);
4521 prev_nr_hw_queues = set->nr_hw_queues;
4522 if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
4526 set->nr_hw_queues = nr_hw_queues;
4528 blk_mq_update_queue_map(set);
4529 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4530 blk_mq_realloc_hw_ctxs(set, q);
4531 if (q->nr_hw_queues != set->nr_hw_queues) {
4532 int i = prev_nr_hw_queues;
4534 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4535 nr_hw_queues, prev_nr_hw_queues);
4536 for (; i < set->nr_hw_queues; i++)
4537 __blk_mq_free_map_and_rqs(set, i);
4539 set->nr_hw_queues = prev_nr_hw_queues;
4540 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4543 blk_mq_map_swqueue(q);
4547 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4548 blk_mq_sysfs_register(q);
4549 blk_mq_debugfs_register_hctxs(q);
4553 list_for_each_entry(q, &set->tag_list, tag_set_list)
4554 blk_mq_elv_switch_back(&head, q);
4556 list_for_each_entry(q, &set->tag_list, tag_set_list)
4557 blk_mq_unfreeze_queue(q);
4560 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4562 mutex_lock(&set->tag_list_lock);
4563 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4564 mutex_unlock(&set->tag_list_lock);
4566 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4568 /* Enable polling stats and return whether they were already enabled. */
4569 static bool blk_poll_stats_enable(struct request_queue *q)
4574 return blk_stats_alloc_enable(q);
4577 static void blk_mq_poll_stats_start(struct request_queue *q)
4580 * We don't arm the callback if polling stats are not enabled or the
4581 * callback is already active.
4583 if (!q->poll_stat || blk_stat_is_active(q->poll_cb))
4586 blk_stat_activate_msecs(q->poll_cb, 100);
4589 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
4591 struct request_queue *q = cb->data;
4594 for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
4595 if (cb->stat[bucket].nr_samples)
4596 q->poll_stat[bucket] = cb->stat[bucket];
4600 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
4603 unsigned long ret = 0;
4607 * If stats collection isn't on, don't sleep but turn it on for
4610 if (!blk_poll_stats_enable(q))
4614 * As an optimistic guess, use half of the mean service time
4615 * for this type of request. We can (and should) make this smarter.
4616 * For instance, if the completion latencies are tight, we can
4617 * get closer than just half the mean. This is especially
4618 * important on devices where the completion latencies are longer
4619 * than ~10 usec. We do use the stats for the relevant IO size
4620 * if available which does lead to better estimates.
4622 bucket = blk_mq_poll_stats_bkt(rq);
4626 if (q->poll_stat[bucket].nr_samples)
4627 ret = (q->poll_stat[bucket].mean + 1) / 2;
4632 static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
4634 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc);
4635 struct request *rq = blk_qc_to_rq(hctx, qc);
4636 struct hrtimer_sleeper hs;
4637 enum hrtimer_mode mode;
4642 * If a request has completed on queue that uses an I/O scheduler, we
4643 * won't get back a request from blk_qc_to_rq.
4645 if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT))
4649 * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
4651 * 0: use half of prev avg
4652 * >0: use this specific value
4654 if (q->poll_nsec > 0)
4655 nsecs = q->poll_nsec;
4657 nsecs = blk_mq_poll_nsecs(q, rq);
4662 rq->rq_flags |= RQF_MQ_POLL_SLEPT;
4665 * This will be replaced with the stats tracking code, using
4666 * 'avg_completion_time / 2' as the pre-sleep target.
4670 mode = HRTIMER_MODE_REL;
4671 hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
4672 hrtimer_set_expires(&hs.timer, kt);
4675 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
4677 set_current_state(TASK_UNINTERRUPTIBLE);
4678 hrtimer_sleeper_start_expires(&hs, mode);
4681 hrtimer_cancel(&hs.timer);
4682 mode = HRTIMER_MODE_ABS;
4683 } while (hs.task && !signal_pending(current));
4685 __set_current_state(TASK_RUNNING);
4686 destroy_hrtimer_on_stack(&hs.timer);
4689 * If we sleep, have the caller restart the poll loop to reset the
4690 * state. Like for the other success return cases, the caller is
4691 * responsible for checking if the IO completed. If the IO isn't
4692 * complete, we'll get called again and will go straight to the busy
4698 static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
4699 struct io_comp_batch *iob, unsigned int flags)
4701 struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
4702 long state = get_current_state();
4706 ret = q->mq_ops->poll(hctx, iob);
4708 __set_current_state(TASK_RUNNING);
4712 if (signal_pending_state(state, current))
4713 __set_current_state(TASK_RUNNING);
4714 if (task_is_running(current))
4717 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4720 } while (!need_resched());
4722 __set_current_state(TASK_RUNNING);
4726 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
4729 if (!(flags & BLK_POLL_NOSLEEP) &&
4730 q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
4731 if (blk_mq_poll_hybrid(q, cookie))
4734 return blk_mq_poll_classic(q, cookie, iob, flags);
4737 unsigned int blk_mq_rq_cpu(struct request *rq)
4739 return rq->mq_ctx->cpu;
4741 EXPORT_SYMBOL(blk_mq_rq_cpu);
4743 void blk_mq_cancel_work_sync(struct request_queue *q)
4745 if (queue_is_mq(q)) {
4746 struct blk_mq_hw_ctx *hctx;
4749 cancel_delayed_work_sync(&q->requeue_work);
4751 queue_for_each_hw_ctx(q, hctx, i)
4752 cancel_delayed_work_sync(&hctx->run_work);
4756 static int __init blk_mq_init(void)
4760 for_each_possible_cpu(i)
4761 init_llist_head(&per_cpu(blk_cpu_done, i));
4762 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4764 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4765 "block/softirq:dead", NULL,
4766 blk_softirq_cpu_dead);
4767 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4768 blk_mq_hctx_notify_dead);
4769 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4770 blk_mq_hctx_notify_online,
4771 blk_mq_hctx_notify_offline);
4774 subsys_initcall(blk_mq_init);